blob: 7ff9dc36c2f803d9829e51a7669db15853920276
1 | /* |
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 | * 08/12/11 beckyb Add highmem support |
18 | */ |
19 | |
20 | #define pr_fmt(fmt) "software IO TLB: " fmt |
21 | |
22 | #include <linux/cache.h> |
23 | #include <linux/dma-mapping.h> |
24 | #include <linux/mm.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 | |
36 | #include <asm/io.h> |
37 | #include <asm/dma.h> |
38 | |
39 | #include <linux/init.h> |
40 | #include <linux/bootmem.h> |
41 | #include <linux/iommu-helper.h> |
42 | |
43 | #define CREATE_TRACE_POINTS |
44 | #include <trace/events/swiotlb.h> |
45 | |
46 | #define OFFSET(val,align) ((unsigned long) \ |
47 | ( (val) & ( (align) - 1))) |
48 | |
49 | #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT)) |
50 | |
51 | /* |
52 | * Minimum IO TLB size to bother booting with. Systems with mainly |
53 | * 64bit capable cards will only lightly use the swiotlb. If we can't |
54 | * allocate a contiguous 1MB, we're probably in trouble anyway. |
55 | */ |
56 | #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT) |
57 | |
58 | enum swiotlb_force swiotlb_force; |
59 | |
60 | /* |
61 | * Used to do a quick range check in swiotlb_tbl_unmap_single and |
62 | * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this |
63 | * API. |
64 | */ |
65 | static phys_addr_t io_tlb_start, io_tlb_end; |
66 | |
67 | /* |
68 | * The number of IO TLB blocks (in groups of 64) between io_tlb_start and |
69 | * io_tlb_end. This is command line adjustable via setup_io_tlb_npages. |
70 | */ |
71 | static unsigned long io_tlb_nslabs; |
72 | |
73 | /* |
74 | * When the IOMMU overflows we return a fallback buffer. This sets the size. |
75 | */ |
76 | static unsigned long io_tlb_overflow = 32*1024; |
77 | |
78 | static phys_addr_t io_tlb_overflow_buffer; |
79 | |
80 | /* |
81 | * This is a free list describing the number of free entries available from |
82 | * each index |
83 | */ |
84 | static unsigned int *io_tlb_list; |
85 | static unsigned int io_tlb_index; |
86 | |
87 | /* |
88 | * We need to save away the original address corresponding to a mapped entry |
89 | * for the sync operations. |
90 | */ |
91 | #define INVALID_PHYS_ADDR (~(phys_addr_t)0) |
92 | static phys_addr_t *io_tlb_orig_addr; |
93 | |
94 | /* |
95 | * Protect the above data structures in the map and unmap calls |
96 | */ |
97 | static DEFINE_SPINLOCK(io_tlb_lock); |
98 | |
99 | static int late_alloc; |
100 | |
101 | static int __init |
102 | setup_io_tlb_npages(char *str) |
103 | { |
104 | if (isdigit(*str)) { |
105 | io_tlb_nslabs = simple_strtoul(str, &str, 0); |
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 = SWIOTLB_FORCE; |
113 | } else if (!strcmp(str, "noforce")) { |
114 | swiotlb_force = SWIOTLB_NO_FORCE; |
115 | io_tlb_nslabs = 1; |
116 | } |
117 | |
118 | return 0; |
119 | } |
120 | early_param("swiotlb", setup_io_tlb_npages); |
121 | /* make io_tlb_overflow tunable too? */ |
122 | |
123 | unsigned long swiotlb_nr_tbl(void) |
124 | { |
125 | return io_tlb_nslabs; |
126 | } |
127 | EXPORT_SYMBOL_GPL(swiotlb_nr_tbl); |
128 | |
129 | /* default to 64MB */ |
130 | #define IO_TLB_DEFAULT_SIZE (64UL<<20) |
131 | unsigned long swiotlb_size_or_default(void) |
132 | { |
133 | unsigned long size; |
134 | |
135 | size = io_tlb_nslabs << IO_TLB_SHIFT; |
136 | |
137 | return size ? size : (IO_TLB_DEFAULT_SIZE); |
138 | } |
139 | |
140 | /* Note that this doesn't work with highmem page */ |
141 | static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev, |
142 | volatile void *address) |
143 | { |
144 | return phys_to_dma(hwdev, virt_to_phys(address)); |
145 | } |
146 | |
147 | static bool no_iotlb_memory; |
148 | |
149 | void swiotlb_print_info(void) |
150 | { |
151 | unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT; |
152 | |
153 | if (no_iotlb_memory) { |
154 | pr_warn("No low mem\n"); |
155 | return; |
156 | } |
157 | |
158 | pr_info("mapped [mem %#010llx-%#010llx] (%luMB)\n", |
159 | (unsigned long long)io_tlb_start, |
160 | (unsigned long long)io_tlb_end, |
161 | bytes >> 20); |
162 | } |
163 | |
