aboutsummaryrefslogtreecommitdiffstats
path: root/arch/sparc64/kernel/sbus.c
diff options
context:
space:
mode:
authorDavid S. Miller <davem@sunset.davemloft.net>2007-04-12 00:38:45 -0400
committerDavid S. Miller <davem@sunset.davemloft.net>2007-04-12 02:56:10 -0400
commit2f3a2efd85b698e51e90f06a37d85066725fb4c4 (patch)
tree64306b683060c7739de68319fd393541095e1ec1 /arch/sparc64/kernel/sbus.c
parent80584ff3b99c36ead7e130e453b3a48b18072d18 (diff)
[SPARC64]: Fix SBUS IOMMU allocation code.
There are several IOMMU allocator bugs. Instead of trying to fix this overly complicated code, just mirror the PCI IOMMU arena allocator which is very stable and well stress tested. I tried to make the code as identical as possible so we can switch sun4u PCI and SBUS over to a common piece of IOMMU code. All that will be need are two callbacks, one to do a full IOMMU flush and one to do a streaming buffer flush. This patch gets rid of a lot of hangs and mysterious crashes on SBUS sparc64 systems, at least for me. Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'arch/sparc64/kernel/sbus.c')
-rw-r--r--arch/sparc64/kernel/sbus.c560
1 files changed, 232 insertions, 328 deletions
diff --git a/arch/sparc64/kernel/sbus.c b/arch/sparc64/kernel/sbus.c
index 01d6d869ea2b..14f78fb5e890 100644
--- a/arch/sparc64/kernel/sbus.c
+++ b/arch/sparc64/kernel/sbus.c
@@ -24,48 +24,25 @@
24 24
25#include "iommu_common.h" 25#include "iommu_common.h"
26 26
27/* These should be allocated on an SMP_CACHE_BYTES
28 * aligned boundary for optimal performance.
29 *
30 * On SYSIO, using an 8K page size we have 1GB of SBUS
31 * DMA space mapped. We divide this space into equally
32 * sized clusters. We allocate a DMA mapping from the
33 * cluster that matches the order of the allocation, or
34 * if the order is greater than the number of clusters,
35 * we try to allocate from the last cluster.
36 */
37
38#define NCLUSTERS 8UL
39#define ONE_GIG (1UL * 1024UL * 1024UL * 1024UL)
40#define CLUSTER_SIZE (ONE_GIG / NCLUSTERS)
41#define CLUSTER_MASK (CLUSTER_SIZE - 1)
42#define CLUSTER_NPAGES (CLUSTER_SIZE >> IO_PAGE_SHIFT)
43#define MAP_BASE ((u32)0xc0000000) 27#define MAP_BASE ((u32)0xc0000000)
44 28
29struct sbus_iommu_arena {
30 unsigned long *map;
31 unsigned int hint;
32 unsigned int limit;
33};
34
45struct sbus_iommu { 35struct sbus_iommu {
46/*0x00*/spinlock_t lock; 36 spinlock_t lock;
47 37
48/*0x08*/iopte_t *page_table; 38 struct sbus_iommu_arena arena;
49/*0x10*/unsigned long strbuf_regs;
50/*0x18*/unsigned long iommu_regs;
51/*0x20*/unsigned long sbus_control_reg;
52 39
53/*0x28*/volatile unsigned long strbuf_flushflag; 40 iopte_t *page_table;
41 unsigned long strbuf_regs;
42 unsigned long iommu_regs;
43 unsigned long sbus_control_reg;
54 44
55 /* If NCLUSTERS is ever decresed to 4 or lower, 45 volatile unsigned long strbuf_flushflag;
56 * you must increase the size of the type of
57 * these counters. You have been duly warned. -DaveM
58 */
59/*0x30*/struct {
60 u16 next;
61 u16 flush;
62 } alloc_info[NCLUSTERS];
63
64 /* The lowest used consistent mapping entry. Since
65 * we allocate consistent maps out of cluster 0 this
66 * is relative to the beginning of closter 0.
