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
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.
30 #define PAGE_HOME_DMA PAGE_HOME_UNCACHED
32 #define PAGE_HOME_DMA PAGE_HOME_HASH
35 static void *tile_dma_alloc_coherent(struct device *dev, size_t size,
36 dma_addr_t *dma_handle, gfp_t gfp,
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);
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)) {
61 pg = homecache_alloc_pages_node(node, gfp, order, PAGE_HOME_DMA);
65 addr = page_to_phys(pg);
66 if (addr + size > dma_mask) {
67 __homecache_free_pages(pg, order);
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,
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);
121 case PAGE_HOME_UNCACHED:
126 case PAGE_HOME_IMMUTABLE:
127 /* Should be going to the device only. */
128 BUG_ON(direction == DMA_FROM_DEVICE ||
129 direction == DMA_BIDIRECTIONAL);
131 case PAGE_HOME_INCOHERENT:
132 /* Incoherent anyway, so no need to work hard here. */
135 BUG_ON(home < 0 || home >= NR_CPUS);
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);
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)
153 switch (page_home(page)) {
155 /* I/O device delivered data the way the cpu wanted it. */
157 case PAGE_HOME_INCOHERENT:
158 /* Incoherent anyway, so no need to work hard here. */
160 case PAGE_HOME_IMMUTABLE:
161 /* Extra read-only copies are not a problem. */
164 /* Flush the bogus hash-for-home I/O entries to memory. */
165 homecache_finv_map_page(page, PAGE_HOME_HASH);
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));
179 __dma_prep_page(page, offset, bytes, direction);
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));
195 __dma_complete_page(page, offset, bytes, direction);
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,
207 struct scatterlist *sg;
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;
219 if (attrs & DMA_ATTR_SKIP_CPU_SYNC)
221 __dma_prep_pa_range(sg->dma_address, sg->length, direction);
227 static void tile_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
228 int nents, enum dma_data_direction direction,
231 struct scatterlist *sg;
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)
239 __dma_complete_pa_range(sg->dma_address, sg->length,
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,
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,
262 BUG_ON(!valid_dma_direction(direction));
264 if (attrs & DMA_ATTR_SKIP_CPU_SYNC)
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,
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;
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;
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);
321 tile_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
327 tile_dma_supported(struct device *dev, u64 mask)
332 static struct dma_map_ops tile_default_dma_map_ops = {
333 .alloc = tile_dma_alloc_coherent,
334 .free = tile_dma_free_coherent,
335 .map_page = tile_dma_map_page,
336 .unmap_page = tile_dma_unmap_page,
337 .map_sg = tile_dma_map_sg,
338 .unmap_sg = tile_dma_unmap_sg,
339 .sync_single_for_cpu = tile_dma_sync_single_for_cpu,
340 .sync_single_for_device = tile_dma_sync_single_for_device,
341 .sync_sg_for_cpu = tile_dma_sync_sg_for_cpu,
342 .sync_sg_for_device = tile_dma_sync_sg_for_device,
343 .mapping_error = tile_dma_mapping_error,
344 .dma_supported = tile_dma_supported
347 struct dma_map_ops *tile_dma_map_ops = &tile_default_dma_map_ops;
348 EXPORT_SYMBOL(tile_dma_map_ops);
350 /* Generic PCI DMA mapping functions */
352 static void *tile_pci_dma_alloc_coherent(struct device *dev, size_t size,
353 dma_addr_t *dma_handle, gfp_t gfp,
356 int node = dev_to_node(dev);
357 int order = get_order(size);
363 pg = homecache_alloc_pages_node(node, gfp, order, PAGE_HOME_DMA);
367 addr = page_to_phys(pg);
369 *dma_handle = addr + get_dma_offset(dev);
371 return page_address(pg);
375 * Free memory that was allocated with tile_pci_dma_alloc_coherent.