164 | int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose) |
165 | { |
166 | void *v_overflow_buffer; |
167 | unsigned long i, bytes; |
168 | |
169 | bytes = nslabs << IO_TLB_SHIFT; |
170 | |
171 | io_tlb_nslabs = nslabs; |
172 | io_tlb_start = __pa(tlb); |
173 | io_tlb_end = io_tlb_start + bytes; |
174 | |
175 | /* |
176 | * Get the overflow emergency buffer |
177 | */ |
178 | v_overflow_buffer = memblock_virt_alloc_low_nopanic( |
179 | PAGE_ALIGN(io_tlb_overflow), |
180 | PAGE_SIZE); |
181 | if (!v_overflow_buffer) |
182 | return -ENOMEM; |
183 | |
184 | io_tlb_overflow_buffer = __pa(v_overflow_buffer); |
185 | |
186 | /* |
187 | * Allocate and initialize the free list array. This array is used |
188 | * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE |
189 | * between io_tlb_start and io_tlb_end. |
190 | */ |
191 | io_tlb_list = memblock_virt_alloc( |
192 | PAGE_ALIGN(io_tlb_nslabs * sizeof(int)), |
193 | PAGE_SIZE); |
194 | io_tlb_orig_addr = memblock_virt_alloc( |
195 | PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)), |
196 | PAGE_SIZE); |
197 | for (i = 0; i < io_tlb_nslabs; i++) { |
198 | io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE); |
199 | io_tlb_orig_addr[i] = INVALID_PHYS_ADDR; |
200 | } |
201 | io_tlb_index = 0; |
202 | |
203 | if (verbose) |
204 | swiotlb_print_info(); |
205 | |
206 | return 0; |
207 | } |
208 | |
209 | /* |
210 | * Statically reserve bounce buffer space and initialize bounce buffer data |
211 | * structures for the software IO TLB used to implement the DMA API. |
212 | */ |
213 | void __init |
214 | swiotlb_init(int verbose) |
215 | { |
216 | size_t default_size = IO_TLB_DEFAULT_SIZE; |
217 | unsigned char *vstart; |
218 | unsigned long bytes; |
219 | |
220 | if (!io_tlb_nslabs) { |
221 | io_tlb_nslabs = (default_size >> IO_TLB_SHIFT); |
222 | io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE); |
223 | } |
224 | |
225 | bytes = io_tlb_nslabs << IO_TLB_SHIFT; |
226 | |
227 | /* Get IO TLB memory from the low pages */ |
228 | vstart = memblock_virt_alloc_low_nopanic(PAGE_ALIGN(bytes), PAGE_SIZE); |
229 | if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose)) |
230 | return; |
231 | |
232 | if (io_tlb_start) |
233 | memblock_free_early(io_tlb_start, |
234 | PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT)); |
235 | pr_warn("Cannot allocate buffer"); |
236 | no_iotlb_memory = true; |
237 | } |
238 | |
239 | /* |
240 | * Systems with larger DMA zones (those that don't support ISA) can |
241 | * initialize the swiotlb later using the slab allocator if needed. |
242 | * This should be just like above, but with some error catching. |
243 | */ |
244 | int |
245 | swiotlb_late_init_with_default_size(size_t default_size) |
246 | { |
247 | unsigned long bytes, req_nslabs = io_tlb_nslabs; |
248 | unsigned char *vstart = NULL; |
249 | unsigned int order; |
250 | int rc = 0; |
251 | |
252 | if (!io_tlb_nslabs) { |
253 | io_tlb_nslabs = (default_size >> IO_TLB_SHIFT); |
254 | io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE); |
255 | } |
256 | |
257 | /* |
258 | * Get IO TLB memory from the low pages |
259 | */ |
260 | order = get_order(io_tlb_nslabs << IO_TLB_SHIFT); |
261 | io_tlb_nslabs = SLABS_PER_PAGE << order; |
262 | bytes = io_tlb_nslabs << IO_TLB_SHIFT; |
263 | |
264 | while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) { |
265 | vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN, |
266 | order); |
267 | if (vstart) |
268 | break; |
269 | order--; |
270 | } |
271 | |
272 | if (!vstart) { |
273 | io_tlb_nslabs = req_nslabs; |
274 | return -ENOMEM; |
275 | } |
276 | if (order != get_order(bytes)) { |
277 | pr_warn("only able to allocate %ld MB\n", |
278 | (PAGE_SIZE << order) >> 20); |
279 | io_tlb_nslabs = SLABS_PER_PAGE << order; |
280 | } |
281 | rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs); |
282 | if (rc) |
283 | free_pages((unsigned long)vstart, order); |
284 | return rc; |
285 | } |
286 | |
287 | int |
288 | swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs) |
289 | { |
290 | unsigned long i, bytes; |
291 | unsigned char *v_overflow_buffer; |
292 | |
