aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorDavid S. Miller <davem@sunset.davemloft.net>2005-10-14 01:15:24 -0400
committerDavid S. Miller <davem@sunset.davemloft.net>2005-10-14 01:15:24 -0400
commit688cb30bdc3e398d97682a6a58f825821ee838c2 (patch)
tree8fc9ca543fc3966eb961743ba4587298fce49279
parent51e8513615ed8202b22ba9a43b0c7376ea4f6868 (diff)
[SPARC64]: Eliminate PCI IOMMU dma mapping size limit.
The hairy fast allocator in the sparc64 PCI IOMMU code has a hard limit of 256 pages. Certain devices can exceed this when performing very large I/Os. So replace with a more simple allocator, based largely upon the arch/ppc64/kernel/iommu.c code. Signed-off-by: David S. Miller <davem@davemloft.net>
-rw-r--r--arch/sparc64/kernel/pci_iommu.c365
-rw-r--r--include/asm-sparc64/pbm.h28
2 files changed, 145 insertions, 248 deletions
diff --git a/arch/sparc64/kernel/pci_iommu.c b/arch/sparc64/kernel/pci_iommu.c
index cdee7d9fed72..a11910be1013 100644
--- a/arch/sparc64/kernel/pci_iommu.c
+++ b/arch/sparc64/kernel/pci_iommu.c
@@ -49,12 +49,6 @@ static void __iommu_flushall(struct pci_iommu *iommu)
49 49
50 /* Ensure completion of previous PIO writes. */ 50 /* Ensure completion of previous PIO writes. */
51 (void) pci_iommu_read(iommu->write_complete_reg); 51 (void) pci_iommu_read(iommu->write_complete_reg);
52
53 /* Now update everyone's flush point. */
54 for (entry = 0; entry < PBM_NCLUSTERS; entry++) {
55 iommu->alloc_info[entry].flush =
56 iommu->alloc_info[entry].next;
57 }
58} 52}
59 53
60#define IOPTE_CONSISTENT(CTX) \ 54#define IOPTE_CONSISTENT(CTX) \
@@ -80,9 +74,61 @@ static void inline iopte_make_dummy(struct pci_iommu *iommu, iopte_t *iopte)
80 iopte_val(*iopte) = val; 74 iopte_val(*iopte) = val;
81} 75}
82 76
77/* Based largely upon the ppc64 iommu allocator. */
78static long pci_arena_alloc(struct pci_iommu *iommu, unsigned long npages)
79{
80 struct pci_iommu_arena *arena = &iommu->arena;
81 unsigned long n, i, start, end, limit;
82 int pass;
83
84 limit = arena->limit;
85 start = arena->hint;
86 pass = 0;
87
88again:
89 n = find_next_zero_bit(arena->map, limit, start);
90 end = n + npages;
91 if (unlikely(end >= limit)) {
92 if (likely(pass < 1)) {
93 limit = start;
94 start = 0;
95 __iommu_flushall(iommu);
96 pass++;
97 goto again;
98 } else {
99 /* Scanned the whole thing, give up. */
100 return -1;
101 }
102 }
103
104 for (i = n; i < end; i++) {
105 if (test_bit(i, arena->map)) {
106 start = i + 1;
107 goto again;
108 }
109 }
110
111 for (i = n; i < end; i++)
112 __set_bit(i, arena->map);
113
114 arena->hint = end;
115
116 return n;
117}
118
119static void pci_arena_free(struct pci_iommu_arena *arena, unsigned long base, unsigned long npages)
120{
121 unsigned long i;
122
123 for (i = base; i < (base + npages); i++)
124 __clear_bit(i, arena->map);
125}
126
83void pci_iommu_table_init(struct pci_iommu *iommu, int tsbsize, u32 dma_offset, u32 dma_addr_mask) 127void pci_iommu_table_init(struct pci_iommu *iommu, int tsbsize, u32 dma_offset, u32 dma_addr_mask)
84{ 128{
