blob: 025922807e6e648bb7dc4eda61bfd8fa609a221c [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.
Linus Torvalds1da177e2005-04-16 15:20:36 -070017 */
18
19#include <linux/cache.h>
Tony Luck17e5ad62005-09-29 15:52:13 -070020#include <linux/dma-mapping.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <linux/mm.h>
22#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/spinlock.h>
24#include <linux/string.h>
25#include <linux/types.h>
26#include <linux/ctype.h>
27
28#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <asm/dma.h>
Tony Luck17e5ad62005-09-29 15:52:13 -070030#include <asm/scatterlist.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070031
32#include <linux/init.h>
33#include <linux/bootmem.h>
34
35#define OFFSET(val,align) ((unsigned long) \
36 ( (val) & ( (align) - 1)))
37
Jens Axboef9527f12007-10-22 19:44:53 +020038#define SG_ENT_VIRT_ADDRESS(sg) (sg_virt((sg)))
Jan Beulich93fbff62007-02-05 18:49:45 -080039#define SG_ENT_PHYS_ADDRESS(sg) virt_to_bus(SG_ENT_VIRT_ADDRESS(sg))
Linus Torvalds1da177e2005-04-16 15:20:36 -070040
41/*
42 * Maximum allowable number of contiguous slabs to map,
43 * must be a power of 2. What is the appropriate value ?
44 * The complexity of {map,unmap}_single is linearly dependent on this value.
45 */
46#define IO_TLB_SEGSIZE 128
47
48/*
49 * log of the size of each IO TLB slab. The number of slabs is command line
50 * controllable.
51 */
52#define IO_TLB_SHIFT 11
53
Alex Williamson0b9afed2005-09-06 11:20:49 -060054#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
55
56/*
57 * Minimum IO TLB size to bother booting with. Systems with mainly
58 * 64bit capable cards will only lightly use the swiotlb. If we can't
59 * allocate a contiguous 1MB, we're probably in trouble anyway.
60 */
61#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
62
John W. Linvillede69e0f2005-09-29 14:44:57 -070063/*
64 * Enumeration for sync targets
65 */
66enum dma_sync_target {
67 SYNC_FOR_CPU = 0,
68 SYNC_FOR_DEVICE = 1,
69};
70
Linus Torvalds1da177e2005-04-16 15:20:36 -070071int swiotlb_force;
72
73/*
74 * Used to do a quick range check in swiotlb_unmap_single and
75 * swiotlb_sync_single_*, to see if the memory was in fact allocated by this
76 * API.
77 */
78static char *io_tlb_start, *io_tlb_end;
79
80/*
81 * The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and
82 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
83 */
84static unsigned long io_tlb_nslabs;
85
86/*
87 * When the IOMMU overflows we return a fallback buffer. This sets the size.
88 */
89static unsigned long io_tlb_overflow = 32*1024;
90
91void *io_tlb_overflow_buffer;
92
93/*
94 * This is a free list describing the number of free entries available from
95 * each index
96 */
97static unsigned int *io_tlb_list;
98static unsigned int io_tlb_index;
99
100/*
101 * We need to save away the original address corresponding to a mapped entry
102 * for the sync operations.
