blob: 196b06984decbc0263b74eda9f34e94f425eb070 [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>
Paul Gortmaker8bc3bcc2011-11-16 21:29:17 -050023#include <linux/export.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>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090031#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070032
33#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070034#include <asm/dma.h>
Tony Luck17e5ad62005-09-29 15:52:13 -070035#include <asm/scatterlist.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036
37#include <linux/init.h>
38#include <linux/bootmem.h>
FUJITA Tomonoria8522502008-04-29 00:59:36 -070039#include <linux/iommu-helper.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040
41#define OFFSET(val,align) ((unsigned long) \
42 ( (val) & ( (align) - 1)))
43
Alex Williamson0b9afed2005-09-06 11:20:49 -060044#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
45
46/*
47 * Minimum IO TLB size to bother booting with. Systems with mainly
48 * 64bit capable cards will only lightly use the swiotlb. If we can't
49 * allocate a contiguous 1MB, we're probably in trouble anyway.
50 */
51#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
52
Linus Torvalds1da177e2005-04-16 15:20:36 -070053int swiotlb_force;
54
55/*
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -040056 * Used to do a quick range check in swiotlb_tbl_unmap_single and
57 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
Linus Torvalds1da177e2005-04-16 15:20:36 -070058 * API.
59 */
Alexander Duyckff7204a2012-10-15 10:19:28 -070060static phys_addr_t io_tlb_start, io_tlb_end;
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
62/*
Uwe Kleine-Königb5950762010-11-01 15:38:34 -040063 * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
Linus Torvalds1da177e2005-04-16 15:20:36 -070064 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
65 */
66static unsigned long io_tlb_nslabs;
67
68/*
69 * When the IOMMU overflows we return a fallback buffer. This sets the size.
70 */
71static unsigned long io_tlb_overflow = 32*1024;
72
Alexander Duyckee3f6ba2012-10-15 10:19:34 -070073static phys_addr_t io_tlb_overflow_buffer;
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
75/*
76 * This is a free list describing the number of free entries available from
77 * each index
78 */
79static unsigned int *io_tlb_list;
80static unsigned int io_tlb_index;
81
82/*
83 * We need to save away the original address corresponding to a mapped entry
84 * for the sync operations.
85 */
Becky Brucebc40ac62008-12-22 10:26:08 -080086static phys_addr_t *io_tlb_orig_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -070087
88/*
89 * Protect the above data structures in the map and unmap calls
90 */
91static DEFINE_SPINLOCK(io_tlb_lock);
92
FUJITA Tomonori5740afd2009-11-10 19:46:18 +090093static int late_alloc;
94
Linus Torvalds1da177e2005-04-16 15:20:36 -070095static int __init
96setup_io_tlb_npages(char *str)
97{
98 if (isdigit(*str)) {
Alex Williamsone8579e72005-08-04 13:06:00 -070099 io_tlb_nslabs = simple_strtoul(str, &str, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100 /* avoid tail segment of size < IO_TLB_SEGSIZE */
101 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
102 }
103 if (*str == ',')
104 ++str;
FUJITA Tomonorib18485e2009-11-12 00:03:28 +0900105 if (!strcmp(str, "force"))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106 swiotlb_force = 1;
FUJITA Tomonorib18485e2009-11-12 00:03:28 +0900107
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108 return 1;
109}
110__setup("swiotlb=", setup_io_tlb_npages);
111/* make io_tlb_overflow tunable too? */
112
Konrad Rzeszutek Wilkf21ffe92011-08-11 16:50:56 -0400113unsigned long swiotlb_nr_tbl(void)
FUJITA Tomonori5f98ecd2011-06-05 11:47:29 +0900114{
115 return io_tlb_nslabs;
116}
Konrad Rzeszutek Wilkf21ffe92011-08-11 16:50:56 -0400117EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900118/* Note that this doesn't work with highmem page */
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800119static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
120 volatile void *address)
Ian Campbelle08e1f72008-12-16 12:17:30 -0800121{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900122 return phys_to_dma(hwdev, virt_to_phys(address));
Ian Campbelle08e1f72008-12-16 12:17:30 -0800123}
124
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900125void swiotlb_print_info(void)
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800126{
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900127 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700128 unsigned char *vstart, *vend;
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800129
Alexander Duyckff7204a2012-10-15 10:19:28 -0700130 vstart = phys_to_virt(io_tlb_start);
Alexander Duyckc40dba02012-10-15 10:19:22 -0700131 vend = phys_to_virt(io_tlb_end);
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800132
Bjorn Helgaas3af684c2012-05-29 15:06:29 -0700133 printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
Alexander Duyckff7204a2012-10-15 10:19:28 -0700134 (unsigned long long)io_tlb_start,
Alexander Duyckc40dba02012-10-15 10:19:22 -0700135 (unsigned long long)io_tlb_end,
Alexander Duyckff7204a2012-10-15 10:19:28 -0700136 bytes >> 20, vstart, vend - 1);
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800137}
138
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400139void __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140{
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700141 void *v_overflow_buffer;
Jan Beulich563aaf02007-02-05 18:51:25 -0800142 unsigned long i, bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400144 bytes = nslabs << IO_TLB_SHIFT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400146 io_tlb_nslabs = nslabs;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700147 io_tlb_start = __pa(tlb);
148 io_tlb_end = io_tlb_start + bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149
150 /*
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700151 * Get the overflow emergency buffer
152 */
153 v_overflow_buffer = alloc_bootmem_low_pages(PAGE_ALIGN(io_tlb_overflow));
154 if (!v_overflow_buffer)
155 panic("Cannot allocate SWIOTLB overflow buffer!\n");
156
157 io_tlb_overflow_buffer = __pa(v_overflow_buffer);
158
159 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160 * Allocate and initialize the free list array. This array is used
161 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
162 * between io_tlb_start and io_tlb_end.
