blob: ac25cd28e8077717263b03a179df2764acd10a50 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Dynamic DMA mapping support.
3 *
Jan Beulich563aaf02007-02-05 18:51:25 -08004 * This implementation is a fallback for platforms that do not support
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 * 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.
John W. Linville569c8bf2005-09-29 14:45:24 -070014 * 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.
Becky Brucefb05a372008-12-22 10:26:09 -080017 * 08/12/11 beckyb Add highmem support
Linus Torvalds1da177e2005-04-16 15:20:36 -070018 */
19
20#include <linux/cache.h>
Tony Luck17e5ad62005-09-29 15:52:13 -070021#include <linux/dma-mapping.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070022#include <linux/mm.h>
23#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/spinlock.h>
25#include <linux/string.h>
Ian Campbell0016fde2008-12-16 12:17:27 -080026#include <linux/swiotlb.h>
Becky Brucefb05a372008-12-22 10:26:09 -080027#include <linux/pfn.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#include <linux/types.h>
29#include <linux/ctype.h>
Jeremy Fitzhardingeef9b1892008-12-16 12:17:33 -080030#include <linux/highmem.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070031
32#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070033#include <asm/dma.h>
Tony Luck17e5ad62005-09-29 15:52:13 -070034#include <asm/scatterlist.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070035
36#include <linux/init.h>
37#include <linux/bootmem.h>
FUJITA Tomonoria8522502008-04-29 00:59:36 -070038#include <linux/iommu-helper.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070039
40#define OFFSET(val,align) ((unsigned long) \
41 ( (val) & ( (align) - 1)))
42
Alex Williamson0b9afed2005-09-06 11:20:49 -060043#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
44
45/*
46 * Minimum IO TLB size to bother booting with. Systems with mainly
47 * 64bit capable cards will only lightly use the swiotlb. If we can't
48 * allocate a contiguous 1MB, we're probably in trouble anyway.
49 */
50#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
51
John W. Linvillede69e0f2005-09-29 14:44:57 -070052/*
53 * Enumeration for sync targets
54 */
55enum dma_sync_target {
56 SYNC_FOR_CPU = 0,
57 SYNC_FOR_DEVICE = 1,
58};
59
Linus Torvalds1da177e2005-04-16 15:20:36 -070060int swiotlb_force;
61
62/*
Becky Bruceceb5ac32009-04-08 09:09:15 -050063 * Used to do a quick range check in unmap_single and
64 * sync_single_*, to see if the memory was in fact allocated by this
Linus Torvalds1da177e2005-04-16 15:20:36 -070065 * API.
66 */
67static char *io_tlb_start, *io_tlb_end;
68
69/*
70 * The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and
71 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
72 */
73static unsigned long io_tlb_nslabs;
74
75/*
76 * When the IOMMU overflows we return a fallback buffer. This sets the size.
77 */
78static unsigned long io_tlb_overflow = 32*1024;
79
80void *io_tlb_overflow_buffer;
81
82/*
83 * This is a free list describing the number of free entries available from
84 * each index
85 */
86static unsigned int *io_tlb_list;
87static unsigned int io_tlb_index;
88
89/*
90 * We need to save away the original address corresponding to a mapped entry
91 * for the sync operations.