103 */
Tony Luck25667d62007-03-06 13:31:45 -0800104static unsigned char **io_tlb_orig_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
106/*
107 * Protect the above data structures in the map and unmap calls
108 */
109static DEFINE_SPINLOCK(io_tlb_lock);
110
111static int __init
112setup_io_tlb_npages(char *str)
113{
114 if (isdigit(*str)) {
Alex Williamsone8579e72005-08-04 13:06:00 -0700115 io_tlb_nslabs = simple_strtoul(str, &str, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116 /* avoid tail segment of size < IO_TLB_SEGSIZE */
117 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
118 }
119 if (*str == ',')
120 ++str;
121 if (!strcmp(str, "force"))
122 swiotlb_force = 1;
123 return 1;
124}
125__setup("swiotlb=", setup_io_tlb_npages);
126/* make io_tlb_overflow tunable too? */
127
128/*
129 * Statically reserve bounce buffer space and initialize bounce buffer data
Tony Luck17e5ad62005-09-29 15:52:13 -0700130 * structures for the software IO TLB used to implement the DMA API.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800132void __init
133swiotlb_init_with_default_size(size_t default_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134{
Jan Beulich563aaf02007-02-05 18:51:25 -0800135 unsigned long i, bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136
137 if (!io_tlb_nslabs) {
Alex Williamsone8579e72005-08-04 13:06:00 -0700138 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
140 }
141
Jan Beulich563aaf02007-02-05 18:51:25 -0800142 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
143
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 /*
145 * Get IO TLB memory from the low pages
146 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800147 io_tlb_start = alloc_bootmem_low_pages(bytes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148 if (!io_tlb_start)
149 panic("Cannot allocate SWIOTLB buffer");
Jan Beulich563aaf02007-02-05 18:51:25 -0800150 io_tlb_end = io_tlb_start + bytes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151
152 /*
153 * Allocate and initialize the free list array. This array is used
154 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
155 * between io_tlb_start and io_tlb_end.
156 */
157 io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
Tony Luck25667d62007-03-06 13:31:45 -0800158 for (i = 0; i < io_tlb_nslabs; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
160 io_tlb_index = 0;
Tony Luck25667d62007-03-06 13:31:45 -0800161 io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(char *));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162
163 /*
164 * Get the overflow emergency buffer
165 */
166 io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
Jan Beulich563aaf02007-02-05 18:51:25 -0800167 if (!io_tlb_overflow_buffer)
168 panic("Cannot allocate SWIOTLB overflow buffer!\n");
169
Tony Luck25667d62007-03-06 13:31:45 -0800170 printk(KERN_INFO "Placing software IO TLB between 0x%lx - 0x%lx\n",
171 virt_to_bus(io_tlb_start), virt_to_bus(io_tlb_end));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172}
173
Jan Beulich563aaf02007-02-05 18:51:25 -0800174void __init
175swiotlb_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176{
Tony Luck25667d62007-03-06 13:31:45 -0800177 swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178}
179
Alex Williamson0b9afed2005-09-06 11:20:49 -0600180/*
181 * Systems with larger DMA zones (those that don't support ISA) can
182 * initialize the swiotlb later using the slab allocator if needed.
183 * This should be just like above, but with some error catching.
184 */
185int
Jan Beulich563aaf02007-02-05 18:51:25 -0800186swiotlb_late_init_with_default_size(size_t default_size)
Alex Williamson0b9afed2005-09-06 11:20:49 -0600187{
Jan Beulich563aaf02007-02-05 18:51:25 -0800188 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600189 unsigned int order;
190
191 if (!io_tlb_nslabs) {
192 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
193 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
194 }
195
196 /*
197 * Get IO TLB memory from the low pages
198 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800199 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600200 io_tlb_nslabs = SLABS_PER_PAGE << order;
Jan Beulich563aaf02007-02-05 18:51:25 -0800201 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600202
203 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
204 io_tlb_start = (char *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
205 order);
206 if (io_tlb_start)
207 break;
208 order--;
209 }
210
211 if (!io_tlb_start)
212 goto cleanup1;
213
Jan Beulich563aaf02007-02-05 18:51:25 -0800214 if (order != get_order(bytes)) {
Alex Williamson0b9afed2005-09-06 11:20:49 -0600215 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
216 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
217 io_tlb_nslabs = SLABS_PER_PAGE << order;
Jan Beulich563aaf02007-02-05 18:51:25 -0800218 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600219 }
Jan Beulich563aaf02007-02-05 18:51:25 -0800220 io_tlb_end = io_tlb_start + bytes;
221 memset(io_tlb_start, 0, bytes);
Alex Williamson0b9afed2005-09-06 11:20:49 -0600222
223 /*
224 * Allocate and initialize the free list array. This array is used
225 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
226 * between io_tlb_start and io_tlb_end.