163 */
Yinghai Lue79f86b2010-10-11 10:40:25 -0700164 io_tlb_list = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
Tony Luck25667d62007-03-06 13:31:45 -0800165 for (i = 0; i < io_tlb_nslabs; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
167 io_tlb_index = 0;
Yinghai Lue79f86b2010-10-11 10:40:25 -0700168 io_tlb_orig_addr = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900170 if (verbose)
171 swiotlb_print_info();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172}
173
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400174/*
175 * Statically reserve bounce buffer space and initialize bounce buffer data
176 * structures for the software IO TLB used to implement the DMA API.
177 */
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400178static void __init
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400179swiotlb_init_with_default_size(size_t default_size, int verbose)
180{
Alexander Duyckff7204a2012-10-15 10:19:28 -0700181 unsigned char *vstart;
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400182 unsigned long bytes;
183
184 if (!io_tlb_nslabs) {
185 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
186 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
187 }
188
189 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
190
191 /*
192 * Get IO TLB memory from the low pages
193 */
Alexander Duyckff7204a2012-10-15 10:19:28 -0700194 vstart = alloc_bootmem_low_pages(PAGE_ALIGN(bytes));
195 if (!vstart)
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400196 panic("Cannot allocate SWIOTLB buffer");
197
Alexander Duyckff7204a2012-10-15 10:19:28 -0700198 swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose);
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400199}
200
Jan Beulich563aaf02007-02-05 18:51:25 -0800201void __init
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900202swiotlb_init(int verbose)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203{
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900204 swiotlb_init_with_default_size(64 * (1<<20), verbose); /* default to 64MB */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205}
206
Alex Williamson0b9afed2005-09-06 11:20:49 -0600207/*
208 * Systems with larger DMA zones (those that don't support ISA) can
209 * initialize the swiotlb later using the slab allocator if needed.
210 * This should be just like above, but with some error catching.
211 */
212int
Jan Beulich563aaf02007-02-05 18:51:25 -0800213swiotlb_late_init_with_default_size(size_t default_size)
Alex Williamson0b9afed2005-09-06 11:20:49 -0600214{
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400215 unsigned long bytes, req_nslabs = io_tlb_nslabs;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700216 unsigned char *vstart = NULL;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600217 unsigned int order;
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400218 int rc = 0;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600219
220 if (!io_tlb_nslabs) {
221 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
222 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
223 }
224
225 /*
226 * Get IO TLB memory from the low pages
227 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800228 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600229 io_tlb_nslabs = SLABS_PER_PAGE << order;
Jan Beulich563aaf02007-02-05 18:51:25 -0800230 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600231
232 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
Alexander Duyckff7204a2012-10-15 10:19:28 -0700233 vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
234 order);
235 if (vstart)
Alex Williamson0b9afed2005-09-06 11:20:49 -0600236 break;
237 order--;
238 }
239
Alexander Duyckff7204a2012-10-15 10:19:28 -0700240 if (!vstart) {
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400241 io_tlb_nslabs = req_nslabs;
242 return -ENOMEM;
243 }
Jan Beulich563aaf02007-02-05 18:51:25 -0800244 if (order != get_order(bytes)) {
Alex Williamson0b9afed2005-09-06 11:20:49 -0600245 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
246 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
247 io_tlb_nslabs = SLABS_PER_PAGE << order;
248 }
Alexander Duyckff7204a2012-10-15 10:19:28 -0700249 rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400250 if (rc)
Alexander Duyckff7204a2012-10-15 10:19:28 -0700251 free_pages((unsigned long)vstart, order);
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400252 return rc;
253}
254
255int
256swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
257{
258 unsigned long i, bytes;
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700259 unsigned char *v_overflow_buffer;
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400260
261 bytes = nslabs << IO_TLB_SHIFT;
262
263 io_tlb_nslabs = nslabs;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700264 io_tlb_start = virt_to_phys(tlb);
265 io_tlb_end = io_tlb_start + bytes;
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400266
Alexander Duyckff7204a2012-10-15 10:19:28 -0700267 memset(tlb, 0, bytes);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600268
269 /*
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700270 * Get the overflow emergency buffer
271 */
272 v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
273 get_order(io_tlb_overflow));
274 if (!v_overflow_buffer)
275 goto cleanup2;
276
277 io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
278
279 /*
Alex Williamson0b9afed2005-09-06 11:20:49 -0600280 * Allocate and initialize the free list array. This array is used
281 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
282 * between io_tlb_start and io_tlb_end.
