1 |
62 |
marcus.erl |
/*
|
2 |
|
|
* Dynamic DMA mapping support.
|
3 |
|
|
*
|
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>
|
10 |
|
|
*
|
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 |
|
|
*/
|
18 |
|
|
|
19 |
|
|
#include <linux/cache.h>
|
20 |
|
|
#include <linux/dma-mapping.h>
|
21 |
|
|
#include <linux/mm.h>
|
22 |
|
|
#include <linux/module.h>
|
23 |
|
|
#include <linux/spinlock.h>
|
24 |
|
|
#include <linux/string.h>
|
25 |
|
|
#include <linux/types.h>
|
26 |
|
|
#include <linux/ctype.h>
|
27 |
|
|
|
28 |
|
|
#include <asm/io.h>
|
29 |
|
|
#include <asm/dma.h>
|
30 |
|
|
#include <asm/scatterlist.h>
|
31 |
|
|
|
32 |
|
|
#include <linux/init.h>
|
33 |
|
|
#include <linux/bootmem.h>
|
34 |
|
|
|
35 |
|
|
#define OFFSET(val,align) ((unsigned long) \
|
36 |
|
|
( (val) & ( (align) - 1)))
|
37 |
|
|
|
38 |
|
|
#define SG_ENT_VIRT_ADDRESS(sg) (sg_virt((sg)))
|
39 |
|
|
#define SG_ENT_PHYS_ADDRESS(sg) virt_to_bus(SG_ENT_VIRT_ADDRESS(sg))
|
40 |
|
|
|
41 |
|
|
/*
|
42 |
|
|
* Maximum allowable number of contiguous slabs to map,
|
43 |
|
|
* must be a power of 2. What is the appropriate value ?
|
44 |
|
|
* The complexity of {map,unmap}_single is linearly dependent on this value.
|
45 |
|
|
*/
|
46 |
|
|
#define IO_TLB_SEGSIZE 128
|
47 |
|
|
|
48 |
|
|
/*
|
49 |
|
|
* log of the size of each IO TLB slab. The number of slabs is command line
|
50 |
|
|
* controllable.
|
51 |
|
|
*/
|
52 |
|
|
#define IO_TLB_SHIFT 11
|
53 |
|
|
|
54 |
|
|
#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
|
55 |
|
|
|
56 |
|
|
/*
|
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.
|
60 |
|
|
*/
|
61 |
|
|
#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
|
62 |
|
|
|
63 |
|
|
/*
|
64 |
|
|
* Enumeration for sync targets
|
65 |
|
|
*/
|
66 |
|
|
enum dma_sync_target {
|
67 |
|
|
SYNC_FOR_CPU = 0,
|
68 |
|
|
SYNC_FOR_DEVICE = 1,
|
69 |
|
|
};
|
70 |
|
|
|
71 |
|
|
int swiotlb_force;
|
72 |
|
|
|
73 |
|
|
/*
|
74 |
|
|
* Used to do a quick range check in swiotlb_unmap_single and
|
75 |
|
|
* swiotlb_sync_single_*, to see if the memory was in fact allocated by this
|
76 |
|
|
* API.
|
77 |
|
|
*/
|
78 |
|
|
static char *io_tlb_start, *io_tlb_end;
|
79 |
|
|
|
80 |
|
|
/*
|
81 |
|
|
* The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and
|
82 |
|
|
* io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
|
83 |
|
|
*/
|
84 |
|
|
static unsigned long io_tlb_nslabs;
|
85 |
|
|
|
86 |
|
|
/*
|
87 |
|
|
* When the IOMMU overflows we return a fallback buffer. This sets the size.
|
88 |
|
|
*/
|
89 |
|
|
static unsigned long io_tlb_overflow = 32*1024;
|
90 |
|
|
|
91 |
|
|
void *io_tlb_overflow_buffer;
|
92 |
|
|
|
93 |
|
|
/*
|
94 |
|
|
* This is a free list describing the number of free entries available from
|
95 |
|
|
* each index
|
96 |
|
|
*/
|
97 |
|
|
static unsigned int *io_tlb_list;
|
98 |
|
|
static unsigned int io_tlb_index;
|
99 |
|
|
|
100 |
|
|
/*
|
101 |
|
|
* We need to save away the original address corresponding to a mapped entry
|
102 |
|
|
* for the sync operations.
|
103 |
|
|
*/
|
104 |
|
|
static unsigned char **io_tlb_orig_addr;
|
105 |
|
|
|
106 |
|
|
/*
|
107 |
|
|
* Protect the above data structures in the map and unmap calls
|
108 |
|
|
*/
|
109 |
|
|
static DEFINE_SPINLOCK(io_tlb_lock);
|
110 |
|
|
|
111 |
|
|
static int __init
|
112 |
|
|
setup_io_tlb_npages(char *str)
|
113 |
|
|
{
|
114 |
|
|
if (isdigit(*str)) {
|
115 |
|
|
io_tlb_nslabs = simple_strtoul(str, &str, 0);
|
116 |
|
|
/* avoid tail segment of size < IO_TLB_SEGSIZE */
|
117 |
|
|
io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
|
118 |
|
|
}
|
119 |
|
|
if (*str == ',')
|
120 |
|
|
++str;
|
121 |
|
|
if (!strcmp(str, "force"))
|
122 |
|
|
swiotlb_force = 1;
|
123 |
|
|
return 1;
|
124 |
|
|
}
|
125 |
|
|
__setup("swiotlb=", setup_io_tlb_npages);
|
126 |
|
|
/* make io_tlb_overflow tunable too? */
|
127 |
|
|
|
128 |
|
|
/*
|
129 |
|
|
* Statically reserve bounce buffer space and initialize bounce buffer data
|
130 |
|
|
* structures for the software IO TLB used to implement the DMA API.