92 */
Becky Brucebc40ac62008-12-22 10:26:08 -080093static phys_addr_t *io_tlb_orig_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -070094
95/*
96 * Protect the above data structures in the map and unmap calls
97 */
98static DEFINE_SPINLOCK(io_tlb_lock);
99
100static int __init
101setup_io_tlb_npages(char *str)
102{
103 if (isdigit(*str)) {
Alex Williamsone8579e72005-08-04 13:06:00 -0700104 io_tlb_nslabs = simple_strtoul(str, &str, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105 /* avoid tail segment of size < IO_TLB_SEGSIZE */
106 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
107 }
108 if (*str == ',')
109 ++str;
110 if (!strcmp(str, "force"))
111 swiotlb_force = 1;
112 return 1;
113}
114__setup("swiotlb=", setup_io_tlb_npages);
115/* make io_tlb_overflow tunable too? */
116
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900117/* Note that this doesn't work with highmem page */
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800118static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
119 volatile void *address)
Ian Campbelle08e1f72008-12-16 12:17:30 -0800120{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900121 return phys_to_dma(hwdev, virt_to_phys(address));
Ian Campbelle08e1f72008-12-16 12:17:30 -0800122}
123
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800124static void swiotlb_print_info(unsigned long bytes)
125{
126 phys_addr_t pstart, pend;
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800127
128 pstart = virt_to_phys(io_tlb_start);
129 pend = virt_to_phys(io_tlb_end);
130
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800131 printk(KERN_INFO "Placing %luMB software IO TLB between %p - %p\n",
132 bytes >> 20, io_tlb_start, io_tlb_end);
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800133 printk(KERN_INFO "software IO TLB at phys %#llx - %#llx\n",
134 (unsigned long long)pstart,
135 (unsigned long long)pend);
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800136}
137
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138/*
139 * Statically reserve bounce buffer space and initialize bounce buffer data
Tony Luck17e5ad62005-09-29 15:52:13 -0700140 * structures for the software IO TLB used to implement the DMA API.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800142void __init
143swiotlb_init_with_default_size(size_t default_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144{
Jan Beulich563aaf02007-02-05 18:51:25 -0800145 unsigned long i, bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146
147 if (!io_tlb_nslabs) {
Alex Williamsone8579e72005-08-04 13:06:00 -0700148 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
150 }
151
Jan Beulich563aaf02007-02-05 18:51:25 -0800152 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
153
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 /*
155 * Get IO TLB memory from the low pages
156 */
FUJITA Tomonori38851232009-07-10 10:04:50 +0900157 io_tlb_start = alloc_bootmem_low_pages(bytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158 if (!io_tlb_start)
159 panic("Cannot allocate SWIOTLB buffer");
Jan Beulich563aaf02007-02-05 18:51:25 -0800160 io_tlb_end = io_tlb_start + bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161
162 /*
163 * Allocate and initialize the free list array. This array is used
164 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
165 * between io_tlb_start and io_tlb_end.
166 */
167 io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
Tony Luck25667d62007-03-06 13:31:45 -0800168 for (i = 0; i < io_tlb_nslabs; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
170 io_tlb_index = 0;
Becky Brucebc40ac62008-12-22 10:26:08 -0800171 io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(phys_addr_t));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172
173 /*
174 * Get the overflow emergency buffer
175 */
176 io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
Jan Beulich563aaf02007-02-05 18:51:25 -0800177 if (!io_tlb_overflow_buffer)
178 panic("Cannot allocate SWIOTLB overflow buffer!\n");
179
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800180 swiotlb_print_info(bytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181}
182
Jan Beulich563aaf02007-02-05 18:51:25 -0800183void __init
184swiotlb_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185{
Tony Luck25667d62007-03-06 13:31:45 -0800186 swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187}
188
Alex Williamson0b9afed2005-09-06 11:20:49 -0600189/*
190 * Systems with larger DMA zones (those that don't support ISA) can
191 * initialize the swiotlb later using the slab allocator if needed.
192 * This should be just like above, but with some error catching.
193 */
194int
Jan Beulich563aaf02007-02-05 18:51:25 -0800195swiotlb_late_init_with_default_size(size_t default_size)
Alex Williamson0b9afed2005-09-06 11:20:49 -0600196{
Jan Beulich563aaf02007-02-05 18:51:25 -0800197 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600198 unsigned int order;
199
200 if (!io_tlb_nslabs) {
201 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
202 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
203 }
204
205 /*
206 * Get IO TLB memory from the low pages
207 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800208 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600209 io_tlb_nslabs = SLABS_PER_PAGE << order;
Jan Beulich563aaf02007-02-05 18:51:25 -0800210 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600211
212 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
FUJITA Tomonoribb521962009-07-10 10:04:51 +0900213 io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
214 order);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600215 if (io_tlb_start)
216 break;
217 order--;
218 }
219
220 if (!io_tlb_start)
221 goto cleanup1;
222
Jan Beulich563aaf02007-02-05 18:51:25 -0800223 if (order != get_order(bytes)) {
Alex Williamson0b9afed2005-09-06 11:20:49 -0600224 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
225 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
226 io_tlb_nslabs = SLABS_PER_PAGE << order;
Jan Beulich563aaf02007-02-05 18:51:25 -0800227 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600228 }
Jan Beulich563aaf02007-02-05 18:51:25 -0800229 io_tlb_end = io_tlb_start + bytes;
230 memset(io_tlb_start, 0, bytes);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600231
232 /*
233 * Allocate and initialize the free list array. This array is used
234 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
235 * between io_tlb_start and io_tlb_end.
