2 * Dynamic DMA mapping support.
4 * This implementation is a fallback for platforms that do not support
5 * I/O TLBs (aka DMA address translation hardware).
6 * Copyright (C) 2000 Asit Mallick <Asit.K.Mallick@intel.com>
7 * Copyright (C) 2000 Goutham Rao <goutham.rao@intel.com>
8 * Copyright (C) 2000, 2003 Hewlett-Packard Co
9 * David Mosberger-Tang <davidm@hpl.hp.com>
11 * 03/05/07 davidm Switch from PCI-DMA to generic device DMA API.
12 * 00/12/13 davidm Rename to swiotlb.c and add mark_clean() to avoid
13 * unnecessary i-cache flushing.
14 * 04/07/.. ak Better overflow handling. Assorted fixes.
15 * 05/09/10 linville Add support for syncing ranges, support syncing for
16 * DMA_BIDIRECTIONAL mappings, miscellaneous cleanup.
19 #include <linux/cache.h>
20 #include <linux/dma-mapping.h>
22 #include <linux/module.h>
23 #include <linux/spinlock.h>
24 #include <linux/swiotlb.h>
25 #include <linux/string.h>
26 #include <linux/types.h>
27 #include <linux/ctype.h>
31 #include <asm/scatterlist.h>
33 #include <linux/init.h>
34 #include <linux/bootmem.h>
35 #include <linux/iommu-helper.h>
37 #define OFFSET(val,align) ((unsigned long) \
38 ( (val) & ( (align) - 1)))
40 #define SG_ENT_VIRT_ADDRESS(sg) (sg_virt((sg)))
41 #define SG_ENT_PHYS_ADDRESS(sg) virt_to_bus(SG_ENT_VIRT_ADDRESS(sg))
44 * Maximum allowable number of contiguous slabs to map,
45 * must be a power of 2. What is the appropriate value ?
46 * The complexity of {map,unmap}_single is linearly dependent on this value.
48 #define IO_TLB_SEGSIZE 128
51 * log of the size of each IO TLB slab. The number of slabs is command line
54 #define IO_TLB_SHIFT 11
56 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
59 * Minimum IO TLB size to bother booting with. Systems with mainly
60 * 64bit capable cards will only lightly use the swiotlb. If we can't
61 * allocate a contiguous 1MB, we're probably in trouble anyway.
63 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
66 * Enumeration for sync targets
68 enum dma_sync_target
{
76 * Used to do a quick range check in swiotlb_unmap_single and
77 * swiotlb_sync_single_*, to see if the memory was in fact allocated by this
80 static char *io_tlb_start
, *io_tlb_end
;
83 * The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and
84 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
86 static unsigned long io_tlb_nslabs
;
89 * When the IOMMU overflows we return a fallback buffer. This sets the size.
91 static unsigned long io_tlb_overflow
= 32*1024;
93 void *io_tlb_overflow_buffer
;
96 * This is a free list describing the number of free entries available from
99 static unsigned int *io_tlb_list
;
100 static unsigned int io_tlb_index
;
103 * We need to save away the original address corresponding to a mapped entry
104 * for the sync operations.
106 static unsigned char **io_tlb_orig_addr
;
109 * Protect the above data structures in the map and unmap calls
111 static DEFINE_SPINLOCK(io_tlb_lock
);
114 setup_io_tlb_npages(char *str
)
117 io_tlb_nslabs
= simple_strtoul(str
, &str
, 0);
118 /* avoid tail segment of size < IO_TLB_SEGSIZE */
119 io_tlb_nslabs
= ALIGN(io_tlb_nslabs
, IO_TLB_SEGSIZE
);
123 if (!strcmp(str
, "force"))
127 __setup("swiotlb=", setup_io_tlb_npages
);
128 /* make io_tlb_overflow tunable too? */
130 void * __weak
swiotlb_alloc_boot(size_t size
, unsigned long nslabs
)
132 return alloc_bootmem_low_pages(size
);
135 void * __weak
swiotlb_alloc(unsigned order
, unsigned long nslabs
)
137 return (void *)__get_free_pages(GFP_DMA
| __GFP_NOWARN
, order
);
141 * Statically reserve bounce buffer space and initialize bounce buffer data
142 * structures for the software IO TLB used to implement the DMA API.