|
131 |
|
|
*/
|
132 |
|
|
void __init
|
133 |
|
|
swiotlb_init_with_default_size(size_t default_size)
|
134 |
|
|
{
|
135 |
|
|
unsigned long i, bytes;
|
136 |
|
|
|
137 |
|
|
if (!io_tlb_nslabs) {
|
138 |
|
|
io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
|
139 |
|
|
io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
|
140 |
|
|
}
|
141 |
|
|
|
142 |
|
|
bytes = io_tlb_nslabs << IO_TLB_SHIFT;
|
143 |
|
|
|
144 |
|
|
/*
|
145 |
|
|
* Get IO TLB memory from the low pages
|
146 |
|
|
*/
|
147 |
|
|
io_tlb_start = alloc_bootmem_low_pages(bytes);
|
148 |
|
|
if (!io_tlb_start)
|
149 |
|
|
panic("Cannot allocate SWIOTLB buffer");
|
150 |
|
|
io_tlb_end = io_tlb_start + bytes;
|
151 |
|
|
|
152 |
|
|
/*
|
153 |
|
|
* Allocate and initialize the free list array. This array is used
|
154 |
|
|
* to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
|
155 |
|
|
* between io_tlb_start and io_tlb_end.
|
156 |
|
|
*/
|
157 |
|
|
io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
|
158 |
|
|
for (i = 0; i < io_tlb_nslabs; i++)
|
159 |
|
|
io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
|
160 |
|
|
io_tlb_index = 0;
|
161 |
|
|
io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(char *));
|
162 |
|
|
|
163 |
|
|
/*
|
164 |
|
|
* Get the overflow emergency buffer
|
165 |
|
|
*/
|
166 |
|
|
io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
|
167 |
|
|
if (!io_tlb_overflow_buffer)
|
168 |
|
|
panic("Cannot allocate SWIOTLB overflow buffer!\n");
|
169 |
|
|
|
170 |
|
|
printk(KERN_INFO "Placing software IO TLB between 0x%lx - 0x%lx\n",
|
171 |
|
|
virt_to_bus(io_tlb_start), virt_to_bus(io_tlb_end));
|
172 |
|
|
}
|
173 |
|
|
|
174 |
|
|
void __init
|
175 |
|
|
swiotlb_init(void)
|
176 |
|
|
{
|
177 |
|
|
swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
|
178 |
|
|
}
|
179 |
|
|
|
180 |
|
|
/*
|
181 |
|
|
* Systems with larger DMA zones (those that don't support ISA) can
|
182 |
|
|
* initialize the swiotlb later using the slab allocator if needed.
|
183 |
|
|
* This should be just like above, but with some error catching.
|
184 |
|
|
*/
|
185 |
|
|
int
|
186 |
|
|
swiotlb_late_init_with_default_size(size_t default_size)
|
187 |
|
|
{
|
188 |
|
|
unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
|
189 |
|
|
unsigned int order;
|
190 |
|
|
|
191 |
|
|
if (!io_tlb_nslabs) {
|
192 |
|
|
io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
|
193 |
|
|
io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
|
194 |
|
|
}
|
195 |
|
|
|
196 |
|
|
/*
|
197 |
|
|
* Get IO TLB memory from the low pages
|
198 |
|
|
*/
|
199 |
|
|
order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
|
200 |
|
|
io_tlb_nslabs = SLABS_PER_PAGE << order;
|
201 |
|
|
bytes = io_tlb_nslabs << IO_TLB_SHIFT;
|
202 |
|
|
|
203 |
|
|
while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
|
204 |
|
|
io_tlb_start = (char *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
|
205 |
|
|
order);
|
206 |
|
|
if (io_tlb_start)
|
207 |
|
|
break;
|
208 |
|
|
order--;
|
209 |
|
|
}
|
210 |
|
|
|
211 |
|
|
if (!io_tlb_start)
|
212 |
|
|
goto cleanup1;
|
213 |
|
|
|
214 |
|
|
if (order != get_order(bytes)) {
|
215 |
|
|
printk(KERN_WARNING "Warning: only able to allocate %ld MB "
|
216 |
|
|
"for software IO TLB\n", (PAGE_SIZE << order) >> 20);
|
217 |
|
|
io_tlb_nslabs = SLABS_PER_PAGE << order;
|
218 |
|
|
bytes = io_tlb_nslabs << IO_TLB_SHIFT;
|
219 |
|
|
}
|
220 |
|
|
io_tlb_end = io_tlb_start + bytes;
|
221 |
|
|
memset(io_tlb_start, 0, bytes);
|
222 |
|
|
|
223 |
|
|
/*
|
224 |
|
|
* Allocate and initialize the free list array. This array is used
|
225 |
|
|
* to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
|
226 |
|
|
* between io_tlb_start and io_tlb_end.
