Linux Audio

Check our new training course

Loading...
v3.1
 
  1/*
  2 *  Copyright 2010
  3 *  by Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
  4 *
  5 * This code provides a IOMMU for Xen PV guests with PCI passthrough.
  6 *
  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
 10 *
 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.
 15 *
 16 * PV guests under Xen are running in an non-contiguous memory architecture.
 17 *
 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).
 22 *
 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.
 33 *
 34 */
 35
 36#include <linux/bootmem.h>
 37#include <linux/dma-mapping.h>
 
 
 
 
 38#include <xen/swiotlb-xen.h>
 39#include <xen/page.h>
 40#include <xen/xen-ops.h>
 41/*
 42 * Used to do a quick range check in swiotlb_tbl_unmap_single and
 43 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
 44 * API.
 45 */
 
 46
 47static char *xen_io_tlb_start, *xen_io_tlb_end;
 48static unsigned long xen_io_tlb_nslabs;
 49/*
 50 * Quick lookup value of the bus address of the IOTLB.
 51 */
 52
 53u64 start_dma_addr;
 54
 55static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
 56{
 57	return phys_to_machine(XPADDR(paddr)).maddr;
 58}
 59
 60static phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
 61{
 62	return machine_to_phys(XMADDR(baddr)).paddr;
 63}
 64
 65static dma_addr_t xen_virt_to_bus(void *address)
 66{
 67	return xen_phys_to_bus(virt_to_phys(address));
 68}
 69
 70static int check_pages_physically_contiguous(unsigned long pfn,
 71					     unsigned int offset,
 72					     size_t length)
 73{
 74	unsigned long next_mfn;
 75	int i;
 76	int nr_pages;
 77
 78	next_mfn = pfn_to_mfn(pfn);
 79	nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
 80
 81	for (i = 1; i < nr_pages; i++) {
 82		if (pfn_to_mfn(++pfn) != ++next_mfn)
 83			return 0;
 84	}
 85	return 1;
 86}
 87
 88static int range_straddles_page_boundary(phys_addr_t p, size_t size)
 89{
 90	unsigned long pfn = PFN_DOWN(p);
 91	unsigned int offset = p & ~PAGE_MASK;
 
 
 92
 93	if (offset + size <= PAGE_SIZE)
 94		return 0;
 95	if (check_pages_physically_contiguous(pfn, offset, size))
 96		return 0;
 97	return 1;
 98}
 99
100static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
101{
102	unsigned long mfn = PFN_DOWN(dma_addr);
103	unsigned long pfn = mfn_to_local_pfn(mfn);
104	phys_addr_t paddr;
105
106	/* If the address is outside our domain, it CAN
107	 * have the same virtual address as another address
108	 * in our domain. Therefore _only_ check address within our domain.
109	 */
110	if (pfn_valid(pfn)) {
111		paddr = PFN_PHYS(pfn);
112		return paddr >= virt_to_phys(xen_io_tlb_start) &&
113		       paddr < virt_to_phys(xen_io_tlb_end);
114	}
115	return 0;
116}
117
118static int max_dma_bits = 32;
119
120static int
121xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
122{
123	int i, rc;
124	int dma_bits;
 
 
 
