]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Dynamic DMA mapping support. | |
3 | * | |
563aaf06 | 4 | * This implementation is a fallback for platforms that do not support |
1da177e4 LT |
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. | |
569c8bf5 JL |
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. | |
1da177e4 LT |
17 | */ |
18 | ||
19 | #include <linux/cache.h> | |
17e5ad6c | 20 | #include <linux/dma-mapping.h> |
1da177e4 LT |
21 | #include <linux/mm.h> |
22 | #include <linux/module.h> | |
1da177e4 | 23 | #include <linux/spinlock.h> |
8c5df16b | 24 | #include <linux/swiotlb.h> |
1da177e4 | 25 | #include <linux/string.h> |
0016fdee | 26 | #include <linux/swiotlb.h> |
1da177e4 LT |
27 | #include <linux/types.h> |
28 | #include <linux/ctype.h> | |
29 | ||
30 | #include <asm/io.h> | |
1da177e4 | 31 | #include <asm/dma.h> |
17e5ad6c | 32 | #include <asm/scatterlist.h> |
1da177e4 LT |
33 | |
34 | #include <linux/init.h> | |
35 | #include <linux/bootmem.h> | |
a8522509 | 36 | #include <linux/iommu-helper.h> |
1da177e4 LT |
37 | |
38 | #define OFFSET(val,align) ((unsigned long) \ | |
39 | ( (val) & ( (align) - 1))) | |
40 | ||
f9527f12 | 41 | #define SG_ENT_VIRT_ADDRESS(sg) (sg_virt((sg))) |
93fbff63 | 42 | #define SG_ENT_PHYS_ADDRESS(sg) virt_to_bus(SG_ENT_VIRT_ADDRESS(sg)) |
1da177e4 | 43 | |
0b9afede AW |
44 | #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT)) |
45 | ||
46 | /* | |
47 | * Minimum IO TLB size to bother booting with. Systems with mainly | |
48 | * 64bit capable cards will only lightly use the swiotlb. If we can't | |
49 | * allocate a contiguous 1MB, we're probably in trouble anyway. | |
50 | */ | |
51 | #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT) | |
52 | ||
de69e0f0 JL |
53 | /* |
54 | * Enumeration for sync targets | |
55 | */ | |
56 | enum dma_sync_target { | |
57 | SYNC_FOR_CPU = 0, | |
58 | SYNC_FOR_DEVICE = 1, | |
59 | }; | |
60 | ||
1da177e4 LT |
61 | int swiotlb_force; |
62 | ||
63 | /* | |
64 | * Used to do a quick range check in swiotlb_unmap_single and | |
65 | * swiotlb_sync_single_*, to see if the memory was in fact allocated by this | |
66 | * API. | |
67 | */ | |
68 | static char *io_tlb_start, *io_tlb_end; | |
69 | ||
70 | /* | |
71 | * The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and | |
72 | * io_tlb_end. This is command line adjustable via setup_io_tlb_npages. | |
73 | */ | |
74 | static unsigned long io_tlb_nslabs; | |
75 | ||
76 | /* | |
77 | * When the IOMMU overflows we return a fallback buffer. This sets the size. | |
78 | */ | |
79 | static unsigned long io_tlb_overflow = 32*1024; | |
80 | ||
81 | void *io_tlb_overflow_buffer; | |
82 | ||
83 | /* | |
84 | * This is a free list describing the number of free entries available from | |
85 | * each index | |
86 | */ | |
87 | static unsigned int *io_tlb_list; | |
88 | static unsigned int io_tlb_index; | |
89 | ||
90 | /* | |
91 | * We need to save away the original address corresponding to a mapped entry | |
92 | * for the sync operations. | |
93 | */ | |
25667d67 | 94 | static unsigned char **io_tlb_orig_addr; |
1da177e4 LT |
95 | |
96 | /* | |
97 | * Protect the above data structures in the map and unmap calls | |
98 | */ | |
99 | static DEFINE_SPINLOCK(io_tlb_lock); | |
100 | ||
101 | static int __init | |
102 | setup_io_tlb_npages(char *str) | |
103 | { | |
104 | if (isdigit(*str)) { | |
e8579e72 | 105 | io_tlb_nslabs = simple_strtoul(str, &str, 0); |
1da177e4 LT |
106 | /* avoid tail segment of size < IO_TLB_SEGSIZE */ |
107 | io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE); | |
108 | } | |
109 | if (*str == ',') | |
110 | ++str; | |
111 | if (!strcmp(str, "force")) | |
112 | swiotlb_force = 1; | |
113 | return 1; | |
114 | } | |
115 | __setup("swiotlb=", setup_io_tlb_npages); | |
116 | /* make io_tlb_overflow tunable too? */ | |
117 | ||
8c5df16b JF |
118 | void * __weak swiotlb_alloc_boot(size_t size, unsigned long nslabs) |
119 | { | |
120 | return alloc_bootmem_low_pages(size); | |
121 | } | |
122 | ||
123 | void * __weak swiotlb_alloc(unsigned order, unsigned long nslabs) | |
124 | { | |
125 | return (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN, order); | |
126 | } | |
127 | ||
e08e1f7a IC |
128 | dma_addr_t __weak swiotlb_phys_to_bus(phys_addr_t paddr) |
129 | { | |
130 | return paddr; | |
131 | } | |
132 | ||
