Linux Audio

Check our new training course

Loading...
v3.15
  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#include <linux/cache.h>
 21#include <linux/dma-mapping.h>
 22#include <linux/mm.h>
 23#include <linux/export.h>
 24#include <linux/spinlock.h>
 25#include <linux/string.h>
 26#include <linux/swiotlb.h>
 27#include <linux/pfn.h>
 28#include <linux/types.h>
 29#include <linux/ctype.h>
 30#include <linux/highmem.h>
 31#include <linux/gfp.h>
 
 
 
 32
 33#include <asm/io.h>
 34#include <asm/dma.h>
 35#include <asm/scatterlist.h>
 36
 37#include <linux/init.h>
 38#include <linux/bootmem.h>
 39#include <linux/iommu-helper.h>
 40
 41#define CREATE_TRACE_POINTS
 42#include <trace/events/swiotlb.h>
 43
 44#define OFFSET(val,align) ((unsigned long)	\
 45	                   ( (val) & ( (align) - 1)))
 46
 47#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
 48
 49/*
 50 * Minimum IO TLB size to bother booting with.  Systems with mainly
 51 * 64bit capable cards will only lightly use the swiotlb.  If we can't
 52 * allocate a contiguous 1MB, we're probably in trouble anyway.
 53 */
 54#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
 55
 56int swiotlb_force;
 57
 58/*
 59 * Used to do a quick range check in swiotlb_tbl_unmap_single and
 60 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
 61 * API.
 62 */
 63static phys_addr_t io_tlb_start, io_tlb_end;
 64
 65/*
 66 * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
 67 * io_tlb_end.  This is command line adjustable via setup_io_tlb_npages.
 68 */
 69static unsigned long io_tlb_nslabs;
 70
 71/*
 72 * When the IOMMU overflows we return a fallback buffer. This sets the size.
 73 */
 74static unsigned long io_tlb_overflow = 32*1024;
 75
 76static phys_addr_t io_tlb_overflow_buffer;
 77
 78/*
 79 * This is a free list describing the number of free entries available from
 80 * each index
 81 */
 82static unsigned int *io_tlb_list;
 83static unsigned int io_tlb_index;
 84
 85/*
 
 
 
 
 
 
 86 * We need to save away the original address corresponding to a mapped entry
 87 * for the sync operations.
 88 */
 
 89static phys_addr_t *io_tlb_orig_addr;
 90
 91/*
 92 * Protect the above data structures in the map and unmap calls
 93 */
 94static DEFINE_SPINLOCK(io_tlb_lock);
 95
 96static int late_alloc;
 97
 98static int __init
 99setup_io_tlb_npages(char *str)
100{
101	if (isdigit(*str)) {
102		io_tlb_nslabs = simple_strtoul(str, &str, 0);
103		/* avoid tail segment of size < IO_TLB_SEGSIZE */
104		io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
105	}
106	if (*str == ',')
107		++str;
108	if (!strcmp(str, "force"))
109		swiotlb_force = 1;
 
 
 
 
110
111	return 0;
112}
113early_param("swiotlb", setup_io_tlb_npages);
114/* make io_tlb_overflow tunable too? */
115
116unsigned long swiotlb_nr_tbl(void)
117{
118	return io_tlb_nslabs;
119}
120EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
121
 
 
 
 
 
 
 
 
 
 
 
 
 
 
122/* default to 64MB */
123#define IO_TLB_DEFAULT_SIZE (64UL<<20)
124unsigned long swiotlb_size_or_default(void)
125{
126	unsigned long size;
127
128	size = io_tlb_nslabs << IO_TLB_SHIFT;
129
130	return size ? size : (IO_TLB_DEFAULT_SIZE);
131}
132
133/* Note that this doesn't work with highmem page */
134static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
135				      volatile void *address)
136{
137	return phys_to_dma(hwdev, virt_to_phys(address));
138}
139
140static bool no_iotlb_memory;
141
142void swiotlb_print_info(void)
143{
144	unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
145	unsigned char *vstart, *vend;
146
147	if (no_iotlb_memory) {
148		pr_warn("software IO TLB: No low mem\n");
149		return;
150	}
151
152	vstart = phys_to_virt(io_tlb_start);
153	vend = phys_to_virt(io_tlb_end);
154
155	printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
156	       (unsigned long long)io_tlb_start,
157	       (unsigned long long)io_tlb_end,
158	       bytes >> 20, vstart, vend - 1);
159}
160
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
161int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
162{
163	void *v_overflow_buffer;
164	unsigned long i, bytes;
165
166	bytes = nslabs << IO_TLB_SHIFT;
167
168	io_tlb_nslabs = nslabs;
169	io_tlb_start = __pa(tlb);
170	io_tlb_end = io_tlb_start + bytes;
171
172	/*
173	 * Get the overflow emergency buffer
174	 */
175	v_overflow_buffer = memblock_virt_alloc_low_nopanic(
176						PAGE_ALIGN(io_tlb_overflow),
177						PAGE_SIZE);
178	if (!v_overflow_buffer)
179		return -ENOMEM;
180
181	io_tlb_overflow_buffer = __pa(v_overflow_buffer);
182
183	/*
184	 * Allocate and initialize the free list array.  This array is used
185	 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
186	 * between io_tlb_start and io_tlb_end.
187	 */
188	io_tlb_list = memblock_virt_alloc(
189				PAGE_ALIGN(io_tlb_nslabs * sizeof(int)),
190				PAGE_SIZE);
191	for (i = 0; i < io_tlb_nslabs; i++)
192 		io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
193	io_tlb_index = 0;
194	io_tlb_orig_addr = memblock_virt_alloc(
195				PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)),
196				PAGE_SIZE);
 
 
 
 
 
197
198	if (verbose)
199		swiotlb_print_info();
200
 
201	return 0;
202}
203
204/*
205 * Statically reserve bounce buffer space and initialize bounce buffer data
206 * structures for the software IO TLB used to implement the DMA API.
207 */
208void  __init
209swiotlb_init(int verbose)
210{
211	size_t default_size = IO_TLB_DEFAULT_SIZE;
212	unsigned char *vstart;
213	unsigned long bytes;
214
215	if (!io_tlb_nslabs) {
216		io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
217		io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
218	}
219
220	bytes = io_tlb_nslabs << IO_TLB_SHIFT;
221
222	/* Get IO TLB memory from the low pages */
223	vstart = memblock_virt_alloc_low_nopanic(PAGE_ALIGN(bytes), PAGE_SIZE);
224	if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose))
225		return;
226
227	if (io_tlb_start)
228		memblock_free_early(io_tlb_start,
229				    PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
230	pr_warn("Cannot allocate SWIOTLB buffer");
231	no_iotlb_memory = true;
232}
233
234/*
235 * Systems with larger DMA zones (those that don't support ISA) can
236 * initialize the swiotlb later using the slab allocator if needed.
237 * This should be just like above, but with some error catching.
238 */
239int
240swiotlb_late_init_with_default_size(size_t default_size)
241{
242	unsigned long bytes, req_nslabs = io_tlb_nslabs;
243	unsigned char *vstart = NULL;
244	unsigned int order;
245	int rc = 0;
246
247	if (!io_tlb_nslabs) {
248		io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
249		io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
250	}
251
252	/*
253	 * Get IO TLB memory from the low pages
254	 */
255	order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
256	io_tlb_nslabs = SLABS_PER_PAGE << order;
257	bytes = io_tlb_nslabs << IO_TLB_SHIFT;
258
259	while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
260		vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
261						  order);
262		if (vstart)
263			break;
264		order--;
265	}
266
267	if (!vstart) {
268		io_tlb_nslabs = req_nslabs;
269		return -ENOMEM;
270	}
271	if (order != get_order(bytes)) {
272		printk(KERN_WARNING "Warning: only able to allocate %ld MB "
273		       "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
274		io_tlb_nslabs = SLABS_PER_PAGE << order;
275	}
276	rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
277	if (rc)
278		free_pages((unsigned long)vstart, order);
 