293 | bytes = nslabs << IO_TLB_SHIFT; |
294 | |
295 | io_tlb_nslabs = nslabs; |
296 | io_tlb_start = virt_to_phys(tlb); |
297 | io_tlb_end = io_tlb_start + bytes; |
298 | |
299 | memset(tlb, 0, bytes); |
300 | |
301 | /* |
302 | * Get the overflow emergency buffer |
303 | */ |
304 | v_overflow_buffer = (void *)__get_free_pages(GFP_DMA, |
305 | get_order(io_tlb_overflow)); |
306 | if (!v_overflow_buffer) |
307 | goto cleanup2; |
308 | |
309 | io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer); |
310 | |
311 | /* |
312 | * Allocate and initialize the free list array. This array is used |
313 | * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE |
314 | * between io_tlb_start and io_tlb_end. |
315 | */ |
316 | io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL, |
317 | get_order(io_tlb_nslabs * sizeof(int))); |
318 | if (!io_tlb_list) |
319 | goto cleanup3; |
320 | |
321 | io_tlb_orig_addr = (phys_addr_t *) |
322 | __get_free_pages(GFP_KERNEL, |
323 | get_order(io_tlb_nslabs * |
324 | sizeof(phys_addr_t))); |
325 | if (!io_tlb_orig_addr) |
326 | goto cleanup4; |
327 | |
328 | for (i = 0; i < io_tlb_nslabs; i++) { |
329 | io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE); |
330 | io_tlb_orig_addr[i] = INVALID_PHYS_ADDR; |
331 | } |
332 | io_tlb_index = 0; |
333 | |
334 | swiotlb_print_info(); |
335 | |
336 | late_alloc = 1; |
337 | |
338 | return 0; |
339 | |
340 | cleanup4: |
341 | free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs * |
342 | sizeof(int))); |
343 | io_tlb_list = NULL; |
344 | cleanup3: |
345 | free_pages((unsigned long)v_overflow_buffer, |
346 | get_order(io_tlb_overflow)); |
347 | io_tlb_overflow_buffer = 0; |
348 | cleanup2: |
349 | io_tlb_end = 0; |
350 | io_tlb_start = 0; |
351 | io_tlb_nslabs = 0; |
352 | return -ENOMEM; |
353 | } |
354 | |
355 | void __init swiotlb_free(void) |
356 | { |
357 | if (!io_tlb_orig_addr) |
358 | return; |
359 | |
360 | if (late_alloc) { |
361 | free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer), |
362 | get_order(io_tlb_overflow)); |
363 | free_pages((unsigned long)io_tlb_orig_addr, |
364 | get_order(io_tlb_nslabs * sizeof(phys_addr_t))); |
365 | free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs * |
366 | sizeof(int))); |
367 | free_pages((unsigned long)phys_to_virt(io_tlb_start), |
368 | get_order(io_tlb_nslabs << IO_TLB_SHIFT)); |
369 | } else { |
370 | memblock_free_late(io_tlb_overflow_buffer, |
371 | PAGE_ALIGN(io_tlb_overflow)); |
372 | memblock_free_late(__pa(io_tlb_orig_addr), |
373 | PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t))); |
374 | memblock_free_late(__pa(io_tlb_list), |
375 | PAGE_ALIGN(io_tlb_nslabs * sizeof(int))); |
376 | memblock_free_late(io_tlb_start, |
377 | PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT)); |
378 | } |
379 | io_tlb_nslabs = 0; |
380 | } |
381 | |
382 | int is_swiotlb_buffer(phys_addr_t paddr) |
383 | { |
384 | return paddr >= io_tlb_start && paddr < io_tlb_end; |
385 | } |
386 | |
387 | /* |
388 | * Bounce: copy the swiotlb buffer back to the original dma location |
389 | */ |
390 | static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr, |
391 | size_t size, enum dma_data_direction dir) |
392 | { |
393 | unsigned long pfn = PFN_DOWN(orig_addr); |
394 | unsigned char *vaddr = phys_to_virt(tlb_addr); |
395 | |
396 | if (PageHighMem(pfn_to_page(pfn))) { |
397 | /* The buffer does not have a mapping. Map it in and copy */ |
398 | unsigned int offset = orig_addr & ~PAGE_MASK; |
399 | char *buffer; |
400 | unsigned int sz = 0; |
401 | unsigned long flags; |
402 | |
403 | while (size) { |
404 | sz = min_t(size_t, PAGE_SIZE - offset, size); |
405 | |
406 | local_irq_save(flags); |
407 | buffer = kmap_atomic(pfn_to_page(pfn)); |
408 | if (dir == DMA_TO_DEVICE) |
409 | memcpy(vaddr, buffer + offset, sz); |
410 | else |
411 | memcpy(buffer + offset, vaddr, sz); |
412 | kunmap_atomic(buffer); |
413 | local_irq_restore(flags); |
414 | |
415 | size -= sz; |
416 | pfn++; |
417 | vaddr += sz; |
418 | offset = 0; |
419 | } |
420 | } else if (dir == DMA_TO_DEVICE) { |
421 | memcpy(vaddr, phys_to_virt(orig_addr), size); |
