blob: 22e13a0e19d76a2b704edab2464f0c2a0e57783f [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>
Christoph Hellwig84be4562015-05-01 12:46:15 +020032#include <linux/scatterlist.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070033
34#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070035#include <asm/dma.h>
36
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 */
Jan Beulich8e0629c2014-06-02 14:58:25 +010089#define INVALID_PHYS_ADDR (~(phys_addr_t)0)
Becky Brucebc40ac62008-12-22 10:26:08 -080090static phys_addr_t *io_tlb_orig_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
92/*
93 * Protect the above data structures in the map and unmap calls
94 */
95static DEFINE_SPINLOCK(io_tlb_lock);
96
FUJITA Tomonori5740afd2009-11-10 19:46:18 +090097static int late_alloc;
98
Linus Torvalds1da177e2005-04-16 15:20:36 -070099static int __init
100setup_io_tlb_npages(char *str)
101{
102 if (isdigit(*str)) {
Alex Williamsone8579e72005-08-04 13:06:00 -0700103 io_tlb_nslabs = simple_strtoul(str, &str, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104 /* avoid tail segment of size < IO_TLB_SEGSIZE */
105 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
106 }
107 if (*str == ',')
108 ++str;
FUJITA Tomonorib18485e2009-11-12 00:03:28 +0900109 if (!strcmp(str, "force"))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110 swiotlb_force = 1;
FUJITA Tomonorib18485e2009-11-12 00:03:28 +0900111
Yinghai Luc729de82013-04-15 22:23:45 -0700112 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113}
Yinghai Luc729de82013-04-15 22:23:45 -0700114early_param("swiotlb", setup_io_tlb_npages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115/* make io_tlb_overflow tunable too? */
116
Konrad Rzeszutek Wilkf21ffe92011-08-11 16:50:56 -0400117unsigned long swiotlb_nr_tbl(void)
FUJITA Tomonori5f98ecd2011-06-05 11:47:29 +0900118{
119 return io_tlb_nslabs;
120}
Konrad Rzeszutek Wilkf21ffe92011-08-11 16:50:56 -0400121EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
Yinghai Luc729de82013-04-15 22:23:45 -0700122
123/* default to 64MB */
124#define IO_TLB_DEFAULT_SIZE (64UL<<20)
125unsigned long swiotlb_size_or_default(void)
126{
127 unsigned long size;
128
129 size = io_tlb_nslabs << IO_TLB_SHIFT;
130
131 return size ? size : (IO_TLB_DEFAULT_SIZE);
132}
133
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900134/* Note that this doesn't work with highmem page */
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800135static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
136 volatile void *address)
Ian Campbelle08e1f72008-12-16 12:17:30 -0800137{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900138 return phys_to_dma(hwdev, virt_to_phys(address));
Ian Campbelle08e1f72008-12-16 12:17:30 -0800139}
140
Yinghai Luac2cbab2013-01-24 12:20:16 -0800141static bool no_iotlb_memory;
142
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900143void swiotlb_print_info(void)
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800144{
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900145 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700146 unsigned char *vstart, *vend;
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800147
Yinghai Luac2cbab2013-01-24 12:20:16 -0800148 if (no_iotlb_memory) {
149 pr_warn("software IO TLB: No low mem\n");
150 return;
151 }
152
Alexander Duyckff7204a2012-10-15 10:19:28 -0700153 vstart = phys_to_virt(io_tlb_start);
Alexander Duyckc40dba02012-10-15 10:19:22 -0700154 vend = phys_to_virt(io_tlb_end);
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800155
Bjorn Helgaas3af684c2012-05-29 15:06:29 -0700156 printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
Alexander Duyckff7204a2012-10-15 10:19:28 -0700157 (unsigned long long)io_tlb_start,
Alexander Duyckc40dba02012-10-15 10:19:22 -0700158 (unsigned long long)io_tlb_end,
Alexander Duyckff7204a2012-10-15 10:19:28 -0700159 bytes >> 20, vstart, vend - 1);
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800160}
161
Yinghai Luac2cbab2013-01-24 12:20:16 -0800162int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163{
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700164 void *v_overflow_buffer;
Jan Beulich563aaf02007-02-05 18:51:25 -0800165 unsigned long i, bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400167 bytes = nslabs << IO_TLB_SHIFT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400169 io_tlb_nslabs = nslabs;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700170 io_tlb_start = __pa(tlb);
171 io_tlb_end = io_tlb_start + bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172
173 /*
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700174 * Get the overflow emergency buffer
175 */
Yinghai Luad6492b2014-01-27 17:06:49 -0800176 v_overflow_buffer = memblock_virt_alloc_low_nopanic(
Santosh Shilimkar457ff1d2014-01-21 15:50:30 -0800177 PAGE_ALIGN(io_tlb_overflow),
178 PAGE_SIZE);
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700179 if (!v_overflow_buffer)
Yinghai Luac2cbab2013-01-24 12:20:16 -0800180 return -ENOMEM;
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700181
182 io_tlb_overflow_buffer = __pa(v_overflow_buffer);
183
184 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185 * Allocate and initialize the free list array. This array is used
186 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
187 * between io_tlb_start and io_tlb_end.