67 */
68/*0x50*/u32 lowest_consistent_map;
69}; 46};
70 47
71/* Offsets from iommu_regs */ 48/* Offsets from iommu_regs */
@@ -91,19 +68,6 @@ static void __iommu_flushall(struct sbus_iommu *iommu)
91 tag += 8UL; 68 tag += 8UL;
92 } 69 }
93 upa_readq(iommu->sbus_control_reg); 70 upa_readq(iommu->sbus_control_reg);
94
95 for (entry = 0; entry < NCLUSTERS; entry++) {
96 iommu->alloc_info[entry].flush =
97 iommu->alloc_info[entry].next;
98 }
99}
100
101static void iommu_flush(struct sbus_iommu *iommu, u32 base, unsigned long npages)
102{
103 while (npages--)
104 upa_writeq(base + (npages << IO_PAGE_SHIFT),
105 iommu->iommu_regs + IOMMU_FLUSH);
106 upa_readq(iommu->sbus_control_reg);
107} 71}
108 72
109/* Offsets from strbuf_regs */ 73/* Offsets from strbuf_regs */
@@ -156,178 +120,115 @@ static void sbus_strbuf_flush(struct sbus_iommu *iommu, u32 base, unsigned long
156 base, npages); 120 base, npages);
157} 121}
158 122
159static iopte_t *alloc_streaming_cluster(struct sbus_iommu *iommu, unsigned long npages) 123/* Based largely upon the ppc64 iommu allocator. */
124static long sbus_arena_alloc(struct sbus_iommu *iommu, unsigned long npages)
160{ 125{
161 iopte_t *iopte, *limit, *first, *cluster; 126 struct sbus_iommu_arena *arena = &iommu->arena;
162 unsigned long cnum, ent, nent, flush_point, found; 127 unsigned long n, i, start, end, limit;
163 128 int pass;
164 cnum = 0; 129
165 nent = 1; 130 limit = arena->limit;
166 while ((1UL << cnum) < npages) 131 start = arena->hint;
167 cnum++; 132 pass = 0;
168 if(cnum >= NCLUSTERS) { 133
169 nent = 1UL << (cnum - NCLUSTERS); 134again:
170 cnum = NCLUSTERS - 1; 135 n = find_next_zero_bit(arena->map, limit, start);
171 } 136 end = n + npages;
172 iopte = iommu->page_table + (cnum * CLUSTER_NPAGES); 137 if (unlikely(end >= limit)) {
173 138 if (likely(pass < 1)) {
174 if (cnum == 0) 139 limit = start;
175 limit = (iommu->page_table + 140 start = 0;
176 iommu->lowest_consistent_map); 141 __iommu_flushall(iommu);
177 else 142 pass++;
178 limit = (iopte + CLUSTER_NPAGES); 143 goto again;
179
180 iopte += ((ent = iommu->alloc_info[cnum].next) << cnum);
181 flush_point = iommu->alloc_info[cnum].flush;
182
183 first = iopte;
184 cluster = NULL;
185 found = 0;
186 for (;;) {
187 if (iopte_val(*iopte) == 0UL) {
188 found++;
189 if (!cluster)
190 cluster = iopte;
191 } else { 144 } else {
192 /* Used cluster in the way */ 145 /* Scanned the whole thing, give up. */
193 cluster = NULL; 146 return -1;
194 found = 0;
195 } 147 }
148 }
196 149
197 if (found == nent) 150 for (i = n; i < end; i++) {
198 break; 151 if (test_bit(i, arena->map)) {
199 152 start = i + 1;
200 iopte += (1 << cnum); 153 goto again;
201 ent++;
202 if (iopte >= limit) {
203 iopte = (iommu->page_table + (cnum * CLUSTER_NPAGES));
204 ent = 0;
205
206 /* Multiple cluster allocations must not wrap */
207 cluster = NULL;
208 found = 0;
209 } 154 }
210 if (ent == flush_point)
211 __iommu_flushall(iommu);
212 if (iopte == first)
213 goto bad;
214 } 155 }
215 156
216 /* ent/iopte points to the last cluster entry we're going to use, 157 for (i = n; i < end; i++)
217 * so save our place for the next allocation. 158 __set_bit(i, arena->map);
218 */ 159
219 if ((iopte + (1 << cnum)) >= limit) 160 arena->hint = end;
220 ent = 0; 161
221 else 162 return n;
222 ent = ent + 1;
223 iommu->alloc_info[cnum].next = ent;
224 if (ent == flush_point)
225 __iommu_flushall(iommu);
226
227 /* I've got your streaming cluster right here buddy boy... */
228 return cluster;
229
230bad:
231 printk(KERN_EMERG "sbus: alloc_streaming_cluster of npages(%ld) failed!\n",
232 npages);
233 return NULL;
234} 163}
235 164
236static void free_streaming_cluster(struct sbus_iommu *iommu, u32 base, unsigned long npages) 165static void sbus_arena_free(struct sbus_iommu_arena *arena, unsigned long base, unsigned long npages)
237{ 166{
238 unsigned long cnum, ent, nent; 167 unsigned long i;
239 iopte_t *iopte;
240 168
241 cnum = 0; 169 for (i = base; i < (base + npages); i++)
242 nent = 1; 170 __clear_bit(i, arena->map);
243 while ((1UL << cnum) < npages)
244 cnum++;
245 if(cnum >= NCLUSTERS) {
246 nent = 1UL << (cnum - NCLUSTERS);
247 cnum = NCLUSTERS - 1;
248 }
249 ent = (base & CLUSTER_MASK) >> (IO_PAGE_SHIFT + cnum);
250 iopte = iommu->page_table + ((base - MAP_BASE) >> IO_PAGE_SHIFT);
251 do {
252 iopte_val(*iopte) = 0UL;
253 iopte += 1 << cnum;
254 } while(--nent);
255
256 /* If the global flush might not have caught this entry,
257 * adjust the flush point such that we will flush before
258 * ever trying to reuse it.