283 */
284 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
285 get_order(io_tlb_nslabs * sizeof(int)));
286 if (!io_tlb_list)
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700287 goto cleanup3;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600288
289 for (i = 0; i < io_tlb_nslabs; i++)
290 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
291 io_tlb_index = 0;
292
Becky Brucebc40ac62008-12-22 10:26:08 -0800293 io_tlb_orig_addr = (phys_addr_t *)
294 __get_free_pages(GFP_KERNEL,
295 get_order(io_tlb_nslabs *
296 sizeof(phys_addr_t)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600297 if (!io_tlb_orig_addr)
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700298 goto cleanup4;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600299
Becky Brucebc40ac62008-12-22 10:26:08 -0800300 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600301
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900302 swiotlb_print_info();
Alex Williamson0b9afed2005-09-06 11:20:49 -0600303
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900304 late_alloc = 1;
305
Alex Williamson0b9afed2005-09-06 11:20:49 -0600306 return 0;
307
308cleanup4:
Tony Luck25667d62007-03-06 13:31:45 -0800309 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
310 sizeof(int)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600311 io_tlb_list = NULL;
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700312cleanup3:
313 free_pages((unsigned long)v_overflow_buffer,
314 get_order(io_tlb_overflow));
315 io_tlb_overflow_buffer = 0;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600316cleanup2:
Alexander Duyckc40dba02012-10-15 10:19:22 -0700317 io_tlb_end = 0;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700318 io_tlb_start = 0;
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400319 io_tlb_nslabs = 0;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600320 return -ENOMEM;
321}
322
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900323void __init swiotlb_free(void)
324{
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700325 if (!io_tlb_orig_addr)
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900326 return;
327
328 if (late_alloc) {
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700329 free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer),
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900330 get_order(io_tlb_overflow));
331 free_pages((unsigned long)io_tlb_orig_addr,
332 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
333 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
334 sizeof(int)));
Alexander Duyckff7204a2012-10-15 10:19:28 -0700335 free_pages((unsigned long)phys_to_virt(io_tlb_start),
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900336 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
337 } else {
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700338 free_bootmem_late(io_tlb_overflow_buffer,
Yinghai Lue79f86b2010-10-11 10:40:25 -0700339 PAGE_ALIGN(io_tlb_overflow));
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900340 free_bootmem_late(__pa(io_tlb_orig_addr),
Yinghai Lue79f86b2010-10-11 10:40:25 -0700341 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900342 free_bootmem_late(__pa(io_tlb_list),
Yinghai Lue79f86b2010-10-11 10:40:25 -0700343 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
Alexander Duyckff7204a2012-10-15 10:19:28 -0700344 free_bootmem_late(io_tlb_start,
Yinghai Lue79f86b2010-10-11 10:40:25 -0700345 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900346 }
Konrad Rzeszutek Wilkf21ffe92011-08-11 16:50:56 -0400347 io_tlb_nslabs = 0;
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900348}
349
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900350static int is_swiotlb_buffer(phys_addr_t paddr)
FUJITA Tomonori640aebf2008-09-08 18:53:50 +0900351{
Alexander Duyckff7204a2012-10-15 10:19:28 -0700352 return paddr >= io_tlb_start && paddr < io_tlb_end;
FUJITA Tomonori640aebf2008-09-08 18:53:50 +0900353}
354
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355/*
Becky Brucefb05a372008-12-22 10:26:09 -0800356 * Bounce: copy the swiotlb buffer back to the original dma location
357 */
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700358static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
359 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360{
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700361 unsigned long pfn = PFN_DOWN(orig_addr);
362 unsigned char *vaddr = phys_to_virt(tlb_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363
