Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* $Id: pci_iommu.c,v 1.17 2001/12/17 07:05:09 davem Exp $ |
2 | * pci_iommu.c: UltraSparc PCI controller IOM/STC support. | |
3 | * | |
4 | * Copyright (C) 1999 David S. Miller (davem@redhat.com) | |
5 | * Copyright (C) 1999, 2000 Jakub Jelinek (jakub@redhat.com) | |
6 | */ | |
7 | ||
8 | #include <linux/kernel.h> | |
9 | #include <linux/sched.h> | |
10 | #include <linux/mm.h> | |
4dbc30fb | 11 | #include <linux/delay.h> |
1da177e4 LT |
12 | |
13 | #include <asm/pbm.h> | |
14 | ||
15 | #include "iommu_common.h" | |
16 | ||
17 | #define PCI_STC_CTXMATCH_ADDR(STC, CTX) \ | |
18 | ((STC)->strbuf_ctxmatch_base + ((CTX) << 3)) | |
19 | ||
20 | /* Accessing IOMMU and Streaming Buffer registers. | |
21 | * REG parameter is a physical address. All registers | |
22 | * are 64-bits in size. | |
23 | */ | |
24 | #define pci_iommu_read(__reg) \ | |
25 | ({ u64 __ret; \ | |
26 | __asm__ __volatile__("ldxa [%1] %2, %0" \ | |
27 | : "=r" (__ret) \ | |
28 | : "r" (__reg), "i" (ASI_PHYS_BYPASS_EC_E) \ | |
29 | : "memory"); \ | |
30 | __ret; \ | |
31 | }) | |
32 | #define pci_iommu_write(__reg, __val) \ | |
33 | __asm__ __volatile__("stxa %0, [%1] %2" \ | |
34 | : /* no outputs */ \ | |
35 | : "r" (__val), "r" (__reg), \ | |
36 | "i" (ASI_PHYS_BYPASS_EC_E)) | |
37 | ||
38 | /* Must be invoked under the IOMMU lock. */ | |
39 | static void __iommu_flushall(struct pci_iommu *iommu) | |
40 | { | |
41 | unsigned long tag; | |
42 | int entry; | |
43 | ||
44 | tag = iommu->iommu_flush + (0xa580UL - 0x0210UL); | |
45 | for (entry = 0; entry < 16; entry++) { | |
46 | pci_iommu_write(tag, 0); | |
47 | tag += 8; | |
48 | } | |
49 | ||
50 | /* Ensure completion of previous PIO writes. */ | |
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 | } | |
59 | ||
60 | #define IOPTE_CONSISTENT(CTX) \ | |
61 | (IOPTE_VALID | IOPTE_CACHE | \ | |
62 | (((CTX) << 47) & IOPTE_CONTEXT)) | |
63 | ||
64 | #define IOPTE_STREAMING(CTX) \ | |
65 | (IOPTE_CONSISTENT(CTX) | IOPTE_STBUF) | |
66 | ||
67 | /* Existing mappings are never marked invalid, instead they | |
68 | * are pointed to a dummy page. | |
69 | */ | |
70 | #define IOPTE_IS_DUMMY(iommu, iopte) \ | |
71 | ((iopte_val(*iopte) & IOPTE_PAGE) == (iommu)->dummy_page_pa) | |
72 | ||
73 | static void inline iopte_make_dummy(struct pci_iommu *iommu, iopte_t *iopte) | |
74 | { | |
75 | unsigned long val = iopte_val(*iopte); | |
76 | ||
77 | val &= ~IOPTE_PAGE; | |
78 | val |= iommu->dummy_page_pa; | |
79 | ||
80 | iopte_val(*iopte) = val; | |
81 | } | |
82 | ||
51e85136 | 83 | void pci_iommu_table_init(struct pci_iommu *iommu, int tsbsize, u32 dma_offset, u32 dma_addr_mask) |
1da177e4 | 84 | { |
51e85136 DM |
85 | unsigned long i, tsbbase, order; |
86 | ||
87 | /* Setup initial software IOMMU state. */ | |
88 | spin_lock_init(&iommu->lock); | |
89 | iommu->ctx_lowest_free = 1; | |
90 | iommu->page_table_map_base = dma_offset; | |
91 | iommu->dma_addr_mask = dma_addr_mask; | |
92 | ||
93 | switch (tsbsize / (8 * 1024)) { | |
94 | case 64: | |
95 | iommu->page_table_sz_bits = 16; | |
96 | break; | |
97 | case 128: | |
98 | iommu->page_table_sz_bits = 17; | |
99 | break; | |
100 | default: | |
101 | prom_printf("PCI_IOMMU: Illegal TSB size %d\n", | |
102 | tsbsize / (8 * 1024)); | |
103 | 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 | } | |
1da177e4 | 114 | |