188 */
Santosh Shilimkar457ff1d2014-01-21 15:50:30 -0800189 io_tlb_list = memblock_virt_alloc(
190 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)),
191 PAGE_SIZE);
Santosh Shilimkar457ff1d2014-01-21 15:50:30 -0800192 io_tlb_orig_addr = memblock_virt_alloc(
193 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)),
194 PAGE_SIZE);
Jan Beulich8e0629c2014-06-02 14:58:25 +0100195 for (i = 0; i < io_tlb_nslabs; i++) {
196 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
197 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
198 }
199 io_tlb_index = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900201 if (verbose)
202 swiotlb_print_info();
Yinghai Luac2cbab2013-01-24 12:20:16 -0800203
204 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205}
206
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400207/*
208 * Statically reserve bounce buffer space and initialize bounce buffer data
209 * structures for the software IO TLB used to implement the DMA API.
210 */
Yinghai Luac2cbab2013-01-24 12:20:16 -0800211void __init
212swiotlb_init(int verbose)
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400213{
Yinghai Luc729de82013-04-15 22:23:45 -0700214 size_t default_size = IO_TLB_DEFAULT_SIZE;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700215 unsigned char *vstart;
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400216 unsigned long bytes;
217
218 if (!io_tlb_nslabs) {
219 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
220 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
221 }
222
223 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
224
Yinghai Luac2cbab2013-01-24 12:20:16 -0800225 /* Get IO TLB memory from the low pages */
Yinghai Luad6492b2014-01-27 17:06:49 -0800226 vstart = memblock_virt_alloc_low_nopanic(PAGE_ALIGN(bytes), PAGE_SIZE);
Yinghai Luac2cbab2013-01-24 12:20:16 -0800227 if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose))
228 return;
FUJITA Tomonoriabbceff2010-05-10 15:15:12 -0400229
Yinghai Luac2cbab2013-01-24 12:20:16 -0800230 if (io_tlb_start)
Santosh Shilimkar457ff1d2014-01-21 15:50:30 -0800231 memblock_free_early(io_tlb_start,
232 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
Yinghai Luac2cbab2013-01-24 12:20:16 -0800233 pr_warn("Cannot allocate SWIOTLB buffer");
234 no_iotlb_memory = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235}
236
Alex Williamson0b9afed2005-09-06 11:20:49 -0600237/*
238 * Systems with larger DMA zones (those that don't support ISA) can
239 * initialize the swiotlb later using the slab allocator if needed.
240 * This should be just like above, but with some error catching.
241 */
242int
Jan Beulich563aaf02007-02-05 18:51:25 -0800243swiotlb_late_init_with_default_size(size_t default_size)
Alex Williamson0b9afed2005-09-06 11:20:49 -0600244{
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400245 unsigned long bytes, req_nslabs = io_tlb_nslabs;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700246 unsigned char *vstart = NULL;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600247 unsigned int order;
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400248 int rc = 0;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600249
250 if (!io_tlb_nslabs) {
251 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
252 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
253 }
254
255 /*
256 * Get IO TLB memory from the low pages
257 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800258 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600259 io_tlb_nslabs = SLABS_PER_PAGE << order;
Jan Beulich563aaf02007-02-05 18:51:25 -0800260 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600261
262 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
Alexander Duyckff7204a2012-10-15 10:19:28 -0700263 vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
264 order);
265 if (vstart)
Alex Williamson0b9afed2005-09-06 11:20:49 -0600266 break;
267 order--;
268 }
269
Alexander Duyckff7204a2012-10-15 10:19:28 -0700270 if (!vstart) {
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400271 io_tlb_nslabs = req_nslabs;
272 return -ENOMEM;
273 }
Jan Beulich563aaf02007-02-05 18:51:25 -0800274 if (order != get_order(bytes)) {
Alex Williamson0b9afed2005-09-06 11:20:49 -0600275 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
276 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
277 io_tlb_nslabs = SLABS_PER_PAGE << order;
278 }
Alexander Duyckff7204a2012-10-15 10:19:28 -0700279 rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400280 if (rc)
Alexander Duyckff7204a2012-10-15 10:19:28 -0700281 free_pages((unsigned long)vstart, order);
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400282 return rc;
283}
284
285int
286swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
287{
288 unsigned long i, bytes;
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700289 unsigned char *v_overflow_buffer;
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400290
291 bytes = nslabs << IO_TLB_SHIFT;
292
293 io_tlb_nslabs = nslabs;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700294 io_tlb_start = virt_to_phys(tlb);
295 io_tlb_end = io_tlb_start + bytes;
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400296
Alexander Duyckff7204a2012-10-15 10:19:28 -0700297 memset(tlb, 0, bytes);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600298
299 /*
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700300 * Get the overflow emergency buffer
301 */
302 v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
303 get_order(io_tlb_overflow));
304 if (!v_overflow_buffer)
305 goto cleanup2;
306
307 io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
308
309 /*
Alex Williamson0b9afed2005-09-06 11:20:49 -0600310 * Allocate and initialize the free list array. This array is used
311 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
312 * between io_tlb_start and io_tlb_end.