227 */
228 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
229 get_order(io_tlb_nslabs * sizeof(int)));
230 if (!io_tlb_list)
231 goto cleanup2;
232
233 for (i = 0; i < io_tlb_nslabs; i++)
234 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
235 io_tlb_index = 0;
236
Tony Luck25667d62007-03-06 13:31:45 -0800237 io_tlb_orig_addr = (unsigned char **)__get_free_pages(GFP_KERNEL,
238 get_order(io_tlb_nslabs * sizeof(char *)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600239 if (!io_tlb_orig_addr)
240 goto cleanup3;
241
Tony Luck25667d62007-03-06 13:31:45 -0800242 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(char *));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600243
244 /*
245 * Get the overflow emergency buffer
246 */
247 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
248 get_order(io_tlb_overflow));
249 if (!io_tlb_overflow_buffer)
250 goto cleanup4;
251
Tony Luck25667d62007-03-06 13:31:45 -0800252 printk(KERN_INFO "Placing %luMB software IO TLB between 0x%lx - "
253 "0x%lx\n", bytes >> 20,
254 virt_to_bus(io_tlb_start), virt_to_bus(io_tlb_end));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600255
256 return 0;
257
258cleanup4:
Tony Luck25667d62007-03-06 13:31:45 -0800259 free_pages((unsigned long)io_tlb_orig_addr, get_order(io_tlb_nslabs *
260 sizeof(char *)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600261 io_tlb_orig_addr = NULL;
262cleanup3:
Tony Luck25667d62007-03-06 13:31:45 -0800263 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
264 sizeof(int)));
Alex Williamson0b9afed2005-09-06 11:20:49 -0600265 io_tlb_list = NULL;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600266cleanup2:
Jan Beulich563aaf02007-02-05 18:51:25 -0800267 io_tlb_end = NULL;
Alex Williamson0b9afed2005-09-06 11:20:49 -0600268 free_pages((unsigned long)io_tlb_start, order);
269 io_tlb_start = NULL;
270cleanup1:
271 io_tlb_nslabs = req_nslabs;
272 return -ENOMEM;
273}
274
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800275static int
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276address_needs_mapping(struct device *hwdev, dma_addr_t addr)
277{
278 dma_addr_t mask = 0xffffffff;
279 /* If the device has a mask, use it, otherwise default to 32 bits */
280 if (hwdev && hwdev->dma_mask)
281 mask = *hwdev->dma_mask;
282 return (addr & ~mask) != 0;
283}
284
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800285static inline unsigned int is_span_boundary(unsigned int index,
286 unsigned int nslots,
287 unsigned long offset_slots,
288 unsigned long max_slots)
289{
290 unsigned long offset = (offset_slots + index) & (max_slots - 1);
291 return offset + nslots > max_slots;
292}
293
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294/*
295 * Allocates bounce buffer and returns its kernel virtual address.
296 */
297static void *
Tony Luck25667d62007-03-06 13:31:45 -0800298map_single(struct device *hwdev, char *buffer, size_t size, int dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299{
300 unsigned long flags;
301 char *dma_addr;
302 unsigned int nslots, stride, index, wrap;
303 int i;
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800304 unsigned long start_dma_addr;
305 unsigned long mask;
306 unsigned long offset_slots;
307 unsigned long max_slots;
308
309 mask = dma_get_seg_boundary(hwdev);
310 start_dma_addr = virt_to_bus(io_tlb_start) & mask;
311
312 offset_slots = ALIGN(start_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
Jan Beulichb15a38912008-03-13 09:13:30 +0000313 max_slots = mask + 1
314 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
315 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316
317 /*
318 * For mappings greater than a page, we limit the stride (and
319 * hence alignment) to a page size.
320 */
321 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
322 if (size > PAGE_SIZE)
323 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
324 else
325 stride = 1;
326
Eric Sesterhenn34814542006-03-24 18:47:11 +0100327 BUG_ON(!nslots);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328
329 /*
330 * Find suitable number of IO TLB entries size that will fit this
331 * request and allocate a buffer from that IO TLB pool.