51e85136 DM |
115 | /* Allocate and initialize the dummy page which we |
116 | * set inactive IO PTEs to point to. | |
117 | */ | |
118 | iommu->dummy_page = __get_free_pages(GFP_KERNEL, 0); | |
119 | if (!iommu->dummy_page) { | |
120 | prom_printf("PCI_IOMMU: Error, gfp(dummy_page) failed.\n"); | |
121 | prom_halt(); | |
122 | } | |
123 | memset((void *)iommu->dummy_page, 0, PAGE_SIZE); | |
124 | iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page); | |
125 | ||
126 | /* Now allocate and setup the IOMMU page table itself. */ | |
127 | order = get_order(tsbsize); | |
128 | tsbbase = __get_free_pages(GFP_KERNEL, order); | |
129 | if (!tsbbase) { | |
130 | prom_printf("PCI_IOMMU: Error, gfp(tsb) failed.\n"); | |
131 | prom_halt(); | |
132 | } | |
133 | iommu->page_table = (iopte_t *)tsbbase; | |
1da177e4 | 134 | |
51e85136 | 135 | for (i = 0; i < tsbsize / sizeof(iopte_t); i++) |
1da177e4 LT |
136 | iopte_make_dummy(iommu, &iommu->page_table[i]); |
137 | } | |
138 | ||
139 | static iopte_t *alloc_streaming_cluster(struct pci_iommu *iommu, unsigned long npages) | |
140 | { | |
141 | iopte_t *iopte, *limit, *first; | |
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 | ||
157 | iopte += ((ent = iommu->alloc_info[cnum].next) << cnum); | |
158 | flush_point = iommu->alloc_info[cnum].flush; | |
159 | ||
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 | ||
189 | bad: | |
190 | printk(KERN_EMERG "pci_iommu: alloc_streaming_cluster of npages(%ld) failed!\n", | |
191 | npages); | |
192 | return NULL; | |
193 | } | |
194 | ||
195 | static void free_streaming_cluster(struct pci_iommu *iommu, dma_addr_t base, | |
196 | unsigned long npages, unsigned long ctx) | |
197 | { | |
198 | unsigned long cnum, ent; | |
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. */ | |
218 | static 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 | } | |
244 | ||
7c963ad1 DM |
245 | static int iommu_alloc_ctx(struct pci_iommu *iommu) |
246 | { | |
247 | int lowest = iommu->ctx_lowest_free; | |
248 | int sz = IOMMU_NUM_CTXS - lowest; | |
249 | int n = find_next_zero_bit(iommu->ctx_bitmap, sz, lowest); | |
250 | ||
251 | if (unlikely(n == sz)) { | |
252 | n = find_next_zero_bit(iommu->ctx_bitmap, lowest, 1); | |
253 | if (unlikely(n == lowest)) { | |
254 | printk(KERN_WARNING "IOMMU: Ran out of contexts.\n"); | |
255 | n = 0; | |
256 | } | |
257 | } | |
258 | if (n) | |
259 | __set_bit(n, iommu->ctx_bitmap); | |
260 | ||
261 | return n; | |
262 | } | |
263 | ||
264 | static inline void iommu_free_ctx(struct pci_iommu *iommu, int ctx) | |
265 | { | |
266 | if (likely(ctx)) { | |
267 | __clear_bit(ctx, iommu->ctx_bitmap); | |
268 | if (ctx < iommu->ctx_lowest_free) | |
269 | iommu->ctx_lowest_free = ctx; | |
270 | } | |
271 | } | |
272 | ||
1da177e4 LT |
273 | /* Allocate and map kernel buffer of size SIZE using consistent mode |
274 | * DMA for PCI device PDEV. Return non-NULL cpu-side address if | |
275 | * successful and set *DMA_ADDRP to the PCI side dma address. | |
276 | */ | |
277 | void *pci_alloc_consistent(struct pci_dev *pdev, size_t size, dma_addr_t *dma_addrp) | |
278 | { | |
279 | struct pcidev_cookie *pcp; | |
280 | struct pci_iommu *iommu; | |
281 | iopte_t *iopte; | |
282 | unsigned long flags, order, first_page, ctx; | |
283 | void *ret; | |
284 | int npages; | |
285 | ||
286 | size = IO_PAGE_ALIGN(size); | |
