]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/arch/alpha/kernel/pci_iommu.c | |
3 | */ | |
4 | ||
5 | #include <linux/kernel.h> | |
6 | #include <linux/mm.h> | |
7 | #include <linux/pci.h> | |
5a0e3ad6 | 8 | #include <linux/gfp.h> |
1da177e4 | 9 | #include <linux/bootmem.h> |
00cd1176 | 10 | #include <linux/export.h> |
944bda26 | 11 | #include <linux/scatterlist.h> |
74fd1b68 | 12 | #include <linux/log2.h> |
7c53664d | 13 | #include <linux/dma-mapping.h> |
fd28841d | 14 | #include <linux/iommu-helper.h> |
1da177e4 LT |
15 | |
16 | #include <asm/io.h> | |
17 | #include <asm/hwrpb.h> | |
18 | ||
19 | #include "proto.h" | |
20 | #include "pci_impl.h" | |
21 | ||
22 | ||
23 | #define DEBUG_ALLOC 0 | |
24 | #if DEBUG_ALLOC > 0 | |
25 | # define DBGA(args...) printk(KERN_DEBUG args) | |
26 | #else | |
27 | # define DBGA(args...) | |
28 | #endif | |
29 | #if DEBUG_ALLOC > 1 | |
30 | # define DBGA2(args...) printk(KERN_DEBUG args) | |
31 | #else | |
32 | # define DBGA2(args...) | |
33 | #endif | |
34 | ||
35 | #define DEBUG_NODIRECT 0 | |
1da177e4 LT |
36 | |
37 | #define ISA_DMA_MASK 0x00ffffff | |
38 | ||
39 | static inline unsigned long | |
40 | mk_iommu_pte(unsigned long paddr) | |
41 | { | |
42 | return (paddr >> (PAGE_SHIFT-1)) | 1; | |
43 | } | |
44 | ||
1da177e4 LT |
45 | /* Return the minimum of MAX or the first power of two larger |
46 | than main memory. */ | |
47 | ||
48 | unsigned long | |
49 | size_for_memory(unsigned long max) | |
50 | { | |
51 | unsigned long mem = max_low_pfn << PAGE_SHIFT; | |
52 | if (mem < max) | |
74fd1b68 | 53 | max = roundup_pow_of_two(mem); |
1da177e4 LT |
54 | return max; |
55 | } | |
56 | \f | |
ed5f6561 | 57 | struct pci_iommu_arena * __init |
1da177e4 LT |
58 | iommu_arena_new_node(int nid, struct pci_controller *hose, dma_addr_t base, |
59 | unsigned long window_size, unsigned long align) | |
60 | { | |
61 | unsigned long mem_size; | |
62 | struct pci_iommu_arena *arena; | |
63 | ||
64 | mem_size = window_size / (PAGE_SIZE / sizeof(unsigned long)); | |
65 | ||
66 | /* Note that the TLB lookup logic uses bitwise concatenation, | |
67 | not addition, so the required arena alignment is based on | |
68 | the size of the window. Retain the align parameter so that | |
69 | particular systems can over-align the arena. */ | |
70 | if (align < mem_size) | |
71 | align = mem_size; | |
72 | ||
73 | ||
74 | #ifdef CONFIG_DISCONTIGMEM | |
75 | ||
bbb8d343 HH |
76 | arena = alloc_bootmem_node(NODE_DATA(nid), sizeof(*arena)); |
77 | if (!NODE_DATA(nid) || !arena) { | |
78 | printk("%s: couldn't allocate arena from node %d\n" | |
79 | " falling back to system-wide allocation\n", | |
80 | __func__, nid); | |
81 | arena = alloc_bootmem(sizeof(*arena)); | |
82 | } | |
83 | ||
84 | arena->ptes = __alloc_bootmem_node(NODE_DATA(nid), mem_size, align, 0); | |
85 | if (!NODE_DATA(nid) || !arena->ptes) { | |
86 | printk("%s: couldn't allocate arena ptes from node %d\n" | |
87 | " falling back to system-wide allocation\n", | |
88 | __func__, nid); | |
89 | arena->ptes = __alloc_bootmem(mem_size, align, 0); | |
90 | } | |
1da177e4 LT |
91 | |
92 | #else /* CONFIG_DISCONTIGMEM */ | |
93 | ||
94 | arena = alloc_bootmem(sizeof(*arena)); | |
95 | arena->ptes = __alloc_bootmem(mem_size, align, 0); | |
96 | ||
97 | #endif /* CONFIG_DISCONTIGMEM */ | |
98 | ||
99 | spin_lock_init(&arena->lock); | |
100 | arena->hose = hose; | |
101 | arena->dma_base = base; | |
102 | arena->size = window_size; | |
103 | arena->next_entry = 0; | |
104 | ||
105 | /* Align allocations to a multiple of a page size. Not needed | |
106 | unless there are chip bugs. */ | |
107 | arena->align_entry = 1; | |
108 | ||
109 | return arena; | |
110 | } | |
111 | ||
ed5f6561 | 112 | struct pci_iommu_arena * __init |
1da177e4 LT |
113 | iommu_arena_new(struct pci_controller *hose, dma_addr_t base, |
114 | unsigned long window_size, unsigned long align) | |
115 | { | |
116 | return iommu_arena_new_node(0, hose, base, window_size, align); | |
117 | } | |
118 | ||
119 | /* Must be called with the arena lock held */ | |
120 | static long | |
cf540145 FT |
121 | iommu_arena_find_pages(struct device *dev, struct pci_iommu_arena *arena, |
122 | long n, long mask) | |
1da177e4 LT |
123 | { |
124 | unsigned long *ptes; | |
125 | long i, p, nent; | |
23d7e039 | 126 | int pass = 0; |
cf540145 FT |
127 | unsigned long base; |
128 | unsigned long boundary_size; | |