145 swiotlb_init_with_default_size(size_t default_size
)
147 unsigned long i
, bytes
;
149 if (!io_tlb_nslabs
) {
150 io_tlb_nslabs
= (default_size
>> IO_TLB_SHIFT
);
151 io_tlb_nslabs
= ALIGN(io_tlb_nslabs
, IO_TLB_SEGSIZE
);
154 bytes
= io_tlb_nslabs
<< IO_TLB_SHIFT
;
157 * Get IO TLB memory from the low pages
159 io_tlb_start
= swiotlb_alloc_boot(bytes
, io_tlb_nslabs
);
161 panic("Cannot allocate SWIOTLB buffer");
162 io_tlb_end
= io_tlb_start
+ bytes
;
165 * Allocate and initialize the free list array. This array is used
166 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
167 * between io_tlb_start and io_tlb_end.
169 io_tlb_list
= alloc_bootmem(io_tlb_nslabs
* sizeof(int));
170 for (i
= 0; i
< io_tlb_nslabs
; i
++)
171 io_tlb_list
[i
] = IO_TLB_SEGSIZE
- OFFSET(i
, IO_TLB_SEGSIZE
);
173 io_tlb_orig_addr
= alloc_bootmem(io_tlb_nslabs
* sizeof(char *));
176 * Get the overflow emergency buffer
178 io_tlb_overflow_buffer
= alloc_bootmem_low(io_tlb_overflow
);
179 if (!io_tlb_overflow_buffer
)
180 panic("Cannot allocate SWIOTLB overflow buffer!\n");
182 printk(KERN_INFO
"Placing software IO TLB between 0x%lx - 0x%lx\n",
183 virt_to_bus(io_tlb_start
), virt_to_bus(io_tlb_end
));
189 swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
193 * Systems with larger DMA zones (those that don't support ISA) can
194 * initialize the swiotlb later using the slab allocator if needed.
195 * This should be just like above, but with some error catching.
198 swiotlb_late_init_with_default_size(size_t default_size
)
200 unsigned long i
, bytes
, req_nslabs
= io_tlb_nslabs
;
203 if (!io_tlb_nslabs
) {
204 io_tlb_nslabs
= (default_size
>> IO_TLB_SHIFT
);
205 io_tlb_nslabs
= ALIGN(io_tlb_nslabs
, IO_TLB_SEGSIZE
);
209 * Get IO TLB memory from the low pages
211 order
= get_order(io_tlb_nslabs
<< IO_TLB_SHIFT
);
212 io_tlb_nslabs
= SLABS_PER_PAGE
<< order
;
213 bytes
= io_tlb_nslabs
<< IO_TLB_SHIFT
;
215 while ((SLABS_PER_PAGE
<< order
) > IO_TLB_MIN_SLABS
) {
216 io_tlb_start
= swiotlb_alloc(order
, io_tlb_nslabs
);
225 if (order
!= get_order(bytes
)) {
226 printk(KERN_WARNING
"Warning: only able to allocate %ld MB "
227 "for software IO TLB\n", (PAGE_SIZE
<< order
) >> 20);
228 io_tlb_nslabs
= SLABS_PER_PAGE
<< order
;
229 bytes
= io_tlb_nslabs
<< IO_TLB_SHIFT
;
231 io_tlb_end
= io_tlb_start
+ bytes
;
232 memset(io_tlb_start
, 0, bytes
);
235 * Allocate and initialize the free list array. This array is used
236 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
237 * between io_tlb_start and io_tlb_end.