332 */
333 spin_lock_irqsave(&io_tlb_lock, flags);
334 {
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800335 index = ALIGN(io_tlb_index, stride);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 if (index >= io_tlb_nslabs)
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800337 index = 0;
FUJITA Tomonori681cc5c2008-02-04 22:28:16 -0800338 wrap = index;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339
340 do {
Jan Beulichb15a38912008-03-13 09:13:30 +0000341 while (is_span_boundary(index, nslots, offset_slots,
342 max_slots)) {
343 index += stride;
344 if (index >= io_tlb_nslabs)
345 index = 0;
346 if (index == wrap)
347 goto not_found;
348 }
349
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350 /*
351 * If we find a slot that indicates we have 'nslots'
352 * number of contiguous buffers, we allocate the
353 * buffers from that slot and mark the entries as '0'
354 * indicating unavailable.
355 */
356 if (io_tlb_list[index] >= nslots) {
357 int count = 0;
358
359 for (i = index; i < (int) (index + nslots); i++)
360 io_tlb_list[i] = 0;
361 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
362 io_tlb_list[i] = ++count;
363 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
364
365 /*
366 * Update the indices to avoid searching in
367 * the next round.
368 */
369 io_tlb_index = ((index + nslots) < io_tlb_nslabs
370 ? (index + nslots) : 0);
371
372 goto found;
373 }
Jan Beulichb15a38912008-03-13 09:13:30 +0000374 index += stride;
375 if (index >= io_tlb_nslabs)
376 index = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 } while (index != wrap);
378
Jan Beulichb15a38912008-03-13 09:13:30 +0000379 not_found:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 spin_unlock_irqrestore(&io_tlb_lock, flags);
381 return NULL;
382 }
383 found:
384 spin_unlock_irqrestore(&io_tlb_lock, flags);
385
386 /*
387 * Save away the mapping from the original address to the DMA address.
388 * This is needed when we sync the memory. Then we sync the buffer if
389 * needed.
390 */
Keir Fraserdf336d12007-07-21 04:37:24 -0700391 for (i = 0; i < nslots; i++)
392 io_tlb_orig_addr[index+i] = buffer + (i << IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
Tony Luck25667d62007-03-06 13:31:45 -0800394 memcpy(dma_addr, buffer, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395
396 return dma_addr;
397}
398
399/*
400 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
401 */
402static void
403unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
404{
405 unsigned long flags;
406 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
407 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
Tony Luck25667d62007-03-06 13:31:45 -0800408 char *buffer = io_tlb_orig_addr[index];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409
410 /*
411 * First, sync the memory before unmapping the entry
412 */
Tony Luck25667d62007-03-06 13:31:45 -0800413 if (buffer && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 /*
415 * bounce... copy the data back into the original buffer * and
416 * delete the bounce buffer.
417 */
Tony Luck25667d62007-03-06 13:31:45 -0800418 memcpy(buffer, dma_addr, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419
420 /*
421 * Return the buffer to the free list by setting the corresponding
422 * entries to indicate the number of contigous entries available.
423 * While returning the entries to the free list, we merge the entries
424 * with slots below and above the pool being returned.
425 */
426 spin_lock_irqsave(&io_tlb_lock, flags);
427 {
428 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
429 io_tlb_list[index + nslots] : 0);
430 /*
431 * Step 1: return the slots to the free list, merging the
432 * slots with superceeding slots
433 */
434 for (i = index + nslots - 1; i >= index; i--)
435 io_tlb_list[i] = ++count;
436 /*
437 * Step 2: merge the returned slots with the preceding slots,
438 * if available (non zero)
439 */
440 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
441 io_tlb_list[i] = ++count;
442 }
443 spin_unlock_irqrestore(&io_tlb_lock, flags);
444}
445
446static void
John W. Linvillede69e0f2005-09-29 14:44:57 -0700447sync_single(struct device *hwdev, char *dma_addr, size_t size,
448 int dir, int target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449{
450 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
Tony Luck25667d62007-03-06 13:31:45 -0800451 char *buffer = io_tlb_orig_addr[index];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452
Keir Fraserdf336d12007-07-21 04:37:24 -0700453 buffer += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
454
John W. Linvillede69e0f2005-09-29 14:44:57 -0700455 switch (target) {
456 case SYNC_FOR_CPU:
457 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
Tony Luck25667d62007-03-06 13:31:45 -0800458 memcpy(buffer, dma_addr, size);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100459 else
460 BUG_ON(dir != DMA_TO_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700461 break;
462 case SYNC_FOR_DEVICE:
463 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
Tony Luck25667d62007-03-06 13:31:45 -0800464 memcpy(dma_addr, buffer, size);
Eric Sesterhenn34814542006-03-24 18:47:11 +0100465 else
466 BUG_ON(dir != DMA_FROM_DEVICE);
John W. Linvillede69e0f2005-09-29 14:44:57 -0700467 break;
468 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 BUG();
John W. Linvillede69e0f2005-09-29 14:44:57 -0700470 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471}
472
473void *
474swiotlb_alloc_coherent(struct device *hwdev, size_t size,
Al Viro06a54492005-10-21 03:21:03 -0400475 dma_addr_t *dma_handle, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476{
Jan Beulich563aaf02007-02-05 18:51:25 -0800477 dma_addr_t dev_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 void *ret;
479 int order = get_order(size);
480
481 /*
482 * XXX fix me: the DMA API should pass us an explicit DMA mask
483 * instead, or use ZONE_DMA32 (ia64 overloads ZONE_DMA to be a ~32
484 * bit range instead of a 16MB one).