|
227 |
|
|
*/
|
228 |
|
|
io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
|
229 |
|
|
get_order(io_tlb_nslabs * sizeof(int)));
|
230 |
|
|
if (!io_tlb_list)
|
231 |
|
|
goto cleanup2;
|
232 |
|
|
|
233 |
|
|
for (i = 0; i < io_tlb_nslabs; i++)
|
234 |
|
|
io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
|
235 |
|
|
io_tlb_index = 0;
|
236 |
|
|
|
237 |
|
|
io_tlb_orig_addr = (unsigned char **)__get_free_pages(GFP_KERNEL,
|
238 |
|
|
get_order(io_tlb_nslabs * sizeof(char *)));
|
239 |
|
|
if (!io_tlb_orig_addr)
|
240 |
|
|
goto cleanup3;
|
241 |
|
|
|
242 |
|
|
memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(char *));
|
243 |
|
|
|
244 |
|
|
/*
|
245 |
|
|
* Get the overflow emergency buffer
|
246 |
|
|
*/
|
247 |
|
|
io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
|
248 |
|
|
get_order(io_tlb_overflow));
|
249 |
|
|
if (!io_tlb_overflow_buffer)
|
250 |
|
|
goto cleanup4;
|
251 |
|
|
|
252 |
|
|
printk(KERN_INFO "Placing %luMB software IO TLB between 0x%lx - "
|
253 |
|
|
"0x%lx\n", bytes >> 20,
|
254 |
|
|
virt_to_bus(io_tlb_start), virt_to_bus(io_tlb_end));
|
255 |
|
|
|
256 |
|
|
return 0;
|
257 |
|
|
|
258 |
|
|
cleanup4:
|
259 |
|
|
free_pages((unsigned long)io_tlb_orig_addr, get_order(io_tlb_nslabs *
|
260 |
|
|
sizeof(char *)));
|
261 |
|
|
io_tlb_orig_addr = NULL;
|
262 |
|
|
cleanup3:
|
263 |
|
|
free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
|
264 |
|
|
sizeof(int)));
|
265 |
|
|
io_tlb_list = NULL;
|
266 |
|
|
cleanup2:
|
267 |
|
|
io_tlb_end = NULL;
|
268 |
|
|
free_pages((unsigned long)io_tlb_start, order);
|
269 |
|
|
io_tlb_start = NULL;
|
270 |
|
|
cleanup1:
|
271 |
|
|
io_tlb_nslabs = req_nslabs;
|
272 |
|
|
return -ENOMEM;
|
273 |
|
|
}
|
274 |
|
|
|
275 |
|
|
static int
|
276 |
|
|
address_needs_mapping(struct device *hwdev, dma_addr_t addr)
|
277 |
|
|
{
|
278 |
|
|
dma_addr_t mask = 0xffffffff;
|
279 |
|
|
/* If the device has a mask, use it, otherwise default to 32 bits */
|
280 |
|
|
if (hwdev && hwdev->dma_mask)
|
281 |
|
|
mask = *hwdev->dma_mask;
|
282 |
|
|
return (addr & ~mask) != 0;
|
283 |
|
|
}
|
284 |
|
|
|
285 |
|
|
/*
|
286 |
|
|
* Allocates bounce buffer and returns its kernel virtual address.
|
287 |
|
|
*/
|
288 |
|
|
static void *
|
289 |
|
|
map_single(struct device *hwdev, char *buffer, size_t size, int dir)
|
290 |
|
|
{
|
291 |
|
|
unsigned long flags;
|
292 |
|
|
char *dma_addr;
|
293 |
|
|
unsigned int nslots, stride, index, wrap;
|
294 |
|
|
int i;
|
295 |
|
|
|
296 |
|
|
/*
|
297 |
|
|
* For mappings greater than a page, we limit the stride (and
|
298 |
|
|
* hence alignment) to a page size.
|
299 |
|
|
*/
|
300 |
|
|
nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
|
301 |
|
|
if (size > PAGE_SIZE)
|
302 |
|
|
stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
|
303 |
|
|
else
|
304 |
|
|
stride = 1;
|
305 |
|
|
|
306 |
|
|
BUG_ON(!nslots);
|
307 |
|
|
|
308 |
|
|
/*
|
309 |
|
|
* Find suitable number of IO TLB entries size that will fit this
|
310 |
|
|
* request and allocate a buffer from that IO TLB pool.