239 io_tlb_list
= (unsigned int *)__get_free_pages(GFP_KERNEL
,
240 get_order(io_tlb_nslabs
* sizeof(int)));
244 for (i
= 0; i
< io_tlb_nslabs
; i
++)
245 io_tlb_list
[i
] = IO_TLB_SEGSIZE
- OFFSET(i
, IO_TLB_SEGSIZE
);
248 io_tlb_orig_addr
= (unsigned char **)__get_free_pages(GFP_KERNEL
,
249 get_order(io_tlb_nslabs
* sizeof(char *)));
250 if (!io_tlb_orig_addr
)
253 memset(io_tlb_orig_addr
, 0, io_tlb_nslabs
* sizeof(char *));
256 * Get the overflow emergency buffer
258 io_tlb_overflow_buffer
= (void *)__get_free_pages(GFP_DMA
,
259 get_order(io_tlb_overflow
));
260 if (!io_tlb_overflow_buffer
)
263 printk(KERN_INFO
"Placing %luMB software IO TLB between 0x%lx - "
264 "0x%lx\n", bytes
>> 20,
265 virt_to_bus(io_tlb_start
), virt_to_bus(io_tlb_end
));
270 free_pages((unsigned long)io_tlb_orig_addr
, get_order(io_tlb_nslabs
*
272 io_tlb_orig_addr
= NULL
;
274 free_pages((unsigned long)io_tlb_list
, get_order(io_tlb_nslabs
*
279 free_pages((unsigned long)io_tlb_start
, order
);
282 io_tlb_nslabs
= req_nslabs
;
287 address_needs_mapping(struct device
*hwdev
, dma_addr_t addr
, size_t size
)
289 return !is_buffer_dma_capable(dma_get_mask(hwdev
), addr
, size
);
292 static int is_swiotlb_buffer(char *addr
)
294 return addr
>= io_tlb_start
&& addr
< io_tlb_end
;
298 * Allocates bounce buffer and returns its kernel virtual address.
301 map_single(struct device
*hwdev
, char *buffer
, size_t size
, int dir
)
305 unsigned int nslots
, stride
, index
, wrap
;
307 unsigned long start_dma_addr
;
309 unsigned long offset_slots
;
310 unsigned long max_slots
;
312 mask
= dma_get_seg_boundary(hwdev
);
313 start_dma_addr
= virt_to_bus(io_tlb_start
) & mask
;
315 offset_slots
= ALIGN(start_dma_addr
, 1 << IO_TLB_SHIFT
) >> IO_TLB_SHIFT
;
317 ? ALIGN(mask
+ 1, 1 << IO_TLB_SHIFT
) >> IO_TLB_SHIFT
318 : 1UL << (BITS_PER_LONG
- IO_TLB_SHIFT
);
321 * For mappings greater than a page, we limit the stride (and
322 * hence alignment) to a page size.
324 nslots
= ALIGN(size
, 1 << IO_TLB_SHIFT
) >> IO_TLB_SHIFT
;
325 if (size
> PAGE_SIZE
)
326 stride
= (1 << (PAGE_SHIFT
- IO_TLB_SHIFT
));
333 * Find suitable number of IO TLB entries size that will fit this
334 * request and allocate a buffer from that IO TLB pool.
336 spin_lock_irqsave(&io_tlb_lock
, flags
);
337 index
= ALIGN(io_tlb_index
, stride
);
338 if (index
>= io_tlb_nslabs
)
343 while (iommu_is_span_boundary(index
, nslots
, offset_slots
,
346 if (index
>= io_tlb_nslabs
)
353 * If we find a slot that indicates we have 'nslots' number of
354 * contiguous buffers, we allocate the buffers from that slot
355 * and mark the entries as '0' indicating unavailable.
357 if (io_tlb_list
[index
] >= nslots
) {
360 for (i
= index
; i
< (int) (index
+ nslots
); i
++)
362 for (i
= index
- 1; (OFFSET(i
, IO_TLB_SEGSIZE
) != IO_TLB_SEGSIZE
- 1) && io_tlb_list
[i
]; i
--)
363 io_tlb_list
[i
] = ++count
;
364 dma_addr
= io_tlb_start
+ (index
<< IO_TLB_SHIFT
);
367 * Update the indices to avoid searching in the next
370 io_tlb_index
= ((index
+ nslots
) < io_tlb_nslabs
371 ? (index
+ nslots
) : 0);
376 if (index
>= io_tlb_nslabs
)
378 } while (index
!= wrap
);
381 spin_unlock_irqrestore(&io_tlb_lock
, flags
);
384 spin_unlock_irqrestore(&io_tlb_lock
, flags
);
387 * Save away the mapping from the original address to the DMA address.
