blob: 649d097853a105d8a1917a7c41bf4d0d37a0c453 [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
Thierry Redingce5be5a2013-10-23 13:32:04 +020041#define CREATE_TRACE_POINTS
Zoltan Kiss2b2b6142013-09-04 21:11:05 +010042#include <trace/events/swiotlb.h>
43
Linus Torvalds1da177e2005-04-16 15:20:36 -070044#define OFFSET(val,align) ((unsigned long) \
45 ( (val) & ( (align) - 1)))
46
Alex Williamson0b9afed2005-09-06 11:20:49 -060047#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
48
49/*
50 * Minimum IO TLB size to bother booting with. Systems with mainly
51 * 64bit capable cards will only lightly use the swiotlb. If we can't
52 * allocate a contiguous 1MB, we're probably in trouble anyway.
53 */
54#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
55
Linus Torvalds1da177e2005-04-16 15:20:36 -070056int swiotlb_force;
57
58/*
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -040059 * Used to do a quick range check in swiotlb_tbl_unmap_single and
60 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
Linus Torvalds1da177e2005-04-16 15:20:36 -070061 * API.
62 */
Alexander Duyckff7204a2012-10-15 10:19:28 -070063static phys_addr_t io_tlb_start, io_tlb_end;
Linus Torvalds1da177e2005-04-16 15:20:36 -070064
65/*
Uwe Kleine-Königb5950762010-11-01 15:38:34 -040066 * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
Linus Torvalds1da177e2005-04-16 15:20:36 -070067 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
68 */
69static unsigned long io_tlb_nslabs;
70
71/*
72 * When the IOMMU overflows we return a fallback buffer. This sets the size.
73 */
74static unsigned long io_tlb_overflow = 32*1024;
75
Alexander Duyckee3f6ba2012-10-15 10:19:34 -070076static phys_addr_t io_tlb_overflow_buffer;
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
78/*
79 * This is a free list describing the number of free entries available from
80 * each index
81 */
82static unsigned int *io_tlb_list;
83static unsigned int io_tlb_index;
84
85/*
86 * We need to save away the original address corresponding to a mapped entry
87 * for the sync operations.
88 */
Becky Brucebc40ac62008-12-22 10:26:08 -080089static phys_addr_t *io_tlb_orig_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -070090
91/*
92 * Protect the above data structures in the map and unmap calls
93 */
94static DEFINE_SPINLOCK(io_tlb_lock);
95
FUJITA Tomonori5740afd2009-11-10 19:46:18 +090096static int late_alloc;
97
Linus Torvalds1da177e2005-04-16 15:20:36 -070098static int __init
99setup_io_tlb_npages(char *str)
100{
101 if (isdigit(*str)) {
Alex Williamsone8579e72005-08-04 13:06:00 -0700102 io_tlb_nslabs = simple_strtoul(str, &str, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 /* avoid tail segment of size < IO_TLB_SEGSIZE */
104 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
105 }
106 if (*str == ',')
107 ++str;
FUJITA Tomonorib18485e2009-11-12 00:03:28 +0900108 if (!strcmp(str, "force"))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109 swiotlb_force = 1;
FUJITA Tomonorib18485e2009-11-12 00:03:28 +0900110
Yinghai Luc729de82013-04-15 22:23:45 -0700111 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112}
Yinghai Luc729de82013-04-15 22:23:45 -0700113early_param("swiotlb", setup_io_tlb_npages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114/* make io_tlb_overflow tunable too? */
115
Konrad Rzeszutek Wilkf21ffe92011-08-11 16:50:56 -0400116unsigned long swiotlb_nr_tbl(void)
FUJITA Tomonori5f98ecd2011-06-05 11:47:29 +0900117{
118 return io_tlb_nslabs;
119}
Konrad Rzeszutek Wilkf21ffe92011-08-11 16:50:56 -0400120EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
Yinghai Luc729de82013-04-15 22:23:45 -0700121
122/* default to 64MB */
123#define IO_TLB_DEFAULT_SIZE (64UL<<20)
124unsigned long swiotlb_size_or_default(void)
125{
126 unsigned long size;
127
128 size = io_tlb_nslabs << IO_TLB_SHIFT;
129
130 return size ? size : (IO_TLB_DEFAULT_SIZE);
131}
132
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900133/* Note that this doesn't work with highmem page */
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800134static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
135 volatile void *address)
Ian Campbelle08e1f72008-12-16 12:17:30 -0800136{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900137 return phys_to_dma(hwdev, virt_to_phys(address));
Ian Campbelle08e1f72008-12-16 12:17:30 -0800138}
139
Yinghai Luac2cbab2013-01-24 12:20:16 -0800140static bool no_iotlb_memory;
141
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900142void swiotlb_print_info(void)
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800143{
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900144 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700145 unsigned char *vstart, *vend;