|
311 |
|
|
*/
|
312 |
|
|
spin_lock_irqsave(&io_tlb_lock, flags);
|
313 |
|
|
{
|
314 |
|
|
wrap = index = ALIGN(io_tlb_index, stride);
|
315 |
|
|
|
316 |
|
|
if (index >= io_tlb_nslabs)
|
317 |
|
|
wrap = index = 0;
|
318 |
|
|
|
319 |
|
|
do {
|
320 |
|
|
/*
|
321 |
|
|
* If we find a slot that indicates we have 'nslots'
|
322 |
|
|
* number of contiguous buffers, we allocate the
|
323 |
|
|
* buffers from that slot and mark the entries as '0'
|
324 |
|
|
* indicating unavailable.
|
325 |
|
|
*/
|
326 |
|
|
if (io_tlb_list[index] >= nslots) {
|
327 |
|
|
int count = 0;
|
328 |
|
|
|
329 |
|
|
for (i = index; i < (int) (index + nslots); i++)
|
330 |
|
|
io_tlb_list[i] = 0;
|
331 |
|
|
for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
|
332 |
|
|
io_tlb_list[i] = ++count;
|
333 |
|
|
dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
|
334 |
|
|
|
335 |
|
|
/*
|
336 |
|
|
* Update the indices to avoid searching in
|
337 |
|
|
* the next round.
|
338 |
|
|
*/
|
339 |
|
|
io_tlb_index = ((index + nslots) < io_tlb_nslabs
|
340 |
|
|
? (index + nslots) : 0);
|
341 |
|
|
|
342 |
|
|
goto found;
|
343 |
|
|
}
|
344 |
|
|
index += stride;
|
345 |
|
|
if (index >= io_tlb_nslabs)
|
346 |
|
|
index = 0;
|
347 |
|
|
} while (index != wrap);
|
348 |
|
|
|
349 |
|
|
spin_unlock_irqrestore(&io_tlb_lock, flags);
|
350 |
|
|
return NULL;
|
351 |
|
|
}
|
352 |
|
|
found:
|
353 |
|
|
spin_unlock_irqrestore(&io_tlb_lock, flags);
|
354 |
|
|
|
355 |
|
|
/*
|
356 |
|
|
* Save away the mapping from the original address to the DMA address.
|
357 |
|
|
* This is needed when we sync the memory. Then we sync the buffer if
|
358 |
|
|
* needed.
|
359 |
|
|
*/
|
360 |
|
|
for (i = 0; i < nslots; i++)
|
361 |
|
|
io_tlb_orig_addr[index+i] = buffer + (i << IO_TLB_SHIFT);
|
362 |
|
|
if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
|
363 |
|
|
memcpy(dma_addr, buffer, size);
|
364 |
|
|
|
365 |
|
|
return dma_addr;
|
366 |
|
|
}
|
367 |
|
|
|
368 |
|
|
/*
|
369 |
|
|
* dma_addr is the kernel virtual address of the bounce buffer to unmap.
|
370 |
|
|
*/
|
371 |
|
|
static void
|
372 |
|
|
unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
|
373 |
|
|
{
|
374 |
|
|
unsigned long flags;
|
375 |
|
|
int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
|
376 |
|
|
int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
|
377 |
|
|
char *buffer = io_tlb_orig_addr[index];
|
378 |
|
|
|
379 |
|
|
/*
|
380 |
|
|
* First, sync the memory before unmapping the entry
|
381 |
|
|
*/
|
382 |
|
|
if (buffer && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
|
383 |
|
|
/*
|
384 |
|
|
* bounce... copy the data back into the original buffer * and
|
385 |
|
|
* delete the bounce buffer.
|
386 |
|
|
*/
|
387 |
|
|
memcpy(buffer, dma_addr, size);
|
388 |
|
|
|
389 |
|
|
/*
|
390 |
|
|
* Return the buffer to the free list by setting the corresponding
|
391 |
|
|
* entries to indicate the number of contigous entries available.
|
392 |
|
|
* While returning the entries to the free list, we merge the entries
|
393 |
|
|
* with slots below and above the pool being returned.