236 */
237 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
238 get_order(io_tlb_nslabs * sizeof(int)));
239 if (!io_tlb_list)
240 goto cleanup2;
241
242 for (i = 0; i < io_tlb_nslabs; i++)
243 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
244 io_tlb_index = 0;
245
Becky Brucebc40ac62008-12-22 10:26:08 -0800246 io_tlb_orig_addr = (phys_addr_t *)
247 __get_free_pages(GFP_KERNEL,
248 get_order(io_tlb_nslabs *
249 sizeof(phys_addr_t)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600250 if (!io_tlb_orig_addr)
251 goto cleanup3;
252
Becky Brucebc40ac62008-12-22 10:26:08 -0800253 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600254
255 /*
256 * Get the overflow emergency buffer
257 */
258 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
259 get_order(io_tlb_overflow));
260 if (!io_tlb_overflow_buffer)
261 goto cleanup4;
262
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800263 swiotlb_print_info(bytes);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600264
265 return 0;
266
267cleanup4:
Becky Brucebc40ac62008-12-22 10:26:08 -0800268 free_pages((unsigned long)io_tlb_orig_addr,
269 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600270 io_tlb_orig_addr = NULL;
271cleanup3:
Tony Luck25667d62007-03-06 13:31:45 -0800272 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
273 sizeof(int)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600274 io_tlb_list = NULL;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600275cleanup2:
Jan Beulich563aaf02007-02-05 18:51:25 -0800276 io_tlb_end = NULL;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600277 free_pages((unsigned long)io_tlb_start, order);
278 io_tlb_start = NULL;
279cleanup1:
280 io_tlb_nslabs = req_nslabs;
281 return -ENOMEM;
282}
283
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900284static int is_swiotlb_buffer(phys_addr_t paddr)
FUJITA Tomonori640aebf2008-09-08 18:53:50 +0900285{
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900286 return paddr >= virt_to_phys(io_tlb_start) &&
287 paddr < virt_to_phys(io_tlb_end);
FUJITA Tomonori640aebf2008-09-08 18:53:50 +0900288}
289
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290/*
Becky Brucefb05a372008-12-22 10:26:09 -0800291 * Bounce: copy the swiotlb buffer back to the original dma location
292 */
293static void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
294 enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295{
Becky Brucefb05a372008-12-22 10:26:09 -0800296 unsigned long pfn = PFN_DOWN(phys);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297
Becky Brucefb05a372008-12-22 10:26:09 -0800298 if (PageHighMem(pfn_to_page(pfn))) {
299 /* The buffer does not have a mapping. Map it in and copy */
300 unsigned int offset = phys & ~PAGE_MASK;
301 char *buffer;
302 unsigned int sz = 0;
303 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304
Becky Brucefb05a372008-12-22 10:26:09 -0800305 while (size) {
Becky Bruce67131ad2009-04-08 09:09:16 -0500306 sz = min_t(size_t, PAGE_SIZE - offset, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307
Becky Brucefb05a372008-12-22 10:26:09 -0800308 local_irq_save(flags);
309 buffer = kmap_atomic(pfn_to_page(pfn),
310 KM_BOUNCE_READ);
311 if (dir == DMA_TO_DEVICE)
312 memcpy(dma_addr, buffer + offset, sz);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313 else
Becky Brucefb05a372008-12-22 10:26:09 -0800314 memcpy(buffer + offset, dma_addr, sz);
315 kunmap_atomic(buffer, KM_BOUNCE_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316 local_irq_restore(flags);
Becky Brucefb05a372008-12-22 10:26:09 -0800317
318 size -= sz;
319 pfn++;
320 dma_addr += sz;
321 offset = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322 }
323 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700324 if (dir == DMA_TO_DEVICE)
Becky Brucefb05a372008-12-22 10:26:09 -0800325 memcpy(dma_addr, phys_to_virt(phys), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326 else
Becky Brucefb05a372008-12-22 10:26:09 -0800327 memcpy(phys_to_virt(phys), dma_addr, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328 }
329}
330
331/*
332 * Allocates bounce buffer and returns its kernel virtual address.