422 | } else { |
423 | memcpy(phys_to_virt(orig_addr), vaddr, size); |
424 | } |
425 | } |
426 | |
427 | phys_addr_t swiotlb_tbl_map_single(struct device *hwdev, |
428 | dma_addr_t tbl_dma_addr, |
429 | phys_addr_t orig_addr, size_t size, |
430 | enum dma_data_direction dir) |
431 | { |
432 | unsigned long flags; |
433 | phys_addr_t tlb_addr; |
434 | unsigned int nslots, stride, index, wrap; |
435 | int i; |
436 | unsigned long mask; |
437 | unsigned long offset_slots; |
438 | unsigned long max_slots; |
439 | |
440 | if (no_iotlb_memory) |
441 | panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer"); |
442 | |
443 | mask = dma_get_seg_boundary(hwdev); |
444 | |
445 | tbl_dma_addr &= mask; |
446 | |
447 | offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT; |
448 | |
449 | /* |
450 | * Carefully handle integer overflow which can occur when mask == ~0UL. |
451 | */ |
452 | max_slots = mask + 1 |
453 | ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT |
454 | : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT); |
455 | |
456 | /* |
457 | * For mappings greater than or equal to a page, we limit the stride |
458 | * (and hence alignment) to a page size. |
459 | */ |
460 | nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT; |
461 | if (size >= PAGE_SIZE) |
462 | stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT)); |
463 | else |
464 | stride = 1; |
465 | |
466 | BUG_ON(!nslots); |
467 | |
468 | /* |
469 | * Find suitable number of IO TLB entries size that will fit this |
470 | * request and allocate a buffer from that IO TLB pool. |
471 | */ |
472 | spin_lock_irqsave(&io_tlb_lock, flags); |
473 | index = ALIGN(io_tlb_index, stride); |
474 | if (index >= io_tlb_nslabs) |
475 | index = 0; |
476 | wrap = index; |
477 | |
478 | do { |
479 | while (iommu_is_span_boundary(index, nslots, offset_slots, |
480 | max_slots)) { |
481 | index += stride; |
482 | if (index >= io_tlb_nslabs) |
483 | index = 0; |
484 | if (index == wrap) |
485 | goto not_found; |
486 | } |
487 | |
488 | /* |
489 | * If we find a slot that indicates we have 'nslots' number of |
490 | * contiguous buffers, we allocate the buffers from that slot |
491 | * and mark the entries as '0' indicating unavailable. |
492 | */ |
493 | if (io_tlb_list[index] >= nslots) { |
494 | int count = 0; |
495 | |
496 | for (i = index; i < (int) (index + nslots); i++) |
497 | io_tlb_list[i] = 0; |
498 | for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--) |
499 | io_tlb_list[i] = ++count; |
500 | tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT); |
501 | |
502 | /* |
503 | * Update the indices to avoid searching in the next |
504 | * round. |
505 | */ |
506 | io_tlb_index = ((index + nslots) < io_tlb_nslabs |
507 | ? (index + nslots) : 0); |
508 | |
509 | goto found; |
510 | } |
511 | index += stride; |
512 | if (index >= io_tlb_nslabs) |
513 | index = 0; |
514 | } while (index != wrap); |
515 | |
516 | not_found: |
517 | spin_unlock_irqrestore(&io_tlb_lock, flags); |
518 | if (printk_ratelimit()) |
519 | dev_warn(hwdev, "swiotlb buffer is full (sz: %zd bytes)\n", size); |
520 | return SWIOTLB_MAP_ERROR; |
521 | found: |
522 | spin_unlock_irqrestore(&io_tlb_lock, flags); |
523 | |
524 | /* |
525 | * Save away the mapping from the original address to the DMA address. |
526 | * This is needed when we sync the memory. Then we sync the buffer if |
527 | * needed. |
528 | */ |
529 | for (i = 0; i < nslots; i++) |
530 | io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT); |
531 | if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL) |
532 | swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE); |
533 | |
534 | return tlb_addr; |
535 | } |
536 | EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single); |
537 | |
538 | /* |
539 | * Allocates bounce buffer and returns its kernel virtual address. |
540 | */ |
541 | |
542 | static phys_addr_t |
543 | map_single(struct device *hwdev, phys_addr_t phys, size_t size, |
544 | enum dma_data_direction dir) |
545 | { |
546 | dma_addr_t start_dma_addr; |
547 | |
548 | if (swiotlb_force == SWIOTLB_NO_FORCE) { |
