• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 int 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
setup_io_tlb_npages(char * str)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 = 1;
113 
114 	return 0;
115 }
116 early_param("swiotlb", setup_io_tlb_npages);
117 /* make io_tlb_overflow tunable too? */
118 
swiotlb_nr_tbl(void)119 unsigned long swiotlb_nr_tbl(void)
120 {
121 	return io_tlb_nslabs;
122 }
123 EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
124 
125 /* default to 64MB */
126 #define IO_TLB_DEFAULT_SIZE (64UL<<20)
swiotlb_size_or_default(void)127 unsigned long swiotlb_size_or_default(void)
128 {
129 	unsigned long size;
130 
131 	size = io_tlb_nslabs << IO_TLB_SHIFT;
132 
133 	return size ? size : (IO_TLB_DEFAULT_SIZE);
134 }
135 
136 /* Note that this doesn't work with highmem page */
swiotlb_virt_to_bus(struct device * hwdev,volatile void * address)137 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
138 				      volatile void *address)
139 {
140 	return phys_to_dma(hwdev, virt_to_phys(address));
141 }
142 
143 static bool no_iotlb_memory;
144 
swiotlb_print_info(void)145 void swiotlb_print_info(void)
146 {
147 	unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
148 
149 	if (no_iotlb_memory) {
150 		pr_warn("No low mem\n");
151 		return;
152 	}
153 
154 	pr_info("mapped [mem %#010llx-%#010llx] (%luMB)\n",
155 	       (unsigned long long)io_tlb_start,
156 	       (unsigned long long)io_tlb_end,
157 	       bytes >> 20);
158 }
159 
swiotlb_init_with_tbl(char * tlb,unsigned long nslabs,int verbose)160 int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
161 {
162 	void *v_overflow_buffer;
163 	unsigned long i, bytes;
164 
165 	bytes = nslabs << IO_TLB_SHIFT;
166 
167 	io_tlb_nslabs = nslabs;
168 	io_tlb_start = __pa(tlb);
169 	io_tlb_end = io_tlb_start + bytes;
170 
171 	/*
172 	 * Get the overflow emergency buffer
173 	 */
174 	v_overflow_buffer = memblock_virt_alloc_low_nopanic(
175 						PAGE_ALIGN(io_tlb_overflow),
176 						PAGE_SIZE);
177 	if (!v_overflow_buffer)
178 		return -ENOMEM;
179 
180 	io_tlb_overflow_buffer = __pa(v_overflow_buffer);
181 
182 	/*
183 	 * Allocate and initialize the free list array.  This array is used
184 	 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
185 	 * between io_tlb_start and io_tlb_end.
186 	 */
187 	io_tlb_list = memblock_virt_alloc(
188 				PAGE_ALIGN(io_tlb_nslabs * sizeof(int)),
189 				PAGE_SIZE);
190 	io_tlb_orig_addr = memblock_virt_alloc(
191 				PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)),
192 				PAGE_SIZE);
193 	for (i = 0; i < io_tlb_nslabs; i++) {
194 		io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
195 		io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
196 	}
197 	io_tlb_index = 0;
198 	no_iotlb_memory = false;
199 
200 	if (verbose)
201 		swiotlb_print_info();
202 
203 	return 0;
204 }
205 
206 /*
207  * Statically reserve bounce buffer space and initialize bounce buffer data
208  * structures for the software IO TLB used to implement the DMA API.
209  */
210 void  __init
swiotlb_init(int verbose)211 swiotlb_init(int verbose)
212 {
213 	size_t default_size = IO_TLB_DEFAULT_SIZE;
214 	unsigned char *vstart;
215 	unsigned long bytes;
216 
217 	if (!io_tlb_nslabs) {
218 		io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
219 		io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
220 	}
221 
222 	bytes = io_tlb_nslabs << IO_TLB_SHIFT;
223 
224 	/* Get IO TLB memory from the low pages */
225 	vstart = memblock_virt_alloc_low_nopanic(PAGE_ALIGN(bytes), PAGE_SIZE);
226 	if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose))
227 		return;
228 
229 	if (io_tlb_start) {
230 		memblock_free_early(io_tlb_start,
231 				    PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
232 		io_tlb_start = 0;
233 	}
234 	pr_warn("Cannot allocate buffer");
235 	no_iotlb_memory = true;
236 }
237 
238 /*
239  * Systems with larger DMA zones (those that don't support ISA) can
240  * initialize the swiotlb later using the slab allocator if needed.