333 */
334static void *
Becky Brucebc40ac62008-12-22 10:26:08 -0800335map_single(struct device *hwdev, phys_addr_t phys, size_t size, int dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336{
337 unsigned long flags;
338 char *dma_addr;
339 unsigned int nslots, stride, index, wrap;
340 int i;
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800341 unsigned long start_dma_addr;
342 unsigned long mask;
343 unsigned long offset_slots;
344 unsigned long max_slots;
345
346 mask = dma_get_seg_boundary(hwdev);
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800347 start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start) & mask;
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800348
349 offset_slots = ALIGN(start_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
Ian Campbella5ddde42008-12-16 12:17:29 -0800350
351 /*
352 * Carefully handle integer overflow which can occur when mask == ~0UL.
353 */
Jan Beulichb15a3892008-03-13 09:13:30 +0000354 max_slots = mask + 1
355 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
356 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357
358 /*
359 * For mappings greater than a page, we limit the stride (and
360 * hence alignment) to a page size.
361 */
362 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
363 if (size > PAGE_SIZE)
364 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
365 else
366 stride = 1;
367
Eric Sesterhenn34814542006-03-24 18:47:11 +0100368 BUG_ON(!nslots);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369
370 /*
371 * Find suitable number of IO TLB entries size that will fit this
372 * request and allocate a buffer from that IO TLB pool.
373 */
374 spin_lock_irqsave(&io_tlb_lock, flags);
Andrew Mortona7133a12008-04-29 00:59:36 -0700375 index = ALIGN(io_tlb_index, stride);
376 if (index >= io_tlb_nslabs)
377 index = 0;
378 wrap = index;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379
Andrew Mortona7133a12008-04-29 00:59:36 -0700380 do {
FUJITA Tomonoria8522502008-04-29 00:59:36 -0700381 while (iommu_is_span_boundary(index, nslots, offset_slots,
382 max_slots)) {
Jan Beulichb15a3892008-03-13 09:13:30 +0000383 index += stride;
384 if (index >= io_tlb_nslabs)
385 index = 0;
Andrew Mortona7133a12008-04-29 00:59:36 -0700386 if (index == wrap)
387 goto not_found;
388 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389
Andrew Mortona7133a12008-04-29 00:59:36 -0700390 /*
391 * If we find a slot that indicates we have 'nslots' number of
392 * contiguous buffers, we allocate the buffers from that slot
393 * and mark the entries as '0' indicating unavailable.
394 */
395 if (io_tlb_list[index] >= nslots) {
396 int count = 0;
397
398 for (i = index; i < (int) (index + nslots); i++)
399 io_tlb_list[i] = 0;
400 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
401 io_tlb_list[i] = ++count;
402 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
403
404 /*
405 * Update the indices to avoid searching in the next
406 * round.
407 */
408 io_tlb_index = ((index + nslots) < io_tlb_nslabs
409 ? (index + nslots) : 0);
410
411 goto found;
412 }
413 index += stride;
414 if (index >= io_tlb_nslabs)
415 index = 0;
416 } while (index != wrap);
417
418not_found:
419 spin_unlock_irqrestore(&io_tlb_lock, flags);
420 return NULL;
421found:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 spin_unlock_irqrestore(&io_tlb_lock, flags);
423
424 /*
425 * Save away the mapping from the original address to the DMA address.
426 * This is needed when we sync the memory. Then we sync the buffer if
427 * needed.
428 */
Becky Brucebc40ac62008-12-22 10:26:08 -0800429 for (i = 0; i < nslots; i++)
430 io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700431 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
Becky Brucefb05a372008-12-22 10:26:09 -0800432 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433
434 return dma_addr;
435}
436
437/*
438 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
439 */
440static void
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500441do_unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442{
443 unsigned long flags;
444 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
445 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
Becky Brucebc40ac62008-12-22 10:26:08 -0800446 phys_addr_t phys = io_tlb_orig_addr[index];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447
448 /*
449 * First, sync the memory before unmapping the entry
450 */
Becky Brucebc40ac62008-12-22 10:26:08 -0800451 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
Becky Brucefb05a372008-12-22 10:26:09 -0800452 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453
454 /*
455 * Return the buffer to the free list by setting the corresponding
456 * entries to indicate the number of contigous entries available.
