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 <linux/export.h>
39 #include <xen/swiotlb-xen.h>
41 #include <xen/xen-ops.h>
42 #include <xen/hvc-console.h>
44 * Used to do a quick range check in swiotlb_tbl_unmap_single and
45 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
49 static char *xen_io_tlb_start
, *xen_io_tlb_end
;
50 static unsigned long xen_io_tlb_nslabs
;
52 * Quick lookup value of the bus address of the IOTLB.
55 static u64 start_dma_addr
;
57 static dma_addr_t
xen_phys_to_bus(phys_addr_t paddr
)
59 return phys_to_machine(XPADDR(paddr
)).maddr
;
62 static phys_addr_t
xen_bus_to_phys(dma_addr_t baddr
)
64 return machine_to_phys(XMADDR(baddr
)).paddr
;
67 static dma_addr_t
xen_virt_to_bus(void *address
)
69 return xen_phys_to_bus(virt_to_phys(address
));
72 static int check_pages_physically_contiguous(unsigned long pfn
,
76 unsigned long next_mfn
;
80 next_mfn
= pfn_to_mfn(pfn
);
81 nr_pages
= (offset
+ length
+ PAGE_SIZE
-1) >> PAGE_SHIFT
;
83 for (i
= 1; i
< nr_pages
; i
++) {
84 if (pfn_to_mfn(++pfn
) != ++next_mfn
)
90 static int range_straddles_page_boundary(phys_addr_t p
, size_t size
)
92 unsigned long pfn
= PFN_DOWN(p
);
93 unsigned int offset
= p
& ~PAGE_MASK
;
95 if (offset
+ size
<= PAGE_SIZE
)
97 if (check_pages_physically_contiguous(pfn
, offset
, size
))
102 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr
)
104 unsigned long mfn
= PFN_DOWN(dma_addr
);
105 unsigned long pfn
= mfn_to_local_pfn(mfn
);
108 /* If the address is outside our domain, it CAN
109 * have the same virtual address as another address
110 * in our domain. Therefore _only_ check address within our domain.
112 if (pfn_valid(pfn
)) {
113 paddr
= PFN_PHYS(pfn
);
114 return paddr
>= virt_to_phys(xen_io_tlb_start
) &&
115 paddr
< virt_to_phys(xen_io_tlb_end
);
120 static int max_dma_bits
= 32;
123 xen_swiotlb_fixup(void *buf
, size_t size
, unsigned long nslabs
)
128 dma_bits
= get_order(IO_TLB_SEGSIZE
<< IO_TLB_SHIFT
) + PAGE_SHIFT
;
132 int slabs
= min(nslabs
- i
, (unsigned long)IO_TLB_SEGSIZE
);
135 rc
= xen_create_contiguous_region(
136 (unsigned long)buf
+ (i
<< IO_TLB_SHIFT
),
137 get_order(slabs
<< IO_TLB_SHIFT
),
139 } while (rc
&& dma_bits
++ < max_dma_bits
);
144 } while (i
< nslabs
);
147 static unsigned long xen_set_nslabs(unsigned long nr_tbl
)
150 xen_io_tlb_nslabs
= (64 * 1024 * 1024 >> IO_TLB_SHIFT
);
151 xen_io_tlb_nslabs
= ALIGN(xen_io_tlb_nslabs
, IO_TLB_SEGSIZE
);
153 xen_io_tlb_nslabs
= nr_tbl
;
155 return xen_io_tlb_nslabs
<< IO_TLB_SHIFT
;
158 enum xen_swiotlb_err
{
159 XEN_SWIOTLB_UNKNOWN
= 0,
164 static const char *xen_swiotlb_error(enum xen_swiotlb_err err
)
167 case XEN_SWIOTLB_ENOMEM
:
168 return "Cannot allocate Xen-SWIOTLB buffer\n";
169 case XEN_SWIOTLB_EFIXUP
:
170 return "Failed to get contiguous memory for DMA from Xen!\n"\
171 "You either: don't have the permissions, do not have"\
172 " enough free memory under 4GB, or the hypervisor memory"\
173 " is too fragmented!";
179 int __ref
xen_swiotlb_init(int verbose
, bool early
)
181 unsigned long bytes
, order
;
183 enum xen_swiotlb_err m_ret
= XEN_SWIOTLB_UNKNOWN
;
184 unsigned int repeat
= 3;
186 xen_io_tlb_nslabs
= swiotlb_nr_tbl();
188 bytes
= xen_set_nslabs(xen_io_tlb_nslabs
);
189 order
= get_order(xen_io_tlb_nslabs
<< IO_TLB_SHIFT
);
191 * Get IO TLB memory from any location.