85 unsigned long i, tsbbase, order; 129 unsigned long i, tsbbase, order, sz, num_tsb_entries;
130
131 num_tsb_entries = tsbsize / sizeof(iopte_t);
86 132
87 /* Setup initial software IOMMU state. */ 133 /* Setup initial software IOMMU state. */
88 spin_lock_init(&iommu->lock); 134 spin_lock_init(&iommu->lock);
@@ -90,27 +136,16 @@ void pci_iommu_table_init(struct pci_iommu *iommu, int tsbsize, u32 dma_offset,
90 iommu->page_table_map_base = dma_offset; 136 iommu->page_table_map_base = dma_offset;
91 iommu->dma_addr_mask = dma_addr_mask; 137 iommu->dma_addr_mask = dma_addr_mask;
92 138
93 switch (tsbsize / (8 * 1024)) { 139 /* Allocate and initialize the free area map. */
94 case 64: 140 sz = num_tsb_entries / 8;
95 iommu->page_table_sz_bits = 16; 141 sz = (sz + 7UL) & ~7UL;
96 break; 142 iommu->arena.map = kmalloc(sz, GFP_KERNEL);
97 case 128: 143 if (!iommu->arena.map) {
98 iommu->page_table_sz_bits = 17; 144 prom_printf("PCI_IOMMU: Error, kmalloc(arena.map) failed.\n");
99 break;
100 default:
101 prom_printf("PCI_IOMMU: Illegal TSB size %d\n",
102 tsbsize / (8 * 1024));
103 prom_halt(); 145 prom_halt();
104 break;
105 };
106
107 iommu->lowest_consistent_map =
108 1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS);
109
110 for (i = 0; i < PBM_NCLUSTERS; i++) {
111 iommu->alloc_info[i].flush = 0;
112 iommu->alloc_info[i].next = 0;
113 } 146 }
147 memset(iommu->arena.map, 0, sz);
148 iommu->arena.limit = num_tsb_entries;
114 149
115 /* Allocate and initialize the dummy page which we 150 /* Allocate and initialize the dummy page which we
116 * set inactive IO PTEs to point to. 151 * set inactive IO PTEs to point to.
@@ -132,114 +167,24 @@ void pci_iommu_table_init(struct pci_iommu *iommu, int tsbsize, u32 dma_offset,
132 } 167 }
133 iommu->page_table = (iopte_t *)tsbbase; 168 iommu->page_table = (iopte_t *)tsbbase;
134 169
135 for (i = 0; i < tsbsize / sizeof(iopte_t); i++) 170 for (i = 0; i < num_tsb_entries; i++)
136 iopte_make_dummy(iommu, &iommu->page_table[i]); 171 iopte_make_dummy(iommu, &iommu->page_table[i]);
137} 172}
138 173
139static iopte_t *alloc_streaming_cluster(struct pci_iommu *iommu, unsigned long npages) 174static inline iopte_t *alloc_npages(struct pci_iommu *iommu, unsigned long npages)
140{ 175{
141 iopte_t *iopte, *limit, *first; 176 long entry;
142 unsigned long cnum, ent, flush_point;
143
144 cnum = 0;
145 while ((1UL << cnum) < npages)
146 cnum++;
147 iopte = (iommu->page_table +
148 (cnum << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS)));
149
150 if (cnum == 0)
151 limit = (iommu->page_table +
152 iommu->lowest_consistent_map);
153 else
154 limit = (iopte +
155 (1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS)));
156 177
157 iopte += ((ent = iommu->alloc_info[cnum].next) << cnum); 178 entry = pci_arena_alloc(iommu, npages);
158 flush_point = iommu->alloc_info[cnum].flush; 179 if (unlikely(entry < 0))
159 180 return NULL;
160 first = iopte;
161 for (;;) {
162 if (IOPTE_IS_DUMMY(iommu, iopte)) {
163 if ((iopte + (1 << cnum)) >= limit)
164 ent = 0;
165 else
