blob: 414f46ed1dcda9ae63be846cc6cf6e1b9bb7bb2e [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 */
60static char *io_tlb_start, *io_tlb_end;
61
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
FUJITA Tomonori03620b22010-08-02 23:48:06 +090073static void *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;
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800128 phys_addr_t pstart, pend;
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800129
130 pstart = virt_to_phys(io_tlb_start);
131 pend = virt_to_phys(io_tlb_end);
132
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800133 printk(KERN_INFO "Placing %luMB software IO TLB between %p - %p\n",
134 bytes >> 20, io_tlb_start, io_tlb_end);
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800135 printk(KERN_INFO "software IO TLB at phys %#llx - %#llx\n",
136 (unsigned long long)pstart,
137 (unsigned long long)pend);
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800138}
139
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400140void __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141{
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;
147 io_tlb_start = tlb;
Jan Beulich563aaf02007-02-05 18:51:25 -0800148 io_tlb_end = io_tlb_start + bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149
150 /*
151 * Allocate and initialize the free list array. This array is used
152 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
153 * between io_tlb_start and io_tlb_end.
154 */
Yinghai Lue79f86b2010-10-11 10:40:25 -0700155 io_tlb_list = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
Tony Luck25667d62007-03-06 13:31:45 -0800156 for (i = 0; i < io_tlb_nslabs; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
158 io_tlb_index = 0;
Yinghai Lue79f86b2010-10-11 10:40:25 -0700159 io_tlb_orig_addr = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160
161 /*
162 * Get the overflow emergency buffer
163 */
Yinghai Lue79f86b2010-10-11 10:40:25 -0700164 io_tlb_overflow_buffer = alloc_bootmem_low_pages(PAGE_ALIGN(io_tlb_overflow));
Jan Beulich563aaf02007-02-05 18:51:25 -0800165 if (!io_tlb_overflow_buffer)
166 panic("Cannot allocate SWIOTLB overflow buffer!\n");
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900167 if (verbose)
168 swiotlb_print_info();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169}
170
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400171/*
172 * Statically reserve bounce buffer space and initialize bounce buffer data
173 * structures for the software IO TLB used to implement the DMA API.
174 */
175void __init
176swiotlb_init_with_default_size(size_t default_size, int verbose)
177{
178 unsigned long bytes;
179
180 if (!io_tlb_nslabs) {
181 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
182 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
183 }
184
185 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
186
187 /*
188 * Get IO TLB memory from the low pages
189 */
Yinghai Lue79f86b2010-10-11 10:40:25 -0700190 io_tlb_start = alloc_bootmem_low_pages(PAGE_ALIGN(bytes));
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400191 if (!io_tlb_start)
192 panic("Cannot allocate SWIOTLB buffer");
193
194 swiotlb_init_with_tbl(io_tlb_start, io_tlb_nslabs, verbose);
195}
196
Jan Beulich563aaf02007-02-05 18:51:25 -0800197void __init
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900198swiotlb_init(int verbose)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199{
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900200 swiotlb_init_with_default_size(64 * (1<<20), verbose); /* default to 64MB */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201}
202
Alex Williamson0b9afed2005-09-06 11:20:49 -0600203/*
204 * Systems with larger DMA zones (those that don't support ISA) can
205 * initialize the swiotlb later using the slab allocator if needed.
206 * This should be just like above, but with some error catching.
207 */
208int
Jan Beulich563aaf02007-02-05 18:51:25 -0800209swiotlb_late_init_with_default_size(size_t default_size)
Alex Williamson0b9afed2005-09-06 11:20:49 -0600210{
Jan Beulich563aaf02007-02-05 18:51:25 -0800211 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600212 unsigned int order;
213
214 if (!io_tlb_nslabs) {
215 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
216 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
217 }
218
219 /*
220 * Get IO TLB memory from the low pages
221 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800222 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600223 io_tlb_nslabs = SLABS_PER_PAGE << order;
Jan Beulich563aaf02007-02-05 18:51:25 -0800224 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600225
226 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
FUJITA Tomonoribb521962009-07-10 10:04:51 +0900227 io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
228 order);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600229 if (io_tlb_start)
230 break;
231 order--;
232 }
233
234 if (!io_tlb_start)
235 goto cleanup1;
236
Jan Beulich563aaf02007-02-05 18:51:25 -0800237 if (order != get_order(bytes)) {
Alex Williamson0b9afed2005-09-06 11:20:49 -0600238 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
239 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
240 io_tlb_nslabs = SLABS_PER_PAGE << order;
Jan Beulich563aaf02007-02-05 18:51:25 -0800241 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600242 }
Jan Beulich563aaf02007-02-05 18:51:25 -0800243 io_tlb_end = io_tlb_start + bytes;
244 memset(io_tlb_start, 0, bytes);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600245
246 /*
247 * Allocate and initialize the free list array. This array is used
248 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
249 * between io_tlb_start and io_tlb_end.
