Merge remote-tracking branch 'moduleh/module.h-split'
[linux-2.6/next.git] / drivers / xen / swiotlb-xen.c
blobbfe438ffb8541362b1673ea0d1ed1d5bd3a37faa
1 /*
2 * Copyright 2010
3 * by Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
5 * This code provides a IOMMU for Xen PV guests with PCI passthrough.
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License v2.0 as published by
9 * the Free Software Foundation
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * PV guests under Xen are running in an non-contiguous memory architecture.
18 * When PCI pass-through is utilized, this necessitates an IOMMU for
19 * translating bus (DMA) to virtual and vice-versa and also providing a
20 * mechanism to have contiguous pages for device drivers operations (say DMA
21 * operations).
23 * Specifically, under Xen the Linux idea of pages is an illusion. It
24 * assumes that pages start at zero and go up to the available memory. To
25 * help with that, the Linux Xen MMU provides a lookup mechanism to
26 * translate the page frame numbers (PFN) to machine frame numbers (MFN)
27 * and vice-versa. The MFN are the "real" frame numbers. Furthermore
28 * memory is not contiguous. Xen hypervisor stitches memory for guests
29 * from different pools, which means there is no guarantee that PFN==MFN
30 * and PFN+1==MFN+1. Lastly with Xen 4.0, pages (in debug mode) are
31 * allocated in descending order (high to low), meaning the guest might
32 * never get any MFN's under the 4GB mark.
36 #include <linux/bootmem.h>
37 #include <linux/dma-mapping.h>
38 #include <linux/export.h>
39 #include <xen/swiotlb-xen.h>
40 #include <xen/page.h>
41 #include <xen/xen-ops.h>
42 #include <xen/hvc-console.h>
44 * Used to do a quick range check in swiotlb_tbl_unmap_single and
45 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
46 * API.
49 static char *xen_io_tlb_start, *xen_io_tlb_end;
50 static unsigned long xen_io_tlb_nslabs;
52 * Quick lookup value of the bus address of the IOTLB.
55 u64 start_dma_addr;
57 static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
59 return phys_to_machine(XPADDR(paddr)).maddr;
62 static phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
64 return machine_to_phys(XMADDR(baddr)).paddr;
67 static dma_addr_t xen_virt_to_bus(void *address)
69 return xen_phys_to_bus(virt_to_phys(address));
72 static int check_pages_physically_contiguous(unsigned long pfn,
73 unsigned int offset,
74 size_t length)
76 unsigned long next_mfn;
77 int i;
78 int nr_pages;
80 next_mfn = pfn_to_mfn(pfn);
81 nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
83 for (i = 1; i < nr_pages; i++) {
84 if (pfn_to_mfn(++pfn) != ++next_mfn)
85 return 0;
87 return 1;
90 static int range_straddles_page_boundary(phys_addr_t p, size_t size)
92 unsigned long pfn = PFN_DOWN(p);
93 unsigned int offset = p & ~PAGE_MASK;
95 if (offset + size <= PAGE_SIZE)
96 return 0;
97 if (check_pages_physically_contiguous(pfn, offset, size))
98 return 0;
99 return 1;
102 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
104 unsigned long mfn = PFN_DOWN(dma_addr);
105 unsigned long pfn = mfn_to_local_pfn(mfn);
106 phys_addr_t paddr;
108 /* If the address is outside our domain, it CAN
109 * have the same virtual address as another address
110 * in our domain. Therefore _only_ check address within our domain.
112 if (pfn_valid(pfn)) {
113 paddr = PFN_PHYS(pfn);
114 return paddr >= virt_to_phys(xen_io_tlb_start) &&
115 paddr < virt_to_phys(xen_io_tlb_end);
117 return 0;
120 static int max_dma_bits = 32;
122 static int
123 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
125 int i, rc;
126 int dma_bits;
128 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
130 i = 0;
131 do {
132 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
134 do {
135 rc = xen_create_contiguous_region(
136 (unsigned long)buf + (i << IO_TLB_SHIFT),
137 get_order(slabs << IO_TLB_SHIFT),
138 dma_bits);
139 } while (rc && dma_bits++ < max_dma_bits);
140 if (rc)
141 return rc;
143 i += slabs;
144 } while (i < nslabs);
145 return 0;
148 void __init xen_swiotlb_init(int verbose)
150 unsigned long bytes;
151 int rc = -ENOMEM;
152 unsigned long nr_tbl;
153 char *m = NULL;
154 unsigned int repeat = 3;
156 nr_tbl = swioltb_nr_tbl();
157 if (nr_tbl)
158 xen_io_tlb_nslabs = nr_tbl;
159 else {
160 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
161 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
163 retry:
164 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
167 * Get IO TLB memory from any location.
