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 #define pr_fmt(fmt) "software IO TLB: " fmt
22 #include <linux/cache.h>
23 #include <linux/dma-direct.h>
25 #include <linux/export.h>
26 #include <linux/spinlock.h>
27 #include <linux/string.h>
28 #include <linux/swiotlb.h>
29 #include <linux/pfn.h>
30 #include <linux/types.h>
31 #include <linux/ctype.h>
32 #include <linux/highmem.h>
33 #include <linux/gfp.h>
34 #include <linux/scatterlist.h>
35 #include <linux/mem_encrypt.h>
36 #include <linux/set_memory.h>
37 #ifdef CONFIG_DEBUG_FS
38 #include <linux/debugfs.h>
44 #include <linux/init.h>
45 #include <linux/memblock.h>
46 #include <linux/iommu-helper.h>
48 #define CREATE_TRACE_POINTS
49 #include <trace/events/swiotlb.h>
51 #define OFFSET(val,align) ((unsigned long) \
52 ( (val) & ( (align) - 1)))
54 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
57 * Minimum IO TLB size to bother booting with. Systems with mainly
58 * 64bit capable cards will only lightly use the swiotlb. If we can't
59 * allocate a contiguous 1MB, we're probably in trouble anyway.
61 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
63 enum swiotlb_force swiotlb_force
;
66 * Used to do a quick range check in swiotlb_tbl_unmap_single and
67 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
70 phys_addr_t io_tlb_start
, io_tlb_end
;
73 * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
74 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
76 static unsigned long io_tlb_nslabs
;
79 * The number of used IO TLB block
81 static unsigned long io_tlb_used
;
84 * This is a free list describing the number of free entries available from
87 static unsigned int *io_tlb_list
;
88 static unsigned int io_tlb_index
;
91 * Max segment that we can provide which (if pages are contingous) will
92 * not be bounced (unless SWIOTLB_FORCE is set).
94 unsigned int max_segment
;
97 * We need to save away the original address corresponding to a mapped entry
98 * for the sync operations.
100 #define INVALID_PHYS_ADDR (~(phys_addr_t)0)
101 static phys_addr_t
*io_tlb_orig_addr
;
104 * Protect the above data structures in the map and unmap calls
106 static DEFINE_SPINLOCK(io_tlb_lock
);
108 static int late_alloc
;
111 setup_io_tlb_npages(char *str
)
114 io_tlb_nslabs
= simple_strtoul(str
, &str
, 0);
115 /* avoid tail segment of size < IO_TLB_SEGSIZE */
116 io_tlb_nslabs
= ALIGN(io_tlb_nslabs
, IO_TLB_SEGSIZE
);
120 if (!strcmp(str
, "force")) {
121 swiotlb_force
= SWIOTLB_FORCE
;
122 } else if (!strcmp(str
, "noforce")) {
123 swiotlb_force
= SWIOTLB_NO_FORCE
;
129 early_param("swiotlb", setup_io_tlb_npages
);
131 unsigned long swiotlb_nr_tbl(void)
133 return io_tlb_nslabs
;
135 EXPORT_SYMBOL_GPL(swiotlb_nr_tbl
);
137 unsigned int swiotlb_max_segment(void)
141 EXPORT_SYMBOL_GPL(swiotlb_max_segment
);
143 void swiotlb_set_max_segment(unsigned int val
)
145 if (swiotlb_force
== SWIOTLB_FORCE
)
148 max_segment
= rounddown(val
, PAGE_SIZE
);
151 /* default to 64MB */
152 #define IO_TLB_DEFAULT_SIZE (64UL<<20)
153 unsigned long swiotlb_size_or_default(void)
157 size
= io_tlb_nslabs
<< IO_TLB_SHIFT
;
159 return size
? size
: (IO_TLB_DEFAULT_SIZE
);
162 static bool no_iotlb_memory
;
164 void swiotlb_print_info(void)
166 unsigned long bytes
= io_tlb_nslabs
<< IO_TLB_SHIFT
;
168 if (no_iotlb_memory
) {
169 pr_warn("No low mem\n");
173 pr_info("mapped [mem %#010llx-%#010llx] (%luMB)\n",
174 (unsigned long long)io_tlb_start
,
175 (unsigned long long)io_tlb_end
,
180 * Early SWIOTLB allocation may be too early to allow an architecture to
181 * perform the desired operations. This function allows the architecture to
182 * call SWIOTLB when the operations are possible. It needs to be called
183 * before the SWIOTLB memory is used.