|
394 |
|
|
*/
|
395 |
|
|
spin_lock_irqsave(&io_tlb_lock, flags);
|
396 |
|
|
{
|
397 |
|
|
count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
|
398 |
|
|
io_tlb_list[index + nslots] : 0);
|
399 |
|
|
/*
|
400 |
|
|
* Step 1: return the slots to the free list, merging the
|
401 |
|
|
* slots with superceeding slots
|
402 |
|
|
*/
|
403 |
|
|
for (i = index + nslots - 1; i >= index; i--)
|
404 |
|
|
io_tlb_list[i] = ++count;
|
405 |
|
|
/*
|
406 |
|
|
* Step 2: merge the returned slots with the preceding slots,
|
407 |
|
|
* if available (non zero)
|
408 |
|
|
*/
|
409 |
|
|
for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
|
410 |
|
|
io_tlb_list[i] = ++count;
|
411 |
|
|
}
|
412 |
|
|
spin_unlock_irqrestore(&io_tlb_lock, flags);
|
413 |
|
|
}
|
414 |
|
|
|
415 |
|
|
static void
|
416 |
|
|
sync_single(struct device *hwdev, char *dma_addr, size_t size,
|
417 |
|
|
int dir, int target)
|
418 |
|
|
{
|
419 |
|
|
int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
|
420 |
|
|
char *buffer = io_tlb_orig_addr[index];
|
421 |
|
|
|
422 |
|
|
buffer += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
|
423 |
|
|
|
424 |
|
|
switch (target) {
|
425 |
|
|
case SYNC_FOR_CPU:
|
426 |
|
|
if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
|
427 |
|
|
memcpy(buffer, dma_addr, size);
|
428 |
|
|
else
|
429 |
|
|
BUG_ON(dir != DMA_TO_DEVICE);
|
430 |
|
|
break;
|
431 |
|
|
case SYNC_FOR_DEVICE:
|
432 |
|
|
if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
|
433 |
|
|
memcpy(dma_addr, buffer, size);
|
434 |
|
|
else
|
435 |
|
|
BUG_ON(dir != DMA_FROM_DEVICE);
|
436 |
|
|
break;
|
437 |
|
|
default:
|
438 |
|
|
BUG();
|
439 |
|
|
}
|
440 |
|
|
}
|
441 |
|
|
|
442 |
|
|
void *
|
443 |
|
|
swiotlb_alloc_coherent(struct device *hwdev, size_t size,
|
444 |
|
|
dma_addr_t *dma_handle, gfp_t flags)
|
445 |
|
|
{
|
446 |
|
|
dma_addr_t dev_addr;
|
447 |
|
|
void *ret;
|
448 |
|
|
int order = get_order(size);
|
449 |
|
|
|
450 |
|
|
/*
|
451 |
|
|
* XXX fix me: the DMA API should pass us an explicit DMA mask
|
452 |
|
|
* instead, or use ZONE_DMA32 (ia64 overloads ZONE_DMA to be a ~32
|
453 |
|
|
* bit range instead of a 16MB one).
|
454 |
|
|
*/
|
455 |
|
|
flags |= GFP_DMA;
|
456 |
|
|
|
457 |
|
|
ret = (void *)__get_free_pages(flags, order);
|
458 |
|
|
if (ret && address_needs_mapping(hwdev, virt_to_bus(ret))) {
|
459 |
|
|
/*
|
460 |
|
|
* The allocated memory isn't reachable by the device.
|
461 |
|
|
* Fall back on swiotlb_map_single().
|
462 |
|
|
*/
|
463 |
|
|
free_pages((unsigned long) ret, order);
|
464 |
|
|
ret = NULL;
|
465 |
|
|
}
|
466 |
|
|
if (!ret) {
|
467 |
|
|
/*
|
468 |
|
|
* We are either out of memory or the device can't DMA
|
469 |
|
|
* to GFP_DMA memory; fall back on
|
470 |
|
|
* swiotlb_map_single(), which will grab memory from
|
471 |
|
|
* the lowest available address range.
|
472 |
|
|
*/
|
473 |
|
|
dma_addr_t handle;
|
474 |
|
|
handle = swiotlb_map_single(NULL, NULL, size, DMA_FROM_DEVICE);
|
475 |
|
|
if (swiotlb_dma_mapping_error(handle))
|
476 |
|
|
return NULL;
|
477 |
|
|
|
478 |
|
|
ret = bus_to_virt(handle);
|
479 |
|
|
}
|
480 |
|
|
|
481 |
|
|
memset(ret, 0, size);
|
482 |
|
|
dev_addr = virt_to_bus(ret);
|
483 |
|
|
|
484 |
|
|
/* Confirm address can be DMA'd by device */
|
485 |
|
|
if (address_needs_mapping(hwdev, dev_addr)) {
|
486 |
|
|
printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
|
487 |
|
|
(unsigned long long)*hwdev->dma_mask,
|
488 |
|
|
(unsigned long long)dev_addr);
|
489 |
|
|
panic("swiotlb_alloc_coherent: allocated memory is out of "
|
490 |
|
|
"range for device");
|
491 |
|
|
}
|
492 |
|
|
*dma_handle = dev_addr;
|
493 |
|
|
return ret;
|
494 |
|
|
}
|
495 |
|
|
|
496 |
|
|
void
|
497 |
|
|
swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
|
498 |
|
|
dma_addr_t dma_handle)
|
499 |
|
|
{
|
500 |
|
|
WARN_ON(irqs_disabled());
|
501 |
|
|
if (!(vaddr >= (void *)io_tlb_start
|
502 |
|
|
&& vaddr < (void *)io_tlb_end))
|
503 |
|
|
free_pages((unsigned long) vaddr, get_order(size));
|
504 |
|
|
else
|
505 |
|
|
/* DMA_TO_DEVICE to avoid memcpy in unmap_single */
|
506 |
|
|
swiotlb_unmap_single (hwdev, dma_handle, size, DMA_TO_DEVICE);
|
507 |
|
|
}
|
508 |
|
|
|
509 |
|
|
static void
|
510 |
|
|
swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
|
511 |
|
|
{
|
512 |
|
|
/*
|
513 |
|
|
* Ran out of IOMMU space for this operation. This is very bad.
|
514 |
|
|
* Unfortunately the drivers cannot handle this operation properly.