250 */
251 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
252 get_order(io_tlb_nslabs * sizeof(int)));
253 if (!io_tlb_list)
254 goto cleanup2;
255
256 for (i = 0; i < io_tlb_nslabs; i++)
257 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
258 io_tlb_index = 0;
259
Becky Brucebc40ac62008-12-22 10:26:08 -0800260 io_tlb_orig_addr = (phys_addr_t *)
261 __get_free_pages(GFP_KERNEL,
262 get_order(io_tlb_nslabs *
263 sizeof(phys_addr_t)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600264 if (!io_tlb_orig_addr)
265 goto cleanup3;
266
Becky Brucebc40ac62008-12-22 10:26:08 -0800267 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600268
269 /*
270 * Get the overflow emergency buffer
271 */
272 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
273 get_order(io_tlb_overflow));
274 if (!io_tlb_overflow_buffer)
275 goto cleanup4;
276
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900277 swiotlb_print_info();
Alex Williamson0b9afed2005-09-06 11:20:49 -0600278
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900279 late_alloc = 1;
280
Alex Williamson0b9afed2005-09-06 11:20:49 -0600281 return 0;
282
283cleanup4:
Becky Brucebc40ac62008-12-22 10:26:08 -0800284 free_pages((unsigned long)io_tlb_orig_addr,
285 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600286 io_tlb_orig_addr = NULL;
287cleanup3:
Tony Luck25667d62007-03-06 13:31:45 -0800288 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
289 sizeof(int)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600290 io_tlb_list = NULL;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600291cleanup2:
Jan Beulich563aaf02007-02-05 18:51:25 -0800292 io_tlb_end = NULL;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600293 free_pages((unsigned long)io_tlb_start, order);
294 io_tlb_start = NULL;
295cleanup1:
296 io_tlb_nslabs = req_nslabs;
297 return -ENOMEM;
298}
299
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900300void __init swiotlb_free(void)
301{
302 if (!io_tlb_overflow_buffer)
303 return;
304
305 if (late_alloc) {
306 free_pages((unsigned long)io_tlb_overflow_buffer,
307 get_order(io_tlb_overflow));
308 free_pages((unsigned long)io_tlb_orig_addr,
309 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
310 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
311 sizeof(int)));
312 free_pages((unsigned long)io_tlb_start,
313 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
314 } else {
315 free_bootmem_late(__pa(io_tlb_overflow_buffer),
Yinghai Lue79f86b2010-10-11 10:40:25 -0700316 PAGE_ALIGN(io_tlb_overflow));
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900317 free_bootmem_late(__pa(io_tlb_orig_addr),
Yinghai Lue79f86b2010-10-11 10:40:25 -0700318 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900319 free_bootmem_late(__pa(io_tlb_list),
Yinghai Lue79f86b2010-10-11 10:40:25 -0700320 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900321 free_bootmem_late(__pa(io_tlb_start),
Yinghai Lue79f86b2010-10-11 10:40:25 -0700322 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900323 }
Konrad Rzeszutek Wilkf21ffe92011-08-11 16:50:56 -0400324 io_tlb_nslabs = 0;
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900325}
326
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900327static int is_swiotlb_buffer(phys_addr_t paddr)
FUJITA Tomonori640aebf2008-09-08 18:53:50 +0900328{
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900329 return paddr >= virt_to_phys(io_tlb_start) &&
330 paddr < virt_to_phys(io_tlb_end);
FUJITA Tomonori640aebf2008-09-08 18:53:50 +0900331}
332
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333/*
Becky Brucefb05a372008-12-22 10:26:09 -0800334 * Bounce: copy the swiotlb buffer back to the original dma location
335 */
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400336void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
337 enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338{
Becky Brucefb05a372008-12-22 10:26:09 -0800339 unsigned long pfn = PFN_DOWN(phys);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340
Becky Brucefb05a372008-12-22 10:26:09 -0800341 if (PageHighMem(pfn_to_page(pfn))) {
342 /* The buffer does not have a mapping. Map it in and copy */
343 unsigned int offset = phys & ~PAGE_MASK;
344 char *buffer;
345 unsigned int sz = 0;
346 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347