185 void __init
swiotlb_update_mem_attributes(void)
190 if (no_iotlb_memory
|| late_alloc
)
193 vaddr
= phys_to_virt(io_tlb_start
);
194 bytes
= PAGE_ALIGN(io_tlb_nslabs
<< IO_TLB_SHIFT
);
195 set_memory_decrypted((unsigned long)vaddr
, bytes
>> PAGE_SHIFT
);
196 memset(vaddr
, 0, bytes
);
199 int __init
swiotlb_init_with_tbl(char *tlb
, unsigned long nslabs
, int verbose
)
201 unsigned long i
, bytes
;
204 bytes
= nslabs
<< IO_TLB_SHIFT
;
206 io_tlb_nslabs
= nslabs
;
207 io_tlb_start
= __pa(tlb
);
208 io_tlb_end
= io_tlb_start
+ bytes
;
211 * Allocate and initialize the free list array. This array is used
212 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
213 * between io_tlb_start and io_tlb_end.
215 alloc_size
= PAGE_ALIGN(io_tlb_nslabs
* sizeof(int));
216 io_tlb_list
= memblock_alloc(alloc_size
, PAGE_SIZE
);
218 panic("%s: Failed to allocate %zu bytes align=0x%lx\n",
219 __func__
, alloc_size
, PAGE_SIZE
);
221 alloc_size
= PAGE_ALIGN(io_tlb_nslabs
* sizeof(phys_addr_t
));
222 io_tlb_orig_addr
= memblock_alloc(alloc_size
, PAGE_SIZE
);
223 if (!io_tlb_orig_addr
)
224 panic("%s: Failed to allocate %zu bytes align=0x%lx\n",
225 __func__
, alloc_size
, PAGE_SIZE
);
227 for (i
= 0; i
< io_tlb_nslabs
; i
++) {
228 io_tlb_list
[i
] = IO_TLB_SEGSIZE
- OFFSET(i
, IO_TLB_SEGSIZE
);
229 io_tlb_orig_addr
[i
] = INVALID_PHYS_ADDR
;
234 swiotlb_print_info();
236 swiotlb_set_max_segment(io_tlb_nslabs
<< IO_TLB_SHIFT
);
241 * Statically reserve bounce buffer space and initialize bounce buffer data
242 * structures for the software IO TLB used to implement the DMA API.
245 swiotlb_init(int verbose
)
247 size_t default_size
= IO_TLB_DEFAULT_SIZE
;
248 unsigned char *vstart
;
251 if (!io_tlb_nslabs
) {
252 io_tlb_nslabs
= (default_size
>> IO_TLB_SHIFT
);
253 io_tlb_nslabs
= ALIGN(io_tlb_nslabs
, IO_TLB_SEGSIZE
);
256 bytes
= io_tlb_nslabs
<< IO_TLB_SHIFT
;
258 /* Get IO TLB memory from the low pages */
259 vstart
= memblock_alloc_low(PAGE_ALIGN(bytes
), PAGE_SIZE
);
260 if (vstart
&& !swiotlb_init_with_tbl(vstart
, io_tlb_nslabs
, verbose
))
264 memblock_free_early(io_tlb_start
,
265 PAGE_ALIGN(io_tlb_nslabs
<< IO_TLB_SHIFT
));
266 pr_warn("Cannot allocate buffer");
267 no_iotlb_memory
= true;
271 * Systems with larger DMA zones (those that don't support ISA) can
272 * initialize the swiotlb later using the slab allocator if needed.