129 | ||
cf540145 | 130 | base = arena->dma_base >> PAGE_SHIFT; |
ab875cf6 IK |
131 | if (dev) { |
132 | boundary_size = dma_get_seg_boundary(dev) + 1; | |
ab875cf6 IK |
133 | boundary_size >>= PAGE_SHIFT; |
134 | } else { | |
135 | boundary_size = 1UL << (32 - PAGE_SHIFT); | |
136 | } | |
1da177e4 LT |
137 | |
138 | /* Search forward for the first mask-aligned sequence of N free ptes */ | |
139 | ptes = arena->ptes; | |
140 | nent = arena->size >> PAGE_SHIFT; | |
3c5f1def | 141 | p = ALIGN(arena->next_entry, mask + 1); |
1da177e4 | 142 | i = 0; |
23d7e039 FT |
143 | |
144 | again: | |
1da177e4 | 145 | while (i < n && p+i < nent) { |
fd28841d | 146 | if (!i && iommu_is_span_boundary(p, n, base, boundary_size)) { |
cf540145 FT |
147 | p = ALIGN(p + 1, mask + 1); |
148 | goto again; | |
149 | } | |
150 | ||
1da177e4 | 151 | if (ptes[p+i]) |
3c5f1def | 152 | p = ALIGN(p + i + 1, mask + 1), i = 0; |
1da177e4 LT |
153 | else |
154 | i = i + 1; | |
155 | } | |
156 | ||
157 | if (i < n) { | |
23d7e039 FT |
158 | if (pass < 1) { |
159 | /* | |
160 | * Reached the end. Flush the TLB and restart | |
161 | * the search from the beginning. | |
162 | */ | |
163 | alpha_mv.mv_pci_tbi(arena->hose, 0, -1); | |
164 | ||
165 | pass++; | |
166 | p = 0; | |
167 | i = 0; | |
168 | goto again; | |
169 | } else | |
1da177e4 LT |
170 | return -1; |
171 | } | |
172 | ||
173 | /* Success. It's the responsibility of the caller to mark them | |
174 | in use before releasing the lock */ | |
175 | return p; | |
176 | } | |
177 | ||
178 | static long | |
cf540145 FT |
179 | iommu_arena_alloc(struct device *dev, struct pci_iommu_arena *arena, long n, |
180 | unsigned int align) | |
1da177e4 LT |
181 | { |
182 | unsigned long flags; | |
183 | unsigned long *ptes; | |
184 | long i, p, mask; | |
185 | ||
186 | spin_lock_irqsave(&arena->lock, flags); | |
187 | ||
188 | /* Search for N empty ptes */ | |
189 | ptes = arena->ptes; | |
190 | mask = max(align, arena->align_entry) - 1; | |
cf540145 | 191 | p = iommu_arena_find_pages(dev, arena, n, mask); |
1da177e4 LT |
192 | if (p < 0) { |
193 | spin_unlock_irqrestore(&arena->lock, flags); | |
194 | return -1; | |
195 | } | |
196 | ||
197 | /* Success. Mark them all in use, ie not zero and invalid | |
198 | for the iommu tlb that could load them from under us. | |
199 | The chip specific bits will fill this in with something | |
200 | kosher when we return. */ | |
201 | for (i = 0; i < n; ++i) | |
202 | ptes[p+i] = IOMMU_INVALID_PTE; | |
203 | ||
204 | arena->next_entry = p + n; | |
205 | spin_unlock_irqrestore(&arena->lock, flags); | |
206 | ||
207 | return p; | |
208 | } | |
209 | ||
210 | static void | |
211 | iommu_arena_free(struct pci_iommu_arena *arena, long ofs, long n) | |
212 | { | |
213 | unsigned long *p; | |
214 | long i; | |
215 | ||
216 | p = arena->ptes + ofs; | |
217 | for (i = 0; i < n; ++i) | |
218 | p[i] = 0; | |
219 | } | |
c186caca FT |
220 | |
221 | /* | |
222 | * True if the machine supports DAC addressing, and DEV can | |
223 | * make use of it given MASK. | |
224 | */ | |
225 | static int pci_dac_dma_supported(struct pci_dev *dev, u64 mask) | |
226 | { | |
f49d2eb3 | 227 | dma_addr_t dac_offset = alpha_mv.pci_dac_offset; |
c186caca FT |
228 | int ok = 1; |
229 | ||
230 | /* If this is not set, the machine doesn't support DAC at all. */ | |
231 | if (dac_offset == 0) | |
232 | ok = 0; | |
233 | ||
234 | /* The device has to be able to address our DAC bit. */ | |
235 | if ((dac_offset & dev->dma_mask) != dac_offset) | |
236 | ok = 0; | |
237 | ||
238 | /* If both conditions above are met, we are fine. */ | |
a19568eb | 239 | DBGA("pci_dac_dma_supported %s from %pf\n", |
c186caca FT |
240 | ok ? "yes" : "no", __builtin_return_address(0)); |
241 | ||
242 | return ok; | |
243 | } | |
caa51716 | 244 | |
1da177e4 LT |
245 | /* Map a single buffer of the indicated size for PCI DMA in streaming |
246 | mode. The 32-bit PCI bus mastering address to use is returned. | |
247 | Once the device is given the dma address, the device owns this memory | |
248 | until either pci_unmap_single or pci_dma_sync_single is performed. */ | |
249 | ||
250 | static dma_addr_t | |
251 | pci_map_single_1(struct pci_dev *pdev, void *cpu_addr, size_t size, | |
252 | int dac_allowed) | |
253 | { | |
254 | struct pci_controller *hose = pdev ? pdev->sysdata : pci_isa_hose; | |
255 | dma_addr_t max_dma = pdev ? pdev->dma_mask : ISA_DMA_MASK; | |