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800146
Yinghai Luac2cbab2013-01-24 12:20:16 -0800147 if (no_iotlb_memory) {
148 pr_warn("software IO TLB: No low mem\n");
149 return;
150 }
151
Alexander Duyckff7204a2012-10-15 10:19:28 -0700152 vstart = phys_to_virt(io_tlb_start);
Alexander Duyckc40dba02012-10-15 10:19:22 -0700153 vend = phys_to_virt(io_tlb_end);
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800154
Bjorn Helgaas3af684c2012-05-29 15:06:29 -0700155 printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
Alexander Duyckff7204a2012-10-15 10:19:28 -0700156 (unsigned long long)io_tlb_start,
Alexander Duyckc40dba02012-10-15 10:19:22 -0700157 (unsigned long long)io_tlb_end,
Alexander Duyckff7204a2012-10-15 10:19:28 -0700158 bytes >> 20, vstart, vend - 1);
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800159}
160
Yinghai Luac2cbab2013-01-24 12:20:16 -0800161int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162{
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700163 void *v_overflow_buffer;
Jan Beulich563aaf02007-02-05 18:51:25 -0800164 unsigned long i, bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400166 bytes = nslabs << IO_TLB_SHIFT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400168 io_tlb_nslabs = nslabs;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700169 io_tlb_start = __pa(tlb);
170 io_tlb_end = io_tlb_start + bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171
172 /*
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700173 * Get the overflow emergency buffer
174 */
Yinghai Luad6492b2014-01-27 17:06:49 -0800175 v_overflow_buffer = memblock_virt_alloc_low_nopanic(
Santosh Shilimkar457ff1d2014-01-21 15:50:30 -0800176 PAGE_ALIGN(io_tlb_overflow),
177 PAGE_SIZE);
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700178 if (!v_overflow_buffer)
Yinghai Luac2cbab2013-01-24 12:20:16 -0800179 return -ENOMEM;
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700180
181 io_tlb_overflow_buffer = __pa(v_overflow_buffer);
182
183 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 * Allocate and initialize the free list array. This array is used
185 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
186 * between io_tlb_start and io_tlb_end.
187 */
Santosh Shilimkar457ff1d2014-01-21 15:50:30 -0800188 io_tlb_list = memblock_virt_alloc(
189 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)),
190 PAGE_SIZE);
Tony Luck25667d62007-03-06 13:31:45 -0800191 for (i = 0; i < io_tlb_nslabs; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
193 io_tlb_index = 0;
Santosh Shilimkar457ff1d2014-01-21 15:50:30 -0800194 io_tlb_orig_addr = memblock_virt_alloc(
195 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)),
196 PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900198 if (verbose)
199 swiotlb_print_info();
Yinghai Luac2cbab2013-01-24 12:20:16 -0800200
201 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202}
203
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400204/*
205 * Statically reserve bounce buffer space and initialize bounce buffer data
206 * structures for the software IO TLB used to implement the DMA API.
207 */
Yinghai Luac2cbab2013-01-24 12:20:16 -0800208void __init
209swiotlb_init(int verbose)
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400210{
Yinghai Luc729de82013-04-15 22:23:45 -0700211 size_t default_size = IO_TLB_DEFAULT_SIZE;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700212 unsigned char *vstart;
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400213 unsigned long bytes;
214
215 if (!io_tlb_nslabs) {
216 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
217 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
218 }
219
220 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
221
Yinghai Luac2cbab2013-01-24 12:20:16 -0800222 /* Get IO TLB memory from the low pages */
Yinghai Luad6492b2014-01-27 17:06:49 -0800223 vstart = memblock_virt_alloc_low_nopanic(PAGE_ALIGN(bytes), PAGE_SIZE);
Yinghai Luac2cbab2013-01-24 12:20:16 -0800224 if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose))
225 return;
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400226
Yinghai Luac2cbab2013-01-24 12:20:16 -0800227 if (io_tlb_start)
Santosh Shilimkar457ff1d2014-01-21 15:50:30 -0800228 memblock_free_early(io_tlb_start,
229 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
Yinghai Luac2cbab2013-01-24 12:20:16 -0800230 pr_warn("Cannot allocate SWIOTLB buffer");
231 no_iotlb_memory = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232}
233
Alex Williamson0b9afed2005-09-06 11:20:49 -0600234/*
235 * Systems with larger DMA zones (those that don't support ISA) can
236 * initialize the swiotlb later using the slab allocator if needed.
237 * This should be just like above, but with some error catching.