194 xen_io_tlb_start
= alloc_bootmem_pages(PAGE_ALIGN(bytes
));
196 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
197 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
198 while ((SLABS_PER_PAGE
<< order
) > IO_TLB_MIN_SLABS
) {
199 xen_io_tlb_start
= (void *)__get_free_pages(__GFP_NOWARN
, order
);
200 if (xen_io_tlb_start
)
204 if (order
!= get_order(bytes
)) {
205 pr_warn("Warning: only able to allocate %ld MB "
206 "for software IO TLB\n", (PAGE_SIZE
<< order
) >> 20);
207 xen_io_tlb_nslabs
= SLABS_PER_PAGE
<< order
;
208 bytes
= xen_io_tlb_nslabs
<< IO_TLB_SHIFT
;
211 if (!xen_io_tlb_start
) {
212 m_ret
= XEN_SWIOTLB_ENOMEM
;
215 xen_io_tlb_end
= xen_io_tlb_start
+ bytes
;
217 * And replace that memory with pages under 4GB.
219 rc
= xen_swiotlb_fixup(xen_io_tlb_start
,
224 free_bootmem(__pa(xen_io_tlb_start
), PAGE_ALIGN(bytes
));
226 free_pages((unsigned long)xen_io_tlb_start
, order
);
227 xen_io_tlb_start
= NULL
;
229 m_ret
= XEN_SWIOTLB_EFIXUP
;
232 start_dma_addr
= xen_virt_to_bus(xen_io_tlb_start
);
234 swiotlb_init_with_tbl(xen_io_tlb_start
, xen_io_tlb_nslabs
, verbose
);
237 rc
= swiotlb_late_init_with_tbl(xen_io_tlb_start
, xen_io_tlb_nslabs
);
241 xen_io_tlb_nslabs
= max(1024UL, /* Min is 2MB */
242 (xen_io_tlb_nslabs
>> 1));
243 printk(KERN_INFO
"Xen-SWIOTLB: Lowering to %luMB\n",
244 (xen_io_tlb_nslabs
<< IO_TLB_SHIFT
) >> 20);
247 pr_err("%s (rc:%d)", xen_swiotlb_error(m_ret
), rc
);
249 panic("%s (rc:%d)", xen_swiotlb_error(m_ret
), rc
);
251 free_pages((unsigned long)xen_io_tlb_start
, order
);
255 xen_swiotlb_alloc_coherent(struct device
*hwdev
, size_t size
,
256 dma_addr_t
*dma_handle
, gfp_t flags
,
257 struct dma_attrs
*attrs
)
260 int order
= get_order(size
);
261 u64 dma_mask
= DMA_BIT_MASK(32);
262 unsigned long vstart
;
267 * Ignore region specifiers - the kernel's ideas of
268 * pseudo-phys memory layout has nothing to do with the
269 * machine physical layout. We can't allocate highmem
270 * because we can't return a pointer to it.
272 flags
&= ~(__GFP_DMA
| __GFP_HIGHMEM
);
274 if (dma_alloc_from_coherent(hwdev
, size
, dma_handle
, &ret
))
277 vstart
= __get_free_pages(flags
, order
);
278 ret
= (void *)vstart
;
283 if (hwdev
&& hwdev
->coherent_dma_mask
)
284 dma_mask
= dma_alloc_coherent_mask(hwdev
, flags
);
286 phys
= virt_to_phys(ret
);
287 dev_addr
= xen_phys_to_bus(phys
);
288 if (((dev_addr
+ size
- 1 <= dma_mask
)) &&
289 !range_straddles_page_boundary(phys
, size
))
290 *dma_handle
= dev_addr
;
292 if (xen_create_contiguous_region(vstart
, order
,
293 fls64(dma_mask
)) != 0) {
294 free_pages(vstart
, order
);
297 *dma_handle
= virt_to_machine(ret
).maddr
;
299 memset(ret
, 0, size
);
302 EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent
);
305 xen_swiotlb_free_coherent(struct device
*hwdev
, size_t size
, void *vaddr
,
306 dma_addr_t dev_addr
, struct dma_attrs
*attrs
)
308 int order
= get_order(size
);
310 u64 dma_mask
= DMA_BIT_MASK(32);
312 if (dma_release_from_coherent(hwdev
, order
, vaddr
))
315 if (hwdev
&& hwdev
->coherent_dma_mask
)
316 dma_mask
= hwdev
->coherent_dma_mask
;
318 phys
= virt_to_phys(vaddr
);
320 if (((dev_addr
+ size
- 1 > dma_mask
)) ||
321 range_straddles_page_boundary(phys
, size
))
322 xen_destroy_contiguous_region((unsigned long)vaddr
, order
);
324 free_pages((unsigned long)vaddr
, order
);
326 EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent
);
330 * Map a single buffer of the indicated size for DMA in streaming mode. The
331 * physical address to use is returned.