287 | order = get_order(size); | |
288 | if (order >= 10) | |
289 | return NULL; | |
290 | ||
291 | first_page = __get_free_pages(GFP_ATOMIC, order); | |
292 | if (first_page == 0UL) | |
293 | return NULL; | |
294 | memset((char *)first_page, 0, PAGE_SIZE << order); | |
295 | ||
296 | pcp = pdev->sysdata; | |
297 | iommu = pcp->pbm->iommu; | |
298 | ||
299 | spin_lock_irqsave(&iommu->lock, flags); | |
300 | iopte = alloc_consistent_cluster(iommu, size >> IO_PAGE_SHIFT); | |
301 | if (iopte == NULL) { | |
302 | spin_unlock_irqrestore(&iommu->lock, flags); | |
303 | free_pages(first_page, order); | |
304 | return NULL; | |
305 | } | |
306 | ||
307 | *dma_addrp = (iommu->page_table_map_base + | |
308 | ((iopte - iommu->page_table) << IO_PAGE_SHIFT)); | |
309 | ret = (void *) first_page; | |
310 | npages = size >> IO_PAGE_SHIFT; | |
311 | ctx = 0; | |
312 | if (iommu->iommu_ctxflush) | |
7c963ad1 | 313 | ctx = iommu_alloc_ctx(iommu); |
1da177e4 LT |
314 | first_page = __pa(first_page); |
315 | while (npages--) { | |
316 | iopte_val(*iopte) = (IOPTE_CONSISTENT(ctx) | | |
317 | IOPTE_WRITE | | |
318 | (first_page & IOPTE_PAGE)); | |
319 | iopte++; | |
320 | first_page += IO_PAGE_SIZE; | |
321 | } | |
322 | ||
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; | |
337 | } | |
338 | ||
339 | /* Free and unmap a consistent DMA translation. */ | |
340 | void pci_free_consistent(struct pci_dev *pdev, size_t size, void *cpu, dma_addr_t dvma) | |
341 | { | |
342 | struct pcidev_cookie *pcp; | |
343 | struct pci_iommu *iommu; | |
344 | iopte_t *iopte; | |
345 | unsigned long flags, order, npages, i, ctx; | |
346 | ||
347 | npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT; | |
348 | pcp = pdev->sysdata; | |
349 | iommu = pcp->pbm->iommu; | |
350 | iopte = iommu->page_table + | |
351 | ((dvma - iommu->page_table_map_base) >> IO_PAGE_SHIFT); | |
352 | ||
353 | spin_lock_irqsave(&iommu->lock, flags); | |
354 | ||
355 | if ((iopte - iommu->page_table) == | |
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 | ||
7c963ad1 DM |
394 | iommu_free_ctx(iommu, ctx); |
395 | ||
1da177e4 LT |
396 | spin_unlock_irqrestore(&iommu->lock, flags); |
397 | ||
398 | order = get_order(size); | |
399 | if (order < 10) | |
400 | free_pages((unsigned long)cpu, order); | |
401 | } | |
402 | ||
403 | /* Map a single buffer at PTR of SZ bytes for PCI DMA | |
404 | * in streaming mode. | |
405 | */ | |
406 | dma_addr_t pci_map_single(struct pci_dev *pdev, void *ptr, size_t sz, int direction) | |
407 | { | |
408 | struct pcidev_cookie *pcp; | |
409 | struct pci_iommu *iommu; | |
410 | struct pci_strbuf *strbuf; | |
411 | iopte_t *base; | |
412 | unsigned long flags, npages, oaddr; | |
413 | unsigned long i, base_paddr, ctx; | |
414 | u32 bus_addr, ret; | |
415 | unsigned long iopte_protection; | |
416 | ||
417 | pcp = pdev->sysdata; | |
418 | iommu = pcp->pbm->iommu; | |
419 | strbuf = &pcp->pbm->stc; | |
420 | ||
421 | if (direction == PCI_DMA_NONE) | |
422 | BUG(); | |
423 | ||
424 | oaddr = (unsigned long)ptr; | |
425 | npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK); | |
426 | npages >>= IO_PAGE_SHIFT; | |
427 | ||
428 | spin_lock_irqsave(&iommu->lock, flags); | |
429 | ||
430 | base = alloc_streaming_cluster(iommu, npages); | |
431 | if (base == NULL) | |
432 | goto bad; | |
433 | bus_addr = (iommu->page_table_map_base + | |