485 */
486 flags |= GFP_DMA;
487
Tony Luck25667d62007-03-06 13:31:45 -0800488 ret = (void *)__get_free_pages(flags, order);
Jan Beulich93fbff62007-02-05 18:49:45 -0800489 if (ret && address_needs_mapping(hwdev, virt_to_bus(ret))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 /*
491 * The allocated memory isn't reachable by the device.
492 * Fall back on swiotlb_map_single().
493 */
494 free_pages((unsigned long) ret, order);
495 ret = NULL;
496 }
497 if (!ret) {
498 /*
499 * We are either out of memory or the device can't DMA
500 * to GFP_DMA memory; fall back on
501 * swiotlb_map_single(), which will grab memory from
502 * the lowest available address range.
503 */
504 dma_addr_t handle;
505 handle = swiotlb_map_single(NULL, NULL, size, DMA_FROM_DEVICE);
Muli Ben-Yehuda17a941d2006-01-11 22:44:42 +0100506 if (swiotlb_dma_mapping_error(handle))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 return NULL;
508
Jan Beulich93fbff62007-02-05 18:49:45 -0800509 ret = bus_to_virt(handle);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 }
511
512 memset(ret, 0, size);
Jan Beulich93fbff62007-02-05 18:49:45 -0800513 dev_addr = virt_to_bus(ret);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514
515 /* Confirm address can be DMA'd by device */
516 if (address_needs_mapping(hwdev, dev_addr)) {
Jan Beulich563aaf02007-02-05 18:51:25 -0800517 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
518 (unsigned long long)*hwdev->dma_mask,
519 (unsigned long long)dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 panic("swiotlb_alloc_coherent: allocated memory is out of "
521 "range for device");
522 }
523 *dma_handle = dev_addr;
524 return ret;
525}
526
527void
528swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
529 dma_addr_t dma_handle)
530{
David Brownellaa248862007-08-10 13:10:27 -0700531 WARN_ON(irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 if (!(vaddr >= (void *)io_tlb_start
533 && vaddr < (void *)io_tlb_end))
534 free_pages((unsigned long) vaddr, get_order(size));
535 else
536 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
537 swiotlb_unmap_single (hwdev, dma_handle, size, DMA_TO_DEVICE);
538}
539
540static void
541swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
542{
543 /*
544 * Ran out of IOMMU space for this operation. This is very bad.
545 * Unfortunately the drivers cannot handle this operation properly.
Tony Luck17e5ad62005-09-29 15:52:13 -0700546 * unless they check for dma_mapping_error (most don't)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 * When the mapping is small enough return a static buffer to limit
548 * the damage, or panic when the transfer is too big.