377 static void tile_pci_dma_free_coherent(struct device *dev, size_t size,
378 void *vaddr, dma_addr_t dma_handle,
381 homecache_free_pages((unsigned long)vaddr, get_order(size));
384 static int tile_pci_dma_map_sg(struct device *dev, struct scatterlist *sglist,
385 int nents, enum dma_data_direction direction,
388 struct scatterlist *sg;
391 BUG_ON(!valid_dma_direction(direction));
393 WARN_ON(nents == 0 || sglist->length == 0);
395 for_each_sg(sglist, sg, nents, i) {
396 sg->dma_address = sg_phys(sg);
397 __dma_prep_pa_range(sg->dma_address, sg->length, direction);
399 sg->dma_address = sg->dma_address + get_dma_offset(dev);
400 #ifdef CONFIG_NEED_SG_DMA_LENGTH
401 sg->dma_length = sg->length;
408 static void tile_pci_dma_unmap_sg(struct device *dev,
409 struct scatterlist *sglist, int nents,
410 enum dma_data_direction direction,
413 struct scatterlist *sg;
416 BUG_ON(!valid_dma_direction(direction));
417 for_each_sg(sglist, sg, nents, i) {
418 sg->dma_address = sg_phys(sg);
419 __dma_complete_pa_range(sg->dma_address, sg->length,
424 static dma_addr_t tile_pci_dma_map_page(struct device *dev, struct page *page,
425 unsigned long offset, size_t size,
426 enum dma_data_direction direction,
429 BUG_ON(!valid_dma_direction(direction));
431 BUG_ON(offset + size > PAGE_SIZE);
432 __dma_prep_page(page, offset, size, direction);
434 return page_to_pa(page) + offset + get_dma_offset(dev);
437 static void tile_pci_dma_unmap_page(struct device *dev, dma_addr_t dma_address,
439 enum dma_data_direction direction,
442 BUG_ON(!valid_dma_direction(direction));
444 dma_address -= get_dma_offset(dev);
446 __dma_complete_page(pfn_to_page(PFN_DOWN(dma_address)),
447 dma_address & (PAGE_SIZE - 1), size, direction);
450 static void tile_pci_dma_sync_single_for_cpu(struct device *dev,
451 dma_addr_t dma_handle,
453 enum dma_data_direction direction)
455 BUG_ON(!valid_dma_direction(direction));
457 dma_handle -= get_dma_offset(dev);
459 __dma_complete_pa_range(dma_handle, size, direction);
462 static void tile_pci_dma_sync_single_for_device(struct device *dev,
463 dma_addr_t dma_handle,
465 enum dma_data_direction
468 dma_handle -= get_dma_offset(dev);
470 __dma_prep_pa_range(dma_handle, size, direction);
473 static void tile_pci_dma_sync_sg_for_cpu(struct device *dev,
474 struct scatterlist *sglist,
476 enum dma_data_direction direction)
478 struct scatterlist *sg;
481 BUG_ON(!valid_dma_direction(direction));
482 WARN_ON(nelems == 0 || sglist->length == 0);
484 for_each_sg(sglist, sg, nelems, i) {
485 dma_sync_single_for_cpu(dev, sg->dma_address,
486 sg_dma_len(sg), direction);
490 static void tile_pci_dma_sync_sg_for_device(struct device *dev,
491 struct scatterlist *sglist,
493 enum dma_data_direction direction)
495 struct scatterlist *sg;
498 BUG_ON(!valid_dma_direction(direction));
499 WARN_ON(nelems == 0 || sglist->length == 0);
501 for_each_sg(sglist, sg, nelems, i) {
502 dma_sync_single_for_device(dev, sg->dma_address,
503 sg_dma_len(sg), direction);
508 tile_pci_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
514 tile_pci_dma_supported(struct device *dev, u64 mask)
519 static struct dma_map_ops tile_pci_default_dma_map_ops = {
520 .alloc = tile_pci_dma_alloc_coherent,
521 .free = tile_pci_dma_free_coherent,
522 .map_page = tile_pci_dma_map_page,
523 .unmap_page = tile_pci_dma_unmap_page,
524 .map_sg = tile_pci_dma_map_sg,
525 .unmap_sg = tile_pci_dma_unmap_sg,
526 .sync_single_for_cpu = tile_pci_dma_sync_single_for_cpu,
527 .sync_single_for_device = tile_pci_dma_sync_single_for_device,
528 .sync_sg_for_cpu = tile_pci_dma_sync_sg_for_cpu,
529 .sync_sg_for_device = tile_pci_dma_sync_sg_for_device,
530 .mapping_error = tile_pci_dma_mapping_error,
531 .dma_supported = tile_pci_dma_supported
534 struct dma_map_ops *gx_pci_dma_map_ops = &tile_pci_default_dma_map_ops;
535 EXPORT_SYMBOL(gx_pci_dma_map_ops);
537 /* PCI DMA mapping functions for legacy PCI devices */
539 #ifdef CONFIG_SWIOTLB
540 static void *tile_swiotlb_alloc_coherent(struct device *dev, size_t size,