388 * This is needed when we sync the memory. Then we sync the buffer if
391 for (i
= 0; i
< nslots
; i
++)
392 io_tlb_orig_addr
[index
+i
] = buffer
+ (i
<< IO_TLB_SHIFT
);
393 if (dir
== DMA_TO_DEVICE
|| dir
== DMA_BIDIRECTIONAL
)
394 memcpy(dma_addr
, buffer
, size
);
400 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
403 unmap_single(struct device
*hwdev
, char *dma_addr
, size_t size
, int dir
)
406 int i
, count
, nslots
= ALIGN(size
, 1 << IO_TLB_SHIFT
) >> IO_TLB_SHIFT
;
407 int index
= (dma_addr
- io_tlb_start
) >> IO_TLB_SHIFT
;
408 char *buffer
= io_tlb_orig_addr
[index
];
411 * First, sync the memory before unmapping the entry
413 if (buffer
&& ((dir
== DMA_FROM_DEVICE
) || (dir
== DMA_BIDIRECTIONAL
)))
415 * bounce... copy the data back into the original buffer * and
416 * delete the bounce buffer.
418 memcpy(buffer
, dma_addr
, size
);
421 * Return the buffer to the free list by setting the corresponding
422 * entries to indicate the number of contigous entries available.
423 * While returning the entries to the free list, we merge the entries
424 * with slots below and above the pool being returned.
426 spin_lock_irqsave(&io_tlb_lock
, flags
);
428 count
= ((index
+ nslots
) < ALIGN(index
+ 1, IO_TLB_SEGSIZE
) ?
429 io_tlb_list
[index
+ nslots
] : 0);
431 * Step 1: return the slots to the free list, merging the
432 * slots with superceeding slots
434 for (i
= index
+ nslots
- 1; i
>= index
; i
--)
435 io_tlb_list
[i
] = ++count
;
437 * Step 2: merge the returned slots with the preceding slots,
438 * if available (non zero)
440 for (i
= index
- 1; (OFFSET(i
, IO_TLB_SEGSIZE
) != IO_TLB_SEGSIZE
-1) && io_tlb_list
[i
]; i
--)
441 io_tlb_list
[i
] = ++count
;
443 spin_unlock_irqrestore(&io_tlb_lock
, flags
);
447 sync_single(struct device
*hwdev
, char *dma_addr
, size_t size
,
450 int index
= (dma_addr
- io_tlb_start
) >> IO_TLB_SHIFT
;
451 char *buffer
= io_tlb_orig_addr
[index
];
453 buffer
+= ((unsigned long)dma_addr
& ((1 << IO_TLB_SHIFT
) - 1));
457 if (likely(dir
== DMA_FROM_DEVICE
|| dir
== DMA_BIDIRECTIONAL
))
458 memcpy(buffer
, dma_addr
, size
);
460 BUG_ON(dir
!= DMA_TO_DEVICE
);
462 case SYNC_FOR_DEVICE
:
463 if (likely(dir
== DMA_TO_DEVICE
|| dir
== DMA_BIDIRECTIONAL
))
464 memcpy(dma_addr
, buffer
, size
);
466 BUG_ON(dir
!= DMA_FROM_DEVICE
);
474 swiotlb_alloc_coherent(struct device
*hwdev
, size_t size
,
475 dma_addr_t
*dma_handle
, gfp_t flags
)
479 int order
= get_order(size
);
480 u64 dma_mask
= DMA_32BIT_MASK
;
482 if (hwdev
&& hwdev
->coherent_dma_mask
)
483 dma_mask
= hwdev
->coherent_dma_mask
;
485 ret
= (void *)__get_free_pages(flags
, order
);
486 if (ret
&& !is_buffer_dma_capable(dma_mask
, virt_to_bus(ret
), size
)) {
488 * The allocated memory isn't reachable by the device.
489 * Fall back on swiotlb_map_single().
491 free_pages((unsigned long) ret
, order
);
496 * We are either out of memory or the device can't DMA
497 * to GFP_DMA memory; fall back on
498 * swiotlb_map_single(), which will grab memory from
499 * the lowest available address range.