166 ent = ent + 1;
167 iommu->alloc_info[cnum].next = ent;
168 if (ent == flush_point)
169 __iommu_flushall(iommu);
170 break;
171 }
172 iopte += (1 << cnum);
173 ent++;
174 if (iopte >= limit) {
175 iopte = (iommu->page_table +
176 (cnum <<
177 (iommu->page_table_sz_bits - PBM_LOGCLUSTERS)));
178 ent = 0;
179 }
180 if (ent == flush_point)
181 __iommu_flushall(iommu);
182 if (iopte == first)
183 goto bad;
184 }
185
186 /* I've got your streaming cluster right here buddy boy... */
187 return iopte;
188 181
189bad: 182 return iommu->page_table + entry;
190 printk(KERN_EMERG "pci_iommu: alloc_streaming_cluster of npages(%ld) failed!\n",
191 npages);
192 return NULL;
193} 183}
194 184
195static void free_streaming_cluster(struct pci_iommu *iommu, dma_addr_t base, 185static inline void free_npages(struct pci_iommu *iommu, dma_addr_t base, unsigned long npages)
196 unsigned long npages, unsigned long ctx)
197{ 186{
198 unsigned long cnum, ent; 187 pci_arena_free(&iommu->arena, base >> IO_PAGE_SHIFT, npages);
199
200 cnum = 0;
201 while ((1UL << cnum) < npages)
202 cnum++;
203
204 ent = (base << (32 - IO_PAGE_SHIFT + PBM_LOGCLUSTERS - iommu->page_table_sz_bits))
205 >> (32 + PBM_LOGCLUSTERS + cnum - iommu->page_table_sz_bits);
206
207 /* If the global flush might not have caught this entry,
208 * adjust the flush point such that we will flush before
209 * ever trying to reuse it.
210 */
211#define between(X,Y,Z) (((Z) - (Y)) >= ((X) - (Y)))
212 if (between(ent, iommu->alloc_info[cnum].next, iommu->alloc_info[cnum].flush))
213 iommu->alloc_info[cnum].flush = ent;
214#undef between
215}
216
217/* We allocate consistent mappings from the end of cluster zero. */
218static iopte_t *alloc_consistent_cluster(struct pci_iommu *iommu, unsigned long npages)
219{
220 iopte_t *iopte;
221
222 iopte = iommu->page_table + (1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS));
223 while (iopte > iommu->page_table) {
224 iopte--;
225 if (IOPTE_IS_DUMMY(iommu, iopte)) {
226 unsigned long tmp = npages;
227
228 while (--tmp) {
229 iopte--;
230 if (!IOPTE_IS_DUMMY(iommu, iopte))
231 break;
232 }
233 if (tmp == 0) {
234 u32 entry = (iopte - iommu->page_table);
235
236 if (entry < iommu->lowest_consistent_map)
237 iommu->lowest_consistent_map = entry;
238 return iopte;
239 }
240 }
241 }
242 return NULL;
243} 188}
244 189
245static int iommu_alloc_ctx(struct pci_iommu *iommu) 190static int iommu_alloc_ctx(struct pci_iommu *iommu)
@@ -279,7 +224,7 @@ void *pci_alloc_consistent(struct pci_dev *pdev, size_t size, dma_addr_t *dma_ad
279 struct pcidev_cookie *pcp; 224 struct pcidev_cookie *pcp;
280 struct pci_iommu *iommu; 225 struct pci_iommu *iommu;
281 iopte_t *iopte; 226 iopte_t *iopte;
282 unsigned long flags, order, first_page, ctx; 227 unsigned long flags, order, first_page;
283 void *ret; 228 void *ret;
284 int npages; 229 int npages;
285 230
@@ -297,9 +242,10 @@ void *pci_alloc_consistent(struct pci_dev *pdev, size_t size, dma_addr_t *dma_ad
297 iommu = pcp->pbm->iommu; 242 iommu = pcp->pbm->iommu;
298 243
