blob: a009055140ecf7a798625d3fee7cd856c0bb8d4c [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Dynamic DMA mapping support.
3 *
Jan Beulich563aaf02007-02-05 18:51:25 -08004 * This implementation is a fallback for platforms that do not support
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 * I/O TLBs (aka DMA address translation hardware).
6 * Copyright (C) 2000 Asit Mallick <Asit.K.Mallick@intel.com>
7 * Copyright (C) 2000 Goutham Rao <goutham.rao@intel.com>
8 * Copyright (C) 2000, 2003 Hewlett-Packard Co
9 * David Mosberger-Tang <davidm@hpl.hp.com>
10 *
11 * 03/05/07 davidm Switch from PCI-DMA to generic device DMA API.
12 * 00/12/13 davidm Rename to swiotlb.c and add mark_clean() to avoid
13 * unnecessary i-cache flushing.
John W. Linville569c8bf2005-09-29 14:45:24 -070014 * 04/07/.. ak Better overflow handling. Assorted fixes.
15 * 05/09/10 linville Add support for syncing ranges, support syncing for
16 * DMA_BIDIRECTIONAL mappings, miscellaneous cleanup.
Becky Brucefb05a372008-12-22 10:26:09 -080017 * 08/12/11 beckyb Add highmem support
Linus Torvalds1da177e2005-04-16 15:20:36 -070018 */
19
20#include <linux/cache.h>
Tony Luck17e5ad62005-09-29 15:52:13 -070021#include <linux/dma-mapping.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070022#include <linux/mm.h>
23#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/spinlock.h>
25#include <linux/string.h>
Ian Campbell0016fde2008-12-16 12:17:27 -080026#include <linux/swiotlb.h>
Becky Brucefb05a372008-12-22 10:26:09 -080027#include <linux/pfn.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070028#include <linux/types.h>
29#include <linux/ctype.h>
Jeremy Fitzhardingeef9b1892008-12-16 12:17:33 -080030#include <linux/highmem.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090031#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070032
33#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070034#include <asm/dma.h>
Tony Luck17e5ad62005-09-29 15:52:13 -070035#include <asm/scatterlist.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036
37#include <linux/init.h>
38#include <linux/bootmem.h>
FUJITA Tomonoria8522502008-04-29 00:59:36 -070039#include <linux/iommu-helper.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040
41#define OFFSET(val,align) ((unsigned long) \
42 ( (val) & ( (align) - 1)))
43
Alex Williamson0b9afed2005-09-06 11:20:49 -060044#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
45
46/*
47 * Minimum IO TLB size to bother booting with. Systems with mainly
48 * 64bit capable cards will only lightly use the swiotlb. If we can't
49 * allocate a contiguous 1MB, we're probably in trouble anyway.
50 */
51#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
52
John W. Linvillede69e0f2005-09-29 14:44:57 -070053/*
54 * Enumeration for sync targets
55 */
56enum dma_sync_target {
57 SYNC_FOR_CPU = 0,
58 SYNC_FOR_DEVICE = 1,
59};
60
Linus Torvalds1da177e2005-04-16 15:20:36 -070061int swiotlb_force;
62
63/*
Becky Bruceceb5ac32009-04-08 09:09:15 -050064 * Used to do a quick range check in unmap_single and
65 * sync_single_*, to see if the memory was in fact allocated by this
Linus Torvalds1da177e2005-04-16 15:20:36 -070066 * API.
67 */
68static char *io_tlb_start, *io_tlb_end;
69
70/*
71 * The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and
72 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
73 */
74static unsigned long io_tlb_nslabs;
75
76/*
77 * When the IOMMU overflows we return a fallback buffer. This sets the size.
78 */
79static unsigned long io_tlb_overflow = 32*1024;
80
81void *io_tlb_overflow_buffer;
82
83/*
84 * This is a free list describing the number of free entries available from
85 * each index
86 */
87static unsigned int *io_tlb_list;
88static unsigned int io_tlb_index;
89
90/*
91 * We need to save away the original address corresponding to a mapped entry
92 * for the sync operations.