259 */
260#define between(X,Y,Z) (((Z) - (Y)) >= ((X) - (Y)))
261 if (between(ent, iommu->alloc_info[cnum].next, iommu->alloc_info[cnum].flush))
262 iommu->alloc_info[cnum].flush = ent;
263#undef between
264} 171}
265 172
266/* We allocate consistent mappings from the end of cluster zero. */ 173static void sbus_iommu_table_init(struct sbus_iommu *iommu, unsigned int tsbsize)
267static iopte_t *alloc_consistent_cluster(struct sbus_iommu *iommu, unsigned long npages)
268{ 174{
269 iopte_t *iopte; 175 unsigned long tsbbase, order, sz, num_tsb_entries;
270 176
271 iopte = iommu->page_table + (1 * CLUSTER_NPAGES); 177 num_tsb_entries = tsbsize / sizeof(iopte_t);
272 while (iopte > iommu->page_table) {
273 iopte--;
274 if (!(iopte_val(*iopte) & IOPTE_VALID)) {
275 unsigned long tmp = npages;
276 178
277 while (--tmp) { 179 /* Setup initial software IOMMU state. */
278 iopte--; 180 spin_lock_init(&iommu->lock);
279 if (iopte_val(*iopte) & IOPTE_VALID)
280 break;
281 }
282 if (tmp == 0) {
283 u32 entry = (iopte - iommu->page_table);
284 181
285 if (entry < iommu->lowest_consistent_map) 182 /* Allocate and initialize the free area map. */
286 iommu->lowest_consistent_map = entry; 183 sz = num_tsb_entries / 8;
287 return iopte; 184 sz = (sz + 7UL) & ~7UL;
288 } 185 iommu->arena.map = kzalloc(sz, GFP_KERNEL);
289 } 186 if (!iommu->arena.map) {
187 prom_printf("PCI_IOMMU: Error, kmalloc(arena.map) failed.\n");
188 prom_halt();
189 }
190 iommu->arena.limit = num_tsb_entries;
191
192 /* Now allocate and setup the IOMMU page table itself. */
193 order = get_order(tsbsize);
194 tsbbase = __get_free_pages(GFP_KERNEL, order);
195 if (!tsbbase) {
196 prom_printf("IOMMU: Error, gfp(tsb) failed.\n");
197 prom_halt();
290 } 198 }
291 return NULL; 199 iommu->page_table = (iopte_t *)tsbbase;
200 memset(iommu->page_table, 0, tsbsize);
292} 201}
293 202
294static void free_consistent_cluster(struct sbus_iommu *iommu, u32 base, unsigned long npages) 203static inline iopte_t *alloc_npages(struct sbus_iommu *iommu, unsigned long npages)
295{ 204{
296 iopte_t *iopte = iommu->page_table + ((base - MAP_BASE) >> IO_PAGE_SHIFT); 205 long entry;
297 206
298 if ((iopte - iommu->page_table) == iommu->lowest_consistent_map) { 207 entry = sbus_arena_alloc(iommu, npages);
299 iopte_t *walk = iopte + npages; 208 if (unlikely(entry < 0))
300 iopte_t *limit; 209 return NULL;
301 210
302 limit = iommu->page_table + CLUSTER_NPAGES; 211 return iommu->page_table + entry;
303 while (walk < limit) { 212}
304 if (iopte_val(*walk) != 0UL)
305 break;
306 walk++;
307 }
308 iommu->lowest_consistent_map =
309 (walk - iommu->page_table);
310 }
311 213
312 while (npages--) 214static inline void free_npages(struct sbus_iommu *iommu, dma_addr_t base, unsigned long npages)
313 *iopte++ = __iopte(0UL); 215{
216 sbus_arena_free(&iommu->arena, base >> IO_PAGE_SHIFT, npages);
314} 217}
315 218
316void *sbus_alloc_consistent(struct sbus_dev *sdev, size_t size, dma_addr_t *dvma_addr) 219void *sbus_alloc_consistent(struct sbus_dev *sdev, size_t size, dma_addr_t *dvma_addr)
317{ 220{
318 unsigned long order, first_page, flags;
319 struct sbus_iommu *iommu; 221 struct sbus_iommu *iommu;
320 iopte_t *iopte; 222 iopte_t *iopte;
223 unsigned long flags, order, first_page;
321 void *ret; 224 void *ret;
322 int npages; 225 int npages;
323 226
324 if (size <= 0 || sdev == NULL || dvma_addr == NULL)
325 return NULL;
326
327 size = IO_PAGE_ALIGN(size); 227 size = IO_PAGE_ALIGN(size);
328 order = get_order(size); 228 order = get_order(size);
329 if (order >= 10) 229 if (order >= 10)
330 return NULL; 230 return NULL;
231