299 spin_lock_irqsave(&iommu->lock, flags); 244 spin_lock_irqsave(&iommu->lock, flags);
300 iopte = alloc_consistent_cluster(iommu, size >> IO_PAGE_SHIFT); 245 iopte = alloc_npages(iommu, size >> IO_PAGE_SHIFT);
301 if (iopte == NULL) { 246 spin_unlock_irqrestore(&iommu->lock, flags);
302 spin_unlock_irqrestore(&iommu->lock, flags); 247
248 if (unlikely(iopte == NULL)) {
303 free_pages(first_page, order); 249 free_pages(first_page, order);
304 return NULL; 250 return NULL;
305 } 251 }
@@ -308,31 +254,15 @@ void *pci_alloc_consistent(struct pci_dev *pdev, size_t size, dma_addr_t *dma_ad
308 ((iopte - iommu->page_table) << IO_PAGE_SHIFT)); 254 ((iopte - iommu->page_table) << IO_PAGE_SHIFT));
309 ret = (void *) first_page; 255 ret = (void *) first_page;
310 npages = size >> IO_PAGE_SHIFT; 256 npages = size >> IO_PAGE_SHIFT;
311 ctx = 0;
312 if (iommu->iommu_ctxflush)
313 ctx = iommu_alloc_ctx(iommu);
314 first_page = __pa(first_page); 257 first_page = __pa(first_page);
315 while (npages--) { 258 while (npages--) {
316 iopte_val(*iopte) = (IOPTE_CONSISTENT(ctx) | 259 iopte_val(*iopte) = (IOPTE_CONSISTENT(0UL) |
317 IOPTE_WRITE | 260 IOPTE_WRITE |
318 (first_page & IOPTE_PAGE)); 261 (first_page & IOPTE_PAGE));
319 iopte++; 262 iopte++;
320 first_page += IO_PAGE_SIZE; 263 first_page += IO_PAGE_SIZE;
321 } 264 }
322 265
323 {
324 int i;
325 u32 daddr = *dma_addrp;
326
327 npages = size >> IO_PAGE_SHIFT;
328 for (i = 0; i < npages; i++) {
329 pci_iommu_write(iommu->iommu_flush, daddr);
330 daddr += IO_PAGE_SIZE;
331 }
332 }
333
334 spin_unlock_irqrestore(&iommu->lock, flags);
335
336 return ret; 266 return ret;
337} 267}
338 268
@@ -342,7 +272,7 @@ void pci_free_consistent(struct pci_dev *pdev, size_t size, void *cpu, dma_addr_
342 struct pcidev_cookie *pcp; 272 struct pcidev_cookie *pcp;
343 struct pci_iommu *iommu; 273 struct pci_iommu *iommu;
344 iopte_t *iopte; 274 iopte_t *iopte;
345 unsigned long flags, order, npages, i, ctx; 275 unsigned long flags, order, npages;
346 276
347 npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT; 277 npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT;
348 pcp = pdev->sysdata; 278 pcp = pdev->sysdata;
@@ -352,46 +282,7 @@ void pci_free_consistent(struct pci_dev *pdev, size_t size, void *cpu, dma_addr_
352 282
353 spin_lock_irqsave(&iommu->lock, flags); 283 spin_lock_irqsave(&iommu->lock, flags);
354 284
355 if ((iopte - iommu->page_table) == 285 free_npages(iommu, dvma, npages);
356 iommu->lowest_consistent_map) {
357 iopte_t *walk = iopte + npages;
358 iopte_t *limit;
359
360 limit = (iommu->page_table +
361 (1 << (iommu->page_table_sz_bits - PBM_LOGCLUSTERS)));
362 while (walk < limit) {
363 if (!IOPTE_IS_DUMMY(iommu, walk))
364 break;
365 walk++;
366 }
367 iommu->lowest_consistent_map =
368 (walk - iommu->page_table);
369 }
370
371 /* Data for consistent mappings cannot enter the streaming
372 * buffers, so we only need to update the TSB. We flush
373 * the IOMMU here as well to prevent conflicts with the
374 * streaming mapping deferred tlb flush scheme.
