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