241  * This should be just like above, but with some error catching.
242  */
243 int
swiotlb_late_init_with_default_size(size_t default_size)244 swiotlb_late_init_with_default_size(size_t default_size)
245 {
246 	unsigned long bytes, req_nslabs = io_tlb_nslabs;
247 	unsigned char *vstart = NULL;
248 	unsigned int order;
249 	int rc = 0;
250 
251 	if (!io_tlb_nslabs) {
252 		io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
253 		io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
254 	}
255 
256 	/*
257 	 * Get IO TLB memory from the low pages
258 	 */
259 	order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
260 	io_tlb_nslabs = SLABS_PER_PAGE << order;
261 	bytes = io_tlb_nslabs << IO_TLB_SHIFT;
262 
263 	while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
264 		vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
265 						  order);
266 		if (vstart)
267 			break;
268 		order--;
269 	}
270 
271 	if (!vstart) {
272 		io_tlb_nslabs = req_nslabs;
273 		return -ENOMEM;
274 	}
275 	if (order != get_order(bytes)) {
276 		pr_warn("only able to allocate %ld MB\n",
277 			(PAGE_SIZE << order) >> 20);
278 		io_tlb_nslabs = SLABS_PER_PAGE << order;
279 	}
280 	rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
281 	if (rc)
282 		free_pages((unsigned long)vstart, order);
283 	return rc;
284 }
285 
286 int
swiotlb_late_init_with_tbl(char * tlb,unsigned long nslabs)287 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
288 {
289 	unsigned long i, bytes;
290 	unsigned char *v_overflow_buffer;
291 
292 	bytes = nslabs << IO_TLB_SHIFT;
293 
294 	io_tlb_nslabs = nslabs;
295 	io_tlb_start = virt_to_phys(tlb);
296 	io_tlb_end = io_tlb_start + bytes;
297 
298 	memset(tlb, 0, bytes);
299 
300 	/*
301 	 * Get the overflow emergency buffer
302 	 */
303 	v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
304 						     get_order(io_tlb_overflow));
305 	if (!v_overflow_buffer)
306 		goto cleanup2;
307 
308 	io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
309 
310 	/*
311 	 * Allocate and initialize the free list array.  This array is used
312 	 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
313 	 * between io_tlb_start and io_tlb_end.
314 	 */
315 	io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
316 	                              get_order(io_tlb_nslabs * sizeof(int)));
317 	if (!io_tlb_list)
318 		goto cleanup3;
319 
320 	io_tlb_orig_addr = (phys_addr_t *)
321 		__get_free_pages(GFP_KERNEL,
322 				 get_order(io_tlb_nslabs *
323 					   sizeof(phys_addr_t)));
324 	if (!io_tlb_orig_addr)
325 		goto cleanup4;
326 
327 	for (i = 0; i < io_tlb_nslabs; i++) {
328 		io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
329 		io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
330 	}
331 	io_tlb_index = 0;
332 	no_iotlb_memory = false;
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 
swiotlb_free(void)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 
is_swiotlb_buffer(phys_addr_t paddr)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  */
swiotlb_bounce(phys_addr_t orig_addr,phys_addr_t tlb_addr,size_t size,enum dma_data_direction dir)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 
swiotlb_tbl_map_single(struct device * hwdev,dma_addr_t tbl_dma_addr,phys_addr_t orig_addr,size_t size,enum dma_data_direction dir)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
map_single(struct device * hwdev,phys_addr_t phys,size_t size,enum dma_data_direction dir)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 = phys_to_dma(hwdev, io_tlb_start);
547 
548 	return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
549 }
550 
551 /*
552  * dma_addr is the kernel virtual address of the bounce buffer to unmap.