133 | phys_addr_t __weak swiotlb_bus_to_phys(dma_addr_t baddr) | |
134 | { | |
135 | return baddr; | |
136 | } | |
137 | ||
138 | static dma_addr_t swiotlb_virt_to_bus(volatile void *address) | |
139 | { | |
140 | return swiotlb_phys_to_bus(virt_to_phys(address)); | |
141 | } | |
142 | ||
143 | static void *swiotlb_bus_to_virt(dma_addr_t address) | |
144 | { | |
145 | return phys_to_virt(swiotlb_bus_to_phys(address)); | |
146 | } | |
147 | ||
b81ea27b IC |
148 | int __weak swiotlb_arch_range_needs_mapping(void *ptr, size_t size) |
149 | { | |
150 | return 0; | |
151 | } | |
152 | ||
1da177e4 LT |
153 | /* |
154 | * Statically reserve bounce buffer space and initialize bounce buffer data | |
17e5ad6c | 155 | * structures for the software IO TLB used to implement the DMA API. |
1da177e4 | 156 | */ |
563aaf06 JB |
157 | void __init |
158 | swiotlb_init_with_default_size(size_t default_size) | |
1da177e4 | 159 | { |
563aaf06 | 160 | unsigned long i, bytes; |
1da177e4 LT |
161 | |
162 | if (!io_tlb_nslabs) { | |
e8579e72 | 163 | io_tlb_nslabs = (default_size >> IO_TLB_SHIFT); |
1da177e4 LT |
164 | io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE); |
165 | } | |
166 | ||
563aaf06 JB |
167 | bytes = io_tlb_nslabs << IO_TLB_SHIFT; |
168 | ||
1da177e4 LT |
169 | /* |
170 | * Get IO TLB memory from the low pages | |
171 | */ | |
8c5df16b | 172 | io_tlb_start = swiotlb_alloc_boot(bytes, io_tlb_nslabs); |
1da177e4 LT |
173 | if (!io_tlb_start) |
174 | panic("Cannot allocate SWIOTLB buffer"); | |
563aaf06 | 175 | io_tlb_end = io_tlb_start + bytes; |
1da177e4 LT |
176 | |
177 | /* | |
178 | * Allocate and initialize the free list array. This array is used | |
179 | * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE | |
180 | * between io_tlb_start and io_tlb_end. | |
181 | */ | |
182 | io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int)); | |
25667d67 | 183 | for (i = 0; i < io_tlb_nslabs; i++) |
1da177e4 LT |
184 | io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE); |
185 | io_tlb_index = 0; | |
25667d67 | 186 | io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(char *)); |
1da177e4 LT |
187 | |
188 | /* | |
189 | * Get the overflow emergency buffer | |
190 | */ | |
191 | io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow); | |
563aaf06 JB |
192 | if (!io_tlb_overflow_buffer) |
193 | panic("Cannot allocate SWIOTLB overflow buffer!\n"); | |
194 | ||
25667d67 | 195 | printk(KERN_INFO "Placing software IO TLB between 0x%lx - 0x%lx\n", |
e08e1f7a | 196 | swiotlb_virt_to_bus(io_tlb_start), swiotlb_virt_to_bus(io_tlb_end)); |
1da177e4 LT |
197 | } |
198 | ||
563aaf06 JB |
199 | void __init |
200 | swiotlb_init(void) | |
1da177e4 | 201 | { |
25667d67 | 202 | swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */ |
1da177e4 LT |
203 | } |
204 | ||
0b9afede AW |
205 | /* |
206 | * Systems with larger DMA zones (those that don't support ISA) can | |
207 | * initialize the swiotlb later using the slab allocator if needed. | |
208 | * This should be just like above, but with some error catching. | |
209 | */ | |
210 | int | |
563aaf06 | 211 | swiotlb_late_init_with_default_size(size_t default_size) |
0b9afede | 212 | { |
563aaf06 | 213 | unsigned long i, bytes, req_nslabs = io_tlb_nslabs; |
0b9afede AW |
214 | unsigned int order; |
215 | ||
216 | if (!io_tlb_nslabs) { | |
217 | io_tlb_nslabs = (default_size >> IO_TLB_SHIFT); | |
218 | io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE); | |
219 | } | |
220 | ||
221 | /* | |
222 | * Get IO TLB memory from the low pages | |
223 | */ | |
563aaf06 | 224 | order = get_order(io_tlb_nslabs << IO_TLB_SHIFT); |
0b9afede | 225 | io_tlb_nslabs = SLABS_PER_PAGE << order; |
563aaf06 | 226 | bytes = io_tlb_nslabs << IO_TLB_SHIFT; |
0b9afede AW |
227 | |
228 | while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) { | |
8c5df16b | 229 | io_tlb_start = swiotlb_alloc(order, io_tlb_nslabs); |
0b9afede AW |
230 | if (io_tlb_start) |
231 | break; | |
232 | order--; | |
233 | } | |
234 | ||
235 | if (!io_tlb_start) | |
236 | goto cleanup1; | |
237 | ||
563aaf06 | 238 | if (order != get_order(bytes)) { |
0b9afede AW |
239 | printk(KERN_WARNING "Warning: only able to allocate %ld MB " |
240 | "for software IO TLB\n", (PAGE_SIZE << order) >> 20); | |
241 | io_tlb_nslabs = SLABS_PER_PAGE << order; | |
563aaf06 | 242 | bytes = io_tlb_nslabs << IO_TLB_SHIFT; |