375 */
376
377 ctx = 0;
378 if (iommu->iommu_ctxflush)
379 ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL;
380
381 for (i = 0; i < npages; i++, iopte++)
382 iopte_make_dummy(iommu, iopte);
383
384 if (iommu->iommu_ctxflush) {
385 pci_iommu_write(iommu->iommu_ctxflush, ctx);
386 } else {
387 for (i = 0; i < npages; i++) {
388 u32 daddr = dvma + (i << IO_PAGE_SHIFT);
389
390 pci_iommu_write(iommu->iommu_flush, daddr);
391 }
392 }
393
394 iommu_free_ctx(iommu, ctx);
395 286
396 spin_unlock_irqrestore(&iommu->lock, flags); 287 spin_unlock_irqrestore(&iommu->lock, flags);
397 288
@@ -418,25 +309,27 @@ dma_addr_t pci_map_single(struct pci_dev *pdev, void *ptr, size_t sz, int direct
418 iommu = pcp->pbm->iommu; 309 iommu = pcp->pbm->iommu;
419 strbuf = &pcp->pbm->stc; 310 strbuf = &pcp->pbm->stc;
420 311
421 if (direction == PCI_DMA_NONE) 312 if (unlikely(direction == PCI_DMA_NONE))
422 BUG(); 313 goto bad_no_ctx;
423 314
424 oaddr = (unsigned long)ptr; 315 oaddr = (unsigned long)ptr;
425 npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK); 316 npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK);
426 npages >>= IO_PAGE_SHIFT; 317 npages >>= IO_PAGE_SHIFT;
427 318
428 spin_lock_irqsave(&iommu->lock, flags); 319 spin_lock_irqsave(&iommu->lock, flags);
320 base = alloc_npages(iommu, npages);
321 ctx = 0;
322 if (iommu->iommu_ctxflush)
323 ctx = iommu_alloc_ctx(iommu);
324 spin_unlock_irqrestore(&iommu->lock, flags);
429 325
430 base = alloc_streaming_cluster(iommu, npages); 326 if (unlikely(!base))
431 if (base == NULL)
432 goto bad; 327 goto bad;
328
433 bus_addr = (iommu->page_table_map_base + 329 bus_addr = (iommu->page_table_map_base +
434 ((base - iommu->page_table) << IO_PAGE_SHIFT)); 330 ((base - iommu->page_table) << IO_PAGE_SHIFT));
435 ret = bus_addr | (oaddr & ~IO_PAGE_MASK); 331 ret = bus_addr | (oaddr & ~IO_PAGE_MASK);
436 base_paddr = __pa(oaddr & IO_PAGE_MASK); 332 base_paddr = __pa(oaddr & IO_PAGE_MASK);
437 ctx = 0;
438 if (iommu->iommu_ctxflush)
439 ctx = iommu_alloc_ctx(iommu);
440 if (strbuf->strbuf_enabled) 333 if (strbuf->strbuf_enabled)
441 iopte_protection = IOPTE_STREAMING(ctx); 334 iopte_protection = IOPTE_STREAMING(ctx);
442 else 335 else
@@ -447,12 +340,13 @@ dma_addr_t pci_map_single(struct pci_dev *pdev, void *ptr, size_t sz, int direct
447 for (i = 0; i < npages; i++, base++, base_paddr += IO_PAGE_SIZE) 340 for (i = 0; i < npages; i++, base++, base_paddr += IO_PAGE_SIZE)
448 iopte_val(*base) = iopte_protection | base_paddr; 341 iopte_val(*base) = iopte_protection | base_paddr;
449 342
450 spin_unlock_irqrestore(&iommu->lock, flags);
451
452 return ret; 343 return ret;
453 344
454bad: 345bad:
455 spin_unlock_irqrestore(&iommu->lock, flags); 346 iommu_free_ctx(iommu, ctx);
347bad_no_ctx:
348 if (printk_ratelimit())
349 WARN_ON(1);
456 return PCI_DMA_ERROR_CODE; 350 return PCI_DMA_ERROR_CODE;
457} 351}
458 352
@@ -527,10 +421,13 @@ void pci_unmap_single(struct pci_dev *pdev, dma_addr_t bus_addr, size_t sz, int
527 struct pci_iommu *iommu; 421 struct pci_iommu *iommu;
528 struct pci_strbuf *strbuf; 422 struct pci_strbuf *strbuf;
529 iopte_t *base; 423 iopte_t *base;
530 unsigned long flags, npages, ctx; 424 unsigned long flags, npages, ctx, i;
531 425