457 * While returning the entries to the free list, we merge the entries
458 * with slots below and above the pool being returned.
459 */
460 spin_lock_irqsave(&io_tlb_lock, flags);
461 {
462 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
463 io_tlb_list[index + nslots] : 0);
464 /*
465 * Step 1: return the slots to the free list, merging the
466 * slots with superceeding slots
467 */
468 for (i = index + nslots - 1; i >= index; i--)
469 io_tlb_list[i] = ++count;
470 /*
471 * Step 2: merge the returned slots with the preceding slots,
472 * if available (non zero)
473 */
474 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
475 io_tlb_list[i] = ++count;
476 }
477 spin_unlock_irqrestore(&io_tlb_lock, flags);
478}
479
480static void
John W. Linvillede69e0f2005-09-29 14:44:57 -0700481sync_single(struct device *hwdev, char *dma_addr, size_t size,
482 int dir, int target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483{
Becky Brucebc40ac62008-12-22 10:26:08 -0800484 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
485 phys_addr_t phys = io_tlb_orig_addr[index];
486
487 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
Keir Fraserdf336d12007-07-21 04:37:24 -0700488
John W. Linvillede69e0f2005-09-29 14:44:57 -0700489 switch (target) {
490 case SYNC_FOR_CPU:
491 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
Becky Brucefb05a372008-12-22 10:26:09 -0800492 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100493 else
494 BUG_ON(dir != DMA_TO_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700495 break;
496 case SYNC_FOR_DEVICE:
497 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
Becky Brucefb05a372008-12-22 10:26:09 -0800498 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100499 else
500 BUG_ON(dir != DMA_FROM_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700501 break;
502 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 BUG();
John W. Linvillede69e0f2005-09-29 14:44:57 -0700504 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505}
506
507void *
508swiotlb_alloc_coherent(struct device *hwdev, size_t size,
Al Viro06a54492005-10-21 03:21:03 -0400509 dma_addr_t *dma_handle, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510{
Jan Beulich563aaf02007-02-05 18:51:25 -0800511 dma_addr_t dev_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 void *ret;
513 int order = get_order(size);
Yang Hongyang284901a2009-04-06 19:01:15 -0700514 u64 dma_mask = DMA_BIT_MASK(32);
FUJITA Tomonori1e74f302008-11-17 16:24:34 +0900515
516 if (hwdev && hwdev->coherent_dma_mask)
517 dma_mask = hwdev->coherent_dma_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518
Tony Luck25667d62007-03-06 13:31:45 -0800519 ret = (void *)__get_free_pages(flags, order);
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900520 if (ret && swiotlb_virt_to_bus(hwdev, ret) + size > dma_mask) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 /*
522 * The allocated memory isn't reachable by the device.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 */
524 free_pages((unsigned long) ret, order);
525 ret = NULL;
526 }
527 if (!ret) {
528 /*
529 * We are either out of memory or the device can't DMA
Becky Bruceceb5ac32009-04-08 09:09:15 -0500530 * to GFP_DMA memory; fall back on map_single(), which
531 * will grab memory from the lowest available address range.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 */
Becky Brucebc40ac62008-12-22 10:26:08 -0800533 ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
FUJITA Tomonori9dfda122008-09-08 18:53:48 +0900534 if (!ret)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 }
537
538 memset(ret, 0, size);
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800539 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540
541 /* Confirm address can be DMA'd by device */
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900542 if (dev_addr + size > dma_mask) {
Jan Beulich563aaf02007-02-05 18:51:25 -0800543 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
FUJITA Tomonori1e74f302008-11-17 16:24:34 +0900544 (unsigned long long)dma_mask,
Jan Beulich563aaf02007-02-05 18:51:25 -0800545 (unsigned long long)dev_addr);
FUJITA Tomonoria2b89b52008-10-23 18:42:03 +0900546
547 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500548 do_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
FUJITA Tomonoria2b89b52008-10-23 18:42:03 +0900549 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 }
551 *dma_handle = dev_addr;
552 return ret;
553}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900554EXPORT_SYMBOL(swiotlb_alloc_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555
556void
557swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900558 dma_addr_t dev_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900560 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900561
David Brownellaa248862007-08-10 13:10:27 -0700562 WARN_ON(irqs_disabled());
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900563 if (!is_swiotlb_buffer(paddr))
564 free_pages((unsigned long)vaddr, get_order(size));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565 else
566 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500567 do_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900569EXPORT_SYMBOL(swiotlb_free_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570
571static void
572swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
573{
574 /*
575 * Ran out of IOMMU space for this operation. This is very bad.
