Linux Audio

Check our new training course

Loading...
v6.2
  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, io_tlb_default_mem.end - 1) <= 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};
v3.15
 
  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#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
 37
 38#include <linux/bootmem.h>
 39#include <linux/dma-mapping.h>
 
 40#include <linux/export.h>
 41#include <xen/swiotlb-xen.h>
 42#include <xen/page.h>
 43#include <xen/xen-ops.h>
 44#include <xen/hvc-console.h>
 45
 46#include <asm/dma-mapping.h>
 47#include <asm/xen/page-coherent.h>
 48
 49#include <trace/events/swiotlb.h>
 
 
 50/*
 51 * Used to do a quick range check in swiotlb_tbl_unmap_single and
 52 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
 53 * API.
 54 */
 55
 56#ifndef CONFIG_X86
 57static unsigned long dma_alloc_coherent_mask(struct device *dev,
 58					    gfp_t gfp)
 59{
 60	unsigned long dma_mask = 0;
 
 61
 62	dma_mask = dev->coherent_dma_mask;
 63	if (!dma_mask)
 64		dma_mask = (gfp & GFP_DMA) ? DMA_BIT_MASK(24) : DMA_BIT_MASK(32);
 65
 66	return dma_mask;
 67}
 68#endif
 69
 70static char *xen_io_tlb_start, *xen_io_tlb_end;
 71static unsigned long xen_io_tlb_nslabs;
 72/*
 73 * Quick lookup value of the bus address of the IOTLB.
 74 */
 75
 76static u64 start_dma_addr;
 77
 78/*
 79 * Both of these functions should avoid PFN_PHYS because phys_addr_t
 80 * can be 32bit when dma_addr_t is 64bit leading to a loss in
 81 * information if the shift is done before casting to 64bit.
 82 */
 83static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
 84{
 85	unsigned long mfn = pfn_to_mfn(PFN_DOWN(paddr));
 86	dma_addr_t dma = (dma_addr_t)mfn << PAGE_SHIFT;
 87
 88	dma |= paddr & ~PAGE_MASK;
 89
 90	return dma;
 91}
 92
 93static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
 
 94{
 95	unsigned long pfn = mfn_to_pfn(PFN_DOWN(baddr));
 96	dma_addr_t dma = (dma_addr_t)pfn << PAGE_SHIFT;
 97	phys_addr_t paddr = dma;
 98
 99	BUG_ON(paddr != dma); /* truncation has occurred, should never happen */
100
101	paddr |= baddr & ~PAGE_MASK;
102
103	return paddr;
104}
105
106static inline dma_addr_t xen_virt_to_bus(void *address)
 
107{
108	return xen_phys_to_bus(virt_to_phys(address));
109}
110
111static int check_pages_physically_contiguous(unsigned long pfn,
112					     unsigned int offset,
113					     size_t length)
114{
115	unsigned long next_mfn;
116	int i;
117	int nr_pages;
118
119	next_mfn = pfn_to_mfn(pfn);
120	nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
121
122	for (i = 1; i < nr_pages; i++) {
123		if (pfn_to_mfn(++pfn) != ++next_mfn)
124			return 0;
125	}
126	return 1;
127}
128
129static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
130{
131	unsigned long pfn = PFN_DOWN(p);
132	unsigned int offset = p & ~PAGE_MASK;
133
134	if (offset + size <= PAGE_SIZE)
135		return 0;
136	if (check_pages_physically_contiguous(pfn, offset, size))
137		return 0;
138	return 1;
139}
140
141static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
142{
143	unsigned long mfn = PFN_DOWN(dma_addr);
144	unsigned long pfn = mfn_to_local_pfn(mfn);
145	phys_addr_t paddr;
146
147	/* If the address is outside our domain, it CAN
148	 * have the same virtual address as another address
149	 * in our domain. Therefore _only_ check address within our domain.
150	 */
151	if (pfn_valid(pfn)) {
152		paddr = PFN_PHYS(pfn);
153		return paddr >= virt_to_phys(xen_io_tlb_start) &&
154		       paddr < virt_to_phys(xen_io_tlb_end);
155	}
156	return 0;
157}
158
159static int max_dma_bits = 32;
160
161static int
162xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
163{
164	int i, rc;
165	int dma_bits;
 
166	dma_addr_t dma_handle;
167	phys_addr_t p = virt_to_phys(buf);
168
169	dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
 