256 | struct pci_iommu_arena *arena; | |
257 | long npages, dma_ofs, i; | |
258 | unsigned long paddr; | |
259 | dma_addr_t ret; | |
260 | unsigned int align = 0; | |
cf540145 | 261 | struct device *dev = pdev ? &pdev->dev : NULL; |
1da177e4 LT |
262 | |
263 | paddr = __pa(cpu_addr); | |
264 | ||
265 | #if !DEBUG_NODIRECT | |
266 | /* First check to see if we can use the direct map window. */ | |
267 | if (paddr + size + __direct_map_base - 1 <= max_dma | |
268 | && paddr + size <= __direct_map_size) { | |
269 | ret = paddr + __direct_map_base; | |
270 | ||
a19568eb | 271 | DBGA2("pci_map_single: [%p,%zx] -> direct %llx from %pf\n", |
1da177e4 LT |
272 | cpu_addr, size, ret, __builtin_return_address(0)); |
273 | ||
274 | return ret; | |
275 | } | |
276 | #endif | |
277 | ||
278 | /* Next, use DAC if selected earlier. */ | |
279 | if (dac_allowed) { | |
280 | ret = paddr + alpha_mv.pci_dac_offset; | |
281 | ||
a19568eb | 282 | DBGA2("pci_map_single: [%p,%zx] -> DAC %llx from %pf\n", |
1da177e4 LT |
283 | cpu_addr, size, ret, __builtin_return_address(0)); |
284 | ||
285 | return ret; | |
286 | } | |
287 | ||
288 | /* If the machine doesn't define a pci_tbi routine, we have to | |
289 | assume it doesn't support sg mapping, and, since we tried to | |
290 | use direct_map above, it now must be considered an error. */ | |
291 | if (! alpha_mv.mv_pci_tbi) { | |
62173198 | 292 | printk_once(KERN_WARNING "pci_map_single: no HW sg\n"); |
1da177e4 LT |
293 | return 0; |
294 | } | |
295 | ||
296 | arena = hose->sg_pci; | |
297 | if (!arena || arena->dma_base + arena->size - 1 > max_dma) | |
298 | arena = hose->sg_isa; | |
299 | ||
eb117d3e | 300 | npages = iommu_num_pages(paddr, size, PAGE_SIZE); |
1da177e4 LT |
301 | |
302 | /* Force allocation to 64KB boundary for ISA bridges. */ | |
303 | if (pdev && pdev == isa_bridge) | |
304 | align = 8; | |
cf540145 | 305 | dma_ofs = iommu_arena_alloc(dev, arena, npages, align); |
1da177e4 LT |
306 | if (dma_ofs < 0) { |
307 | printk(KERN_WARNING "pci_map_single failed: " | |
308 | "could not allocate dma page tables\n"); | |
309 | return 0; | |
310 | } | |
311 | ||
312 | paddr &= PAGE_MASK; | |
313 | for (i = 0; i < npages; ++i, paddr += PAGE_SIZE) | |
314 | arena->ptes[i + dma_ofs] = mk_iommu_pte(paddr); | |
315 | ||
316 | ret = arena->dma_base + dma_ofs * PAGE_SIZE; | |
317 | ret += (unsigned long)cpu_addr & ~PAGE_MASK; | |
318 | ||
a19568eb | 319 | DBGA2("pci_map_single: [%p,%zx] np %ld -> sg %llx from %pf\n", |
1da177e4 LT |
320 | cpu_addr, size, npages, ret, __builtin_return_address(0)); |
321 | ||
322 | return ret; | |
323 | } | |
324 | ||
c186caca FT |
325 | /* Helper for generic DMA-mapping functions. */ |
326 | static struct pci_dev *alpha_gendev_to_pci(struct device *dev) | |
1da177e4 | 327 | { |
09296c0b | 328 | if (dev && dev_is_pci(dev)) |
c186caca | 329 | return to_pci_dev(dev); |
1da177e4 | 330 | |
c186caca FT |
331 | /* Assume that non-PCI devices asking for DMA are either ISA or EISA, |
332 | BUG() otherwise. */ | |
333 | BUG_ON(!isa_bridge); | |
1da177e4 | 334 | |
c186caca FT |
335 | /* Assume non-busmaster ISA DMA when dma_mask is not set (the ISA |
336 | bridge is bus master then). */ | |
337 | if (!dev || !dev->dma_mask || !*dev->dma_mask) | |
338 | return isa_bridge; | |
339 | ||
340 | /* For EISA bus masters, return isa_bridge (it might have smaller | |
341 | dma_mask due to wiring limitations). */ | |
342 | if (*dev->dma_mask >= isa_bridge->dma_mask) | |
343 | return isa_bridge; | |
344 | ||
345 | /* This assumes ISA bus master with dma_mask 0xffffff. */ | |
346 | return NULL; | |
1da177e4 LT |
347 | } |
348 | ||
c186caca FT |
349 | static dma_addr_t alpha_pci_map_page(struct device *dev, struct page *page, |
350 | unsigned long offset, size_t size, | |
351 | enum dma_data_direction dir, | |
00085f1e | 352 | unsigned long attrs) |
1da177e4 | 353 | { |
c186caca | 354 | struct pci_dev *pdev = alpha_gendev_to_pci(dev); |
1da177e4 LT |
355 | int dac_allowed; |
356 | ||
8166ea07 | 357 | BUG_ON(dir == PCI_DMA_NONE); |
1da177e4 LT |
358 | |
359 | dac_allowed = pdev ? pci_dac_dma_supported(pdev, pdev->dma_mask) : 0; | |
360 | return pci_map_single_1(pdev, (char *)page_address(page) + offset, | |
361 | size, dac_allowed); | |
362 | } | |
363 | ||
364 | /* Unmap a single streaming mode DMA translation. The DMA_ADDR and | |
365 | SIZE must match what was provided for in a previous pci_map_single | |