532 if (direction == PCI_DMA_NONE) 426 if (unlikely(direction == PCI_DMA_NONE)) {
533 BUG(); 427 if (printk_ratelimit())
428 WARN_ON(1);
429 return;
430 }
534 431
535 pcp = pdev->sysdata; 432 pcp = pdev->sysdata;
536 iommu = pcp->pbm->iommu; 433 iommu = pcp->pbm->iommu;
@@ -556,13 +453,14 @@ void pci_unmap_single(struct pci_dev *pdev, dma_addr_t bus_addr, size_t sz, int
556 453
557 /* Step 1: Kick data out of streaming buffers if necessary. */ 454 /* Step 1: Kick data out of streaming buffers if necessary. */
558 if (strbuf->strbuf_enabled) 455 if (strbuf->strbuf_enabled)
559 pci_strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction); 456 pci_strbuf_flush(strbuf, iommu, bus_addr, ctx,
457 npages, direction);
560 458
561 /* Step 2: Clear out first TSB entry. */ 459 /* Step 2: Clear out TSB entries. */
562 iopte_make_dummy(iommu, base); 460 for (i = 0; i < npages; i++)
461 iopte_make_dummy(iommu, base + i);
563 462
564 free_streaming_cluster(iommu, bus_addr - iommu->page_table_map_base, 463 free_npages(iommu, bus_addr - iommu->page_table_map_base, npages);
565 npages, ctx);
566 464
567 iommu_free_ctx(iommu, ctx); 465 iommu_free_ctx(iommu, ctx);
568 466
@@ -667,6 +565,8 @@ int pci_map_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int
667 pci_map_single(pdev, 565 pci_map_single(pdev,
668 (page_address(sglist->page) + sglist->offset), 566 (page_address(sglist->page) + sglist->offset),
669 sglist->length, direction); 567 sglist->length, direction);
568 if (unlikely(sglist->dma_address == PCI_DMA_ERROR_CODE))
569 return 0;
670 sglist->dma_length = sglist->length; 570 sglist->dma_length = sglist->length;
671 return 1; 571 return 1;
672 } 572 }
@@ -675,21 +575,29 @@ int pci_map_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int
675 iommu = pcp->pbm->iommu; 575 iommu = pcp->pbm->iommu;
676 strbuf = &pcp->pbm->stc; 576 strbuf = &pcp->pbm->stc;
677 577
678 if (direction == PCI_DMA_NONE) 578 if (unlikely(direction == PCI_DMA_NONE))
679 BUG(); 579 goto bad_no_ctx;
680 580
681 /* Step 1: Prepare scatter list. */ 581 /* Step 1: Prepare scatter list. */
682 582
683 npages = prepare_sg(sglist, nelems); 583 npages = prepare_sg(sglist, nelems);
684 584
685 /* Step 2: Allocate a cluster. */ 585 /* Step 2: Allocate a cluster and context, if necessary. */
686 586
687 spin_lock_irqsave(&iommu->lock, flags); 587 spin_lock_irqsave(&iommu->lock, flags);
688 588
689 base = alloc_streaming_cluster(iommu, npages); 589 base = alloc_npages(iommu, npages);
590 ctx = 0;
591 if (iommu->iommu_ctxflush)
592 ctx = iommu_alloc_ctx(iommu);
593
594 spin_unlock_irqrestore(&iommu->lock, flags);
595
690 if (base == NULL) 596 if (base == NULL)
691 goto bad; 597 goto bad;
692 dma_base = iommu->page_table_map_base + ((base - iommu->page_table) << IO_PAGE_SHIFT); 598
599 dma_base = iommu->page_table_map_base +
600 ((base - iommu->page_table) << IO_PAGE_SHIFT);
693 601
694 /* Step 3: Normalize DMA addresses. */ 602 /* Step 3: Normalize DMA addresses. */
695 used = nelems; 603 used = nelems;
@@ -702,30 +610,28 @@ int pci_map_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int
702 } 610 }
703 used = nelems - used; 611 used = nelems - used;
704 612
705 /* Step 4: Choose a context if necessary. */ 613 /* Step 4: Create the mappings. */
706 ctx = 0;
707 if (iommu->iommu_ctxflush)