125
126	dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
 
127
128	i = 0;
129	do {
130		int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
131
132		do {
133			rc = xen_create_contiguous_region(
134				(unsigned long)buf + (i << IO_TLB_SHIFT),
135				get_order(slabs << IO_TLB_SHIFT),
136				dma_bits);
137		} while (rc && dma_bits++ < max_dma_bits);
138		if (rc)
139			return rc;
140
141		i += slabs;
142	} while (i < nslabs);
143	return 0;
144}
145
146void __init xen_swiotlb_init(int verbose)
147{
148	unsigned long bytes;
149	int rc;
150	unsigned long nr_tbl;
151
152	nr_tbl = swioltb_nr_tbl();
153	if (nr_tbl)
154		xen_io_tlb_nslabs = nr_tbl;
155	else {
156		xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
157		xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
158	}
159
160	bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
161
162	/*
163	 * Get IO TLB memory from any location.
164	 */
165	xen_io_tlb_start = alloc_bootmem(bytes);
166	if (!xen_io_tlb_start)
167		panic("Cannot allocate SWIOTLB buffer");
168
169	xen_io_tlb_end = xen_io_tlb_start + bytes;
170	/*
171	 * And replace that memory with pages under 4GB.
172	 */
173	rc = xen_swiotlb_fixup(xen_io_tlb_start,
174			       bytes,
175			       xen_io_tlb_nslabs);
176	if (rc)
177		goto error;
178
179	start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
180	swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs, verbose);
181
182	return;
183error:
184	panic("DMA(%d): Failed to exchange pages allocated for DMA with Xen! "\
185	      "We either don't have the permission or you do not have enough"\
186	      "free memory under 4GB!\n", rc);
187}
188
189void *
190xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
191			   dma_addr_t *dma_handle, gfp_t flags)
192{
193	void *ret;
194	int order = get_order(size);
195	u64 dma_mask = DMA_BIT_MASK(32);
196	unsigned long vstart;
197
198	/*
199	* Ignore region specifiers - the kernel's ideas of
200	* pseudo-phys memory layout has nothing to do with the
201	* machine physical layout.  We can't allocate highmem
202	* because we can't return a pointer to it.
203	*/
204	flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
205
206	if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
 
207		return ret;
 
208
209	vstart = __get_free_pages(flags, order);
210	ret = (void *)vstart;
211
212	if (hwdev && hwdev->coherent_dma_mask)
213		dma_mask = dma_alloc_coherent_mask(hwdev, flags);
214
215	if (ret) {
216		if (xen_create_contiguous_region(vstart, order,
217						 fls64(dma_mask)) != 0) {
218			free_pages(vstart, order);
219			return NULL;
220		}
221		memset(ret, 0, size);
222		*dma_handle = virt_to_machine(ret).maddr;
223	}
 
 
224	return ret;
 
 
 
 
225}
226EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
227
228void
229xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
230			  dma_addr_t dev_addr)
231{
 
232	int order = get_order(size);
233
234	if (dma_release_from_coherent(hwdev, order, vaddr))
235		return;
236
237	xen_destroy_contiguous_region((unsigned long)vaddr, order);
238	free_pages((unsigned long)vaddr, order);
 
 
 
 
 
239}
240EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
241
242
243/*
244 * Map a single buffer of the indicated size for DMA in streaming mode.  The
245 * physical address to use is returned.
246 *
247 * Once the device is given the dma address, the device owns this memory until
248 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
249 */
250dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
251				unsigned long offset, size_t size,
252				enum dma_data_direction dir,
253				struct dma_attrs *attrs)
254{
255	phys_addr_t phys = page_to_phys(page) + offset;
256	dma_addr_t dev_addr = xen_phys_to_bus(phys);
257	void *map;
258
259	BUG_ON(dir == DMA_NONE);
260	/*
261	 * If the address happens to be in the device's DMA window,
262	 * we can safely return the device addr and not worry about bounce
263	 * buffering it.
264	 */
265	if (dma_capable(dev, dev_addr, size) &&
266	    !range_straddles_page_boundary(phys, size) && !swiotlb_force)
267		return dev_addr;
 
 
268
269	/*
270	 * Oh well, have to allocate and map a bounce buffer.
271	 */
272	map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
273	if (!map)
274		return DMA_ERROR_CODE;
 
 
275
276	dev_addr = xen_virt_to_bus(map);
 
277
278	/*
279	 * Ensure that the address returned is DMA'ble
280	 */
281	if (!dma_capable(dev, dev_addr, size))
282		panic("map_single: bounce buffer is not DMA'ble");
 
 
 
283
 
 
 
 
 
 
 
284	return dev_addr;
285}
286EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
287
288/*
289 * Unmap a single streaming mode DMA translation.  The dma_addr and size must
290 * match what was provided for in a previous xen_swiotlb_map_page call.  All
291 * other usages are undefined.
292 *
293 * After this call, reads by the cpu to the buffer are guaranteed to see
294 * whatever the device wrote there.
295 */
296static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
297			     size_t size, enum dma_data_direction dir)
298{
299	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
300
301	BUG_ON(dir == DMA_NONE);
302
303	/* NOTE: We use dev_addr here, not paddr! */
304	if (is_xen_swiotlb_buffer(dev_addr)) {
305		swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
306		return;
 