313 */
314 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
315 get_order(io_tlb_nslabs * sizeof(int)));
316 if (!io_tlb_list)
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700317 goto cleanup3;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600318
Becky Brucebc40ac62008-12-22 10:26:08 -0800319 io_tlb_orig_addr = (phys_addr_t *)
320 __get_free_pages(GFP_KERNEL,
321 get_order(io_tlb_nslabs *
322 sizeof(phys_addr_t)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600323 if (!io_tlb_orig_addr)
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700324 goto cleanup4;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600325
Jan Beulich8e0629c2014-06-02 14:58:25 +0100326 for (i = 0; i < io_tlb_nslabs; i++) {
327 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
328 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
329 }
330 io_tlb_index = 0;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600331
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900332 swiotlb_print_info();
Alex Williamson0b9afed2005-09-06 11:20:49 -0600333
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900334 late_alloc = 1;
335
Alex Williamson0b9afed2005-09-06 11:20:49 -0600336 return 0;
337
338cleanup4:
Tony Luck25667d62007-03-06 13:31:45 -0800339 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
340 sizeof(int)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600341 io_tlb_list = NULL;
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700342cleanup3:
343 free_pages((unsigned long)v_overflow_buffer,
344 get_order(io_tlb_overflow));
345 io_tlb_overflow_buffer = 0;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600346cleanup2:
Alexander Duyckc40dba02012-10-15 10:19:22 -0700347 io_tlb_end = 0;
Alexander Duyckff7204a2012-10-15 10:19:28 -0700348 io_tlb_start = 0;
Konrad Rzeszutek Wilk74838b72012-07-27 20:55:27 -0400349 io_tlb_nslabs = 0;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600350 return -ENOMEM;
351}
352
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900353void __init swiotlb_free(void)
354{
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700355 if (!io_tlb_orig_addr)
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900356 return;
357
358 if (late_alloc) {
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700359 free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer),
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900360 get_order(io_tlb_overflow));
361 free_pages((unsigned long)io_tlb_orig_addr,
362 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
363 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
364 sizeof(int)));
Alexander Duyckff7204a2012-10-15 10:19:28 -0700365 free_pages((unsigned long)phys_to_virt(io_tlb_start),
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900366 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
367 } else {
Santosh Shilimkar457ff1d2014-01-21 15:50:30 -0800368 memblock_free_late(io_tlb_overflow_buffer,
369 PAGE_ALIGN(io_tlb_overflow));
370 memblock_free_late(__pa(io_tlb_orig_addr),
371 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
372 memblock_free_late(__pa(io_tlb_list),
373 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
374 memblock_free_late(io_tlb_start,
375 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900376 }
Konrad Rzeszutek Wilkf21ffe92011-08-11 16:50:56 -0400377 io_tlb_nslabs = 0;
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900378}
379
Akinobu Mita9c5a3622014-06-04 16:06:50 -0700380int is_swiotlb_buffer(phys_addr_t paddr)
FUJITA Tomonori640aebf2008-09-08 18:53:50 +0900381{
Alexander Duyckff7204a2012-10-15 10:19:28 -0700382 return paddr >= io_tlb_start && paddr < io_tlb_end;
FUJITA Tomonori640aebf2008-09-08 18:53:50 +0900383}
384
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385/*
Becky Brucefb05a372008-12-22 10:26:09 -0800386 * Bounce: copy the swiotlb buffer back to the original dma location
387 */
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700388static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
389 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390{
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700391 unsigned long pfn = PFN_DOWN(orig_addr);
392 unsigned char *vaddr = phys_to_virt(tlb_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393
Becky Brucefb05a372008-12-22 10:26:09 -0800394 if (PageHighMem(pfn_to_page(pfn))) {
395 /* The buffer does not have a mapping. Map it in and copy */
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700396 unsigned int offset = orig_addr & ~PAGE_MASK;
Becky Brucefb05a372008-12-22 10:26:09 -0800397 char *buffer;
398 unsigned int sz = 0;
399 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400
Becky Brucefb05a372008-12-22 10:26:09 -0800401 while (size) {
Becky Bruce67131ad2009-04-08 09:09:16 -0500402 sz = min_t(size_t, PAGE_SIZE - offset, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403
Becky Brucefb05a372008-12-22 10:26:09 -0800404 local_irq_save(flags);
Cong Wangc3eede82011-11-25 23:14:39 +0800405 buffer = kmap_atomic(pfn_to_page(pfn));
Becky Brucefb05a372008-12-22 10:26:09 -0800406 if (dir == DMA_TO_DEVICE)