366 | call. All other usages are undefined. After this call, reads by | |
367 | the cpu to the buffer are guaranteed to see whatever the device | |
368 | wrote there. */ | |
369 | ||
c186caca FT |
370 | static void alpha_pci_unmap_page(struct device *dev, dma_addr_t dma_addr, |
371 | size_t size, enum dma_data_direction dir, | |
00085f1e | 372 | unsigned long attrs) |
1da177e4 LT |
373 | { |
374 | unsigned long flags; | |
c186caca | 375 | struct pci_dev *pdev = alpha_gendev_to_pci(dev); |
1da177e4 LT |
376 | struct pci_controller *hose = pdev ? pdev->sysdata : pci_isa_hose; |
377 | struct pci_iommu_arena *arena; | |
378 | long dma_ofs, npages; | |
379 | ||
8166ea07 | 380 | BUG_ON(dir == PCI_DMA_NONE); |
1da177e4 LT |
381 | |
382 | if (dma_addr >= __direct_map_base | |
383 | && dma_addr < __direct_map_base + __direct_map_size) { | |
384 | /* Nothing to do. */ | |
385 | ||
a19568eb | 386 | DBGA2("pci_unmap_single: direct [%llx,%zx] from %pf\n", |
1da177e4 LT |
387 | dma_addr, size, __builtin_return_address(0)); |
388 | ||
389 | return; | |
390 | } | |
391 | ||
392 | if (dma_addr > 0xffffffff) { | |
a19568eb | 393 | DBGA2("pci64_unmap_single: DAC [%llx,%zx] from %pf\n", |
1da177e4 LT |
394 | dma_addr, size, __builtin_return_address(0)); |
395 | return; | |
396 | } | |
397 | ||
398 | arena = hose->sg_pci; | |
399 | if (!arena || dma_addr < arena->dma_base) | |
400 | arena = hose->sg_isa; | |
401 | ||
402 | dma_ofs = (dma_addr - arena->dma_base) >> PAGE_SHIFT; | |
403 | if (dma_ofs * PAGE_SIZE >= arena->size) { | |
5f0e3da6 RD |
404 | printk(KERN_ERR "Bogus pci_unmap_single: dma_addr %llx " |
405 | " base %llx size %x\n", | |
406 | dma_addr, arena->dma_base, arena->size); | |
1da177e4 LT |
407 | return; |
408 | BUG(); | |
409 | } | |
410 | ||
eb117d3e | 411 | npages = iommu_num_pages(dma_addr, size, PAGE_SIZE); |
1da177e4 LT |
412 | |
413 | spin_lock_irqsave(&arena->lock, flags); | |
414 | ||
415 | iommu_arena_free(arena, dma_ofs, npages); | |
416 | ||
417 | /* If we're freeing ptes above the `next_entry' pointer (they | |
418 | may have snuck back into the TLB since the last wrap flush), | |
419 | we need to flush the TLB before reallocating the latter. */ | |
420 | if (dma_ofs >= arena->next_entry) | |
421 | alpha_mv.mv_pci_tbi(hose, dma_addr, dma_addr + size - 1); | |
422 | ||
423 | spin_unlock_irqrestore(&arena->lock, flags); | |
424 | ||
a19568eb | 425 | DBGA2("pci_unmap_single: sg [%llx,%zx] np %ld from %pf\n", |
1da177e4 LT |
426 | dma_addr, size, npages, __builtin_return_address(0)); |
427 | } | |
1da177e4 LT |
428 | |
429 | /* Allocate and map kernel buffer using consistent mode DMA for PCI | |
430 | device. Returns non-NULL cpu-view pointer to the buffer if | |
431 | successful and sets *DMA_ADDRP to the pci side dma address as well, | |
432 | else DMA_ADDRP is undefined. */ | |
433 | ||
c186caca | 434 | static void *alpha_pci_alloc_coherent(struct device *dev, size_t size, |
4ce9a91f | 435 | dma_addr_t *dma_addrp, gfp_t gfp, |
00085f1e | 436 | unsigned long attrs) |
1da177e4 | 437 | { |
c186caca | 438 | struct pci_dev *pdev = alpha_gendev_to_pci(dev); |
1da177e4 LT |
439 | void *cpu_addr; |
440 | long order = get_order(size); | |
c143d43a IK |
441 | |
442 | gfp &= ~GFP_DMA; | |
1da177e4 LT |
443 | |
444 | try_again: | |
445 | cpu_addr = (void *)__get_free_pages(gfp, order); | |
446 | if (! cpu_addr) { | |
447 | printk(KERN_INFO "pci_alloc_consistent: " | |
a19568eb | 448 | "get_free_pages failed from %pf\n", |
1da177e4 LT |
449 | __builtin_return_address(0)); |
450 | /* ??? Really atomic allocation? Otherwise we could play | |
451 | with vmalloc and sg if we can't find contiguous memory. */ | |
452 | return NULL; | |
453 | } | |
454 | memset(cpu_addr, 0, size); | |
455 | ||
456 | *dma_addrp = pci_map_single_1(pdev, cpu_addr, size, 0); | |
457 | if (*dma_addrp == 0) { | |
458 | free_pages((unsigned long)cpu_addr, order); | |
459 | if (alpha_mv.mv_pci_tbi || (gfp & GFP_DMA)) | |
460 | return NULL; | |
461 | /* The address doesn't fit required mask and we | |
462 | do not have iommu. Try again with GFP_DMA. */ | |
463 | gfp |= GFP_DMA; | |
464 | goto try_again; | |
465 | } | |
c186caca | 466 | |
a19568eb | 467 | DBGA2("pci_alloc_consistent: %zx -> [%p,%llx] from %pf\n", |
1da177e4 LT |
468 | size, cpu_addr, *dma_addrp, __builtin_return_address(0)); |
469 | ||
470 | return cpu_addr; | |
471 | } | |
472 | ||