0b9afede | 243 | } |
563aaf06 JB |
244 | io_tlb_end = io_tlb_start + bytes; |
245 | memset(io_tlb_start, 0, bytes); | |
0b9afede AW |
246 | |
247 | /* | |
248 | * Allocate and initialize the free list array. This array is used | |
249 | * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE | |
250 | * between io_tlb_start and io_tlb_end. | |
251 | */ | |
252 | io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL, | |
253 | get_order(io_tlb_nslabs * sizeof(int))); | |
254 | if (!io_tlb_list) | |
255 | goto cleanup2; | |
256 | ||
257 | for (i = 0; i < io_tlb_nslabs; i++) | |
258 | io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE); | |
259 | io_tlb_index = 0; | |
260 | ||
25667d67 TL |
261 | io_tlb_orig_addr = (unsigned char **)__get_free_pages(GFP_KERNEL, |
262 | get_order(io_tlb_nslabs * sizeof(char *))); | |
0b9afede AW |
263 | if (!io_tlb_orig_addr) |
264 | goto cleanup3; | |
265 | ||
25667d67 | 266 | memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(char *)); |
0b9afede AW |
267 | |
268 | /* | |
269 | * Get the overflow emergency buffer | |
270 | */ | |
271 | io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA, | |
272 | get_order(io_tlb_overflow)); | |
273 | if (!io_tlb_overflow_buffer) | |
274 | goto cleanup4; | |
275 | ||
25667d67 TL |
276 | printk(KERN_INFO "Placing %luMB software IO TLB between 0x%lx - " |
277 | "0x%lx\n", bytes >> 20, | |
e08e1f7a | 278 | swiotlb_virt_to_bus(io_tlb_start), swiotlb_virt_to_bus(io_tlb_end)); |
0b9afede AW |
279 | |
280 | return 0; | |
281 | ||
282 | cleanup4: | |
25667d67 TL |
283 | free_pages((unsigned long)io_tlb_orig_addr, get_order(io_tlb_nslabs * |
284 | sizeof(char *))); | |
0b9afede AW |
285 | io_tlb_orig_addr = NULL; |
286 | cleanup3: | |
25667d67 TL |
287 | free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs * |
288 | sizeof(int))); | |
0b9afede | 289 | io_tlb_list = NULL; |
0b9afede | 290 | cleanup2: |
563aaf06 | 291 | io_tlb_end = NULL; |
0b9afede AW |
292 | free_pages((unsigned long)io_tlb_start, order); |
293 | io_tlb_start = NULL; | |
294 | cleanup1: | |
295 | io_tlb_nslabs = req_nslabs; | |
296 | return -ENOMEM; | |
297 | } | |
298 | ||
be6b0267 | 299 | static int |
2797982e | 300 | address_needs_mapping(struct device *hwdev, dma_addr_t addr, size_t size) |
1da177e4 | 301 | { |
07a2c01a | 302 | return !is_buffer_dma_capable(dma_get_mask(hwdev), addr, size); |
1da177e4 LT |
303 | } |
304 | ||
b81ea27b IC |
305 | static inline int range_needs_mapping(void *ptr, size_t size) |
306 | { | |
307 | return swiotlb_force || swiotlb_arch_range_needs_mapping(ptr, size); | |
308 | } | |
309 | ||
640aebfe FT |
310 | static int is_swiotlb_buffer(char *addr) |
311 | { | |
312 | return addr >= io_tlb_start && addr < io_tlb_end; | |
313 | } | |
314 | ||
1da177e4 LT |
315 | /* |
316 | * Allocates bounce buffer and returns its kernel virtual address. | |
317 | */ | |
318 | static void * | |
25667d67 | 319 | map_single(struct device *hwdev, char *buffer, size_t size, int dir) |
1da177e4 LT |
320 | { |
321 | unsigned long flags; | |
322 | char *dma_addr; | |
323 | unsigned int nslots, stride, index, wrap; | |
324 | int i; | |
681cc5cd FT |
325 | unsigned long start_dma_addr; |
326 | unsigned long mask; | |
327 | unsigned long offset_slots; | |
328 | unsigned long max_slots; | |
329 | ||
330 | mask = dma_get_seg_boundary(hwdev); | |
e08e1f7a | 331 | start_dma_addr = swiotlb_virt_to_bus(io_tlb_start) & mask; |
681cc5cd FT |
332 | |
333 | offset_slots = ALIGN(start_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT; | |
a5ddde4a IC |
334 | |
335 | /* | |
336 | * Carefully handle integer overflow which can occur when mask == ~0UL. | |
337 | */ | |
b15a3891 JB |
338 | max_slots = mask + 1 |
339 | ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT | |
340 | : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT); | |
1da177e4 LT |
341 | |
342 | /* | |
343 | * For mappings greater than a page, we limit the stride (and | |
344 | * hence alignment) to a page size. | |
345 | */ | |
346 | nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT; | |
347 | if (size > PAGE_SIZE) | |
348 | stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT)); | |
349 | else | |
350 | stride = 1; | |
351 | ||
34814545 | 352 | BUG_ON(!nslots); |
1da177e4 LT |
353 | |
354 | /* | |
355 | * Find suitable number of IO TLB entries size that will fit this | |