Becky Brucefb05a372008-12-22 10:26:09 -0800364 if (PageHighMem(pfn_to_page(pfn))) {
365 /* The buffer does not have a mapping. Map it in and copy */
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700366 unsigned int offset = orig_addr & ~PAGE_MASK;
Becky Brucefb05a372008-12-22 10:26:09 -0800367 char *buffer;
368 unsigned int sz = 0;
369 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370
Becky Brucefb05a372008-12-22 10:26:09 -0800371 while (size) {
Becky Bruce67131ad2009-04-08 09:09:16 -0500372 sz = min_t(size_t, PAGE_SIZE - offset, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373
Becky Brucefb05a372008-12-22 10:26:09 -0800374 local_irq_save(flags);
Cong Wangc3eede82011-11-25 23:14:39 +0800375 buffer = kmap_atomic(pfn_to_page(pfn));
Becky Brucefb05a372008-12-22 10:26:09 -0800376 if (dir == DMA_TO_DEVICE)
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700377 memcpy(vaddr, buffer + offset, sz);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 else
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700379 memcpy(buffer + offset, vaddr, sz);
Cong Wangc3eede82011-11-25 23:14:39 +0800380 kunmap_atomic(buffer);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381 local_irq_restore(flags);
Becky Brucefb05a372008-12-22 10:26:09 -0800382
383 size -= sz;
384 pfn++;
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700385 vaddr += sz;
Becky Brucefb05a372008-12-22 10:26:09 -0800386 offset = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 }
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700388 } else if (dir == DMA_TO_DEVICE) {
389 memcpy(vaddr, phys_to_virt(orig_addr), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390 } else {
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700391 memcpy(phys_to_virt(orig_addr), vaddr, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392 }
393}
394
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700395phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
396 dma_addr_t tbl_dma_addr,
397 phys_addr_t orig_addr, size_t size,
398 enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399{
400 unsigned long flags;
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700401 phys_addr_t tlb_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402 unsigned int nslots, stride, index, wrap;
403 int i;
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800404 unsigned long mask;
405 unsigned long offset_slots;
406 unsigned long max_slots;
407
408 mask = dma_get_seg_boundary(hwdev);
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800409
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400410 tbl_dma_addr &= mask;
411
412 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
Ian Campbella5ddde4a2008-12-16 12:17:29 -0800413
414 /*
415 * Carefully handle integer overflow which can occur when mask == ~0UL.
416 */
Jan Beulichb15a38912008-03-13 09:13:30 +0000417 max_slots = mask + 1
418 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
419 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420
421 /*
422 * For mappings greater than a page, we limit the stride (and
423 * hence alignment) to a page size.
424 */
425 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
426 if (size > PAGE_SIZE)
427 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
428 else
429 stride = 1;
430
Eric Sesterhenn34814542006-03-24 18:47:11 +0100431 BUG_ON(!nslots);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432
433 /*
434 * Find suitable number of IO TLB entries size that will fit this
435 * request and allocate a buffer from that IO TLB pool.
436 */
437 spin_lock_irqsave(&io_tlb_lock, flags);
Andrew Mortona7133a12008-04-29 00:59:36 -0700438 index = ALIGN(io_tlb_index, stride);
439 if (index >= io_tlb_nslabs)
440 index = 0;
441 wrap = index;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442
Andrew Mortona7133a12008-04-29 00:59:36 -0700443 do {
FUJITA Tomonoria8522502008-04-29 00:59:36 -0700444 while (iommu_is_span_boundary(index, nslots, offset_slots,
445 max_slots)) {
Jan Beulichb15a38912008-03-13 09:13:30 +0000446 index += stride;
447 if (index >= io_tlb_nslabs)
448 index = 0;
Andrew Mortona7133a12008-04-29 00:59:36 -0700449 if (index == wrap)
450 goto not_found;
451 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452
Andrew Mortona7133a12008-04-29 00:59:36 -0700453 /*
454 * If we find a slot that indicates we have 'nslots' number of
455 * contiguous buffers, we allocate the buffers from that slot
456 * and mark the entries as '0' indicating unavailable.