|
515 |
|
|
* unless they check for dma_mapping_error (most don't)
|
516 |
|
|
* When the mapping is small enough return a static buffer to limit
|
517 |
|
|
* the damage, or panic when the transfer is too big.
|
518 |
|
|
*/
|
519 |
|
|
printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
|
520 |
|
|
"device %s\n", size, dev ? dev->bus_id : "?");
|
521 |
|
|
|
522 |
|
|
if (size > io_tlb_overflow && do_panic) {
|
523 |
|
|
if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
|
524 |
|
|
panic("DMA: Memory would be corrupted\n");
|
525 |
|
|
if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
|
526 |
|
|
panic("DMA: Random memory would be DMAed\n");
|
527 |
|
|
}
|
528 |
|
|
}
|
529 |
|
|
|
530 |
|
|
/*
|
531 |
|
|
* Map a single buffer of the indicated size for DMA in streaming mode. The
|
532 |
|
|
* physical address to use is returned.
|
533 |
|
|
*
|
534 |
|
|
* Once the device is given the dma address, the device owns this memory until
|
535 |
|
|
* either swiotlb_unmap_single or swiotlb_dma_sync_single is performed.
|
536 |
|
|
*/
|
537 |
|
|
dma_addr_t
|
538 |
|
|
swiotlb_map_single(struct device *hwdev, void *ptr, size_t size, int dir)
|
539 |
|
|
{
|
540 |
|
|
dma_addr_t dev_addr = virt_to_bus(ptr);
|
541 |
|
|
void *map;
|
542 |
|
|
|
543 |
|
|
BUG_ON(dir == DMA_NONE);
|
544 |
|
|
/*
|
545 |
|
|
* If the pointer passed in happens to be in the device's DMA window,
|
546 |
|
|
* we can safely return the device addr and not worry about bounce
|
547 |
|
|
* buffering it.
|
548 |
|
|
*/
|
549 |
|
|
if (!address_needs_mapping(hwdev, dev_addr) && !swiotlb_force)
|
550 |
|
|
return dev_addr;
|
551 |
|
|
|
552 |
|
|
/*
|
553 |
|
|
* Oh well, have to allocate and map a bounce buffer.
|
554 |
|
|
*/
|
555 |
|
|
map = map_single(hwdev, ptr, size, dir);
|
556 |
|
|
if (!map) {
|
557 |
|
|
swiotlb_full(hwdev, size, dir, 1);
|
558 |
|
|
map = io_tlb_overflow_buffer;
|
559 |
|
|
}
|
560 |
|
|
|
561 |
|
|
dev_addr = virt_to_bus(map);
|
562 |
|
|
|
563 |
|
|
/*
|
564 |
|
|
* Ensure that the address returned is DMA'ble
|
565 |
|
|
*/
|
566 |
|
|
if (address_needs_mapping(hwdev, dev_addr))
|
567 |
|
|
panic("map_single: bounce buffer is not DMA'ble");
|
568 |
|
|
|
569 |
|
|
return dev_addr;
|
570 |
|
|
}
|
571 |
|
|
|
572 |
|
|
/*
|
573 |
|
|
* Unmap a single streaming mode DMA translation. The dma_addr and size must
|
574 |
|
|
* match what was provided for in a previous swiotlb_map_single call. All
|
575 |
|
|
* other usages are undefined.
|
576 |
|
|
*
|
577 |
|
|
* After this call, reads by the cpu to the buffer are guaranteed to see
|
578 |
|
|
* whatever the device wrote there.
|
579 |
|
|
*/
|
580 |
|
|
void
|
581 |
|
|
swiotlb_unmap_single(struct device *hwdev, dma_addr_t dev_addr, size_t size,
|
582 |
|
|
int dir)
|
583 |
|
|
{
|
584 |
|
|
char *dma_addr = bus_to_virt(dev_addr);
|
585 |
|
|
|
586 |
|
|
BUG_ON(dir == DMA_NONE);
|
587 |
|
|
if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
|
588 |
|
|
unmap_single(hwdev, dma_addr, size, dir);
|
589 |
|
|
else if (dir == DMA_FROM_DEVICE)
|
590 |
|
|
dma_mark_clean(dma_addr, size);
|
591 |
|
|
}
|
592 |
|
|
|
593 |
|
|
/*
|
594 |
|
|
* Make physical memory consistent for a single streaming mode DMA translation
|
595 |
|
|
* after a transfer.