273 * This should be just like above, but with some error catching.
276 swiotlb_late_init_with_default_size(size_t default_size
)
278 unsigned long bytes
, req_nslabs
= io_tlb_nslabs
;
279 unsigned char *vstart
= NULL
;
283 if (!io_tlb_nslabs
) {
284 io_tlb_nslabs
= (default_size
>> IO_TLB_SHIFT
);
285 io_tlb_nslabs
= ALIGN(io_tlb_nslabs
, IO_TLB_SEGSIZE
);
289 * Get IO TLB memory from the low pages
291 order
= get_order(io_tlb_nslabs
<< IO_TLB_SHIFT
);
292 io_tlb_nslabs
= SLABS_PER_PAGE
<< order
;
293 bytes
= io_tlb_nslabs
<< IO_TLB_SHIFT
;
295 while ((SLABS_PER_PAGE
<< order
) > IO_TLB_MIN_SLABS
) {
296 vstart
= (void *)__get_free_pages(GFP_DMA
| __GFP_NOWARN
,
304 io_tlb_nslabs
= req_nslabs
;
307 if (order
!= get_order(bytes
)) {
308 pr_warn("only able to allocate %ld MB\n",
309 (PAGE_SIZE
<< order
) >> 20);
310 io_tlb_nslabs
= SLABS_PER_PAGE
<< order
;
312 rc
= swiotlb_late_init_with_tbl(vstart
, io_tlb_nslabs
);
314 free_pages((unsigned long)vstart
, order
);
320 swiotlb_late_init_with_tbl(char *tlb
, unsigned long nslabs
)
322 unsigned long i
, bytes
;
324 bytes
= nslabs
<< IO_TLB_SHIFT
;
326 io_tlb_nslabs
= nslabs
;
327 io_tlb_start
= virt_to_phys(tlb
);
328 io_tlb_end
= io_tlb_start
+ bytes
;
330 set_memory_decrypted((unsigned long)tlb
, bytes
>> PAGE_SHIFT
);
331 memset(tlb
, 0, bytes
);
334 * Allocate and initialize the free list array. This array is used
335 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
336 * between io_tlb_start and io_tlb_end.
338 io_tlb_list
= (unsigned int *)__get_free_pages(GFP_KERNEL
,
339 get_order(io_tlb_nslabs
* sizeof(int)));
343 io_tlb_orig_addr
= (phys_addr_t
*)
344 __get_free_pages(GFP_KERNEL
,
345 get_order(io_tlb_nslabs
*
346 sizeof(phys_addr_t
)));
347 if (!io_tlb_orig_addr
)
350 for (i
= 0; i
< io_tlb_nslabs
; i
++) {
351 io_tlb_list
[i
] = IO_TLB_SEGSIZE
- OFFSET(i
, IO_TLB_SEGSIZE
);
352 io_tlb_orig_addr
[i
] = INVALID_PHYS_ADDR
;
356 swiotlb_print_info();
360 swiotlb_set_max_segment(io_tlb_nslabs
<< IO_TLB_SHIFT
);
365 free_pages((unsigned long)io_tlb_list
, get_order(io_tlb_nslabs
*
376 void __init
swiotlb_exit(void)
378 if (!io_tlb_orig_addr
)
382 free_pages((unsigned long)io_tlb_orig_addr
,
383 get_order(io_tlb_nslabs
* sizeof(phys_addr_t
)));
384 free_pages((unsigned long)io_tlb_list
, get_order(io_tlb_nslabs
*
386 free_pages((unsigned long)phys_to_virt(io_tlb_start
),
387 get_order(io_tlb_nslabs
<< IO_TLB_SHIFT
));
389 memblock_free_late(__pa(io_tlb_orig_addr
),
390 PAGE_ALIGN(io_tlb_nslabs
* sizeof(phys_addr_t
)));
391 memblock_free_late(__pa(io_tlb_list
),
392 PAGE_ALIGN(io_tlb_nslabs
* sizeof(int)));
393 memblock_free_late(io_tlb_start