457 */
458 if (io_tlb_list[index] >= nslots) {
459 int count = 0;
460
461 for (i = index; i < (int) (index + nslots); i++)
462 io_tlb_list[i] = 0;
463 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
464 io_tlb_list[i] = ++count;
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700465 tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
Andrew Mortona7133a12008-04-29 00:59:36 -0700466
467 /*
468 * Update the indices to avoid searching in the next
469 * round.
470 */
471 io_tlb_index = ((index + nslots) < io_tlb_nslabs
472 ? (index + nslots) : 0);
473
474 goto found;
475 }
476 index += stride;
477 if (index >= io_tlb_nslabs)
478 index = 0;
479 } while (index != wrap);
480
481not_found:
482 spin_unlock_irqrestore(&io_tlb_lock, flags);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700483 return SWIOTLB_MAP_ERROR;
Andrew Mortona7133a12008-04-29 00:59:36 -0700484found:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485 spin_unlock_irqrestore(&io_tlb_lock, flags);
486
487 /*
488 * Save away the mapping from the original address to the DMA address.
489 * This is needed when we sync the memory. Then we sync the buffer if
490 * needed.
491 */
Becky Brucebc40ac62008-12-22 10:26:08 -0800492 for (i = 0; i < nslots; i++)
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700493 io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700495 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700496
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700497 return tlb_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400499EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500
501/*
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400502 * Allocates bounce buffer and returns its kernel virtual address.
503 */
504
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700505phys_addr_t map_single(struct device *hwdev, phys_addr_t phys, size_t size,
506 enum dma_data_direction dir)
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400507{
Alexander Duyckff7204a2012-10-15 10:19:28 -0700508 dma_addr_t start_dma_addr = phys_to_dma(hwdev, io_tlb_start);
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400509
510 return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
511}
512
513/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
515 */
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700516void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
517 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518{
519 unsigned long flags;
520 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700521 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
522 phys_addr_t orig_addr = io_tlb_orig_addr[index];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523
524 /*
525 * First, sync the memory before unmapping the entry
526 */
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700527 if (orig_addr && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
528 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529
530 /*
531 * Return the buffer to the free list by setting the corresponding
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200532 * entries to indicate the number of contiguous entries available.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 * While returning the entries to the free list, we merge the entries
534 * with slots below and above the pool being returned.
535 */
536 spin_lock_irqsave(&io_tlb_lock, flags);
537 {
538 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
539 io_tlb_list[index + nslots] : 0);
540 /*
541 * Step 1: return the slots to the free list, merging the
542 * slots with superceeding slots
543 */
544 for (i = index + nslots - 1; i >= index; i--)
545 io_tlb_list[i] = ++count;
546 /*
547 * Step 2: merge the returned slots with the preceding slots,
548 * if available (non zero)
549 */
550 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
551 io_tlb_list[i] = ++count;
552 }
553 spin_unlock_irqrestore(&io_tlb_lock, flags);
554}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400555EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556
Alexander Duyckfbfda892012-10-15 10:19:49 -0700557void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
558 size_t size, enum dma_data_direction dir,
559 enum dma_sync_target target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560{
Alexander Duyckfbfda892012-10-15 10:19:49 -0700561 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
562 phys_addr_t orig_addr = io_tlb_orig_addr[index];
Becky Brucebc40ac62008-12-22 10:26:08 -0800563
Alexander Duyckfbfda892012-10-15 10:19:49 -0700564 orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
Keir Fraserdf336d12007-07-21 04:37:24 -0700565
John W. Linvillede69e0f2005-09-29 14:44:57 -0700566 switch (target) {
567 case SYNC_FOR_CPU:
568 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700569 swiotlb_bounce(orig_addr, tlb_addr,
Alexander Duyckfbfda892012-10-15 10:19:49 -0700570 size, DMA_FROM_DEVICE);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100571 else
572 BUG_ON(dir != DMA_TO_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700573 break;
574 case SYNC_FOR_DEVICE:
575 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700576 swiotlb_bounce(orig_addr, tlb_addr,
Alexander Duyckfbfda892012-10-15 10:19:49 -0700577 size, DMA_TO_DEVICE);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100578 else
579 BUG_ON(dir != DMA_FROM_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700580 break;
581 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 BUG();
John W. Linvillede69e0f2005-09-29 14:44:57 -0700583 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400585EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586
587void *
588swiotlb_alloc_coherent(struct device *hwdev, size_t size,
Al Viro06a54492005-10-21 03:21:03 -0400589 dma_addr_t *dma_handle, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590{
Jan Beulich563aaf02007-02-05 18:51:25 -0800591 dma_addr_t dev_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592 void *ret;
593 int order = get_order(size);
Yang Hongyang284901a2009-04-06 19:01:15 -0700594 u64 dma_mask = DMA_BIT_MASK(32);
FUJITA Tomonori1e74f302008-11-17 16:24:34 +0900595
596 if (hwdev && hwdev->coherent_dma_mask)
597 dma_mask = hwdev->coherent_dma_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598
Tony Luck25667d62007-03-06 13:31:45 -0800599 ret = (void *)__get_free_pages(flags, order);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700600 if (ret) {
601 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
602 if (dev_addr + size - 1 > dma_mask) {
603 /*
604 * The allocated memory isn't reachable by the device.