473 | /* Free and unmap a consistent DMA buffer. CPU_ADDR and DMA_ADDR must | |
474 | be values that were returned from pci_alloc_consistent. SIZE must | |
475 | be the same as what as passed into pci_alloc_consistent. | |
476 | References to the memory and mappings associated with CPU_ADDR or | |
477 | DMA_ADDR past this call are illegal. */ | |
478 | ||
c186caca | 479 | static void alpha_pci_free_coherent(struct device *dev, size_t size, |
4ce9a91f | 480 | void *cpu_addr, dma_addr_t dma_addr, |
00085f1e | 481 | unsigned long attrs) |
1da177e4 | 482 | { |
c186caca | 483 | struct pci_dev *pdev = alpha_gendev_to_pci(dev); |
1da177e4 LT |
484 | pci_unmap_single(pdev, dma_addr, size, PCI_DMA_BIDIRECTIONAL); |
485 | free_pages((unsigned long)cpu_addr, get_order(size)); | |
486 | ||
a19568eb | 487 | DBGA2("pci_free_consistent: [%llx,%zx] from %pf\n", |
1da177e4 LT |
488 | dma_addr, size, __builtin_return_address(0)); |
489 | } | |
1da177e4 LT |
490 | |
491 | /* Classify the elements of the scatterlist. Write dma_address | |
492 | of each element with: | |
493 | 0 : Followers all physically adjacent. | |
494 | 1 : Followers all virtually adjacent. | |
495 | -1 : Not leader, physically adjacent to previous. | |
496 | -2 : Not leader, virtually adjacent to previous. | |
497 | Write dma_length of each leader with the combined lengths of | |
498 | the mergable followers. */ | |
499 | ||
58b053e4 | 500 | #define SG_ENT_VIRT_ADDRESS(SG) (sg_virt((SG))) |
1da177e4 LT |
501 | #define SG_ENT_PHYS_ADDRESS(SG) __pa(SG_ENT_VIRT_ADDRESS(SG)) |
502 | ||
503 | static void | |
7c53664d FT |
504 | sg_classify(struct device *dev, struct scatterlist *sg, struct scatterlist *end, |
505 | int virt_ok) | |
1da177e4 LT |
506 | { |
507 | unsigned long next_paddr; | |
508 | struct scatterlist *leader; | |
509 | long leader_flag, leader_length; | |
7c53664d | 510 | unsigned int max_seg_size; |
1da177e4 LT |
511 | |
512 | leader = sg; | |
513 | leader_flag = 0; | |
514 | leader_length = leader->length; | |
515 | next_paddr = SG_ENT_PHYS_ADDRESS(leader) + leader_length; | |
516 | ||
7c53664d FT |
517 | /* we will not marge sg without device. */ |
518 | max_seg_size = dev ? dma_get_max_seg_size(dev) : 0; | |
1da177e4 LT |
519 | for (++sg; sg < end; ++sg) { |
520 | unsigned long addr, len; | |
521 | addr = SG_ENT_PHYS_ADDRESS(sg); | |
522 | len = sg->length; | |
523 | ||
7c53664d FT |
524 | if (leader_length + len > max_seg_size) |
525 | goto new_segment; | |
526 | ||
1da177e4 LT |
527 | if (next_paddr == addr) { |
528 | sg->dma_address = -1; | |
529 | leader_length += len; | |
530 | } else if (((next_paddr | addr) & ~PAGE_MASK) == 0 && virt_ok) { | |
531 | sg->dma_address = -2; | |
532 | leader_flag = 1; | |
533 | leader_length += len; | |
534 | } else { | |
7c53664d | 535 | new_segment: |
1da177e4 LT |
536 | leader->dma_address = leader_flag; |
537 | leader->dma_length = leader_length; | |
538 | leader = sg; | |
539 | leader_flag = 0; | |
540 | leader_length = len; | |
541 | } | |
542 | ||
543 | next_paddr = addr + len; | |
544 | } | |
545 | ||
546 | leader->dma_address = leader_flag; | |
547 | leader->dma_length = leader_length; | |
548 | } | |
549 | ||
550 | /* Given a scatterlist leader, choose an allocation method and fill | |
551 | in the blanks. */ | |
552 | ||
553 | static int | |
7c53664d | 554 | sg_fill(struct device *dev, struct scatterlist *leader, struct scatterlist *end, |
1da177e4 LT |
555 | struct scatterlist *out, struct pci_iommu_arena *arena, |
556 | dma_addr_t max_dma, int dac_allowed) | |
557 | { | |
558 | unsigned long paddr = SG_ENT_PHYS_ADDRESS(leader); | |
559 | long size = leader->dma_length; | |
560 | struct scatterlist *sg; | |
561 | unsigned long *ptes; | |
562 | long npages, dma_ofs, i; | |
563 | ||
564 | #if !DEBUG_NODIRECT | |
565 | /* If everything is physically contiguous, and the addresses | |
566 | fall into the direct-map window, use it. */ | |
567 | if (leader->dma_address == 0 | |
568 | && paddr + size + __direct_map_base - 1 <= max_dma | |
569 | && paddr + size <= __direct_map_size) { | |
570 | out->dma_address = paddr + __direct_map_base; | |
571 | out->dma_length = size; | |
572 | ||
5f0e3da6 | 573 | DBGA(" sg_fill: [%p,%lx] -> direct %llx\n", |
1da177e4 LT |
574 | __va(paddr), size, out->dma_address); |
575 | ||
576 | return 0; | |
577 | } | |
578 | #endif | |
579 | ||
580 | /* If physically contiguous and DAC is available, use it. */ | |