279	return rc;
280}
281
282int
283swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
284{
285	unsigned long i, bytes;
286	unsigned char *v_overflow_buffer;
287
288	bytes = nslabs << IO_TLB_SHIFT;
289
290	io_tlb_nslabs = nslabs;
291	io_tlb_start = virt_to_phys(tlb);
292	io_tlb_end = io_tlb_start + bytes;
293
 
294	memset(tlb, 0, bytes);
295
296	/*
297	 * Get the overflow emergency buffer
298	 */
299	v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
300						     get_order(io_tlb_overflow));
301	if (!v_overflow_buffer)
302		goto cleanup2;
303
 
 
 
304	io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
305
306	/*
307	 * Allocate and initialize the free list array.  This array is used
308	 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
309	 * between io_tlb_start and io_tlb_end.
310	 */
311	io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
312	                              get_order(io_tlb_nslabs * sizeof(int)));
313	if (!io_tlb_list)
314		goto cleanup3;
315
316	for (i = 0; i < io_tlb_nslabs; i++)
317 		io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
318	io_tlb_index = 0;
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	memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
 
 
 
 
328
329	swiotlb_print_info();
330
331	late_alloc = 1;
332
 
 
333	return 0;
334
335cleanup4:
336	free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
337	                                                 sizeof(int)));
338	io_tlb_list = NULL;
339cleanup3:
340	free_pages((unsigned long)v_overflow_buffer,
341		   get_order(io_tlb_overflow));
342	io_tlb_overflow_buffer = 0;
343cleanup2:
344	io_tlb_end = 0;
345	io_tlb_start = 0;
346	io_tlb_nslabs = 0;
 
347	return -ENOMEM;
348}
349
350void __init swiotlb_free(void)
351{
352	if (!io_tlb_orig_addr)
353		return;
354
355	if (late_alloc) {
356		free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer),
357			   get_order(io_tlb_overflow));
358		free_pages((unsigned long)io_tlb_orig_addr,
359			   get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
360		free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
361								 sizeof(int)));
362		free_pages((unsigned long)phys_to_virt(io_tlb_start),
363			   get_order(io_tlb_nslabs << IO_TLB_SHIFT));
364	} else {
365		memblock_free_late(io_tlb_overflow_buffer,
366				   PAGE_ALIGN(io_tlb_overflow));
367		memblock_free_late(__pa(io_tlb_orig_addr),
368				   PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
369		memblock_free_late(__pa(io_tlb_list),
370				   PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
371		memblock_free_late(io_tlb_start,
372				   PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
373	}
374	io_tlb_nslabs = 0;
 
375}
376
377static int is_swiotlb_buffer(phys_addr_t paddr)
378{
379	return paddr >= io_tlb_start && paddr < io_tlb_end;
380}
381
382/*
383 * Bounce: copy the swiotlb buffer back to the original dma location
384 */
385static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
386			   size_t size, enum dma_data_direction dir)
387{
388	unsigned long pfn = PFN_DOWN(orig_addr);
389	unsigned char *vaddr = phys_to_virt(tlb_addr);
390
391	if (PageHighMem(pfn_to_page(pfn))) {
392		/* The buffer does not have a mapping.  Map it in and copy */
393		unsigned int offset = orig_addr & ~PAGE_MASK;
394		char *buffer;
395		unsigned int sz = 0;
396		unsigned long flags;
397
398		while (size) {
399			sz = min_t(size_t, PAGE_SIZE - offset, size);
400
401			local_irq_save(flags);
402			buffer = kmap_atomic(pfn_to_page(pfn));
403			if (dir == DMA_TO_DEVICE)
404				memcpy(vaddr, buffer + offset, sz);
405			else
406				memcpy(buffer + offset, vaddr, sz);
407			kunmap_atomic(buffer);
408			local_irq_restore(flags);
409
410			size -= sz;
411			pfn++;
412			vaddr += sz;
413			offset = 0;
414		}
415	} else if (dir == DMA_TO_DEVICE) {
416		memcpy(vaddr, phys_to_virt(orig_addr), size);
417	} else {
418		memcpy(phys_to_virt(orig_addr), vaddr, size);
419	}
420}
421
422phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
423				   dma_addr_t tbl_dma_addr,
424				   phys_addr_t orig_addr, size_t size,
425				   enum dma_data_direction dir)
 
426{
427	unsigned long flags;
428	phys_addr_t tlb_addr;
429	unsigned int nslots, stride, index, wrap;
430	int i;
431	unsigned long mask;
432	unsigned long offset_slots;
433	unsigned long max_slots;
434
435	if (no_iotlb_memory)
436		panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
437
 
 
 
 
438	mask = dma_get_seg_boundary(hwdev);
439
440	tbl_dma_addr &= mask;
441
442	offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
443
444	/*
445 	 * Carefully handle integer overflow which can occur when mask == ~0UL.
446 	 */
447	max_slots = mask + 1
448		    ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
449		    : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
450
451	/*
452	 * For mappings greater than a page, we limit the stride (and
453	 * hence alignment) to a page size.
454	 */
455	nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
456	if (size > PAGE_SIZE)
457		stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
458	else
459		stride = 1;
460
461	BUG_ON(!nslots);
462
463	/*
464	 * Find suitable number of IO TLB entries size that will fit this
465	 * request and allocate a buffer from that IO TLB pool.
466	 */
467	spin_lock_irqsave(&io_tlb_lock, flags);
468	index = ALIGN(io_tlb_index, stride);
469	if (index >= io_tlb_nslabs)
470		index = 0;
471	wrap = index;
472
473	do {
474		while (iommu_is_span_boundary(index, nslots, offset_slots,
475					      max_slots)) {
476			index += stride;
477			if (index >= io_tlb_nslabs)
478				index = 0;
479			if (index == wrap)
480				goto not_found;
481		}
482
483		/*
484		 * If we find a slot that indicates we have 'nslots' number of
485		 * contiguous buffers, we allocate the buffers from that slot
486		 * and mark the entries as '0' indicating unavailable.
487		 */
488		if (io_tlb_list[index] >= nslots) {
489			int count = 0;
490
491			for (i = index; i < (int) (index + nslots); i++)
492				io_tlb_list[i] = 0;
493			for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
494				io_tlb_list[i] = ++count;
495			tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
496
497			/*
498			 * Update the indices to avoid searching in the next
499			 * round.
500			 */
501			io_tlb_index = ((index + nslots) < io_tlb_nslabs
502					? (index + nslots) : 0);
503
504			goto found;
505		}
506		index += stride;
507		if (index >= io_tlb_nslabs)
508			index = 0;
509	} while (index != wrap);
510
511not_found:
512	spin_unlock_irqrestore(&io_tlb_lock, flags);
513	if (printk_ratelimit())
514		dev_warn(hwdev, "swiotlb buffer is full (sz: %zd bytes)\n", size);
515	return SWIOTLB_MAP_ERROR;
516found:
517	spin_unlock_irqrestore(&io_tlb_lock, flags);
518
519	/*
520	 * Save away the mapping from the original address to the DMA address.
521	 * This is needed when we sync the memory.  Then we sync the buffer if
522	 * needed.
523	 */
524	for (i = 0; i < nslots; i++)
525		io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
526	if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
 
527		swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE);
528
529	return tlb_addr;
530}
531EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
532
533/*
534 * Allocates bounce buffer and returns its kernel virtual address.
535 */
536
537phys_addr_t map_single(struct device *hwdev, phys_addr_t phys, size_t size,
538		       enum dma_data_direction dir)
539{
540	dma_addr_t start_dma_addr = phys_to_dma(hwdev, io_tlb_start);
 
 
 
 
 
 
 
541
542	return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
 
 
543}
544
545/*
546 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
547 */
548void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
549			      size_t size, enum dma_data_direction dir)
 
550{
551	unsigned long flags;
552	int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
553	int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
554	phys_addr_t orig_addr = io_tlb_orig_addr[index];
555
556	/*
557	 * First, sync the memory before unmapping the entry
558	 */
559	if (orig_addr && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
 