605 */
606 free_pages((unsigned long) ret, order);
607 ret = NULL;
608 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 }
610 if (!ret) {
611 /*
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400612 * We are either out of memory or the device can't DMA to
613 * GFP_DMA memory; fall back on map_single(), which
Becky Bruceceb5ac32009-04-08 09:09:15 -0500614 * will grab memory from the lowest available address range.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615 */
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700616 phys_addr_t paddr = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
617 if (paddr == SWIOTLB_MAP_ERROR)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700620 ret = phys_to_virt(paddr);
621 dev_addr = phys_to_dma(hwdev, paddr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700623 /* Confirm address can be DMA'd by device */
624 if (dev_addr + size - 1 > dma_mask) {
625 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
626 (unsigned long long)dma_mask,
627 (unsigned long long)dev_addr);
FUJITA Tomonoria2b89b52008-10-23 18:42:03 +0900628
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700629 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
630 swiotlb_tbl_unmap_single(hwdev, paddr,
631 size, DMA_TO_DEVICE);
632 return NULL;
633 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 }
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700635
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636 *dma_handle = dev_addr;
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700637 memset(ret, 0, size);
638
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639 return ret;
640}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900641EXPORT_SYMBOL(swiotlb_alloc_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642
643void
644swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900645 dma_addr_t dev_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900647 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900648
David Brownellaa248862007-08-10 13:10:27 -0700649 WARN_ON(irqs_disabled());
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900650 if (!is_swiotlb_buffer(paddr))
651 free_pages((unsigned long)vaddr, get_order(size));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700652 else
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400653 /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700654 swiotlb_tbl_unmap_single(hwdev, paddr, size, DMA_TO_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900656EXPORT_SYMBOL(swiotlb_free_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657
658static void
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500659swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
660 int do_panic)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661{
662 /*
663 * Ran out of IOMMU space for this operation. This is very bad.
664 * Unfortunately the drivers cannot handle this operation properly.
Tony Luck17e5ad62005-09-29 15:52:13 -0700665 * unless they check for dma_mapping_error (most don't)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 * When the mapping is small enough return a static buffer to limit
667 * the damage, or panic when the transfer is too big.
668 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800669 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
Kay Sievers94b32482009-01-06 10:44:37 -0800670 "device %s\n", size, dev ? dev_name(dev) : "?");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671
Casey Dahlinc7084b32009-08-20 16:27:56 -0700672 if (size <= io_tlb_overflow || !do_panic)
673 return;
674
675 if (dir == DMA_BIDIRECTIONAL)
676 panic("DMA: Random memory could be DMA accessed\n");
677 if (dir == DMA_FROM_DEVICE)
678 panic("DMA: Random memory could be DMA written\n");
679 if (dir == DMA_TO_DEVICE)
680 panic("DMA: Random memory could be DMA read\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681}
682
683/*
684 * Map a single buffer of the indicated size for DMA in streaming mode. The
Tony Luck17e5ad62005-09-29 15:52:13 -0700685 * physical address to use is returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686 *
687 * Once the device is given the dma address, the device owns this memory until
Becky Bruceceb5ac32009-04-08 09:09:15 -0500688 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 */
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900690dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
691 unsigned long offset, size_t size,
692 enum dma_data_direction dir,
693 struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694{
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700695 phys_addr_t map, phys = page_to_phys(page) + offset;
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900696 dma_addr_t dev_addr = phys_to_dma(dev, phys);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697
Eric Sesterhenn34814542006-03-24 18:47:11 +0100698 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 /*
Becky Bruceceb5ac32009-04-08 09:09:15 -0500700 * If the address happens to be in the device's DMA window,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701 * we can safely return the device addr and not worry about bounce
702 * buffering it.
