mm: fix exec activate_mm vs TLB shootdown and lazy tlb switching race
[linux/fpc-iii.git] / arch / tile / kernel / pci-dma.c
blobf2abedc8a080e2ab8ac2f2d269167b88634fd4c7
1 /*
2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation, version 2.
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
11 * NON INFRINGEMENT. See the GNU General Public License for
12 * more details.
15 #include <linux/mm.h>
16 #include <linux/dma-mapping.h>
17 #include <linux/swiotlb.h>
18 #include <linux/vmalloc.h>
19 #include <linux/export.h>
20 #include <asm/tlbflush.h>
21 #include <asm/homecache.h>
23 /* Generic DMA mapping functions: */
26 * Allocate what Linux calls "coherent" memory. On TILEPro this is
27 * uncached memory; on TILE-Gx it is hash-for-home memory.
29 #ifdef __tilepro__
30 #define PAGE_HOME_DMA PAGE_HOME_UNCACHED
31 #else
32 #define PAGE_HOME_DMA PAGE_HOME_HASH
33 #endif
35 static void *tile_dma_alloc_coherent(struct device *dev, size_t size,
36 dma_addr_t *dma_handle, gfp_t gfp,
37 unsigned long attrs)
39 u64 dma_mask = (dev && dev->coherent_dma_mask) ?
40 dev->coherent_dma_mask : DMA_BIT_MASK(32);
41 int node = dev ? dev_to_node(dev) : 0;
42 int order = get_order(size);
43 struct page *pg;
44 dma_addr_t addr;
46 gfp |= __GFP_ZERO;
49 * If the mask specifies that the memory be in the first 4 GB, then
50 * we force the allocation to come from the DMA zone. We also
51 * force the node to 0 since that's the only node where the DMA
52 * zone isn't empty. If the mask size is smaller than 32 bits, we
53 * may still not be able to guarantee a suitable memory address, in
54 * which case we will return NULL. But such devices are uncommon.
56 if (dma_mask <= DMA_BIT_MASK(32)) {
57 gfp |= GFP_DMA;
58 node = 0;
61 pg = homecache_alloc_pages_node(node, gfp, order, PAGE_HOME_DMA);
62 if (pg == NULL)
63 return NULL;
65 addr = page_to_phys(pg);
66 if (addr + size > dma_mask) {
67 __homecache_free_pages(pg, order);
68 return NULL;
71 *dma_handle = addr;
73 return page_address(pg);
77 * Free memory that was allocated with tile_dma_alloc_coherent.
79 static void tile_dma_free_coherent(struct device *dev, size_t size,
80 void *vaddr, dma_addr_t dma_handle,
81 unsigned long attrs)
83 homecache_free_pages((unsigned long)vaddr, get_order(size));
87 * The map routines "map" the specified address range for DMA
88 * accesses. The memory belongs to the device after this call is
89 * issued, until it is unmapped with dma_unmap_single.
91 * We don't need to do any mapping, we just flush the address range
92 * out of the cache and return a DMA address.
94 * The unmap routines do whatever is necessary before the processor
95 * accesses the memory again, and must be called before the driver
96 * touches the memory. We can get away with a cache invalidate if we
97 * can count on nothing having been touched.
100 /* Set up a single page for DMA access. */
101 static void __dma_prep_page(struct page *page, unsigned long offset,
102 size_t size, enum dma_data_direction direction)
105 * Flush the page from cache if necessary.
106 * On tilegx, data is delivered to hash-for-home L3; on tilepro,
107 * data is delivered direct to memory.
109 * NOTE: If we were just doing DMA_TO_DEVICE we could optimize
110 * this to be a "flush" not a "finv" and keep some of the
111 * state in cache across the DMA operation, but it doesn't seem
112 * worth creating the necessary flush_buffer_xxx() infrastructure.