 
560		swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
561
562	/*
563	 * Return the buffer to the free list by setting the corresponding
564	 * entries to indicate the number of contiguous entries available.
565	 * While returning the entries to the free list, we merge the entries
566	 * with slots below and above the pool being returned.
567	 */
568	spin_lock_irqsave(&io_tlb_lock, flags);
569	{
570		count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
571			 io_tlb_list[index + nslots] : 0);
572		/*
573		 * Step 1: return the slots to the free list, merging the
574		 * slots with superceeding slots
575		 */
576		for (i = index + nslots - 1; i >= index; i--)
577			io_tlb_list[i] = ++count;
 
 
578		/*
579		 * Step 2: merge the returned slots with the preceding slots,
580		 * if available (non zero)
581		 */
582		for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
583			io_tlb_list[i] = ++count;
584	}
585	spin_unlock_irqrestore(&io_tlb_lock, flags);
586}
587EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
588
589void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
590			     size_t size, enum dma_data_direction dir,
591			     enum dma_sync_target target)
592{
593	int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
594	phys_addr_t orig_addr = io_tlb_orig_addr[index];
595
 
 
596	orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
597
598	switch (target) {
599	case SYNC_FOR_CPU:
600		if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
601			swiotlb_bounce(orig_addr, tlb_addr,
602				       size, DMA_FROM_DEVICE);
603		else
604			BUG_ON(dir != DMA_TO_DEVICE);
605		break;
606	case SYNC_FOR_DEVICE:
607		if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
608			swiotlb_bounce(orig_addr, tlb_addr,
609				       size, DMA_TO_DEVICE);
610		else
611			BUG_ON(dir != DMA_FROM_DEVICE);
612		break;
613	default:
614		BUG();
615	}
616}
617EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
618
619void *
620swiotlb_alloc_coherent(struct device *hwdev, size_t size,
621		       dma_addr_t *dma_handle, gfp_t flags)
622{
623	dma_addr_t dev_addr;
624	void *ret;
625	int order = get_order(size);
626	u64 dma_mask = DMA_BIT_MASK(32);
627
628	if (hwdev && hwdev->coherent_dma_mask)
629		dma_mask = hwdev->coherent_dma_mask;
630
631	ret = (void *)__get_free_pages(flags, order);
632	if (ret) {
633		dev_addr = swiotlb_virt_to_bus(hwdev, ret);
634		if (dev_addr + size - 1 > dma_mask) {
635			/*
636			 * The allocated memory isn't reachable by the device.
637			 */
638			free_pages((unsigned long) ret, order);
639			ret = NULL;
640		}
641	}
642	if (!ret) {
643		/*
644		 * We are either out of memory or the device can't DMA to
645		 * GFP_DMA memory; fall back on map_single(), which
646		 * will grab memory from the lowest available address range.
647		 */
648		phys_addr_t paddr = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
649		if (paddr == SWIOTLB_MAP_ERROR)
650			return NULL;
651
652		ret = phys_to_virt(paddr);
653		dev_addr = phys_to_dma(hwdev, paddr);
654
655		/* Confirm address can be DMA'd by device */
656		if (dev_addr + size - 1 > dma_mask) {
657			printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
658			       (unsigned long long)dma_mask,
659			       (unsigned long long)dev_addr);
660
661			/* DMA_TO_DEVICE to avoid memcpy in unmap_single */
662			swiotlb_tbl_unmap_single(hwdev, paddr,
663						 size, DMA_TO_DEVICE);
664			return NULL;
665		}
666	}
667
668	*dma_handle = dev_addr;
669	memset(ret, 0, size);
 
 
670
671	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
672}
673EXPORT_SYMBOL(swiotlb_alloc_coherent);
674
675void
676swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
677		      dma_addr_t dev_addr)
678{
679	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
680
681	WARN_ON(irqs_disabled());
682	if (!is_swiotlb_buffer(paddr))
683		free_pages((unsigned long)vaddr, get_order(size));
684	else
685		/* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
686		swiotlb_tbl_unmap_single(hwdev, paddr, size, DMA_TO_DEVICE);
 
 
 
 
 
 
687}
688EXPORT_SYMBOL(swiotlb_free_coherent);
689
690static void
691swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
692	     int do_panic)
693{
 
 
 
694	/*
695	 * Ran out of IOMMU space for this operation. This is very bad.
696	 * Unfortunately the drivers cannot handle this operation properly.
697	 * unless they check for dma_mapping_error (most don't)
698	 * When the mapping is small enough return a static buffer to limit
699	 * the damage, or panic when the transfer is too big.
700	 */
701	printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
702	       "device %s\n", size, dev ? dev_name(dev) : "?");
703
704	if (size <= io_tlb_overflow || !do_panic)
705		return;
706
707	if (dir == DMA_BIDIRECTIONAL)
708		panic("DMA: Random memory could be DMA accessed\n");
709	if (dir == DMA_FROM_DEVICE)
710		panic("DMA: Random memory could be DMA written\n");
711	if (dir == DMA_TO_DEVICE)
712		panic("DMA: Random memory could be DMA read\n");
713}
714
715/*
716 * Map a single buffer of the indicated size for DMA in streaming mode.  The
717 * physical address to use is returned.
718 *
719 * Once the device is given the dma address, the device owns this memory until
720 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
721 */
722dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
723			    unsigned long offset, size_t size,
724			    enum dma_data_direction dir,
725			    struct dma_attrs *attrs)
726{
727	phys_addr_t map, phys = page_to_phys(page) + offset;
728	dma_addr_t dev_addr = phys_to_dma(dev, phys);
729
730	BUG_ON(dir == DMA_NONE);
731	/*
732	 * If the address happens to be in the device's DMA window,
733	 * we can safely return the device addr and not worry about bounce
734	 * buffering it.
735	 */
736	if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
737		return dev_addr;
738
739	trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
740
741	/* Oh well, have to allocate and map a bounce buffer. */
742	map = map_single(dev, phys, size, dir);
743	if (map == SWIOTLB_MAP_ERROR) {
744		swiotlb_full(dev, size, dir, 1);
745		return phys_to_dma(dev, io_tlb_overflow_buffer);
746	}
747
748	dev_addr = phys_to_dma(dev, map);
749
750	/* Ensure that the address returned is DMA'ble */
751	if (!dma_capable(dev, dev_addr, size)) {
752		swiotlb_tbl_unmap_single(dev, map, size, dir);
753		return phys_to_dma(dev, io_tlb_overflow_buffer);
754	}
 
755
756	return dev_addr;
757}
758EXPORT_SYMBOL_GPL(swiotlb_map_page);
759
760/*
761 * Unmap a single streaming mode DMA translation.  The dma_addr and size must
762 * match what was provided for in a previous swiotlb_map_page call.  All
763 * other usages are undefined.
764 *
765 * After this call, reads by the cpu to the buffer are guaranteed to see
766 * whatever the device wrote there.
767 */
768static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
769			 size_t size, enum dma_data_direction dir)
 