,
394 PAGE_ALIGN(io_tlb_nslabs
<< IO_TLB_SHIFT
));
403 * Bounce: copy the swiotlb buffer from or back to the original dma location
405 static void swiotlb_bounce(phys_addr_t orig_addr
, phys_addr_t tlb_addr
,
406 size_t size
, enum dma_data_direction dir
)
408 unsigned long pfn
= PFN_DOWN(orig_addr
);
409 unsigned char *vaddr
= phys_to_virt(tlb_addr
);
411 if (PageHighMem(pfn_to_page(pfn
))) {
412 /* The buffer does not have a mapping. Map it in and copy */
413 unsigned int offset
= orig_addr
& ~PAGE_MASK
;
419 sz
= min_t(size_t, PAGE_SIZE
- offset
, size
);
421 local_irq_save(flags
);
422 buffer
= kmap_atomic(pfn_to_page(pfn
));
423 if (dir
== DMA_TO_DEVICE
)
424 memcpy(vaddr
, buffer
+ offset
, sz
);
426 memcpy(buffer
+ offset
, vaddr
, sz
);
427 kunmap_atomic(buffer
);
428 local_irq_restore(flags
);
435 } else if (dir
== DMA_TO_DEVICE
) {
436 memcpy(vaddr
, phys_to_virt(orig_addr
), size
);
438 memcpy(phys_to_virt(orig_addr
), vaddr
, size
);
442 phys_addr_t
swiotlb_tbl_map_single(struct device
*hwdev
,
443 dma_addr_t tbl_dma_addr
,
444 phys_addr_t orig_addr
, size_t size
,
445 enum dma_data_direction dir
,
449 phys_addr_t tlb_addr
;
450 unsigned int nslots
, stride
, index
, wrap
;
453 unsigned long offset_slots
;
454 unsigned long max_slots
;
457 panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
459 if (mem_encrypt_active())
460 pr_warn_once("%s is active and system is using DMA bounce buffers\n",
461 sme_active() ? "SME" : "SEV");
463 mask
= dma_get_seg_boundary(hwdev
);
465 tbl_dma_addr
&= mask
;
467 offset_slots
= ALIGN(tbl_dma_addr
, 1 << IO_TLB_SHIFT
) >> IO_TLB_SHIFT
;
470 * Carefully handle integer overflow which can occur when mask == ~0UL.
473 ? ALIGN(mask
+ 1, 1 << IO_TLB_SHIFT
) >> IO_TLB_SHIFT
474 : 1UL << (BITS_PER_LONG
- IO_TLB_SHIFT
);
477 * For mappings greater than or equal to a page, we limit the stride
478 * (and hence alignment) to a page size.
480 nslots
= ALIGN(size
, 1 << IO_TLB_SHIFT
) >> IO_TLB_SHIFT
;
481 if (size
>= PAGE_SIZE
)
482 stride
= (1 << (PAGE_SHIFT
- IO_TLB_SHIFT
));
489 * Find suitable number of IO TLB entries size that will fit this
490 * request and allocate a buffer from that IO TLB pool.
492 spin_lock_irqsave(&io_tlb_lock
, flags
);
494 if (unlikely(nslots
> io_tlb_nslabs
- io_tlb_used
))
497 index
= ALIGN(io_tlb_index
, stride
);
498 if (index
>= io_tlb_nslabs
)
503 while (iommu_is_span_boundary(index
, nslots
, offset_slots
,
506 if (index
>= io_tlb_nslabs
)
513 * If we find a slot that indicates we have 'nslots' number of
514 * contiguous buffers, we allocate the buffers from that slot
515 * and mark the entries as '0' indicating unavailable.