114 int home = page_home(page);
115 switch (home) {
116 case PAGE_HOME_HASH:
117 #ifdef __tilegx__
118 return;
119 #endif
120 break;
121 case PAGE_HOME_UNCACHED:
122 #ifdef __tilepro__
123 return;
124 #endif
125 break;
126 case PAGE_HOME_IMMUTABLE:
127 /* Should be going to the device only. */
128 BUG_ON(direction == DMA_FROM_DEVICE ||
129 direction == DMA_BIDIRECTIONAL);
130 return;
131 case PAGE_HOME_INCOHERENT:
132 /* Incoherent anyway, so no need to work hard here. */
133 return;
134 default:
135 BUG_ON(home < 0 || home >= NR_CPUS);
136 break;
138 homecache_finv_page(page);
140 #ifdef DEBUG_ALIGNMENT
141 /* Warn if the region isn't cacheline aligned. */
142 if (offset & (L2_CACHE_BYTES - 1) || (size & (L2_CACHE_BYTES - 1)))
143 pr_warn("Unaligned DMA to non-hfh memory: PA %#llx/%#lx\n",
144 PFN_PHYS(page_to_pfn(page)) + offset, size);
145 #endif
148 /* Make the page ready to be read by the core. */
149 static void __dma_complete_page(struct page *page, unsigned long offset,
150 size_t size, enum dma_data_direction direction)
152 #ifdef __tilegx__
153 switch (page_home(page)) {
154 case PAGE_HOME_HASH:
155 /* I/O device delivered data the way the cpu wanted it. */
156 break;
157 case PAGE_HOME_INCOHERENT:
158 /* Incoherent anyway, so no need to work hard here. */
159 break;
160 case PAGE_HOME_IMMUTABLE:
161 /* Extra read-only copies are not a problem. */
162 break;
163 default:
164 /* Flush the bogus hash-for-home I/O entries to memory. */
165 homecache_finv_map_page(page, PAGE_HOME_HASH);
166 break;
168 #endif
171 static void __dma_prep_pa_range(dma_addr_t dma_addr, size_t size,
172 enum dma_data_direction direction)
174 struct page *page = pfn_to_page(PFN_DOWN(dma_addr));
175 unsigned long offset = dma_addr & (PAGE_SIZE - 1);
176 size_t bytes = min(size, (size_t)(PAGE_SIZE - offset));
178 while (size != 0) {
179 __dma_prep_page(page, offset, bytes, direction);
180 size -= bytes;
181 ++page;
182 offset = 0;
183 bytes = min((size_t)PAGE_SIZE, size);
187 static void __dma_complete_pa_range(dma_addr_t dma_addr, size_t size,
188 enum dma_data_direction direction)
190 struct page *page = pfn_to_page(PFN_DOWN(dma_addr));
191 unsigned long offset = dma_addr & (PAGE_SIZE - 1);
192 size_t bytes = min(size, (size_t)(PAGE_SIZE - offset));
194 while (size != 0) {
195 __dma_complete_page(page, offset, bytes, direction);
196 size -= bytes;
197 ++page;
198 offset = 0;
199 bytes = min((size_t)PAGE_SIZE, size);
203 static int tile_dma_map_sg(struct device *dev, struct scatterlist *sglist,
204 int nents, enum dma_data_direction direction,
205 unsigned long attrs)
207 struct scatterlist *sg;
208 int i;
210 BUG_ON(!valid_dma_direction(direction));
212 WARN_ON(nents == 0 || sglist->length == 0);
214 for_each_sg(sglist, sg, nents, i) {
215 sg->dma_address = sg_phys(sg);
216 #ifdef CONFIG_NEED_SG_DMA_LENGTH
217 sg->dma_length = sg->length;
218 #endif
219 if (attrs & DMA_ATTR_SKIP_CPU_SYNC)
220 continue;
221 __dma_prep_pa_range(sg->dma_address, sg->length, direction);
224 return nents;
227 static void tile_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
228 int nents, enum dma_data_direction direction,
229 unsigned long attrs)
231 struct scatterlist *sg;
232 int i;
234 BUG_ON(!valid_dma_direction(direction));
235 for_each_sg(sglist, sg, nents, i) {
236 sg->dma_address = sg_phys(sg);
237 if (attrs & DMA_ATTR_SKIP_CPU_SYNC)