549 | dev_warn_ratelimited(hwdev, "Cannot do DMA to address %pa\n", |
550 | &phys); |
551 | return SWIOTLB_MAP_ERROR; |
552 | } |
553 | |
554 | start_dma_addr = phys_to_dma(hwdev, io_tlb_start); |
555 | return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir); |
556 | } |
557 | |
558 | /* |
559 | * dma_addr is the kernel virtual address of the bounce buffer to unmap. |
560 | */ |
561 | void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr, |
562 | size_t size, enum dma_data_direction dir) |
563 | { |
564 | unsigned long flags; |
565 | int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT; |
566 | int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT; |
567 | phys_addr_t orig_addr = io_tlb_orig_addr[index]; |
568 | |
569 | /* |
570 | * First, sync the memory before unmapping the entry |
571 | */ |
572 | if (orig_addr != INVALID_PHYS_ADDR && |
573 | ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL))) |
574 | swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE); |
575 | |
576 | /* |
577 | * Return the buffer to the free list by setting the corresponding |
578 | * entries to indicate the number of contiguous entries available. |
579 | * While returning the entries to the free list, we merge the entries |
580 | * with slots below and above the pool being returned. |
581 | */ |
582 | spin_lock_irqsave(&io_tlb_lock, flags); |
583 | { |
584 | count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ? |
585 | io_tlb_list[index + nslots] : 0); |
586 | /* |
587 | * Step 1: return the slots to the free list, merging the |
588 | * slots with superceeding slots |
589 | */ |
590 | for (i = index + nslots - 1; i >= index; i--) { |
591 | io_tlb_list[i] = ++count; |
592 | io_tlb_orig_addr[i] = INVALID_PHYS_ADDR; |
593 | } |
594 | /* |
595 | * Step 2: merge the returned slots with the preceding slots, |
596 | * if available (non zero) |
597 | */ |
598 | for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--) |
599 | io_tlb_list[i] = ++count; |
600 | } |
601 | spin_unlock_irqrestore(&io_tlb_lock, flags); |
602 | } |
603 | EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single); |
604 | |
605 | void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr, |
606 | size_t size, enum dma_data_direction dir, |
607 | enum dma_sync_target target) |
608 | { |
609 | int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT; |
610 | phys_addr_t orig_addr = io_tlb_orig_addr[index]; |
611 | |
612 | if (orig_addr == INVALID_PHYS_ADDR) |
613 | return; |
614 | orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1); |
615 | |
616 | switch (target) { |
617 | case SYNC_FOR_CPU: |
618 | if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)) |
619 | swiotlb_bounce(orig_addr, tlb_addr, |
620 | size, DMA_FROM_DEVICE); |
621 | else |
622 | BUG_ON(dir != DMA_TO_DEVICE); |
623 | break; |
624 | case SYNC_FOR_DEVICE: |
625 | if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)) |
626 | swiotlb_bounce(orig_addr, tlb_addr, |
627 | size, DMA_TO_DEVICE); |
628 | else |
629 | BUG_ON(dir != DMA_FROM_DEVICE); |
630 | break; |
631 | default: |
632 | BUG(); |
633 | } |
634 | } |
635 | EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single); |
636 | |
637 | void * |
638 | swiotlb_alloc_coherent(struct device *hwdev, size_t size, |
639 | dma_addr_t *dma_handle, gfp_t flags) |
640 | { |
641 | dma_addr_t dev_addr; |
642 | void *ret; |
643 | int order = get_order(size); |
644 | u64 dma_mask = DMA_BIT_MASK(32); |
645 | |
646 | if (hwdev && hwdev->coherent_dma_mask) |
647 | dma_mask = hwdev->coherent_dma_mask; |
648 | |
649 | ret = (void *)__get_free_pages(flags, order); |
650 | if (ret) { |
651 | dev_addr = swiotlb_virt_to_bus(hwdev, ret); |
652 | if (dev_addr + size - 1 > dma_mask) { |
653 | /* |
654 | * The allocated memory isn't reachable by the device. |
655 | */ |
656 | free_pages((unsigned long) ret, order); |
657 | ret = NULL; |
658 | } |
659 | } |
660 | if (!ret) { |
661 | /* |
662 | * We are either out of memory or the device can't DMA to |
663 | * GFP_DMA memory; fall back on map_single(), which |
664 | * will grab memory from the lowest available address range. |