Becky Brucefb05a372008-12-22 10:26:09 -0800348 while (size) {
Becky Bruce67131ad2009-04-08 09:09:16 -0500349 sz = min_t(size_t, PAGE_SIZE - offset, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350
Becky Brucefb05a372008-12-22 10:26:09 -0800351 local_irq_save(flags);
Cong Wangc3eede82011-11-25 23:14:39 +0800352 buffer = kmap_atomic(pfn_to_page(pfn));
Becky Brucefb05a372008-12-22 10:26:09 -0800353 if (dir == DMA_TO_DEVICE)
354 memcpy(dma_addr, buffer + offset, sz);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355 else
Becky Brucefb05a372008-12-22 10:26:09 -0800356 memcpy(buffer + offset, dma_addr, sz);
Cong Wangc3eede82011-11-25 23:14:39 +0800357 kunmap_atomic(buffer);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358 local_irq_restore(flags);
Becky Brucefb05a372008-12-22 10:26:09 -0800359
360 size -= sz;
361 pfn++;
362 dma_addr += sz;
363 offset = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 }
365 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366 if (dir == DMA_TO_DEVICE)
Becky Brucefb05a372008-12-22 10:26:09 -0800367 memcpy(dma_addr, phys_to_virt(phys), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368 else
Becky Brucefb05a372008-12-22 10:26:09 -0800369 memcpy(phys_to_virt(phys), dma_addr, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370 }
371}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400372EXPORT_SYMBOL_GPL(swiotlb_bounce);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400374void *swiotlb_tbl_map_single(struct device *hwdev, dma_addr_t tbl_dma_addr,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500375 phys_addr_t phys, size_t size,
376 enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377{
378 unsigned long flags;
379 char *dma_addr;
380 unsigned int nslots, stride, index, wrap;
381 int i;
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800382 unsigned long mask;
383 unsigned long offset_slots;
384 unsigned long max_slots;
385
386 mask = dma_get_seg_boundary(hwdev);
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800387
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400388 tbl_dma_addr &= mask;
389
390 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
Ian Campbella5ddde4a2008-12-16 12:17:29 -0800391
392 /*
393 * Carefully handle integer overflow which can occur when mask == ~0UL.
394 */
Jan Beulichb15a38912008-03-13 09:13:30 +0000395 max_slots = mask + 1
396 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
397 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398
399 /*
400 * For mappings greater than a page, we limit the stride (and
401 * hence alignment) to a page size.
402 */
403 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
404 if (size > PAGE_SIZE)
405 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
406 else
407 stride = 1;
408
Eric Sesterhenn34814542006-03-24 18:47:11 +0100409 BUG_ON(!nslots);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410
411 /*
412 * Find suitable number of IO TLB entries size that will fit this
413 * request and allocate a buffer from that IO TLB pool.
414 */
415 spin_lock_irqsave(&io_tlb_lock, flags);
Andrew Mortona7133a12008-04-29 00:59:36 -0700416 index = ALIGN(io_tlb_index, stride);
417 if (index >= io_tlb_nslabs)
418 index = 0;
419 wrap = index;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420
Andrew Mortona7133a12008-04-29 00:59:36 -0700421 do {
FUJITA Tomonoria8522502008-04-29 00:59:36 -0700422 while (iommu_is_span_boundary(index, nslots, offset_slots,
423 max_slots)) {
Jan Beulichb15a38912008-03-13 09:13:30 +0000424 index += stride;
425 if (index >= io_tlb_nslabs)
426 index = 0;
Andrew Mortona7133a12008-04-29 00:59:36 -0700427 if (index == wrap)
428 goto not_found;
429 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430
Andrew Mortona7133a12008-04-29 00:59:36 -0700431 /*
432 * If we find a slot that indicates we have 'nslots' number of
433 * contiguous buffers, we allocate the buffers from that slot
434 * and mark the entries as '0' indicating unavailable.
435 */
436 if (io_tlb_list[index] >= nslots) {
437 int count = 0;
438
439 for (i = index; i < (int) (index + nslots); i++)
440 io_tlb_list[i] = 0;
441 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
442 io_tlb_list[i] = ++count;
443 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
444
445 /*
446 * Update the indices to avoid searching in the next
447 * round.