238 */
239int
Jan Beulich563aaf02007-02-05 18:51:25 -0800240swiotlb_late_init_with_default_size(size_t default_size)
Alex Williamson0b9afed2005-09-06 11:20:49 -0600241{
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400242 unsigned long bytes, req_nslabs = io_tlb_nslabs;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700243 unsigned char *vstart = NULL;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600244 unsigned int order;
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400245 int rc = 0;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600246
247 if (!io_tlb_nslabs) {
248 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
249 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
250 }
251
252 /*
253 * Get IO TLB memory from the low pages
254 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800255 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600256 io_tlb_nslabs = SLABS_PER_PAGE << order;
Jan Beulich563aaf02007-02-05 18:51:25 -0800257 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600258
259 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
Alexander Duyckff7204a2012-10-15 10:19:28 -0700260 vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
261 order);
262 if (vstart)
Alex Williamson0b9afed2005-09-06 11:20:49 -0600263 break;
264 order--;
265 }
266
Alexander Duyckff7204a2012-10-15 10:19:28 -0700267 if (!vstart) {
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400268 io_tlb_nslabs = req_nslabs;
269 return -ENOMEM;
270 }
Jan Beulich563aaf02007-02-05 18:51:25 -0800271 if (order != get_order(bytes)) {
Alex Williamson0b9afed2005-09-06 11:20:49 -0600272 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
273 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
274 io_tlb_nslabs = SLABS_PER_PAGE << order;
275 }
Alexander Duyckff7204a2012-10-15 10:19:28 -0700276 rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400277 if (rc)
Alexander Duyckff7204a2012-10-15 10:19:28 -0700278 free_pages((unsigned long)vstart, order);
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400279 return rc;
280}
281
282int
283swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
284{
285 unsigned long i, bytes;
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700286 unsigned char *v_overflow_buffer;
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400287
288 bytes = nslabs << IO_TLB_SHIFT;
289
290 io_tlb_nslabs = nslabs;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700291 io_tlb_start = virt_to_phys(tlb);
292 io_tlb_end = io_tlb_start + bytes;
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400293
Alexander Duyckff7204a2012-10-15 10:19:28 -0700294 memset(tlb, 0, bytes);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600295
296 /*
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700297 * Get the overflow emergency buffer
298 */
299 v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
300 get_order(io_tlb_overflow));
301 if (!v_overflow_buffer)
302 goto cleanup2;
303
304 io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
305
306 /*
Alex Williamson0b9afed2005-09-06 11:20:49 -0600307 * Allocate and initialize the free list array. This array is used
308 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
309 * between io_tlb_start and io_tlb_end.
310 */
311 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
312 get_order(io_tlb_nslabs * sizeof(int)));
313 if (!io_tlb_list)
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700314 goto cleanup3;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600315
316 for (i = 0; i < io_tlb_nslabs; i++)
317 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
318 io_tlb_index = 0;
319
Becky Brucebc40ac62008-12-22 10:26:08 -0800320 io_tlb_orig_addr = (phys_addr_t *)
321 __get_free_pages(GFP_KERNEL,
322 get_order(io_tlb_nslabs *
323 sizeof(phys_addr_t)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600324 if (!io_tlb_orig_addr)
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700325 goto cleanup4;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600326
Becky Brucebc40ac62008-12-22 10:26:08 -0800327 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600328
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900329 swiotlb_print_info();
Alex Williamson0b9afed2005-09-06 11:20:49 -0600330
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900331 late_alloc = 1;
332
Alex Williamson0b9afed2005-09-06 11:20:49 -0600333 return 0;
334
335cleanup4:
Tony Luck25667d62007-03-06 13:31:45 -0800336 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
337 sizeof(int)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600338 io_tlb_list = NULL;
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700339cleanup3:
340 free_pages((unsigned long)v_overflow_buffer,
341 get_order(io_tlb_overflow));
342 io_tlb_overflow_buffer = 0;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600343cleanup2:
Alexander Duyckc40dba02012-10-15 10:19:22 -0700344 io_tlb_end = 0;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700345 io_tlb_start = 0;
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400346 io_tlb_nslabs = 0;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600347 return -ENOMEM;
348}
349
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900350void __init swiotlb_free(void)
351{
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700352 if (!io_tlb_orig_addr)
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900353 return;
354
355 if (late_alloc) {
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700356 free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer),
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900357 get_order(io_tlb_overflow));
358 free_pages((unsigned long)io_tlb_orig_addr,
359 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
360 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
361 sizeof(int)));
Alexander Duyckff7204a2012-10-15 10:19:28 -0700362 free_pages((unsigned long)phys_to_virt(io_tlb_start),
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900363 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
364 } else {