331 first_page = __get_free_pages(GFP_KERNEL|__GFP_COMP, order); 232 first_page = __get_free_pages(GFP_KERNEL|__GFP_COMP, order);
332 if (first_page == 0UL) 233 if (first_page == 0UL)
333 return NULL; 234 return NULL;
@@ -336,108 +237,121 @@ void *sbus_alloc_consistent(struct sbus_dev *sdev, size_t size, dma_addr_t *dvma
336 iommu = sdev->bus->iommu; 237 iommu = sdev->bus->iommu;
337 238
338 spin_lock_irqsave(&iommu->lock, flags); 239 spin_lock_irqsave(&iommu->lock, flags);
339 iopte = alloc_consistent_cluster(iommu, size >> IO_PAGE_SHIFT); 240 iopte = alloc_npages(iommu, size >> IO_PAGE_SHIFT);
340 if (iopte == NULL) { 241 spin_unlock_irqrestore(&iommu->lock, flags);
341 spin_unlock_irqrestore(&iommu->lock, flags); 242
243 if (unlikely(iopte == NULL)) {
342 free_pages(first_page, order); 244 free_pages(first_page, order);
343 return NULL; 245 return NULL;
344 } 246 }
345 247
346 /* Ok, we're committed at this point. */ 248 *dvma_addr = (MAP_BASE +
347 *dvma_addr = MAP_BASE + ((iopte - iommu->page_table) << IO_PAGE_SHIFT); 249 ((iopte - iommu->page_table) << IO_PAGE_SHIFT));
348 ret = (void *) first_page; 250 ret = (void *) first_page;
349 npages = size >> IO_PAGE_SHIFT; 251 npages = size >> IO_PAGE_SHIFT;
252 first_page = __pa(first_page);
350 while (npages--) { 253 while (npages--) {
351 *iopte++ = __iopte(IOPTE_VALID | IOPTE_CACHE | IOPTE_WRITE | 254 iopte_val(*iopte) = (IOPTE_VALID | IOPTE_CACHE |
352 (__pa(first_page) & IOPTE_PAGE)); 255 IOPTE_WRITE |
256 (first_page & IOPTE_PAGE));
257 iopte++;
353 first_page += IO_PAGE_SIZE; 258 first_page += IO_PAGE_SIZE;
354 } 259 }
355 iommu_flush(iommu, *dvma_addr, size >> IO_PAGE_SHIFT);
356 spin_unlock_irqrestore(&iommu->lock, flags);
357 260
358 return ret; 261 return ret;
359} 262}
360 263
361void sbus_free_consistent(struct sbus_dev *sdev, size_t size, void *cpu, dma_addr_t dvma) 264void sbus_free_consistent(struct sbus_dev *sdev, size_t size, void *cpu, dma_addr_t dvma)
362{ 265{
363 unsigned long order, npages;
364 struct sbus_iommu *iommu; 266 struct sbus_iommu *iommu;
365 267 iopte_t *iopte;
366 if (size <= 0 || sdev == NULL || cpu == NULL) 268 unsigned long flags, order, npages;
367 return;
368 269
369 npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT; 270 npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT;
370 iommu = sdev->bus->iommu; 271 iommu = sdev->bus->iommu;
272 iopte = iommu->page_table +
273 ((dvma - MAP_BASE) >> IO_PAGE_SHIFT);
274
275 spin_lock_irqsave(&iommu->lock, flags);
276
277 free_npages(iommu, dvma - MAP_BASE, npages);
371 278
372 spin_lock_irq(&iommu->lock); 279 spin_unlock_irqrestore(&iommu->lock, flags);
373 free_consistent_cluster(iommu, dvma, npages);
374 iommu_flush(iommu, dvma, npages);
375 spin_unlock_irq(&iommu->lock);
376 280
377 order = get_order(size); 281 order = get_order(size);
378 if (order < 10) 282 if (order < 10)
379 free_pages((unsigned long)cpu, order); 283 free_pages((unsigned long)cpu, order);
380} 284}
381 285
382dma_addr_t sbus_map_single(struct sbus_dev *sdev, void *ptr, size_t size, int dir) 286dma_addr_t sbus_map_single(struct sbus_dev *sdev, void *ptr, size_t sz, int direction)
383{ 287{
384 struct sbus_iommu *iommu = sdev->bus->iommu; 288 struct sbus_iommu *iommu;
385 unsigned long npages, pbase, flags; 289 iopte_t *base;
386 iopte_t *iopte; 290 unsigned long flags, npages, oaddr;
387 u32 dma_base, offset; 291 unsigned long i, base_paddr;
388 unsigned long iopte_bits; 292 u32 bus_addr, ret;
293 unsigned long iopte_protection;
294
295 iommu = sdev->bus->iommu;
389 296
390 if (dir == SBUS_DMA_NONE) 297 if (unlikely(direction == SBUS_DMA_NONE))
391 BUG(); 298 BUG();
392 299