93 */
Becky Brucebc40ac62008-12-22 10:26:08 -080094static phys_addr_t *io_tlb_orig_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -070095
96/*
97 * Protect the above data structures in the map and unmap calls
98 */
99static DEFINE_SPINLOCK(io_tlb_lock);
100
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900101static int late_alloc;
102
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103static int __init
104setup_io_tlb_npages(char *str)
105{
106 if (isdigit(*str)) {
Alex Williamsone8579e72005-08-04 13:06:00 -0700107 io_tlb_nslabs = simple_strtoul(str, &str, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108 /* avoid tail segment of size < IO_TLB_SEGSIZE */
109 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
110 }
111 if (*str == ',')
112 ++str;
FUJITA Tomonorib18485e2009-11-12 00:03:28 +0900113 if (!strcmp(str, "force"))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114 swiotlb_force = 1;
FUJITA Tomonorib18485e2009-11-12 00:03:28 +0900115
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116 return 1;
117}
118__setup("swiotlb=", setup_io_tlb_npages);
119/* make io_tlb_overflow tunable too? */
120
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900121/* Note that this doesn't work with highmem page */
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800122static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
123 volatile void *address)
Ian Campbelle08e1f72008-12-16 12:17:30 -0800124{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900125 return phys_to_dma(hwdev, virt_to_phys(address));
Ian Campbelle08e1f72008-12-16 12:17:30 -0800126}
127
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900128void swiotlb_print_info(void)
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800129{
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900130 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800131 phys_addr_t pstart, pend;
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800132
133 pstart = virt_to_phys(io_tlb_start);
134 pend = virt_to_phys(io_tlb_end);
135
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800136 printk(KERN_INFO "Placing %luMB software IO TLB between %p - %p\n",
137 bytes >> 20, io_tlb_start, io_tlb_end);
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800138 printk(KERN_INFO "software IO TLB at phys %#llx - %#llx\n",
139 (unsigned long long)pstart,
140 (unsigned long long)pend);
Ian Campbell2e5b2b82008-12-16 12:17:34 -0800141}
142
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143/*
144 * Statically reserve bounce buffer space and initialize bounce buffer data
Tony Luck17e5ad62005-09-29 15:52:13 -0700145 * structures for the software IO TLB used to implement the DMA API.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800147void __init
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900148swiotlb_init_with_default_size(size_t default_size, int verbose)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149{
Jan Beulich563aaf02007-02-05 18:51:25 -0800150 unsigned long i, bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151
152 if (!io_tlb_nslabs) {
Alex Williamsone8579e72005-08-04 13:06:00 -0700153 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
155 }
156
Jan Beulich563aaf02007-02-05 18:51:25 -0800157 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
158
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 /*
160 * Get IO TLB memory from the low pages
161 */
FUJITA Tomonori38851232009-07-10 10:04:50 +0900162 io_tlb_start = alloc_bootmem_low_pages(bytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163 if (!io_tlb_start)
164 panic("Cannot allocate SWIOTLB buffer");
Jan Beulich563aaf02007-02-05 18:51:25 -0800165 io_tlb_end = io_tlb_start + bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166
167 /*
168 * Allocate and initialize the free list array. This array is used
169 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
170 * between io_tlb_start and io_tlb_end.
171 */
172 io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
Tony Luck25667d62007-03-06 13:31:45 -0800173 for (i = 0; i < io_tlb_nslabs; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
175 io_tlb_index = 0;
Becky Brucebc40ac62008-12-22 10:26:08 -0800176 io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(phys_addr_t));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177
178 /*
179 * Get the overflow emergency buffer
180 */
181 io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
Jan Beulich563aaf02007-02-05 18:51:25 -0800182 if (!io_tlb_overflow_buffer)
183 panic("Cannot allocate SWIOTLB overflow buffer!\n");
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900184 if (verbose)
185 swiotlb_print_info();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186}
187
Jan Beulich563aaf02007-02-05 18:51:25 -0800188void __init
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900189swiotlb_init(int verbose)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190{
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900191 swiotlb_init_with_default_size(64 * (1<<20), verbose); /* default to 64MB */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192}
193
Alex Williamson0b9afed2005-09-06 11:20:49 -0600194/*
195 * Systems with larger DMA zones (those that don't support ISA) can
196 * initialize the swiotlb later using the slab allocator if needed.
197 * This should be just like above, but with some error catching.