Santosh Shilimkar457ff1d2014-01-21 15:50:30 -0800365 memblock_free_late(io_tlb_overflow_buffer,
366 PAGE_ALIGN(io_tlb_overflow));
367 memblock_free_late(__pa(io_tlb_orig_addr),
368 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
369 memblock_free_late(__pa(io_tlb_list),
370 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
371 memblock_free_late(io_tlb_start,
372 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900373 }
Konrad Rzeszutek Wilkf21ffe92011-08-11 16:50:56 -0400374 io_tlb_nslabs = 0;
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900375}
376
Akinobu Mita9c5a3622014-06-04 16:06:50 -0700377int is_swiotlb_buffer(phys_addr_t paddr)
FUJITA Tomonori640aebf2008-09-08 18:53:50 +0900378{
Alexander Duyckff7204a2012-10-15 10:19:28 -0700379 return paddr >= io_tlb_start && paddr < io_tlb_end;
FUJITA Tomonori640aebf2008-09-08 18:53:50 +0900380}
381
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382/*
Becky Brucefb05a372008-12-22 10:26:09 -0800383 * Bounce: copy the swiotlb buffer back to the original dma location
384 */
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700385static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
386 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387{
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700388 unsigned long pfn = PFN_DOWN(orig_addr);
389 unsigned char *vaddr = phys_to_virt(tlb_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390
Becky Brucefb05a372008-12-22 10:26:09 -0800391 if (PageHighMem(pfn_to_page(pfn))) {
392 /* The buffer does not have a mapping. Map it in and copy */
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700393 unsigned int offset = orig_addr & ~PAGE_MASK;
Becky Brucefb05a372008-12-22 10:26:09 -0800394 char *buffer;
395 unsigned int sz = 0;
396 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397
Becky Brucefb05a372008-12-22 10:26:09 -0800398 while (size) {
Becky Bruce67131ad2009-04-08 09:09:16 -0500399 sz = min_t(size_t, PAGE_SIZE - offset, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400
Becky Brucefb05a372008-12-22 10:26:09 -0800401 local_irq_save(flags);
Cong Wangc3eede82011-11-25 23:14:39 +0800402 buffer = kmap_atomic(pfn_to_page(pfn));
Becky Brucefb05a372008-12-22 10:26:09 -0800403 if (dir == DMA_TO_DEVICE)
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700404 memcpy(vaddr, buffer + offset, sz);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 else
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700406 memcpy(buffer + offset, vaddr, sz);
Cong Wangc3eede82011-11-25 23:14:39 +0800407 kunmap_atomic(buffer);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408 local_irq_restore(flags);
Becky Brucefb05a372008-12-22 10:26:09 -0800409
410 size -= sz;
411 pfn++;
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700412 vaddr += sz;
Becky Brucefb05a372008-12-22 10:26:09 -0800413 offset = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 }
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700415 } else if (dir == DMA_TO_DEVICE) {
416 memcpy(vaddr, phys_to_virt(orig_addr), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 } else {
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700418 memcpy(phys_to_virt(orig_addr), vaddr, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 }
420}
421
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700422phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
423 dma_addr_t tbl_dma_addr,
424 phys_addr_t orig_addr, size_t size,
425 enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426{
427 unsigned long flags;
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700428 phys_addr_t tlb_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429 unsigned int nslots, stride, index, wrap;
430 int i;
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800431 unsigned long mask;
432 unsigned long offset_slots;
433 unsigned long max_slots;
434
Yinghai Luac2cbab2013-01-24 12:20:16 -0800435 if (no_iotlb_memory)
436 panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
437
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800438 mask = dma_get_seg_boundary(hwdev);
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800439
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400440 tbl_dma_addr &= mask;
441
442 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
Ian Campbella5ddde4a2008-12-16 12:17:29 -0800443
444 /*
445 * Carefully handle integer overflow which can occur when mask == ~0UL.
446 */
Jan Beulichb15a38912008-03-13 09:13:30 +0000447 max_slots = mask + 1
448 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
449 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450
451 /*
452 * For mappings greater than a page, we limit the stride (and
453 * hence alignment) to a page size.
454 */
455 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
456 if (size > PAGE_SIZE)
457 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
458 else
459 stride = 1;
460
Eric Sesterhenn34814542006-03-24 18:47:11 +0100461 BUG_ON(!nslots);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462
463 /*
464 * Find suitable number of IO TLB entries size that will fit this
465 * request and allocate a buffer from that IO TLB pool.
466 */
467 spin_lock_irqsave(&io_tlb_lock, flags);
Andrew Mortona7133a12008-04-29 00:59:36 -0700468 index = ALIGN(io_tlb_index, stride);
469 if (index >= io_tlb_nslabs)
470 index = 0;
471 wrap = index;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472
Andrew Mortona7133a12008-04-29 00:59:36 -0700473 do {
FUJITA Tomonoria8522502008-04-29 00:59:36 -0700474 while (iommu_is_span_boundary(index, nslots, offset_slots,
475 max_slots)) {
Jan Beulichb15a38912008-03-13 09:13:30 +0000476 index += stride;
477 if (index >= io_tlb_nslabs)
478 index = 0;
Andrew Mortona7133a12008-04-29 00:59:36 -0700479 if (index == wrap)
480 goto not_found;
481 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482
Andrew Mortona7133a12008-04-29 00:59:36 -0700483 /*
484 * If we find a slot that indicates we have 'nslots' number of
485 * contiguous buffers, we allocate the buffers from that slot
486 * and mark the entries as '0' indicating unavailable.