356 | * request and allocate a buffer from that IO TLB pool. | |
357 | */ | |
358 | spin_lock_irqsave(&io_tlb_lock, flags); | |
a7133a15 AM |
359 | index = ALIGN(io_tlb_index, stride); |
360 | if (index >= io_tlb_nslabs) | |
361 | index = 0; | |
362 | wrap = index; | |
363 | ||
364 | do { | |
a8522509 FT |
365 | while (iommu_is_span_boundary(index, nslots, offset_slots, |
366 | max_slots)) { | |
b15a3891 JB |
367 | index += stride; |
368 | if (index >= io_tlb_nslabs) | |
369 | index = 0; | |
a7133a15 AM |
370 | if (index == wrap) |
371 | goto not_found; | |
372 | } | |
373 | ||
374 | /* | |
375 | * If we find a slot that indicates we have 'nslots' number of | |
376 | * contiguous buffers, we allocate the buffers from that slot | |
377 | * and mark the entries as '0' indicating unavailable. | |
378 | */ | |
379 | if (io_tlb_list[index] >= nslots) { | |
380 | int count = 0; | |
381 | ||
382 | for (i = index; i < (int) (index + nslots); i++) | |
383 | io_tlb_list[i] = 0; | |
384 | for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--) | |
385 | io_tlb_list[i] = ++count; | |
386 | dma_addr = io_tlb_start + (index << IO_TLB_SHIFT); | |
1da177e4 | 387 | |
a7133a15 AM |
388 | /* |
389 | * Update the indices to avoid searching in the next | |
390 | * round. | |
391 | */ | |
392 | io_tlb_index = ((index + nslots) < io_tlb_nslabs | |
393 | ? (index + nslots) : 0); | |
394 | ||
395 | goto found; | |
396 | } | |
397 | index += stride; | |
398 | if (index >= io_tlb_nslabs) | |
399 | index = 0; | |
400 | } while (index != wrap); | |
401 | ||
402 | not_found: | |
403 | spin_unlock_irqrestore(&io_tlb_lock, flags); | |
404 | return NULL; | |
405 | found: | |
1da177e4 LT |
406 | spin_unlock_irqrestore(&io_tlb_lock, flags); |
407 | ||
408 | /* | |
409 | * Save away the mapping from the original address to the DMA address. | |
410 | * This is needed when we sync the memory. Then we sync the buffer if | |
411 | * needed. | |
412 | */ | |
df336d1c KF |
413 | for (i = 0; i < nslots; i++) |
414 | io_tlb_orig_addr[index+i] = buffer + (i << IO_TLB_SHIFT); | |
1da177e4 | 415 | if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL) |
25667d67 | 416 | memcpy(dma_addr, buffer, size); |
1da177e4 LT |
417 | |
418 | return dma_addr; | |
419 | } | |
420 | ||
421 | /* | |
422 | * dma_addr is the kernel virtual address of the bounce buffer to unmap. | |
423 | */ | |
424 | static void | |
425 | unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir) | |
426 | { | |
427 | unsigned long flags; | |
428 | int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT; | |
429 | int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT; | |
25667d67 | 430 | char *buffer = io_tlb_orig_addr[index]; |
1da177e4 LT |
431 | |
432 | /* | |
433 | * First, sync the memory before unmapping the entry | |
434 | */ | |
25667d67 | 435 | if (buffer && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL))) |
1da177e4 LT |
436 | /* |
437 | * bounce... copy the data back into the original buffer * and | |
438 | * delete the bounce buffer. | |
439 | */ | |
25667d67 | 440 | memcpy(buffer, dma_addr, size); |
1da177e4 LT |
441 | |
442 | /* | |
443 | * Return the buffer to the free list by setting the corresponding | |
444 | * entries to indicate the number of contigous entries available. | |
445 | * While returning the entries to the free list, we merge the entries | |
446 | * with slots below and above the pool being returned. | |
447 | */ | |
448 | spin_lock_irqsave(&io_tlb_lock, flags); | |
449 | { | |
450 | count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ? | |
451 | io_tlb_list[index + nslots] : 0); | |
452 | /* | |
453 | * Step 1: return the slots to the free list, merging the | |
454 | * slots with superceeding slots | |
455 | */ | |
456 | for (i = index + nslots - 1; i >= index; i--) | |
457 | io_tlb_list[i] = ++count; | |
458 | /* | |
459 | * Step 2: merge the returned slots with the preceding slots, | |
460 | * if available (non zero) | |
461 | */ | |
462 | for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--) | |
463 | io_tlb_list[i] = ++count; | |
464 | } | |
465 | spin_unlock_irqrestore(&io_tlb_lock, flags); | |
466 | } | |
467 | ||
468 | static void | |
de69e0f0 JL |
469 | sync_single(struct device *hwdev, char *dma_addr, size_t size, |
470 | int dir, int target) | |
1da177e4 LT |
471 | { |