198 */
199int
Jan Beulich563aaf02007-02-05 18:51:25 -0800200swiotlb_late_init_with_default_size(size_t default_size)
Alex Williamson0b9afed2005-09-06 11:20:49 -0600201{
Jan Beulich563aaf02007-02-05 18:51:25 -0800202 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600203 unsigned int order;
204
205 if (!io_tlb_nslabs) {
206 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
207 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
208 }
209
210 /*
211 * Get IO TLB memory from the low pages
212 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800213 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600214 io_tlb_nslabs = SLABS_PER_PAGE << order;
Jan Beulich563aaf02007-02-05 18:51:25 -0800215 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600216
217 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
FUJITA Tomonoribb521962009-07-10 10:04:51 +0900218 io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
219 order);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600220 if (io_tlb_start)
221 break;
222 order--;
223 }
224
225 if (!io_tlb_start)
226 goto cleanup1;
227
Jan Beulich563aaf02007-02-05 18:51:25 -0800228 if (order != get_order(bytes)) {
Alex Williamson0b9afed2005-09-06 11:20:49 -0600229 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
230 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
231 io_tlb_nslabs = SLABS_PER_PAGE << order;
Jan Beulich563aaf02007-02-05 18:51:25 -0800232 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600233 }
Jan Beulich563aaf02007-02-05 18:51:25 -0800234 io_tlb_end = io_tlb_start + bytes;
235 memset(io_tlb_start, 0, bytes);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600236
237 /*
238 * Allocate and initialize the free list array. This array is used
239 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
240 * between io_tlb_start and io_tlb_end.
241 */
242 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
243 get_order(io_tlb_nslabs * sizeof(int)));
244 if (!io_tlb_list)
245 goto cleanup2;
246
247 for (i = 0; i < io_tlb_nslabs; i++)
248 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
249 io_tlb_index = 0;
250
Becky Brucebc40ac62008-12-22 10:26:08 -0800251 io_tlb_orig_addr = (phys_addr_t *)
252 __get_free_pages(GFP_KERNEL,
253 get_order(io_tlb_nslabs *
254 sizeof(phys_addr_t)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600255 if (!io_tlb_orig_addr)
256 goto cleanup3;
257
Becky Brucebc40ac62008-12-22 10:26:08 -0800258 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600259
260 /*
261 * Get the overflow emergency buffer
262 */
263 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
264 get_order(io_tlb_overflow));
265 if (!io_tlb_overflow_buffer)
266 goto cleanup4;
267
FUJITA Tomonoriad32e8c2009-11-10 19:46:19 +0900268 swiotlb_print_info();
Alex Williamson0b9afed2005-09-06 11:20:49 -0600269
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900270 late_alloc = 1;
271
Alex Williamson0b9afed2005-09-06 11:20:49 -0600272 return 0;
273
274cleanup4:
Becky Brucebc40ac62008-12-22 10:26:08 -0800275 free_pages((unsigned long)io_tlb_orig_addr,
276 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600277 io_tlb_orig_addr = NULL;
278cleanup3:
Tony Luck25667d62007-03-06 13:31:45 -0800279 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
280 sizeof(int)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600281 io_tlb_list = NULL;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600282cleanup2:
Jan Beulich563aaf02007-02-05 18:51:25 -0800283 io_tlb_end = NULL;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600284 free_pages((unsigned long)io_tlb_start, order);
285 io_tlb_start = NULL;
286cleanup1:
287 io_tlb_nslabs = req_nslabs;
288 return -ENOMEM;
289}
290
FUJITA Tomonori5740afd2009-11-10 19:46:18 +0900291void __init swiotlb_free(void)
292{
293 if (!io_tlb_overflow_buffer)
294 return;
295
296 if (late_alloc) {
297 free_pages((unsigned long)io_tlb_overflow_buffer,
298 get_order(io_tlb_overflow));
299 free_pages((unsigned long)io_tlb_orig_addr,
300 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
301 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
302 sizeof(int)));
303 free_pages((unsigned long)io_tlb_start,
304 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
305 } else {
306 free_bootmem_late(__pa(io_tlb_overflow_buffer),
307 io_tlb_overflow);
308 free_bootmem_late(__pa(io_tlb_orig_addr),
309 io_tlb_nslabs * sizeof(phys_addr_t));
310 free_bootmem_late(__pa(io_tlb_list),
311 io_tlb_nslabs * sizeof(int));
312 free_bootmem_late(__pa(io_tlb_start),
313 io_tlb_nslabs << IO_TLB_SHIFT);
314 }
315}
316
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900317static int is_swiotlb_buffer(phys_addr_t paddr)
FUJITA Tomonori640aebf2008-09-08 18:53:50 +0900318{