703 */
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900704 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 return dev_addr;
706
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700707 /* Oh well, have to allocate and map a bounce buffer. */
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900708 map = map_single(dev, phys, size, dir);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700709 if (map == SWIOTLB_MAP_ERROR) {
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900710 swiotlb_full(dev, size, dir, 1);
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700711 return phys_to_dma(dev, io_tlb_overflow_buffer);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 }
713
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700714 dev_addr = phys_to_dma(dev, map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700716 /* Ensure that the address returned is DMA'ble */
FUJITA Tomonorifba99fa2011-02-25 14:44:16 -0800717 if (!dma_capable(dev, dev_addr, size)) {
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700718 swiotlb_tbl_unmap_single(dev, map, size, dir);
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700719 return phys_to_dma(dev, io_tlb_overflow_buffer);
FUJITA Tomonorifba99fa2011-02-25 14:44:16 -0800720 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721
722 return dev_addr;
723}
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900724EXPORT_SYMBOL_GPL(swiotlb_map_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725
726/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 * Unmap a single streaming mode DMA translation. The dma_addr and size must
Becky Bruceceb5ac32009-04-08 09:09:15 -0500728 * match what was provided for in a previous swiotlb_map_page call. All
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 * other usages are undefined.
730 *
731 * After this call, reads by the cpu to the buffer are guaranteed to see
732 * whatever the device wrote there.
733 */
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500734static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500735 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900737 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738
Eric Sesterhenn34814542006-03-24 18:47:11 +0100739 BUG_ON(dir == DMA_NONE);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500740
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900741 if (is_swiotlb_buffer(paddr)) {
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700742 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500743 return;
744 }
745
746 if (dir != DMA_FROM_DEVICE)
747 return;
748
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900749 /*
750 * phys_to_virt doesn't work with hihgmem page but we could
751 * call dma_mark_clean() with hihgmem page here. However, we
752 * are fine since dma_mark_clean() is null on POWERPC. We can
753 * make dma_mark_clean() take a physical address if necessary.
754 */
755 dma_mark_clean(phys_to_virt(paddr), size);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500756}
757
758void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
759 size_t size, enum dma_data_direction dir,
760 struct dma_attrs *attrs)
761{
762 unmap_single(hwdev, dev_addr, size, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763}
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900764EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900765
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766/*
767 * Make physical memory consistent for a single streaming mode DMA translation
768 * after a transfer.
769 *
Becky Bruceceb5ac32009-04-08 09:09:15 -0500770 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
Tony Luck17e5ad62005-09-29 15:52:13 -0700771 * using the cpu, yet do not wish to teardown the dma mapping, you must
772 * call this function before doing so. At the next point you give the dma
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773 * address back to the card, you must first perform a
774 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
775 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800776static void
John W. Linville8270f3f2005-09-29 14:43:32 -0700777swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400778 size_t size, enum dma_data_direction dir,
779 enum dma_sync_target target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900781 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782
Eric Sesterhenn34814542006-03-24 18:47:11 +0100783 BUG_ON(dir == DMA_NONE);
Becky Bruce380d6872009-04-08 09:09:20 -0500784
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900785 if (is_swiotlb_buffer(paddr)) {
Alexander Duyckfbfda892012-10-15 10:19:49 -0700786 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
Becky Bruce380d6872009-04-08 09:09:20 -0500787 return;
788 }
789
790 if (dir != DMA_FROM_DEVICE)
791 return;
792
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900793 dma_mark_clean(phys_to_virt(paddr), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794}
795
796void
John W. Linville8270f3f2005-09-29 14:43:32 -0700797swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900798 size_t size, enum dma_data_direction dir)
John W. Linville8270f3f2005-09-29 14:43:32 -0700799{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700800 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700801}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900802EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
John W. Linville8270f3f2005-09-29 14:43:32 -0700803
804void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900806 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700808 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900810EXPORT_SYMBOL(swiotlb_sync_single_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811
812/*
813 * Map a set of buffers described by scatterlist in streaming mode for DMA.
Becky Bruceceb5ac32009-04-08 09:09:15 -0500814 * This is the scatter-gather version of the above swiotlb_map_page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 * interface. Here the scatter gather list elements are each tagged with the
816 * appropriate dma address and length. They are obtained via
817 * sg_dma_{address,length}(SG).
818 *
819 * NOTE: An implementation may be able to use a smaller number of
820 * DMA address/length pairs than there are SG table elements.
821 * (for example via virtual mapping capabilities)
822 * The routine returns the number of addr/length pairs actually
823 * used, at most nents.
824 *
Becky Bruceceb5ac32009-04-08 09:09:15 -0500825 * Device ownership issues as mentioned above for swiotlb_map_page are the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 * same here.