553  */
swiotlb_tbl_unmap_single(struct device * hwdev,phys_addr_t tlb_addr,size_t size,enum dma_data_direction dir)554 void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
555 			      size_t size, enum dma_data_direction dir)
556 {
557 	unsigned long flags;
558 	int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
559 	int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
560 	phys_addr_t orig_addr = io_tlb_orig_addr[index];
561 
562 	/*
563 	 * First, sync the memory before unmapping the entry
564 	 */
565 	if (orig_addr != INVALID_PHYS_ADDR &&
566 	    ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
567 		swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
568 
569 	/*
570 	 * Return the buffer to the free list by setting the corresponding
571 	 * entries to indicate the number of contiguous entries available.
572 	 * While returning the entries to the free list, we merge the entries
573 	 * with slots below and above the pool being returned.
574 	 */
575 	spin_lock_irqsave(&io_tlb_lock, flags);
576 	{
577 		count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
578 			 io_tlb_list[index + nslots] : 0);
579 		/*
580 		 * Step 1: return the slots to the free list, merging the
581 		 * slots with superceeding slots
582 		 */
583 		for (i = index + nslots - 1; i >= index; i--) {
584 			io_tlb_list[i] = ++count;
585 			io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
586 		}
587 		/*
588 		 * Step 2: merge the returned slots with the preceding slots,
589 		 * if available (non zero)
590 		 */
591 		for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
592 			io_tlb_list[i] = ++count;
593 	}
594 	spin_unlock_irqrestore(&io_tlb_lock, flags);
595 }
596 EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
597 
swiotlb_tbl_sync_single(struct device * hwdev,phys_addr_t tlb_addr,size_t size,enum dma_data_direction dir,enum dma_sync_target target)598 void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
599 			     size_t size, enum dma_data_direction dir,
600 			     enum dma_sync_target target)
601 {
602 	int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
603 	phys_addr_t orig_addr = io_tlb_orig_addr[index];
604 
605 	if (orig_addr == INVALID_PHYS_ADDR)
606 		return;
607 	orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
608 
609 	switch (target) {
610 	case SYNC_FOR_CPU:
611 		if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
612 			swiotlb_bounce(orig_addr, tlb_addr,
613 				       size, DMA_FROM_DEVICE);
614 		else
615 			BUG_ON(dir != DMA_TO_DEVICE);
616 		break;
617 	case SYNC_FOR_DEVICE:
618 		if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
619 			swiotlb_bounce(orig_addr, tlb_addr,
620 				       size, DMA_TO_DEVICE);
621 		else
622 			BUG_ON(dir != DMA_FROM_DEVICE);
623 		break;
624 	default:
625 		BUG();
626 	}
627 }
628 EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
629 
630 void *
swiotlb_alloc_coherent(struct device * hwdev,size_t size,dma_addr_t * dma_handle,gfp_t flags)631 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
632 		       dma_addr_t *dma_handle, gfp_t flags)
633 {
634 	dma_addr_t dev_addr;
635 	void *ret;
636 	int order = get_order(size);
637 	u64 dma_mask = DMA_BIT_MASK(32);
638 
639 	if (hwdev && hwdev->coherent_dma_mask)
640 		dma_mask = hwdev->coherent_dma_mask;
641 
642 	ret = (void *)__get_free_pages(flags, order);
643 	if (ret) {
644 		dev_addr = swiotlb_virt_to_bus(hwdev, ret);
645 		if (dev_addr + size - 1 > dma_mask) {
646 			/*
647 			 * The allocated memory isn't reachable by the device.
648 			 */
649 			free_pages((unsigned long) ret, order);
650 			ret = NULL;
651 		}
652 	}
653 	if (!ret) {
654 		/*
655 		 * We are either out of memory or the device can't DMA to
656 		 * GFP_DMA memory; fall back on map_single(), which
657 		 * will grab memory from the lowest available address range.