770{
771	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
772
773	BUG_ON(dir == DMA_NONE);
774
775	if (is_swiotlb_buffer(paddr)) {
776		swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
777		return;
778	}
779
780	if (dir != DMA_FROM_DEVICE)
781		return;
782
783	/*
784	 * phys_to_virt doesn't work with hihgmem page but we could
785	 * call dma_mark_clean() with hihgmem page here. However, we
786	 * are fine since dma_mark_clean() is null on POWERPC. We can
787	 * make dma_mark_clean() take a physical address if necessary.
788	 */
789	dma_mark_clean(phys_to_virt(paddr), size);
790}
791
792void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
793			size_t size, enum dma_data_direction dir,
794			struct dma_attrs *attrs)
795{
796	unmap_single(hwdev, dev_addr, size, dir);
797}
798EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
799
800/*
801 * Make physical memory consistent for a single streaming mode DMA translation
802 * after a transfer.
803 *
804 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
805 * using the cpu, yet do not wish to teardown the dma mapping, you must
806 * call this function before doing so.  At the next point you give the dma
807 * address back to the card, you must first perform a
808 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
809 */
810static void
811swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
812		    size_t size, enum dma_data_direction dir,
813		    enum dma_sync_target target)
814{
815	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
816
817	BUG_ON(dir == DMA_NONE);
818
819	if (is_swiotlb_buffer(paddr)) {
820		swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
821		return;
822	}
823
824	if (dir != DMA_FROM_DEVICE)
825		return;
826
827	dma_mark_clean(phys_to_virt(paddr), size);
828}
829
830void
831swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
832			    size_t size, enum dma_data_direction dir)
833{
834	swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
835}
836EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
837
838void
839swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
840			       size_t size, enum dma_data_direction dir)
841{
842	swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
843}
844EXPORT_SYMBOL(swiotlb_sync_single_for_device);
845
846/*
847 * Map a set of buffers described by scatterlist in streaming mode for DMA.
848 * This is the scatter-gather version of the above swiotlb_map_page
849 * interface.  Here the scatter gather list elements are each tagged with the
850 * appropriate dma address and length.  They are obtained via
851 * sg_dma_{address,length}(SG).
852 *
853 * NOTE: An implementation may be able to use a smaller number of
854 *       DMA address/length pairs than there are SG table elements.
855 *       (for example via virtual mapping capabilities)
856 *       The routine returns the number of addr/length pairs actually
857 *       used, at most nents.
858 *
859 * Device ownership issues as mentioned above for swiotlb_map_page are the
860 * same here.
861 */
862int
863swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
864		     enum dma_data_direction dir, struct dma_attrs *attrs)
865{
866	struct scatterlist *sg;
867	int i;
868
869	BUG_ON(dir == DMA_NONE);
870
871	for_each_sg(sgl, sg, nelems, i) {
872		phys_addr_t paddr = sg_phys(sg);
873		dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
874
875		if (swiotlb_force ||
876		    !dma_capable(hwdev, dev_addr, sg->length)) {
877			phys_addr_t map = map_single(hwdev, sg_phys(sg),
878						     sg->length, dir);
879			if (map == SWIOTLB_MAP_ERROR) {
880				/* Don't panic here, we expect map_sg users
881				   to do proper error handling. */
882				swiotlb_full(hwdev, sg->length, dir, 0);
 
883				swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
884						       attrs);
885				sg_dma_len(sgl) = 0;
886				return 0;
887			}
888			sg->dma_address = phys_to_dma(hwdev, map);
889		} else
890			sg->dma_address = dev_addr;
891		sg_dma_len(sg) = sg->length;
892	}
893	return nelems;
894}
895EXPORT_SYMBOL(swiotlb_map_sg_attrs);
896
897int
898swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
899	       enum dma_data_direction dir)
900{
901	return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
902}
903EXPORT_SYMBOL(swiotlb_map_sg);
904
905/*
906 * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
907 * concerning calls here are the same as for swiotlb_unmap_page() above.
908 */
909void
910swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
911		       int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
 
912{
913	struct scatterlist *sg;
914	int i;
915
916	BUG_ON(dir == DMA_NONE);
917
918	for_each_sg(sgl, sg, nelems, i)
919		unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir);
920
921}
922EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
923
924void
925swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
926		 enum dma_data_direction dir)
927{
928	return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
929}
930EXPORT_SYMBOL(swiotlb_unmap_sg);
931
932/*
933 * Make physical memory consistent for a set of streaming mode DMA translations
934 * after a transfer.
935 *
936 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
937 * and usage.
938 */
939static void
940swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
941		int nelems, enum dma_data_direction dir,
942		enum dma_sync_target target)
943{
944	struct scatterlist *sg;
945	int i;
946
947	for_each_sg(sgl, sg, nelems, i)
948		swiotlb_sync_single(hwdev, sg->dma_address,
949				    sg_dma_len(sg), dir, target);
950}
951
952void
953swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
954			int nelems, enum dma_data_direction dir)
955{
956	swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
957}
958EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
959
960void
961swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
962			   int nelems, enum dma_data_direction dir)
963{
964	swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
965}
966EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
967
968int
969swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
970{
971	return (dma_addr == phys_to_dma(hwdev, io_tlb_overflow_buffer));
972}
973EXPORT_SYMBOL(swiotlb_dma_mapping_error);
974
975/*
976 * Return whether the given device DMA address mask can be supported
977 * properly.  For example, if your device can only drive the low 24-bits
978 * during bus mastering, then you would pass 0x00ffffff as the mask to
979 * this function.
980 */
981int
982swiotlb_dma_supported(struct device *hwdev, u64 mask)
983{
984	return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
985}
986EXPORT_SYMBOL(swiotlb_dma_supported);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
v4.17
   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#include <linux/cache.h>
  21#include <linux/dma-direct.h>
  22#include <linux/mm.h>
  23#include <linux/export.h>
  24#include <linux/spinlock.h>
  25#include <linux/string.h>
  26#include <linux/swiotlb.h>
  27#include <linux/pfn.h>
  28#include <linux/types.h>
  29#include <linux/ctype.h>
  30#include <linux/highmem.h>
  31#include <linux/gfp.h>
  32#include <linux/scatterlist.h>
  33#include <linux/mem_encrypt.h>
  34#include <linux/set_memory.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
  58enum 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 */
  65static 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 */
  71static unsigned long io_tlb_nslabs;
  72
  73/*
  74 * When the IOMMU overflows we return a fallback buffer. This sets the size.
  75 */
  76static unsigned long io_tlb_overflow = 32*1024;
  77
  78static 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 */
  84static unsigned int *io_tlb_list;
  85static unsigned int io_tlb_index;
  86
  87/*
  88 * Max segment that we can provide which (if pages are contingous) will
  89 * not be bounced (unless SWIOTLB_FORCE is set).
  90 */
  91unsigned int max_segment;
  92
  93/*
  94 * We need to save away the original address corresponding to a mapped entry
  95 * for the sync operations.
  96 */
  97#define INVALID_PHYS_ADDR (~(phys_addr_t)0)
  98static phys_addr_t *io_tlb_orig_addr;
  99
 100/*
 101 * Protect the above data structures in the map and unmap calls
 102 */
 103static DEFINE_SPINLOCK(io_tlb_lock);
 104
 105static int late_alloc;
 106
 107static int __init
 108setup_io_tlb_npages(char *str)
 109{
 110	if (isdigit(*str)) {
 111		io_tlb_nslabs = simple_strtoul(str, &str, 0);
 112		/* avoid tail segment of size < IO_TLB_SEGSIZE */
 113		io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
 114	}
 115	if (*str == ',')
 116		++str;
 117	if (!strcmp(str, "force")) {
 118		swiotlb_force = SWIOTLB_FORCE;
 119	} else if (!strcmp(str, "noforce")) {
 120		swiotlb_force = SWIOTLB_NO_FORCE;
 121		io_tlb_nslabs = 1;
 122	}
 123
 124	return 0;
 125}
 126early_param("swiotlb", setup_io_tlb_npages);
 127/* make io_tlb_overflow tunable too? */
 128
 129unsigned long swiotlb_nr_tbl(void)
 130{
 131	return io_tlb_nslabs;
 132}
 133EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
 134
 135unsigned int swiotlb_max_segment(void)
 136{
 137	return max_segment;
 138}
 139EXPORT_SYMBOL_GPL(swiotlb_max_segment);
 140
 141void swiotlb_set_max_segment(unsigned int val)
 142{
 143	if (swiotlb_force == SWIOTLB_FORCE)
 144		max_segment = 1;
 145	else
 146		max_segment = rounddown(val, PAGE_SIZE);
 147}
 148
 149/* default to 64MB */
 150#define IO_TLB_DEFAULT_SIZE (64UL<<20)
 151unsigned long swiotlb_size_or_default(void)
 152{
 153	unsigned long size;
 154
 155	size = io_tlb_nslabs << IO_TLB_SHIFT;
 156
 157	return size ? size : (IO_TLB_DEFAULT_SIZE);
 158}
 159
 