434 | ((base - iommu->page_table) << IO_PAGE_SHIFT)); | |
435 | ret = bus_addr | (oaddr & ~IO_PAGE_MASK); | |
436 | base_paddr = __pa(oaddr & IO_PAGE_MASK); | |
437 | ctx = 0; | |
438 | if (iommu->iommu_ctxflush) | |
7c963ad1 | 439 | ctx = iommu_alloc_ctx(iommu); |
1da177e4 LT |
440 | if (strbuf->strbuf_enabled) |
441 | iopte_protection = IOPTE_STREAMING(ctx); | |
442 | else | |
443 | iopte_protection = IOPTE_CONSISTENT(ctx); | |
444 | if (direction != PCI_DMA_TODEVICE) | |
445 | iopte_protection |= IOPTE_WRITE; | |
446 | ||
447 | for (i = 0; i < npages; i++, base++, base_paddr += IO_PAGE_SIZE) | |
448 | iopte_val(*base) = iopte_protection | base_paddr; | |
449 | ||
450 | spin_unlock_irqrestore(&iommu->lock, flags); | |
451 | ||
452 | return ret; | |
453 | ||
454 | bad: | |
455 | spin_unlock_irqrestore(&iommu->lock, flags); | |
456 | return PCI_DMA_ERROR_CODE; | |
457 | } | |
458 | ||
7c963ad1 | 459 | static void pci_strbuf_flush(struct pci_strbuf *strbuf, struct pci_iommu *iommu, u32 vaddr, unsigned long ctx, unsigned long npages, int direction) |
4dbc30fb DM |
460 | { |
461 | int limit; | |
462 | ||
4dbc30fb DM |
463 | if (strbuf->strbuf_ctxflush && |
464 | iommu->iommu_ctxflush) { | |
465 | unsigned long matchreg, flushreg; | |
7c963ad1 | 466 | u64 val; |
4dbc30fb DM |
467 | |
468 | flushreg = strbuf->strbuf_ctxflush; | |
469 | matchreg = PCI_STC_CTXMATCH_ADDR(strbuf, ctx); | |
470 | ||
a228dfd5 | 471 | pci_iommu_write(flushreg, ctx); |
88314ee7 DM |
472 | val = pci_iommu_read(matchreg); |
473 | val &= 0xffff; | |
474 | if (!val) | |
7c963ad1 DM |
475 | goto do_flush_sync; |
476 | ||
7c963ad1 DM |
477 | while (val) { |
478 | if (val & 0x1) | |
479 | pci_iommu_write(flushreg, ctx); | |
480 | val >>= 1; | |
a228dfd5 | 481 | } |
7c963ad1 DM |
482 | val = pci_iommu_read(matchreg); |
483 | if (unlikely(val)) { | |
4dbc30fb | 484 | printk(KERN_WARNING "pci_strbuf_flush: ctx flush " |
7c963ad1 DM |
485 | "timeout matchreg[%lx] ctx[%lx]\n", |
486 | val, ctx); | |
487 | goto do_page_flush; | |
488 | } | |
4dbc30fb DM |
489 | } else { |
490 | unsigned long i; | |
491 | ||
7c963ad1 | 492 | do_page_flush: |
4dbc30fb DM |
493 | for (i = 0; i < npages; i++, vaddr += IO_PAGE_SIZE) |
494 | pci_iommu_write(strbuf->strbuf_pflush, vaddr); | |
495 | } | |
496 | ||
7c963ad1 DM |
497 | do_flush_sync: |
498 | /* If the device could not have possibly put dirty data into | |
499 | * the streaming cache, no flush-flag synchronization needs | |
500 | * to be performed. | |
501 | */ | |
502 | if (direction == PCI_DMA_TODEVICE) | |
503 | return; | |
504 | ||
505 | PCI_STC_FLUSHFLAG_INIT(strbuf); | |
4dbc30fb DM |
506 | pci_iommu_write(strbuf->strbuf_fsync, strbuf->strbuf_flushflag_pa); |
507 | (void) pci_iommu_read(iommu->write_complete_reg); | |
508 | ||
a228dfd5 | 509 | limit = 100000; |
4dbc30fb DM |
510 | while (!PCI_STC_FLUSHFLAG_SET(strbuf)) { |
511 | limit--; | |
512 | if (!limit) | |
513 | break; | |
a228dfd5 | 514 | udelay(1); |
4f07118f | 515 | rmb(); |
4dbc30fb DM |
516 | } |
517 | if (!limit) | |
518 | printk(KERN_WARNING "pci_strbuf_flush: flushflag timeout " | |
519 | "vaddr[%08x] ctx[%lx] npages[%ld]\n", | |
520 | vaddr, ctx, npages); | |
521 | } | |
522 | ||
1da177e4 LT |
523 | /* Unmap a single streaming mode DMA translation. */ |
524 | void pci_unmap_single(struct pci_dev *pdev, dma_addr_t bus_addr, size_t sz, int direction) | |
525 | { | |