448 */
449 io_tlb_index = ((index + nslots) < io_tlb_nslabs
450 ? (index + nslots) : 0);
451
452 goto found;
453 }
454 index += stride;
455 if (index >= io_tlb_nslabs)
456 index = 0;
457 } while (index != wrap);
458
459not_found:
460 spin_unlock_irqrestore(&io_tlb_lock, flags);
461 return NULL;
462found:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 spin_unlock_irqrestore(&io_tlb_lock, flags);
464
465 /*
466 * Save away the mapping from the original address to the DMA address.
467 * This is needed when we sync the memory. Then we sync the buffer if
468 * needed.
469 */
Becky Brucebc40ac62008-12-22 10:26:08 -0800470 for (i = 0; i < nslots; i++)
471 io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
Becky Brucefb05a372008-12-22 10:26:09 -0800473 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474
475 return dma_addr;
476}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400477EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478
479/*
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400480 * Allocates bounce buffer and returns its kernel virtual address.
481 */
482
483static void *
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500484map_single(struct device *hwdev, phys_addr_t phys, size_t size,
485 enum dma_data_direction dir)
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400486{
487 dma_addr_t start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start);
488
489 return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
490}
491
492/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
494 */
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400495void
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400496swiotlb_tbl_unmap_single(struct device *hwdev, char *dma_addr, size_t size,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500497 enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498{
499 unsigned long flags;
500 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
501 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
Becky Brucebc40ac62008-12-22 10:26:08 -0800502 phys_addr_t phys = io_tlb_orig_addr[index];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503
504 /*
505 * First, sync the memory before unmapping the entry
506 */
Becky Brucebc40ac62008-12-22 10:26:08 -0800507 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
Becky Brucefb05a372008-12-22 10:26:09 -0800508 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509
510 /*
511 * Return the buffer to the free list by setting the corresponding
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200512 * entries to indicate the number of contiguous entries available.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 * While returning the entries to the free list, we merge the entries
514 * with slots below and above the pool being returned.
515 */
516 spin_lock_irqsave(&io_tlb_lock, flags);
517 {
518 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
519 io_tlb_list[index + nslots] : 0);
520 /*
521 * Step 1: return the slots to the free list, merging the
522 * slots with superceeding slots
523 */
524 for (i = index + nslots - 1; i >= index; i--)
525 io_tlb_list[i] = ++count;
526 /*
527 * Step 2: merge the returned slots with the preceding slots,
528 * if available (non zero)
529 */
530 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
531 io_tlb_list[i] = ++count;
532 }
533 spin_unlock_irqrestore(&io_tlb_lock, flags);
534}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400535EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400537void
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400538swiotlb_tbl_sync_single(struct device *hwdev, char *dma_addr, size_t size,
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400539 enum dma_data_direction dir,
540 enum dma_sync_target target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541{
Becky Brucebc40ac62008-12-22 10:26:08 -0800542 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
543 phys_addr_t phys = io_tlb_orig_addr[index];
544
545 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
Keir Fraserdf336d12007-07-21 04:37:24 -0700546
John W. Linvillede69e0f2005-09-29 14:44:57 -0700547 switch (target) {
548 case SYNC_FOR_CPU:
549 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
Becky Brucefb05a372008-12-22 10:26:09 -0800550 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100551 else
552 BUG_ON(dir != DMA_TO_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700553 break;
554 case SYNC_FOR_DEVICE:
555 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
Becky Brucefb05a372008-12-22 10:26:09 -0800556 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100557 else
558 BUG_ON(dir != DMA_FROM_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700559 break;
560 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 BUG();
John W. Linvillede69e0f2005-09-29 14:44:57 -0700562 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400564EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565
566void *
567swiotlb_alloc_coherent(struct device *hwdev, size_t size,
Al Viro06a54492005-10-21 03:21:03 -0400568 dma_addr_t *dma_handle, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569{
Jan Beulich563aaf02007-02-05 18:51:25 -0800570 dma_addr_t dev_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 void *ret;
572 int order = get_order(size);
Yang Hongyang284901a2009-04-06 19:01:15 -0700573 u64 dma_mask = DMA_BIT_MASK(32);
FUJITA Tomonori1e74f302008-11-17 16:24:34 +0900574
575 if (hwdev && hwdev->coherent_dma_mask)
576 dma_mask = hwdev->coherent_dma_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577