|
596 |
|
|
*
|
597 |
|
|
* If you perform a swiotlb_map_single() but wish to interrogate the buffer
|
598 |
|
|
* using the cpu, yet do not wish to teardown the dma mapping, you must
|
599 |
|
|
* call this function before doing so. At the next point you give the dma
|
600 |
|
|
* address back to the card, you must first perform a
|
601 |
|
|
* swiotlb_dma_sync_for_device, and then the device again owns the buffer
|
602 |
|
|
*/
|
603 |
|
|
static void
|
604 |
|
|
swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
|
605 |
|
|
size_t size, int dir, int target)
|
606 |
|
|
{
|
607 |
|
|
char *dma_addr = bus_to_virt(dev_addr);
|
608 |
|
|
|
609 |
|
|
BUG_ON(dir == DMA_NONE);
|
610 |
|
|
if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
|
611 |
|
|
sync_single(hwdev, dma_addr, size, dir, target);
|
612 |
|
|
else if (dir == DMA_FROM_DEVICE)
|
613 |
|
|
dma_mark_clean(dma_addr, size);
|
614 |
|
|
}
|
615 |
|
|
|
616 |
|
|
void
|
617 |
|
|
swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
|
618 |
|
|
size_t size, int dir)
|
619 |
|
|
{
|
620 |
|
|
swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
|
621 |
|
|
}
|
622 |
|
|
|
623 |
|
|
void
|
624 |
|
|
swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
|
625 |
|
|
size_t size, int dir)
|
626 |
|
|
{
|
627 |
|
|
swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
|
628 |
|
|
}
|
629 |
|
|
|
630 |
|
|
/*
|
631 |
|
|
* Same as above, but for a sub-range of the mapping.
|
632 |
|
|
*/
|
633 |
|
|
static void
|
634 |
|
|
swiotlb_sync_single_range(struct device *hwdev, dma_addr_t dev_addr,
|
635 |
|
|
unsigned long offset, size_t size,
|
636 |
|
|
int dir, int target)
|
637 |
|
|
{
|
638 |
|
|
char *dma_addr = bus_to_virt(dev_addr) + offset;
|
639 |
|
|
|
640 |
|
|
BUG_ON(dir == DMA_NONE);
|
641 |
|
|
if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
|
642 |
|
|
sync_single(hwdev, dma_addr, size, dir, target);
|
643 |
|
|
else if (dir == DMA_FROM_DEVICE)
|
644 |
|
|
dma_mark_clean(dma_addr, size);
|
645 |
|
|
}
|
646 |
|
|
|
647 |
|
|
void
|
648 |
|
|
swiotlb_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
|
649 |
|
|
unsigned long offset, size_t size, int dir)
|
650 |
|
|
{
|
651 |
|
|
swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
|
652 |
|
|
SYNC_FOR_CPU);
|
653 |
|
|
}
|
654 |
|
|
|
655 |
|
|
void
|
656 |
|
|
swiotlb_sync_single_range_for_device(struct device *hwdev, dma_addr_t dev_addr,
|
657 |
|
|
unsigned long offset, size_t size, int dir)
|
658 |
|
|
{
|
659 |
|
|
swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
|
660 |
|
|
SYNC_FOR_DEVICE);
|
661 |
|
|
}
|
662 |
|
|
|
663 |
|
|
/*
|
664 |
|
|
* Map a set of buffers described by scatterlist in streaming mode for DMA.
|
665 |
|
|
* This is the scatter-gather version of the above swiotlb_map_single
|
666 |
|
|
* interface. Here the scatter gather list elements are each tagged with the
|
667 |
|
|
* appropriate dma address and length. They are obtained via
|
668 |
|
|
* sg_dma_{address,length}(SG).
|
669 |
|
|
*
|
670 |
|
|
* NOTE: An implementation may be able to use a smaller number of
|
671 |
|
|
* DMA address/length pairs than there are SG table elements.
|
672 |
|
|
* (for example via virtual mapping capabilities)
|
673 |
|
|
* The routine returns the number of addr/length pairs actually
|
674 |
|
|
* used, at most nents.
|
675 |
|
|
*
|
676 |
|
|
* Device ownership issues as mentioned above for swiotlb_map_single are the
|
677 |
|
|
* same here.
|
678 |
|
|
*/
|
679 |
|
|
int
|
680 |
|
|
swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
|
681 |
|
|
int dir)
|
682 |
|
|
{
|
683 |
|
|
struct scatterlist *sg;
|
684 |
|
|
void *addr;
|
685 |
|
|
dma_addr_t dev_addr;
|
686 |
|
|
int i;
|
687 |
|
|
|
688 |
|
|
BUG_ON(dir == DMA_NONE);
|
689 |
|
|
|
690 |
|
|
for_each_sg(sgl, sg, nelems, i) {
|
691 |
|
|
addr = SG_ENT_VIRT_ADDRESS(sg);
|
692 |
|
|
dev_addr = virt_to_bus(addr);
|
693 |
|
|
if (swiotlb_force || address_needs_mapping(hwdev, dev_addr)) {
|
694 |
|
|
void *map = map_single(hwdev, addr, sg->length, dir);
|
695 |
|
|
if (!map) {
|
696 |
|
|
/* Don't panic here, we expect map_sg users
|
697 |
|
|
to do proper error handling. */
|
698 |
|
|
swiotlb_full(hwdev, sg->length, dir, 0);
|
699 |
|
|
swiotlb_unmap_sg(hwdev, sgl, i, dir);
|
700 |
|
|
sgl[0].dma_length = 0;
|
701 |
|
|
return 0;
|
702 |
|
|
}
|
703 |
|
|
sg->dma_address = virt_to_bus(map);
|
704 |
|
|
} else
|
705 |
|
|
sg->dma_address = dev_addr;
|
706 |
|
|
sg->dma_length = sg->length;
|
707 |
|
|
}
|
708 |
|
|
return nelems;
|
709 |
|
|
}
|
710 |
|
|
|
711 |
|
|
/*
|
712 |
|
|
* Unmap a set of streaming mode DMA translations. Again, cpu read rules
|
713 |
|
|
* concerning calls here are the same as for swiotlb_unmap_single() above.
