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.
17 * 08/12/11 beckyb Add highmem support
20 #include <linux/cache.h>
21 #include <linux/dma-mapping.h>
23 #include <linux/module.h>
24 #include <linux/spinlock.h>
25 #include <linux/string.h>
26 #include <linux/swiotlb.h>
27 #include <linux/pfn.h>
28 #include <linux/types.h>
29 #include <linux/ctype.h>
30 #include <linux/highmem.h>
34 #include <asm/scatterlist.h>
36 #include <linux/init.h>
37 #include <linux/bootmem.h>
38 #include <linux/iommu-helper.h>
40 #define OFFSET(val,align) ((unsigned long) \
41 ( (val) & ( (align) - 1)))
43 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
46 * Minimum IO TLB size to bother booting with. Systems with mainly
47 * 64bit capable cards will only lightly use the swiotlb. If we can't
48 * allocate a contiguous 1MB, we're probably in trouble anyway.
50 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
53 * Enumeration for sync targets
55 enum dma_sync_target
{
63 * Used to do a quick range check in unmap_single and
64 * sync_single_*, to see if the memory was in fact allocated by this
67 static char *io_tlb_start
, *io_tlb_end
;
70 * The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and
71 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
73 static unsigned long io_tlb_nslabs
;
76 * When the IOMMU overflows we return a fallback buffer. This sets the size.
78 static unsigned long io_tlb_overflow
= 32*1024;
80 void *io_tlb_overflow_buffer
;
83 * This is a free list describing the number of free entries available from
86 static unsigned int *io_tlb_list
;
87 static unsigned int io_tlb_index
;
90 * We need to save away the original address corresponding to a mapped entry
91 * for the sync operations.
93 static phys_addr_t
*io_tlb_orig_addr
;
96 * Protect the above data structures in the map and unmap calls
98 static DEFINE_SPINLOCK(io_tlb_lock
);
100 static int late_alloc
;
103 setup_io_tlb_npages(char *str
)
106 io_tlb_nslabs
= simple_strtoul(str
, &str
, 0);
107 /* avoid tail segment of size < IO_TLB_SEGSIZE */
108 io_tlb_nslabs
= ALIGN(io_tlb_nslabs
, IO_TLB_SEGSIZE
);
112 if (!strcmp(str
, "force"))
116 __setup("swiotlb=", setup_io_tlb_npages
);
117 /* make io_tlb_overflow tunable too? */
119 /* Note that this doesn't work with highmem page */
120 static dma_addr_t
swiotlb_virt_to_bus(struct device
*hwdev
,
121 volatile void *address
)
123 return phys_to_dma(hwdev
, virt_to_phys(address
));
126 static void swiotlb_print_info(unsigned long bytes
)
128 phys_addr_t pstart
, pend
;
130 pstart
= virt_to_phys(io_tlb_start
);
131 pend
= virt_to_phys(io_tlb_end
);
133 printk(KERN_INFO
"Placing %luMB software IO TLB between %p - %p\n",
134 bytes
>> 20, io_tlb_start
, io_tlb_end
);
135 printk(KERN_INFO
"software IO TLB at phys %#llx - %#llx\n",
136 (unsigned long long)pstart
,
137 (unsigned long long)pend
);
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
= alloc_bootmem_low_pages(bytes
);
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(phys_addr_t
));
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 swiotlb_print_info(bytes
);
188 swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
192 * Systems with larger DMA zones (those that don't support ISA) can
193 * initialize the swiotlb later using the slab allocator if needed.
194 * This should be just like above, but with some error catching.