Tony Luck25667d62007-03-06 13:31:45 -0800578 ret = (void *)__get_free_pages(flags, order);
Jan Beulichac2b3e62009-12-15 16:47:43 -0800579 if (ret && swiotlb_virt_to_bus(hwdev, ret) + size - 1 > dma_mask) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 /*
581 * The allocated memory isn't reachable by the device.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 */
583 free_pages((unsigned long) ret, order);
584 ret = NULL;
585 }
586 if (!ret) {
587 /*
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400588 * We are either out of memory or the device can't DMA to
589 * GFP_DMA memory; fall back on map_single(), which
Becky Bruceceb5ac32009-04-08 09:09:15 -0500590 * will grab memory from the lowest available address range.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 */
Becky Brucebc40ac62008-12-22 10:26:08 -0800592 ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
FUJITA Tomonori9dfda122008-09-08 18:53:48 +0900593 if (!ret)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595 }
596
597 memset(ret, 0, size);
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800598 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599
600 /* Confirm address can be DMA'd by device */
Jan Beulichac2b3e62009-12-15 16:47:43 -0800601 if (dev_addr + size - 1 > dma_mask) {
Jan Beulich563aaf02007-02-05 18:51:25 -0800602 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
FUJITA Tomonori1e74f302008-11-17 16:24:34 +0900603 (unsigned long long)dma_mask,
Jan Beulich563aaf02007-02-05 18:51:25 -0800604 (unsigned long long)dev_addr);
FUJITA Tomonoria2b89b52008-10-23 18:42:03 +0900605
606 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400607 swiotlb_tbl_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
FUJITA Tomonoria2b89b52008-10-23 18:42:03 +0900608 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 }
610 *dma_handle = dev_addr;
611 return ret;
612}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900613EXPORT_SYMBOL(swiotlb_alloc_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614
615void
616swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900617 dma_addr_t dev_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900619 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900620
David Brownellaa248862007-08-10 13:10:27 -0700621 WARN_ON(irqs_disabled());
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900622 if (!is_swiotlb_buffer(paddr))
623 free_pages((unsigned long)vaddr, get_order(size));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 else
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400625 /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
626 swiotlb_tbl_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900628EXPORT_SYMBOL(swiotlb_free_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629
630static void
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500631swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
632 int do_panic)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633{
634 /*
635 * Ran out of IOMMU space for this operation. This is very bad.
636 * Unfortunately the drivers cannot handle this operation properly.
Tony Luck17e5ad62005-09-29 15:52:13 -0700637 * unless they check for dma_mapping_error (most don't)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 * When the mapping is small enough return a static buffer to limit
639 * the damage, or panic when the transfer is too big.
640 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800641 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
Kay Sievers94b32482009-01-06 10:44:37 -0800642 "device %s\n", size, dev ? dev_name(dev) : "?");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643
Casey Dahlinc7084b32009-08-20 16:27:56 -0700644 if (size <= io_tlb_overflow || !do_panic)
645 return;
646
647 if (dir == DMA_BIDIRECTIONAL)
648 panic("DMA: Random memory could be DMA accessed\n");
649 if (dir == DMA_FROM_DEVICE)
650 panic("DMA: Random memory could be DMA written\n");
651 if (dir == DMA_TO_DEVICE)
652 panic("DMA: Random memory could be DMA read\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653}
654
655/*
656 * Map a single buffer of the indicated size for DMA in streaming mode. The
Tony Luck17e5ad62005-09-29 15:52:13 -0700657 * physical address to use is returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 *
659 * Once the device is given the dma address, the device owns this memory until
Becky Bruceceb5ac32009-04-08 09:09:15 -0500660 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661 */
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900662dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
663 unsigned long offset, size_t size,
664 enum dma_data_direction dir,
665 struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666{
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900667 phys_addr_t phys = page_to_phys(page) + offset;
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900668 dma_addr_t dev_addr = phys_to_dma(dev, phys);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669 void *map;
670
Eric Sesterhenn34814542006-03-24 18:47:11 +0100671 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672 /*
Becky Bruceceb5ac32009-04-08 09:09:15 -0500673 * If the address happens to be in the device's DMA window,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674 * we can safely return the device addr and not worry about bounce
675 * buffering it.
676 */
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900677 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 return dev_addr;
679
680 /*
681 * Oh well, have to allocate and map a bounce buffer.