501 ret
= map_single(hwdev
, NULL
, size
, DMA_FROM_DEVICE
);
506 memset(ret
, 0, size
);
507 dev_addr
= virt_to_bus(ret
);
509 /* Confirm address can be DMA'd by device */
510 if (!is_buffer_dma_capable(dma_mask
, dev_addr
, size
)) {
511 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
512 (unsigned long long)dma_mask
,
513 (unsigned long long)dev_addr
);
515 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
516 unmap_single(hwdev
, ret
, size
, DMA_TO_DEVICE
);
519 *dma_handle
= dev_addr
;
524 swiotlb_free_coherent(struct device
*hwdev
, size_t size
, void *vaddr
,
525 dma_addr_t dma_handle
)
527 WARN_ON(irqs_disabled());
528 if (!is_swiotlb_buffer(vaddr
))
529 free_pages((unsigned long) vaddr
, get_order(size
));
531 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
532 unmap_single(hwdev
, vaddr
, size
, DMA_TO_DEVICE
);
536 swiotlb_full(struct device
*dev
, size_t size
, int dir
, int do_panic
)
539 * Ran out of IOMMU space for this operation. This is very bad.
540 * Unfortunately the drivers cannot handle this operation properly.
541 * unless they check for dma_mapping_error (most don't)
542 * When the mapping is small enough return a static buffer to limit
543 * the damage, or panic when the transfer is too big.
545 printk(KERN_ERR
"DMA: Out of SW-IOMMU space for %zu bytes at "
546 "device %s\n", size
, dev
? dev
->bus_id
: "?");
548 if (size
> io_tlb_overflow
&& do_panic
) {
549 if (dir
== DMA_FROM_DEVICE
|| dir
== DMA_BIDIRECTIONAL
)
550 panic("DMA: Memory would be corrupted\n");
551 if (dir
== DMA_TO_DEVICE
|| dir
== DMA_BIDIRECTIONAL
)
552 panic("DMA: Random memory would be DMAed\n");
557 * Map a single buffer of the indicated size for DMA in streaming mode. The
558 * physical address to use is returned.
560 * Once the device is given the dma address, the device owns this memory until
561 * either swiotlb_unmap_single or swiotlb_dma_sync_single is performed.
564 swiotlb_map_single_attrs(struct device
*hwdev
, void *ptr
, size_t size
,
565 int dir
, struct dma_attrs
*attrs
)
567 dma_addr_t dev_addr
= virt_to_bus(ptr
);
570 BUG_ON(dir
== DMA_NONE
);
572 * If the pointer passed in happens to be in the device's DMA window,
573 * we can safely return the device addr and not worry about bounce
576 if (!address_needs_mapping(hwdev
, dev_addr
, size
) && !swiotlb_force
)
580 * Oh well, have to allocate and map a bounce buffer.
582 map
= map_single(hwdev
, ptr
, size
, dir
);
584 swiotlb_full(hwdev
, size
, dir
, 1);
585 map
= io_tlb_overflow_buffer
;
588 dev_addr
= virt_to_bus(map
);
591 * Ensure that the address returned is DMA'ble
593 if (address_needs_mapping(hwdev
, dev_addr
, size
))
594 panic("map_single: bounce buffer is not DMA'ble");
598 EXPORT_SYMBOL(swiotlb_map_single_attrs
);
601 swiotlb_map_single(struct device
*hwdev
, void *ptr
, size_t size
, int dir
)
603 return swiotlb_map_single_attrs(hwdev
, ptr
, size
, dir
, NULL
);
607 * Unmap a single streaming mode DMA translation. The dma_addr and size must
608 * match what was provided for in a previous swiotlb_map_single call. All
609 * other usages are undefined.
611 * After this call, reads by the cpu to the buffer are guaranteed to see
612 * whatever the device wrote there.