238 continue;
239 __dma_complete_pa_range(sg->dma_address, sg->length,
240 direction);
244 static dma_addr_t tile_dma_map_page(struct device *dev, struct page *page,
245 unsigned long offset, size_t size,
246 enum dma_data_direction direction,
247 unsigned long attrs)
249 BUG_ON(!valid_dma_direction(direction));
251 BUG_ON(offset + size > PAGE_SIZE);
252 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
253 __dma_prep_page(page, offset, size, direction);
255 return page_to_pa(page) + offset;
258 static void tile_dma_unmap_page(struct device *dev, dma_addr_t dma_address,
259 size_t size, enum dma_data_direction direction,
260 unsigned long attrs)
262 BUG_ON(!valid_dma_direction(direction));
264 if (attrs & DMA_ATTR_SKIP_CPU_SYNC)
265 return;
267 __dma_complete_page(pfn_to_page(PFN_DOWN(dma_address)),
268 dma_address & (PAGE_SIZE - 1), size, direction);
271 static void tile_dma_sync_single_for_cpu(struct device *dev,
272 dma_addr_t dma_handle,
273 size_t size,
274 enum dma_data_direction direction)
276 BUG_ON(!valid_dma_direction(direction));
278 __dma_complete_pa_range(dma_handle, size, direction);
281 static void tile_dma_sync_single_for_device(struct device *dev,
282 dma_addr_t dma_handle, size_t size,
283 enum dma_data_direction direction)
285 __dma_prep_pa_range(dma_handle, size, direction);
288 static void tile_dma_sync_sg_for_cpu(struct device *dev,
289 struct scatterlist *sglist, int nelems,
290 enum dma_data_direction direction)
292 struct scatterlist *sg;
293 int i;
295 BUG_ON(!valid_dma_direction(direction));
296 WARN_ON(nelems == 0 || sglist->length == 0);
298 for_each_sg(sglist, sg, nelems, i) {
299 dma_sync_single_for_cpu(dev, sg->dma_address,
300 sg_dma_len(sg), direction);
304 static void tile_dma_sync_sg_for_device(struct device *dev,
305 struct scatterlist *sglist, int nelems,
306 enum dma_data_direction direction)
308 struct scatterlist *sg;
309 int i;
311 BUG_ON(!valid_dma_direction(direction));
312 WARN_ON(nelems == 0 || sglist->length == 0);
314 for_each_sg(sglist, sg, nelems, i) {
315 dma_sync_single_for_device(dev, sg->dma_address,
316 sg_dma_len(sg), direction);
320 static const struct dma_map_ops tile_default_dma_map_ops = {
321 .alloc = tile_dma_alloc_coherent,
322 .free = tile_dma_free_coherent,
323 .map_page = tile_dma_map_page,
324 .unmap_page = tile_dma_unmap_page,
325 .map_sg = tile_dma_map_sg,
326 .unmap_sg = tile_dma_unmap_sg,
327 .sync_single_for_cpu = tile_dma_sync_single_for_cpu,
328 .sync_single_for_device = tile_dma_sync_single_for_device,
329 .sync_sg_for_cpu = tile_dma_sync_sg_for_cpu,
330 .sync_sg_for_device = tile_dma_sync_sg_for_device,
333 const struct dma_map_ops *tile_dma_map_ops = &tile_default_dma_map_ops;
334 EXPORT_SYMBOL(tile_dma_map_ops);
336 /* Generic PCI DMA mapping functions */
338 static void *tile_pci_dma_alloc_coherent(struct device *dev, size_t size,
339 dma_addr_t *dma_handle, gfp_t gfp,
340 unsigned long attrs)
342 int node = dev_to_node(dev);
343 int order = get_order(size);
344 struct page *pg;
345 dma_addr_t addr;
347 gfp |= __GFP_ZERO;
349 pg = homecache_alloc_pages_node(node, gfp, order, PAGE_HOME_DMA);
350 if (pg == NULL)
351 return NULL;
353 addr = page_to_phys(pg);
355 *dma_handle = addr + get_dma_offset(dev);
357 return page_address(pg);
361 * Free memory that was allocated with tile_pci_dma_alloc_coherent.
