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 <asm/dma-mapping.h>
47 #include <asm/xen/page-coherent.h>
49 #include <trace/events/swiotlb.h>
51 * Used to do a quick range check in swiotlb_tbl_unmap_single and
52 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
57 static unsigned long dma_alloc_coherent_mask(struct device
*dev
,
60 unsigned long dma_mask
= 0;
62 dma_mask
= dev
->coherent_dma_mask
;
64 dma_mask
= (gfp
& GFP_DMA
) ? DMA_BIT_MASK(24) : DMA_BIT_MASK(32);
70 static char *xen_io_tlb_start
, *xen_io_tlb_end
;
71 static unsigned long xen_io_tlb_nslabs
;
73 * Quick lookup value of the bus address of the IOTLB.
76 static u64 start_dma_addr
;
79 * Both of these functions should avoid PFN_PHYS because phys_addr_t
80 * can be 32bit when dma_addr_t is 64bit leading to a loss in
81 * information if the shift is done before casting to 64bit.
83 static inline dma_addr_t
xen_phys_to_bus(phys_addr_t paddr
)
85 unsigned long mfn
= pfn_to_mfn(PFN_DOWN(paddr
));
86 dma_addr_t dma
= (dma_addr_t
)mfn
<< PAGE_SHIFT
;
88 dma
|= paddr
& ~PAGE_MASK
;
93 static inline phys_addr_t
xen_bus_to_phys(dma_addr_t baddr
)
95 unsigned long pfn
= mfn_to_pfn(PFN_DOWN(baddr
));
96 dma_addr_t dma
= (dma_addr_t
)pfn
<< PAGE_SHIFT
;
97 phys_addr_t paddr
= dma
;
99 BUG_ON(paddr
!= dma
); /* truncation has occurred, should never happen */
101 paddr
|= baddr
& ~PAGE_MASK
;
106 static inline dma_addr_t
xen_virt_to_bus(void *address
)
108 return xen_phys_to_bus(virt_to_phys(address
));
111 static int check_pages_physically_contiguous(unsigned long pfn
,
115 unsigned long next_mfn
;
119 next_mfn
= pfn_to_mfn(pfn
);
120 nr_pages
= (offset
+ length
+ PAGE_SIZE
-1) >> PAGE_SHIFT
;
122 for (i
= 1; i
< nr_pages
; i
++) {
123 if (pfn_to_mfn(++pfn
) != ++next_mfn
)
129 static inline int range_straddles_page_boundary(phys_addr_t p
, size_t size
)
131 unsigned long pfn
= PFN_DOWN(p
);
132 unsigned int offset
= p
& ~PAGE_MASK
;
134 if (offset
+ size
<= PAGE_SIZE
)
136 if (check_pages_physically_contiguous(pfn
, offset
, size
))
141 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr
)
143 unsigned long mfn
= PFN_DOWN(dma_addr
);
144 unsigned long pfn
= mfn_to_local_pfn(mfn
);
147 /* If the address is outside our domain, it CAN
148 * have the same virtual address as another address
149 * in our domain. Therefore _only_ check address within our domain.
