Merge branch 'core-objtool-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[deliverable/linux.git] / arch / powerpc / sysdev / dart_iommu.c
1 /*
2 * arch/powerpc/sysdev/dart_iommu.c
3 *
4 * Copyright (C) 2004 Olof Johansson <olof@lixom.net>, IBM Corporation
5 * Copyright (C) 2005 Benjamin Herrenschmidt <benh@kernel.crashing.org>,
6 * IBM Corporation
7 *
8 * Based on pSeries_iommu.c:
9 * Copyright (C) 2001 Mike Corrigan & Dave Engebretsen, IBM Corporation
10 * Copyright (C) 2004 Olof Johansson <olof@lixom.net>, IBM Corporation
11 *
12 * Dynamic DMA mapping support, Apple U3, U4 & IBM CPC925 "DART" iommu.
13 *
14 *
15 * This program is free software; you can redistribute it and/or modify
16 * it under the terms of the GNU General Public License as published by
17 * the Free Software Foundation; either version 2 of the License, or
18 * (at your option) any later version.
19 *
20 * This program is distributed in the hope that it will be useful,
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
23 * GNU General Public License for more details.
24 *
25 * You should have received a copy of the GNU General Public License
26 * along with this program; if not, write to the Free Software
27 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
28 */
29
30 #include <linux/init.h>
31 #include <linux/types.h>
32 #include <linux/mm.h>
33 #include <linux/spinlock.h>
34 #include <linux/string.h>
35 #include <linux/pci.h>
36 #include <linux/dma-mapping.h>
37 #include <linux/vmalloc.h>
38 #include <linux/suspend.h>
39 #include <linux/memblock.h>
40 #include <linux/gfp.h>
41 #include <asm/io.h>
42 #include <asm/prom.h>
43 #include <asm/iommu.h>
44 #include <asm/pci-bridge.h>
45 #include <asm/machdep.h>
46 #include <asm/cacheflush.h>
47 #include <asm/ppc-pci.h>
48
49 #include "dart.h"
50
51 /* Physical base address and size of the DART table */
52 unsigned long dart_tablebase; /* exported to htab_initialize */
53 static unsigned long dart_tablesize;
54
55 /* Virtual base address of the DART table */
56 static u32 *dart_vbase;
57 #ifdef CONFIG_PM
58 static u32 *dart_copy;
59 #endif
60
61 /* Mapped base address for the dart */
62 static unsigned int __iomem *dart;
63
64 /* Dummy val that entries are set to when unused */
65 static unsigned int dart_emptyval;
66
67 static struct iommu_table iommu_table_dart;
68 static int iommu_table_dart_inited;
69 static int dart_dirty;
70 static int dart_is_u4;
71
72 #define DART_U4_BYPASS_BASE 0x8000000000ull
73
74 #define DBG(...)
75
76 static DEFINE_SPINLOCK(invalidate_lock);
77
78 static inline void dart_tlb_invalidate_all(void)
79 {
80 unsigned long l = 0;
81 unsigned int reg, inv_bit;
82 unsigned long limit;
83 unsigned long flags;
84
85 spin_lock_irqsave(&invalidate_lock, flags);
86
87 DBG("dart: flush\n");
88
89 /* To invalidate the DART, set the DARTCNTL_FLUSHTLB bit in the
90 * control register and wait for it to clear.
91 *
92 * Gotcha: Sometimes, the DART won't detect that the bit gets
93 * set. If so, clear it and set it again.