363 static void tile_pci_dma_free_coherent(struct device *dev, size_t size,
364 void *vaddr, dma_addr_t dma_handle,
365 unsigned long attrs)
367 homecache_free_pages((unsigned long)vaddr, get_order(size));
370 static int tile_pci_dma_map_sg(struct device *dev, struct scatterlist *sglist,
371 int nents, enum dma_data_direction direction,
372 unsigned long attrs)
374 struct scatterlist *sg;
375 int i;
377 BUG_ON(!valid_dma_direction(direction));
379 WARN_ON(nents == 0 || sglist->length == 0);
381 for_each_sg(sglist, sg, nents, i) {
382 sg->dma_address = sg_phys(sg);
383 __dma_prep_pa_range(sg->dma_address, sg->length, direction);
385 sg->dma_address = sg->dma_address + get_dma_offset(dev);
386 #ifdef CONFIG_NEED_SG_DMA_LENGTH
387 sg->dma_length = sg->length;
388 #endif
391 return nents;
394 static void tile_pci_dma_unmap_sg(struct device *dev,
395 struct scatterlist *sglist, int nents,
396 enum dma_data_direction direction,
397 unsigned long attrs)
399 struct scatterlist *sg;
400 int i;
402 BUG_ON(!valid_dma_direction(direction));
403 for_each_sg(sglist, sg, nents, i) {
404 sg->dma_address = sg_phys(sg);
405 __dma_complete_pa_range(sg->dma_address, sg->length,
406 direction);
410 static dma_addr_t tile_pci_dma_map_page(struct device *dev, struct page *page,
411 unsigned long offset, size_t size,
412 enum dma_data_direction direction,
413 unsigned long attrs)
415 BUG_ON(!valid_dma_direction(direction));
417 BUG_ON(offset + size > PAGE_SIZE);
418 __dma_prep_page(page, offset, size, direction);
420 return page_to_pa(page) + offset + get_dma_offset(dev);
423 static void tile_pci_dma_unmap_page(struct device *dev, dma_addr_t dma_address,
424 size_t size,
425 enum dma_data_direction direction,
426 unsigned long attrs)
428 BUG_ON(!valid_dma_direction(direction));
430 dma_address -= get_dma_offset(dev);
432 __dma_complete_page(pfn_to_page(PFN_DOWN(dma_address)),
433 dma_address & (PAGE_SIZE - 1), size, direction);
436 static void tile_pci_dma_sync_single_for_cpu(struct device *dev,
437 dma_addr_t dma_handle,
438 size_t size,
439 enum dma_data_direction direction)
441 BUG_ON(!valid_dma_direction(direction));
443 dma_handle -= get_dma_offset(dev);
445 __dma_complete_pa_range(dma_handle, size, direction);
448 static void tile_pci_dma_sync_single_for_device(struct device *dev,
449 dma_addr_t dma_handle,
450 size_t size,
451 enum dma_data_direction
452 direction)
454 dma_handle -= get_dma_offset(dev);
456 __dma_prep_pa_range(dma_handle, size, direction);
459 static void tile_pci_dma_sync_sg_for_cpu(struct device *dev,
460 struct scatterlist *sglist,
461 int nelems,
462 enum dma_data_direction direction)
464 struct scatterlist *sg;
465 int i;
467 BUG_ON(!valid_dma_direction(direction));
468 WARN_ON(nelems == 0 || sglist->length == 0);
470 for_each_sg(sglist, sg, nelems, i) {