151 if (pfn_valid(pfn
)) {
152 paddr
= PFN_PHYS(pfn
);
153 return paddr
>= virt_to_phys(xen_io_tlb_start
) &&
154 paddr
< virt_to_phys(xen_io_tlb_end
);
159 static int max_dma_bits
= 32;
162 xen_swiotlb_fixup(void *buf
, size_t size
, unsigned long nslabs
)
166 dma_addr_t dma_handle
;
167 phys_addr_t p
= virt_to_phys(buf
);
169 dma_bits
= get_order(IO_TLB_SEGSIZE
<< IO_TLB_SHIFT
) + PAGE_SHIFT
;
173 int slabs
= min(nslabs
- i
, (unsigned long)IO_TLB_SEGSIZE
);
176 rc
= xen_create_contiguous_region(
177 p
+ (i
<< IO_TLB_SHIFT
),
178 get_order(slabs
<< IO_TLB_SHIFT
),
179 dma_bits
, &dma_handle
);
180 } while (rc
&& dma_bits
++ < max_dma_bits
);
185 } while (i
< nslabs
);
188 static unsigned long xen_set_nslabs(unsigned long nr_tbl
)
191 xen_io_tlb_nslabs
= (64 * 1024 * 1024 >> IO_TLB_SHIFT
);
192 xen_io_tlb_nslabs
= ALIGN(xen_io_tlb_nslabs
, IO_TLB_SEGSIZE
);
194 xen_io_tlb_nslabs
= nr_tbl
;
196 return xen_io_tlb_nslabs
<< IO_TLB_SHIFT
;
199 enum xen_swiotlb_err
{
200 XEN_SWIOTLB_UNKNOWN
= 0,
205 static const char *xen_swiotlb_error(enum xen_swiotlb_err err
)
208 case XEN_SWIOTLB_ENOMEM
:
209 return "Cannot allocate Xen-SWIOTLB buffer\n";
210 case XEN_SWIOTLB_EFIXUP
:
211 return "Failed to get contiguous memory for DMA from Xen!\n"\
212 "You either: don't have the permissions, do not have"\
213 " enough free memory under 4GB, or the hypervisor memory"\
214 " is too fragmented!";
220 int __ref
xen_swiotlb_init(int verbose
, bool early
)
222 unsigned long bytes
, order
;
224 enum xen_swiotlb_err m_ret
= XEN_SWIOTLB_UNKNOWN
;
225 unsigned int repeat
= 3;
227 xen_io_tlb_nslabs
= swiotlb_nr_tbl();
229 bytes
= xen_set_nslabs(xen_io_tlb_nslabs
);
230 order
= get_order(xen_io_tlb_nslabs
<< IO_TLB_SHIFT
);
232 * Get IO TLB memory from any location.
235 xen_io_tlb_start
= alloc_bootmem_pages(PAGE_ALIGN(bytes
));
237 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
238 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
239 while ((SLABS_PER_PAGE
<< order
) > IO_TLB_MIN_SLABS
) {
240 xen_io_tlb_start
= (void *)__get_free_pages(__GFP_NOWARN
, order
);
241 if (xen_io_tlb_start
)
245 if (order
!= get_order(bytes
)) {
246 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
247 (PAGE_SIZE
<< order
) >> 20);
248 xen_io_tlb_nslabs
= SLABS_PER_PAGE
<< order
;
249 bytes
= xen_io_tlb_nslabs
<< IO_TLB_SHIFT
;
252 if (!xen_io_tlb_start
) {
253 m_ret
= XEN_SWIOTLB_ENOMEM
;
256 xen_io_tlb_end
= xen_io_tlb_start
+ bytes
;
258 * And replace that memory with pages under 4GB.
260 rc
= xen_swiotlb_fixup(xen_io_tlb_start
,
265 free_bootmem(__pa(xen_io_tlb_start
), PAGE_ALIGN(bytes
));
267 free_pages((unsigned long)xen_io_tlb_start
, order
);
268 xen_io_tlb_start
= NULL
;
270 m_ret
= XEN_SWIOTLB_EFIXUP
;
273 start_dma_addr
= xen_virt_to_bus(xen_io_tlb_start
);
275 if (swiotlb_init_with_tbl(xen_io_tlb_start
, xen_io_tlb_nslabs
,
277 panic("Cannot allocate SWIOTLB buffer");
280 rc
= swiotlb_late_init_with_tbl(xen_io_tlb_start
, xen_io_tlb_nslabs
);
284 xen_io_tlb_nslabs
= max(1024UL, /* Min is 2MB */
285 (xen_io_tlb_nslabs
>> 1));
286 pr_info("Lowering to %luMB\n",
287 (xen_io_tlb_nslabs
<< IO_TLB_SHIFT
) >> 20);
290 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret
), rc
);
292 panic("%s (rc:%d)", xen_swiotlb_error(m_ret
), rc
);
294 free_pages((unsigned long)xen_io_tlb_start
, order
);
298 xen_swiotlb_alloc_coherent(struct device
*hwdev
, size_t size
,
299 dma_addr_t
*dma_handle
, gfp_t flags
,
300 struct dma_attrs
*attrs
)
303 int order
= get_order(size
);
304 u64 dma_mask
= DMA_BIT_MASK(32);
309 * Ignore region specifiers - the kernel's ideas of
310 * pseudo-phys memory layout has nothing to do with the
311 * machine physical layout. We can't allocate highmem
312 * because we can't return a pointer to it.