94 */
95
96 limit = 0;
97
98 inv_bit = dart_is_u4 ? DART_CNTL_U4_FLUSHTLB : DART_CNTL_U3_FLUSHTLB;
99 retry:
100 l = 0;
101 reg = DART_IN(DART_CNTL);
102 reg |= inv_bit;
103 DART_OUT(DART_CNTL, reg);
104
105 while ((DART_IN(DART_CNTL) & inv_bit) && l < (1L << limit))
106 l++;
107 if (l == (1L << limit)) {
108 if (limit < 4) {
109 limit++;
110 reg = DART_IN(DART_CNTL);
111 reg &= ~inv_bit;
112 DART_OUT(DART_CNTL, reg);
113 goto retry;
114 } else
115 panic("DART: TLB did not flush after waiting a long "
116 "time. Buggy U3 ?");
117 }
118
119 spin_unlock_irqrestore(&invalidate_lock, flags);
120 }
121
122 static inline void dart_tlb_invalidate_one(unsigned long bus_rpn)
123 {
124 unsigned int reg;
125 unsigned int l, limit;
126 unsigned long flags;
127
128 spin_lock_irqsave(&invalidate_lock, flags);
129
130 reg = DART_CNTL_U4_ENABLE | DART_CNTL_U4_IONE |
131 (bus_rpn & DART_CNTL_U4_IONE_MASK);
132 DART_OUT(DART_CNTL, reg);
133
134 limit = 0;
135 wait_more:
136 l = 0;
137 while ((DART_IN(DART_CNTL) & DART_CNTL_U4_IONE) && l < (1L << limit)) {
138 rmb();
139 l++;
140 }
141
142 if (l == (1L << limit)) {
143 if (limit < 4) {
144 limit++;
145 goto wait_more;
146 } else
147 panic("DART: TLB did not flush after waiting a long "
148 "time. Buggy U4 ?");
149 }
150
151 spin_unlock_irqrestore(&invalidate_lock, flags);
152 }
153
154 static void dart_flush(struct iommu_table *tbl)
155 {
156 mb();
157 if (dart_dirty) {
158 dart_tlb_invalidate_all();
159 dart_dirty = 0;
160 }
161 }
162
163 static int dart_build(struct iommu_table *tbl, long index,
164 long npages, unsigned long uaddr,
165 enum dma_data_direction direction,
166 struct dma_attrs *attrs)
167 {
168 unsigned int *dp;
169 unsigned int rpn;
170 long l;
171
172 DBG("dart: build at: %lx, %lx, addr: %x\n", index, npages, uaddr);
173
174 dp = ((unsigned int*)tbl->it_base) + index;
175
176 /* On U3, all memory is contiguous, so we can move this
177 * out of the loop.
178 */
179 l = npages;
180 while (l--) {
181 rpn = __pa(uaddr) >> DART_PAGE_SHIFT;
182
183 *(dp++) = DARTMAP_VALID | (rpn & DARTMAP_RPNMASK);
184
185 uaddr += DART_PAGE_SIZE;
186 }
187
188 /* make sure all updates have reached memory */
189 mb();
190 in_be32((unsigned __iomem *)dp);
191 mb();
192
193 if (dart_is_u4) {
194 rpn = index;
195 while (npages--)
196 dart_tlb_invalidate_one(rpn++);
197 } else {
198 dart_dirty = 1;
199 }
200 return 0;
201 }
202
203
204 static void dart_free(struct iommu_table *tbl, long index, long npages)
205 {
206 unsigned int *dp;
207
208 /* We don't worry about flushing the TLB cache. The only drawback of
209 * not doing it is that we won't catch buggy device drivers doing
210 * bad DMAs, but then no 32-bit architecture ever does either.
211 */
212
213 DBG("dart: free at: %lx, %lx\n", index, npages);
214
215 dp = ((unsigned int *)tbl->it_base) + index;
216
217 while (npages--)
218 *(dp++) = dart_emptyval;
219 }
220
221
222 static int __init dart_init(struct device_node *dart_node)
223 {
224 unsigned int i;
225 unsigned long tmp, base, size;
226 struct resource r;
227
228 if (dart_tablebase == 0 || dart_tablesize == 0) {
229 printk(KERN_INFO "DART: table not allocated, using "
230 "direct DMA\n");
231 return -ENODEV;
232 }
233
234 if (of_address_to_resource(dart_node, 0, &r))
235 panic("DART: can't get register base ! ");
236
237 /* Make sure nothing from the DART range remains in the CPU cache
238 * from a previous mapping that existed before the kernel took
239 * over
240 */
241 flush_dcache_phys_range(dart_tablebase,