708 ctx = iommu_alloc_ctx(iommu);
709
710 /* Step 5: Create the mappings. */
711 if (strbuf->strbuf_enabled) 614 if (strbuf->strbuf_enabled)
712 iopte_protection = IOPTE_STREAMING(ctx); 615 iopte_protection = IOPTE_STREAMING(ctx);
713 else 616 else
714 iopte_protection = IOPTE_CONSISTENT(ctx); 617 iopte_protection = IOPTE_CONSISTENT(ctx);
715 if (direction != PCI_DMA_TODEVICE) 618 if (direction != PCI_DMA_TODEVICE)
716 iopte_protection |= IOPTE_WRITE; 619 iopte_protection |= IOPTE_WRITE;
717 fill_sg (base, sglist, used, nelems, iopte_protection); 620
621 fill_sg(base, sglist, used, nelems, iopte_protection);
622
718#ifdef VERIFY_SG 623#ifdef VERIFY_SG
719 verify_sglist(sglist, nelems, base, npages); 624 verify_sglist(sglist, nelems, base, npages);
720#endif 625#endif
721 626
722 spin_unlock_irqrestore(&iommu->lock, flags);
723
724 return used; 627 return used;
725 628
726bad: 629bad:
727 spin_unlock_irqrestore(&iommu->lock, flags); 630 iommu_free_ctx(iommu, ctx);
728 return PCI_DMA_ERROR_CODE; 631bad_no_ctx:
632 if (printk_ratelimit())
633 WARN_ON(1);
634 return 0;
729} 635}
730 636
731/* Unmap a set of streaming mode DMA translations. */ 637/* Unmap a set of streaming mode DMA translations. */
@@ -738,8 +644,10 @@ void pci_unmap_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems,
738 unsigned long flags, ctx, i, npages; 644 unsigned long flags, ctx, i, npages;
739 u32 bus_addr; 645 u32 bus_addr;
740 646
741 if (direction == PCI_DMA_NONE) 647 if (unlikely(direction == PCI_DMA_NONE)) {
742 BUG(); 648 if (printk_ratelimit())
649 WARN_ON(1);
650 }
743 651
744 pcp = pdev->sysdata; 652 pcp = pdev->sysdata;
745 iommu = pcp->pbm->iommu; 653 iommu = pcp->pbm->iommu;
@@ -751,7 +659,8 @@ void pci_unmap_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems,
751 if (sglist[i].dma_length == 0) 659 if (sglist[i].dma_length == 0)
752 break; 660 break;
753 i--; 661 i--;
754 npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length) - bus_addr) >> IO_PAGE_SHIFT; 662 npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length) -
663 bus_addr) >> IO_PAGE_SHIFT;
755 664
756 base = iommu->page_table + 665 base = iommu->page_table +
757 ((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT); 666 ((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
@@ -772,11 +681,11 @@ void pci_unmap_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems,
772 if (strbuf->strbuf_enabled) 681 if (strbuf->strbuf_enabled)
773 pci_strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction); 682 pci_strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
774 683
775 /* Step 2: Clear out first TSB entry. */ 684 /* Step 2: Clear out the TSB entries. */
776 iopte_make_dummy(iommu, base); 685 for (i = 0; i < npages; i++)
686 iopte_make_dummy(iommu, base + i);
777 687
778 free_streaming_cluster(iommu, bus_addr - iommu->page_table_map_base, 688 free_npages(iommu, bus_addr - iommu->page_table_map_base, npages);
779 npages, ctx);
780 689
781 iommu_free_ctx(iommu, ctx); 690 iommu_free_ctx(iommu, ctx);
782 691
diff --git a/include/asm-sparc64/pbm.h b/include/asm-sparc64/pbm.h
index c067407de0b0..dd35a2c7798a 100644
--- a/include/asm-sparc64/pbm.h
+++ b/include/asm-sparc64/pbm.h
@@ -27,23 +27,27 @@
27 * PCI bus. 27 * PCI bus.