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900319 return paddr >= virt_to_phys(io_tlb_start) &&
320 paddr < virt_to_phys(io_tlb_end);
FUJITA Tomonori640aebf2008-09-08 18:53:50 +0900321}
322
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323/*
Becky Brucefb05a372008-12-22 10:26:09 -0800324 * Bounce: copy the swiotlb buffer back to the original dma location
325 */
326static void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
327 enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328{
Becky Brucefb05a372008-12-22 10:26:09 -0800329 unsigned long pfn = PFN_DOWN(phys);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330
Becky Brucefb05a372008-12-22 10:26:09 -0800331 if (PageHighMem(pfn_to_page(pfn))) {
332 /* The buffer does not have a mapping. Map it in and copy */
333 unsigned int offset = phys & ~PAGE_MASK;
334 char *buffer;
335 unsigned int sz = 0;
336 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337
Becky Brucefb05a372008-12-22 10:26:09 -0800338 while (size) {
Becky Bruce67131ad2009-04-08 09:09:16 -0500339 sz = min_t(size_t, PAGE_SIZE - offset, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340
Becky Brucefb05a372008-12-22 10:26:09 -0800341 local_irq_save(flags);
342 buffer = kmap_atomic(pfn_to_page(pfn),
343 KM_BOUNCE_READ);
344 if (dir == DMA_TO_DEVICE)
345 memcpy(dma_addr, buffer + offset, sz);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346 else
Becky Brucefb05a372008-12-22 10:26:09 -0800347 memcpy(buffer + offset, dma_addr, sz);
348 kunmap_atomic(buffer, KM_BOUNCE_READ);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349 local_irq_restore(flags);
Becky Brucefb05a372008-12-22 10:26:09 -0800350
351 size -= sz;
352 pfn++;
353 dma_addr += sz;
354 offset = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355 }
356 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357 if (dir == DMA_TO_DEVICE)
Becky Brucefb05a372008-12-22 10:26:09 -0800358 memcpy(dma_addr, phys_to_virt(phys), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 else
Becky Brucefb05a372008-12-22 10:26:09 -0800360 memcpy(phys_to_virt(phys), dma_addr, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 }
362}
363
364/*
365 * Allocates bounce buffer and returns its kernel virtual address.
366 */
367static void *
Becky Brucebc40ac62008-12-22 10:26:08 -0800368map_single(struct device *hwdev, phys_addr_t phys, size_t size, int dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369{
370 unsigned long flags;
371 char *dma_addr;
372 unsigned int nslots, stride, index, wrap;
373 int i;
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800374 unsigned long start_dma_addr;
375 unsigned long mask;
376 unsigned long offset_slots;
377 unsigned long max_slots;
378
379 mask = dma_get_seg_boundary(hwdev);
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800380 start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start) & mask;
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800381
382 offset_slots = ALIGN(start_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
Ian Campbella5ddde42008-12-16 12:17:29 -0800383
384 /*
385 * Carefully handle integer overflow which can occur when mask == ~0UL.
386 */
Jan Beulichb15a3892008-03-13 09:13:30 +0000387 max_slots = mask + 1
388 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
389 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390
391 /*
392 * For mappings greater than a page, we limit the stride (and
393 * hence alignment) to a page size.
394 */
395 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
396 if (size > PAGE_SIZE)
397 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
398 else
399 stride = 1;
400
Eric Sesterhenn34814542006-03-24 18:47:11 +0100401 BUG_ON(!nslots);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402
403 /*
404 * Find suitable number of IO TLB entries size that will fit this
405 * request and allocate a buffer from that IO TLB pool.
406 */
407 spin_lock_irqsave(&io_tlb_lock, flags);
Andrew Mortona7133a12008-04-29 00:59:36 -0700408 index = ALIGN(io_tlb_index, stride);
409 if (index >= io_tlb_nslabs)
410 index = 0;
411 wrap = index;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412
Andrew Mortona7133a12008-04-29 00:59:36 -0700413 do {
FUJITA Tomonoria8522502008-04-29 00:59:36 -0700414 while (iommu_is_span_boundary(index, nslots, offset_slots,
415 max_slots)) {
Jan Beulichb15a3892008-03-13 09:13:30 +0000416 index += stride;
417 if (index >= io_tlb_nslabs)
418 index = 0;
Andrew Mortona7133a12008-04-29 00:59:36 -0700419 if (index == wrap)
420 goto not_found;
421 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422
Andrew Mortona7133a12008-04-29 00:59:36 -0700423 /*
424 * If we find a slot that indicates we have 'nslots' number of
425 * contiguous buffers, we allocate the buffers from that slot
426 * and mark the entries as '0' indicating unavailable.