549 */
Jan Beulich563aaf02007-02-05 18:51:25 -0800550 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 "device %s\n", size, dev ? dev->bus_id : "?");
552
553 if (size > io_tlb_overflow && do_panic) {
Tony Luck17e5ad62005-09-29 15:52:13 -0700554 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
555 panic("DMA: Memory would be corrupted\n");
556 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
557 panic("DMA: Random memory would be DMAed\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558 }
559}
560
561/*
562 * Map a single buffer of the indicated size for DMA in streaming mode. The
Tony Luck17e5ad62005-09-29 15:52:13 -0700563 * physical address to use is returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 *
565 * Once the device is given the dma address, the device owns this memory until
566 * either swiotlb_unmap_single or swiotlb_dma_sync_single is performed.
567 */
568dma_addr_t
569swiotlb_map_single(struct device *hwdev, void *ptr, size_t size, int dir)
570{
Jan Beulich563aaf02007-02-05 18:51:25 -0800571 dma_addr_t dev_addr = virt_to_bus(ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 void *map;
573
Eric Sesterhenn34814542006-03-24 18:47:11 +0100574 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 /*
576 * If the pointer passed in happens to be in the device's DMA window,
577 * we can safely return the device addr and not worry about bounce
578 * buffering it.
579 */
Tony Luck25667d62007-03-06 13:31:45 -0800580 if (!address_needs_mapping(hwdev, dev_addr) && !swiotlb_force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581 return dev_addr;
582
583 /*
584 * Oh well, have to allocate and map a bounce buffer.
585 */
Tony Luck25667d62007-03-06 13:31:45 -0800586 map = map_single(hwdev, ptr, size, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 if (!map) {
588 swiotlb_full(hwdev, size, dir, 1);
589 map = io_tlb_overflow_buffer;
590 }
591
Jan Beulich93fbff62007-02-05 18:49:45 -0800592 dev_addr = virt_to_bus(map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593
594 /*
595 * Ensure that the address returned is DMA'ble
596 */
597 if (address_needs_mapping(hwdev, dev_addr))
598 panic("map_single: bounce buffer is not DMA'ble");
599
600 return dev_addr;
601}
602
603/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604 * Unmap a single streaming mode DMA translation. The dma_addr and size must
605 * match what was provided for in a previous swiotlb_map_single call. All
606 * other usages are undefined.
607 *
608 * After this call, reads by the cpu to the buffer are guaranteed to see
609 * whatever the device wrote there.
610 */
611void
612swiotlb_unmap_single(struct device *hwdev, dma_addr_t dev_addr, size_t size,
613 int dir)
614{
Jan Beulich93fbff62007-02-05 18:49:45 -0800615 char *dma_addr = bus_to_virt(dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616
Eric Sesterhenn34814542006-03-24 18:47:11 +0100617 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
619 unmap_single(hwdev, dma_addr, size, dir);
620 else if (dir == DMA_FROM_DEVICE)
Jan Beulichcde14bb2007-02-05 18:46:40 -0800621 dma_mark_clean(dma_addr, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622}
623
624/*
625 * Make physical memory consistent for a single streaming mode DMA translation
626 * after a transfer.
627 *
628 * If you perform a swiotlb_map_single() but wish to interrogate the buffer
Tony Luck17e5ad62005-09-29 15:52:13 -0700629 * using the cpu, yet do not wish to teardown the dma mapping, you must
630 * call this function before doing so. At the next point you give the dma
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631 * address back to the card, you must first perform a
632 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
633 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800634static void
John W. Linville8270f3f2005-09-29 14:43:32 -0700635swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700636 size_t size, int dir, int target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637{
Jan Beulich93fbff62007-02-05 18:49:45 -0800638 char *dma_addr = bus_to_virt(dev_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
Eric Sesterhenn34814542006-03-24 18:47:11 +0100640 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
John W. Linvillede69e0f2005-09-29 14:44:57 -0700642 sync_single(hwdev, dma_addr, size, dir, target);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 else if (dir == DMA_FROM_DEVICE)
Jan Beulichcde14bb2007-02-05 18:46:40 -0800644 dma_mark_clean(dma_addr, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645}
646
647void
John W. Linville8270f3f2005-09-29 14:43:32 -0700648swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
649 size_t size, int dir)
650{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700651 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700652}
653
654void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
656 size_t size, int dir)
657{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700658 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659}
660
661/*
John W. Linville878a97c2005-09-29 14:44:23 -0700662 * Same as above, but for a sub-range of the mapping.