487 */
488 if (io_tlb_list[index] >= nslots) {
489 int count = 0;
490
491 for (i = index; i < (int) (index + nslots); i++)
492 io_tlb_list[i] = 0;
493 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
494 io_tlb_list[i] = ++count;
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700495 tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
Andrew Mortona7133a12008-04-29 00:59:36 -0700496
497 /*
498 * Update the indices to avoid searching in the next
499 * round.
500 */
501 io_tlb_index = ((index + nslots) < io_tlb_nslabs
502 ? (index + nslots) : 0);
503
504 goto found;
505 }
506 index += stride;
507 if (index >= io_tlb_nslabs)
508 index = 0;
509 } while (index != wrap);
510
511not_found:
512 spin_unlock_irqrestore(&io_tlb_lock, flags);
Konrad Rzeszutek Wilk0cb637b2013-12-16 14:05:01 -0500513 if (printk_ratelimit())
514 dev_warn(hwdev, "swiotlb buffer is full (sz: %zd bytes)\n", size);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700515 return SWIOTLB_MAP_ERROR;
Andrew Mortona7133a12008-04-29 00:59:36 -0700516found:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 spin_unlock_irqrestore(&io_tlb_lock, flags);
518
519 /*
520 * Save away the mapping from the original address to the DMA address.
521 * This is needed when we sync the memory. Then we sync the buffer if
522 * needed.
523 */
Becky Brucebc40ac62008-12-22 10:26:08 -0800524 for (i = 0; i < nslots; i++)
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700525 io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700527 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700529 return tlb_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400531EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532
533/*
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400534 * Allocates bounce buffer and returns its kernel virtual address.
535 */
536
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700537phys_addr_t map_single(struct device *hwdev, phys_addr_t phys, size_t size,
538 enum dma_data_direction dir)
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400539{
Alexander Duyckff7204a2012-10-15 10:19:28 -0700540 dma_addr_t start_dma_addr = phys_to_dma(hwdev, io_tlb_start);
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400541
542 return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
543}
544
545/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
547 */
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700548void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
549 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550{
551 unsigned long flags;
552 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700553 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
554 phys_addr_t orig_addr = io_tlb_orig_addr[index];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555
556 /*
557 * First, sync the memory before unmapping the entry
558 */
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700559 if (orig_addr && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
560 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561
562 /*
563 * Return the buffer to the free list by setting the corresponding
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200564 * entries to indicate the number of contiguous entries available.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565 * While returning the entries to the free list, we merge the entries
566 * with slots below and above the pool being returned.
567 */
568 spin_lock_irqsave(&io_tlb_lock, flags);
569 {
570 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
571 io_tlb_list[index + nslots] : 0);
572 /*
573 * Step 1: return the slots to the free list, merging the
574 * slots with superceeding slots
575 */
576 for (i = index + nslots - 1; i >= index; i--)
577 io_tlb_list[i] = ++count;
578 /*
579 * Step 2: merge the returned slots with the preceding slots,
580 * if available (non zero)
581 */
582 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
583 io_tlb_list[i] = ++count;
584 }
585 spin_unlock_irqrestore(&io_tlb_lock, flags);
586}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400587EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588
Alexander Duyckfbfda892012-10-15 10:19:49 -0700589void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
590 size_t size, enum dma_data_direction dir,
591 enum dma_sync_target target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592{
Alexander Duyckfbfda892012-10-15 10:19:49 -0700593 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
594 phys_addr_t orig_addr = io_tlb_orig_addr[index];
Becky Brucebc40ac62008-12-22 10:26:08 -0800595
Alexander Duyckfbfda892012-10-15 10:19:49 -0700596 orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
Keir Fraserdf336d12007-07-21 04:37:24 -0700597
John W. Linvillede69e0f2005-09-29 14:44:57 -0700598 switch (target) {
599 case SYNC_FOR_CPU:
600 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700601 swiotlb_bounce(orig_addr, tlb_addr,
Alexander Duyckfbfda892012-10-15 10:19:49 -0700602 size, DMA_FROM_DEVICE);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100603 else
604 BUG_ON(dir != DMA_TO_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700605 break;
606 case SYNC_FOR_DEVICE:
607 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700608 swiotlb_bounce(orig_addr, tlb_addr,
Alexander Duyckfbfda892012-10-15 10:19:49 -0700609 size, DMA_TO_DEVICE);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100610 else
611 BUG_ON(dir != DMA_FROM_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700612 break;
613 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 BUG();
John W. Linvillede69e0f2005-09-29 14:44:57 -0700615 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400617EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618
619void *
620swiotlb_alloc_coherent(struct device *hwdev, size_t size,
Al Viro06a54492005-10-21 03:21:03 -0400621 dma_addr_t *dma_handle, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622{
Jan Beulich563aaf02007-02-05 18:51:25 -0800623 dma_addr_t dev_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 void *ret;
625 int order = get_order(size);
Yang Hongyang284901a2009-04-06 19:01:15 -0700626 u64 dma_mask = DMA_BIT_MASK(32);
FUJITA Tomonori1e74f302008-11-17 16:24:34 +0900627
628 if (hwdev && hwdev->coherent_dma_mask)
629 dma_mask = hwdev->coherent_dma_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630
Tony Luck25667d62007-03-06 13:31:45 -0800631 ret = (void *)__get_free_pages(flags, order);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700632 if (ret) {
633 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
634 if (dev_addr + size - 1 > dma_mask) {
635 /*
636 * The allocated memory isn't reachable by the device.