576 * Unfortunately the drivers cannot handle this operation properly.
Tony Luck17e5ad62005-09-29 15:52:13 -0700577 * unless they check for dma_mapping_error (most don't)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578 * When the mapping is small enough return a static buffer to limit
579 * the damage, or panic when the transfer is too big.
580 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800581 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
Kay Sievers94b32482009-01-06 10:44:37 -0800582 "device %s\n", size, dev ? dev_name(dev) : "?");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583
Casey Dahlinc7084b32009-08-20 16:27:56 -0700584 if (size <= io_tlb_overflow || !do_panic)
585 return;
586
587 if (dir == DMA_BIDIRECTIONAL)
588 panic("DMA: Random memory could be DMA accessed\n");
589 if (dir == DMA_FROM_DEVICE)
590 panic("DMA: Random memory could be DMA written\n");
591 if (dir == DMA_TO_DEVICE)
592 panic("DMA: Random memory could be DMA read\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593}
594
595/*
596 * Map a single buffer of the indicated size for DMA in streaming mode. The
Tony Luck17e5ad62005-09-29 15:52:13 -0700597 * physical address to use is returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 *
599 * Once the device is given the dma address, the device owns this memory until
Becky Bruceceb5ac32009-04-08 09:09:15 -0500600 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601 */
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900602dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
603 unsigned long offset, size_t size,
604 enum dma_data_direction dir,
605 struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606{
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900607 phys_addr_t phys = page_to_phys(page) + offset;
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900608 dma_addr_t dev_addr = phys_to_dma(dev, phys);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 void *map;
610
Eric Sesterhenn34814542006-03-24 18:47:11 +0100611 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612 /*
Becky Bruceceb5ac32009-04-08 09:09:15 -0500613 * If the address happens to be in the device's DMA window,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 * we can safely return the device addr and not worry about bounce
615 * buffering it.
616 */
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900617 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 return dev_addr;
619
620 /*
621 * Oh well, have to allocate and map a bounce buffer.
622 */
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900623 map = map_single(dev, phys, size, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 if (!map) {
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900625 swiotlb_full(dev, size, dir, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 map = io_tlb_overflow_buffer;
627 }
628
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900629 dev_addr = swiotlb_virt_to_bus(dev, map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630
631 /*
632 * Ensure that the address returned is DMA'ble
633 */
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900634 if (!dma_capable(dev, dev_addr, size))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 panic("map_single: bounce buffer is not DMA'ble");
636
637 return dev_addr;
638}
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900639EXPORT_SYMBOL_GPL(swiotlb_map_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640
641/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642 * Unmap a single streaming mode DMA translation. The dma_addr and size must
Becky Bruceceb5ac32009-04-08 09:09:15 -0500643 * match what was provided for in a previous swiotlb_map_page call. All
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644 * other usages are undefined.
645 *
646 * After this call, reads by the cpu to the buffer are guaranteed to see
647 * whatever the device wrote there.
648 */
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500649static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
650 size_t size, int dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900652 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653
Eric Sesterhenn34814542006-03-24 18:47:11 +0100654 BUG_ON(dir == DMA_NONE);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500655
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900656 if (is_swiotlb_buffer(paddr)) {
657 do_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500658 return;
659 }
660
661 if (dir != DMA_FROM_DEVICE)
662 return;
663
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900664 /*
665 * phys_to_virt doesn't work with hihgmem page but we could
666 * call dma_mark_clean() with hihgmem page here. However, we
667 * are fine since dma_mark_clean() is null on POWERPC. We can
668 * make dma_mark_clean() take a physical address if necessary.