242 dart_tablebase + dart_tablesize);
243
244 /* Allocate a spare page to map all invalid DART pages. We need to do
245 * that to work around what looks like a problem with the HT bridge
246 * prefetching into invalid pages and corrupting data
247 */
248 tmp = memblock_alloc(DART_PAGE_SIZE, DART_PAGE_SIZE);
249 dart_emptyval = DARTMAP_VALID | ((tmp >> DART_PAGE_SHIFT) &
250 DARTMAP_RPNMASK);
251
252 /* Map in DART registers */
253 dart = ioremap(r.start, resource_size(&r));
254 if (dart == NULL)
255 panic("DART: Cannot map registers!");
256
257 /* Map in DART table */
258 dart_vbase = ioremap(__pa(dart_tablebase), dart_tablesize);
259
260 /* Fill initial table */
261 for (i = 0; i < dart_tablesize/4; i++)
262 dart_vbase[i] = dart_emptyval;
263
264 /* Initialize DART with table base and enable it. */
265 base = dart_tablebase >> DART_PAGE_SHIFT;
266 size = dart_tablesize >> DART_PAGE_SHIFT;
267 if (dart_is_u4) {
268 size &= DART_SIZE_U4_SIZE_MASK;
269 DART_OUT(DART_BASE_U4, base);
270 DART_OUT(DART_SIZE_U4, size);
271 DART_OUT(DART_CNTL, DART_CNTL_U4_ENABLE);
272 } else {
273 size &= DART_CNTL_U3_SIZE_MASK;
274 DART_OUT(DART_CNTL,
275 DART_CNTL_U3_ENABLE |
276 (base << DART_CNTL_U3_BASE_SHIFT) |
277 (size << DART_CNTL_U3_SIZE_SHIFT));
278 }
279
280 /* Invalidate DART to get rid of possible stale TLBs */
281 dart_tlb_invalidate_all();
282
283 printk(KERN_INFO "DART IOMMU initialized for %s type chipset\n",
284 dart_is_u4 ? "U4" : "U3");
285
286 return 0;
287 }
288
289 static struct iommu_table_ops iommu_dart_ops = {
290 .set = dart_build,
291 .clear = dart_free,
292 .flush = dart_flush,
293 };
294
295 static void iommu_table_dart_setup(void)
296 {
297 iommu_table_dart.it_busno = 0;
298 iommu_table_dart.it_offset = 0;
299 /* it_size is in number of entries */
300 iommu_table_dart.it_size = dart_tablesize / sizeof(u32);
301 iommu_table_dart.it_page_shift = IOMMU_PAGE_SHIFT_4K;
302
303 /* Initialize the common IOMMU code */
304 iommu_table_dart.it_base = (unsigned long)dart_vbase;
305 iommu_table_dart.it_index = 0;
306 iommu_table_dart.it_blocksize = 1;
307 iommu_table_dart.it_ops = &iommu_dart_ops;
308 iommu_init_table(&iommu_table_dart, -1);
309
310 /* Reserve the last page of the DART to avoid possible prefetch
311 * past the DART mapped area
312 */
313 set_bit(iommu_table_dart.it_size - 1, iommu_table_dart.it_map);
314 }
315
316 static void pci_dma_dev_setup_dart(struct pci_dev *dev)
317 {
318 if (dart_is_u4)
319 set_dma_offset(&dev->dev, DART_U4_BYPASS_BASE);
320 set_iommu_table_base(&dev->dev, &iommu_table_dart);
321 }
322
323 static void pci_dma_bus_setup_dart(struct pci_bus *bus)
324 {
325 if (!iommu_table_dart_inited) {
326 iommu_table_dart_inited = 1;
327 iommu_table_dart_setup();
328 }
329 }
330
331 static bool dart_device_on_pcie(struct device *dev)
332 {
333 struct device_node *np = of_node_get(dev->of_node);
334
335 while(np) {
336 if (of_device_is_compatible(np, "U4-pcie") ||
337 of_device_is_compatible(np, "u4-pcie")) {
338 of_node_put(np);
339 return true;
340 }
341 np = of_get_next_parent(np);
342 }
343 return false;
344 }
345
346 static int dart_dma_set_mask(struct device *dev, u64 dma_mask)
347 {
348 if (!dev->dma_mask || !dma_supported(dev, dma_mask))
349 return -EIO;
350
351 /* U4 supports a DART bypass, we use it for 64-bit capable
352 * devices to improve performances. However, that only works
353 * for devices connected to U4 own PCIe interface, not bridged
354 * through hypertransport. We need the device to support at
355 * least 40 bits of addresses.