314 flags
&= ~(__GFP_DMA
| __GFP_HIGHMEM
);
316 if (dma_alloc_from_coherent(hwdev
, size
, dma_handle
, &ret
))
319 /* On ARM this function returns an ioremap'ped virtual address for
320 * which virt_to_phys doesn't return the corresponding physical
321 * address. In fact on ARM virt_to_phys only works for kernel direct
322 * mapped RAM memory. Also see comment below.
324 ret
= xen_alloc_coherent_pages(hwdev
, size
, dma_handle
, flags
, attrs
);
329 if (hwdev
&& hwdev
->coherent_dma_mask
)
330 dma_mask
= dma_alloc_coherent_mask(hwdev
, flags
);
332 /* At this point dma_handle is the physical address, next we are
333 * going to set it to the machine address.
334 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
337 dev_addr
= xen_phys_to_bus(phys
);
338 if (((dev_addr
+ size
- 1 <= dma_mask
)) &&
339 !range_straddles_page_boundary(phys
, size
))
340 *dma_handle
= dev_addr
;
342 if (xen_create_contiguous_region(phys
, order
,
343 fls64(dma_mask
), dma_handle
) != 0) {
344 xen_free_coherent_pages(hwdev
, size
, ret
, (dma_addr_t
)phys
, attrs
);
348 memset(ret
, 0, size
);
351 EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent
);
354 xen_swiotlb_free_coherent(struct device
*hwdev
, size_t size
, void *vaddr
,
355 dma_addr_t dev_addr
, struct dma_attrs
*attrs
)
357 int order
= get_order(size
);
359 u64 dma_mask
= DMA_BIT_MASK(32);
361 if (dma_release_from_coherent(hwdev
, order
, vaddr
))
364 if (hwdev
&& hwdev
->coherent_dma_mask
)
365 dma_mask
= hwdev
->coherent_dma_mask
;
367 /* do not use virt_to_phys because on ARM it doesn't return you the
368 * physical address */
369 phys
= xen_bus_to_phys(dev_addr
);
371 if (((dev_addr
+ size
- 1 > dma_mask
)) ||
372 range_straddles_page_boundary(phys
, size
))
373 xen_destroy_contiguous_region(phys
, order
);
375 xen_free_coherent_pages(hwdev
, size
, vaddr
, (dma_addr_t
)phys
, attrs
);
377 EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent
);
381 * Map a single buffer of the indicated size for DMA in streaming mode. The
382 * physical address to use is returned.
384 * Once the device is given the dma address, the device owns this memory until
385 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
387 dma_addr_t
xen_swiotlb_map_page(struct device
*dev
, struct page
*page
,
388 unsigned long offset
, size_t size
,
389 enum dma_data_direction dir
,
390 struct dma_attrs
*attrs
)
392 phys_addr_t map
, phys
= page_to_phys(page
) + offset
;
393 dma_addr_t dev_addr
= xen_phys_to_bus(phys
);
395 BUG_ON(dir
== DMA_NONE
);
397 * If the address happens to be in the device's DMA window,
398 * we can safely return the device addr and not worry about bounce
401 if (dma_capable(dev
, dev_addr
, size
) &&
402 !range_straddles_page_boundary(phys
, size
) && !swiotlb_force
) {
403 /* we are not interested in the dma_addr returned by
404 * xen_dma_map_page, only in the potential cache flushes executed
405 * by the function. */
406 xen_dma_map_page(dev
, page
, offset
, size
, dir
, attrs
);
411 * Oh well, have to allocate and map a bounce buffer.
413 trace_swiotlb_bounced(dev
, dev_addr
, size
, swiotlb_force
);
415 map
= swiotlb_tbl_map_single(dev
, start_dma_addr
, phys
, size
, dir
);
416 if (map
== SWIOTLB_MAP_ERROR
)
417 return DMA_ERROR_CODE
;
419 xen_dma_map_page(dev
, pfn_to_page(map
>> PAGE_SHIFT
),
420 map
& ~PAGE_MASK
, size
, dir
, attrs
);
421 dev_addr
= xen_phys_to_bus(map
);
424 * Ensure that the address returned is DMA'ble
426 if (!dma_capable(dev
, dev_addr
, size
)) {
427 swiotlb_tbl_unmap_single(dev
, map
, size
, dir
);
432 EXPORT_SYMBOL_GPL(xen_swiotlb_map_page
);
435 * Unmap a single streaming mode DMA translation. The dma_addr and size must
436 * match what was provided for in a previous xen_swiotlb_map_page call. All
437 * other usages are undefined.