393 pbase = (unsigned long) ptr; 300 oaddr = (unsigned long)ptr;
394 offset = (u32) (pbase & ~IO_PAGE_MASK); 301 npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK);
395 size = (IO_PAGE_ALIGN(pbase + size) - (pbase & IO_PAGE_MASK)); 302 npages >>= IO_PAGE_SHIFT;
396 pbase = (unsigned long) __pa(pbase & IO_PAGE_MASK);
397 303
398 spin_lock_irqsave(&iommu->lock, flags); 304 spin_lock_irqsave(&iommu->lock, flags);
399 npages = size >> IO_PAGE_SHIFT; 305 base = alloc_npages(iommu, npages);
400 iopte = alloc_streaming_cluster(iommu, npages);
401 if (iopte == NULL)
402 goto bad;
403 dma_base = MAP_BASE + ((iopte - iommu->page_table) << IO_PAGE_SHIFT);
404 npages = size >> IO_PAGE_SHIFT;
405 iopte_bits = IOPTE_VALID | IOPTE_STBUF | IOPTE_CACHE;
406 if (dir != SBUS_DMA_TODEVICE)
407 iopte_bits |= IOPTE_WRITE;
408 while (npages--) {
409 *iopte++ = __iopte(iopte_bits | (pbase & IOPTE_PAGE));
410 pbase += IO_PAGE_SIZE;
411 }
412 npages = size >> IO_PAGE_SHIFT;
413 spin_unlock_irqrestore(&iommu->lock, flags); 306 spin_unlock_irqrestore(&iommu->lock, flags);
414 307
415 return (dma_base | offset); 308 if (unlikely(!base))
309 BUG();
416 310
417bad: 311 bus_addr = (MAP_BASE +
418 spin_unlock_irqrestore(&iommu->lock, flags); 312 ((base - iommu->page_table) << IO_PAGE_SHIFT));
419 BUG(); 313 ret = bus_addr | (oaddr & ~IO_PAGE_MASK);
420 return 0; 314 base_paddr = __pa(oaddr & IO_PAGE_MASK);
315
316 iopte_protection = IOPTE_VALID | IOPTE_STBUF | IOPTE_CACHE;
317 if (direction != SBUS_DMA_TODEVICE)
318 iopte_protection |= IOPTE_WRITE;
319
320 for (i = 0; i < npages; i++, base++, base_paddr += IO_PAGE_SIZE)
321 iopte_val(*base) = iopte_protection | base_paddr;
322
323 return ret;
421} 324}
422 325
423void sbus_unmap_single(struct sbus_dev *sdev, dma_addr_t dma_addr, size_t size, int direction) 326void sbus_unmap_single(struct sbus_dev *sdev, dma_addr_t bus_addr, size_t sz, int direction)
424{ 327{
425 struct sbus_iommu *iommu = sdev->bus->iommu; 328 struct sbus_iommu *iommu = sdev->bus->iommu;
426 u32 dma_base = dma_addr & IO_PAGE_MASK; 329 iopte_t *base;
427 unsigned long flags; 330 unsigned long flags, npages, i;
331
332 if (unlikely(direction == SBUS_DMA_NONE))
333 BUG();
334
335 npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
336 npages >>= IO_PAGE_SHIFT;
337 base = iommu->page_table +
338 ((bus_addr - MAP_BASE) >> IO_PAGE_SHIFT);
428 339
429 size = (IO_PAGE_ALIGN(dma_addr + size) - dma_base); 340 bus_addr &= IO_PAGE_MASK;
430 341
431 spin_lock_irqsave(&iommu->lock, flags); 342 spin_lock_irqsave(&iommu->lock, flags);
432 free_streaming_cluster(iommu, dma_base, size >> IO_PAGE_SHIFT); 343 sbus_strbuf_flush(iommu, bus_addr, npages, direction);
433 sbus_strbuf_flush(iommu, dma_base, size >> IO_PAGE_SHIFT, direction); 344 for (i = 0; i < npages; i++)
345 iopte_val(base[i]) = 0UL;
346 free_npages(iommu, bus_addr - MAP_BASE, npages);
434 spin_unlock_irqrestore(&iommu->lock, flags); 347 spin_unlock_irqrestore(&iommu->lock, flags);
435} 348}
436 349
437#define SG_ENT_PHYS_ADDRESS(SG) \ 350#define SG_ENT_PHYS_ADDRESS(SG) \
438 (__pa(page_address((SG)->page)) + (SG)->offset) 351 (__pa(page_address((SG)->page)) + (SG)->offset)
439 352
440static inline void fill_sg(iopte_t *iopte, struct scatterlist *sg, int nused, int nelems, unsigned long iopte_bits) 353static inline void fill_sg(iopte_t *iopte, struct scatterlist *sg,
354 int nused, int nelems, unsigned long iopte_protection)
441{ 355{
442 struct scatterlist *dma_sg = sg; 356 struct scatterlist *dma_sg = sg;
443 struct scatterlist *sg_end = sg + nelems; 357 struct scatterlist *sg_end = sg + nelems;