427 */
428 if (io_tlb_list[index] >= nslots) {
429 int count = 0;
430
431 for (i = index; i < (int) (index + nslots); i++)
432 io_tlb_list[i] = 0;
433 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
434 io_tlb_list[i] = ++count;
435 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
436
437 /*
438 * Update the indices to avoid searching in the next
439 * round.
440 */
441 io_tlb_index = ((index + nslots) < io_tlb_nslabs
442 ? (index + nslots) : 0);
443
444 goto found;
445 }
446 index += stride;
447 if (index >= io_tlb_nslabs)
448 index = 0;
449 } while (index != wrap);
450
451not_found:
452 spin_unlock_irqrestore(&io_tlb_lock, flags);
453 return NULL;
454found:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455 spin_unlock_irqrestore(&io_tlb_lock, flags);
456
457 /*
458 * Save away the mapping from the original address to the DMA address.
459 * This is needed when we sync the memory. Then we sync the buffer if
460 * needed.
461 */
Becky Brucebc40ac62008-12-22 10:26:08 -0800462 for (i = 0; i < nslots; i++)
463 io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
Becky Brucefb05a372008-12-22 10:26:09 -0800465 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466
467 return dma_addr;
468}
469
470/*
471 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
472 */
473static void
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500474do_unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475{
476 unsigned long flags;
477 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
478 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
Becky Brucebc40ac62008-12-22 10:26:08 -0800479 phys_addr_t phys = io_tlb_orig_addr[index];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480
481 /*
482 * First, sync the memory before unmapping the entry
483 */
Becky Brucebc40ac62008-12-22 10:26:08 -0800484 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
Becky Brucefb05a372008-12-22 10:26:09 -0800485 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486
487 /*
488 * Return the buffer to the free list by setting the corresponding
André Goddard Rosaaf901ca2009-11-14 13:09:05 -0200489 * entries to indicate the number of contiguous entries available.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 * While returning the entries to the free list, we merge the entries
491 * with slots below and above the pool being returned.
492 */
493 spin_lock_irqsave(&io_tlb_lock, flags);
494 {
495 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
496 io_tlb_list[index + nslots] : 0);
497 /*
498 * Step 1: return the slots to the free list, merging the
499 * slots with superceeding slots
500 */
501 for (i = index + nslots - 1; i >= index; i--)
502 io_tlb_list[i] = ++count;
503 /*
504 * Step 2: merge the returned slots with the preceding slots,
505 * if available (non zero)
506 */
507 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
508 io_tlb_list[i] = ++count;
509 }
510 spin_unlock_irqrestore(&io_tlb_lock, flags);
511}
512
513static void
John W. Linvillede69e0f2005-09-29 14:44:57 -0700514sync_single(struct device *hwdev, char *dma_addr, size_t size,
515 int dir, int target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516{
Becky Brucebc40ac62008-12-22 10:26:08 -0800517 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
518 phys_addr_t phys = io_tlb_orig_addr[index];
519
520 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
Keir Fraserdf336d12007-07-21 04:37:24 -0700521
John W. Linvillede69e0f2005-09-29 14:44:57 -0700522 switch (target) {
523 case SYNC_FOR_CPU:
524 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
Becky Brucefb05a372008-12-22 10:26:09 -0800525 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100526 else
527 BUG_ON(dir != DMA_TO_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700528 break;
529 case SYNC_FOR_DEVICE:
530 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
Becky Brucefb05a372008-12-22 10:26:09 -0800531 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100532 else
533 BUG_ON(dir != DMA_FROM_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700534 break;
535 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 BUG();
John W. Linvillede69e0f2005-09-29 14:44:57 -0700537 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538}
539
540void *
541swiotlb_alloc_coherent(struct device *hwdev, size_t size,
Al Viro06a54492005-10-21 03:21:03 -0400542 dma_addr_t *dma_handle, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543{
Jan Beulich563aaf02007-02-05 18:51:25 -0800544 dma_addr_t dev_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 void *ret;
546 int order = get_order(size);
Yang Hongyang284901a2009-04-06 19:01:15 -0700547 u64 dma_mask = DMA_BIT_MASK(32);
FUJITA Tomonori1e74f302008-11-17 16:24:34 +0900548
549 if (hwdev && hwdev->coherent_dma_mask)
550 dma_mask = hwdev->coherent_dma_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551
Tony Luck25667d62007-03-06 13:31:45 -0800552 ret = (void *)__get_free_pages(flags, order);
Jan Beulichac2b3e62009-12-15 16:47:43 -0800553 if (ret && swiotlb_virt_to_bus(hwdev, ret) + size - 1 > dma_mask) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554 /*