541 dma_addr_t *dma_handle, gfp_t gfp,
545 return swiotlb_alloc_coherent(dev, size, dma_handle, gfp);
548 static void tile_swiotlb_free_coherent(struct device *dev, size_t size,
549 void *vaddr, dma_addr_t dma_addr,
552 swiotlb_free_coherent(dev, size, vaddr, dma_addr);
555 static struct dma_map_ops pci_swiotlb_dma_ops = {
556 .alloc = tile_swiotlb_alloc_coherent,
557 .free = tile_swiotlb_free_coherent,
558 .map_page = swiotlb_map_page,
559 .unmap_page = swiotlb_unmap_page,
560 .map_sg = swiotlb_map_sg_attrs,
561 .unmap_sg = swiotlb_unmap_sg_attrs,
562 .sync_single_for_cpu = swiotlb_sync_single_for_cpu,
563 .sync_single_for_device = swiotlb_sync_single_for_device,
564 .sync_sg_for_cpu = swiotlb_sync_sg_for_cpu,
565 .sync_sg_for_device = swiotlb_sync_sg_for_device,
566 .dma_supported = swiotlb_dma_supported,
567 .mapping_error = swiotlb_dma_mapping_error,
570 static struct dma_map_ops pci_hybrid_dma_ops = {
571 .alloc = tile_swiotlb_alloc_coherent,
572 .free = tile_swiotlb_free_coherent,
573 .map_page = tile_pci_dma_map_page,
574 .unmap_page = tile_pci_dma_unmap_page,
575 .map_sg = tile_pci_dma_map_sg,
576 .unmap_sg = tile_pci_dma_unmap_sg,
577 .sync_single_for_cpu = tile_pci_dma_sync_single_for_cpu,
578 .sync_single_for_device = tile_pci_dma_sync_single_for_device,
579 .sync_sg_for_cpu = tile_pci_dma_sync_sg_for_cpu,
580 .sync_sg_for_device = tile_pci_dma_sync_sg_for_device,
581 .mapping_error = tile_pci_dma_mapping_error,
582 .dma_supported = tile_pci_dma_supported
585 struct dma_map_ops *gx_legacy_pci_dma_map_ops = &pci_swiotlb_dma_ops;
586 struct dma_map_ops *gx_hybrid_pci_dma_map_ops = &pci_hybrid_dma_ops;
588 struct dma_map_ops *gx_legacy_pci_dma_map_ops;
589 struct dma_map_ops *gx_hybrid_pci_dma_map_ops;
591 EXPORT_SYMBOL(gx_legacy_pci_dma_map_ops);
592 EXPORT_SYMBOL(gx_hybrid_pci_dma_map_ops);
594 int dma_set_mask(struct device *dev, u64 mask)
596 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 hybrid, with the consistent memory DMA space limited
601 * to 32-bit. For 32-bit capable devices, limit the streaming DMA
602 * address 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 dma_ops == gx_legacy_pci_dma_map_ops)
609 set_dma_ops(dev, gx_hybrid_pci_dma_map_ops);
610 else if (mask > dev->archdata.max_direct_dma_addr)
611 mask = dev->archdata.max_direct_dma_addr;
614 if (!dev->dma_mask || !dma_supported(dev, mask))
617 *dev->dma_mask = mask;
621 EXPORT_SYMBOL(dma_set_mask);
623 #ifdef CONFIG_ARCH_HAS_DMA_SET_COHERENT_MASK
624 int dma_set_coherent_mask(struct device *dev, u64 mask)
626 struct dma_map_ops *dma_ops = get_dma_ops(dev);
629 * For PCI devices with 64-bit DMA addressing capability, promote
630 * the dma_ops to full capability for both streams and consistent
631 * memory access. For 32-bit capable devices, limit the consistent
632 * memory DMA range to max_direct_dma_addr.
634 if (dma_ops == gx_pci_dma_map_ops ||
635 dma_ops == gx_hybrid_pci_dma_map_ops ||
636 dma_ops == gx_legacy_pci_dma_map_ops) {
637 if (mask == DMA_BIT_MASK(64))
638 set_dma_ops(dev, gx_pci_dma_map_ops);
639 else if (mask > dev->archdata.max_direct_dma_addr)
640 mask = dev->archdata.max_direct_dma_addr;
643 if (!dma_supported(dev, mask))
645 dev->coherent_dma_mask = mask;
648 EXPORT_SYMBOL(dma_set_coherent_mask);
651 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
653 * The generic dma_get_required_mask() uses the highest physical address
654 * (max_pfn) to provide the hint to the PCI drivers regarding 32-bit or
655 * 64-bit DMA configuration. Since TILEGx has I/O TLB/MMU, allowing the
656 * DMAs to use the full 64-bit PCI address space and not limited by
657 * the physical memory space, we always let the PCI devices use
658 * 64-bit DMA if they have that capability, by returning the 64-bit
659 * DMA mask here. The device driver has the option to use 32-bit DMA if
660 * the device is not capable of 64-bit DMA.
662 u64 dma_get_required_mask(struct device *dev)
664 return DMA_BIT_MASK(64);
666 EXPORT_SYMBOL_GPL(dma_get_required_mask);