307	}
308
309	if (dir != DMA_FROM_DEVICE)
310		return;
311
312	/*
313	 * phys_to_virt doesn't work with hihgmem page but we could
314	 * call dma_mark_clean() with hihgmem page here. However, we
315	 * are fine since dma_mark_clean() is null on POWERPC. We can
316	 * make dma_mark_clean() take a physical address if necessary.
317	 */
318	dma_mark_clean(phys_to_virt(paddr), size);
319}
320
321void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
322			    size_t size, enum dma_data_direction dir,
323			    struct dma_attrs *attrs)
324{
325	xen_unmap_single(hwdev, dev_addr, size, dir);
326}
327EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
328
329/*
330 * Make physical memory consistent for a single streaming mode DMA translation
331 * after a transfer.
332 *
333 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
334 * using the cpu, yet do not wish to teardown the dma mapping, you must
335 * call this function before doing so.  At the next point you give the dma
336 * address back to the card, you must first perform a
337 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
338 */
339static void
340xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
341			size_t size, enum dma_data_direction dir,
342			enum dma_sync_target target)
343{
344	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
345
346	BUG_ON(dir == DMA_NONE);
347
348	/* NOTE: We use dev_addr here, not paddr! */
349	if (is_xen_swiotlb_buffer(dev_addr)) {
350		swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
351				       target);
352		return;
353	}
354
355	if (dir != DMA_FROM_DEVICE)
356		return;
357
358	dma_mark_clean(phys_to_virt(paddr), size);
359}
360
361void
362xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
363				size_t size, enum dma_data_direction dir)
364{
365	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
366}
367EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
368
369void
370xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
371				   size_t size, enum dma_data_direction dir)
372{
373	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
 
 
 
 
374}
375EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
376
377/*
378 * Map a set of buffers described by scatterlist in streaming mode for DMA.
379 * This is the scatter-gather version of the above xen_swiotlb_map_page
380 * interface.  Here the scatter gather list elements are each tagged with the
381 * appropriate dma address and length.  They are obtained via
382 * sg_dma_{address,length}(SG).
383 *
384 * NOTE: An implementation may be able to use a smaller number of
385 *       DMA address/length pairs than there are SG table elements.
386 *       (for example via virtual mapping capabilities)
387 *       The routine returns the number of addr/length pairs actually
388 *       used, at most nents.
389 *
390 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
391 * same here.
392 */
393int
394xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
395			 int nelems, enum dma_data_direction dir,
396			 struct dma_attrs *attrs)
397{
398	struct scatterlist *sg;
399	int i;
400
401	BUG_ON(dir == DMA_NONE);
402
403	for_each_sg(sgl, sg, nelems, i) {
404		phys_addr_t paddr = sg_phys(sg);
405		dma_addr_t dev_addr = xen_phys_to_bus(paddr);
406
407		if (swiotlb_force ||
408		    !dma_capable(hwdev, dev_addr, sg->length) ||
409		    range_straddles_page_boundary(paddr, sg->length)) {
410			void *map = swiotlb_tbl_map_single(hwdev,
411							   start_dma_addr,
412							   sg_phys(sg),
413							   sg->length, dir);
414			if (!map) {
415				/* Don't panic here, we expect map_sg users
416				   to do proper error handling. */
417				xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
418							   attrs);
419				sgl[0].dma_length = 0;
420				return DMA_ERROR_CODE;
421			}
422			sg->dma_address = xen_virt_to_bus(map);
423		} else
424			sg->dma_address = dev_addr;
425		sg->dma_length = sg->length;
426	}
427	return nelems;
428}
429EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
430
431int
432xen_swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
433		   enum dma_data_direction dir)
434{
435	return xen_swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
436}
437EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg);
438
439/*
440 * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
441 * concerning calls here are the same as for swiotlb_unmap_page() above.
442 */
443void
444xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
445			   int nelems, enum dma_data_direction dir,
446			   struct dma_attrs *attrs)
447{
448	struct scatterlist *sg;
449	int i;
450
451	BUG_ON(dir == DMA_NONE);
452
453	for_each_sg(sgl, sg, nelems, i)
454		xen_unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
455
456}
457EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
 