526 | struct pcidev_cookie *pcp; | |
527 | struct pci_iommu *iommu; | |
528 | struct pci_strbuf *strbuf; | |
529 | iopte_t *base; | |
4dbc30fb | 530 | unsigned long flags, npages, ctx; |
1da177e4 LT |
531 | |
532 | if (direction == PCI_DMA_NONE) | |
533 | BUG(); | |
534 | ||
535 | pcp = pdev->sysdata; | |
536 | iommu = pcp->pbm->iommu; | |
537 | strbuf = &pcp->pbm->stc; | |
538 | ||
539 | npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK); | |
540 | npages >>= IO_PAGE_SHIFT; | |
541 | base = iommu->page_table + | |
542 | ((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT); | |
543 | #ifdef DEBUG_PCI_IOMMU | |
544 | if (IOPTE_IS_DUMMY(iommu, base)) | |
545 | printk("pci_unmap_single called on non-mapped region %08x,%08x from %016lx\n", | |
546 | bus_addr, sz, __builtin_return_address(0)); | |
547 | #endif | |
548 | bus_addr &= IO_PAGE_MASK; | |
549 | ||
550 | spin_lock_irqsave(&iommu->lock, flags); | |
551 | ||
552 | /* Record the context, if any. */ | |
553 | ctx = 0; | |
554 | if (iommu->iommu_ctxflush) | |
555 | ctx = (iopte_val(*base) & IOPTE_CONTEXT) >> 47UL; | |
556 | ||
557 | /* Step 1: Kick data out of streaming buffers if necessary. */ | |
4dbc30fb | 558 | if (strbuf->strbuf_enabled) |
7c963ad1 | 559 | pci_strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction); |
1da177e4 LT |
560 | |
561 | /* Step 2: Clear out first TSB entry. */ | |
562 | iopte_make_dummy(iommu, base); | |
563 | ||
564 | free_streaming_cluster(iommu, bus_addr - iommu->page_table_map_base, | |
565 | npages, ctx); | |
566 | ||
7c963ad1 DM |
567 | iommu_free_ctx(iommu, ctx); |
568 | ||
1da177e4 LT |
569 | spin_unlock_irqrestore(&iommu->lock, flags); |
570 | } | |
571 | ||
572 | #define SG_ENT_PHYS_ADDRESS(SG) \ | |
573 | (__pa(page_address((SG)->page)) + (SG)->offset) | |
574 | ||
575 | static inline void fill_sg(iopte_t *iopte, struct scatterlist *sg, | |
576 | int nused, int nelems, unsigned long iopte_protection) | |
577 | { | |
578 | struct scatterlist *dma_sg = sg; | |
579 | struct scatterlist *sg_end = sg + nelems; | |
580 | int i; | |
581 | ||
582 | for (i = 0; i < nused; i++) { | |
583 | unsigned long pteval = ~0UL; | |
584 | u32 dma_npages; | |
585 | ||
586 | dma_npages = ((dma_sg->dma_address & (IO_PAGE_SIZE - 1UL)) + | |
587 | dma_sg->dma_length + | |
588 | ((IO_PAGE_SIZE - 1UL))) >> IO_PAGE_SHIFT; | |
589 | do { | |
590 | unsigned long offset; | |
591 | signed int len; | |
592 | ||
593 | /* If we are here, we know we have at least one | |
594 | * more page to map. So walk forward until we | |
595 | * hit a page crossing, and begin creating new | |
596 | * mappings from that spot. | |
597 | */ | |
598 | for (;;) { | |
599 | unsigned long tmp; | |
600 | ||
601 | tmp = SG_ENT_PHYS_ADDRESS(sg); | |
602 | len = sg->length; | |
603 | if (((tmp ^ pteval) >> IO_PAGE_SHIFT) != 0UL) { | |
604 | pteval = tmp & IO_PAGE_MASK; | |
605 | offset = tmp & (IO_PAGE_SIZE - 1UL); | |
606 | break; | |
607 | } | |
608 | if (((tmp ^ (tmp + len - 1UL)) >> IO_PAGE_SHIFT) != 0UL) { | |
609 | pteval = (tmp + IO_PAGE_SIZE) & IO_PAGE_MASK; | |
610 | offset = 0UL; | |
611 | len -= (IO_PAGE_SIZE - (tmp & (IO_PAGE_SIZE - 1UL))); | |
612 | break; | |
613 | } | |
614 | sg++; | |
615 | } | |
616 | ||
617 | pteval = iopte_protection | (pteval & IOPTE_PAGE); | |
618 | while (len > 0) { | |
619 | *iopte++ = __iopte(pteval); | |
620 | pteval += IO_PAGE_SIZE; | |