@@ -462,7 +376,7 @@ static inline void fill_sg(iopte_t *iopte, struct scatterlist *sg, int nused, in
462 for (;;) { 376 for (;;) {
463 unsigned long tmp; 377 unsigned long tmp;
464 378
465 tmp = (unsigned long) SG_ENT_PHYS_ADDRESS(sg); 379 tmp = SG_ENT_PHYS_ADDRESS(sg);
466 len = sg->length; 380 len = sg->length;
467 if (((tmp ^ pteval) >> IO_PAGE_SHIFT) != 0UL) { 381 if (((tmp ^ pteval) >> IO_PAGE_SHIFT) != 0UL) {
468 pteval = tmp & IO_PAGE_MASK; 382 pteval = tmp & IO_PAGE_MASK;
@@ -478,7 +392,7 @@ static inline void fill_sg(iopte_t *iopte, struct scatterlist *sg, int nused, in
478 sg++; 392 sg++;
479 } 393 }
480 394
481 pteval = ((pteval & IOPTE_PAGE) | iopte_bits); 395 pteval = iopte_protection | (pteval & IOPTE_PAGE);
482 while (len > 0) { 396 while (len > 0) {
483 *iopte++ = __iopte(pteval); 397 *iopte++ = __iopte(pteval);
484 pteval += IO_PAGE_SIZE; 398 pteval += IO_PAGE_SIZE;
@@ -509,103 +423,111 @@ static inline void fill_sg(iopte_t *iopte, struct scatterlist *sg, int nused, in
509 } 423 }
510} 424}
511 425
512int sbus_map_sg(struct sbus_dev *sdev, struct scatterlist *sg, int nents, int dir) 426int sbus_map_sg(struct sbus_dev *sdev, struct scatterlist *sglist, int nelems, int direction)
513{ 427{
514 struct sbus_iommu *iommu = sdev->bus->iommu; 428 struct sbus_iommu *iommu;
515 unsigned long flags, npages; 429 unsigned long flags, npages, iopte_protection;
516 iopte_t *iopte; 430 iopte_t *base;
517 u32 dma_base; 431 u32 dma_base;
518 struct scatterlist *sgtmp; 432 struct scatterlist *sgtmp;
519 int used; 433 int used;
520 unsigned long iopte_bits;
521
522 if (dir == SBUS_DMA_NONE)
523 BUG();
524 434
525 /* Fast path single entry scatterlists. */ 435 /* Fast path single entry scatterlists. */
526 if (nents == 1) { 436 if (nelems == 1) {
527 sg->dma_address = 437 sglist->dma_address =
528 sbus_map_single(sdev, 438 sbus_map_single(sdev,
529 (page_address(sg->page) + sg->offset), 439 (page_address(sglist->page) + sglist->offset),
530 sg->length, dir); 440 sglist->length, direction);
531 sg->dma_length = sg->length; 441 sglist->dma_length = sglist->length;
532 return 1; 442 return 1;
533 } 443 }
534 444
535 npages = prepare_sg(sg, nents); 445 iommu = sdev->bus->iommu;
446
447 if (unlikely(direction == SBUS_DMA_NONE))
448 BUG();
449
450 npages = prepare_sg(sglist, nelems);
536 451
537 spin_lock_irqsave(&iommu->lock, flags); 452 spin_lock_irqsave(&iommu->lock, flags);
538 iopte = alloc_streaming_cluster(iommu, npages); 453 base = alloc_npages(iommu, npages);
539 if (iopte == NULL) 454 spin_unlock_irqrestore(&iommu->lock, flags);
540 goto bad; 455
541 dma_base = MAP_BASE + ((iopte - iommu->page_table) << IO_PAGE_SHIFT); 456 if (unlikely(base == NULL))
457 BUG();
458
459 dma_base = MAP_BASE +
460 ((base - iommu->page_table) << IO_PAGE_SHIFT);
542 461
543 /* Normalize DVMA addresses. */ 462 /* Normalize DVMA addresses. */
544 sgtmp = sg; 463 used = nelems;
545 used = nents;
546 464
465 sgtmp = sglist;
547 while (used && sgtmp->dma_length) { 466 while (used && sgtmp->dma_length) {
548 sgtmp->dma_address += dma_base; 467 sgtmp->dma_address += dma_base;
549 sgtmp++; 468 sgtmp++;
550 used--; 469 used--;
551 } 470 }
552 used = nents - used; 471 used = nelems - used;
553 472
554 iopte_bits = IOPTE_VALID | IOPTE_STBUF | IOPTE_CACHE; 473 iopte_protection = IOPTE_VALID | IOPTE_STBUF | IOPTE_CACHE;
555 if (dir != SBUS_DMA_TODEVICE) 474 if (direction != SBUS_DMA_TODEVICE)
556 iopte_bits |= IOPTE_WRITE; 475 iopte_protection |= IOPTE_WRITE;
476
477 fill_sg(base, sglist, used, nelems, iopte_protection);
557 478
558 fill_sg(iopte, sg, used, nents, iopte_bits);
559#ifdef VERIFY_SG 479#ifdef VERIFY_SG