517 if (io_tlb_list
[index
] >= nslots
) {
520 for (i
= index
; i
< (int) (index
+ nslots
); i
++)
522 for (i
= index
- 1; (OFFSET(i
, IO_TLB_SEGSIZE
) != IO_TLB_SEGSIZE
- 1) && io_tlb_list
[i
]; i
--)
523 io_tlb_list
[i
] = ++count
;
524 tlb_addr
= io_tlb_start
+ (index
<< IO_TLB_SHIFT
);
527 * Update the indices to avoid searching in the next
530 io_tlb_index
= ((index
+ nslots
) < io_tlb_nslabs
531 ? (index
+ nslots
) : 0);
536 if (index
>= io_tlb_nslabs
)
538 } while (index
!= wrap
);
541 spin_unlock_irqrestore(&io_tlb_lock
, flags
);
542 if (!(attrs
& DMA_ATTR_NO_WARN
) && printk_ratelimit())
543 dev_warn(hwdev
, "swiotlb buffer is full (sz: %zd bytes)\n", size
);
544 return DMA_MAPPING_ERROR
;
546 io_tlb_used
+= nslots
;
547 spin_unlock_irqrestore(&io_tlb_lock
, flags
);
550 * Save away the mapping from the original address to the DMA address.
551 * This is needed when we sync the memory. Then we sync the buffer if
554 for (i
= 0; i
< nslots
; i
++)
555 io_tlb_orig_addr
[index
+i
] = orig_addr
+ (i
<< IO_TLB_SHIFT
);
556 if (!(attrs
& DMA_ATTR_SKIP_CPU_SYNC
) &&
557 (dir
== DMA_TO_DEVICE
|| dir
== DMA_BIDIRECTIONAL
))
558 swiotlb_bounce(orig_addr
, tlb_addr
, size
, DMA_TO_DEVICE
);
564 * tlb_addr is the physical address of the bounce buffer to unmap.
566 void swiotlb_tbl_unmap_single(struct device
*hwdev
, phys_addr_t tlb_addr
,
567 size_t size
, enum dma_data_direction dir
,
571 int i
, count
, nslots
= ALIGN(size
, 1 << IO_TLB_SHIFT
) >> IO_TLB_SHIFT
;
572 int index
= (tlb_addr
- io_tlb_start
) >> IO_TLB_SHIFT
;
573 phys_addr_t orig_addr
= io_tlb_orig_addr
[index
];
576 * First, sync the memory before unmapping the entry
578 if (orig_addr
!= INVALID_PHYS_ADDR
&&
579 !(attrs
& DMA_ATTR_SKIP_CPU_SYNC
) &&
580 ((dir
== DMA_FROM_DEVICE
) || (dir
== DMA_BIDIRECTIONAL
)))
581 swiotlb_bounce(orig_addr
, tlb_addr
, size
, DMA_FROM_DEVICE
);
584 * Return the buffer to the free list by setting the corresponding
585 * entries to indicate the number of contiguous entries available.
586 * While returning the entries to the free list, we merge the entries
587 * with slots below and above the pool being returned.
589 spin_lock_irqsave(&io_tlb_lock
, flags
);
591 count
= ((index
+ nslots
) < ALIGN(index
+ 1, IO_TLB_SEGSIZE
) ?