472 | int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT; | |
25667d67 | 473 | char *buffer = io_tlb_orig_addr[index]; |
1da177e4 | 474 | |
df336d1c KF |
475 | buffer += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1)); |
476 | ||
de69e0f0 JL |
477 | switch (target) { |
478 | case SYNC_FOR_CPU: | |
479 | if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)) | |
25667d67 | 480 | memcpy(buffer, dma_addr, size); |
34814545 ES |
481 | else |
482 | BUG_ON(dir != DMA_TO_DEVICE); | |
de69e0f0 JL |
483 | break; |
484 | case SYNC_FOR_DEVICE: | |
485 | if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)) | |
25667d67 | 486 | memcpy(dma_addr, buffer, size); |
34814545 ES |
487 | else |
488 | BUG_ON(dir != DMA_FROM_DEVICE); | |
de69e0f0 JL |
489 | break; |
490 | default: | |
1da177e4 | 491 | BUG(); |
de69e0f0 | 492 | } |
1da177e4 LT |
493 | } |
494 | ||
495 | void * | |
496 | swiotlb_alloc_coherent(struct device *hwdev, size_t size, | |
06a54497 | 497 | dma_addr_t *dma_handle, gfp_t flags) |
1da177e4 | 498 | { |
563aaf06 | 499 | dma_addr_t dev_addr; |
1da177e4 LT |
500 | void *ret; |
501 | int order = get_order(size); | |
1e74f300 FT |
502 | u64 dma_mask = DMA_32BIT_MASK; |
503 | ||
504 | if (hwdev && hwdev->coherent_dma_mask) | |
505 | dma_mask = hwdev->coherent_dma_mask; | |
1da177e4 | 506 | |
25667d67 | 507 | ret = (void *)__get_free_pages(flags, order); |
e08e1f7a | 508 | if (ret && !is_buffer_dma_capable(dma_mask, swiotlb_virt_to_bus(ret), size)) { |
1da177e4 LT |
509 | /* |
510 | * The allocated memory isn't reachable by the device. | |
511 | * Fall back on swiotlb_map_single(). | |
512 | */ | |
513 | free_pages((unsigned long) ret, order); | |
514 | ret = NULL; | |
515 | } | |
516 | if (!ret) { | |
517 | /* | |
518 | * We are either out of memory or the device can't DMA | |
519 | * to GFP_DMA memory; fall back on | |
520 | * swiotlb_map_single(), which will grab memory from | |
521 | * the lowest available address range. | |
522 | */ | |
9dfda12b FT |
523 | ret = map_single(hwdev, NULL, size, DMA_FROM_DEVICE); |
524 | if (!ret) | |
1da177e4 | 525 | return NULL; |
1da177e4 LT |
526 | } |
527 | ||
528 | memset(ret, 0, size); | |
e08e1f7a | 529 | dev_addr = swiotlb_virt_to_bus(ret); |
1da177e4 LT |
530 | |
531 | /* Confirm address can be DMA'd by device */ | |
1e74f300 | 532 | if (!is_buffer_dma_capable(dma_mask, dev_addr, size)) { |
563aaf06 | 533 | printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n", |
1e74f300 | 534 | (unsigned long long)dma_mask, |
563aaf06 | 535 | (unsigned long long)dev_addr); |
a2b89b59 FT |
536 | |
537 | /* DMA_TO_DEVICE to avoid memcpy in unmap_single */ | |
538 | unmap_single(hwdev, ret, size, DMA_TO_DEVICE); | |
539 | return NULL; | |
1da177e4 LT |
540 | } |
541 | *dma_handle = dev_addr; | |
542 | return ret; | |
543 | } | |
544 | ||
545 | void | |
546 | swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr, | |
547 | dma_addr_t dma_handle) | |
548 | { | |
aa24886e | 549 | WARN_ON(irqs_disabled()); |
640aebfe | 550 | if (!is_swiotlb_buffer(vaddr)) |
1da177e4 LT |
551 | free_pages((unsigned long) vaddr, get_order(size)); |
552 | else | |
553 | /* DMA_TO_DEVICE to avoid memcpy in unmap_single */ | |
21f6c4de | 554 | unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE); |
1da177e4 LT |
555 | } |
556 | ||
557 | static void | |
558 | swiotlb_full(struct device *dev, size_t size, int dir, int do_panic) | |
559 | { | |
560 | /* | |
561 | * Ran out of IOMMU space for this operation. This is very bad. | |
562 | * Unfortunately the drivers cannot handle this operation properly. | |
17e5ad6c | 563 | * unless they check for dma_mapping_error (most don't) |
1da177e4 LT |
564 | * When the mapping is small enough return a static buffer to limit |
565 | * the damage, or panic when the transfer is too big. | |
566 | */ | |
563aaf06 | 567 | printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at " |
1da177e4 LT |
568 | "device %s\n", size, dev ? dev->bus_id : "?"); |
569 | ||
570 | if (size > io_tlb_overflow && do_panic) { | |
17e5ad6c TL |
571 | if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL) |
572 | panic("DMA: Memory would be corrupted\n"); | |
573 | if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL) | |
574 | panic("DMA: Random memory would be DMAed\n"); | |
1da177e4 LT |
575 | } |
576 | } | |
577 | ||
578 | /* | |