 
 
 
 
 
 
 160static bool no_iotlb_memory;
 161
 162void swiotlb_print_info(void)
 163{
 164	unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
 165	unsigned char *vstart, *vend;
 166
 167	if (no_iotlb_memory) {
 168		pr_warn("software IO TLB: No low mem\n");
 169		return;
 170	}
 171
 172	vstart = phys_to_virt(io_tlb_start);
 173	vend = phys_to_virt(io_tlb_end);
 174
 175	printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
 176	       (unsigned long long)io_tlb_start,
 177	       (unsigned long long)io_tlb_end,
 178	       bytes >> 20, vstart, vend - 1);
 179}
 180
 181/*
 182 * Early SWIOTLB allocation may be too early to allow an architecture to
 183 * perform the desired operations.  This function allows the architecture to
 184 * call SWIOTLB when the operations are possible.  It needs to be called
 185 * before the SWIOTLB memory is used.
 186 */
 187void __init swiotlb_update_mem_attributes(void)
 188{
 189	void *vaddr;
 190	unsigned long bytes;
 191
 192	if (no_iotlb_memory || late_alloc)
 193		return;
 194
 195	vaddr = phys_to_virt(io_tlb_start);
 196	bytes = PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT);
 197	set_memory_decrypted((unsigned long)vaddr, bytes >> PAGE_SHIFT);
 198	memset(vaddr, 0, bytes);
 199
 200	vaddr = phys_to_virt(io_tlb_overflow_buffer);
 201	bytes = PAGE_ALIGN(io_tlb_overflow);
 202	set_memory_decrypted((unsigned long)vaddr, bytes >> PAGE_SHIFT);
 203	memset(vaddr, 0, bytes);
 204}
 205
 206int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
 207{
 208	void *v_overflow_buffer;
 209	unsigned long i, bytes;
 210
 211	bytes = nslabs << IO_TLB_SHIFT;
 212
 213	io_tlb_nslabs = nslabs;
 214	io_tlb_start = __pa(tlb);
 215	io_tlb_end = io_tlb_start + bytes;
 216
 217	/*
 218	 * Get the overflow emergency buffer
 219	 */
 220	v_overflow_buffer = memblock_virt_alloc_low_nopanic(
 221						PAGE_ALIGN(io_tlb_overflow),
 222						PAGE_SIZE);
 223	if (!v_overflow_buffer)
 224		return -ENOMEM;
 225
 226	io_tlb_overflow_buffer = __pa(v_overflow_buffer);
 227
 228	/*
 229	 * Allocate and initialize the free list array.  This array is used
 230	 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
 231	 * between io_tlb_start and io_tlb_end.
 232	 */
 233	io_tlb_list = memblock_virt_alloc(
 234				PAGE_ALIGN(io_tlb_nslabs * sizeof(int)),
 235				PAGE_SIZE);
 
 
 
 236	io_tlb_orig_addr = memblock_virt_alloc(
 237				PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)),
 238				PAGE_SIZE);
 239	for (i = 0; i < io_tlb_nslabs; i++) {
 240		io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
 241		io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
 242	}
 243	io_tlb_index = 0;
 244
 245	if (verbose)
 246		swiotlb_print_info();
 247
 248	swiotlb_set_max_segment(io_tlb_nslabs << IO_TLB_SHIFT);
 249	return 0;
 250}
 251
 252/*
 253 * Statically reserve bounce buffer space and initialize bounce buffer data
 254 * structures for the software IO TLB used to implement the DMA API.
 255 */
 256void  __init
 257swiotlb_init(int verbose)
 258{
 259	size_t default_size = IO_TLB_DEFAULT_SIZE;
 260	unsigned char *vstart;
 261	unsigned long bytes;
 262
 263	if (!io_tlb_nslabs) {
 264		io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
 265		io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
 266	}
 267
 268	bytes = io_tlb_nslabs << IO_TLB_SHIFT;
 269
 270	/* Get IO TLB memory from the low pages */
 271	vstart = memblock_virt_alloc_low_nopanic(PAGE_ALIGN(bytes), PAGE_SIZE);
 272	if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose))
 273		return;
 274
 275	if (io_tlb_start)
 276		memblock_free_early(io_tlb_start,
 277				    PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
 278	pr_warn("Cannot allocate SWIOTLB buffer");
 279	no_iotlb_memory = true;
 280}
 281
 282/*
 283 * Systems with larger DMA zones (those that don't support ISA) can
 284 * initialize the swiotlb later using the slab allocator if needed.
 285 * This should be just like above, but with some error catching.
 286 */
 287int
 288swiotlb_late_init_with_default_size(size_t default_size)
 289{
 290	unsigned long bytes, req_nslabs = io_tlb_nslabs;
 291	unsigned char *vstart = NULL;
 292	unsigned int order;
 293	int rc = 0;
 294
 295	if (!io_tlb_nslabs) {
 296		io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
 297		io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
 298	}
 299
 300	/*
 301	 * Get IO TLB memory from the low pages
 302	 */
 303	order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
 304	io_tlb_nslabs = SLABS_PER_PAGE << order;
 305	bytes = io_tlb_nslabs << IO_TLB_SHIFT;
 306
 307	while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
 308		vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
 309						  order);
 310		if (vstart)
 311			break;
 312		order--;
 313	}
 314
 315	if (!vstart) {
 316		io_tlb_nslabs = req_nslabs;
 317		return -ENOMEM;
 318	}
 319	if (order != get_order(bytes)) {
 320		printk(KERN_WARNING "Warning: only able to allocate %ld MB "
 321		       "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
 322		io_tlb_nslabs = SLABS_PER_PAGE << order;
 323	}
 324	rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
 325	if (rc)
 326		free_pages((unsigned long)vstart, order);
 327
 328	return rc;
 329}
 330
 331int
 332swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
 333{
 334	unsigned long i, bytes;
 335	unsigned char *v_overflow_buffer;
 336
 337	bytes = nslabs << IO_TLB_SHIFT;
 338
 339	io_tlb_nslabs = nslabs;
 340	io_tlb_start = virt_to_phys(tlb);
 341	io_tlb_end = io_tlb_start + bytes;
 342
 343	set_memory_decrypted((unsigned long)tlb, bytes >> PAGE_SHIFT);
 344	memset(tlb, 0, bytes);
 345
 346	/*
 347	 * Get the overflow emergency buffer
 348	 */
 349	v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
 350						     get_order(io_tlb_overflow));
 351	if (!v_overflow_buffer)
 352		goto cleanup2;
 353
 354	set_memory_decrypted((unsigned long)v_overflow_buffer,
 355			io_tlb_overflow >> PAGE_SHIFT);
 356	memset(v_overflow_buffer, 0, io_tlb_overflow);
 357	io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
 358
 359	/*
 360	 * Allocate and initialize the free list array.  This array is used
 361	 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
 362	 * between io_tlb_start and io_tlb_end.
 363	 */
 364	io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
 365	                              get_order(io_tlb_nslabs * sizeof(int)));
 366	if (!io_tlb_list)
 367		goto cleanup3;
 368
 
 
 