621 | len -= (IO_PAGE_SIZE - offset); | |
622 | offset = 0; | |
623 | dma_npages--; | |
624 | } | |
625 | ||
626 | pteval = (pteval & IOPTE_PAGE) + len; | |
627 | sg++; | |
628 | ||
629 | /* Skip over any tail mappings we've fully mapped, | |
630 | * adjusting pteval along the way. Stop when we | |
631 | * detect a page crossing event. | |
632 | */ | |
633 | while (sg < sg_end && | |
634 | (pteval << (64 - IO_PAGE_SHIFT)) != 0UL && | |
635 | (pteval == SG_ENT_PHYS_ADDRESS(sg)) && | |
636 | ((pteval ^ | |
637 | (SG_ENT_PHYS_ADDRESS(sg) + sg->length - 1UL)) >> IO_PAGE_SHIFT) == 0UL) { | |
638 | pteval += sg->length; | |
639 | sg++; | |
640 | } | |
641 | if ((pteval << (64 - IO_PAGE_SHIFT)) == 0UL) | |
642 | pteval = ~0UL; | |
643 | } while (dma_npages != 0); | |
644 | dma_sg++; | |
645 | } | |
646 | } | |
647 | ||
648 | /* Map a set of buffers described by SGLIST with NELEMS array | |
649 | * elements in streaming mode for PCI DMA. | |
650 | * When making changes here, inspect the assembly output. I was having | |
651 | * hard time to kepp this routine out of using stack slots for holding variables. | |
652 | */ | |
653 | int pci_map_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int direction) | |
654 | { | |
655 | struct pcidev_cookie *pcp; | |
656 | struct pci_iommu *iommu; | |
657 | struct pci_strbuf *strbuf; | |
658 | unsigned long flags, ctx, npages, iopte_protection; | |
659 | iopte_t *base; | |
660 | u32 dma_base; | |
661 | struct scatterlist *sgtmp; | |
662 | int used; | |
663 | ||
664 | /* Fast path single entry scatterlists. */ | |
665 | if (nelems == 1) { | |
666 | sglist->dma_address = | |
667 | pci_map_single(pdev, | |
668 | (page_address(sglist->page) + sglist->offset), | |
669 | sglist->length, direction); | |
670 | sglist->dma_length = sglist->length; | |
671 | return 1; | |
672 | } | |
673 | ||
674 | pcp = pdev->sysdata; | |
675 | iommu = pcp->pbm->iommu; | |
676 | strbuf = &pcp->pbm->stc; | |
677 | ||
678 | if (direction == PCI_DMA_NONE) | |
679 | BUG(); | |
680 | ||
681 | /* Step 1: Prepare scatter list. */ | |
682 | ||
683 | npages = prepare_sg(sglist, nelems); | |
684 | ||
685 | /* Step 2: Allocate a cluster. */ | |
686 | ||
687 | spin_lock_irqsave(&iommu->lock, flags); | |
688 | ||
689 | base = alloc_streaming_cluster(iommu, npages); | |
690 | if (base == NULL) | |
691 | goto bad; | |
692 | dma_base = iommu->page_table_map_base + ((base - iommu->page_table) << IO_PAGE_SHIFT); | |
693 | ||
694 | /* Step 3: Normalize DMA addresses. */ | |
695 | used = nelems; | |
696 | ||
697 | sgtmp = sglist; | |
698 | while (used && sgtmp->dma_length) { | |
699 | sgtmp->dma_address += dma_base; | |
700 | sgtmp++; | |
701 | used--; | |
702 | } | |
703 | used = nelems - used; | |
704 | ||
705 | /* Step 4: Choose a context if necessary. */ | |
706 | ctx = 0; | |
707 | if (iommu->iommu_ctxflush) | |
7c963ad1 | 708 | ctx = iommu_alloc_ctx(iommu); |
1da177e4 LT |
709 | |
710 | /* Step 5: Create the mappings. */ | |
711 | if (strbuf->strbuf_enabled) | |
712 | iopte_protection = IOPTE_STREAMING(ctx); | |
713 | else | |
714 | iopte_protection = IOPTE_CONSISTENT(ctx); | |
715 | if (direction != PCI_DMA_TODEVICE) | |
716 | iopte_protection |= IOPTE_WRITE; | |
717 | fill_sg (base, sglist, used, nelems, iopte_protection); | |
718 | #ifdef VERIFY_SG | |
719 | verify_sglist(sglist, nelems, base, npages); | |
720 | #endif | |
721 | ||