658 		 */
659 		phys_addr_t paddr = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
660 		if (paddr == SWIOTLB_MAP_ERROR)
661 			goto err_warn;
662 
663 		ret = phys_to_virt(paddr);
664 		dev_addr = phys_to_dma(hwdev, paddr);
665 
666 		/* Confirm address can be DMA'd by device */
667 		if (dev_addr + size - 1 > dma_mask) {
668 			printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
669 			       (unsigned long long)dma_mask,
670 			       (unsigned long long)dev_addr);
671 
672 			/* DMA_TO_DEVICE to avoid memcpy in unmap_single */
673 			swiotlb_tbl_unmap_single(hwdev, paddr,
674 						 size, DMA_TO_DEVICE);
675 			goto err_warn;
676 		}
677 	}
678 
679 	*dma_handle = dev_addr;
680 	memset(ret, 0, size);
681 
682 	return ret;
683 
684 err_warn:
685 	pr_warn("coherent allocation failed for device %s size=%zu\n",
686 		dev_name(hwdev), size);
687 	dump_stack();
688 
689 	return NULL;
690 }
691 EXPORT_SYMBOL(swiotlb_alloc_coherent);
692 
693 void
swiotlb_free_coherent(struct device * hwdev,size_t size,void * vaddr,dma_addr_t dev_addr)694 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
695 		      dma_addr_t dev_addr)
696 {
697 	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
698 
699 	WARN_ON(irqs_disabled());
700 	if (!is_swiotlb_buffer(paddr))
701 		free_pages((unsigned long)vaddr, get_order(size));
702 	else
703 		/* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
704 		swiotlb_tbl_unmap_single(hwdev, paddr, size, DMA_TO_DEVICE);
705 }
706 EXPORT_SYMBOL(swiotlb_free_coherent);
707 
708 static void
swiotlb_full(struct device * dev,size_t size,enum dma_data_direction dir,int do_panic)709 swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
710 	     int do_panic)
711 {
712 	/*
713 	 * Ran out of IOMMU space for this operation. This is very bad.
714 	 * Unfortunately the drivers cannot handle this operation properly.
715 	 * unless they check for dma_mapping_error (most don't)
716 	 * When the mapping is small enough return a static buffer to limit
717 	 * the damage, or panic when the transfer is too big.
718 	 */
719 	printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
720 	       "device %s\n", size, dev ? dev_name(dev) : "?");
721 
722 	if (size <= io_tlb_overflow || !do_panic)
723 		return;
724 
725 	if (dir == DMA_BIDIRECTIONAL)
726 		panic("DMA: Random memory could be DMA accessed\n");
727 	if (dir == DMA_FROM_DEVICE)
728 		panic("DMA: Random memory could be DMA written\n");
729 	if (dir == DMA_TO_DEVICE)
730 		panic("DMA: Random memory could be DMA read\n");
731 }
732 
733 /*
734  * Map a single buffer of the indicated size for DMA in streaming mode.  The
735  * physical address to use is returned.
736  *
737  * Once the device is given the dma address, the device owns this memory until
738  * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
739  */
swiotlb_map_page(struct device * dev,struct page * page,unsigned long offset,size_t size,enum dma_data_direction dir,struct dma_attrs * attrs)740 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
741 			    unsigned long offset, size_t size,
742 			    enum dma_data_direction dir,
743 			    struct dma_attrs *attrs)
744 {
745 	phys_addr_t map, phys = page_to_phys(page) + offset;
746 	dma_addr_t dev_addr = phys_to_dma(dev, phys);
747 
748 	BUG_ON(dir == DMA_NONE);
749 	/*
750 	 * If the address happens to be in the device's DMA window,
751 	 * we can safely return the device addr and not worry about bounce
752 	 * buffering it.
753 	 */
754 	if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
755 		return dev_addr;
756 
757 	trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
758 
759 	/* Oh well, have to allocate and map a bounce buffer. */
760 	map = map_single(dev, phys, size, dir);
761 	if (map == SWIOTLB_MAP_ERROR) {
762 		swiotlb_full(dev, size, dir, 1);
763 		return phys_to_dma(dev, io_tlb_overflow_buffer);
764 	}
765 
766 	dev_addr = phys_to_dma(dev, map);
767 
768 	/* Ensure that the address returned is DMA'ble */
769 	if (!dma_capable(dev, dev_addr, size)) {
770 		swiotlb_tbl_unmap_single(dev, map, size, dir);
771 		return phys_to_dma(dev, io_tlb_overflow_buffer);
772 	}
773 
774 	return dev_addr;
775 }
776 EXPORT_SYMBOL_GPL(swiotlb_map_page);
777 
778 /*
779  * Unmap a single streaming mode DMA translation.  The dma_addr and size must
780  * match what was provided for in a previous swiotlb_map_page call.  All
781  * other usages are undefined.