615 swiotlb_unmap_single_attrs(struct device
*hwdev
, dma_addr_t dev_addr
,
616 size_t size
, int dir
, struct dma_attrs
*attrs
)
618 char *dma_addr
= bus_to_virt(dev_addr
);
620 BUG_ON(dir
== DMA_NONE
);
621 if (is_swiotlb_buffer(dma_addr
))
622 unmap_single(hwdev
, dma_addr
, size
, dir
);
623 else if (dir
== DMA_FROM_DEVICE
)
624 dma_mark_clean(dma_addr
, size
);
626 EXPORT_SYMBOL(swiotlb_unmap_single_attrs
);
629 swiotlb_unmap_single(struct device
*hwdev
, dma_addr_t dev_addr
, size_t size
,
632 return swiotlb_unmap_single_attrs(hwdev
, dev_addr
, size
, dir
, NULL
);
635 * Make physical memory consistent for a single streaming mode DMA translation
638 * If you perform a swiotlb_map_single() but wish to interrogate the buffer
639 * using the cpu, yet do not wish to teardown the dma mapping, you must
640 * call this function before doing so. At the next point you give the dma
641 * address back to the card, you must first perform a
642 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
645 swiotlb_sync_single(struct device
*hwdev
, dma_addr_t dev_addr
,
646 size_t size
, int dir
, int target
)
648 char *dma_addr
= bus_to_virt(dev_addr
);
650 BUG_ON(dir
== DMA_NONE
);
651 if (is_swiotlb_buffer(dma_addr
))
652 sync_single(hwdev
, dma_addr
, size
, dir
, target
);
653 else if (dir
== DMA_FROM_DEVICE
)
654 dma_mark_clean(dma_addr
, size
);
658 swiotlb_sync_single_for_cpu(struct device
*hwdev
, dma_addr_t dev_addr
,
659 size_t size
, int dir
)
661 swiotlb_sync_single(hwdev
, dev_addr
, size
, dir
, SYNC_FOR_CPU
);
665 swiotlb_sync_single_for_device(struct device
*hwdev
, dma_addr_t dev_addr
,
666 size_t size
, int dir
)
668 swiotlb_sync_single(hwdev
, dev_addr
, size
, dir
, SYNC_FOR_DEVICE
);
672 * Same as above, but for a sub-range of the mapping.
675 swiotlb_sync_single_range(struct device
*hwdev
, dma_addr_t dev_addr
,
676 unsigned long offset
, size_t size
,
679 char *dma_addr
= bus_to_virt(dev_addr
) + offset
;
681 BUG_ON(dir
== DMA_NONE
);
682 if (is_swiotlb_buffer(dma_addr
))
683 sync_single(hwdev
, dma_addr
, size
, dir
, target
);
684 else if (dir
== DMA_FROM_DEVICE
)
685 dma_mark_clean(dma_addr
, size
);
689 swiotlb_sync_single_range_for_cpu(struct device
*hwdev
, dma_addr_t dev_addr
,
690 unsigned long offset
, size_t size
, int dir
)
692 swiotlb_sync_single_range(hwdev
, dev_addr
, offset
, size
, dir
,
697 swiotlb_sync_single_range_for_device(struct device
*hwdev
, dma_addr_t dev_addr
,
698 unsigned long offset
, size_t size
, int dir
)
700 swiotlb_sync_single_range(hwdev
, dev_addr
, offset
, size
, dir
,
704 void swiotlb_unmap_sg_attrs(struct device
*, struct scatterlist
*, int, int,
707 * Map a set of buffers described by scatterlist in streaming mode for DMA.
708 * This is the scatter-gather version of the above swiotlb_map_single
709 * interface. Here the scatter gather list elements are each tagged with the
710 * appropriate dma address and length. They are obtained via
711 * sg_dma_{address,length}(SG).
713 * NOTE: An implementation may be able to use a smaller number of
714 * DMA address/length pairs than there are SG table elements.
715 * (for example via virtual mapping capabilities)
716 * The routine returns the number of addr/length pairs actually
717 * used, at most nents.
719 * Device ownership issues as mentioned above for swiotlb_map_single are the
723 swiotlb_map_sg_attrs(struct device
*hwdev
, struct scatterlist
*sgl
, int nelems
,
724 int dir
, struct dma_attrs
*attrs
)
726 struct scatterlist
*sg
;
731 BUG_ON(dir
== DMA_NONE
);
733 for_each_sg(sgl
, sg
, nelems
, i
) {
734 addr
= SG_ENT_VIRT_ADDRESS(sg
);
735 dev_addr
= virt_to_bus(addr
);
737 address_needs_mapping(hwdev
, dev_addr
, sg
->length
)) {
738 void *map
= map_single(hwdev
, addr
, sg
->length
, dir
);
740 /* Don't panic here, we expect map_sg users
741 to do proper error handling. */
742 swiotlb_full(hwdev
, sg
->length
, dir
, 0);
743 swiotlb_unmap_sg_attrs(hwdev
, sgl
, i
, dir
,
745 sgl
[0].dma_length
= 0;
748 sg
->dma_address
= virt_to_bus(map
);
750 sg
->dma_address
= dev_addr
;
751 sg
->dma_length
= sg
->length
;
755 EXPORT_SYMBOL(swiotlb_map_sg_attrs
);
758 swiotlb_map_sg(struct device
*hwdev
, struct scatterlist
*sgl
, int nelems
,
761 return swiotlb_map_sg_attrs(hwdev
, sgl
, nelems
, dir
, NULL
);
765 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
766 * concerning calls here are the same as for swiotlb_unmap_single() above.