669 */
670 dma_mark_clean(phys_to_virt(paddr), size);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500671}
672
673void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
674 size_t size, enum dma_data_direction dir,
675 struct dma_attrs *attrs)
676{
677 unmap_single(hwdev, dev_addr, size, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678}
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900679EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900680
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681/*
682 * Make physical memory consistent for a single streaming mode DMA translation
683 * after a transfer.
684 *
Becky Bruceceb5ac32009-04-08 09:09:15 -0500685 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
Tony Luck17e5ad62005-09-29 15:52:13 -0700686 * using the cpu, yet do not wish to teardown the dma mapping, you must
687 * call this function before doing so. At the next point you give the dma
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 * address back to the card, you must first perform a
689 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
690 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800691static void
John W. Linville8270f3f2005-09-29 14:43:32 -0700692swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700693 size_t size, int dir, int target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900695 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696
Eric Sesterhenn34814542006-03-24 18:47:11 +0100697 BUG_ON(dir == DMA_NONE);
Becky Bruce380d6872009-04-08 09:09:20 -0500698
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900699 if (is_swiotlb_buffer(paddr)) {
700 sync_single(hwdev, phys_to_virt(paddr), size, dir, target);
Becky Bruce380d6872009-04-08 09:09:20 -0500701 return;
702 }
703
704 if (dir != DMA_FROM_DEVICE)
705 return;
706
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900707 dma_mark_clean(phys_to_virt(paddr), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708}
709
710void
John W. Linville8270f3f2005-09-29 14:43:32 -0700711swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900712 size_t size, enum dma_data_direction dir)
John W. Linville8270f3f2005-09-29 14:43:32 -0700713{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700714 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700715}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900716EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
John W. Linville8270f3f2005-09-29 14:43:32 -0700717
718void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900720 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700722 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900724EXPORT_SYMBOL(swiotlb_sync_single_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725
726/*
John W. Linville878a97c2005-09-29 14:44:23 -0700727 * Same as above, but for a sub-range of the mapping.
728 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800729static void
John W. Linville878a97c2005-09-29 14:44:23 -0700730swiotlb_sync_single_range(struct device *hwdev, dma_addr_t dev_addr,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700731 unsigned long offset, size_t size,
732 int dir, int target)
John W. Linville878a97c2005-09-29 14:44:23 -0700733{
Becky Bruce380d6872009-04-08 09:09:20 -0500734 swiotlb_sync_single(hwdev, dev_addr + offset, size, dir, target);
John W. Linville878a97c2005-09-29 14:44:23 -0700735}
736
737void
738swiotlb_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900739 unsigned long offset, size_t size,
740 enum dma_data_direction dir)
John W. Linville878a97c2005-09-29 14:44:23 -0700741{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700742 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
743 SYNC_FOR_CPU);
John W. Linville878a97c2005-09-29 14:44:23 -0700744}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900745EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu);
John W. Linville878a97c2005-09-29 14:44:23 -0700746
747void
748swiotlb_sync_single_range_for_device(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900749 unsigned long offset, size_t size,
750 enum dma_data_direction dir)
John W. Linville878a97c2005-09-29 14:44:23 -0700751{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700752 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
753 SYNC_FOR_DEVICE);
John W. Linville878a97c2005-09-29 14:44:23 -0700754}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900755EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device);
John W. Linville878a97c2005-09-29 14:44:23 -0700756
757/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 * Map a set of buffers described by scatterlist in streaming mode for DMA.
Becky Bruceceb5ac32009-04-08 09:09:15 -0500759 * This is the scatter-gather version of the above swiotlb_map_page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760 * interface. Here the scatter gather list elements are each tagged with the
761 * appropriate dma address and length. They are obtained via
762 * sg_dma_{address,length}(SG).
763 *
764 * NOTE: An implementation may be able to use a smaller number of
765 * DMA address/length pairs than there are SG table elements.
766 * (for example via virtual mapping capabilities)
767 * The routine returns the number of addr/length pairs actually
768 * used, at most nents.
769 *
Becky Bruceceb5ac32009-04-08 09:09:15 -0500770 * Device ownership issues as mentioned above for swiotlb_map_page are the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771 * same here.