579 | * Map a single buffer of the indicated size for DMA in streaming mode. The | |
17e5ad6c | 580 | * physical address to use is returned. |
1da177e4 LT |
581 | * |
582 | * Once the device is given the dma address, the device owns this memory until | |
583 | * either swiotlb_unmap_single or swiotlb_dma_sync_single is performed. | |
584 | */ | |
585 | dma_addr_t | |
309df0c5 AK |
586 | swiotlb_map_single_attrs(struct device *hwdev, void *ptr, size_t size, |
587 | int dir, struct dma_attrs *attrs) | |
1da177e4 | 588 | { |
e08e1f7a | 589 | dma_addr_t dev_addr = swiotlb_virt_to_bus(ptr); |
1da177e4 LT |
590 | void *map; |
591 | ||
34814545 | 592 | BUG_ON(dir == DMA_NONE); |
1da177e4 LT |
593 | /* |
594 | * If the pointer passed in happens to be in the device's DMA window, | |
595 | * we can safely return the device addr and not worry about bounce | |
596 | * buffering it. | |
597 | */ | |
b81ea27b IC |
598 | if (!address_needs_mapping(hwdev, dev_addr, size) && |
599 | !range_needs_mapping(ptr, size)) | |
1da177e4 LT |
600 | return dev_addr; |
601 | ||
602 | /* | |
603 | * Oh well, have to allocate and map a bounce buffer. | |
604 | */ | |
25667d67 | 605 | map = map_single(hwdev, ptr, size, dir); |
1da177e4 LT |
606 | if (!map) { |
607 | swiotlb_full(hwdev, size, dir, 1); | |
608 | map = io_tlb_overflow_buffer; | |
609 | } | |
610 | ||
e08e1f7a | 611 | dev_addr = swiotlb_virt_to_bus(map); |
1da177e4 LT |
612 | |
613 | /* | |
614 | * Ensure that the address returned is DMA'ble | |
615 | */ | |
2797982e | 616 | if (address_needs_mapping(hwdev, dev_addr, size)) |
1da177e4 LT |
617 | panic("map_single: bounce buffer is not DMA'ble"); |
618 | ||
619 | return dev_addr; | |
620 | } | |
309df0c5 AK |
621 | EXPORT_SYMBOL(swiotlb_map_single_attrs); |
622 | ||
623 | dma_addr_t | |
624 | swiotlb_map_single(struct device *hwdev, void *ptr, size_t size, int dir) | |
625 | { | |
626 | return swiotlb_map_single_attrs(hwdev, ptr, size, dir, NULL); | |
627 | } | |
1da177e4 | 628 | |
1da177e4 LT |
629 | /* |
630 | * Unmap a single streaming mode DMA translation. The dma_addr and size must | |
631 | * match what was provided for in a previous swiotlb_map_single call. All | |
632 | * other usages are undefined. | |
633 | * | |
634 | * After this call, reads by the cpu to the buffer are guaranteed to see | |
635 | * whatever the device wrote there. | |
636 | */ | |
637 | void | |
309df0c5 AK |
638 | swiotlb_unmap_single_attrs(struct device *hwdev, dma_addr_t dev_addr, |
639 | size_t size, int dir, struct dma_attrs *attrs) | |
1da177e4 | 640 | { |
e08e1f7a | 641 | char *dma_addr = swiotlb_bus_to_virt(dev_addr); |
1da177e4 | 642 | |
34814545 | 643 | BUG_ON(dir == DMA_NONE); |
640aebfe | 644 | if (is_swiotlb_buffer(dma_addr)) |
1da177e4 LT |
645 | unmap_single(hwdev, dma_addr, size, dir); |
646 | else if (dir == DMA_FROM_DEVICE) | |
cde14bbf | 647 | dma_mark_clean(dma_addr, size); |
1da177e4 | 648 | } |
309df0c5 | 649 | EXPORT_SYMBOL(swiotlb_unmap_single_attrs); |
1da177e4 | 650 | |
309df0c5 AK |
651 | void |
652 | swiotlb_unmap_single(struct device *hwdev, dma_addr_t dev_addr, size_t size, | |
653 | int dir) | |
654 | { | |
655 | return swiotlb_unmap_single_attrs(hwdev, dev_addr, size, dir, NULL); | |
656 | } | |
1da177e4 LT |
657 | /* |
658 | * Make physical memory consistent for a single streaming mode DMA translation | |
659 | * after a transfer. | |
660 | * | |
661 | * If you perform a swiotlb_map_single() but wish to interrogate the buffer | |
17e5ad6c TL |
662 | * using the cpu, yet do not wish to teardown the dma mapping, you must |
663 | * call this function before doing so. At the next point you give the dma | |
1da177e4 LT |
664 | * address back to the card, you must first perform a |
665 | * swiotlb_dma_sync_for_device, and then the device again owns the buffer | |
666 | */ | |
be6b0267 | 667 | static void |
8270f3f1 | 668 | swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr, |
de69e0f0 | 669 | size_t size, int dir, int target) |
1da177e4 | 670 | { |
e08e1f7a | 671 | char *dma_addr = swiotlb_bus_to_virt(dev_addr); |
1da177e4 | 672 | |
34814545 | 673 | BUG_ON(dir == DMA_NONE); |
640aebfe | 674 | if (is_swiotlb_buffer(dma_addr)) |
de69e0f0 | 675 | sync_single(hwdev, dma_addr, size, dir, target); |
1da177e4 | 676 | else if (dir == DMA_FROM_DEVICE) |
cde14bbf | 677 | dma_mark_clean(dma_addr, size); |
1da177e4 LT |
678 | } |
679 | ||