637 */
638 free_pages((unsigned long) ret, order);
639 ret = NULL;
640 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 }
642 if (!ret) {
643 /*
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400644 * We are either out of memory or the device can't DMA to
645 * GFP_DMA memory; fall back on map_single(), which
Becky Bruceceb5ac32009-04-08 09:09:15 -0500646 * will grab memory from the lowest available address range.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 */
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700648 phys_addr_t paddr = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
649 if (paddr == SWIOTLB_MAP_ERROR)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700652 ret = phys_to_virt(paddr);
653 dev_addr = phys_to_dma(hwdev, paddr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700655 /* Confirm address can be DMA'd by device */
656 if (dev_addr + size - 1 > dma_mask) {
657 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
658 (unsigned long long)dma_mask,
659 (unsigned long long)dev_addr);
FUJITA Tomonoria2b89b52008-10-23 18:42:03 +0900660
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700661 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
662 swiotlb_tbl_unmap_single(hwdev, paddr,
663 size, DMA_TO_DEVICE);
664 return NULL;
665 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666 }
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700667
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 *dma_handle = dev_addr;
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700669 memset(ret, 0, size);
670
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 return ret;
672}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900673EXPORT_SYMBOL(swiotlb_alloc_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674
675void
676swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900677 dma_addr_t dev_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900679 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900680
David Brownellaa248862007-08-10 13:10:27 -0700681 WARN_ON(irqs_disabled());
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900682 if (!is_swiotlb_buffer(paddr))
683 free_pages((unsigned long)vaddr, get_order(size));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 else
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400685 /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700686 swiotlb_tbl_unmap_single(hwdev, paddr, size, DMA_TO_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900688EXPORT_SYMBOL(swiotlb_free_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689
690static void
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500691swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
692 int do_panic)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693{
694 /*
695 * Ran out of IOMMU space for this operation. This is very bad.
696 * Unfortunately the drivers cannot handle this operation properly.
Tony Luck17e5ad62005-09-29 15:52:13 -0700697 * unless they check for dma_mapping_error (most don't)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 * When the mapping is small enough return a static buffer to limit
699 * the damage, or panic when the transfer is too big.
700 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800701 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
Kay Sievers94b32482009-01-06 10:44:37 -0800702 "device %s\n", size, dev ? dev_name(dev) : "?");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703
Casey Dahlinc7084b32009-08-20 16:27:56 -0700704 if (size <= io_tlb_overflow || !do_panic)
705 return;
706
707 if (dir == DMA_BIDIRECTIONAL)
708 panic("DMA: Random memory could be DMA accessed\n");
709 if (dir == DMA_FROM_DEVICE)
710 panic("DMA: Random memory could be DMA written\n");
711 if (dir == DMA_TO_DEVICE)
712 panic("DMA: Random memory could be DMA read\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713}
714
715/*
716 * Map a single buffer of the indicated size for DMA in streaming mode. The
Tony Luck17e5ad62005-09-29 15:52:13 -0700717 * physical address to use is returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718 *
719 * Once the device is given the dma address, the device owns this memory until
Becky Bruceceb5ac32009-04-08 09:09:15 -0500720 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 */
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900722dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
723 unsigned long offset, size_t size,
724 enum dma_data_direction dir,
725 struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726{
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700727 phys_addr_t map, phys = page_to_phys(page) + offset;
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900728 dma_addr_t dev_addr = phys_to_dma(dev, phys);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729
Eric Sesterhenn34814542006-03-24 18:47:11 +0100730 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731 /*
Becky Bruceceb5ac32009-04-08 09:09:15 -0500732 * If the address happens to be in the device's DMA window,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 * we can safely return the device addr and not worry about bounce
734 * buffering it.