581 | if (leader->dma_address == 0 && dac_allowed) { | |
582 | out->dma_address = paddr + alpha_mv.pci_dac_offset; | |
583 | out->dma_length = size; | |
584 | ||
5f0e3da6 | 585 | DBGA(" sg_fill: [%p,%lx] -> DAC %llx\n", |
1da177e4 LT |
586 | __va(paddr), size, out->dma_address); |
587 | ||
588 | return 0; | |
589 | } | |
590 | ||
591 | /* Otherwise, we'll use the iommu to make the pages virtually | |
592 | contiguous. */ | |
593 | ||
594 | paddr &= ~PAGE_MASK; | |
eb117d3e | 595 | npages = iommu_num_pages(paddr, size, PAGE_SIZE); |
cf540145 | 596 | dma_ofs = iommu_arena_alloc(dev, arena, npages, 0); |
1da177e4 LT |
597 | if (dma_ofs < 0) { |
598 | /* If we attempted a direct map above but failed, die. */ | |
599 | if (leader->dma_address == 0) | |
600 | return -1; | |
601 | ||
602 | /* Otherwise, break up the remaining virtually contiguous | |
603 | hunks into individual direct maps and retry. */ | |
7c53664d FT |
604 | sg_classify(dev, leader, end, 0); |
605 | return sg_fill(dev, leader, end, out, arena, max_dma, dac_allowed); | |
1da177e4 LT |
606 | } |
607 | ||
608 | out->dma_address = arena->dma_base + dma_ofs*PAGE_SIZE + paddr; | |
609 | out->dma_length = size; | |
610 | ||
5f0e3da6 | 611 | DBGA(" sg_fill: [%p,%lx] -> sg %llx np %ld\n", |
1da177e4 LT |
612 | __va(paddr), size, out->dma_address, npages); |
613 | ||
614 | /* All virtually contiguous. We need to find the length of each | |
615 | physically contiguous subsegment to fill in the ptes. */ | |
616 | ptes = &arena->ptes[dma_ofs]; | |
617 | sg = leader; | |
618 | do { | |
619 | #if DEBUG_ALLOC > 0 | |
620 | struct scatterlist *last_sg = sg; | |
621 | #endif | |
622 | ||
623 | size = sg->length; | |
624 | paddr = SG_ENT_PHYS_ADDRESS(sg); | |
625 | ||
626 | while (sg+1 < end && (int) sg[1].dma_address == -1) { | |
627 | size += sg[1].length; | |
628 | sg++; | |
629 | } | |
630 | ||
eb117d3e | 631 | npages = iommu_num_pages(paddr, size, PAGE_SIZE); |
1da177e4 LT |
632 | |
633 | paddr &= PAGE_MASK; | |
634 | for (i = 0; i < npages; ++i, paddr += PAGE_SIZE) | |
635 | *ptes++ = mk_iommu_pte(paddr); | |
636 | ||
637 | #if DEBUG_ALLOC > 0 | |
638 | DBGA(" (%ld) [%p,%x] np %ld\n", | |
639 | last_sg - leader, SG_ENT_VIRT_ADDRESS(last_sg), | |
640 | last_sg->length, npages); | |
641 | while (++last_sg <= sg) { | |
642 | DBGA(" (%ld) [%p,%x] cont\n", | |
643 | last_sg - leader, SG_ENT_VIRT_ADDRESS(last_sg), | |
644 | last_sg->length); | |
645 | } | |
646 | #endif | |
647 | } while (++sg < end && (int) sg->dma_address < 0); | |
648 | ||
649 | return 1; | |
650 | } | |
651 | ||
c186caca FT |
652 | static int alpha_pci_map_sg(struct device *dev, struct scatterlist *sg, |
653 | int nents, enum dma_data_direction dir, | |
00085f1e | 654 | unsigned long attrs) |
1da177e4 | 655 | { |
c186caca | 656 | struct pci_dev *pdev = alpha_gendev_to_pci(dev); |
1da177e4 LT |
657 | struct scatterlist *start, *end, *out; |
658 | struct pci_controller *hose; | |
659 | struct pci_iommu_arena *arena; | |
660 | dma_addr_t max_dma; | |
661 | int dac_allowed; | |
662 | ||
8166ea07 | 663 | BUG_ON(dir == PCI_DMA_NONE); |
1da177e4 | 664 | |
c186caca | 665 | dac_allowed = dev ? pci_dac_dma_supported(pdev, pdev->dma_mask) : 0; |
7c53664d | 666 | |
1da177e4 LT |
667 | /* Fast path single entry scatterlists. */ |
668 | if (nents == 1) { | |
669 | sg->dma_length = sg->length; | |
670 | sg->dma_address | |
671 | = pci_map_single_1(pdev, SG_ENT_VIRT_ADDRESS(sg), | |
672 | sg->length, dac_allowed); | |
673 | return sg->dma_address != 0; | |
674 | } | |
675 | ||
676 | start = sg; | |
677 | end = sg + nents; | |
678 | ||
679 | /* First, prepare information about the entries. */ | |
7c53664d | 680 | sg_classify(dev, sg, end, alpha_mv.mv_pci_tbi != 0); |
1da177e4 LT |
681 | |
682 | /* Second, figure out where we're going to map things. */ | |
683 | if (alpha_mv.mv_pci_tbi) { | |
684 | hose = pdev ? pdev->sysdata : pci_isa_hose; | |
685 | max_dma = pdev ? pdev->dma_mask : ISA_DMA_MASK; | |
686 | arena = hose->sg_pci; | |
687 | if (!arena || arena->dma_base + arena->size - 1 > max_dma) | |
688 | arena = hose->sg_isa; | |
689 | } else { | |
690 | max_dma = -1; | |
691 | arena = NULL; | |
692 | hose = NULL; | |
693 | } | |
694 | ||
695 | /* Third, iterate over the scatterlist leaders and allocate | |
696 | dma space as needed. */ | |
697 | for (out = sg; sg < end; ++sg) { | |
698 | if ((int) sg->dma_address < 0) | |