 
 369	io_tlb_orig_addr = (phys_addr_t *)
 370		__get_free_pages(GFP_KERNEL,
 371				 get_order(io_tlb_nslabs *
 372					   sizeof(phys_addr_t)));
 373	if (!io_tlb_orig_addr)
 374		goto cleanup4;
 375
 376	for (i = 0; i < io_tlb_nslabs; i++) {
 377		io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
 378		io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
 379	}
 380	io_tlb_index = 0;
 381
 382	swiotlb_print_info();
 383
 384	late_alloc = 1;
 385
 386	swiotlb_set_max_segment(io_tlb_nslabs << IO_TLB_SHIFT);
 387
 388	return 0;
 389
 390cleanup4:
 391	free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
 392	                                                 sizeof(int)));
 393	io_tlb_list = NULL;
 394cleanup3:
 395	free_pages((unsigned long)v_overflow_buffer,
 396		   get_order(io_tlb_overflow));
 397	io_tlb_overflow_buffer = 0;
 398cleanup2:
 399	io_tlb_end = 0;
 400	io_tlb_start = 0;
 401	io_tlb_nslabs = 0;
 402	max_segment = 0;
 403	return -ENOMEM;
 404}
 405
 406void __init swiotlb_exit(void)
 407{
 408	if (!io_tlb_orig_addr)
 409		return;
 410
 411	if (late_alloc) {
 412		free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer),
 413			   get_order(io_tlb_overflow));
 414		free_pages((unsigned long)io_tlb_orig_addr,
 415			   get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
 416		free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
 417								 sizeof(int)));
 418		free_pages((unsigned long)phys_to_virt(io_tlb_start),
 419			   get_order(io_tlb_nslabs << IO_TLB_SHIFT));
 420	} else {
 421		memblock_free_late(io_tlb_overflow_buffer,
 422				   PAGE_ALIGN(io_tlb_overflow));
 423		memblock_free_late(__pa(io_tlb_orig_addr),
 424				   PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
 425		memblock_free_late(__pa(io_tlb_list),
 426				   PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
 427		memblock_free_late(io_tlb_start,
 428				   PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
 429	}
 430	io_tlb_nslabs = 0;
 431	max_segment = 0;
 432}
 433
 434int is_swiotlb_buffer(phys_addr_t paddr)
 435{
 436	return paddr >= io_tlb_start && paddr < io_tlb_end;
 437}
 438
 439/*
 440 * Bounce: copy the swiotlb buffer back to the original dma location
 441 */
 442static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
 443			   size_t size, enum dma_data_direction dir)
 444{
 445	unsigned long pfn = PFN_DOWN(orig_addr);
 446	unsigned char *vaddr = phys_to_virt(tlb_addr);
 447
 448	if (PageHighMem(pfn_to_page(pfn))) {
 449		/* The buffer does not have a mapping.  Map it in and copy */
 450		unsigned int offset = orig_addr & ~PAGE_MASK;
 451		char *buffer;
 452		unsigned int sz = 0;
 453		unsigned long flags;
 454
 455		while (size) {
 456			sz = min_t(size_t, PAGE_SIZE - offset, size);
 457
 458			local_irq_save(flags);
 459			buffer = kmap_atomic(pfn_to_page(pfn));
 460			if (dir == DMA_TO_DEVICE)
 461				memcpy(vaddr, buffer + offset, sz);
 462			else
 463				memcpy(buffer + offset, vaddr, sz);
 464			kunmap_atomic(buffer);
 465			local_irq_restore(flags);
 466
 467			size -= sz;
 468			pfn++;
 469			vaddr += sz;
 470			offset = 0;
 471		}
 472	} else if (dir == DMA_TO_DEVICE) {
 473		memcpy(vaddr, phys_to_virt(orig_addr), size);
 474	} else {
 475		memcpy(phys_to_virt(orig_addr), vaddr, size);
 476	}
 477}
 478
 479phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
 480				   dma_addr_t tbl_dma_addr,
 481				   phys_addr_t orig_addr, size_t size,
 482				   enum dma_data_direction dir,
 483				   unsigned long attrs)
 484{
 485	unsigned long flags;
 486	phys_addr_t tlb_addr;
 487	unsigned int nslots, stride, index, wrap;
 488	int i;
 489	unsigned long mask;
 490	unsigned long offset_slots;
 491	unsigned long max_slots;
 492
 493	if (no_iotlb_memory)
 494		panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
 495
 496	if (mem_encrypt_active())
 497		pr_warn_once("%s is active and system is using DMA bounce buffers\n",
 498			     sme_active() ? "SME" : "SEV");
 499
 500	mask = dma_get_seg_boundary(hwdev);
 501
 502	tbl_dma_addr &= mask;
 503
 504	offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
 505
 506	/*
 507 	 * Carefully handle integer overflow which can occur when mask == ~0UL.
 508 	 */
 509	max_slots = mask + 1
 510		    ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
 511		    : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
 512
 513	/*
 514	 * For mappings greater than or equal to a page, we limit the stride
 515	 * (and hence alignment) to a page size.
 516	 */
 517	nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
 518	if (size >= PAGE_SIZE)
 519		stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
 520	else
 521		stride = 1;
 522
 523	BUG_ON(!nslots);
 524
 525	/*
 526	 * Find suitable number of IO TLB entries size that will fit this
 527	 * request and allocate a buffer from that IO TLB pool.
 528	 */
 529	spin_lock_irqsave(&io_tlb_lock, flags);
 530	index = ALIGN(io_tlb_index, stride);
 531	if (index >= io_tlb_nslabs)
 532		index = 0;
 533	wrap = index;
 534
 535	do {
 536		while (iommu_is_span_boundary(index, nslots, offset_slots,
 537					      max_slots)) {
 538			index += stride;
 539			if (index >= io_tlb_nslabs)
 540				index = 0;
 541			if (index == wrap)
 542				goto not_found;
 543		}
 544
 545		/*
 546		 * If we find a slot that indicates we have 'nslots' number of
 547		 * contiguous buffers, we allocate the buffers from that slot
 548		 * and mark the entries as '0' indicating unavailable.
 549		 */
 550		if (io_tlb_list[index] >= nslots) {
 551			int count = 0;
 552
 553			for (i = index; i < (int) (index + nslots); i++)
 554				io_tlb_list[i] = 0;
 555			for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
 556				io_tlb_list[i] = ++count;
 557			tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
 558
 559			/*
 560			 * Update the indices to avoid searching in the next
 561			 * round.
 562			 */
 563			io_tlb_index = ((index + nslots) < io_tlb_nslabs
 564					? (index + nslots) : 0);
 565
 566			goto found;
 567		}
 568		index += stride;
 569		if (index >= io_tlb_nslabs)
 570			index = 0;
 571	} while (index != wrap);
 572
 573not_found:
 574	spin_unlock_irqrestore(&io_tlb_lock, flags);
 575	if (!(attrs & DMA_ATTR_NO_WARN) && printk_ratelimit())
 576		dev_warn(hwdev, "swiotlb buffer is full (sz: %zd bytes)\n", size);
 577	return SWIOTLB_MAP_ERROR;
 578found:
 579	spin_unlock_irqrestore(&io_tlb_lock, flags);
 580
 581	/*
 582	 * Save away the mapping from the original address to the DMA address.
 583	 * This is needed when we sync the memory.  Then we sync the buffer if
 584	 * needed.
 585	 */
 586	for (i = 0; i < nslots; i++)
 587		io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
 588	if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
 589	    (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
 590		swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE);
 591
 592	return tlb_addr;
 593}
 
 594
 595/*
 596 * Allocates bounce buffer and returns its kernel virtual address.
 597 */
 598
 599static phys_addr_t
 600map_single(struct device *hwdev, phys_addr_t phys, size_t size,
 601	   enum dma_data_direction dir, unsigned long attrs)
 602{
 603	dma_addr_t start_dma_addr;
 604
 605	if (swiotlb_force == SWIOTLB_NO_FORCE) {
 606		dev_warn_ratelimited(hwdev, "Cannot do DMA to address %pa\n",
 607				     &phys);
 608		return SWIOTLB_MAP_ERROR;
 609	}
 610
 611	start_dma_addr = __phys_to_dma(hwdev, io_tlb_start);
 612	return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size,
 613				      dir, attrs);
 614}
 615
 616/*
 617 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
 618 */
 619void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
 620			      size_t size, enum dma_data_direction dir,
 621			      unsigned long attrs)
 622{
 623	unsigned long flags;
 624	int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
 625	int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
 626	phys_addr_t orig_addr = io_tlb_orig_addr[index];
 627
 628	/*
 629	 * First, sync the memory before unmapping the entry
 630	 */
 631	if (orig_addr != INVALID_PHYS_ADDR &&
 632	    !(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
 633	    ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
 634		swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
 635
 636	/*
 637	 * Return the buffer to the free list by setting the corresponding
 638	 * entries to indicate the number of contiguous entries available.
 639	 * While returning the entries to the free list, we merge the entries
 640	 * with slots below and above the pool being returned.
 641	 */
 642	spin_lock_irqsave(&io_tlb_lock, flags);
 643	{
 644		count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
 645			 io_tlb_list[index + nslots] : 0);
 646		/*
 647		 * Step 1: return the slots to the free list, merging the
 648		 * slots with superceeding slots
 649		 */
 650		for (i = index + nslots - 1; i >= index; i--) {
 651			io_tlb_list[i] = ++count;
 652			io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
 653		}
 654		/*
 655		 * Step 2: merge the returned slots with the preceding slots,
 656		 * if available (non zero)
 657		 */
 658		for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
 659			io_tlb_list[i] = ++count;
 660	}
 661	spin_unlock_irqrestore(&io_tlb_lock, flags);
 662}
 