555 * The allocated memory isn't reachable by the device.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556 */
557 free_pages((unsigned long) ret, order);
558 ret = NULL;
559 }
560 if (!ret) {
561 /*
562 * We are either out of memory or the device can't DMA
Becky Bruceceb5ac32009-04-08 09:09:15 -0500563 * to GFP_DMA memory; fall back on map_single(), which
564 * will grab memory from the lowest available address range.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565 */
Becky Brucebc40ac62008-12-22 10:26:08 -0800566 ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
FUJITA Tomonori9dfda122008-09-08 18:53:48 +0900567 if (!ret)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 }
570
571 memset(ret, 0, size);
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800572 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573
574 /* Confirm address can be DMA'd by device */
Jan Beulichac2b3e62009-12-15 16:47:43 -0800575 if (dev_addr + size - 1 > dma_mask) {
Jan Beulich563aaf02007-02-05 18:51:25 -0800576 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
FUJITA Tomonori1e74f302008-11-17 16:24:34 +0900577 (unsigned long long)dma_mask,
Jan Beulich563aaf02007-02-05 18:51:25 -0800578 (unsigned long long)dev_addr);
FUJITA Tomonoria2b89b52008-10-23 18:42:03 +0900579
580 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500581 do_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
FUJITA Tomonoria2b89b52008-10-23 18:42:03 +0900582 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 }
584 *dma_handle = dev_addr;
585 return ret;
586}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900587EXPORT_SYMBOL(swiotlb_alloc_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588
589void
590swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900591 dma_addr_t dev_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900593 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900594
David Brownellaa248862007-08-10 13:10:27 -0700595 WARN_ON(irqs_disabled());
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900596 if (!is_swiotlb_buffer(paddr))
597 free_pages((unsigned long)vaddr, get_order(size));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 else
599 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500600 do_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900602EXPORT_SYMBOL(swiotlb_free_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603
604static void
605swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
606{
607 /*
608 * Ran out of IOMMU space for this operation. This is very bad.
609 * Unfortunately the drivers cannot handle this operation properly.
Tony Luck17e5ad62005-09-29 15:52:13 -0700610 * unless they check for dma_mapping_error (most don't)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611 * When the mapping is small enough return a static buffer to limit
612 * the damage, or panic when the transfer is too big.
613 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800614 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
Kay Sievers94b32482009-01-06 10:44:37 -0800615 "device %s\n", size, dev ? dev_name(dev) : "?");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616
Casey Dahlinc7084b32009-08-20 16:27:56 -0700617 if (size <= io_tlb_overflow || !do_panic)
618 return;
619
620 if (dir == DMA_BIDIRECTIONAL)
621 panic("DMA: Random memory could be DMA accessed\n");
622 if (dir == DMA_FROM_DEVICE)
623 panic("DMA: Random memory could be DMA written\n");
624 if (dir == DMA_TO_DEVICE)
625 panic("DMA: Random memory could be DMA read\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626}
627
628/*
629 * Map a single buffer of the indicated size for DMA in streaming mode. The
Tony Luck17e5ad62005-09-29 15:52:13 -0700630 * physical address to use is returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631 *
632 * Once the device is given the dma address, the device owns this memory until
Becky Bruceceb5ac32009-04-08 09:09:15 -0500633 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 */
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900635dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
636 unsigned long offset, size_t size,
637 enum dma_data_direction dir,
638 struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639{
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900640 phys_addr_t phys = page_to_phys(page) + offset;
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900641 dma_addr_t dev_addr = phys_to_dma(dev, phys);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642 void *map;
643
Eric Sesterhenn34814542006-03-24 18:47:11 +0100644 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 /*
Becky Bruceceb5ac32009-04-08 09:09:15 -0500646 * If the address happens to be in the device's DMA window,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 * we can safely return the device addr and not worry about bounce
648 * buffering it.
649 */
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900650 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 return dev_addr;
652
653 /*
654 * Oh well, have to allocate and map a bounce buffer.