197 swiotlb_late_init_with_default_size(size_t default_size
)
199 unsigned long i
, bytes
, req_nslabs
= io_tlb_nslabs
;
202 if (!io_tlb_nslabs
) {
203 io_tlb_nslabs
= (default_size
>> IO_TLB_SHIFT
);
204 io_tlb_nslabs
= ALIGN(io_tlb_nslabs
, IO_TLB_SEGSIZE
);
208 * Get IO TLB memory from the low pages
210 order
= get_order(io_tlb_nslabs
<< IO_TLB_SHIFT
);
211 io_tlb_nslabs
= SLABS_PER_PAGE
<< order
;
212 bytes
= io_tlb_nslabs
<< IO_TLB_SHIFT
;
214 while ((SLABS_PER_PAGE
<< order
) > IO_TLB_MIN_SLABS
) {
215 io_tlb_start
= (void *)__get_free_pages(GFP_DMA
| __GFP_NOWARN
,
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
= (phys_addr_t
*)
249 __get_free_pages(GFP_KERNEL
,
250 get_order(io_tlb_nslabs
*
251 sizeof(phys_addr_t
)));
252 if (!io_tlb_orig_addr
)
255 memset(io_tlb_orig_addr
, 0, io_tlb_nslabs
* sizeof(phys_addr_t
));
258 * Get the overflow emergency buffer
260 io_tlb_overflow_buffer
= (void *)__get_free_pages(GFP_DMA
,
261 get_order(io_tlb_overflow
));
262 if (!io_tlb_overflow_buffer
)
265 swiotlb_print_info(bytes
);
272 free_pages((unsigned long)io_tlb_orig_addr
,
273 get_order(io_tlb_nslabs
* sizeof(phys_addr_t
)));
274 io_tlb_orig_addr
= NULL
;
276 free_pages((unsigned long)io_tlb_list
, get_order(io_tlb_nslabs
*
281 free_pages((unsigned long)io_tlb_start
, order
);
284 io_tlb_nslabs
= req_nslabs
;
288 void __init
swiotlb_free(void)
290 if (!io_tlb_overflow_buffer
)
294 free_pages((unsigned long)io_tlb_overflow_buffer
,
295 get_order(io_tlb_overflow
));
296 free_pages((unsigned long)io_tlb_orig_addr
,
297 get_order(io_tlb_nslabs
* sizeof(phys_addr_t
)));
298 free_pages((unsigned long)io_tlb_list
, get_order(io_tlb_nslabs
*
300 free_pages((unsigned long)io_tlb_start
,
301 get_order(io_tlb_nslabs
<< IO_TLB_SHIFT
));
303 free_bootmem_late(__pa(io_tlb_overflow_buffer
),
305 free_bootmem_late(__pa(io_tlb_orig_addr
),
306 io_tlb_nslabs
* sizeof(phys_addr_t
));
307 free_bootmem_late(__pa(io_tlb_list
),
308 io_tlb_nslabs
* sizeof(int));
309 free_bootmem_late(__pa(io_tlb_start
),
310 io_tlb_nslabs
<< IO_TLB_SHIFT
);
314 static int is_swiotlb_buffer(phys_addr_t paddr
)
316 return paddr
>= virt_to_phys(io_tlb_start
) &&
317 paddr
< virt_to_phys(io_tlb_end
);
321 * Bounce: copy the swiotlb buffer back to the original dma location
323 static void swiotlb_bounce(phys_addr_t phys
, char *dma_addr
, size_t size
,
324 enum dma_data_direction dir
)
326 unsigned long pfn
= PFN_DOWN(phys
);
328 if (PageHighMem(pfn_to_page(pfn
))) {
329 /* The buffer does not have a mapping. Map it in and copy */
330 unsigned int offset
= phys
& ~PAGE_MASK
;
336 sz
= min_t(size_t, PAGE_SIZE
- offset
, size
);
338 local_irq_save(flags
);
339 buffer
= kmap_atomic(pfn_to_page(pfn
),
341 if (dir
== DMA_TO_DEVICE
)
342 memcpy(dma_addr
, buffer
+ offset
, sz
);
344 memcpy(buffer
+ offset
, dma_addr
, sz
);
345 kunmap_atomic(buffer
, KM_BOUNCE_READ
);
346 local_irq_restore(flags
);
354 if (dir
== DMA_TO_DEVICE
)
355 memcpy(dma_addr
, phys_to_virt(phys
), size
);
357 memcpy(phys_to_virt(phys
), dma_addr
, size
);
362 * Allocates bounce buffer and returns its kernel virtual address.