772 */
773int
Arthur Kepner309df0c2008-04-29 01:00:32 -0700774swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900775 enum dma_data_direction dir, struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200777 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 int i;
779
Eric Sesterhenn34814542006-03-24 18:47:11 +0100780 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781
Jens Axboedbfd49f2007-05-11 14:56:18 +0200782 for_each_sg(sgl, sg, nelems, i) {
Ian Campbell961d7d02009-01-09 18:32:10 +0000783 phys_addr_t paddr = sg_phys(sg);
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900784 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
Becky Brucebc40ac62008-12-22 10:26:08 -0800785
FUJITA Tomonoricf56e3f2009-07-10 10:04:52 +0900786 if (swiotlb_force ||
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900787 !dma_capable(hwdev, dev_addr, sg->length)) {
Becky Brucebc40ac62008-12-22 10:26:08 -0800788 void *map = map_single(hwdev, sg_phys(sg),
789 sg->length, dir);
Andi Kleen7e870232005-12-20 14:45:19 +0100790 if (!map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 /* Don't panic here, we expect map_sg users
792 to do proper error handling. */
793 swiotlb_full(hwdev, sg->length, dir, 0);
Arthur Kepner309df0c2008-04-29 01:00:32 -0700794 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
795 attrs);
Jens Axboedbfd49f2007-05-11 14:56:18 +0200796 sgl[0].dma_length = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 return 0;
798 }
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800799 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 } else
801 sg->dma_address = dev_addr;
802 sg->dma_length = sg->length;
803 }
804 return nelems;
805}
Arthur Kepner309df0c2008-04-29 01:00:32 -0700806EXPORT_SYMBOL(swiotlb_map_sg_attrs);
807
808int
809swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
810 int dir)
811{
812 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
813}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900814EXPORT_SYMBOL(swiotlb_map_sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815
816/*
817 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
Becky Bruceceb5ac32009-04-08 09:09:15 -0500818 * concerning calls here are the same as for swiotlb_unmap_page() above.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819 */
820void
Arthur Kepner309df0c2008-04-29 01:00:32 -0700821swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900822 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200824 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 int i;
826
Eric Sesterhenn34814542006-03-24 18:47:11 +0100827 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500829 for_each_sg(sgl, sg, nelems, i)
830 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
831
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832}
Arthur Kepner309df0c2008-04-29 01:00:32 -0700833EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
834
835void
836swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
837 int dir)
838{
839 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
840}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900841EXPORT_SYMBOL(swiotlb_unmap_sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842
843/*
844 * Make physical memory consistent for a set of streaming mode DMA translations
845 * after a transfer.
846 *
847 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
848 * and usage.
849 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800850static void
Jens Axboedbfd49f2007-05-11 14:56:18 +0200851swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700852 int nelems, int dir, int target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200854 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 int i;
856
Becky Bruce380d6872009-04-08 09:09:20 -0500857 for_each_sg(sgl, sg, nelems, i)
858 swiotlb_sync_single(hwdev, sg->dma_address,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700859 sg->dma_length, dir, target);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860}
861
862void
John W. Linville8270f3f2005-09-29 14:43:32 -0700863swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900864 int nelems, enum dma_data_direction dir)
John W. Linville8270f3f2005-09-29 14:43:32 -0700865{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700866 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700867}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900868EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
John W. Linville8270f3f2005-09-29 14:43:32 -0700869
870void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900872 int nelems, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700874 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900876EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877
878int
FUJITA Tomonori8d8bb392008-07-25 19:44:49 -0700879swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880{
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800881 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900883EXPORT_SYMBOL(swiotlb_dma_mapping_error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884
885/*
Tony Luck17e5ad62005-09-29 15:52:13 -0700886 * Return whether the given device DMA address mask can be supported
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887 * properly. For example, if your device can only drive the low 24-bits
Tony Luck17e5ad62005-09-29 15:52:13 -0700888 * during bus mastering, then you would pass 0x00ffffff as the mask to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 * this function.
890 */
891int
Jan Beulich563aaf02007-02-05 18:51:25 -0800892swiotlb_dma_supported(struct device *hwdev, u64 mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893{
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800894 return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896EXPORT_SYMBOL(swiotlb_dma_supported);