560 verify_sglist(sg, nents, iopte, npages); 480 verify_sglist(sglist, nelems, base, npages);
561#endif 481#endif
562 spin_unlock_irqrestore(&iommu->lock, flags);
563 482
564 return used; 483 return used;
565
566bad:
567 spin_unlock_irqrestore(&iommu->lock, flags);
568 BUG();
569 return 0;
570} 484}
571 485
572void sbus_unmap_sg(struct sbus_dev *sdev, struct scatterlist *sg, int nents, int direction) 486void sbus_unmap_sg(struct sbus_dev *sdev, struct scatterlist *sglist, int nelems, int direction)
573{ 487{
574 unsigned long size, flags;
575 struct sbus_iommu *iommu; 488 struct sbus_iommu *iommu;
576 u32 dvma_base; 489 iopte_t *base;
577 int i; 490 unsigned long flags, i, npages;
491 u32 bus_addr;
578 492
579 /* Fast path single entry scatterlists. */ 493 if (unlikely(direction == SBUS_DMA_NONE))
580 if (nents == 1) { 494 BUG();
581 sbus_unmap_single(sdev, sg->dma_address, sg->dma_length, direction); 495
582 return; 496 iommu = sdev->bus->iommu;
583 } 497
498 bus_addr = sglist->dma_address & IO_PAGE_MASK;
584 499
585 dvma_base = sg[0].dma_address & IO_PAGE_MASK; 500 for (i = 1; i < nelems; i++)
586 for (i = 0; i < nents; i++) { 501 if (sglist[i].dma_length == 0)
587 if (sg[i].dma_length == 0)
588 break; 502 break;
589 }
590 i--; 503 i--;
591 size = IO_PAGE_ALIGN(sg[i].dma_address + sg[i].dma_length) - dvma_base; 504 npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length) -
505 bus_addr) >> IO_PAGE_SHIFT;
506
507 base = iommu->page_table +
508 ((bus_addr - MAP_BASE) >> IO_PAGE_SHIFT);
592 509
593 iommu = sdev->bus->iommu;
594 spin_lock_irqsave(&iommu->lock, flags); 510 spin_lock_irqsave(&iommu->lock, flags);
595 free_streaming_cluster(iommu, dvma_base, size >> IO_PAGE_SHIFT); 511 sbus_strbuf_flush(iommu, bus_addr, npages, direction);
596 sbus_strbuf_flush(iommu, dvma_base, size >> IO_PAGE_SHIFT, direction); 512 for (i = 0; i < npages; i++)
513 iopte_val(base[i]) = 0UL;
514 free_npages(iommu, bus_addr - MAP_BASE, npages);
597 spin_unlock_irqrestore(&iommu->lock, flags); 515 spin_unlock_irqrestore(&iommu->lock, flags);
598} 516}
599 517
600void sbus_dma_sync_single_for_cpu(struct sbus_dev *sdev, dma_addr_t base, size_t size, int direction) 518void sbus_dma_sync_single_for_cpu(struct sbus_dev *sdev, dma_addr_t bus_addr, size_t sz, int direction)
601{ 519{
602 struct sbus_iommu *iommu = sdev->bus->iommu; 520 struct sbus_iommu *iommu;
603 unsigned long flags; 521 unsigned long flags, npages;
522
523 iommu = sdev->bus->iommu;
604 524
605 size = (IO_PAGE_ALIGN(base + size) - (base & IO_PAGE_MASK)); 525 npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
526 npages >>= IO_PAGE_SHIFT;
527 bus_addr &= IO_PAGE_MASK;
606 528
607 spin_lock_irqsave(&iommu->lock, flags); 529 spin_lock_irqsave(&iommu->lock, flags);
608 sbus_strbuf_flush(iommu, base & IO_PAGE_MASK, size >> IO_PAGE_SHIFT, direction); 530 sbus_strbuf_flush(iommu, bus_addr, npages, direction);
609 spin_unlock_irqrestore(&iommu->lock, flags); 531 spin_unlock_irqrestore(&iommu->lock, flags);
610} 532}
611 533
@@ -613,23 +535,25 @@ void sbus_dma_sync_single_for_device(struct sbus_dev *sdev, dma_addr_t base, siz
613{ 535{
614} 536}
615 537
616void sbus_dma_sync_sg_for_cpu(struct sbus_dev *sdev, struct scatterlist *sg, int nents, int direction) 538void sbus_dma_sync_sg_for_cpu(struct sbus_dev *sdev, struct scatterlist *sglist, int nelems, int direction)
617{ 539{
618 struct sbus_iommu *iommu = sdev->bus->iommu; 540 struct sbus_iommu *iommu;
619 unsigned long flags, size; 541 unsigned long flags, npages, i;
620 u32 base; 542 u32 bus_addr;
621 int i; 543
544 iommu = sdev->bus->iommu;
622 545
623 base = sg[0].dma_address & IO_PAGE_MASK; 546 bus_addr = sglist[0].dma_address & IO_PAGE_MASK;