665 | */ |
666 | phys_addr_t paddr = map_single(hwdev, 0, size, DMA_FROM_DEVICE); |
667 | if (paddr == SWIOTLB_MAP_ERROR) |
668 | goto err_warn; |
669 | |
670 | ret = phys_to_virt(paddr); |
671 | dev_addr = phys_to_dma(hwdev, paddr); |
672 | |
673 | /* Confirm address can be DMA'd by device */ |
674 | if (dev_addr + size - 1 > dma_mask) { |
675 | printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n", |
676 | (unsigned long long)dma_mask, |
677 | (unsigned long long)dev_addr); |
678 | |
679 | /* DMA_TO_DEVICE to avoid memcpy in unmap_single */ |
680 | swiotlb_tbl_unmap_single(hwdev, paddr, |
681 | size, DMA_TO_DEVICE); |
682 | goto err_warn; |
683 | } |
684 | } |
685 | |
686 | *dma_handle = dev_addr; |
687 | memset(ret, 0, size); |
688 | |
689 | return ret; |
690 | |
691 | err_warn: |
692 | pr_warn("coherent allocation failed for device %s size=%zu\n", |
693 | dev_name(hwdev), size); |
694 | dump_stack(); |
695 | |
696 | return NULL; |
697 | } |
698 | EXPORT_SYMBOL(swiotlb_alloc_coherent); |
699 | |
700 | void |
701 | swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr, |
702 | dma_addr_t dev_addr) |
703 | { |
704 | phys_addr_t paddr = dma_to_phys(hwdev, dev_addr); |
705 | |
706 | WARN_ON(irqs_disabled()); |
707 | if (!is_swiotlb_buffer(paddr)) |
708 | free_pages((unsigned long)vaddr, get_order(size)); |
709 | else |
710 | /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */ |
711 | swiotlb_tbl_unmap_single(hwdev, paddr, size, DMA_TO_DEVICE); |
712 | } |
713 | EXPORT_SYMBOL(swiotlb_free_coherent); |
714 | |
715 | static void |
716 | swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir, |
717 | int do_panic) |
718 | { |
719 | if (swiotlb_force == SWIOTLB_NO_FORCE) |
720 | return; |
721 | |
722 | /* |
723 | * Ran out of IOMMU space for this operation. This is very bad. |
724 | * Unfortunately the drivers cannot handle this operation properly. |
725 | * unless they check for dma_mapping_error (most don't) |
726 | * When the mapping is small enough return a static buffer to limit |
727 | * the damage, or panic when the transfer is too big. |
728 | */ |
729 | printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at " |
730 | "device %s\n", size, dev ? dev_name(dev) : "?"); |
731 | |
732 | if (size <= io_tlb_overflow || !do_panic) |
733 | return; |
734 | |
735 | if (dir == DMA_BIDIRECTIONAL) |
736 | panic("DMA: Random memory could be DMA accessed\n"); |
737 | if (dir == DMA_FROM_DEVICE) |
738 | panic("DMA: Random memory could be DMA written\n"); |
739 | if (dir == DMA_TO_DEVICE) |
740 | panic("DMA: Random memory could be DMA read\n"); |
741 | } |
742 | |
743 | /* |
744 | * Map a single buffer of the indicated size for DMA in streaming mode. The |
745 | * physical address to use is returned. |
746 | * |
747 | * Once the device is given the dma address, the device owns this memory until |
748 | * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed. |
749 | */ |
750 | dma_addr_t swiotlb_map_page(struct device *dev, struct page *page, |
751 | unsigned long offset, size_t size, |
752 | enum dma_data_direction dir, |
753 | unsigned long attrs) |
754 | { |
755 | phys_addr_t map, phys = page_to_phys(page) + offset; |
756 | dma_addr_t dev_addr = phys_to_dma(dev, phys); |
757 | |
758 | BUG_ON(dir == DMA_NONE); |
759 | /* |
760 | * If the address happens to be in the device's DMA window, |
761 | * we can safely return the device addr and not worry about bounce |
762 | * buffering it. |
763 | */ |
764 | if (dma_capable(dev, dev_addr, size) && swiotlb_force != SWIOTLB_FORCE) |
765 | return dev_addr; |
766 | |
767 | trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force); |
768 | |
769 | /* Oh well, have to allocate and map a bounce buffer. */ |
770 | map = map_single(dev, phys, size, dir); |
771 | if (map == SWIOTLB_MAP_ERROR) { |
772 | swiotlb_full(dev, size, dir, 1); |
773 | return phys_to_dma(dev, io_tlb_overflow_buffer); |
774 | } |
775 | |
776 | dev_addr = phys_to_dma(dev, map); |
777 | |
778 | /* Ensure that the address returned is DMA'ble */ |
779 | if (!dma_capable(dev, dev_addr, size)) { |