170
171	i = 0;
172	do {
173		int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
174
175		do {
176			rc = xen_create_contiguous_region(
177				p + (i << IO_TLB_SHIFT),
178				get_order(slabs << IO_TLB_SHIFT),
179				dma_bits, &dma_handle);
180		} while (rc && dma_bits++ < max_dma_bits);
181		if (rc)
182			return rc;
183
184		i += slabs;
185	} while (i < nslabs);
186	return 0;
187}
188static unsigned long xen_set_nslabs(unsigned long nr_tbl)
189{
190	if (!nr_tbl) {
191		xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
192		xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
193	} else
194		xen_io_tlb_nslabs = nr_tbl;
195
196	return xen_io_tlb_nslabs << IO_TLB_SHIFT;
197}
198
199enum xen_swiotlb_err {
200	XEN_SWIOTLB_UNKNOWN = 0,
201	XEN_SWIOTLB_ENOMEM,
202	XEN_SWIOTLB_EFIXUP
203};
204
205static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
 
 
206{
207	switch (err) {
208	case XEN_SWIOTLB_ENOMEM:
209		return "Cannot allocate Xen-SWIOTLB buffer\n";
210	case XEN_SWIOTLB_EFIXUP:
211		return "Failed to get contiguous memory for DMA from Xen!\n"\
212		    "You either: don't have the permissions, do not have"\
213		    " enough free memory under 4GB, or the hypervisor memory"\
214		    " is too fragmented!";
215	default:
216		break;
217	}
218	return "";
219}
220int __ref xen_swiotlb_init(int verbose, bool early)
221{
222	unsigned long bytes, order;
223	int rc = -ENOMEM;
224	enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
225	unsigned int repeat = 3;
226
227	xen_io_tlb_nslabs = swiotlb_nr_tbl();
228retry:
229	bytes = xen_set_nslabs(xen_io_tlb_nslabs);
230	order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
231	/*
232	 * Get IO TLB memory from any location.
233	 */
234	if (early)
235		xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes));
236	else {
237#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
238#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
239		while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
240			xen_io_tlb_start = (void *)__get_free_pages(__GFP_NOWARN, order);
241			if (xen_io_tlb_start)
242				break;
243			order--;
244		}
245		if (order != get_order(bytes)) {
246			pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
247				(PAGE_SIZE << order) >> 20);
248			xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
249			bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
250		}
251	}
252	if (!xen_io_tlb_start) {
253		m_ret = XEN_SWIOTLB_ENOMEM;
254		goto error;
255	}
256	xen_io_tlb_end = xen_io_tlb_start + bytes;
257	/*
258	 * And replace that memory with pages under 4GB.
259	 */
260	rc = xen_swiotlb_fixup(xen_io_tlb_start,
261			       bytes,
262			       xen_io_tlb_nslabs);
263	if (rc) {
264		if (early)
265			free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes));
266		else {
267			free_pages((unsigned long)xen_io_tlb_start, order);
268			xen_io_tlb_start = NULL;
269		}
270		m_ret = XEN_SWIOTLB_EFIXUP;
271		goto error;
272	}
273	start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
274	if (early) {
275		if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
276			 verbose))
277			panic("Cannot allocate SWIOTLB buffer");
278		rc = 0;
279	} else
280		rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
281	return rc;
282error:
283	if (repeat--) {
284		xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
285					(xen_io_tlb_nslabs >> 1));
286		pr_info("Lowering to %luMB\n",
287			(xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
288		goto retry;
289	}
290	pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
291	if (early)
292		panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
293	else
294		free_pages((unsigned long)xen_io_tlb_start, order);
295	return rc;
296}
297void *
298xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
299			   dma_addr_t *dma_handle, gfp_t flags,
300			   struct dma_attrs *attrs)
301{
302	void *ret;
303	int order = get_order(size);
304	u64 dma_mask = DMA_BIT_MASK(32);
305	phys_addr_t phys;
306	dma_addr_t dev_addr;
307
308	/*
309	* Ignore region specifiers - the kernel's ideas of
310	* pseudo-phys memory layout has nothing to do with the
311	* machine physical layout.  We can't allocate highmem
312	* because we can't return a pointer to it.
313	*/
314	flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
315
316	if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
317		return ret;
318
319	/* On ARM this function returns an ioremap'ped virtual address for
320	 * which virt_to_phys doesn't return the corresponding physical
321	 * address. In fact on ARM virt_to_phys only works for kernel direct
322	 * mapped RAM memory. Also see comment below.
323	 */
324	ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
325
 