624 for (i = 0; i < nents; i++) { 547 for (i = 0; i < nelems; i++) {
625 if (sg[i].dma_length == 0) 548 if (!sglist[i].dma_length)
626 break; 549 break;
627 } 550 }
628 i--; 551 i--;
629 size = IO_PAGE_ALIGN(sg[i].dma_address + sg[i].dma_length) - base; 552 npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length)
553 - bus_addr) >> IO_PAGE_SHIFT;
630 554
631 spin_lock_irqsave(&iommu->lock, flags); 555 spin_lock_irqsave(&iommu->lock, flags);
632 sbus_strbuf_flush(iommu, base, size >> IO_PAGE_SHIFT, direction); 556 sbus_strbuf_flush(iommu, bus_addr, npages, direction);
633 spin_unlock_irqrestore(&iommu->lock, flags); 557 spin_unlock_irqrestore(&iommu->lock, flags);
634} 558}
635 559
@@ -1104,7 +1028,7 @@ static void __init sbus_iommu_init(int __node, struct sbus_bus *sbus)
1104 struct linux_prom64_registers *pr; 1028 struct linux_prom64_registers *pr;
1105 struct device_node *dp; 1029 struct device_node *dp;
1106 struct sbus_iommu *iommu; 1030 struct sbus_iommu *iommu;
1107 unsigned long regs, tsb_base; 1031 unsigned long regs;
1108 u64 control; 1032 u64 control;
1109 int i; 1033 int i;
1110 1034
@@ -1132,14 +1056,6 @@ static void __init sbus_iommu_init(int __node, struct sbus_bus *sbus)
1132 1056
1133 memset(iommu, 0, sizeof(*iommu)); 1057 memset(iommu, 0, sizeof(*iommu));
1134 1058
1135 /* We start with no consistent mappings. */
1136 iommu->lowest_consistent_map = CLUSTER_NPAGES;
1137
1138 for (i = 0; i < NCLUSTERS; i++) {
1139 iommu->alloc_info[i].flush = 0;
1140 iommu->alloc_info[i].next = 0;
1141 }
1142
1143 /* Setup spinlock. */ 1059 /* Setup spinlock. */
1144 spin_lock_init(&iommu->lock); 1060 spin_lock_init(&iommu->lock);
1145 1061
@@ -1159,25 +1075,13 @@ static void __init sbus_iommu_init(int __node, struct sbus_bus *sbus)
1159 sbus->portid, regs); 1075 sbus->portid, regs);
1160 1076
1161 /* Setup for TSB_SIZE=7, TBW_SIZE=0, MMU_DE=1, MMU_EN=1 */ 1077 /* Setup for TSB_SIZE=7, TBW_SIZE=0, MMU_DE=1, MMU_EN=1 */
1078 sbus_iommu_table_init(iommu, IO_TSB_SIZE);
1079
1162 control = upa_readq(iommu->iommu_regs + IOMMU_CONTROL); 1080 control = upa_readq(iommu->iommu_regs + IOMMU_CONTROL);
1163 control = ((7UL << 16UL) | 1081 control = ((7UL << 16UL) |
1164 (0UL << 2UL) | 1082 (0UL << 2UL) |
1165 (1UL << 1UL) | 1083 (1UL << 1UL) |
1166 (1UL << 0UL)); 1084 (1UL << 0UL));
1167
1168 /* Using the above configuration we need 1MB iommu page
1169 * table (128K ioptes * 8 bytes per iopte). This is
1170 * page order 7 on UltraSparc.
1171 */
1172 tsb_base = __get_free_pages(GFP_ATOMIC, get_order(IO_TSB_SIZE));
1173 if (tsb_base == 0UL) {
1174 prom_printf("sbus_iommu_init: Fatal error, cannot alloc TSB table.\n");
1175 prom_halt();
1176 }
1177
1178 iommu->page_table = (iopte_t *) tsb_base;
1179 memset(iommu->page_table, 0, IO_TSB_SIZE);
1180
1181 upa_writeq(control, iommu->iommu_regs + IOMMU_CONTROL); 1085 upa_writeq(control, iommu->iommu_regs + IOMMU_CONTROL);
1182 1086
1183 /* Clean out any cruft in the IOMMU using 1087 /* Clean out any cruft in the IOMMU using
@@ -1195,7 +1099,7 @@ static void __init sbus_iommu_init(int __node, struct sbus_bus *sbus)
1195 upa_readq(iommu->sbus_control_reg); 1099 upa_readq(iommu->sbus_control_reg);
1196 1100
1197 /* Give the TSB to SYSIO. */ 1101 /* Give the TSB to SYSIO. */
1198 upa_writeq(__pa(tsb_base), iommu->iommu_regs + IOMMU_TSBBASE); 1102 upa_writeq(__pa(iommu->page_table), iommu->iommu_regs + IOMMU_TSBBASE);
1199 1103
1200 /* Setup streaming buffer, DE=1 SB_EN=1 */ 1104 /* Setup streaming buffer, DE=1 SB_EN=1 */
1201 control = (1UL << 1UL) | (1UL << 0UL); 1105 control = (1UL << 1UL) | (1UL << 0UL);