 
458
459void
460xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
461		     enum dma_data_direction dir)
462{
463	return xen_swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
464}
465EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg);
466
467/*
468 * Make physical memory consistent for a set of streaming mode DMA translations
469 * after a transfer.
470 *
471 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
472 * and usage.
473 */
474static void
475xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
476		    int nelems, enum dma_data_direction dir,
477		    enum dma_sync_target target)
478{
479	struct scatterlist *sg;
480	int i;
481
482	for_each_sg(sgl, sg, nelems, i)
483		xen_swiotlb_sync_single(hwdev, sg->dma_address,
484					sg->dma_length, dir, target);
485}
486
487void
488xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
489			    int nelems, enum dma_data_direction dir)
490{
491	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
492}
493EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
494
495void
496xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
497			       int nelems, enum dma_data_direction dir)
498{
499	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
500}
501EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
502
503int
504xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
505{
506	return !dma_addr;
507}
508EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
509
510/*
511 * Return whether the given device DMA address mask can be supported
512 * properly.  For example, if your device can only drive the low 24-bits
513 * during bus mastering, then you would pass 0x00ffffff as the mask to
514 * this function.
515 */
516int
517xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
518{
519	return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
520}
521EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
v6.8
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 *  Copyright 2010
  4 *  by Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
  5 *
  6 * This code provides a IOMMU for Xen PV guests with PCI passthrough.
  7 *
 
 
 
 
 
 
 
 
 
  8 * PV guests under Xen are running in an non-contiguous memory architecture.
  9 *
 10 * When PCI pass-through is utilized, this necessitates an IOMMU for
 11 * translating bus (DMA) to virtual and vice-versa and also providing a
 12 * mechanism to have contiguous pages for device drivers operations (say DMA
 13 * operations).
 14 *
 15 * Specifically, under Xen the Linux idea of pages is an illusion. It
 16 * assumes that pages start at zero and go up to the available memory. To
 17 * help with that, the Linux Xen MMU provides a lookup mechanism to
 18 * translate the page frame numbers (PFN) to machine frame numbers (MFN)
 19 * and vice-versa. The MFN are the "real" frame numbers. Furthermore
 20 * memory is not contiguous. Xen hypervisor stitches memory for guests
 21 * from different pools, which means there is no guarantee that PFN==MFN
 22 * and PFN+1==MFN+1. Lastly with Xen 4.0, pages (in debug mode) are
 23 * allocated in descending order (high to low), meaning the guest might
 24 * never get any MFN's under the 4GB mark.
 
 25 */
 26
 27#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
 28
 29#include <linux/memblock.h>
 30#include <linux/dma-direct.h>
 31#include <linux/dma-map-ops.h>
 32#include <linux/export.h>
 33#include <xen/swiotlb-xen.h>
 34#include <xen/page.h>
 35#include <xen/xen-ops.h>
 36#include <xen/hvc-console.h>
 37
 38#include <asm/dma-mapping.h>
 39
 40#include <trace/events/swiotlb.h>
 41#define MAX_DMA_BITS 32
 42
 
 
 43/*
 44 * Quick lookup value of the bus address of the IOTLB.
 45 */
 46
 47static inline phys_addr_t xen_phys_to_bus(struct device *dev, phys_addr_t paddr)
 
 
 48{
 49	unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
 50	phys_addr_t baddr = (phys_addr_t)bfn << XEN_PAGE_SHIFT;
 51
 52	baddr |= paddr & ~XEN_PAGE_MASK;
 53	return baddr;
 
 54}
 55
 56static inline dma_addr_t xen_phys_to_dma(struct device *dev, phys_addr_t paddr)
 57{
 58	return phys_to_dma(dev, xen_phys_to_bus(dev, paddr));
 59}
 60
 61static inline phys_addr_t xen_bus_to_phys(struct device *dev,
 62					  phys_addr_t baddr)
 
 63{
 64	unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
 65	phys_addr_t paddr = (xen_pfn << XEN_PAGE_SHIFT) |
 66			    (baddr & ~XEN_PAGE_MASK);
 67
 68	return paddr;
 69}
 70
 71static inline phys_addr_t xen_dma_to_phys(struct device *dev,
 72					  dma_addr_t dma_addr)
 73{
 74	return xen_bus_to_phys(dev, dma_to_phys(dev, dma_addr));
 