722 | spin_unlock_irqrestore(&iommu->lock, flags); | |
723 | ||
724 | return used; | |
725 | ||
726 | bad: | |
727 | spin_unlock_irqrestore(&iommu->lock, flags); | |
728 | return PCI_DMA_ERROR_CODE; | |
729 | } | |
730 | ||
731 | /* Unmap a set of streaming mode DMA translations. */ | |
732 | void pci_unmap_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int direction) | |
733 | { | |
734 | struct pcidev_cookie *pcp; | |
735 | struct pci_iommu *iommu; | |
736 | struct pci_strbuf *strbuf; | |
737 | iopte_t *base; | |
738 | unsigned long flags, ctx, i, npages; | |
739 | u32 bus_addr; | |
740 | ||
741 | if (direction == PCI_DMA_NONE) | |
742 | BUG(); | |
743 | ||
744 | pcp = pdev->sysdata; | |
745 | iommu = pcp->pbm->iommu; | |
746 | strbuf = &pcp->pbm->stc; | |
747 | ||
748 | bus_addr = sglist->dma_address & IO_PAGE_MASK; | |
749 | ||
750 | for (i = 1; i < nelems; i++) | |
751 | if (sglist[i].dma_length == 0) | |
752 | break; | |
753 | i--; | |
754 | npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length) - bus_addr) >> IO_PAGE_SHIFT; | |
755 | ||
756 | base = iommu->page_table + | |
757 | ((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT); | |
758 | ||
759 | #ifdef DEBUG_PCI_IOMMU | |
760 | if (IOPTE_IS_DUMMY(iommu, base)) | |
761 | printk("pci_unmap_sg called on non-mapped region %016lx,%d from %016lx\n", sglist->dma_address, nelems, __builtin_return_address(0)); | |
762 | #endif | |
763 | ||
764 | spin_lock_irqsave(&iommu->lock, flags); | |
765 | ||
766 | /* Record the context, if any. */ | |
767 | ctx = 0; | |
768 | if (iommu->iommu_ctxflush) | |
769 | ctx = (iopte_val(*base) & IOPTE_CONTEXT) >> 47UL; | |
770 | ||
771 | /* Step 1: Kick data out of streaming buffers if necessary. */ | |
4dbc30fb | 772 | if (strbuf->strbuf_enabled) |
7c963ad1 | 773 | pci_strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction); |
1da177e4 LT |
774 | |
775 | /* Step 2: Clear out first TSB entry. */ | |
776 | iopte_make_dummy(iommu, base); | |
777 | ||
778 | free_streaming_cluster(iommu, bus_addr - iommu->page_table_map_base, | |
779 | npages, ctx); | |
780 | ||
7c963ad1 DM |
781 | iommu_free_ctx(iommu, ctx); |
782 | ||
1da177e4 LT |
783 | spin_unlock_irqrestore(&iommu->lock, flags); |
784 | } | |
785 | ||
786 | /* Make physical memory consistent for a single | |
787 | * streaming mode DMA translation after a transfer. | |
788 | */ | |
789 | void pci_dma_sync_single_for_cpu(struct pci_dev *pdev, dma_addr_t bus_addr, size_t sz, int direction) | |
790 | { | |
791 | struct pcidev_cookie *pcp; | |
792 | struct pci_iommu *iommu; | |
793 | struct pci_strbuf *strbuf; | |
794 | unsigned long flags, ctx, npages; | |
795 | ||
796 | pcp = pdev->sysdata; | |
797 | iommu = pcp->pbm->iommu; | |
798 | strbuf = &pcp->pbm->stc; | |
799 | ||
800 | if (!strbuf->strbuf_enabled) | |
801 | return; | |
802 | ||
803 | spin_lock_irqsave(&iommu->lock, flags); | |
804 | ||
805 | npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK); | |
806 | npages >>= IO_PAGE_SHIFT; | |
807 | bus_addr &= IO_PAGE_MASK; | |
808 | ||
809 | /* Step 1: Record the context, if any. */ | |
810 | ctx = 0; | |
811 | if (iommu->iommu_ctxflush && | |
812 | strbuf->strbuf_ctxflush) { | |
813 | iopte_t *iopte; | |
814 | ||
815 | iopte = iommu->page_table + | |
816 | ((bus_addr - iommu->page_table_map_base)>>IO_PAGE_SHIFT); | |
817 | ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL; | |