780 | swiotlb_tbl_unmap_single(dev, map, size, dir); |
781 | return phys_to_dma(dev, io_tlb_overflow_buffer); |
782 | } |
783 | |
784 | return dev_addr; |
785 | } |
786 | EXPORT_SYMBOL_GPL(swiotlb_map_page); |
787 | |
788 | /* |
789 | * Unmap a single streaming mode DMA translation. The dma_addr and size must |
790 | * match what was provided for in a previous swiotlb_map_page call. All |
791 | * other usages are undefined. |
792 | * |
793 | * After this call, reads by the cpu to the buffer are guaranteed to see |
794 | * whatever the device wrote there. |
795 | */ |
796 | static void unmap_single(struct device *hwdev, dma_addr_t dev_addr, |
797 | size_t size, enum dma_data_direction dir) |
798 | { |
799 | phys_addr_t paddr = dma_to_phys(hwdev, dev_addr); |
800 | |
801 | BUG_ON(dir == DMA_NONE); |
802 | |
803 | if (is_swiotlb_buffer(paddr)) { |
804 | swiotlb_tbl_unmap_single(hwdev, paddr, size, dir); |
805 | return; |
806 | } |
807 | |
808 | if (dir != DMA_FROM_DEVICE) |
809 | return; |
810 | |
811 | /* |
812 | * phys_to_virt doesn't work with hihgmem page but we could |
813 | * call dma_mark_clean() with hihgmem page here. However, we |
814 | * are fine since dma_mark_clean() is null on POWERPC. We can |
815 | * make dma_mark_clean() take a physical address if necessary. |
816 | */ |
817 | dma_mark_clean(phys_to_virt(paddr), size); |
818 | } |
819 | |
820 | void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr, |
821 | size_t size, enum dma_data_direction dir, |
822 | unsigned long attrs) |
823 | { |
824 | unmap_single(hwdev, dev_addr, size, dir); |
825 | } |
826 | EXPORT_SYMBOL_GPL(swiotlb_unmap_page); |
827 | |
828 | /* |
829 | * Make physical memory consistent for a single streaming mode DMA translation |
830 | * after a transfer. |
831 | * |
832 | * If you perform a swiotlb_map_page() but wish to interrogate the buffer |
833 | * using the cpu, yet do not wish to teardown the dma mapping, you must |
834 | * call this function before doing so. At the next point you give the dma |
835 | * address back to the card, you must first perform a |
836 | * swiotlb_dma_sync_for_device, and then the device again owns the buffer |
837 | */ |
838 | static void |
839 | swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr, |
840 | size_t size, enum dma_data_direction dir, |
841 | enum dma_sync_target target) |
842 | { |
843 | phys_addr_t paddr = dma_to_phys(hwdev, dev_addr); |
844 | |
845 | BUG_ON(dir == DMA_NONE); |
846 | |
847 | if (is_swiotlb_buffer(paddr)) { |
848 | swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target); |
849 | return; |
850 | } |
851 | |
852 | if (dir != DMA_FROM_DEVICE) |
853 | return; |
854 | |
855 | dma_mark_clean(phys_to_virt(paddr), size); |
856 | } |
857 | |
858 | void |
859 | swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr, |
860 | size_t size, enum dma_data_direction dir) |
861 | { |
862 | swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU); |
863 | } |
864 | EXPORT_SYMBOL(swiotlb_sync_single_for_cpu); |
865 | |
866 | void |
867 | swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr, |
868 | size_t size, enum dma_data_direction dir) |
869 | { |
870 | swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE); |
871 | } |
872 | EXPORT_SYMBOL(swiotlb_sync_single_for_device); |
873 | |
874 | /* |
875 | * Map a set of buffers described by scatterlist in streaming mode for DMA. |
876 | * This is the scatter-gather version of the above swiotlb_map_page |
877 | * interface. Here the scatter gather list elements are each tagged with the |
878 | * appropriate dma address and length. They are obtained via |
879 | * sg_dma_{address,length}(SG). |
880 | * |
881 | * NOTE: An implementation may be able to use a smaller number of |
882 | * DMA address/length pairs than there are SG table elements. |
883 | * (for example via virtual mapping capabilities) |
884 | * The routine returns the number of addr/length pairs actually |
885 | * used, at most nents. |
886 | * |
887 | * Device ownership issues as mentioned above for swiotlb_map_page are the |
888 | * same here. |
889 | */ |
890 | int |
891 | swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems, |