735 */
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900736 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 return dev_addr;
738
Zoltan Kiss2b2b6142013-09-04 21:11:05 +0100739 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
740
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700741 /* Oh well, have to allocate and map a bounce buffer. */
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900742 map = map_single(dev, phys, size, dir);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700743 if (map == SWIOTLB_MAP_ERROR) {
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900744 swiotlb_full(dev, size, dir, 1);
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700745 return phys_to_dma(dev, io_tlb_overflow_buffer);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 }
747
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700748 dev_addr = phys_to_dma(dev, map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700750 /* Ensure that the address returned is DMA'ble */
FUJITA Tomonorifba99fa2011-02-25 14:44:16 -0800751 if (!dma_capable(dev, dev_addr, size)) {
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700752 swiotlb_tbl_unmap_single(dev, map, size, dir);
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700753 return phys_to_dma(dev, io_tlb_overflow_buffer);
FUJITA Tomonorifba99fa2011-02-25 14:44:16 -0800754 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755
756 return dev_addr;
757}
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900758EXPORT_SYMBOL_GPL(swiotlb_map_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759
760/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 * Unmap a single streaming mode DMA translation. The dma_addr and size must
Becky Bruceceb5ac32009-04-08 09:09:15 -0500762 * match what was provided for in a previous swiotlb_map_page call. All
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 * other usages are undefined.
764 *
765 * After this call, reads by the cpu to the buffer are guaranteed to see
766 * whatever the device wrote there.
767 */
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500768static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500769 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900771 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772
Eric Sesterhenn34814542006-03-24 18:47:11 +0100773 BUG_ON(dir == DMA_NONE);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500774
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900775 if (is_swiotlb_buffer(paddr)) {
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700776 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500777 return;
778 }
779
780 if (dir != DMA_FROM_DEVICE)
781 return;
782
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900783 /*
784 * phys_to_virt doesn't work with hihgmem page but we could
785 * call dma_mark_clean() with hihgmem page here. However, we
786 * are fine since dma_mark_clean() is null on POWERPC. We can
787 * make dma_mark_clean() take a physical address if necessary.
788 */
789 dma_mark_clean(phys_to_virt(paddr), size);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500790}
791
792void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
793 size_t size, enum dma_data_direction dir,
794 struct dma_attrs *attrs)
795{
796 unmap_single(hwdev, dev_addr, size, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797}
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900798EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900799
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800/*
801 * Make physical memory consistent for a single streaming mode DMA translation
802 * after a transfer.
803 *
Becky Bruceceb5ac32009-04-08 09:09:15 -0500804 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
Tony Luck17e5ad62005-09-29 15:52:13 -0700805 * using the cpu, yet do not wish to teardown the dma mapping, you must
806 * call this function before doing so. At the next point you give the dma
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 * address back to the card, you must first perform a
808 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
809 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800810static void
John W. Linville8270f3f2005-09-29 14:43:32 -0700811swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400812 size_t size, enum dma_data_direction dir,
813 enum dma_sync_target target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900815 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816
Eric Sesterhenn34814542006-03-24 18:47:11 +0100817 BUG_ON(dir == DMA_NONE);
Becky Bruce380d6872009-04-08 09:09:20 -0500818
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900819 if (is_swiotlb_buffer(paddr)) {
Alexander Duyckfbfda892012-10-15 10:19:49 -0700820 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
Becky Bruce380d6872009-04-08 09:09:20 -0500821 return;
822 }
823
824 if (dir != DMA_FROM_DEVICE)
825 return;
826
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900827 dma_mark_clean(phys_to_virt(paddr), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828}
829
830void
John W. Linville8270f3f2005-09-29 14:43:32 -0700831swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900832 size_t size, enum dma_data_direction dir)
John W. Linville8270f3f2005-09-29 14:43:32 -0700833{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700834 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700835}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900836EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
John W. Linville8270f3f2005-09-29 14:43:32 -0700837
838void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900840 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700842 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900844EXPORT_SYMBOL(swiotlb_sync_single_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845
846/*
847 * Map a set of buffers described by scatterlist in streaming mode for DMA.
Becky Bruceceb5ac32009-04-08 09:09:15 -0500848 * This is the scatter-gather version of the above swiotlb_map_page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849 * interface. Here the scatter gather list elements are each tagged with the
850 * appropriate dma address and length. They are obtained via
851 * sg_dma_{address,length}(SG).
852 *
853 * NOTE: An implementation may be able to use a smaller number of
854 * DMA address/length pairs than there are SG table elements.
855 * (for example via virtual mapping capabilities)
856 * The routine returns the number of addr/length pairs actually
857 * used, at most nents.
858 *
Becky Bruceceb5ac32009-04-08 09:09:15 -0500859 * Device ownership issues as mentioned above for swiotlb_map_page are the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 * same here.