28 */ 28 */
29 29
30#define PBM_LOGCLUSTERS 3
31#define PBM_NCLUSTERS (1 << PBM_LOGCLUSTERS)
32
33struct pci_controller_info; 30struct pci_controller_info;
34 31
35/* This contains the software state necessary to drive a PCI 32/* This contains the software state necessary to drive a PCI
36 * controller's IOMMU. 33 * controller's IOMMU.
37 */ 34 */
35struct pci_iommu_arena {
36 unsigned long *map;
37 unsigned int hint;
38 unsigned int limit;
39};
40
38struct pci_iommu { 41struct pci_iommu {
39 /* This protects the controller's IOMMU and all 42 /* This protects the controller's IOMMU and all
40 * streaming buffers underneath. 43 * streaming buffers underneath.
41 */ 44 */
42 spinlock_t lock; 45 spinlock_t lock;
43 46
47 struct pci_iommu_arena arena;
48
44 /* IOMMU page table, a linear array of ioptes. */ 49 /* IOMMU page table, a linear array of ioptes. */
45 iopte_t *page_table; /* The page table itself. */ 50 iopte_t *page_table; /* The page table itself. */
46 int page_table_sz_bits; /* log2 of ow many pages does it map? */
47 51
48 /* Base PCI memory space address where IOMMU mappings 52 /* Base PCI memory space address where IOMMU mappings
49 * begin. 53 * begin.
@@ -62,12 +66,6 @@ struct pci_iommu {
62 */ 66 */
63 unsigned long write_complete_reg; 67 unsigned long write_complete_reg;
64 68
65 /* The lowest used consistent mapping entry. Since
66 * we allocate consistent maps out of cluster 0 this
67 * is relative to the beginning of closter 0.
68 */
69 u32 lowest_consistent_map;
70
71 /* In order to deal with some buggy third-party PCI bridges that 69 /* In order to deal with some buggy third-party PCI bridges that
72 * do wrong prefetching, we never mark valid mappings as invalid. 70 * do wrong prefetching, we never mark valid mappings as invalid.
73 * Instead we point them at this dummy page. 71 * Instead we point them at this dummy page.
@@ -75,16 +73,6 @@ struct pci_iommu {
75 unsigned long dummy_page; 73 unsigned long dummy_page;
76 unsigned long dummy_page_pa; 74 unsigned long dummy_page_pa;
77 75
78 /* If PBM_NCLUSTERS is ever decreased to 4 or lower,
79 * or if largest supported page_table_sz * 8K goes above
80 * 2GB, you must increase the size of the type of
81 * these counters. You have been duly warned. -DaveM
82 */
83 struct {
84 u16 next;
85 u16 flush;
86 } alloc_info[PBM_NCLUSTERS];
87
88 /* CTX allocation. */ 76 /* CTX allocation. */
89 unsigned long ctx_lowest_free; 77 unsigned long ctx_lowest_free;
90 unsigned long ctx_bitmap[IOMMU_NUM_CTXS / (sizeof(unsigned long) * 8)]; 78 unsigned long ctx_bitmap[IOMMU_NUM_CTXS / (sizeof(unsigned long) * 8)];