655 */
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900656 map = map_single(dev, phys, size, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 if (!map) {
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900658 swiotlb_full(dev, size, dir, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 map = io_tlb_overflow_buffer;
660 }
661
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900662 dev_addr = swiotlb_virt_to_bus(dev, map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663
664 /*
665 * Ensure that the address returned is DMA'ble
666 */
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900667 if (!dma_capable(dev, dev_addr, size))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668 panic("map_single: bounce buffer is not DMA'ble");
669
670 return dev_addr;
671}
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900672EXPORT_SYMBOL_GPL(swiotlb_map_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673
674/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 * Unmap a single streaming mode DMA translation. The dma_addr and size must
Becky Bruceceb5ac32009-04-08 09:09:15 -0500676 * match what was provided for in a previous swiotlb_map_page call. All
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 * other usages are undefined.
678 *
679 * After this call, reads by the cpu to the buffer are guaranteed to see
680 * whatever the device wrote there.
681 */
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500682static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
683 size_t size, int dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900685 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700686
Eric Sesterhenn34814542006-03-24 18:47:11 +0100687 BUG_ON(dir == DMA_NONE);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500688
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900689 if (is_swiotlb_buffer(paddr)) {
690 do_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500691 return;
692 }
693
694 if (dir != DMA_FROM_DEVICE)
695 return;
696
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900697 /*
698 * phys_to_virt doesn't work with hihgmem page but we could
699 * call dma_mark_clean() with hihgmem page here. However, we
700 * are fine since dma_mark_clean() is null on POWERPC. We can
701 * make dma_mark_clean() take a physical address if necessary.
702 */
703 dma_mark_clean(phys_to_virt(paddr), size);
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500704}
705
706void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
707 size_t size, enum dma_data_direction dir,
708 struct dma_attrs *attrs)
709{
710 unmap_single(hwdev, dev_addr, size, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711}
FUJITA Tomonorif98eee82009-01-05 23:59:03 +0900712EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900713
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714/*
715 * Make physical memory consistent for a single streaming mode DMA translation
716 * after a transfer.
717 *
Becky Bruceceb5ac32009-04-08 09:09:15 -0500718 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
Tony Luck17e5ad62005-09-29 15:52:13 -0700719 * using the cpu, yet do not wish to teardown the dma mapping, you must
720 * call this function before doing so. At the next point you give the dma
Linus Torvalds1da177e2005-04-16 15:20:36 -0700721 * address back to the card, you must first perform a
722 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
723 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800724static void
John W. Linville8270f3f2005-09-29 14:43:32 -0700725swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700726 size_t size, int dir, int target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727{
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900728 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729
Eric Sesterhenn34814542006-03-24 18:47:11 +0100730 BUG_ON(dir == DMA_NONE);
Becky Bruce380d6872009-04-08 09:09:20 -0500731
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900732 if (is_swiotlb_buffer(paddr)) {
733 sync_single(hwdev, phys_to_virt(paddr), size, dir, target);
Becky Bruce380d6872009-04-08 09:09:20 -0500734 return;
735 }
736
737 if (dir != DMA_FROM_DEVICE)
738 return;
739
FUJITA Tomonori02ca6462009-07-23 11:18:49 +0900740 dma_mark_clean(phys_to_virt(paddr), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741}
742
743void
John W. Linville8270f3f2005-09-29 14:43:32 -0700744swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900745 size_t size, enum dma_data_direction dir)
John W. Linville8270f3f2005-09-29 14:43:32 -0700746{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700747 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700748}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900749EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
John W. Linville8270f3f2005-09-29 14:43:32 -0700750
751void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900753 size_t size, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700755 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900757EXPORT_SYMBOL(swiotlb_sync_single_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758
759/*
760 * Map a set of buffers described by scatterlist in streaming mode for DMA.
Becky Bruceceb5ac32009-04-08 09:09:15 -0500761 * This is the scatter-gather version of the above swiotlb_map_page
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762 * interface. Here the scatter gather list elements are each tagged with the
763 * appropriate dma address and length. They are obtained via
764 * sg_dma_{address,length}(SG).
765 *
766 * NOTE: An implementation may be able to use a smaller number of
767 * DMA address/length pairs than there are SG table elements.
768 * (for example via virtual mapping capabilities)
769 * The routine returns the number of addr/length pairs actually
770 * used, at most nents.
771 *
Becky Bruceceb5ac32009-04-08 09:09:15 -0500772 * Device ownership issues as mentioned above for swiotlb_map_page are the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773 * same here.