663 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800664static void
John W. Linville878a97c2005-09-29 14:44:23 -0700665swiotlb_sync_single_range(struct device *hwdev, dma_addr_t dev_addr,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700666 unsigned long offset, size_t size,
667 int dir, int target)
John W. Linville878a97c2005-09-29 14:44:23 -0700668{
Jan Beulich93fbff62007-02-05 18:49:45 -0800669 char *dma_addr = bus_to_virt(dev_addr) + offset;
John W. Linville878a97c2005-09-29 14:44:23 -0700670
Eric Sesterhenn34814542006-03-24 18:47:11 +0100671 BUG_ON(dir == DMA_NONE);
John W. Linville878a97c2005-09-29 14:44:23 -0700672 if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
John W. Linvillede69e0f2005-09-29 14:44:57 -0700673 sync_single(hwdev, dma_addr, size, dir, target);
John W. Linville878a97c2005-09-29 14:44:23 -0700674 else if (dir == DMA_FROM_DEVICE)
Jan Beulichcde14bb2007-02-05 18:46:40 -0800675 dma_mark_clean(dma_addr, size);
John W. Linville878a97c2005-09-29 14:44:23 -0700676}
677
678void
679swiotlb_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
680 unsigned long offset, size_t size, int dir)
681{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700682 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
683 SYNC_FOR_CPU);
John W. Linville878a97c2005-09-29 14:44:23 -0700684}
685
686void
687swiotlb_sync_single_range_for_device(struct device *hwdev, dma_addr_t dev_addr,
688 unsigned long offset, size_t size, int dir)
689{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700690 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
691 SYNC_FOR_DEVICE);
John W. Linville878a97c2005-09-29 14:44:23 -0700692}
693
694/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 * Map a set of buffers described by scatterlist in streaming mode for DMA.
696 * This is the scatter-gather version of the above swiotlb_map_single
697 * interface. Here the scatter gather list elements are each tagged with the
698 * appropriate dma address and length. They are obtained via
699 * sg_dma_{address,length}(SG).
700 *
701 * NOTE: An implementation may be able to use a smaller number of
702 * DMA address/length pairs than there are SG table elements.
703 * (for example via virtual mapping capabilities)
704 * The routine returns the number of addr/length pairs actually
705 * used, at most nents.
706 *
707 * Device ownership issues as mentioned above for swiotlb_map_single are the
708 * same here.
709 */
710int
Jens Axboedbfd49f2007-05-11 14:56:18 +0200711swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 int dir)
713{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200714 struct scatterlist *sg;
Tony Luck25667d62007-03-06 13:31:45 -0800715 void *addr;
Jan Beulich563aaf02007-02-05 18:51:25 -0800716 dma_addr_t dev_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717 int i;
718
Eric Sesterhenn34814542006-03-24 18:47:11 +0100719 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720
Jens Axboedbfd49f2007-05-11 14:56:18 +0200721 for_each_sg(sgl, sg, nelems, i) {
Tony Luck25667d62007-03-06 13:31:45 -0800722 addr = SG_ENT_VIRT_ADDRESS(sg);
723 dev_addr = virt_to_bus(addr);
724 if (swiotlb_force || address_needs_mapping(hwdev, dev_addr)) {
725 void *map = map_single(hwdev, addr, sg->length, dir);
Andi Kleen7e870232005-12-20 14:45:19 +0100726 if (!map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 /* Don't panic here, we expect map_sg users
728 to do proper error handling. */
729 swiotlb_full(hwdev, sg->length, dir, 0);
FUJITA Tomonori9e91fdb2007-10-17 10:06:18 +0200730 swiotlb_unmap_sg(hwdev, sgl, i, dir);
Jens Axboedbfd49f2007-05-11 14:56:18 +0200731 sgl[0].dma_length = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 return 0;
733 }
Jan Beulichcde14bb2007-02-05 18:46:40 -0800734 sg->dma_address = virt_to_bus(map);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735 } else
736 sg->dma_address = dev_addr;
737 sg->dma_length = sg->length;
738 }
739 return nelems;
740}
741
742/*
743 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
744 * concerning calls here are the same as for swiotlb_unmap_single() above.