169 xen_io_tlb_start = alloc_bootmem(bytes);
170 if (!xen_io_tlb_start) {
171 m = "Cannot allocate Xen-SWIOTLB buffer!\n";
172 goto error;
174 xen_io_tlb_end = xen_io_tlb_start + bytes;
176 * And replace that memory with pages under 4GB.
178 rc = xen_swiotlb_fixup(xen_io_tlb_start,
179 bytes,
180 xen_io_tlb_nslabs);
181 if (rc) {
182 free_bootmem(__pa(xen_io_tlb_start), bytes);
183 m = "Failed to get contiguous memory for DMA from Xen!\n"\
184 "You either: don't have the permissions, do not have"\
185 " enough free memory under 4GB, or the hypervisor memory"\
186 "is too fragmented!";
187 goto error;
189 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
190 swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs, verbose);
192 return;
193 error:
194 if (repeat--) {
195 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
196 (xen_io_tlb_nslabs >> 1));
197 printk(KERN_INFO "Xen-SWIOTLB: Lowering to %luMB\n",
198 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
199 goto retry;
201 xen_raw_printk("%s (rc:%d)", m, rc);
202 panic("%s (rc:%d)", m, rc);
205 void *
206 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
207 dma_addr_t *dma_handle, gfp_t flags)
209 void *ret;
210 int order = get_order(size);
211 u64 dma_mask = DMA_BIT_MASK(32);
212 unsigned long vstart;
215 * Ignore region specifiers - the kernel's ideas of
216 * pseudo-phys memory layout has nothing to do with the
217 * machine physical layout. We can't allocate highmem
218 * because we can't return a pointer to it.
220 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
222 if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
223 return ret;
225 vstart = __get_free_pages(flags, order);
226 ret = (void *)vstart;
228 if (hwdev && hwdev->coherent_dma_mask)
229 dma_mask = dma_alloc_coherent_mask(hwdev, flags);
231 if (ret) {
232 if (xen_create_contiguous_region(vstart, order,
233 fls64(dma_mask)) != 0) {
234 free_pages(vstart, order);
235 return NULL;
237 memset(ret, 0, size);
238 *dma_handle = virt_to_machine(ret).maddr;
240 return ret;
242 EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
244 void
245 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
246 dma_addr_t dev_addr)
248 int order = get_order(size);
250 if (dma_release_from_coherent(hwdev, order, vaddr))
251 return;
253 xen_destroy_contiguous_region((unsigned long)vaddr, order);
254 free_pages((unsigned long)vaddr, order);
256 EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
260 * Map a single buffer of the indicated size for DMA in streaming mode. The
261 * physical address to use is returned.
263 * Once the device is given the dma address, the device owns this memory until
264 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
266 dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
267 unsigned long offset, size_t size,
268 enum dma_data_direction dir,
269 struct dma_attrs *attrs)
271 phys_addr_t phys = page_to_phys(page) + offset;
272 dma_addr_t dev_addr = xen_phys_to_bus(phys);
273 void *map;
275 BUG_ON(dir == DMA_NONE);
277 * If the address happens to be in the device's DMA window,
278 * we can safely return the device addr and not worry about bounce
279 * buffering it.
281 if (dma_capable(dev, dev_addr, size) &&
282 !range_straddles_page_boundary(phys, size) && !swiotlb_force)
283 return dev_addr;
286 * Oh well, have to allocate and map a bounce buffer.
288 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
289 if (!map)
290 return DMA_ERROR_CODE;
292 dev_addr = xen_virt_to_bus(map);
295 * Ensure that the address returned is DMA'ble
297 if (!dma_capable(dev, dev_addr, size)) {
298 swiotlb_tbl_unmap_single(dev, map, size, dir);
299 dev_addr = 0;
301 return dev_addr;
303 EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
306 * Unmap a single streaming mode DMA translation. The dma_addr and size must
307 * match what was provided for in a previous xen_swiotlb_map_page call. All
308 * other usages are undefined.
310 * After this call, reads by the cpu to the buffer are guaranteed to see
311 * whatever the device wrote there.
313 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
314 size_t size, enum dma_data_direction dir)
316 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
318 BUG_ON(dir == DMA_NONE);
320 /* NOTE: We use dev_addr here, not paddr! */
321 if (is_xen_swiotlb_buffer(dev_addr)) {
322 swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
323 return;
326 if (dir != DMA_FROM_DEVICE)
327 return;
330 * phys_to_virt doesn't work with hihgmem page but we could
331 * call dma_mark_clean() with hihgmem page here. However, we
332 * are fine since dma_mark_clean() is null on POWERPC. We can
333 * make dma_mark_clean() take a physical address if necessary.