 75}
 76
 77static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
 78{
 79	unsigned long next_bfn, xen_pfn = XEN_PFN_DOWN(p);
 80	unsigned int i, nr_pages = XEN_PFN_UP(xen_offset_in_page(p) + size);
 81
 82	next_bfn = pfn_to_bfn(xen_pfn);
 83
 84	for (i = 1; i < nr_pages; i++)
 85		if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
 86			return 1;
 87
 88	return 0;
 89}
 90
 91static int is_xen_swiotlb_buffer(struct device *dev, dma_addr_t dma_addr)
 92{
 93	unsigned long bfn = XEN_PFN_DOWN(dma_to_phys(dev, dma_addr));
 94	unsigned long xen_pfn = bfn_to_local_pfn(bfn);
 95	phys_addr_t paddr = (phys_addr_t)xen_pfn << XEN_PAGE_SHIFT;
 96
 97	/* If the address is outside our domain, it CAN
 98	 * have the same virtual address as another address
 99	 * in our domain. Therefore _only_ check address within our domain.
100	 */
101	if (pfn_valid(PFN_DOWN(paddr)))
102		return is_swiotlb_buffer(dev, paddr);
 
 
 
103	return 0;
104}
105
106#ifdef CONFIG_X86
107int xen_swiotlb_fixup(void *buf, unsigned long nslabs)
 
 
108{
109	int rc;
110	unsigned int order = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT);
111	unsigned int i, dma_bits = order + PAGE_SHIFT;
112	dma_addr_t dma_handle;
113	phys_addr_t p = virt_to_phys(buf);
114
115	BUILD_BUG_ON(IO_TLB_SEGSIZE & (IO_TLB_SEGSIZE - 1));
116	BUG_ON(nslabs % IO_TLB_SEGSIZE);
117
118	i = 0;
119	do {
 
 
120		do {
121			rc = xen_create_contiguous_region(
122				p + (i << IO_TLB_SHIFT), order,
123				dma_bits, &dma_handle);
124		} while (rc && dma_bits++ < MAX_DMA_BITS);
 
125		if (rc)
126			return rc;
127
128		i += IO_TLB_SEGSIZE;
129	} while (i < nslabs);
130	return 0;
131}
132
133static void *
134xen_swiotlb_alloc_coherent(struct device *dev, size_t size,
135		dma_addr_t *dma_handle, gfp_t flags, unsigned long attrs)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
136{
137	u64 dma_mask = dev->coherent_dma_mask;
138	int order = get_order(size);
139	phys_addr_t phys;
140	void *ret;
141
142	/* Align the allocation to the Xen page size */
143	size = 1UL << (order + XEN_PAGE_SHIFT);
 
 
 
 
 
144
145	ret = (void *)__get_free_pages(flags, get_order(size));
146	if (!ret)
147		return ret;
148	phys = virt_to_phys(ret);
149
150	*dma_handle = xen_phys_to_dma(dev, phys);
151	if (*dma_handle + size - 1 > dma_mask ||
152	    range_straddles_page_boundary(phys, size)) {
153		if (xen_create_contiguous_region(phys, order, fls64(dma_mask),
154				dma_handle) != 0)
155			goto out_free_pages;
156		SetPageXenRemapped(virt_to_page(ret));
 
 
 
 
 
 
 
157	}
158
159	memset(ret, 0, size);
160	return ret;
161
162out_free_pages:
163	free_pages((unsigned long)ret, get_order(size));
164	return NULL;
165}
 
166
167static void
168xen_swiotlb_free_coherent(struct device *dev, size_t size, void *vaddr,
169		dma_addr_t dma_handle, unsigned long attrs)
170{
171	phys_addr_t phys = virt_to_phys(vaddr);
172	int order = get_order(size);
173
174	/* Convert the size to actually allocated. */
175	size = 1UL << (order + XEN_PAGE_SHIFT);
176
177	if (WARN_ON_ONCE(dma_handle + size - 1 > dev->coherent_dma_mask) ||
178	    WARN_ON_ONCE(range_straddles_page_boundary(phys, size)))
179	    	return;
180
181	if (TestClearPageXenRemapped(virt_to_page(vaddr)))
182		xen_destroy_contiguous_region(phys, order);
183	free_pages((unsigned long)vaddr, get_order(size));
184}
185#endif /* CONFIG_X86 */
 