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700407 memcpy(vaddr, buffer + offset, sz);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408 else
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700409 memcpy(buffer + offset, vaddr, sz);
Cong Wangc3eede82011-11-25 23:14:39 +0800410 kunmap_atomic(buffer);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411 local_irq_restore(flags);
Becky Brucefb05a372008-12-22 10:26:09 -0800412
413 size -= sz;
414 pfn++;
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700415 vaddr += sz;
Becky Brucefb05a372008-12-22 10:26:09 -0800416 offset = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 }
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700418 } else if (dir == DMA_TO_DEVICE) {
419 memcpy(vaddr, phys_to_virt(orig_addr), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 } else {
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700421 memcpy(phys_to_virt(orig_addr), vaddr, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 }
423}
424
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700425phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
426 dma_addr_t tbl_dma_addr,
427 phys_addr_t orig_addr, size_t size,
428 enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429{
430 unsigned long flags;
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700431 phys_addr_t tlb_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 unsigned int nslots, stride, index, wrap;
433 int i;
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800434 unsigned long mask;
435 unsigned long offset_slots;
436 unsigned long max_slots;
437
Yinghai Luac2cbab2013-01-24 12:20:16 -0800438 if (no_iotlb_memory)
439 panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
440
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800441 mask = dma_get_seg_boundary(hwdev);
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800442
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400443 tbl_dma_addr &= mask;
444
445 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
Ian Campbella5ddde4a2008-12-16 12:17:29 -0800446
447 /*
448 * Carefully handle integer overflow which can occur when mask == ~0UL.
449 */
Jan Beulichb15a38912008-03-13 09:13:30 +0000450 max_slots = mask + 1
451 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
452 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453
454 /*
455 * For mappings greater than a page, we limit the stride (and
456 * hence alignment) to a page size.
457 */
458 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
459 if (size > PAGE_SIZE)
460 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
461 else
462 stride = 1;
463
Eric Sesterhenn34814542006-03-24 18:47:11 +0100464 BUG_ON(!nslots);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465
466 /*
467 * Find suitable number of IO TLB entries size that will fit this
468 * request and allocate a buffer from that IO TLB pool.
469 */
470 spin_lock_irqsave(&io_tlb_lock, flags);
Andrew Mortona7133a12008-04-29 00:59:36 -0700471 index = ALIGN(io_tlb_index, stride);
472 if (index >= io_tlb_nslabs)
473 index = 0;
474 wrap = index;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475
Andrew Mortona7133a12008-04-29 00:59:36 -0700476 do {
FUJITA Tomonoria8522502008-04-29 00:59:36 -0700477 while (iommu_is_span_boundary(index, nslots, offset_slots,
478 max_slots)) {
Jan Beulichb15a38912008-03-13 09:13:30 +0000479 index += stride;
480 if (index >= io_tlb_nslabs)
481 index = 0;
Andrew Mortona7133a12008-04-29 00:59:36 -0700482 if (index == wrap)
483 goto not_found;
484 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485
Andrew Mortona7133a12008-04-29 00:59:36 -0700486 /*
487 * If we find a slot that indicates we have 'nslots' number of
488 * contiguous buffers, we allocate the buffers from that slot
489 * and mark the entries as '0' indicating unavailable.
490 */
491 if (io_tlb_list[index] >= nslots) {
492 int count = 0;
493
494 for (i = index; i < (int) (index + nslots); i++)
495 io_tlb_list[i] = 0;
496 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
497 io_tlb_list[i] = ++count;
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700498 tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
Andrew Mortona7133a12008-04-29 00:59:36 -0700499
500 /*
501 * Update the indices to avoid searching in the next
502 * round.
503 */
504 io_tlb_index = ((index + nslots) < io_tlb_nslabs
505 ? (index + nslots) : 0);
506
507 goto found;
508 }
509 index += stride;
510 if (index >= io_tlb_nslabs)
511 index = 0;
512 } while (index != wrap);
513
514not_found:
515 spin_unlock_irqrestore(&io_tlb_lock, flags);
Konrad Rzeszutek Wilk0cb637b2013-12-16 14:05:01 -0500516 if (printk_ratelimit())
517 dev_warn(hwdev, "swiotlb buffer is full (sz: %zd bytes)\n", size);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700518 return SWIOTLB_MAP_ERROR;
Andrew Mortona7133a12008-04-29 00:59:36 -0700519found:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 spin_unlock_irqrestore(&io_tlb_lock, flags);
521
522 /*
523 * Save away the mapping from the original address to the DMA address.