365 map_single(struct device
*hwdev
, phys_addr_t phys
, size_t size
, int dir
)
369 unsigned int nslots
, stride
, index
, wrap
;
371 unsigned long start_dma_addr
;
373 unsigned long offset_slots
;
374 unsigned long max_slots
;
376 mask
= dma_get_seg_boundary(hwdev
);
377 start_dma_addr
= swiotlb_virt_to_bus(hwdev
, io_tlb_start
) & mask
;
379 offset_slots
= ALIGN(start_dma_addr
, 1 << IO_TLB_SHIFT
) >> IO_TLB_SHIFT
;
382 * Carefully handle integer overflow which can occur when mask == ~0UL.
385 ? ALIGN(mask
+ 1, 1 << IO_TLB_SHIFT
) >> IO_TLB_SHIFT
386 : 1UL << (BITS_PER_LONG
- IO_TLB_SHIFT
);
389 * For mappings greater than a page, we limit the stride (and
390 * hence alignment) to a page size.
392 nslots
= ALIGN(size
, 1 << IO_TLB_SHIFT
) >> IO_TLB_SHIFT
;
393 if (size
> PAGE_SIZE
)
394 stride
= (1 << (PAGE_SHIFT
- IO_TLB_SHIFT
));
401 * Find suitable number of IO TLB entries size that will fit this
402 * request and allocate a buffer from that IO TLB pool.
404 spin_lock_irqsave(&io_tlb_lock
, flags
);
405 index
= ALIGN(io_tlb_index
, stride
);
406 if (index
>= io_tlb_nslabs
)
411 while (iommu_is_span_boundary(index
, nslots
, offset_slots
,
414 if (index
>= io_tlb_nslabs
)
421 * If we find a slot that indicates we have 'nslots' number of
422 * contiguous buffers, we allocate the buffers from that slot
423 * and mark the entries as '0' indicating unavailable.
425 if (io_tlb_list
[index
] >= nslots
) {
428 for (i
= index
; i
< (int) (index
+ nslots
); i
++)
430 for (i
= index
- 1; (OFFSET(i
, IO_TLB_SEGSIZE
) != IO_TLB_SEGSIZE
- 1) && io_tlb_list
[i
]; i
--)
431 io_tlb_list
[i
] = ++count
;
432 dma_addr
= io_tlb_start
+ (index
<< IO_TLB_SHIFT
);
435 * Update the indices to avoid searching in the next
438 io_tlb_index
= ((index
+ nslots
) < io_tlb_nslabs
439 ? (index
+ nslots
) : 0);
444 if (index
>= io_tlb_nslabs
)
446 } while (index
!= wrap
);
449 spin_unlock_irqrestore(&io_tlb_lock
, flags
);
452 spin_unlock_irqrestore(&io_tlb_lock
, flags
);
455 * Save away the mapping from the original address to the DMA address.
456 * This is needed when we sync the memory. Then we sync the buffer if
459 for (i
= 0; i
< nslots
; i
++)
460 io_tlb_orig_addr
[index
+i
] = phys
+ (i
<< IO_TLB_SHIFT
);
461 if (dir
== DMA_TO_DEVICE
|| dir
== DMA_BIDIRECTIONAL
)
462 swiotlb_bounce(phys
, dma_addr
, size
, DMA_TO_DEVICE
);
468 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
471 do_unmap_single(struct device
*hwdev
, char *dma_addr
, size_t size
, int dir
)
474 int i
, count
, nslots
= ALIGN(size
, 1 << IO_TLB_SHIFT
) >> IO_TLB_SHIFT
;
475 int index
= (dma_addr
- io_tlb_start
) >> IO_TLB_SHIFT
;
476 phys_addr_t phys
= io_tlb_orig_addr
[index
];
479 * First, sync the memory before unmapping the entry
481 if (phys
&& ((dir
== DMA_FROM_DEVICE
) || (dir
== DMA_BIDIRECTIONAL
)))
482 swiotlb_bounce(phys
, dma_addr
, size
, DMA_FROM_DEVICE
);
485 * Return the buffer to the free list by setting the corresponding
486 * entries to indicate the number of contigous entries available.