699 | continue; | |
7c53664d | 700 | if (sg_fill(dev, sg, end, out, arena, max_dma, dac_allowed) < 0) |
1da177e4 LT |
701 | goto error; |
702 | out++; | |
703 | } | |
704 | ||
705 | /* Mark the end of the list for pci_unmap_sg. */ | |
706 | if (out < end) | |
707 | out->dma_length = 0; | |
708 | ||
709 | if (out - start == 0) | |
710 | printk(KERN_WARNING "pci_map_sg failed: no entries?\n"); | |
711 | DBGA("pci_map_sg: %ld entries\n", out - start); | |
712 | ||
713 | return out - start; | |
714 | ||
715 | error: | |
716 | printk(KERN_WARNING "pci_map_sg failed: " | |
717 | "could not allocate dma page tables\n"); | |
718 | ||
719 | /* Some allocation failed while mapping the scatterlist | |
720 | entries. Unmap them now. */ | |
721 | if (out > start) | |
c186caca | 722 | pci_unmap_sg(pdev, start, out - start, dir); |
1da177e4 LT |
723 | return 0; |
724 | } | |
725 | ||
726 | /* Unmap a set of streaming mode DMA translations. Again, cpu read | |
727 | rules concerning calls here are the same as for pci_unmap_single() | |
728 | above. */ | |
729 | ||
c186caca FT |
730 | static void alpha_pci_unmap_sg(struct device *dev, struct scatterlist *sg, |
731 | int nents, enum dma_data_direction dir, | |
00085f1e | 732 | unsigned long attrs) |
1da177e4 | 733 | { |
c186caca | 734 | struct pci_dev *pdev = alpha_gendev_to_pci(dev); |
1da177e4 LT |
735 | unsigned long flags; |
736 | struct pci_controller *hose; | |
737 | struct pci_iommu_arena *arena; | |
738 | struct scatterlist *end; | |
739 | dma_addr_t max_dma; | |
740 | dma_addr_t fbeg, fend; | |
741 | ||
8166ea07 | 742 | BUG_ON(dir == PCI_DMA_NONE); |
1da177e4 LT |
743 | |
744 | if (! alpha_mv.mv_pci_tbi) | |
745 | return; | |
746 | ||
747 | hose = pdev ? pdev->sysdata : pci_isa_hose; | |
748 | max_dma = pdev ? pdev->dma_mask : ISA_DMA_MASK; | |
749 | arena = hose->sg_pci; | |
750 | if (!arena || arena->dma_base + arena->size - 1 > max_dma) | |
751 | arena = hose->sg_isa; | |
752 | ||
753 | fbeg = -1, fend = 0; | |
754 | ||
755 | spin_lock_irqsave(&arena->lock, flags); | |
756 | ||
757 | for (end = sg + nents; sg < end; ++sg) { | |
f49d2eb3 | 758 | dma_addr_t addr; |
1da177e4 LT |
759 | size_t size; |
760 | long npages, ofs; | |
761 | dma_addr_t tend; | |
762 | ||
763 | addr = sg->dma_address; | |
764 | size = sg->dma_length; | |
765 | if (!size) | |
766 | break; | |
767 | ||
768 | if (addr > 0xffffffff) { | |
769 | /* It's a DAC address -- nothing to do. */ | |
5f0e3da6 | 770 | DBGA(" (%ld) DAC [%llx,%zx]\n", |
1da177e4 LT |
771 | sg - end + nents, addr, size); |
772 | continue; | |
773 | } | |
774 | ||
775 | if (addr >= __direct_map_base | |
776 | && addr < __direct_map_base + __direct_map_size) { | |
777 | /* Nothing to do. */ | |
5f0e3da6 | 778 | DBGA(" (%ld) direct [%llx,%zx]\n", |
1da177e4 LT |
779 | sg - end + nents, addr, size); |
780 | continue; | |
781 | } | |
782 | ||
5f0e3da6 | 783 | DBGA(" (%ld) sg [%llx,%zx]\n", |
1da177e4 LT |
784 | sg - end + nents, addr, size); |
785 | ||
eb117d3e | 786 | npages = iommu_num_pages(addr, size, PAGE_SIZE); |
1da177e4 LT |
787 | ofs = (addr - arena->dma_base) >> PAGE_SHIFT; |
788 | iommu_arena_free(arena, ofs, npages); | |
789 | ||
790 | tend = addr + size - 1; | |
791 | if (fbeg > addr) fbeg = addr; | |
792 | if (fend < tend) fend = tend; | |
793 | } | |
794 | ||
795 | /* If we're freeing ptes above the `next_entry' pointer (they | |
796 | may have snuck back into the TLB since the last wrap flush), | |
797 | we need to flush the TLB before reallocating the latter. */ | |
798 | if ((fend - arena->dma_base) >> PAGE_SHIFT >= arena->next_entry) | |
799 | alpha_mv.mv_pci_tbi(hose, fbeg, fend); | |
800 | ||
801 | spin_unlock_irqrestore(&arena->lock, flags); | |
802 | ||
803 | DBGA("pci_unmap_sg: %ld entries\n", nents - (end - sg)); | |
804 | } | |
1da177e4 LT |
805 | |
806 | /* Return whether the given PCI device DMA address mask can be | |
807 | supported properly. */ | |
808 | ||
c186caca | 809 | static int alpha_pci_supported(struct device *dev, u64 mask) |
1da177e4 | 810 | { |
c186caca | 811 | struct pci_dev *pdev = alpha_gendev_to_pci(dev); |
1da177e4 LT |
812 | struct pci_controller *hose; |
813 | struct pci_iommu_arena *arena; | |
814 | ||
815 | /* If there exists a direct map, and the mask fits either | |
816 | the entire direct mapped space or the total system memory as | |
817 | shifted by the map base */ | |
818 | if (__direct_map_size != 0 | |