524 * This is needed when we sync the memory. Then we sync the buffer if
525 * needed.
526 */
Becky Brucebc40ac62008-12-22 10:26:08 -0800527 for (i = 0; i < nslots; i++)
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700528 io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700530 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700532 return tlb_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400534EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535
536/*
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400537 * Allocates bounce buffer and returns its kernel virtual address.
538 */
539
Alexandre Courbot023600f2015-05-26 18:02:54 +0900540static phys_addr_t
541map_single(struct device *hwdev, phys_addr_t phys, size_t size,
542 enum dma_data_direction dir)
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400543{
Alexander Duyckff7204a2012-10-15 10:19:28 -0700544 dma_addr_t start_dma_addr = phys_to_dma(hwdev, io_tlb_start);
FUJITA Tomonorieb605a52010-05-10 15:14:54 -0400545
546 return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
547}
548
549/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
551 */
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700552void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
553 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554{
555 unsigned long flags;
556 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700557 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
558 phys_addr_t orig_addr = io_tlb_orig_addr[index];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559
560 /*
561 * First, sync the memory before unmapping the entry
562 */
Jan Beulich8e0629c2014-06-02 14:58:25 +0100563 if (orig_addr != INVALID_PHYS_ADDR &&
564 ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700565 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566
567 /*
568 * Return the buffer to the free list by setting the corresponding
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200569 * entries to indicate the number of contiguous entries available.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 * While returning the entries to the free list, we merge the entries
571 * with slots below and above the pool being returned.
572 */
573 spin_lock_irqsave(&io_tlb_lock, flags);
574 {
575 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
576 io_tlb_list[index + nslots] : 0);
577 /*
578 * Step 1: return the slots to the free list, merging the
579 * slots with superceeding slots
580 */
Jan Beulich8e0629c2014-06-02 14:58:25 +0100581 for (i = index + nslots - 1; i >= index; i--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 io_tlb_list[i] = ++count;
Jan Beulich8e0629c2014-06-02 14:58:25 +0100583 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
584 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 /*
586 * Step 2: merge the returned slots with the preceding slots,
587 * if available (non zero)
588 */
589 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
590 io_tlb_list[i] = ++count;
591 }
592 spin_unlock_irqrestore(&io_tlb_lock, flags);
593}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400594EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595
Alexander Duyckfbfda892012-10-15 10:19:49 -0700596void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
597 size_t size, enum dma_data_direction dir,
598 enum dma_sync_target target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599{
Alexander Duyckfbfda892012-10-15 10:19:49 -0700600 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
601 phys_addr_t orig_addr = io_tlb_orig_addr[index];
Becky Brucebc40ac62008-12-22 10:26:08 -0800602
Jan Beulich8e0629c2014-06-02 14:58:25 +0100603 if (orig_addr == INVALID_PHYS_ADDR)
604 return;
Alexander Duyckfbfda892012-10-15 10:19:49 -0700605 orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
Keir Fraserdf336d12007-07-21 04:37:24 -0700606
John W. Linvillede69e0f2005-09-29 14:44:57 -0700607 switch (target) {
608 case SYNC_FOR_CPU:
609 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700610 swiotlb_bounce(orig_addr, tlb_addr,
Alexander Duyckfbfda892012-10-15 10:19:49 -0700611 size, DMA_FROM_DEVICE);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100612 else
613 BUG_ON(dir != DMA_TO_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700614 break;
615 case SYNC_FOR_DEVICE:
616 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
Alexander Duyckaf51a9f2012-10-15 10:19:55 -0700617 swiotlb_bounce(orig_addr, tlb_addr,
Alexander Duyckfbfda892012-10-15 10:19:49 -0700618 size, DMA_TO_DEVICE);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100619 else
620 BUG_ON(dir != DMA_FROM_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700621 break;
622 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623 BUG();
John W. Linvillede69e0f2005-09-29 14:44:57 -0700624 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625}
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400626EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627
628void *
629swiotlb_alloc_coherent(struct device *hwdev, size_t size,
Al Viro06a54492005-10-21 03:21:03 -0400630 dma_addr_t *dma_handle, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631{
Jan Beulich563aaf02007-02-05 18:51:25 -0800632 dma_addr_t dev_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700633 void *ret;
634 int order = get_order(size);
Yang Hongyang284901a2009-04-06 19:01:15 -0700635 u64 dma_mask = DMA_BIT_MASK(32);
FUJITA Tomonori1e74f302008-11-17 16:24:34 +0900636
637 if (hwdev && hwdev->coherent_dma_mask)
638 dma_mask = hwdev->coherent_dma_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
Tony Luck25667d62007-03-06 13:31:45 -0800640 ret = (void *)__get_free_pages(flags, order);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700641 if (ret) {
642 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
643 if (dev_addr + size - 1 > dma_mask) {
644 /*
645 * The allocated memory isn't reachable by the device.