487 * While returning the entries to the free list, we merge the entries
488 * with slots below and above the pool being returned.
490 spin_lock_irqsave(&io_tlb_lock
, flags
);
492 count
= ((index
+ nslots
) < ALIGN(index
+ 1, IO_TLB_SEGSIZE
) ?
493 io_tlb_list
[index
+ nslots
] : 0);
495 * Step 1: return the slots to the free list, merging the
496 * slots with superceeding slots
498 for (i
= index
+ nslots
- 1; i
>= index
; i
--)
499 io_tlb_list
[i
] = ++count
;
501 * Step 2: merge the returned slots with the preceding slots,
502 * if available (non zero)
504 for (i
= index
- 1; (OFFSET(i
, IO_TLB_SEGSIZE
) != IO_TLB_SEGSIZE
-1) && io_tlb_list
[i
]; i
--)
505 io_tlb_list
[i
] = ++count
;
507 spin_unlock_irqrestore(&io_tlb_lock
, flags
);
511 sync_single(struct device
*hwdev
, char *dma_addr
, size_t size
,
514 int index
= (dma_addr
- io_tlb_start
) >> IO_TLB_SHIFT
;
515 phys_addr_t phys
= io_tlb_orig_addr
[index
];
517 phys
+= ((unsigned long)dma_addr
& ((1 << IO_TLB_SHIFT
) - 1));
521 if (likely(dir
== DMA_FROM_DEVICE
|| dir
== DMA_BIDIRECTIONAL
))
522 swiotlb_bounce(phys
, dma_addr
, size
, DMA_FROM_DEVICE
);
524 BUG_ON(dir
!= DMA_TO_DEVICE
);
526 case SYNC_FOR_DEVICE
:
527 if (likely(dir
== DMA_TO_DEVICE
|| dir
== DMA_BIDIRECTIONAL
))
528 swiotlb_bounce(phys
, dma_addr
, size
, DMA_TO_DEVICE
);
530 BUG_ON(dir
!= DMA_FROM_DEVICE
);
538 swiotlb_alloc_coherent(struct device
*hwdev
, size_t size
,
539 dma_addr_t
*dma_handle
, gfp_t flags
)
543 int order
= get_order(size
);
544 u64 dma_mask
= DMA_BIT_MASK(32);
546 if (hwdev
&& hwdev
->coherent_dma_mask
)
547 dma_mask
= hwdev
->coherent_dma_mask
;
549 ret
= (void *)__get_free_pages(flags
, order
);
550 if (ret
&& swiotlb_virt_to_bus(hwdev
, ret
) + size
> dma_mask
) {
552 * The allocated memory isn't reachable by the device.
554 free_pages((unsigned long) ret
, order
);
559 * We are either out of memory or the device can't DMA
560 * to GFP_DMA memory; fall back on map_single(), which
561 * will grab memory from the lowest available address range.