 663
 664void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
 665			     size_t size, enum dma_data_direction dir,
 666			     enum dma_sync_target target)
 667{
 668	int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
 669	phys_addr_t orig_addr = io_tlb_orig_addr[index];
 670
 671	if (orig_addr == INVALID_PHYS_ADDR)
 672		return;
 673	orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
 674
 675	switch (target) {
 676	case SYNC_FOR_CPU:
 677		if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
 678			swiotlb_bounce(orig_addr, tlb_addr,
 679				       size, DMA_FROM_DEVICE);
 680		else
 681			BUG_ON(dir != DMA_TO_DEVICE);
 682		break;
 683	case SYNC_FOR_DEVICE:
 684		if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
 685			swiotlb_bounce(orig_addr, tlb_addr,
 686				       size, DMA_TO_DEVICE);
 687		else
 688			BUG_ON(dir != DMA_FROM_DEVICE);
 689		break;
 690	default:
 691		BUG();
 692	}
 693}
 
 694
 695#ifdef CONFIG_DMA_DIRECT_OPS
 696static inline bool dma_coherent_ok(struct device *dev, dma_addr_t addr,
 697		size_t size)
 698{
 699	u64 mask = DMA_BIT_MASK(32);
 700
 701	if (dev && dev->coherent_dma_mask)
 702		mask = dev->coherent_dma_mask;
 703	return addr + size - 1 <= mask;
 704}
 705
 706static void *
 707swiotlb_alloc_buffer(struct device *dev, size_t size, dma_addr_t *dma_handle,
 708		unsigned long attrs)
 709{
 710	phys_addr_t phys_addr;
 711
 712	if (swiotlb_force == SWIOTLB_NO_FORCE)
 713		goto out_warn;
 714
 715	phys_addr = swiotlb_tbl_map_single(dev,
 716			__phys_to_dma(dev, io_tlb_start),
 717			0, size, DMA_FROM_DEVICE, attrs);
 718	if (phys_addr == SWIOTLB_MAP_ERROR)
 719		goto out_warn;
 720
 721	*dma_handle = __phys_to_dma(dev, phys_addr);
 722	if (!dma_coherent_ok(dev, *dma_handle, size))
 723		goto out_unmap;
 724
 725	memset(phys_to_virt(phys_addr), 0, size);
 726	return phys_to_virt(phys_addr);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 727
 728out_unmap:
 729	dev_warn(dev, "hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
 730		(unsigned long long)(dev ? dev->coherent_dma_mask : 0),
 731		(unsigned long long)*dma_handle);
 732
 733	/*
 734	 * DMA_TO_DEVICE to avoid memcpy in unmap_single.
 735	 * DMA_ATTR_SKIP_CPU_SYNC is optional.
 736	 */
 737	swiotlb_tbl_unmap_single(dev, phys_addr, size, DMA_TO_DEVICE,
 738			DMA_ATTR_SKIP_CPU_SYNC);
 739out_warn:
 740	if (!(attrs & DMA_ATTR_NO_WARN) && printk_ratelimit()) {
 741		dev_warn(dev,
 742			"swiotlb: coherent allocation failed, size=%zu\n",
 743			size);
 744		dump_stack();
 745	}
 746	return NULL;
 747}
 
 748
 749static bool swiotlb_free_buffer(struct device *dev, size_t size,
 750		dma_addr_t dma_addr)
 
 751{
 752	phys_addr_t phys_addr = dma_to_phys(dev, dma_addr);
 753
 754	WARN_ON_ONCE(irqs_disabled());
 755
 756	if (!is_swiotlb_buffer(phys_addr))
 757		return false;
 758
 759	/*
 760	 * DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single.
 761	 * DMA_ATTR_SKIP_CPU_SYNC is optional.
 762	 */
 763	swiotlb_tbl_unmap_single(dev, phys_addr, size, DMA_TO_DEVICE,
 764				 DMA_ATTR_SKIP_CPU_SYNC);
 765	return true;
 766}
 767#endif
 768
 769static void
 770swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
 771	     int do_panic)
 772{
 773	if (swiotlb_force == SWIOTLB_NO_FORCE)
 774		return;
 775
 776	/*
 777	 * Ran out of IOMMU space for this operation. This is very bad.
 778	 * Unfortunately the drivers cannot handle this operation properly.
 779	 * unless they check for dma_mapping_error (most don't)
 780	 * When the mapping is small enough return a static buffer to limit
 781	 * the damage, or panic when the transfer is too big.
 782	 */
 783	dev_err_ratelimited(dev, "DMA: Out of SW-IOMMU space for %zu bytes\n",
 784			    size);
 785
 786	if (size <= io_tlb_overflow || !do_panic)
 787		return;
 788
 789	if (dir == DMA_BIDIRECTIONAL)
 790		panic("DMA: Random memory could be DMA accessed\n");
 791	if (dir == DMA_FROM_DEVICE)
 792		panic("DMA: Random memory could be DMA written\n");
 793	if (dir == DMA_TO_DEVICE)
 794		panic("DMA: Random memory could be DMA read\n");
 795}
 796
 797/*
 798 * Map a single buffer of the indicated size for DMA in streaming mode.  The
 799 * physical address to use is returned.
 800 *
 801 * Once the device is given the dma address, the device owns this memory until
 802 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
 803 */
 804dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
 805			    unsigned long offset, size_t size,
 806			    enum dma_data_direction dir,
 807			    unsigned long attrs)
 808{
 809	phys_addr_t map, phys = page_to_phys(page) + offset;
 810	dma_addr_t dev_addr = phys_to_dma(dev, phys);
 811
 812	BUG_ON(dir == DMA_NONE);
 813	/*
 814	 * If the address happens to be in the device's DMA window,
 815	 * we can safely return the device addr and not worry about bounce
 816	 * buffering it.
 817	 */
 818	if (dma_capable(dev, dev_addr, size) && swiotlb_force != SWIOTLB_FORCE)
 819		return dev_addr;
 820
 821	trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
 822
 823	/* Oh well, have to allocate and map a bounce buffer. */
 824	map = map_single(dev, phys, size, dir, attrs);
 825	if (map == SWIOTLB_MAP_ERROR) {
 826		swiotlb_full(dev, size, dir, 1);
 827		return __phys_to_dma(dev, io_tlb_overflow_buffer);
 828	}
 829
 830	dev_addr = __phys_to_dma(dev, map);
 831
 832	/* Ensure that the address returned is DMA'ble */
 833	if (dma_capable(dev, dev_addr, size))
 834		return dev_addr;
 835
 836	attrs |= DMA_ATTR_SKIP_CPU_SYNC;
 837	swiotlb_tbl_unmap_single(dev, map, size, dir, attrs);
 838
 839	return __phys_to_dma(dev, io_tlb_overflow_buffer);
 840}
 