774 */
775int
Arthur Kepner309df0c2008-04-29 01:00:32 -0700776swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900777 enum dma_data_direction dir, struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200779 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 int i;
781
Eric Sesterhenn34814542006-03-24 18:47:11 +0100782 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783
Jens Axboedbfd49f2007-05-11 14:56:18 +0200784 for_each_sg(sgl, sg, nelems, i) {
Ian Campbell961d7d02009-01-09 18:32:10 +0000785 phys_addr_t paddr = sg_phys(sg);
FUJITA Tomonori862d1962009-07-10 10:05:02 +0900786 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
Becky Brucebc40ac62008-12-22 10:26:08 -0800787
FUJITA Tomonoricf56e3f2009-07-10 10:04:52 +0900788 if (swiotlb_force ||
FUJITA Tomonorib9394642009-07-10 10:04:58 +0900789 !dma_capable(hwdev, dev_addr, sg->length)) {
Becky Brucebc40ac62008-12-22 10:26:08 -0800790 void *map = map_single(hwdev, sg_phys(sg),
791 sg->length, dir);
Andi Kleen7e870232005-12-20 14:45:19 +0100792 if (!map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 /* Don't panic here, we expect map_sg users
794 to do proper error handling. */
795 swiotlb_full(hwdev, sg->length, dir, 0);
Arthur Kepner309df0c2008-04-29 01:00:32 -0700796 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
797 attrs);
Jens Axboedbfd49f2007-05-11 14:56:18 +0200798 sgl[0].dma_length = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 return 0;
800 }
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800801 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 } else
803 sg->dma_address = dev_addr;
804 sg->dma_length = sg->length;
805 }
806 return nelems;
807}
Arthur Kepner309df0c2008-04-29 01:00:32 -0700808EXPORT_SYMBOL(swiotlb_map_sg_attrs);
809
810int
811swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
812 int dir)
813{
814 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
815}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900816EXPORT_SYMBOL(swiotlb_map_sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817
818/*
819 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
Becky Bruceceb5ac32009-04-08 09:09:15 -0500820 * concerning calls here are the same as for swiotlb_unmap_page() above.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821 */
822void
Arthur Kepner309df0c2008-04-29 01:00:32 -0700823swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900824 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200826 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 int i;
828
Eric Sesterhenn34814542006-03-24 18:47:11 +0100829 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830
Becky Bruce7fcebbd2009-04-08 09:09:19 -0500831 for_each_sg(sgl, sg, nelems, i)
832 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
833
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834}
Arthur Kepner309df0c2008-04-29 01:00:32 -0700835EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
836
837void
838swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
839 int dir)
840{
841 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
842}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900843EXPORT_SYMBOL(swiotlb_unmap_sg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844
845/*
846 * Make physical memory consistent for a set of streaming mode DMA translations
847 * after a transfer.
848 *
849 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
850 * and usage.
851 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800852static void
Jens Axboedbfd49f2007-05-11 14:56:18 +0200853swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700854 int nelems, int dir, int target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200856 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 int i;
858
Becky Bruce380d6872009-04-08 09:09:20 -0500859 for_each_sg(sgl, sg, nelems, i)
860 swiotlb_sync_single(hwdev, sg->dma_address,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700861 sg->dma_length, dir, target);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862}
863
864void
John W. Linville8270f3f2005-09-29 14:43:32 -0700865swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900866 int nelems, enum dma_data_direction dir)
John W. Linville8270f3f2005-09-29 14:43:32 -0700867{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700868 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700869}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900870EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
John W. Linville8270f3f2005-09-29 14:43:32 -0700871
872void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
FUJITA Tomonori160c1d82009-01-05 23:59:02 +0900874 int nelems, enum dma_data_direction dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700876 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900878EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879
880int
FUJITA Tomonori8d8bb392008-07-25 19:44:49 -0700881swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882{
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800883 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884}
FUJITA Tomonori874d6a92008-12-28 15:02:07 +0900885EXPORT_SYMBOL(swiotlb_dma_mapping_error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886
887/*
Tony Luck17e5ad62005-09-29 15:52:13 -0700888 * Return whether the given device DMA address mask can be supported
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 * properly. For example, if your device can only drive the low 24-bits
Tony Luck17e5ad62005-09-29 15:52:13 -0700890 * during bus mastering, then you would pass 0x00ffffff as the mask to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 * this function.
892 */
893int
Jan Beulich563aaf02007-02-05 18:51:25 -0800894swiotlb_dma_supported(struct device *hwdev, u64 mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895{
Jeremy Fitzhardinge70a7d3c2008-12-22 10:26:05 -0800896 return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898EXPORT_SYMBOL(swiotlb_dma_supported);