563 ret
= map_single(hwdev
, 0, size
, DMA_FROM_DEVICE
);
568 memset(ret
, 0, size
);
569 dev_addr
= swiotlb_virt_to_bus(hwdev
, ret
);
571 /* Confirm address can be DMA'd by device */
572 if (dev_addr
+ size
> dma_mask
) {
573 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
574 (unsigned long long)dma_mask
,
575 (unsigned long long)dev_addr
);
577 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
578 do_unmap_single(hwdev
, ret
, size
, DMA_TO_DEVICE
);
581 *dma_handle
= dev_addr
;
584 EXPORT_SYMBOL(swiotlb_alloc_coherent
);
587 swiotlb_free_coherent(struct device
*hwdev
, size_t size
, void *vaddr
,
590 phys_addr_t paddr
= dma_to_phys(hwdev
, dev_addr
);
592 WARN_ON(irqs_disabled());
593 if (!is_swiotlb_buffer(paddr
))
594 free_pages((unsigned long)vaddr
, get_order(size
));
596 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
597 do_unmap_single(hwdev
, vaddr
, size
, DMA_TO_DEVICE
);
599 EXPORT_SYMBOL(swiotlb_free_coherent
);
602 swiotlb_full(struct device
*dev
, size_t size
, int dir
, int do_panic
)
605 * Ran out of IOMMU space for this operation. This is very bad.
606 * Unfortunately the drivers cannot handle this operation properly.
607 * unless they check for dma_mapping_error (most don't)
608 * When the mapping is small enough return a static buffer to limit
609 * the damage, or panic when the transfer is too big.
611 printk(KERN_ERR
"DMA: Out of SW-IOMMU space for %zu bytes at "
612 "device %s\n", size
, dev
? dev_name(dev
) : "?");
614 if (size
<= io_tlb_overflow
|| !do_panic
)
617 if (dir
== DMA_BIDIRECTIONAL
)
618 panic("DMA: Random memory could be DMA accessed\n");
619 if (dir
== DMA_FROM_DEVICE
)
620 panic("DMA: Random memory could be DMA written\n");
621 if (dir
== DMA_TO_DEVICE
)
622 panic("DMA: Random memory could be DMA read\n");
626 * Map a single buffer of the indicated size for DMA in streaming mode. The
627 * physical address to use is returned.
629 * Once the device is given the dma address, the device owns this memory until
630 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
632 dma_addr_t
swiotlb_map_page(struct device
*dev
, struct page
*page
,
633 unsigned long offset
, size_t size
,
634 enum dma_data_direction dir
,
635 struct dma_attrs
*attrs
)
637 phys_addr_t phys
= page_to_phys(page
) + offset
;
638 dma_addr_t dev_addr
= phys_to_dma(dev
, phys
);
641 BUG_ON(dir
== DMA_NONE
);
643 * If the address happens to be in the device's DMA window,
644 * we can safely return the device addr and not worry about bounce
647 if (dma_capable(dev
, dev_addr
, size
) && !swiotlb_force
)
651 * Oh well, have to allocate and map a bounce buffer.
653 map
= map_single(dev
, phys
, size
, dir
);
655 swiotlb_full(dev
, size
, dir
, 1);
656 map
= io_tlb_overflow_buffer
;
659 dev_addr
= swiotlb_virt_to_bus(dev
, map
);
662 * Ensure that the address returned is DMA'ble
664 if (!dma_capable(dev
, dev_addr
, size
))
665 panic("map_single: bounce buffer is not DMA'ble");
669 EXPORT_SYMBOL_GPL(swiotlb_map_page
);
672 * Unmap a single streaming mode DMA translation. The dma_addr and size must
673 * match what was provided for in a previous swiotlb_map_page call. All
674 * other usages are undefined.
676 * After this call, reads by the cpu to the buffer are guaranteed to see
677 * whatever the device wrote there.
679 static void unmap_single(struct device
*hwdev
, dma_addr_t dev_addr
,
680 size_t size
, int dir
)
682 phys_addr_t paddr
= dma_to_phys(hwdev
, dev_addr
);
684 BUG_ON(dir
== DMA_NONE
);
686 if (is_swiotlb_buffer(paddr
)) {
687 do_unmap_single(hwdev
, phys_to_virt(paddr
), size
, dir
);
691 if (dir
!= DMA_FROM_DEVICE
)
695 * phys_to_virt doesn't work with hihgmem page but we could
696 * call dma_mark_clean() with hihgmem page here. However, we
697 * are fine since dma_mark_clean() is null on POWERPC. We can
698 * make dma_mark_clean() take a physical address if necessary.