471 dma_sync_single_for_cpu(dev, sg->dma_address,
472 sg_dma_len(sg), direction);
476 static void tile_pci_dma_sync_sg_for_device(struct device *dev,
477 struct scatterlist *sglist,
478 int nelems,
479 enum dma_data_direction direction)
481 struct scatterlist *sg;
482 int i;
484 BUG_ON(!valid_dma_direction(direction));
485 WARN_ON(nelems == 0 || sglist->length == 0);
487 for_each_sg(sglist, sg, nelems, i) {
488 dma_sync_single_for_device(dev, sg->dma_address,
489 sg_dma_len(sg), direction);
493 static const struct dma_map_ops tile_pci_default_dma_map_ops = {
494 .alloc = tile_pci_dma_alloc_coherent,
495 .free = tile_pci_dma_free_coherent,
496 .map_page = tile_pci_dma_map_page,
497 .unmap_page = tile_pci_dma_unmap_page,
498 .map_sg = tile_pci_dma_map_sg,
499 .unmap_sg = tile_pci_dma_unmap_sg,
500 .sync_single_for_cpu = tile_pci_dma_sync_single_for_cpu,
501 .sync_single_for_device = tile_pci_dma_sync_single_for_device,
502 .sync_sg_for_cpu = tile_pci_dma_sync_sg_for_cpu,
503 .sync_sg_for_device = tile_pci_dma_sync_sg_for_device,
506 const struct dma_map_ops *gx_pci_dma_map_ops = &tile_pci_default_dma_map_ops;
507 EXPORT_SYMBOL(gx_pci_dma_map_ops);
509 /* PCI DMA mapping functions for legacy PCI devices */
511 #ifdef CONFIG_SWIOTLB
512 static void *tile_swiotlb_alloc_coherent(struct device *dev, size_t size,
513 dma_addr_t *dma_handle, gfp_t gfp,
514 unsigned long attrs)
516 gfp |= GFP_DMA;
517 return swiotlb_alloc_coherent(dev, size, dma_handle, gfp);
520 static void tile_swiotlb_free_coherent(struct device *dev, size_t size,
521 void *vaddr, dma_addr_t dma_addr,
522 unsigned long attrs)
524 swiotlb_free_coherent(dev, size, vaddr, dma_addr);
527 static const struct dma_map_ops pci_swiotlb_dma_ops = {
528 .alloc = tile_swiotlb_alloc_coherent,
529 .free = tile_swiotlb_free_coherent,
530 .map_page = swiotlb_map_page,
531 .unmap_page = swiotlb_unmap_page,
532 .map_sg = swiotlb_map_sg_attrs,
533 .unmap_sg = swiotlb_unmap_sg_attrs,
534 .sync_single_for_cpu = swiotlb_sync_single_for_cpu,
535 .sync_single_for_device = swiotlb_sync_single_for_device,
536 .sync_sg_for_cpu = swiotlb_sync_sg_for_cpu,
537 .sync_sg_for_device = swiotlb_sync_sg_for_device,
538 .dma_supported = swiotlb_dma_supported,
539 .mapping_error = swiotlb_dma_mapping_error,
542 static const struct dma_map_ops pci_hybrid_dma_ops = {
543 .alloc = tile_swiotlb_alloc_coherent,
544 .free = tile_swiotlb_free_coherent,
545 .map_page = tile_pci_dma_map_page,
546 .unmap_page = tile_pci_dma_unmap_page,
547 .map_sg = tile_pci_dma_map_sg,
548 .unmap_sg = tile_pci_dma_unmap_sg,
549 .sync_single_for_cpu = tile_pci_dma_sync_single_for_cpu,
550 .sync_single_for_device = tile_pci_dma_sync_single_for_device,
551 .sync_sg_for_cpu = tile_pci_dma_sync_sg_for_cpu,
552 .sync_sg_for_device = tile_pci_dma_sync_sg_for_device,
555 const struct dma_map_ops *gx_legacy_pci_dma_map_ops = &pci_swiotlb_dma_ops;