333 * Once the device is given the dma address, the device owns this memory until
334 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
336 dma_addr_t
xen_swiotlb_map_page(struct device
*dev
, struct page
*page
,
337 unsigned long offset
, size_t size
,
338 enum dma_data_direction dir
,
339 struct dma_attrs
*attrs
)
341 phys_addr_t phys
= page_to_phys(page
) + offset
;
342 dma_addr_t dev_addr
= xen_phys_to_bus(phys
);
345 BUG_ON(dir
== DMA_NONE
);
347 * If the address happens to be in the device's DMA window,
348 * we can safely return the device addr and not worry about bounce
351 if (dma_capable(dev
, dev_addr
, size
) &&
352 !range_straddles_page_boundary(phys
, size
) && !swiotlb_force
)
356 * Oh well, have to allocate and map a bounce buffer.
358 map
= swiotlb_tbl_map_single(dev
, start_dma_addr
, phys
, size
, dir
);
360 return DMA_ERROR_CODE
;
362 dev_addr
= xen_virt_to_bus(map
);
365 * Ensure that the address returned is DMA'ble
367 if (!dma_capable(dev
, dev_addr
, size
)) {
368 swiotlb_tbl_unmap_single(dev
, map
, size
, dir
);
373 EXPORT_SYMBOL_GPL(xen_swiotlb_map_page
);
376 * Unmap a single streaming mode DMA translation. The dma_addr and size must
377 * match what was provided for in a previous xen_swiotlb_map_page call. All
378 * other usages are undefined.
380 * After this call, reads by the cpu to the buffer are guaranteed to see
381 * whatever the device wrote there.
383 static void xen_unmap_single(struct device
*hwdev
, dma_addr_t dev_addr
,
384 size_t size
, enum dma_data_direction dir
)
386 phys_addr_t paddr
= xen_bus_to_phys(dev_addr
);
388 BUG_ON(dir
== DMA_NONE
);
390 /* NOTE: We use dev_addr here, not paddr! */
391 if (is_xen_swiotlb_buffer(dev_addr
)) {
392 swiotlb_tbl_unmap_single(hwdev
, phys_to_virt(paddr
), size
, dir
);
396 if (dir
!= DMA_FROM_DEVICE
)
400 * phys_to_virt doesn't work with hihgmem page but we could
401 * call dma_mark_clean() with hihgmem page here. However, we
402 * are fine since dma_mark_clean() is null on POWERPC. We can
403 * make dma_mark_clean() take a physical address if necessary.
405 dma_mark_clean(phys_to_virt(paddr
), size
);
408 void xen_swiotlb_unmap_page(struct device
*hwdev
, dma_addr_t dev_addr
,
409 size_t size
, enum dma_data_direction dir
,
410 struct dma_attrs
*attrs
)
412 xen_unmap_single(hwdev
, dev_addr
, size
, dir
);
414 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page
);
417 * Make physical memory consistent for a single streaming mode DMA translation
420 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
421 * using the cpu, yet do not wish to teardown the dma mapping, you must
422 * call this function before doing so. At the next point you give the dma
423 * address back to the card, you must first perform a
424 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
427 xen_swiotlb_sync_single(struct device
*hwdev
, dma_addr_t dev_addr
,
428 size_t size
, enum dma_data_direction dir
,
429 enum dma_sync_target target
)
431 phys_addr_t paddr
= xen_bus_to_phys(dev_addr
);
433 BUG_ON(dir
== DMA_NONE
);
435 /* NOTE: We use dev_addr here, not paddr! */
436 if (is_xen_swiotlb_buffer(dev_addr
)) {
437 swiotlb_tbl_sync_single(hwdev
, phys_to_virt(paddr
), size
, dir
,
442 if (dir
!= DMA_FROM_DEVICE
)
445 dma_mark_clean(phys_to_virt(paddr
), size
);
449 xen_swiotlb_sync_single_for_cpu(struct device
*hwdev
, dma_addr_t dev_addr
,
450 size_t size
, enum dma_data_direction dir
)
452 xen_swiotlb_sync_single(hwdev
, dev_addr
, size
, dir
, SYNC_FOR_CPU
);
454 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu
);
457 xen_swiotlb_sync_single_for_device(struct device
*hwdev
, dma_addr_t dev_addr
,
458 size_t size
, enum dma_data_direction dir
)
460 xen_swiotlb_sync_single(hwdev
, dev_addr
, size
, dir
, SYNC_FOR_DEVICE
);
462 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device
);
465 * Map a set of buffers described by scatterlist in streaming mode for DMA.