439 * After this call, reads by the cpu to the buffer are guaranteed to see
440 * whatever the device wrote there.
442 static void xen_unmap_single(struct device
*hwdev
, dma_addr_t dev_addr
,
443 size_t size
, enum dma_data_direction dir
,
444 struct dma_attrs
*attrs
)
446 phys_addr_t paddr
= xen_bus_to_phys(dev_addr
);
448 BUG_ON(dir
== DMA_NONE
);
450 xen_dma_unmap_page(hwdev
, paddr
, size
, dir
, attrs
);
452 /* NOTE: We use dev_addr here, not paddr! */
453 if (is_xen_swiotlb_buffer(dev_addr
)) {
454 swiotlb_tbl_unmap_single(hwdev
, paddr
, size
, dir
);
458 if (dir
!= DMA_FROM_DEVICE
)
462 * phys_to_virt doesn't work with hihgmem page but we could
463 * call dma_mark_clean() with hihgmem page here. However, we
464 * are fine since dma_mark_clean() is null on POWERPC. We can
465 * make dma_mark_clean() take a physical address if necessary.
467 dma_mark_clean(phys_to_virt(paddr
), size
);
470 void xen_swiotlb_unmap_page(struct device
*hwdev
, dma_addr_t dev_addr
,
471 size_t size
, enum dma_data_direction dir
,
472 struct dma_attrs
*attrs
)
474 xen_unmap_single(hwdev
, dev_addr
, size
, dir
, attrs
);
476 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page
);
479 * Make physical memory consistent for a single streaming mode DMA translation
482 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
483 * using the cpu, yet do not wish to teardown the dma mapping, you must
484 * call this function before doing so. At the next point you give the dma
485 * address back to the card, you must first perform a
486 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
489 xen_swiotlb_sync_single(struct device
*hwdev
, dma_addr_t dev_addr
,
490 size_t size
, enum dma_data_direction dir
,
491 enum dma_sync_target target
)
493 phys_addr_t paddr
= xen_bus_to_phys(dev_addr
);
495 BUG_ON(dir
== DMA_NONE
);
497 if (target
== SYNC_FOR_CPU
)
498 xen_dma_sync_single_for_cpu(hwdev
, paddr
, size
, dir
);
500 /* NOTE: We use dev_addr here, not paddr! */
501 if (is_xen_swiotlb_buffer(dev_addr
))
502 swiotlb_tbl_sync_single(hwdev
, paddr
, size
, dir
, target
);
504 if (target
== SYNC_FOR_DEVICE
)
505 xen_dma_sync_single_for_cpu(hwdev
, paddr
, size
, dir
);
507 if (dir
!= DMA_FROM_DEVICE
)
510 dma_mark_clean(phys_to_virt(paddr
), size
);
514 xen_swiotlb_sync_single_for_cpu(struct device
*hwdev
, dma_addr_t dev_addr
,
515 size_t size
, enum dma_data_direction dir
)
517 xen_swiotlb_sync_single(hwdev
, dev_addr
, size
, dir
, SYNC_FOR_CPU
);
519 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu
);
522 xen_swiotlb_sync_single_for_device(struct device
*hwdev
, dma_addr_t dev_addr
,
523 size_t size
, enum dma_data_direction dir
)
525 xen_swiotlb_sync_single(hwdev
, dev_addr
, size
, dir
, SYNC_FOR_DEVICE
);
527 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device
);
530 * Map a set of buffers described by scatterlist in streaming mode for DMA.
531 * This is the scatter-gather version of the above xen_swiotlb_map_page
532 * interface. Here the scatter gather list elements are each tagged with the
533 * appropriate dma address and length. They are obtained via
534 * sg_dma_{address,length}(SG).
536 * NOTE: An implementation may be able to use a smaller number of
537 * DMA address/length pairs than there are SG table elements.
538 * (for example via virtual mapping capabilities)
539 * The routine returns the number of addr/length pairs actually
540 * used, at most nents.