356 */
357 if (dart_device_on_pcie(dev) && dma_mask >= DMA_BIT_MASK(40)) {
358 dev_info(dev, "Using 64-bit DMA iommu bypass\n");
359 set_dma_ops(dev, &dma_direct_ops);
360 } else {
361 dev_info(dev, "Using 32-bit DMA via iommu\n");
362 set_dma_ops(dev, &dma_iommu_ops);
363 }
364
365 *dev->dma_mask = dma_mask;
366 return 0;
367 }
368
369 void __init iommu_init_early_dart(struct pci_controller_ops *controller_ops)
370 {
371 struct device_node *dn;
372
373 /* Find the DART in the device-tree */
374 dn = of_find_compatible_node(NULL, "dart", "u3-dart");
375 if (dn == NULL) {
376 dn = of_find_compatible_node(NULL, "dart", "u4-dart");
377 if (dn == NULL)
378 return; /* use default direct_dma_ops */
379 dart_is_u4 = 1;
380 }
381
382 /* Initialize the DART HW */
383 if (dart_init(dn) != 0)
384 goto bail;
385
386 /* Setup bypass if supported */
387 if (dart_is_u4)
388 ppc_md.dma_set_mask = dart_dma_set_mask;
389
390 controller_ops->dma_dev_setup = pci_dma_dev_setup_dart;
391 controller_ops->dma_bus_setup = pci_dma_bus_setup_dart;
392
393 /* Setup pci_dma ops */
394 set_pci_dma_ops(&dma_iommu_ops);
395 return;
396
397 bail:
398 /* If init failed, use direct iommu and null setup functions */
399 controller_ops->dma_dev_setup = NULL;
400 controller_ops->dma_bus_setup = NULL;
401
402 /* Setup pci_dma ops */
403 set_pci_dma_ops(&dma_direct_ops);
404 }
405
406 #ifdef CONFIG_PM
407 static void iommu_dart_save(void)
408 {
409 memcpy(dart_copy, dart_vbase, 2*1024*1024);
410 }
411
412 static void iommu_dart_restore(void)
413 {
414 memcpy(dart_vbase, dart_copy, 2*1024*1024);
415 dart_tlb_invalidate_all();
416 }
417
418 static int __init iommu_init_late_dart(void)
419 {
420 unsigned long tbasepfn;
421 struct page *p;
422
423 /* if no dart table exists then we won't need to save it
424 * and the area has also not been reserved */
425 if (!dart_tablebase)
426 return 0;
427
428 tbasepfn = __pa(dart_tablebase) >> PAGE_SHIFT;
429 register_nosave_region_late(tbasepfn,
430 tbasepfn + ((1<<24) >> PAGE_SHIFT));
431
432 /* For suspend we need to copy the dart contents because
433 * it is not part of the regular mapping (see above) and
434 * thus not saved automatically. The memory for this copy
435 * must be allocated early because we need 2 MB. */
436 p = alloc_pages(GFP_KERNEL, 21 - PAGE_SHIFT);
437 BUG_ON(!p);
438 dart_copy = page_address(p);
439
440 ppc_md.iommu_save = iommu_dart_save;
441 ppc_md.iommu_restore = iommu_dart_restore;
442
443 return 0;
444 }
445
446 late_initcall(iommu_init_late_dart);
447 #endif
448
449 void __init alloc_dart_table(void)
450 {
451 /* Only reserve DART space if machine has more than 1GB of RAM
452 * or if requested with iommu=on on cmdline.
453 *
454 * 1GB of RAM is picked as limit because some default devices
455 * (i.e. Airport Extreme) have 30 bit address range limits.
456 */
457
458 if (iommu_is_off)
459 return;
460
461 if (!iommu_force_on && memblock_end_of_DRAM() <= 0x40000000ull)
462 return;
463
464 /* 512 pages (2MB) is max DART tablesize. */
465 dart_tablesize = 1UL << 21;
466 /* 16MB (1 << 24) alignment. We allocate a full 16Mb chuck since we
467 * will blow up an entire large page anyway in the kernel mapping
468 */
469 dart_tablebase = (unsigned long)
470 __va(memblock_alloc_base(1UL<<24, 1UL<<24, 0x80000000L));
471 /*
472 * The DART space is later unmapped from the kernel linear mapping and
473 * accessing dart_tablebase during kmemleak scanning will fault.
474 */
475 kmemleak_no_scan((void *)dart_tablebase);
476
477 printk(KERN_INFO "DART table allocated at: %lx\n", dart_tablebase);
478 }
This page took 0.04931 seconds and 5 git commands to generate.