|
714 |
|
|
*/
|
715 |
|
|
void
|
716 |
|
|
swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
|
717 |
|
|
int dir)
|
718 |
|
|
{
|
719 |
|
|
struct scatterlist *sg;
|
720 |
|
|
int i;
|
721 |
|
|
|
722 |
|
|
BUG_ON(dir == DMA_NONE);
|
723 |
|
|
|
724 |
|
|
for_each_sg(sgl, sg, nelems, i) {
|
725 |
|
|
if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
|
726 |
|
|
unmap_single(hwdev, bus_to_virt(sg->dma_address),
|
727 |
|
|
sg->dma_length, dir);
|
728 |
|
|
else if (dir == DMA_FROM_DEVICE)
|
729 |
|
|
dma_mark_clean(SG_ENT_VIRT_ADDRESS(sg), sg->dma_length);
|
730 |
|
|
}
|
731 |
|
|
}
|
732 |
|
|
|
733 |
|
|
/*
|
734 |
|
|
* Make physical memory consistent for a set of streaming mode DMA translations
|
735 |
|
|
* after a transfer.
|
736 |
|
|
*
|
737 |
|
|
* The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
|
738 |
|
|
* and usage.
|
739 |
|
|
*/
|
740 |
|
|
static void
|
741 |
|
|
swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
|
742 |
|
|
int nelems, int dir, int target)
|
743 |
|
|
{
|
744 |
|
|
struct scatterlist *sg;
|
745 |
|
|
int i;
|
746 |
|
|
|
747 |
|
|
BUG_ON(dir == DMA_NONE);
|
748 |
|
|
|
749 |
|
|
for_each_sg(sgl, sg, nelems, i) {
|
750 |
|
|
if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
|
751 |
|
|
sync_single(hwdev, bus_to_virt(sg->dma_address),
|
752 |
|
|
sg->dma_length, dir, target);
|
753 |
|
|
else if (dir == DMA_FROM_DEVICE)
|
754 |
|
|
dma_mark_clean(SG_ENT_VIRT_ADDRESS(sg), sg->dma_length);
|
755 |
|
|
}
|
756 |
|
|
}
|
757 |
|
|
|
758 |
|
|
void
|
759 |
|
|
swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
|
760 |
|
|
int nelems, int dir)
|
761 |
|
|
{
|
762 |
|
|
swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
|
763 |
|
|
}
|
764 |
|
|
|
765 |
|
|
void
|
766 |
|
|
swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
|
767 |
|
|
int nelems, int dir)
|
768 |
|
|
{
|
769 |
|
|
swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
|
770 |
|
|
}
|
771 |
|
|
|
772 |
|
|
int
|
773 |
|
|
swiotlb_dma_mapping_error(dma_addr_t dma_addr)
|
774 |
|
|
{
|
775 |
|
|
return (dma_addr == virt_to_bus(io_tlb_overflow_buffer));
|
776 |
|
|
}
|
777 |
|
|
|
778 |
|
|
/*
|
779 |
|
|
* Return whether the given device DMA address mask can be supported
|
780 |
|
|
* properly. For example, if your device can only drive the low 24-bits
|
781 |
|
|
* during bus mastering, then you would pass 0x00ffffff as the mask to
|
782 |
|
|
* this function.
|
783 |
|
|
*/
|
784 |
|
|
int
|
785 |
|
|
swiotlb_dma_supported(struct device *hwdev, u64 mask)
|
786 |
|
|
{
|
787 |
|
|
return virt_to_bus(io_tlb_end - 1) <= mask;
|
788 |
|
|
}
|
789 |
|
|
|
790 |
|
|
EXPORT_SYMBOL(swiotlb_map_single);
|
791 |
|
|
EXPORT_SYMBOL(swiotlb_unmap_single);
|
792 |
|
|
EXPORT_SYMBOL(swiotlb_map_sg);
|
793 |
|
|
EXPORT_SYMBOL(swiotlb_unmap_sg);
|
794 |
|
|
EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
|
795 |
|
|
EXPORT_SYMBOL(swiotlb_sync_single_for_device);
|
796 |
|
|
EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu);
|
797 |
|
|
EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device);
|
798 |
|
|
EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
|
799 |
|
|
EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
|
800 |
|
|
EXPORT_SYMBOL(swiotlb_dma_mapping_error);
|
801 |
|
|
EXPORT_SYMBOL(swiotlb_alloc_coherent);
|
802 |
|
|
EXPORT_SYMBOL(swiotlb_free_coherent);
|
803 |
|
|
EXPORT_SYMBOL(swiotlb_dma_supported);
|