782  *
783  * After this call, reads by the cpu to the buffer are guaranteed to see
784  * whatever the device wrote there.
785  */
unmap_single(struct device * hwdev,dma_addr_t dev_addr,size_t size,enum dma_data_direction dir)786 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
787 			 size_t size, enum dma_data_direction dir)
788 {
789 	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
790 
791 	BUG_ON(dir == DMA_NONE);
792 
793 	if (is_swiotlb_buffer(paddr)) {
794 		swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
795 		return;
796 	}
797 
798 	if (dir != DMA_FROM_DEVICE)
799 		return;
800 
801 	/*
802 	 * phys_to_virt doesn't work with hihgmem page but we could
803 	 * call dma_mark_clean() with hihgmem page here. However, we
804 	 * are fine since dma_mark_clean() is null on POWERPC. We can
805 	 * make dma_mark_clean() take a physical address if necessary.
806 	 */
807 	dma_mark_clean(phys_to_virt(paddr), size);
808 }
809 
swiotlb_unmap_page(struct device * hwdev,dma_addr_t dev_addr,size_t size,enum dma_data_direction dir,struct dma_attrs * attrs)810 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
811 			size_t size, enum dma_data_direction dir,
812 			struct dma_attrs *attrs)
813 {
814 	unmap_single(hwdev, dev_addr, size, dir);
815 }
816 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
817 
818 /*
819  * Make physical memory consistent for a single streaming mode DMA translation
820  * after a transfer.
821  *
822  * If you perform a swiotlb_map_page() but wish to interrogate the buffer
823  * using the cpu, yet do not wish to teardown the dma mapping, you must
824  * call this function before doing so.  At the next point you give the dma
825  * address back to the card, you must first perform a
826  * swiotlb_dma_sync_for_device, and then the device again owns the buffer
827  */
828 static void
swiotlb_sync_single(struct device * hwdev,dma_addr_t dev_addr,size_t size,enum dma_data_direction dir,enum dma_sync_target target)829 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
830 		    size_t size, enum dma_data_direction dir,
831 		    enum dma_sync_target target)
832 {
833 	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
834 
835 	BUG_ON(dir == DMA_NONE);
836 
837 	if (is_swiotlb_buffer(paddr)) {
838 		swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
839 		return;
840 	}
841 
842 	if (dir != DMA_FROM_DEVICE)
843 		return;
844 
845 	dma_mark_clean(phys_to_virt(paddr), size);
846 }
847 
848 void
swiotlb_sync_single_for_cpu(struct device * hwdev,dma_addr_t dev_addr,size_t size,enum dma_data_direction dir)849 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
850 			    size_t size, enum dma_data_direction dir)
851 {
852 	swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
853 }
854 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
855 
856 void
swiotlb_sync_single_for_device(struct device * hwdev,dma_addr_t dev_addr,size_t size,enum dma_data_direction dir)857 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
858 			       size_t size, enum dma_data_direction dir)
859 {
860 	swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
861 }
862 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
863 
864 /*
865  * Map a set of buffers described by scatterlist in streaming mode for DMA.
866  * This is the scatter-gather version of the above swiotlb_map_page
867  * interface.  Here the scatter gather list elements are each tagged with the
868  * appropriate dma address and length.  They are obtained via
869  * sg_dma_{address,length}(SG).
870  *
871  * NOTE: An implementation may be able to use a smaller number of
872  *       DMA address/length pairs than there are SG table elements.
873  *       (for example via virtual mapping capabilities)
874  *       The routine returns the number of addr/length pairs actually
875  *       used, at most nents.
876  *
877  * Device ownership issues as mentioned above for swiotlb_map_page are the
878  * same here.