335 dma_mark_clean(phys_to_virt(paddr), size);
338 void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
339 size_t size, enum dma_data_direction dir,
340 struct dma_attrs *attrs)
342 xen_unmap_single(hwdev, dev_addr, size, dir);
344 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
347 * Make physical memory consistent for a single streaming mode DMA translation
348 * after a transfer.
350 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
351 * using the cpu, yet do not wish to teardown the dma mapping, you must
352 * call this function before doing so. At the next point you give the dma
353 * address back to the card, you must first perform a
354 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
356 static void
357 xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
358 size_t size, enum dma_data_direction dir,
359 enum dma_sync_target target)
361 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
363 BUG_ON(dir == DMA_NONE);
365 /* NOTE: We use dev_addr here, not paddr! */
366 if (is_xen_swiotlb_buffer(dev_addr)) {
367 swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
368 target);
369 return;
372 if (dir != DMA_FROM_DEVICE)
373 return;
375 dma_mark_clean(phys_to_virt(paddr), size);
378 void
379 xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
380 size_t size, enum dma_data_direction dir)
382 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
384 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
386 void
387 xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
388 size_t size, enum dma_data_direction dir)
390 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
392 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
395 * Map a set of buffers described by scatterlist in streaming mode for DMA.
396 * This is the scatter-gather version of the above xen_swiotlb_map_page
397 * interface. Here the scatter gather list elements are each tagged with the
398 * appropriate dma address and length. They are obtained via
399 * sg_dma_{address,length}(SG).
401 * NOTE: An implementation may be able to use a smaller number of
402 * DMA address/length pairs than there are SG table elements.
403 * (for example via virtual mapping capabilities)
404 * The routine returns the number of addr/length pairs actually
405 * used, at most nents.
407 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
408 * same here.
411 xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
412 int nelems, enum dma_data_direction dir,
413 struct dma_attrs *attrs)
415 struct scatterlist *sg;
416 int i;
418 BUG_ON(dir == DMA_NONE);
420 for_each_sg(sgl, sg, nelems, i) {
421 phys_addr_t paddr = sg_phys(sg);
422 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
424 if (swiotlb_force ||
425 !dma_capable(hwdev, dev_addr, sg->length) ||
426 range_straddles_page_boundary(paddr, sg->length)) {
427 void *map = swiotlb_tbl_map_single(hwdev,
428 start_dma_addr,
429 sg_phys(sg),
430 sg->length, dir);
431 if (!map) {
432 /* Don't panic here, we expect map_sg users
433 to do proper error handling. */
434 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
435 attrs);
436 sgl[0].dma_length = 0;
437 return DMA_ERROR_CODE;
439 sg->dma_address = xen_virt_to_bus(map);
440 } else
441 sg->dma_address = dev_addr;
442 sg->dma_length = sg->length;
444 return nelems;
446 EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
449 xen_swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
450 enum dma_data_direction dir)
452 return xen_swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
454 EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg);
457 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
458 * concerning calls here are the same as for swiotlb_unmap_page() above.
460 void
461 xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
462 int nelems, enum dma_data_direction dir,
463 struct dma_attrs *attrs)
465 struct scatterlist *sg;
466 int i;
468 BUG_ON(dir == DMA_NONE);
470 for_each_sg(sgl, sg, nelems, i)
471 xen_unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
474 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
476 void
477 xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
478 enum dma_data_direction dir)
480 return xen_swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
482 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg);
485 * Make physical memory consistent for a set of streaming mode DMA translations
486 * after a transfer.
488 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
489 * and usage.
491 static void
492 xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
493 int nelems, enum dma_data_direction dir,
494 enum dma_sync_target target)
496 struct scatterlist *sg;
497 int i;
499 for_each_sg(sgl, sg, nelems, i)
500 xen_swiotlb_sync_single(hwdev, sg->dma_address,
501 sg->dma_length, dir, target);
504 void
505 xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
506 int nelems, enum dma_data_direction dir)
508 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
510 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
512 void
513 xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
514 int nelems, enum dma_data_direction dir)
516 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
518 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
521 xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
523 return !dma_addr;
525 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
528 * Return whether the given device DMA address mask can be supported
529 * properly. For example, if your device can only drive the low 24-bits
530 * during bus mastering, then you would pass 0x00ffffff as the mask to
531 * this function.
534 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
536 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
538 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);