542 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
546 xen_swiotlb_map_sg_attrs(struct device
*hwdev
, struct scatterlist
*sgl
,
547 int nelems
, enum dma_data_direction dir
,
548 struct dma_attrs
*attrs
)
550 struct scatterlist
*sg
;
553 BUG_ON(dir
== DMA_NONE
);
555 for_each_sg(sgl
, sg
, nelems
, i
) {
556 phys_addr_t paddr
= sg_phys(sg
);
557 dma_addr_t dev_addr
= xen_phys_to_bus(paddr
);
560 !dma_capable(hwdev
, dev_addr
, sg
->length
) ||
561 range_straddles_page_boundary(paddr
, sg
->length
)) {
562 phys_addr_t map
= swiotlb_tbl_map_single(hwdev
,
567 if (map
== SWIOTLB_MAP_ERROR
) {
568 dev_warn(hwdev
, "swiotlb buffer is full\n");
569 /* Don't panic here, we expect map_sg users
570 to do proper error handling. */
571 xen_swiotlb_unmap_sg_attrs(hwdev
, sgl
, i
, dir
,
576 xen_dma_map_page(hwdev
, pfn_to_page(map
>> PAGE_SHIFT
),
581 sg
->dma_address
= xen_phys_to_bus(map
);
583 /* we are not interested in the dma_addr returned by
584 * xen_dma_map_page, only in the potential cache flushes executed
585 * by the function. */
586 xen_dma_map_page(hwdev
, pfn_to_page(paddr
>> PAGE_SHIFT
),
591 sg
->dma_address
= dev_addr
;
593 sg_dma_len(sg
) = sg
->length
;
597 EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs
);
600 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
601 * concerning calls here are the same as for swiotlb_unmap_page() above.
604 xen_swiotlb_unmap_sg_attrs(struct device
*hwdev
, struct scatterlist
*sgl
,
605 int nelems
, enum dma_data_direction dir
,
606 struct dma_attrs
*attrs
)
608 struct scatterlist
*sg
;
611 BUG_ON(dir
== DMA_NONE
);
613 for_each_sg(sgl
, sg
, nelems
, i
)
614 xen_unmap_single(hwdev
, sg
->dma_address
, sg_dma_len(sg
), dir
, attrs
);
617 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs
);
620 * Make physical memory consistent for a set of streaming mode DMA translations
623 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
627 xen_swiotlb_sync_sg(struct device
*hwdev
, struct scatterlist
*sgl
,
628 int nelems
, enum dma_data_direction dir
,
629 enum dma_sync_target target
)
631 struct scatterlist
*sg
;
634 for_each_sg(sgl
, sg
, nelems
, i
)
635 xen_swiotlb_sync_single(hwdev
, sg
->dma_address
,
636 sg_dma_len(sg
), dir
, target
);
640 xen_swiotlb_sync_sg_for_cpu(struct device
*hwdev
, struct scatterlist
*sg
,
641 int nelems
, enum dma_data_direction dir
)
643 xen_swiotlb_sync_sg(hwdev
, sg
, nelems
, dir
, SYNC_FOR_CPU
);
645 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu
);
648 xen_swiotlb_sync_sg_for_device(struct device
*hwdev
, struct scatterlist
*sg
,
649 int nelems
, enum dma_data_direction dir
)
651 xen_swiotlb_sync_sg(hwdev
, sg
, nelems
, dir
, SYNC_FOR_DEVICE
);
653 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device
);
656 xen_swiotlb_dma_mapping_error(struct device
*hwdev
, dma_addr_t dma_addr
)
660 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error
);
663 * Return whether the given device DMA address mask can be supported
664 * properly. For example, if your device can only drive the low 24-bits
665 * during bus mastering, then you would pass 0x00ffffff as the mask to
669 xen_swiotlb_dma_supported(struct device
*hwdev
, u64 mask
)
671 return xen_virt_to_bus(xen_io_tlb_end
- 1) <= mask
;
673 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported
);
676 xen_swiotlb_set_dma_mask(struct device
*dev
, u64 dma_mask
)
678 if (!dev
->dma_mask
|| !xen_swiotlb_dma_supported(dev
, dma_mask
))
681 *dev
->dma_mask
= dma_mask
;
685 EXPORT_SYMBOL_GPL(xen_swiotlb_set_dma_mask
);