769 swiotlb_unmap_sg_attrs(struct device
*hwdev
, struct scatterlist
*sgl
,
770 int nelems
, int dir
, struct dma_attrs
*attrs
)
772 struct scatterlist
*sg
;
775 BUG_ON(dir
== DMA_NONE
);
777 for_each_sg(sgl
, sg
, nelems
, i
) {
778 if (sg
->dma_address
!= SG_ENT_PHYS_ADDRESS(sg
))
779 unmap_single(hwdev
, bus_to_virt(sg
->dma_address
),
780 sg
->dma_length
, dir
);
781 else if (dir
== DMA_FROM_DEVICE
)
782 dma_mark_clean(SG_ENT_VIRT_ADDRESS(sg
), sg
->dma_length
);
785 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs
);
788 swiotlb_unmap_sg(struct device
*hwdev
, struct scatterlist
*sgl
, int nelems
,
791 return swiotlb_unmap_sg_attrs(hwdev
, sgl
, nelems
, dir
, NULL
);
795 * Make physical memory consistent for a set of streaming mode DMA translations
798 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
802 swiotlb_sync_sg(struct device
*hwdev
, struct scatterlist
*sgl
,
803 int nelems
, int dir
, int target
)
805 struct scatterlist
*sg
;
808 BUG_ON(dir
== DMA_NONE
);
810 for_each_sg(sgl
, sg
, nelems
, i
) {
811 if (sg
->dma_address
!= SG_ENT_PHYS_ADDRESS(sg
))
812 sync_single(hwdev
, bus_to_virt(sg
->dma_address
),
813 sg
->dma_length
, dir
, target
);
814 else if (dir
== DMA_FROM_DEVICE
)
815 dma_mark_clean(SG_ENT_VIRT_ADDRESS(sg
), sg
->dma_length
);
820 swiotlb_sync_sg_for_cpu(struct device
*hwdev
, struct scatterlist
*sg
,
823 swiotlb_sync_sg(hwdev
, sg
, nelems
, dir
, SYNC_FOR_CPU
);
827 swiotlb_sync_sg_for_device(struct device
*hwdev
, struct scatterlist
*sg
,
830 swiotlb_sync_sg(hwdev
, sg
, nelems
, dir
, SYNC_FOR_DEVICE
);
834 swiotlb_dma_mapping_error(struct device
*hwdev
, dma_addr_t dma_addr
)
836 return (dma_addr
== virt_to_bus(io_tlb_overflow_buffer
));
840 * Return whether the given device DMA address mask can be supported
841 * properly. For example, if your device can only drive the low 24-bits
842 * during bus mastering, then you would pass 0x00ffffff as the mask to
846 swiotlb_dma_supported(struct device
*hwdev
, u64 mask
)
848 return virt_to_bus(io_tlb_end
- 1) <= mask
;
851 EXPORT_SYMBOL(swiotlb_map_single
);
852 EXPORT_SYMBOL(swiotlb_unmap_single
);
853 EXPORT_SYMBOL(swiotlb_map_sg
);
854 EXPORT_SYMBOL(swiotlb_unmap_sg
);
855 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu
);
856 EXPORT_SYMBOL(swiotlb_sync_single_for_device
);
857 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu
);
858 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device
);
859 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu
);
860 EXPORT_SYMBOL(swiotlb_sync_sg_for_device
);
861 EXPORT_SYMBOL(swiotlb_dma_mapping_error
);
862 EXPORT_SYMBOL(swiotlb_alloc_coherent
);
863 EXPORT_SYMBOL(swiotlb_free_coherent
);
864 EXPORT_SYMBOL(swiotlb_dma_supported
);