818 | } | |
819 | ||
820 | /* Step 2: Kick data out of streaming buffers. */ | |
7c963ad1 | 821 | pci_strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction); |
1da177e4 LT |
822 | |
823 | spin_unlock_irqrestore(&iommu->lock, flags); | |
824 | } | |
825 | ||
826 | /* Make physical memory consistent for a set of streaming | |
827 | * mode DMA translations after a transfer. | |
828 | */ | |
829 | void pci_dma_sync_sg_for_cpu(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int direction) | |
830 | { | |
831 | struct pcidev_cookie *pcp; | |
832 | struct pci_iommu *iommu; | |
833 | struct pci_strbuf *strbuf; | |
4dbc30fb DM |
834 | unsigned long flags, ctx, npages, i; |
835 | u32 bus_addr; | |
1da177e4 LT |
836 | |
837 | pcp = pdev->sysdata; | |
838 | iommu = pcp->pbm->iommu; | |
839 | strbuf = &pcp->pbm->stc; | |
840 | ||
841 | if (!strbuf->strbuf_enabled) | |
842 | return; | |
843 | ||
844 | spin_lock_irqsave(&iommu->lock, flags); | |
845 | ||
846 | /* Step 1: Record the context, if any. */ | |
847 | ctx = 0; | |
848 | if (iommu->iommu_ctxflush && | |
849 | strbuf->strbuf_ctxflush) { | |
850 | iopte_t *iopte; | |
851 | ||
852 | iopte = iommu->page_table + | |
853 | ((sglist[0].dma_address - iommu->page_table_map_base) >> IO_PAGE_SHIFT); | |
854 | ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL; | |
855 | } | |
856 | ||
857 | /* Step 2: Kick data out of streaming buffers. */ | |
4dbc30fb DM |
858 | bus_addr = sglist[0].dma_address & IO_PAGE_MASK; |
859 | for(i = 1; i < nelems; i++) | |
860 | if (!sglist[i].dma_length) | |
861 | break; | |
862 | i--; | |
863 | npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length) | |
864 | - bus_addr) >> IO_PAGE_SHIFT; | |
7c963ad1 | 865 | pci_strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction); |
1da177e4 LT |
866 | |
867 | spin_unlock_irqrestore(&iommu->lock, flags); | |
868 | } | |
869 | ||
870 | static void ali_sound_dma_hack(struct pci_dev *pdev, int set_bit) | |
871 | { | |
872 | struct pci_dev *ali_isa_bridge; | |
873 | u8 val; | |
874 | ||
875 | /* ALI sound chips generate 31-bits of DMA, a special register | |
876 | * determines what bit 31 is emitted as. | |
877 | */ | |
878 | ali_isa_bridge = pci_get_device(PCI_VENDOR_ID_AL, | |
879 | PCI_DEVICE_ID_AL_M1533, | |
880 | NULL); | |
881 | ||
882 | pci_read_config_byte(ali_isa_bridge, 0x7e, &val); | |
883 | if (set_bit) | |
884 | val |= 0x01; | |
885 | else | |
886 | val &= ~0x01; | |
887 | pci_write_config_byte(ali_isa_bridge, 0x7e, val); | |
888 | pci_dev_put(ali_isa_bridge); | |
889 | } | |
890 | ||
891 | int pci_dma_supported(struct pci_dev *pdev, u64 device_mask) | |
892 | { | |
893 | struct pcidev_cookie *pcp = pdev->sysdata; | |
894 | u64 dma_addr_mask; | |
895 | ||
896 | if (pdev == NULL) { | |
897 | dma_addr_mask = 0xffffffff; | |
898 | } else { | |
899 | struct pci_iommu *iommu = pcp->pbm->iommu; | |
900 | ||
901 | dma_addr_mask = iommu->dma_addr_mask; | |
902 | ||
903 | if (pdev->vendor == PCI_VENDOR_ID_AL && | |
904 | pdev->device == PCI_DEVICE_ID_AL_M5451 && | |
905 | device_mask == 0x7fffffff) { | |
906 | ali_sound_dma_hack(pdev, | |
907 | (dma_addr_mask & 0x80000000) != 0); | |
908 | return 1; | |
909 | } | |
910 | } | |
911 | ||
912 | if (device_mask >= (1UL << 32UL)) | |
913 | return 0; | |
914 | ||
915 | return (device_mask & dma_addr_mask) == dma_addr_mask; | |
916 | } |