682 */
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900683 map = map_single(dev, phys, size, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 if (!map) {
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900685 swiotlb_full(dev, size, dir, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686 map = io_tlb_overflow_buffer;
687 }
688
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900689 dev_addr = swiotlb_virt_to_bus(dev, map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690
691 /*
692 * Ensure that the address returned is DMA'ble
693 */
FUJITA Tomonorifba99fa2011-02-25 14:44:16 -0800694 if (!dma_capable(dev, dev_addr, size)) {
695 swiotlb_tbl_unmap_single(dev, map, size, dir);
696 dev_addr = swiotlb_virt_to_bus(dev, io_tlb_overflow_buffer);
697 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698
699 return dev_addr;
700}
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900701EXPORT_SYMBOL_GPL(swiotlb_map_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702
703/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 * Unmap a single streaming mode DMA translation. The dma_addr and size must
Becky Bruceceb5ac32009-04-08 09:09:15 -0500705 * match what was provided for in a previous swiotlb_map_page call. All
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 * other usages are undefined.
707 *
708 * After this call, reads by the cpu to the buffer are guaranteed to see
709 * whatever the device wrote there.
710 */
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500711static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500712 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900714 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715
Eric Sesterhenn34814542006-03-24 18:47:11 +0100716 BUG_ON(dir == DMA_NONE);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500717
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900718 if (is_swiotlb_buffer(paddr)) {
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400719 swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500720 return;
721 }
722
723 if (dir != DMA_FROM_DEVICE)
724 return;
725
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900726 /*
727 * phys_to_virt doesn't work with hihgmem page but we could
728 * call dma_mark_clean() with hihgmem page here. However, we
729 * are fine since dma_mark_clean() is null on POWERPC. We can
730 * make dma_mark_clean() take a physical address if necessary.
731 */
732 dma_mark_clean(phys_to_virt(paddr), size);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500733}
734
735void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
736 size_t size, enum dma_data_direction dir,
737 struct dma_attrs *attrs)
738{
739 unmap_single(hwdev, dev_addr, size, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740}
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900741EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900742
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743/*
744 * Make physical memory consistent for a single streaming mode DMA translation
745 * after a transfer.
746 *
Becky Bruceceb5ac32009-04-08 09:09:15 -0500747 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
Tony Luck17e5ad62005-09-29 15:52:13 -0700748 * using the cpu, yet do not wish to teardown the dma mapping, you must
749 * call this function before doing so. At the next point you give the dma
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 * address back to the card, you must first perform a
751 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
752 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800753static void
John W. Linville8270f3f2005-09-29 14:43:32 -0700754swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400755 size_t size, enum dma_data_direction dir,
756 enum dma_sync_target target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900758 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759
Eric Sesterhenn34814542006-03-24 18:47:11 +0100760 BUG_ON(dir == DMA_NONE);
Becky Bruce380d6872009-04-08 09:09:20 -0500761
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900762 if (is_swiotlb_buffer(paddr)) {
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400763 swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
764 target);
Becky Bruce380d6872009-04-08 09:09:20 -0500765 return;
766 }
767
768 if (dir != DMA_FROM_DEVICE)
769 return;
770
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900771 dma_mark_clean(phys_to_virt(paddr), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772}
773
774void
John W. Linville8270f3f2005-09-29 14:43:32 -0700775swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900776 size_t size, enum dma_data_direction dir)
John W. Linville8270f3f2005-09-29 14:43:32 -0700777{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700778 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700779}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900780EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
John W. Linville8270f3f2005-09-29 14:43:32 -0700781
782void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900784 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700786 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900788EXPORT_SYMBOL(swiotlb_sync_single_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789
790/*
791 * Map a set of buffers described by scatterlist in streaming mode for DMA.
Becky Bruceceb5ac32009-04-08 09:09:15 -0500792 * This is the scatter-gather version of the above swiotlb_map_page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 * interface. Here the scatter gather list elements are each tagged with the
794 * appropriate dma address and length. They are obtained via
795 * sg_dma_{address,length}(SG).
796 *
797 * NOTE: An implementation may be able to use a smaller number of
798 * DMA address/length pairs than there are SG table elements.
799 * (for example via virtual mapping capabilities)
800 * The routine returns the number of addr/length pairs actually
801 * used, at most nents.
802 *
Becky Bruceceb5ac32009-04-08 09:09:15 -0500803 * Device ownership issues as mentioned above for swiotlb_map_page are the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 * same here.