861 */
862int
Arthur Kepner309df0c2008-04-29 01:00:32 -0700863swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900864 enum dma_data_direction dir, struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200866 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 int i;
868
Eric Sesterhenn34814542006-03-24 18:47:11 +0100869 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870
Jens Axboedbfd49f2007-05-11 14:56:18 +0200871 for_each_sg(sgl, sg, nelems, i) {
Ian Campbell961d7d02009-01-09 18:32:10 +0000872 phys_addr_t paddr = sg_phys(sg);
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900873 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
Becky Brucebc40ac62008-12-22 10:26:08 -0800874
FUJITA Tomonoricf56e3f2009-07-10 10:04:52 +0900875 if (swiotlb_force ||
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900876 !dma_capable(hwdev, dev_addr, sg->length)) {
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700877 phys_addr_t map = map_single(hwdev, sg_phys(sg),
878 sg->length, dir);
879 if (map == SWIOTLB_MAP_ERROR) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880 /* Don't panic here, we expect map_sg users
881 to do proper error handling. */
882 swiotlb_full(hwdev, sg->length, dir, 0);
Arthur Kepner309df0c2008-04-29 01:00:32 -0700883 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
884 attrs);
EunBong Song4d86ec72013-08-05 17:30:47 +0100885 sg_dma_len(sgl) = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886 return 0;
887 }
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700888 sg->dma_address = phys_to_dma(hwdev, map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 } else
890 sg->dma_address = dev_addr;
EunBong Song4d86ec72013-08-05 17:30:47 +0100891 sg_dma_len(sg) = sg->length;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 }
893 return nelems;
894}
Arthur Kepner309df0c2008-04-29 01:00:32 -0700895EXPORT_SYMBOL(swiotlb_map_sg_attrs);
896
897int
898swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500899 enum dma_data_direction dir)
Arthur Kepner309df0c2008-04-29 01:00:32 -0700900{
901 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
902}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900903EXPORT_SYMBOL(swiotlb_map_sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904
905/*
906 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
Becky Bruceceb5ac32009-04-08 09:09:15 -0500907 * concerning calls here are the same as for swiotlb_unmap_page() above.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908 */
909void
Arthur Kepner309df0c2008-04-29 01:00:32 -0700910swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900911 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200913 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 int i;
915
Eric Sesterhenn34814542006-03-24 18:47:11 +0100916 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500918 for_each_sg(sgl, sg, nelems, i)
EunBong Song4d86ec72013-08-05 17:30:47 +0100919 unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500920
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921}
Arthur Kepner309df0c2008-04-29 01:00:32 -0700922EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
923
924void
925swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500926 enum dma_data_direction dir)
Arthur Kepner309df0c2008-04-29 01:00:32 -0700927{
928 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
929}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900930EXPORT_SYMBOL(swiotlb_unmap_sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931
932/*
933 * Make physical memory consistent for a set of streaming mode DMA translations
934 * after a transfer.
935 *
936 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
937 * and usage.
938 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800939static void
Jens Axboedbfd49f2007-05-11 14:56:18 +0200940swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400941 int nelems, enum dma_data_direction dir,
942 enum dma_sync_target target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200944 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 int i;
946
Becky Bruce380d6872009-04-08 09:09:20 -0500947 for_each_sg(sgl, sg, nelems, i)
948 swiotlb_sync_single(hwdev, sg->dma_address,
EunBong Song4d86ec72013-08-05 17:30:47 +0100949 sg_dma_len(sg), dir, target);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950}
951
952void
John W. Linville8270f3f2005-09-29 14:43:32 -0700953swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900954 int nelems, enum dma_data_direction dir)
John W. Linville8270f3f2005-09-29 14:43:32 -0700955{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700956 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700957}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900958EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
John W. Linville8270f3f2005-09-29 14:43:32 -0700959
960void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900962 int nelems, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700964 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900966EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967
968int
FUJITA Tomonori8d8bb392008-07-25 19:44:49 -0700969swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970{
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700971 return (dma_addr == phys_to_dma(hwdev, io_tlb_overflow_buffer));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900973EXPORT_SYMBOL(swiotlb_dma_mapping_error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974
975/*
Tony Luck17e5ad62005-09-29 15:52:13 -0700976 * Return whether the given device DMA address mask can be supported
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 * properly. For example, if your device can only drive the low 24-bits
Tony Luck17e5ad62005-09-29 15:52:13 -0700978 * during bus mastering, then you would pass 0x00ffffff as the mask to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979 * this function.
980 */
981int
Jan Beulich563aaf02007-02-05 18:51:25 -0800982swiotlb_dma_supported(struct device *hwdev, u64 mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983{
Alexander Duyckc40dba02012-10-15 10:19:22 -0700984 return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986EXPORT_SYMBOL(swiotlb_dma_supported);