646 */
647 free_pages((unsigned long) ret, order);
648 ret = NULL;
649 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650 }
651 if (!ret) {
652 /*
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400653 * We are either out of memory or the device can't DMA to
654 * GFP_DMA memory; fall back on map_single(), which
Becky Bruceceb5ac32009-04-08 09:09:15 -0500655 * will grab memory from the lowest available address range.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 */
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700657 phys_addr_t paddr = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
658 if (paddr == SWIOTLB_MAP_ERROR)
Joerg Roedel94cc81f2015-06-10 17:49:40 +0200659 goto err_warn;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700661 ret = phys_to_virt(paddr);
662 dev_addr = phys_to_dma(hwdev, paddr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700664 /* Confirm address can be DMA'd by device */
665 if (dev_addr + size - 1 > dma_mask) {
666 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
667 (unsigned long long)dma_mask,
668 (unsigned long long)dev_addr);
FUJITA Tomonoria2b89b52008-10-23 18:42:03 +0900669
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700670 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
671 swiotlb_tbl_unmap_single(hwdev, paddr,
672 size, DMA_TO_DEVICE);
Joerg Roedel94cc81f2015-06-10 17:49:40 +0200673 goto err_warn;
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700674 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 }
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700676
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 *dma_handle = dev_addr;
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700678 memset(ret, 0, size);
679
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 return ret;
Joerg Roedel94cc81f2015-06-10 17:49:40 +0200681
682err_warn:
683 pr_warn("swiotlb: coherent allocation failed for device %s size=%zu\n",
684 dev_name(hwdev), size);
685 dump_stack();
686
687 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900689EXPORT_SYMBOL(swiotlb_alloc_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690
691void
692swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900693 dma_addr_t dev_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900695 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900696
David Brownellaa248862007-08-10 13:10:27 -0700697 WARN_ON(irqs_disabled());
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900698 if (!is_swiotlb_buffer(paddr))
699 free_pages((unsigned long)vaddr, get_order(size));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 else
Konrad Rzeszutek Wilkbfc55012010-05-10 15:54:20 -0400701 /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700702 swiotlb_tbl_unmap_single(hwdev, paddr, size, DMA_TO_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900704EXPORT_SYMBOL(swiotlb_free_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705
706static void
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500707swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
708 int do_panic)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709{
710 /*
711 * Ran out of IOMMU space for this operation. This is very bad.
712 * Unfortunately the drivers cannot handle this operation properly.
Tony Luck17e5ad62005-09-29 15:52:13 -0700713 * unless they check for dma_mapping_error (most don't)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 * When the mapping is small enough return a static buffer to limit
715 * the damage, or panic when the transfer is too big.
716 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800717 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
Kay Sievers94b32482009-01-06 10:44:37 -0800718 "device %s\n", size, dev ? dev_name(dev) : "?");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719
Casey Dahlinc7084b32009-08-20 16:27:56 -0700720 if (size <= io_tlb_overflow || !do_panic)
721 return;
722
723 if (dir == DMA_BIDIRECTIONAL)
724 panic("DMA: Random memory could be DMA accessed\n");
725 if (dir == DMA_FROM_DEVICE)
726 panic("DMA: Random memory could be DMA written\n");
727 if (dir == DMA_TO_DEVICE)
728 panic("DMA: Random memory could be DMA read\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729}
730
731/*
732 * Map a single buffer of the indicated size for DMA in streaming mode. The
Tony Luck17e5ad62005-09-29 15:52:13 -0700733 * physical address to use is returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734 *
735 * Once the device is given the dma address, the device owns this memory until
Becky Bruceceb5ac32009-04-08 09:09:15 -0500736 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 */
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900738dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
739 unsigned long offset, size_t size,
740 enum dma_data_direction dir,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700741 unsigned long attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742{
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700743 phys_addr_t map, phys = page_to_phys(page) + offset;
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900744 dma_addr_t dev_addr = phys_to_dma(dev, phys);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745
Eric Sesterhenn34814542006-03-24 18:47:11 +0100746 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 /*
Becky Bruceceb5ac32009-04-08 09:09:15 -0500748 * If the address happens to be in the device's DMA window,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749 * we can safely return the device addr and not worry about bounce
750 * buffering it.