805 */
806int
Arthur Kepner309df0c2008-04-29 01:00:32 -0700807swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900808 enum dma_data_direction dir, struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200810 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 int i;
812
Eric Sesterhenn34814542006-03-24 18:47:11 +0100813 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814
Jens Axboedbfd49f2007-05-11 14:56:18 +0200815 for_each_sg(sgl, sg, nelems, i) {
Ian Campbell961d7d02009-01-09 18:32:10 +0000816 phys_addr_t paddr = sg_phys(sg);
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900817 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
Becky Brucebc40ac62008-12-22 10:26:08 -0800818
FUJITA Tomonoricf56e3f2009-07-10 10:04:52 +0900819 if (swiotlb_force ||
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900820 !dma_capable(hwdev, dev_addr, sg->length)) {
Becky Brucebc40ac62008-12-22 10:26:08 -0800821 void *map = map_single(hwdev, sg_phys(sg),
822 sg->length, dir);
Andi Kleen7e870232005-12-20 14:45:19 +0100823 if (!map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824 /* Don't panic here, we expect map_sg users
825 to do proper error handling. */
826 swiotlb_full(hwdev, sg->length, dir, 0);
Arthur Kepner309df0c2008-04-29 01:00:32 -0700827 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
828 attrs);
Jens Axboedbfd49f2007-05-11 14:56:18 +0200829 sgl[0].dma_length = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 return 0;
831 }
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800832 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700833 } else
834 sg->dma_address = dev_addr;
835 sg->dma_length = sg->length;
836 }
837 return nelems;
838}
Arthur Kepner309df0c2008-04-29 01:00:32 -0700839EXPORT_SYMBOL(swiotlb_map_sg_attrs);
840
841int
842swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500843 enum dma_data_direction dir)
Arthur Kepner309df0c2008-04-29 01:00:32 -0700844{
845 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
846}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900847EXPORT_SYMBOL(swiotlb_map_sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848
849/*
850 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
Becky Bruceceb5ac32009-04-08 09:09:15 -0500851 * concerning calls here are the same as for swiotlb_unmap_page() above.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 */
853void
Arthur Kepner309df0c2008-04-29 01:00:32 -0700854swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900855 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200857 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858 int i;
859
Eric Sesterhenn34814542006-03-24 18:47:11 +0100860 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500862 for_each_sg(sgl, sg, nelems, i)
863 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
864
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865}
Arthur Kepner309df0c2008-04-29 01:00:32 -0700866EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
867
868void
869swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500870 enum dma_data_direction dir)
Arthur Kepner309df0c2008-04-29 01:00:32 -0700871{
872 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
873}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900874EXPORT_SYMBOL(swiotlb_unmap_sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875
876/*
877 * Make physical memory consistent for a set of streaming mode DMA translations
878 * after a transfer.
879 *
880 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
881 * and usage.
882 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800883static void
Jens Axboedbfd49f2007-05-11 14:56:18 +0200884swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400885 int nelems, enum dma_data_direction dir,
886 enum dma_sync_target target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200888 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 int i;
890
Becky Bruce380d6872009-04-08 09:09:20 -0500891 for_each_sg(sgl, sg, nelems, i)
892 swiotlb_sync_single(hwdev, sg->dma_address,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700893 sg->dma_length, dir, target);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894}
895
896void
John W. Linville8270f3f2005-09-29 14:43:32 -0700897swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900898 int nelems, enum dma_data_direction dir)
John W. Linville8270f3f2005-09-29 14:43:32 -0700899{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700900 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700901}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900902EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
John W. Linville8270f3f2005-09-29 14:43:32 -0700903
904void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900906 int nelems, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700908 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900910EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911
912int
FUJITA Tomonori8d8bb392008-07-25 19:44:49 -0700913swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914{
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800915 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900917EXPORT_SYMBOL(swiotlb_dma_mapping_error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918
919/*
Tony Luck17e5ad62005-09-29 15:52:13 -0700920 * Return whether the given device DMA address mask can be supported
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 * properly. For example, if your device can only drive the low 24-bits
Tony Luck17e5ad62005-09-29 15:52:13 -0700922 * during bus mastering, then you would pass 0x00ffffff as the mask to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 * this function.
924 */
925int
Jan Beulich563aaf02007-02-05 18:51:25 -0800926swiotlb_dma_supported(struct device *hwdev, u64 mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927{
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800928 return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930EXPORT_SYMBOL(swiotlb_dma_supported);