8270f3f1 JL |
680 | void |
681 | swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr, | |
682 | size_t size, int dir) | |
683 | { | |
de69e0f0 | 684 | swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU); |
8270f3f1 JL |
685 | } |
686 | ||
1da177e4 LT |
687 | void |
688 | swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr, | |
689 | size_t size, int dir) | |
690 | { | |
de69e0f0 | 691 | swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE); |
1da177e4 LT |
692 | } |
693 | ||
878a97cf JL |
694 | /* |
695 | * Same as above, but for a sub-range of the mapping. | |
696 | */ | |
be6b0267 | 697 | static void |
878a97cf | 698 | swiotlb_sync_single_range(struct device *hwdev, dma_addr_t dev_addr, |
de69e0f0 JL |
699 | unsigned long offset, size_t size, |
700 | int dir, int target) | |
878a97cf | 701 | { |
e08e1f7a | 702 | char *dma_addr = swiotlb_bus_to_virt(dev_addr) + offset; |
878a97cf | 703 | |
34814545 | 704 | BUG_ON(dir == DMA_NONE); |
640aebfe | 705 | if (is_swiotlb_buffer(dma_addr)) |
de69e0f0 | 706 | sync_single(hwdev, dma_addr, size, dir, target); |
878a97cf | 707 | else if (dir == DMA_FROM_DEVICE) |
cde14bbf | 708 | dma_mark_clean(dma_addr, size); |
878a97cf JL |
709 | } |
710 | ||
711 | void | |
712 | swiotlb_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dev_addr, | |
713 | unsigned long offset, size_t size, int dir) | |
714 | { | |
de69e0f0 JL |
715 | swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir, |
716 | SYNC_FOR_CPU); | |
878a97cf JL |
717 | } |
718 | ||
719 | void | |
720 | swiotlb_sync_single_range_for_device(struct device *hwdev, dma_addr_t dev_addr, | |
721 | unsigned long offset, size_t size, int dir) | |
722 | { | |
de69e0f0 JL |
723 | swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir, |
724 | SYNC_FOR_DEVICE); | |
878a97cf JL |
725 | } |
726 | ||
309df0c5 AK |
727 | void swiotlb_unmap_sg_attrs(struct device *, struct scatterlist *, int, int, |
728 | struct dma_attrs *); | |
1da177e4 LT |
729 | /* |
730 | * Map a set of buffers described by scatterlist in streaming mode for DMA. | |
731 | * This is the scatter-gather version of the above swiotlb_map_single | |
732 | * interface. Here the scatter gather list elements are each tagged with the | |
733 | * appropriate dma address and length. They are obtained via | |
734 | * sg_dma_{address,length}(SG). | |
735 | * | |
736 | * NOTE: An implementation may be able to use a smaller number of | |
737 | * DMA address/length pairs than there are SG table elements. | |
738 | * (for example via virtual mapping capabilities) | |
739 | * The routine returns the number of addr/length pairs actually | |
740 | * used, at most nents. | |
741 | * | |
742 | * Device ownership issues as mentioned above for swiotlb_map_single are the | |
743 | * same here. | |
744 | */ | |
745 | int | |
309df0c5 AK |
746 | swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems, |
747 | int dir, struct dma_attrs *attrs) | |
1da177e4 | 748 | { |
dbfd49fe | 749 | struct scatterlist *sg; |
25667d67 | 750 | void *addr; |
563aaf06 | 751 | dma_addr_t dev_addr; |
1da177e4 LT |
752 | int i; |
753 | ||
34814545 | 754 | BUG_ON(dir == DMA_NONE); |
1da177e4 | 755 | |
dbfd49fe | 756 | for_each_sg(sgl, sg, nelems, i) { |
25667d67 | 757 | addr = SG_ENT_VIRT_ADDRESS(sg); |
e08e1f7a | 758 | dev_addr = swiotlb_virt_to_bus(addr); |
b81ea27b | 759 | if (range_needs_mapping(sg_virt(sg), sg->length) || |
2797982e | 760 | address_needs_mapping(hwdev, dev_addr, sg->length)) { |
25667d67 | 761 | void *map = map_single(hwdev, addr, sg->length, dir); |
7e870233 | 762 | if (!map) { |
1da177e4 LT |
763 | /* Don't panic here, we expect map_sg users |
764 | to do proper error handling. */ | |
765 | swiotlb_full(hwdev, sg->length, dir, 0); | |
309df0c5 AK |
766 | swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir, |
767 | attrs); | |
dbfd49fe | 768 | sgl[0].dma_length = 0; |
1da177e4 LT |
769 | return 0; |
770 | } | |
e08e1f7a | 771 | sg->dma_address = swiotlb_virt_to_bus(map); |
1da177e4 LT |
772 | } else |
773 | sg->dma_address = dev_addr; | |
774 | sg->dma_length = sg->length; | |
775 | } | |
776 | return nelems; | |
777 | } | |
309df0c5 AK |
778 | EXPORT_SYMBOL(swiotlb_map_sg_attrs); |
779 | ||
780 | int | |
781 | swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems, | |
782 | int dir) | |
783 | { | |