186
187/*
188 * Map a single buffer of the indicated size for DMA in streaming mode.  The
189 * physical address to use is returned.
190 *
191 * Once the device is given the dma address, the device owns this memory until
192 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
193 */
194static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
195				unsigned long offset, size_t size,
196				enum dma_data_direction dir,
197				unsigned long attrs)
198{
199	phys_addr_t map, phys = page_to_phys(page) + offset;
200	dma_addr_t dev_addr = xen_phys_to_dma(dev, phys);
 
201
202	BUG_ON(dir == DMA_NONE);
203	/*
204	 * If the address happens to be in the device's DMA window,
205	 * we can safely return the device addr and not worry about bounce
206	 * buffering it.
207	 */
208	if (dma_capable(dev, dev_addr, size, true) &&
209	    !range_straddles_page_boundary(phys, size) &&
210		!xen_arch_need_swiotlb(dev, phys, dev_addr) &&
211		!is_swiotlb_force_bounce(dev))
212		goto done;
213
214	/*
215	 * Oh well, have to allocate and map a bounce buffer.
216	 */
217	trace_swiotlb_bounced(dev, dev_addr, size);
218
219	map = swiotlb_tbl_map_single(dev, phys, size, size, 0, dir, attrs);
220	if (map == (phys_addr_t)DMA_MAPPING_ERROR)
221		return DMA_MAPPING_ERROR;
222
223	phys = map;
224	dev_addr = xen_phys_to_dma(dev, map);
225
226	/*
227	 * Ensure that the address returned is DMA'ble
228	 */
229	if (unlikely(!dma_capable(dev, dev_addr, size, true))) {
230		swiotlb_tbl_unmap_single(dev, map, size, dir,
231				attrs | DMA_ATTR_SKIP_CPU_SYNC);
232		return DMA_MAPPING_ERROR;
233	}
234
235done:
236	if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
237		if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dev_addr))))
238			arch_sync_dma_for_device(phys, size, dir);
239		else
240			xen_dma_sync_for_device(dev, dev_addr, size, dir);
241	}
242	return dev_addr;
243}
 
244
245/*
246 * Unmap a single streaming mode DMA translation.  The dma_addr and size must
247 * match what was provided for in a previous xen_swiotlb_map_page call.  All
248 * other usages are undefined.
249 *
250 * After this call, reads by the cpu to the buffer are guaranteed to see
251 * whatever the device wrote there.
252 */
253static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
254		size_t size, enum dma_data_direction dir, unsigned long attrs)
255{
256	phys_addr_t paddr = xen_dma_to_phys(hwdev, dev_addr);
257
258	BUG_ON(dir == DMA_NONE);
259
260	if (!dev_is_dma_coherent(hwdev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
261		if (pfn_valid(PFN_DOWN(dma_to_phys(hwdev, dev_addr))))
262			arch_sync_dma_for_cpu(paddr, size, dir);
263		else
264			xen_dma_sync_for_cpu(hwdev, dev_addr, size, dir);
265	}
266
267	/* NOTE: We use dev_addr here, not paddr! */
268	if (is_xen_swiotlb_buffer(hwdev, dev_addr))
269		swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
270}
 
271
 
 
 
 
 
 
 
 
 
 
272static void
273xen_swiotlb_sync_single_for_cpu(struct device *dev, dma_addr_t dma_addr,
274		size_t size, enum dma_data_direction dir)
 
275{
276	phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
 
 
277
278	if (!dev_is_dma_coherent(dev)) {
279		if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
280			arch_sync_dma_for_cpu(paddr, size, dir);
281		else
282			xen_dma_sync_for_cpu(dev, dma_addr, size, dir);
283	}
284
285	if (is_xen_swiotlb_buffer(dev, dma_addr))
286		swiotlb_sync_single_for_cpu(dev, paddr, size, dir);
 
 
287}
288
289static void
290xen_swiotlb_sync_single_for_device(struct device *dev, dma_addr_t dma_addr,
291		size_t size, enum dma_data_direction dir)
292{
293	phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
 
 
294
295	if (is_xen_swiotlb_buffer(dev, dma_addr))
296		swiotlb_sync_single_for_device(dev, paddr, size, dir);
297
298	if (!dev_is_dma_coherent(dev)) {
299		if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
300			arch_sync_dma_for_device(paddr, size, dir);
301		else
302			xen_dma_sync_for_device(dev, dma_addr, size, dir);
303	}
304}
 
305
306/*
307 * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
308 * concerning calls here are the same as for swiotlb_unmap_page() above.
 