751 */
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900752 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753 return dev_addr;
754
Zoltan Kiss2b2b6142013-09-04 21:11:05 +0100755 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
756
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700757 /* Oh well, have to allocate and map a bounce buffer. */
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900758 map = map_single(dev, phys, size, dir);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700759 if (map == SWIOTLB_MAP_ERROR) {
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900760 swiotlb_full(dev, size, dir, 1);
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700761 return phys_to_dma(dev, io_tlb_overflow_buffer);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762 }
763
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700764 dev_addr = phys_to_dma(dev, map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700766 /* Ensure that the address returned is DMA'ble */
FUJITA Tomonorifba99fa2011-02-25 14:44:16 -0800767 if (!dma_capable(dev, dev_addr, size)) {
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700768 swiotlb_tbl_unmap_single(dev, map, size, dir);
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700769 return phys_to_dma(dev, io_tlb_overflow_buffer);
FUJITA Tomonorifba99fa2011-02-25 14:44:16 -0800770 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771
772 return dev_addr;
773}
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900774EXPORT_SYMBOL_GPL(swiotlb_map_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775
776/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 * Unmap a single streaming mode DMA translation. The dma_addr and size must
Becky Bruceceb5ac32009-04-08 09:09:15 -0500778 * match what was provided for in a previous swiotlb_map_page call. All
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 * other usages are undefined.
780 *
781 * After this call, reads by the cpu to the buffer are guaranteed to see
782 * whatever the device wrote there.
783 */
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500784static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500785 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900787 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788
Eric Sesterhenn34814542006-03-24 18:47:11 +0100789 BUG_ON(dir == DMA_NONE);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500790
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900791 if (is_swiotlb_buffer(paddr)) {
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700792 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500793 return;
794 }
795
796 if (dir != DMA_FROM_DEVICE)
797 return;
798
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900799 /*
800 * phys_to_virt doesn't work with hihgmem page but we could
801 * call dma_mark_clean() with hihgmem page here. However, we
802 * are fine since dma_mark_clean() is null on POWERPC. We can
803 * make dma_mark_clean() take a physical address if necessary.
804 */
805 dma_mark_clean(phys_to_virt(paddr), size);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500806}
807
808void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
809 size_t size, enum dma_data_direction dir,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700810 unsigned long attrs)
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500811{
812 unmap_single(hwdev, dev_addr, size, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813}
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900814EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900815
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816/*
817 * Make physical memory consistent for a single streaming mode DMA translation
818 * after a transfer.
819 *
Becky Bruceceb5ac32009-04-08 09:09:15 -0500820 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
Tony Luck17e5ad62005-09-29 15:52:13 -0700821 * using the cpu, yet do not wish to teardown the dma mapping, you must
822 * call this function before doing so. At the next point you give the dma
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823 * address back to the card, you must first perform a
824 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
825 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800826static void
John W. Linville8270f3f2005-09-29 14:43:32 -0700827swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400828 size_t size, enum dma_data_direction dir,
829 enum dma_sync_target target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900831 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832
Eric Sesterhenn34814542006-03-24 18:47:11 +0100833 BUG_ON(dir == DMA_NONE);
Becky Bruce380d6872009-04-08 09:09:20 -0500834
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900835 if (is_swiotlb_buffer(paddr)) {
Alexander Duyckfbfda892012-10-15 10:19:49 -0700836 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
Becky Bruce380d6872009-04-08 09:09:20 -0500837 return;
838 }
839
840 if (dir != DMA_FROM_DEVICE)
841 return;
842
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900843 dma_mark_clean(phys_to_virt(paddr), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844}
845
846void
John W. Linville8270f3f2005-09-29 14:43:32 -0700847swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900848 size_t size, enum dma_data_direction dir)
John W. Linville8270f3f2005-09-29 14:43:32 -0700849{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700850 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700851}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900852EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
John W. Linville8270f3f2005-09-29 14:43:32 -0700853
854void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900856 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700858 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900860EXPORT_SYMBOL(swiotlb_sync_single_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
862/*
863 * Map a set of buffers described by scatterlist in streaming mode for DMA.
Becky Bruceceb5ac32009-04-08 09:09:15 -0500864 * This is the scatter-gather version of the above swiotlb_map_page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865 * interface. Here the scatter gather list elements are each tagged with the
866 * appropriate dma address and length. They are obtained via
867 * sg_dma_{address,length}(SG).
868 *
869 * NOTE: An implementation may be able to use a smaller number of
870 * DMA address/length pairs than there are SG table elements.
871 * (for example via virtual mapping capabilities)
872 * The routine returns the number of addr/length pairs actually
873 * used, at most nents.
874 *
Becky Bruceceb5ac32009-04-08 09:09:15 -0500875 * Device ownership issues as mentioned above for swiotlb_map_page are the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 * same here.