784 | return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL); | |
785 | } | |
1da177e4 LT |
786 | |
787 | /* | |
788 | * Unmap a set of streaming mode DMA translations. Again, cpu read rules | |
789 | * concerning calls here are the same as for swiotlb_unmap_single() above. | |
790 | */ | |
791 | void | |
309df0c5 AK |
792 | swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl, |
793 | int nelems, int dir, struct dma_attrs *attrs) | |
1da177e4 | 794 | { |
dbfd49fe | 795 | struct scatterlist *sg; |
1da177e4 LT |
796 | int i; |
797 | ||
34814545 | 798 | BUG_ON(dir == DMA_NONE); |
1da177e4 | 799 | |
dbfd49fe | 800 | for_each_sg(sgl, sg, nelems, i) { |
1da177e4 | 801 | if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg)) |
e08e1f7a | 802 | unmap_single(hwdev, swiotlb_bus_to_virt(sg->dma_address), |
93fbff63 | 803 | sg->dma_length, dir); |
1da177e4 | 804 | else if (dir == DMA_FROM_DEVICE) |
cde14bbf | 805 | dma_mark_clean(SG_ENT_VIRT_ADDRESS(sg), sg->dma_length); |
dbfd49fe | 806 | } |
1da177e4 | 807 | } |
309df0c5 AK |
808 | EXPORT_SYMBOL(swiotlb_unmap_sg_attrs); |
809 | ||
810 | void | |
811 | swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems, | |
812 | int dir) | |
813 | { | |
814 | return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL); | |
815 | } | |
1da177e4 LT |
816 | |
817 | /* | |
818 | * Make physical memory consistent for a set of streaming mode DMA translations | |
819 | * after a transfer. | |
820 | * | |
821 | * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules | |
822 | * and usage. | |
823 | */ | |
be6b0267 | 824 | static void |
dbfd49fe | 825 | swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl, |
de69e0f0 | 826 | int nelems, int dir, int target) |
1da177e4 | 827 | { |
dbfd49fe | 828 | struct scatterlist *sg; |
1da177e4 LT |
829 | int i; |
830 | ||
34814545 | 831 | BUG_ON(dir == DMA_NONE); |
1da177e4 | 832 | |
dbfd49fe | 833 | for_each_sg(sgl, sg, nelems, i) { |
1da177e4 | 834 | if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg)) |
e08e1f7a | 835 | sync_single(hwdev, swiotlb_bus_to_virt(sg->dma_address), |
de69e0f0 | 836 | sg->dma_length, dir, target); |
cde14bbf JB |
837 | else if (dir == DMA_FROM_DEVICE) |
838 | dma_mark_clean(SG_ENT_VIRT_ADDRESS(sg), sg->dma_length); | |
dbfd49fe | 839 | } |
1da177e4 LT |
840 | } |
841 | ||
8270f3f1 JL |
842 | void |
843 | swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg, | |
844 | int nelems, int dir) | |
845 | { | |
de69e0f0 | 846 | swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU); |
8270f3f1 JL |
847 | } |
848 | ||
1da177e4 LT |
849 | void |
850 | swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg, | |
851 | int nelems, int dir) | |
852 | { | |
de69e0f0 | 853 | swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE); |
1da177e4 LT |
854 | } |
855 | ||
856 | int | |
8d8bb39b | 857 | swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr) |
1da177e4 | 858 | { |
e08e1f7a | 859 | return (dma_addr == swiotlb_virt_to_bus(io_tlb_overflow_buffer)); |
1da177e4 LT |
860 | } |
861 | ||
862 | /* | |
17e5ad6c | 863 | * Return whether the given device DMA address mask can be supported |
1da177e4 | 864 | * properly. For example, if your device can only drive the low 24-bits |
17e5ad6c | 865 | * during bus mastering, then you would pass 0x00ffffff as the mask to |
1da177e4 LT |
866 | * this function. |
867 | */ | |
868 | int | |
563aaf06 | 869 | swiotlb_dma_supported(struct device *hwdev, u64 mask) |
1da177e4 | 870 | { |
e08e1f7a | 871 | return swiotlb_virt_to_bus(io_tlb_end - 1) <= mask; |
1da177e4 LT |
872 | } |
873 | ||
1da177e4 LT |
874 | EXPORT_SYMBOL(swiotlb_map_single); |
875 | EXPORT_SYMBOL(swiotlb_unmap_single); | |
876 | EXPORT_SYMBOL(swiotlb_map_sg); | |
877 | EXPORT_SYMBOL(swiotlb_unmap_sg); | |
878 | EXPORT_SYMBOL(swiotlb_sync_single_for_cpu); | |
879 | EXPORT_SYMBOL(swiotlb_sync_single_for_device); | |
878a97cf JL |
880 | EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu); |
881 | EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device); | |
1da177e4 LT |
882 | EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu); |
883 | EXPORT_SYMBOL(swiotlb_sync_sg_for_device); | |
884 | EXPORT_SYMBOL(swiotlb_dma_mapping_error); | |
25667d67 TL |
885 | EXPORT_SYMBOL(swiotlb_alloc_coherent); |
886 | EXPORT_SYMBOL(swiotlb_free_coherent); | |
1da177e4 | 887 | EXPORT_SYMBOL(swiotlb_dma_supported); |