592 io_tlb_list
[index
+ nslots
] : 0);
594 * Step 1: return the slots to the free list, merging the
595 * slots with superceeding slots
597 for (i
= index
+ nslots
- 1; i
>= index
; i
--) {
598 io_tlb_list
[i
] = ++count
;
599 io_tlb_orig_addr
[i
] = INVALID_PHYS_ADDR
;
602 * Step 2: merge the returned slots with the preceding slots,
603 * if available (non zero)
605 for (i
= index
- 1; (OFFSET(i
, IO_TLB_SEGSIZE
) != IO_TLB_SEGSIZE
-1) && io_tlb_list
[i
]; i
--)
606 io_tlb_list
[i
] = ++count
;
608 io_tlb_used
-= nslots
;
610 spin_unlock_irqrestore(&io_tlb_lock
, flags
);
613 void swiotlb_tbl_sync_single(struct device
*hwdev
, phys_addr_t tlb_addr
,
614 size_t size
, enum dma_data_direction dir
,
615 enum dma_sync_target target
)
617 int index
= (tlb_addr
- io_tlb_start
) >> IO_TLB_SHIFT
;
618 phys_addr_t orig_addr
= io_tlb_orig_addr
[index
];
620 if (orig_addr
== INVALID_PHYS_ADDR
)
622 orig_addr
+= (unsigned long)tlb_addr
& ((1 << IO_TLB_SHIFT
) - 1);
626 if (likely(dir
== DMA_FROM_DEVICE
|| dir
== DMA_BIDIRECTIONAL
))
627 swiotlb_bounce(orig_addr
, tlb_addr
,
628 size
, DMA_FROM_DEVICE
);
630 BUG_ON(dir
!= DMA_TO_DEVICE
);
632 case SYNC_FOR_DEVICE
:
633 if (likely(dir
== DMA_TO_DEVICE
|| dir
== DMA_BIDIRECTIONAL
))
634 swiotlb_bounce(orig_addr
, tlb_addr
,
635 size
, DMA_TO_DEVICE
);
637 BUG_ON(dir
!= DMA_FROM_DEVICE
);
645 * Create a swiotlb mapping for the buffer at @phys, and in case of DMAing
646 * to the device copy the data into it as well.
648 bool swiotlb_map(struct device
*dev
, phys_addr_t
*phys
, dma_addr_t
*dma_addr
,
649 size_t size
, enum dma_data_direction dir
, unsigned long attrs
)
651 trace_swiotlb_bounced(dev
, *dma_addr
, size
, swiotlb_force
);
653 if (unlikely(swiotlb_force
== SWIOTLB_NO_FORCE
)) {
654 dev_warn_ratelimited(dev
,
655 "Cannot do DMA to address %pa\n", phys
);
659 /* Oh well, have to allocate and map a bounce buffer. */
660 *phys
= swiotlb_tbl_map_single(dev
, __phys_to_dma(dev
, io_tlb_start
),
661 *phys
, size
, dir
, attrs
);
662 if (*phys
== DMA_MAPPING_ERROR
)
665 /* Ensure that the address returned is DMA'ble */
666 *dma_addr
= __phys_to_dma(dev
, *phys
);
667 if (unlikely(!dma_capable(dev
, *dma_addr
, size
))) {
668 swiotlb_tbl_unmap_single(dev
, *phys
, size
, dir
,
669 attrs
| DMA_ATTR_SKIP_CPU_SYNC
);
676 size_t swiotlb_max_mapping_size(struct device
*dev
)
678 return ((size_t)1 << IO_TLB_SHIFT
) * IO_TLB_SEGSIZE
;
681 bool is_swiotlb_active(void)
684 * When SWIOTLB is initialized, even if io_tlb_start points to physical
685 * address zero, io_tlb_end surely doesn't.
687 return io_tlb_end
!= 0;
690 #ifdef CONFIG_DEBUG_FS
692 static int __init
swiotlb_create_debugfs(void)
694 struct dentry
*d_swiotlb_usage
;
697 d_swiotlb_usage
= debugfs_create_dir("swiotlb", NULL
);
699 if (!d_swiotlb_usage
)
702 ent
= debugfs_create_ulong("io_tlb_nslabs", 0400,
703 d_swiotlb_usage
, &io_tlb_nslabs
);
707 ent
= debugfs_create_ulong("io_tlb_used", 0400,
708 d_swiotlb_usage
, &io_tlb_used
);
715 debugfs_remove_recursive(d_swiotlb_usage
);
719 late_initcall(swiotlb_create_debugfs
);