466 * This is the scatter-gather version of the above xen_swiotlb_map_page
467 * interface. Here the scatter gather list elements are each tagged with the
468 * appropriate dma address and length. They are obtained via
469 * sg_dma_{address,length}(SG).
471 * NOTE: An implementation may be able to use a smaller number of
472 * DMA address/length pairs than there are SG table elements.
473 * (for example via virtual mapping capabilities)
474 * The routine returns the number of addr/length pairs actually
475 * used, at most nents.
477 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
481 xen_swiotlb_map_sg_attrs(struct device
*hwdev
, struct scatterlist
*sgl
,
482 int nelems
, enum dma_data_direction dir
,
483 struct dma_attrs
*attrs
)
485 struct scatterlist
*sg
;
488 BUG_ON(dir
== DMA_NONE
);
490 for_each_sg(sgl
, sg
, nelems
, i
) {
491 phys_addr_t paddr
= sg_phys(sg
);
492 dma_addr_t dev_addr
= xen_phys_to_bus(paddr
);
495 !dma_capable(hwdev
, dev_addr
, sg
->length
) ||
496 range_straddles_page_boundary(paddr
, sg
->length
)) {
497 void *map
= swiotlb_tbl_map_single(hwdev
,
502 /* Don't panic here, we expect map_sg users
503 to do proper error handling. */
504 xen_swiotlb_unmap_sg_attrs(hwdev
, sgl
, i
, dir
,
506 sgl
[0].dma_length
= 0;
507 return DMA_ERROR_CODE
;
509 sg
->dma_address
= xen_virt_to_bus(map
);
511 sg
->dma_address
= dev_addr
;
512 sg
->dma_length
= sg
->length
;
516 EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs
);
519 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
520 * concerning calls here are the same as for swiotlb_unmap_page() above.
523 xen_swiotlb_unmap_sg_attrs(struct device
*hwdev
, struct scatterlist
*sgl
,
524 int nelems
, enum dma_data_direction dir
,
525 struct dma_attrs
*attrs
)
527 struct scatterlist
*sg
;
530 BUG_ON(dir
== DMA_NONE
);
532 for_each_sg(sgl
, sg
, nelems
, i
)
533 xen_unmap_single(hwdev
, sg
->dma_address
, sg
->dma_length
, dir
);
536 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs
);
539 * Make physical memory consistent for a set of streaming mode DMA translations
542 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
546 xen_swiotlb_sync_sg(struct device
*hwdev
, struct scatterlist
*sgl
,
547 int nelems
, enum dma_data_direction dir
,
548 enum dma_sync_target target
)
550 struct scatterlist
*sg
;
553 for_each_sg(sgl
, sg
, nelems
, i
)
554 xen_swiotlb_sync_single(hwdev
, sg
->dma_address
,
555 sg
->dma_length
, dir
, target
);
559 xen_swiotlb_sync_sg_for_cpu(struct device
*hwdev
, struct scatterlist
*sg
,
560 int nelems
, enum dma_data_direction dir
)
562 xen_swiotlb_sync_sg(hwdev
, sg
, nelems
, dir
, SYNC_FOR_CPU
);
564 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu
);
567 xen_swiotlb_sync_sg_for_device(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_DEVICE
);
572 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device
);
575 xen_swiotlb_dma_mapping_error(struct device
*hwdev
, dma_addr_t dma_addr
)
579 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error
);
582 * Return whether the given device DMA address mask can be supported
583 * properly. For example, if your device can only drive the low 24-bits
584 * during bus mastering, then you would pass 0x00ffffff as the mask to
588 xen_swiotlb_dma_supported(struct device
*hwdev
, u64 mask
)
590 return xen_virt_to_bus(xen_io_tlb_end
- 1) <= mask
;
592 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported
);