892 | enum dma_data_direction dir, unsigned long attrs) |
893 | { |
894 | struct scatterlist *sg; |
895 | int i; |
896 | |
897 | BUG_ON(dir == DMA_NONE); |
898 | |
899 | for_each_sg(sgl, sg, nelems, i) { |
900 | phys_addr_t paddr = sg_phys(sg); |
901 | dma_addr_t dev_addr = phys_to_dma(hwdev, paddr); |
902 | |
903 | if (swiotlb_force == SWIOTLB_FORCE || |
904 | !dma_capable(hwdev, dev_addr, sg->length)) { |
905 | phys_addr_t map = map_single(hwdev, sg_phys(sg), |
906 | sg->length, dir); |
907 | if (map == SWIOTLB_MAP_ERROR) { |
908 | /* Don't panic here, we expect map_sg users |
909 | to do proper error handling. */ |
910 | swiotlb_full(hwdev, sg->length, dir, 0); |
911 | swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir, |
912 | attrs); |
913 | sg_dma_len(sgl) = 0; |
914 | return 0; |
915 | } |
916 | sg->dma_address = phys_to_dma(hwdev, map); |
917 | } else |
918 | sg->dma_address = dev_addr; |
919 | sg_dma_len(sg) = sg->length; |
920 | } |
921 | return nelems; |
922 | } |
923 | EXPORT_SYMBOL(swiotlb_map_sg_attrs); |
924 | |
925 | int |
926 | swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems, |
927 | enum dma_data_direction dir) |
928 | { |
929 | return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, 0); |
930 | } |
931 | EXPORT_SYMBOL(swiotlb_map_sg); |
932 | |
933 | /* |
934 | * Unmap a set of streaming mode DMA translations. Again, cpu read rules |
935 | * concerning calls here are the same as for swiotlb_unmap_page() above. |
936 | */ |
937 | void |
938 | swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl, |
939 | int nelems, enum dma_data_direction dir, |
940 | unsigned long attrs) |
941 | { |
942 | struct scatterlist *sg; |
943 | int i; |
944 | |
945 | BUG_ON(dir == DMA_NONE); |
946 | |
947 | for_each_sg(sgl, sg, nelems, i) |
948 | unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir); |
949 | |
950 | } |
951 | EXPORT_SYMBOL(swiotlb_unmap_sg_attrs); |
952 | |
953 | void |
954 | swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems, |
955 | enum dma_data_direction dir) |
956 | { |
957 | return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, 0); |
958 | } |
959 | EXPORT_SYMBOL(swiotlb_unmap_sg); |
960 | |
961 | /* |
962 | * Make physical memory consistent for a set of streaming mode DMA translations |
963 | * after a transfer. |
964 | * |
965 | * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules |
966 | * and usage. |
967 | */ |
968 | static void |
969 | swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl, |
970 | int nelems, enum dma_data_direction dir, |
971 | enum dma_sync_target target) |
972 | { |
973 | struct scatterlist *sg; |
974 | int i; |
975 | |
976 | for_each_sg(sgl, sg, nelems, i) |
977 | swiotlb_sync_single(hwdev, sg->dma_address, |
978 | sg_dma_len(sg), dir, target); |
979 | } |
980 | |
981 | void |
982 | swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg, |
983 | int nelems, enum dma_data_direction dir) |
984 | { |
985 | swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU); |
986 | } |
987 | EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu); |
988 | |
989 | void |
990 | swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg, |
991 | int nelems, enum dma_data_direction dir) |
992 | { |
993 | swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE); |
994 | } |
995 | EXPORT_SYMBOL(swiotlb_sync_sg_for_device); |
996 | |
997 | int |
998 | swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr) |
999 | { |
1000 | return (dma_addr == phys_to_dma(hwdev, io_tlb_overflow_buffer)); |
1001 | } |
1002 | EXPORT_SYMBOL(swiotlb_dma_mapping_error); |
1003 | |
1004 | /* |
1005 | * Return whether the given device DMA address mask can be supported |
1006 | * properly. For example, if your device can only drive the low 24-bits |
1007 | * during bus mastering, then you would pass 0x00ffffff as the mask to |
1008 | * this function. |
1009 | */ |
1010 | int |
1011 | swiotlb_dma_supported(struct device *hwdev, u64 mask) |
1012 | { |
1013 | return phys_to_dma(hwdev, io_tlb_end - 1) <= mask; |
1014 | } |
1015 | EXPORT_SYMBOL(swiotlb_dma_supported); |
1016 |