 841
 842/*
 843 * Unmap a single streaming mode DMA translation.  The dma_addr and size must
 844 * match what was provided for in a previous swiotlb_map_page call.  All
 845 * other usages are undefined.
 846 *
 847 * After this call, reads by the cpu to the buffer are guaranteed to see
 848 * whatever the device wrote there.
 849 */
 850static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
 851			 size_t size, enum dma_data_direction dir,
 852			 unsigned long attrs)
 853{
 854	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
 855
 856	BUG_ON(dir == DMA_NONE);
 857
 858	if (is_swiotlb_buffer(paddr)) {
 859		swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
 860		return;
 861	}
 862
 863	if (dir != DMA_FROM_DEVICE)
 864		return;
 865
 866	/*
 867	 * phys_to_virt doesn't work with hihgmem page but we could
 868	 * call dma_mark_clean() with hihgmem page here. However, we
 869	 * are fine since dma_mark_clean() is null on POWERPC. We can
 870	 * make dma_mark_clean() take a physical address if necessary.
 871	 */
 872	dma_mark_clean(phys_to_virt(paddr), size);
 873}
 874
 875void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
 876			size_t size, enum dma_data_direction dir,
 877			unsigned long attrs)
 878{
 879	unmap_single(hwdev, dev_addr, size, dir, attrs);
 880}
 
 881
 882/*
 883 * Make physical memory consistent for a single streaming mode DMA translation
 884 * after a transfer.
 885 *
 886 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
 887 * using the cpu, yet do not wish to teardown the dma mapping, you must
 888 * call this function before doing so.  At the next point you give the dma
 889 * address back to the card, you must first perform a
 890 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
 891 */
 892static void
 893swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
 894		    size_t size, enum dma_data_direction dir,
 895		    enum dma_sync_target target)
 896{
 897	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
 898
 899	BUG_ON(dir == DMA_NONE);
 900
 901	if (is_swiotlb_buffer(paddr)) {
 902		swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
 903		return;
 904	}
 905
 906	if (dir != DMA_FROM_DEVICE)
 907		return;
 908
 909	dma_mark_clean(phys_to_virt(paddr), size);
 910}
 911
 912void
 913swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
 914			    size_t size, enum dma_data_direction dir)
 915{
 916	swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
 917}
 
 918
 919void
 920swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
 921			       size_t size, enum dma_data_direction dir)
 922{
 923	swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
 924}
 
 925
 926/*
 927 * Map a set of buffers described by scatterlist in streaming mode for DMA.
 928 * This is the scatter-gather version of the above swiotlb_map_page
 929 * interface.  Here the scatter gather list elements are each tagged with the
 930 * appropriate dma address and length.  They are obtained via
 931 * sg_dma_{address,length}(SG).
 932 *
 933 * NOTE: An implementation may be able to use a smaller number of
 934 *       DMA address/length pairs than there are SG table elements.
 935 *       (for example via virtual mapping capabilities)
 936 *       The routine returns the number of addr/length pairs actually
 937 *       used, at most nents.
 938 *
 939 * Device ownership issues as mentioned above for swiotlb_map_page are the
 940 * same here.
 941 */
 942int
 943swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
 944		     enum dma_data_direction dir, unsigned long attrs)
 945{
 946	struct scatterlist *sg;
 947	int i;
 948
 949	BUG_ON(dir == DMA_NONE);
 950
 951	for_each_sg(sgl, sg, nelems, i) {
 952		phys_addr_t paddr = sg_phys(sg);
 953		dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
 954
 955		if (swiotlb_force == SWIOTLB_FORCE ||
 956		    !dma_capable(hwdev, dev_addr, sg->length)) {
 957			phys_addr_t map = map_single(hwdev, sg_phys(sg),
 958						     sg->length, dir, attrs);
 959			if (map == SWIOTLB_MAP_ERROR) {
 960				/* Don't panic here, we expect map_sg users
 961				   to do proper error handling. */
 962				swiotlb_full(hwdev, sg->length, dir, 0);
 963				attrs |= DMA_ATTR_SKIP_CPU_SYNC;
 964				swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
 965						       attrs);
 966				sg_dma_len(sgl) = 0;
 967				return 0;
 968			}
 969			sg->dma_address = __phys_to_dma(hwdev, map);
 970		} else
 971			sg->dma_address = dev_addr;
 972		sg_dma_len(sg) = sg->length;
 973	}
 974	return nelems;
 975}
 
 
 
 
 
 
 
 
 
 976
 977/*
 978 * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
 979 * concerning calls here are the same as for swiotlb_unmap_page() above.
 980 */
 981void
 982swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
 983		       int nelems, enum dma_data_direction dir,
 984		       unsigned long attrs)
 985{
 986	struct scatterlist *sg;
 987	int i;
 988
 989	BUG_ON(dir == DMA_NONE);
 990
 991	for_each_sg(sgl, sg, nelems, i)
 992		unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir,
 993			     attrs);
 994}
 
 
 
 
 
 
 
 
 
 995
 996/*
 997 * Make physical memory consistent for a set of streaming mode DMA translations
 998 * after a transfer.
 999 *
1000 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
1001 * and usage.
1002 */
1003static void
1004swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
1005		int nelems, enum dma_data_direction dir,
1006		enum dma_sync_target target)
1007{
1008	struct scatterlist *sg;
1009	int i;
1010
1011	for_each_sg(sgl, sg, nelems, i)
1012		swiotlb_sync_single(hwdev, sg->dma_address,
1013				    sg_dma_len(sg), dir, target);
1014}
1015
1016void
1017swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
1018			int nelems, enum dma_data_direction dir)
1019{
1020	swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
1021}
 
1022
1023void
1024swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
1025			   int nelems, enum dma_data_direction dir)
1026{
1027	swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
1028}
 
1029
1030int
1031swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
1032{
1033	return (dma_addr == __phys_to_dma(hwdev, io_tlb_overflow_buffer));
1034}
 
1035
1036/*
1037 * Return whether the given device DMA address mask can be supported
1038 * properly.  For example, if your device can only drive the low 24-bits
1039 * during bus mastering, then you would pass 0x00ffffff as the mask to
1040 * this function.
1041 */
1042int
1043swiotlb_dma_supported(struct device *hwdev, u64 mask)
1044{
1045	return __phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
1046}
1047
1048#ifdef CONFIG_DMA_DIRECT_OPS
1049void *swiotlb_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
1050		gfp_t gfp, unsigned long attrs)
1051{
1052	void *vaddr;
1053
1054	/* temporary workaround: */
1055	if (gfp & __GFP_NOWARN)
1056		attrs |= DMA_ATTR_NO_WARN;
1057
1058	/*
1059	 * Don't print a warning when the first allocation attempt fails.
1060	 * swiotlb_alloc_coherent() will print a warning when the DMA memory
1061	 * allocation ultimately failed.
1062	 */
1063	gfp |= __GFP_NOWARN;
1064
1065	vaddr = dma_direct_alloc(dev, size, dma_handle, gfp, attrs);
1066	if (!vaddr)
1067		vaddr = swiotlb_alloc_buffer(dev, size, dma_handle, attrs);
1068	return vaddr;
1069}
1070
1071void swiotlb_free(struct device *dev, size_t size, void *vaddr,
1072		dma_addr_t dma_addr, unsigned long attrs)
1073{
1074	if (!swiotlb_free_buffer(dev, size, dma_addr))
1075		dma_direct_free(dev, size, vaddr, dma_addr, attrs);
1076}
1077
1078const struct dma_map_ops swiotlb_dma_ops = {
1079	.mapping_error		= swiotlb_dma_mapping_error,
1080	.alloc			= swiotlb_alloc,
1081	.free			= swiotlb_free,
1082	.sync_single_for_cpu	= swiotlb_sync_single_for_cpu,
1083	.sync_single_for_device	= swiotlb_sync_single_for_device,
1084	.sync_sg_for_cpu	= swiotlb_sync_sg_for_cpu,
1085	.sync_sg_for_device	= swiotlb_sync_sg_for_device,
1086	.map_sg			= swiotlb_map_sg_attrs,
1087	.unmap_sg		= swiotlb_unmap_sg_attrs,
1088	.map_page		= swiotlb_map_page,
1089	.unmap_page		= swiotlb_unmap_page,
1090	.dma_supported		= dma_direct_supported,
1091};
1092#endif /* CONFIG_DMA_DIRECT_OPS */