827 */
828int
Arthur Kepner309df0c2008-04-29 01:00:32 -0700829swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900830 enum dma_data_direction dir, struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700831{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200832 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 int i;
834
Eric Sesterhenn34814542006-03-24 18:47:11 +0100835 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836
Jens Axboedbfd49f2007-05-11 14:56:18 +0200837 for_each_sg(sgl, sg, nelems, i) {
Ian Campbell961d7d02009-01-09 18:32:10 +0000838 phys_addr_t paddr = sg_phys(sg);
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900839 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
Becky Brucebc40ac62008-12-22 10:26:08 -0800840
FUJITA Tomonoricf56e3f2009-07-10 10:04:52 +0900841 if (swiotlb_force ||
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900842 !dma_capable(hwdev, dev_addr, sg->length)) {
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700843 phys_addr_t map = map_single(hwdev, sg_phys(sg),
844 sg->length, dir);
845 if (map == SWIOTLB_MAP_ERROR) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 /* Don't panic here, we expect map_sg users
847 to do proper error handling. */
848 swiotlb_full(hwdev, sg->length, dir, 0);
Arthur Kepner309df0c2008-04-29 01:00:32 -0700849 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
850 attrs);
Jens Axboedbfd49f2007-05-11 14:56:18 +0200851 sgl[0].dma_length = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 return 0;
853 }
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700854 sg->dma_address = phys_to_dma(hwdev, map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 } else
856 sg->dma_address = dev_addr;
857 sg->dma_length = sg->length;
858 }
859 return nelems;
860}
Arthur Kepner309df0c2008-04-29 01:00:32 -0700861EXPORT_SYMBOL(swiotlb_map_sg_attrs);
862
863int
864swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500865 enum dma_data_direction dir)
Arthur Kepner309df0c2008-04-29 01:00:32 -0700866{
867 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
868}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900869EXPORT_SYMBOL(swiotlb_map_sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870
871/*
872 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
Becky Bruceceb5ac32009-04-08 09:09:15 -0500873 * concerning calls here are the same as for swiotlb_unmap_page() above.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874 */
875void
Arthur Kepner309df0c2008-04-29 01:00:32 -0700876swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900877 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200879 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880 int i;
881
Eric Sesterhenn34814542006-03-24 18:47:11 +0100882 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500884 for_each_sg(sgl, sg, nelems, i)
885 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
886
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887}
Arthur Kepner309df0c2008-04-29 01:00:32 -0700888EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
889
890void
891swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500892 enum dma_data_direction dir)
Arthur Kepner309df0c2008-04-29 01:00:32 -0700893{
894 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
895}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900896EXPORT_SYMBOL(swiotlb_unmap_sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897
898/*
899 * Make physical memory consistent for a set of streaming mode DMA translations
900 * after a transfer.
901 *
902 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
903 * and usage.
904 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800905static void
Jens Axboedbfd49f2007-05-11 14:56:18 +0200906swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400907 int nelems, enum dma_data_direction dir,
908 enum dma_sync_target target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200910 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 int i;
912
Becky Bruce380d6872009-04-08 09:09:20 -0500913 for_each_sg(sgl, sg, nelems, i)
914 swiotlb_sync_single(hwdev, sg->dma_address,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700915 sg->dma_length, dir, target);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916}
917
918void
John W. Linville8270f3f2005-09-29 14:43:32 -0700919swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900920 int nelems, enum dma_data_direction dir)
John W. Linville8270f3f2005-09-29 14:43:32 -0700921{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700922 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700923}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900924EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
John W. Linville8270f3f2005-09-29 14:43:32 -0700925
926void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900928 int nelems, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700930 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900932EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933
934int
FUJITA Tomonori8d8bb392008-07-25 19:44:49 -0700935swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936{
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700937 return (dma_addr == phys_to_dma(hwdev, io_tlb_overflow_buffer));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900939EXPORT_SYMBOL(swiotlb_dma_mapping_error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940
941/*
Tony Luck17e5ad62005-09-29 15:52:13 -0700942 * Return whether the given device DMA address mask can be supported
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 * properly. For example, if your device can only drive the low 24-bits
Tony Luck17e5ad62005-09-29 15:52:13 -0700944 * during bus mastering, then you would pass 0x00ffffff as the mask to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 * this function.
946 */
947int
Jan Beulich563aaf02007-02-05 18:51:25 -0800948swiotlb_dma_supported(struct device *hwdev, u64 mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949{
Alexander Duyckc40dba02012-10-15 10:19:22 -0700950 return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952EXPORT_SYMBOL(swiotlb_dma_supported);