700 dma_mark_clean(phys_to_virt(paddr
), size
);
703 void swiotlb_unmap_page(struct device
*hwdev
, dma_addr_t dev_addr
,
704 size_t size
, enum dma_data_direction dir
,
705 struct dma_attrs
*attrs
)
707 unmap_single(hwdev
, dev_addr
, size
, dir
);
709 EXPORT_SYMBOL_GPL(swiotlb_unmap_page
);
712 * Make physical memory consistent for a single streaming mode DMA translation
715 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
716 * using the cpu, yet do not wish to teardown the dma mapping, you must
717 * call this function before doing so. At the next point you give the dma
718 * address back to the card, you must first perform a
719 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
722 swiotlb_sync_single(struct device
*hwdev
, dma_addr_t dev_addr
,
723 size_t size
, int dir
, int target
)
725 phys_addr_t paddr
= dma_to_phys(hwdev
, dev_addr
);
727 BUG_ON(dir
== DMA_NONE
);
729 if (is_swiotlb_buffer(paddr
)) {
730 sync_single(hwdev
, phys_to_virt(paddr
), size
, dir
, target
);
734 if (dir
!= DMA_FROM_DEVICE
)
737 dma_mark_clean(phys_to_virt(paddr
), size
);
741 swiotlb_sync_single_for_cpu(struct device
*hwdev
, dma_addr_t dev_addr
,
742 size_t size
, enum dma_data_direction dir
)
744 swiotlb_sync_single(hwdev
, dev_addr
, size
, dir
, SYNC_FOR_CPU
);
746 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu
);
749 swiotlb_sync_single_for_device(struct device
*hwdev
, dma_addr_t dev_addr
,
750 size_t size
, enum dma_data_direction dir
)
752 swiotlb_sync_single(hwdev
, dev_addr
, size
, dir
, SYNC_FOR_DEVICE
);
754 EXPORT_SYMBOL(swiotlb_sync_single_for_device
);
757 * Same as above, but for a sub-range of the mapping.
760 swiotlb_sync_single_range(struct device
*hwdev
, dma_addr_t dev_addr
,
761 unsigned long offset
, size_t size
,
764 swiotlb_sync_single(hwdev
, dev_addr
+ offset
, size
, dir
, target
);
768 swiotlb_sync_single_range_for_cpu(struct device
*hwdev
, dma_addr_t dev_addr
,
769 unsigned long offset
, size_t size
,
770 enum dma_data_direction dir
)
772 swiotlb_sync_single_range(hwdev
, dev_addr
, offset
, size
, dir
,
775 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu
);
778 swiotlb_sync_single_range_for_device(struct device
*hwdev
, dma_addr_t dev_addr
,
779 unsigned long offset
, size_t size
,
780 enum dma_data_direction dir
)
782 swiotlb_sync_single_range(hwdev
, dev_addr
, offset
, size
, dir
,
785 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device
);
788 * Map a set of buffers described by scatterlist in streaming mode for DMA.
789 * This is the scatter-gather version of the above swiotlb_map_page
790 * interface. Here the scatter gather list elements are each tagged with the
791 * appropriate dma address and length. They are obtained via
792 * sg_dma_{address,length}(SG).
794 * NOTE: An implementation may be able to use a smaller number of
795 * DMA address/length pairs than there are SG table elements.
796 * (for example via virtual mapping capabilities)
797 * The routine returns the number of addr/length pairs actually
798 * used, at most nents.