819 | && (__direct_map_base + __direct_map_size - 1 <= mask || | |
820 | __direct_map_base + (max_low_pfn << PAGE_SHIFT) - 1 <= mask)) | |
821 | return 1; | |
822 | ||
823 | /* Check that we have a scatter-gather arena that fits. */ | |
824 | hose = pdev ? pdev->sysdata : pci_isa_hose; | |
825 | arena = hose->sg_isa; | |
826 | if (arena && arena->dma_base + arena->size - 1 <= mask) | |
827 | return 1; | |
828 | arena = hose->sg_pci; | |
829 | if (arena && arena->dma_base + arena->size - 1 <= mask) | |
830 | return 1; | |
831 | ||
832 | /* As last resort try ZONE_DMA. */ | |
833 | if (!__direct_map_base && MAX_DMA_ADDRESS - IDENT_ADDR - 1 <= mask) | |
834 | return 1; | |
835 | ||
836 | return 0; | |
837 | } | |
838 | ||
839 | \f | |
840 | /* | |
841 | * AGP GART extensions to the IOMMU | |
842 | */ | |
843 | int | |
844 | iommu_reserve(struct pci_iommu_arena *arena, long pg_count, long align_mask) | |
845 | { | |
846 | unsigned long flags; | |
847 | unsigned long *ptes; | |
848 | long i, p; | |
849 | ||
850 | if (!arena) return -EINVAL; | |
851 | ||
852 | spin_lock_irqsave(&arena->lock, flags); | |
853 | ||
854 | /* Search for N empty ptes. */ | |
855 | ptes = arena->ptes; | |
cf540145 | 856 | p = iommu_arena_find_pages(NULL, arena, pg_count, align_mask); |
1da177e4 LT |
857 | if (p < 0) { |
858 | spin_unlock_irqrestore(&arena->lock, flags); | |
859 | return -1; | |
860 | } | |
861 | ||
862 | /* Success. Mark them all reserved (ie not zero and invalid) | |
863 | for the iommu tlb that could load them from under us. | |
864 | They will be filled in with valid bits by _bind() */ | |
865 | for (i = 0; i < pg_count; ++i) | |
866 | ptes[p+i] = IOMMU_RESERVED_PTE; | |
867 | ||
868 | arena->next_entry = p + pg_count; | |
869 | spin_unlock_irqrestore(&arena->lock, flags); | |
870 | ||
871 | return p; | |
872 | } | |
873 | ||
874 | int | |
875 | iommu_release(struct pci_iommu_arena *arena, long pg_start, long pg_count) | |
876 | { | |
877 | unsigned long *ptes; | |
878 | long i; | |
879 | ||
880 | if (!arena) return -EINVAL; | |
881 | ||
882 | ptes = arena->ptes; | |
883 | ||
884 | /* Make sure they're all reserved first... */ | |
885 | for(i = pg_start; i < pg_start + pg_count; i++) | |
886 | if (ptes[i] != IOMMU_RESERVED_PTE) | |
887 | return -EBUSY; | |
888 | ||
889 | iommu_arena_free(arena, pg_start, pg_count); | |
890 | return 0; | |
891 | } | |
892 | ||
893 | int | |
894 | iommu_bind(struct pci_iommu_arena *arena, long pg_start, long pg_count, | |
d68721eb | 895 | struct page **pages) |
1da177e4 LT |
896 | { |
897 | unsigned long flags; | |
898 | unsigned long *ptes; | |
899 | long i, j; | |
900 | ||
901 | if (!arena) return -EINVAL; | |
902 | ||
903 | spin_lock_irqsave(&arena->lock, flags); | |
904 | ||
905 | ptes = arena->ptes; | |
906 | ||
907 | for(j = pg_start; j < pg_start + pg_count; j++) { | |
908 | if (ptes[j] != IOMMU_RESERVED_PTE) { | |
909 | spin_unlock_irqrestore(&arena->lock, flags); | |
910 | return -EBUSY; | |
911 | } | |
912 | } | |
913 | ||
914 | for(i = 0, j = pg_start; i < pg_count; i++, j++) | |
d68721eb | 915 | ptes[j] = mk_iommu_pte(page_to_phys(pages[i])); |
1da177e4 LT |
916 | |
917 | spin_unlock_irqrestore(&arena->lock, flags); | |
918 | ||
919 | return 0; | |
920 | } | |
921 | ||
922 | int | |
923 | iommu_unbind(struct pci_iommu_arena *arena, long pg_start, long pg_count) | |
924 | { | |
925 | unsigned long *p; | |
926 | long i; | |
927 | ||
928 | if (!arena) return -EINVAL; | |
929 | ||
930 | p = arena->ptes + pg_start; | |
931 | for(i = 0; i < pg_count; i++) | |
932 | p[i] = IOMMU_RESERVED_PTE; | |
933 | ||
934 | return 0; | |
935 | } | |
936 | ||
c186caca | 937 | static int alpha_pci_mapping_error(struct device *dev, dma_addr_t dma_addr) |
1da177e4 | 938 | { |
c186caca | 939 | return dma_addr == 0; |
1da177e4 LT |
940 | } |
941 | ||
5299709d | 942 | const struct dma_map_ops alpha_pci_ops = { |
4ce9a91f AP |
943 | .alloc = alpha_pci_alloc_coherent, |
944 | .free = alpha_pci_free_coherent, | |
c186caca FT |
945 | .map_page = alpha_pci_map_page, |
946 | .unmap_page = alpha_pci_unmap_page, | |
947 | .map_sg = alpha_pci_map_sg, | |
948 | .unmap_sg = alpha_pci_unmap_sg, | |
949 | .mapping_error = alpha_pci_mapping_error, | |
950 | .dma_supported = alpha_pci_supported, | |
c186caca FT |
951 | }; |
952 | ||
5299709d | 953 | const struct dma_map_ops *dma_ops = &alpha_pci_ops; |
c186caca | 954 | EXPORT_SYMBOL(dma_ops); |