877 */
878int
Arthur Kepner309df0c2008-04-29 01:00:32 -0700879swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700880 enum dma_data_direction dir, unsigned long attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200882 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 int i;
884
Eric Sesterhenn34814542006-03-24 18:47:11 +0100885 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886
Jens Axboedbfd49f2007-05-11 14:56:18 +0200887 for_each_sg(sgl, sg, nelems, i) {
Ian Campbell961d7d02009-01-09 18:32:10 +0000888 phys_addr_t paddr = sg_phys(sg);
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900889 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
Becky Brucebc40ac62008-12-22 10:26:08 -0800890
FUJITA Tomonoricf56e3f2009-07-10 10:04:52 +0900891 if (swiotlb_force ||
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900892 !dma_capable(hwdev, dev_addr, sg->length)) {
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700893 phys_addr_t map = map_single(hwdev, sg_phys(sg),
894 sg->length, dir);
895 if (map == SWIOTLB_MAP_ERROR) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 /* Don't panic here, we expect map_sg users
897 to do proper error handling. */
898 swiotlb_full(hwdev, sg->length, dir, 0);
Arthur Kepner309df0c2008-04-29 01:00:32 -0700899 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
900 attrs);
EunBong Song4d86ec72013-08-05 17:30:47 +0100901 sg_dma_len(sgl) = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 return 0;
903 }
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700904 sg->dma_address = phys_to_dma(hwdev, map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905 } else
906 sg->dma_address = dev_addr;
EunBong Song4d86ec72013-08-05 17:30:47 +0100907 sg_dma_len(sg) = sg->length;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908 }
909 return nelems;
910}
Arthur Kepner309df0c2008-04-29 01:00:32 -0700911EXPORT_SYMBOL(swiotlb_map_sg_attrs);
912
913int
914swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500915 enum dma_data_direction dir)
Arthur Kepner309df0c2008-04-29 01:00:32 -0700916{
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700917 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, 0);
Arthur Kepner309df0c2008-04-29 01:00:32 -0700918}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900919EXPORT_SYMBOL(swiotlb_map_sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920
921/*
922 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
Becky Bruceceb5ac32009-04-08 09:09:15 -0500923 * concerning calls here are the same as for swiotlb_unmap_page() above.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924 */
925void
Arthur Kepner309df0c2008-04-29 01:00:32 -0700926swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700927 int nelems, enum dma_data_direction dir,
928 unsigned long attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200930 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 int i;
932
Eric Sesterhenn34814542006-03-24 18:47:11 +0100933 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500935 for_each_sg(sgl, sg, nelems, i)
EunBong Song4d86ec72013-08-05 17:30:47 +0100936 unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500937
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938}
Arthur Kepner309df0c2008-04-29 01:00:32 -0700939EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
940
941void
942swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
Konrad Rzeszutek Wilk22d48262010-05-10 15:01:15 -0500943 enum dma_data_direction dir)
Arthur Kepner309df0c2008-04-29 01:00:32 -0700944{
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700945 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, 0);
Arthur Kepner309df0c2008-04-29 01:00:32 -0700946}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900947EXPORT_SYMBOL(swiotlb_unmap_sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948
949/*
950 * Make physical memory consistent for a set of streaming mode DMA translations
951 * after a transfer.
952 *
953 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
954 * and usage.
955 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800956static void
Jens Axboedbfd49f2007-05-11 14:56:18 +0200957swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
Konrad Rzeszutek Wilkd7ef1532010-05-28 11:37:10 -0400958 int nelems, enum dma_data_direction dir,
959 enum dma_sync_target target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200961 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 int i;
963
Becky Bruce380d6872009-04-08 09:09:20 -0500964 for_each_sg(sgl, sg, nelems, i)
965 swiotlb_sync_single(hwdev, sg->dma_address,
EunBong Song4d86ec72013-08-05 17:30:47 +0100966 sg_dma_len(sg), dir, target);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967}
968
969void
John W. Linville8270f3f2005-09-29 14:43:32 -0700970swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900971 int nelems, enum dma_data_direction dir)
John W. Linville8270f3f2005-09-29 14:43:32 -0700972{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700973 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700974}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900975EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
John W. Linville8270f3f2005-09-29 14:43:32 -0700976
977void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900979 int nelems, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700981 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900983EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984
985int
FUJITA Tomonori8d8bb392008-07-25 19:44:49 -0700986swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987{
Alexander Duyckee3f6ba2012-10-15 10:19:34 -0700988 return (dma_addr == phys_to_dma(hwdev, io_tlb_overflow_buffer));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900990EXPORT_SYMBOL(swiotlb_dma_mapping_error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991
992/*
Tony Luck17e5ad62005-09-29 15:52:13 -0700993 * Return whether the given device DMA address mask can be supported
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994 * properly. For example, if your device can only drive the low 24-bits
Tony Luck17e5ad62005-09-29 15:52:13 -0700995 * during bus mastering, then you would pass 0x00ffffff as the mask to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 * this function.
997 */
998int
Jan Beulich563aaf02007-02-05 18:51:25 -0800999swiotlb_dma_supported(struct device *hwdev, u64 mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000{
Alexander Duyckc40dba02012-10-15 10:19:22 -07001001 return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003EXPORT_SYMBOL(swiotlb_dma_supported);