 
 
 
 
 
 
 
 
 
 
 
309 */
310static void
311xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
312		enum dma_data_direction dir, unsigned long attrs)
 
313{
314	struct scatterlist *sg;
315	int i;
316
317	BUG_ON(dir == DMA_NONE);
318
319	for_each_sg(sgl, sg, nelems, i)
320		xen_swiotlb_unmap_page(hwdev, sg->dma_address, sg_dma_len(sg),
321				dir, attrs);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
322
 
 
 
 
 
323}
 
324
325static int
326xen_swiotlb_map_sg(struct device *dev, struct scatterlist *sgl, int nelems,
327		enum dma_data_direction dir, unsigned long attrs)
 
 
 
 
 
328{
329	struct scatterlist *sg;
330	int i;
331
332	BUG_ON(dir == DMA_NONE);
333
334	for_each_sg(sgl, sg, nelems, i) {
335		sg->dma_address = xen_swiotlb_map_page(dev, sg_page(sg),
336				sg->offset, sg->length, dir, attrs);
337		if (sg->dma_address == DMA_MAPPING_ERROR)
338			goto out_unmap;
339		sg_dma_len(sg) = sg->length;
340	}
341
342	return nelems;
343out_unmap:
344	xen_swiotlb_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
345	sg_dma_len(sgl) = 0;
346	return -EIO;
347}
 
348
 
 
 
 
 
 
 
349static void
350xen_swiotlb_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl,
351			    int nelems, enum dma_data_direction dir)
 
352{
353	struct scatterlist *sg;
354	int i;
355
356	for_each_sg(sgl, sg, nelems, i) {
357		xen_swiotlb_sync_single_for_cpu(dev, sg->dma_address,
358				sg->length, dir);
359	}
 
 
 
 
 
 
360}
 
361
362static void
363xen_swiotlb_sync_sg_for_device(struct device *dev, struct scatterlist *sgl,
364			       int nelems, enum dma_data_direction dir)
365{
366	struct scatterlist *sg;
367	int i;
 
368
369	for_each_sg(sgl, sg, nelems, i) {
370		xen_swiotlb_sync_single_for_device(dev, sg->dma_address,
371				sg->length, dir);
372	}
373}
 
374
375/*
376 * Return whether the given device DMA address mask can be supported
377 * properly.  For example, if your device can only drive the low 24-bits
378 * during bus mastering, then you would pass 0x00ffffff as the mask to
379 * this function.
380 */
381static int
382xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
383{
384	return xen_phys_to_dma(hwdev, default_swiotlb_limit()) <= mask;
385}
386
387const struct dma_map_ops xen_swiotlb_dma_ops = {
388#ifdef CONFIG_X86
389	.alloc = xen_swiotlb_alloc_coherent,
390	.free = xen_swiotlb_free_coherent,
391#else
392	.alloc = dma_direct_alloc,
393	.free = dma_direct_free,
394#endif
395	.sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
396	.sync_single_for_device = xen_swiotlb_sync_single_for_device,
397	.sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
398	.sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
399	.map_sg = xen_swiotlb_map_sg,
400	.unmap_sg = xen_swiotlb_unmap_sg,
401	.map_page = xen_swiotlb_map_page,
402	.unmap_page = xen_swiotlb_unmap_page,
403	.dma_supported = xen_swiotlb_dma_supported,
404	.mmap = dma_common_mmap,
405	.get_sgtable = dma_common_get_sgtable,
406	.alloc_pages = dma_common_alloc_pages,
407	.free_pages = dma_common_free_pages,
408	.max_mapping_size = swiotlb_max_mapping_size,
409};