879  */
880 int
swiotlb_map_sg_attrs(struct device * hwdev,struct scatterlist * sgl,int nelems,enum dma_data_direction dir,struct dma_attrs * attrs)881 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
882 		     enum dma_data_direction dir, struct dma_attrs *attrs)
883 {
884 	struct scatterlist *sg;
885 	int i;
886 
887 	BUG_ON(dir == DMA_NONE);
888 
889 	for_each_sg(sgl, sg, nelems, i) {
890 		phys_addr_t paddr = sg_phys(sg);
891 		dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
892 
893 		if (swiotlb_force ||
894 		    !dma_capable(hwdev, dev_addr, sg->length)) {
895 			phys_addr_t map = map_single(hwdev, sg_phys(sg),
896 						     sg->length, dir);
897 			if (map == SWIOTLB_MAP_ERROR) {
898 				/* Don't panic here, we expect map_sg users
899 				   to do proper error handling. */
900 				swiotlb_full(hwdev, sg->length, dir, 0);
901 				swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
902 						       attrs);
903 				sg_dma_len(sgl) = 0;
904 				return 0;
905 			}
906 			sg->dma_address = phys_to_dma(hwdev, map);
907 		} else
908 			sg->dma_address = dev_addr;
909 		sg_dma_len(sg) = sg->length;
910 	}
911 	return nelems;
912 }
913 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
914 
915 int
swiotlb_map_sg(struct device * hwdev,struct scatterlist * sgl,int nelems,enum dma_data_direction dir)916 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
917 	       enum dma_data_direction dir)
918 {
919 	return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
920 }
921 EXPORT_SYMBOL(swiotlb_map_sg);
922 
923 /*
924  * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
925  * concerning calls here are the same as for swiotlb_unmap_page() above.
926  */
927 void
swiotlb_unmap_sg_attrs(struct device * hwdev,struct scatterlist * sgl,int nelems,enum dma_data_direction dir,struct dma_attrs * attrs)928 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
929 		       int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
930 {
931 	struct scatterlist *sg;
932 	int i;
933 
934 	BUG_ON(dir == DMA_NONE);
935 
936 	for_each_sg(sgl, sg, nelems, i)
937 		unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir);
938 
939 }
940 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
941 
942 void
swiotlb_unmap_sg(struct device * hwdev,struct scatterlist * sgl,int nelems,enum dma_data_direction dir)943 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
944 		 enum dma_data_direction dir)
945 {
946 	return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
947 }
948 EXPORT_SYMBOL(swiotlb_unmap_sg);
949 
950 /*
951  * Make physical memory consistent for a set of streaming mode DMA translations
952  * after a transfer.
953  *
954  * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
955  * and usage.
956  */
957 static void
swiotlb_sync_sg(struct device * hwdev,struct scatterlist * sgl,int nelems,enum dma_data_direction dir,enum dma_sync_target target)958 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
959 		int nelems, enum dma_data_direction dir,
960 		enum dma_sync_target target)
961 {
962 	struct scatterlist *sg;
963 	int i;
964 
965 	for_each_sg(sgl, sg, nelems, i)
966 		swiotlb_sync_single(hwdev, sg->dma_address,
967 				    sg_dma_len(sg), dir, target);
968 }
969 
970 void
swiotlb_sync_sg_for_cpu(struct device * hwdev,struct scatterlist * sg,int nelems,enum dma_data_direction dir)971 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
972 			int nelems, enum dma_data_direction dir)
973 {
974 	swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
975 }
976 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
977 
978 void
swiotlb_sync_sg_for_device(struct device * hwdev,struct scatterlist * sg,int nelems,enum dma_data_direction dir)979 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
980 			   int nelems, enum dma_data_direction dir)
981 {
982 	swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
983 }
984 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
985 
986 int
swiotlb_dma_mapping_error(struct device * hwdev,dma_addr_t dma_addr)987 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
988 {
989 	return (dma_addr == phys_to_dma(hwdev, io_tlb_overflow_buffer));
990 }
991 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
992 
993 /*
994  * Return whether the given device DMA address mask can be supported
995  * properly.  For example, if your device can only drive the low 24-bits
996  * during bus mastering, then you would pass 0x00ffffff as the mask to
997  * this function.
998  */
999 int
swiotlb_dma_supported(struct device * hwdev,u64 mask)1000 swiotlb_dma_supported(struct device *hwdev, u64 mask)
1001 {
1002 	return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
1003 }
1004 EXPORT_SYMBOL(swiotlb_dma_supported);
1005