800 * Device ownership issues as mentioned above for swiotlb_map_page are the
804 swiotlb_map_sg_attrs(struct device
*hwdev
, struct scatterlist
*sgl
, int nelems
,
805 enum dma_data_direction dir
, struct dma_attrs
*attrs
)
807 struct scatterlist
*sg
;
810 BUG_ON(dir
== DMA_NONE
);
812 for_each_sg(sgl
, sg
, nelems
, i
) {
813 phys_addr_t paddr
= sg_phys(sg
);
814 dma_addr_t dev_addr
= phys_to_dma(hwdev
, paddr
);
817 !dma_capable(hwdev
, dev_addr
, sg
->length
)) {
818 void *map
= map_single(hwdev
, sg_phys(sg
),
821 /* Don't panic here, we expect map_sg users
822 to do proper error handling. */
823 swiotlb_full(hwdev
, sg
->length
, dir
, 0);
824 swiotlb_unmap_sg_attrs(hwdev
, sgl
, i
, dir
,
826 sgl
[0].dma_length
= 0;
829 sg
->dma_address
= swiotlb_virt_to_bus(hwdev
, map
);
831 sg
->dma_address
= dev_addr
;
832 sg
->dma_length
= sg
->length
;
836 EXPORT_SYMBOL(swiotlb_map_sg_attrs
);
839 swiotlb_map_sg(struct device
*hwdev
, struct scatterlist
*sgl
, int nelems
,
842 return swiotlb_map_sg_attrs(hwdev
, sgl
, nelems
, dir
, NULL
);
844 EXPORT_SYMBOL(swiotlb_map_sg
);
847 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
848 * concerning calls here are the same as for swiotlb_unmap_page() above.
851 swiotlb_unmap_sg_attrs(struct device
*hwdev
, struct scatterlist
*sgl
,
852 int nelems
, enum dma_data_direction dir
, struct dma_attrs
*attrs
)
854 struct scatterlist
*sg
;
857 BUG_ON(dir
== DMA_NONE
);
859 for_each_sg(sgl
, sg
, nelems
, i
)
860 unmap_single(hwdev
, sg
->dma_address
, sg
->dma_length
, dir
);
863 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs
);
866 swiotlb_unmap_sg(struct device
*hwdev
, struct scatterlist
*sgl
, int nelems
,
869 return swiotlb_unmap_sg_attrs(hwdev
, sgl
, nelems
, dir
, NULL
);
871 EXPORT_SYMBOL(swiotlb_unmap_sg
);
874 * Make physical memory consistent for a set of streaming mode DMA translations
877 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
881 swiotlb_sync_sg(struct device
*hwdev
, struct scatterlist
*sgl
,
882 int nelems
, int dir
, int target
)
884 struct scatterlist
*sg
;
887 for_each_sg(sgl
, sg
, nelems
, i
)
888 swiotlb_sync_single(hwdev
, sg
->dma_address
,
889 sg
->dma_length
, dir
, target
);
893 swiotlb_sync_sg_for_cpu(struct device
*hwdev
, struct scatterlist
*sg
,
894 int nelems
, enum dma_data_direction dir
)
896 swiotlb_sync_sg(hwdev
, sg
, nelems
, dir
, SYNC_FOR_CPU
);
898 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu
);
901 swiotlb_sync_sg_for_device(struct device
*hwdev
, struct scatterlist
*sg
,
902 int nelems
, enum dma_data_direction dir
)
904 swiotlb_sync_sg(hwdev
, sg
, nelems
, dir
, SYNC_FOR_DEVICE
);
906 EXPORT_SYMBOL(swiotlb_sync_sg_for_device
);
909 swiotlb_dma_mapping_error(struct device
*hwdev
, dma_addr_t dma_addr
)
911 return (dma_addr
== swiotlb_virt_to_bus(hwdev
, io_tlb_overflow_buffer
));
913 EXPORT_SYMBOL(swiotlb_dma_mapping_error
);
916 * Return whether the given device DMA address mask can be supported
917 * properly. For example, if your device can only drive the low 24-bits
918 * during bus mastering, then you would pass 0x00ffffff as the mask to
922 swiotlb_dma_supported(struct device
*hwdev
, u64 mask
)
924 return swiotlb_virt_to_bus(hwdev
, io_tlb_end
- 1) <= mask
;
926 EXPORT_SYMBOL(swiotlb_dma_supported
);