556 const struct dma_map_ops *gx_hybrid_pci_dma_map_ops = &pci_hybrid_dma_ops;
557 #else
558 const struct dma_map_ops *gx_legacy_pci_dma_map_ops;
559 const struct dma_map_ops *gx_hybrid_pci_dma_map_ops;
560 #endif
561 EXPORT_SYMBOL(gx_legacy_pci_dma_map_ops);
562 EXPORT_SYMBOL(gx_hybrid_pci_dma_map_ops);
564 int dma_set_mask(struct device *dev, u64 mask)
566 const struct dma_map_ops *dma_ops = get_dma_ops(dev);
569 * For PCI devices with 64-bit DMA addressing capability, promote
570 * the dma_ops to hybrid, with the consistent memory DMA space limited
571 * to 32-bit. For 32-bit capable devices, limit the streaming DMA
572 * address range to max_direct_dma_addr.
574 if (dma_ops == gx_pci_dma_map_ops ||
575 dma_ops == gx_hybrid_pci_dma_map_ops ||
576 dma_ops == gx_legacy_pci_dma_map_ops) {
577 if (mask == DMA_BIT_MASK(64) &&
578 dma_ops == gx_legacy_pci_dma_map_ops)
579 set_dma_ops(dev, gx_hybrid_pci_dma_map_ops);
580 else if (mask > dev->archdata.max_direct_dma_addr)
581 mask = dev->archdata.max_direct_dma_addr;
584 if (!dev->dma_mask || !dma_supported(dev, mask))
585 return -EIO;
587 *dev->dma_mask = mask;
589 return 0;
591 EXPORT_SYMBOL(dma_set_mask);
593 #ifdef CONFIG_ARCH_HAS_DMA_SET_COHERENT_MASK
594 int dma_set_coherent_mask(struct device *dev, u64 mask)
596 const struct dma_map_ops *dma_ops = get_dma_ops(dev);
599 * For PCI devices with 64-bit DMA addressing capability, promote
600 * the dma_ops to full capability for both streams and consistent
601 * memory access. For 32-bit capable devices, limit the consistent
602 * memory DMA range to max_direct_dma_addr.
604 if (dma_ops == gx_pci_dma_map_ops ||
605 dma_ops == gx_hybrid_pci_dma_map_ops ||
606 dma_ops == gx_legacy_pci_dma_map_ops) {
607 if (mask == DMA_BIT_MASK(64))
608 set_dma_ops(dev, gx_pci_dma_map_ops);
609 else if (mask > dev->archdata.max_direct_dma_addr)
610 mask = dev->archdata.max_direct_dma_addr;
613 if (!dma_supported(dev, mask))
614 return -EIO;
615 dev->coherent_dma_mask = mask;
616 return 0;
618 EXPORT_SYMBOL(dma_set_coherent_mask);
619 #endif
621 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
623 * The generic dma_get_required_mask() uses the highest physical address
624 * (max_pfn) to provide the hint to the PCI drivers regarding 32-bit or
625 * 64-bit DMA configuration. Since TILEGx has I/O TLB/MMU, allowing the
626 * DMAs to use the full 64-bit PCI address space and not limited by
627 * the physical memory space, we always let the PCI devices use
628 * 64-bit DMA if they have that capability, by returning the 64-bit
629 * DMA mask here. The device driver has the option to use 32-bit DMA if
630 * the device is not capable of 64-bit DMA.
632 u64 dma_get_required_mask(struct device *dev)
634 return DMA_BIT_MASK(64);
636 EXPORT_SYMBOL_GPL(dma_get_required_mask);
637 #endif