745 */
746void
Jens Axboedbfd49f2007-05-11 14:56:18 +0200747swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 int dir)
749{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200750 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 int i;
752
Eric Sesterhenn34814542006-03-24 18:47:11 +0100753 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754
Jens Axboedbfd49f2007-05-11 14:56:18 +0200755 for_each_sg(sgl, sg, nelems, i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
Jan Beulich93fbff62007-02-05 18:49:45 -0800757 unmap_single(hwdev, bus_to_virt(sg->dma_address),
758 sg->dma_length, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 else if (dir == DMA_FROM_DEVICE)
Jan Beulichcde14bb2007-02-05 18:46:40 -0800760 dma_mark_clean(SG_ENT_VIRT_ADDRESS(sg), sg->dma_length);
Jens Axboedbfd49f2007-05-11 14:56:18 +0200761 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762}
763
764/*
765 * Make physical memory consistent for a set of streaming mode DMA translations
766 * after a transfer.
767 *
768 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
769 * and usage.
770 */
Andrew Mortonbe6b0262007-02-12 00:52:17 -0800771static void
Jens Axboedbfd49f2007-05-11 14:56:18 +0200772swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700773 int nelems, int dir, int target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774{
Jens Axboedbfd49f2007-05-11 14:56:18 +0200775 struct scatterlist *sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 int i;
777
Eric Sesterhenn34814542006-03-24 18:47:11 +0100778 BUG_ON(dir == DMA_NONE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779
Jens Axboedbfd49f2007-05-11 14:56:18 +0200780 for_each_sg(sgl, sg, nelems, i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
Jan Beulich93fbff62007-02-05 18:49:45 -0800782 sync_single(hwdev, bus_to_virt(sg->dma_address),
John W. Linvillede69e0f2005-09-29 14:44:57 -0700783 sg->dma_length, dir, target);
Jan Beulichcde14bb2007-02-05 18:46:40 -0800784 else if (dir == DMA_FROM_DEVICE)
785 dma_mark_clean(SG_ENT_VIRT_ADDRESS(sg), sg->dma_length);
Jens Axboedbfd49f2007-05-11 14:56:18 +0200786 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787}
788
789void
John W. Linville8270f3f2005-09-29 14:43:32 -0700790swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
791 int nelems, int dir)
792{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700793 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700794}
795
796void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
798 int nelems, int dir)
799{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700800 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801}
802
803int
804swiotlb_dma_mapping_error(dma_addr_t dma_addr)
805{
Jan Beulich93fbff62007-02-05 18:49:45 -0800806 return (dma_addr == virt_to_bus(io_tlb_overflow_buffer));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807}
808
809/*
Tony Luck17e5ad62005-09-29 15:52:13 -0700810 * Return whether the given device DMA address mask can be supported
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 * properly. For example, if your device can only drive the low 24-bits
Tony Luck17e5ad62005-09-29 15:52:13 -0700812 * during bus mastering, then you would pass 0x00ffffff as the mask to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813 * this function.
814 */
815int
Jan Beulich563aaf02007-02-05 18:51:25 -0800816swiotlb_dma_supported(struct device *hwdev, u64 mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817{
Tony Luck25667d62007-03-06 13:31:45 -0800818 return virt_to_bus(io_tlb_end - 1) <= mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819}
820
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821EXPORT_SYMBOL(swiotlb_map_single);
822EXPORT_SYMBOL(swiotlb_unmap_single);
823EXPORT_SYMBOL(swiotlb_map_sg);
824EXPORT_SYMBOL(swiotlb_unmap_sg);
825EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
826EXPORT_SYMBOL(swiotlb_sync_single_for_device);
John W. Linville878a97c2005-09-29 14:44:23 -0700827EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu);
828EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
830EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
831EXPORT_SYMBOL(swiotlb_dma_mapping_error);
Tony Luck25667d62007-03-06 13:31:45 -0800832EXPORT_SYMBOL(swiotlb_alloc_coherent);
833EXPORT_SYMBOL(swiotlb_free_coherent);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834EXPORT_SYMBOL(swiotlb_dma_supported);