326	if (!ret)
327		return ret;
 
328
329	if (hwdev && hwdev->coherent_dma_mask)
330		dma_mask = dma_alloc_coherent_mask(hwdev, flags);
 
 
 
 
 
 
331
332	/* At this point dma_handle is the physical address, next we are
333	 * going to set it to the machine address.
334	 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
335	 * to *dma_handle. */
336	phys = *dma_handle;
337	dev_addr = xen_phys_to_bus(phys);
338	if (((dev_addr + size - 1 <= dma_mask)) &&
339	    !range_straddles_page_boundary(phys, size))
340		*dma_handle = dev_addr;
341	else {
342		if (xen_create_contiguous_region(phys, order,
343						 fls64(dma_mask), dma_handle) != 0) {
344			xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
345			return NULL;
346		}
347	}
348	memset(ret, 0, size);
349	return ret;
 
 
 
 
350}
351EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
352
353void
354xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
355			  dma_addr_t dev_addr, struct dma_attrs *attrs)
356{
 
357	int order = get_order(size);
358	phys_addr_t phys;
359	u64 dma_mask = DMA_BIT_MASK(32);
360
361	if (dma_release_from_coherent(hwdev, order, vaddr))
362		return;
363
364	if (hwdev && hwdev->coherent_dma_mask)
365		dma_mask = hwdev->coherent_dma_mask;
366
367	/* do not use virt_to_phys because on ARM it doesn't return you the
368	 * physical address */
369	phys = xen_bus_to_phys(dev_addr);
370
371	if (((dev_addr + size - 1 > dma_mask)) ||
372	    range_straddles_page_boundary(phys, size))
373		xen_destroy_contiguous_region(phys, order);
374
375	xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
376}
377EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
378
379
380/*
381 * Map a single buffer of the indicated size for DMA in streaming mode.  The
382 * physical address to use is returned.
383 *
384 * Once the device is given the dma address, the device owns this memory until
385 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
386 */
387dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
388				unsigned long offset, size_t size,
389				enum dma_data_direction dir,
390				struct dma_attrs *attrs)
391{
392	phys_addr_t map, phys = page_to_phys(page) + offset;
393	dma_addr_t dev_addr = xen_phys_to_bus(phys);
394
395	BUG_ON(dir == DMA_NONE);
396	/*
397	 * If the address happens to be in the device's DMA window,
398	 * we can safely return the device addr and not worry about bounce
399	 * buffering it.
400	 */
401	if (dma_capable(dev, dev_addr, size) &&
402	    !range_straddles_page_boundary(phys, size) && !swiotlb_force) {
403		/* we are not interested in the dma_addr returned by
404		 * xen_dma_map_page, only in the potential cache flushes executed
405		 * by the function. */
406		xen_dma_map_page(dev, page, offset, size, dir, attrs);
407		return dev_addr;
408	}
409
410	/*
411	 * Oh well, have to allocate and map a bounce buffer.
412	 */
413	trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
 
 
 
 
414
415	map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
416	if (map == SWIOTLB_MAP_ERROR)
417		return DMA_ERROR_CODE;
418
419	xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
420					map & ~PAGE_MASK, size, dir, attrs);
421	dev_addr = xen_phys_to_bus(map);
422
423	/*
424	 * Ensure that the address returned is DMA'ble
425	 */
426	if (!dma_capable(dev, dev_addr, size)) {
427		swiotlb_tbl_unmap_single(dev, map, size, dir);
428		dev_addr = 0;
 
 
 
 
 
 
 
 
 
429	}
430	return dev_addr;
431}
432EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
433
434/*
435 * Unmap a single streaming mode DMA translation.  The dma_addr and size must
436 * match what was provided for in a previous xen_swiotlb_map_page call.  All
437 * other usages are undefined.
438 *
439 * After this call, reads by the cpu to the buffer are guaranteed to see
440 * whatever the device wrote there.
441 */
442static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
443			     size_t size, enum dma_data_direction dir,
444				 struct dma_attrs *attrs)
445{
446	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
447
448	BUG_ON(dir == DMA_NONE);
449
450	xen_dma_unmap_page(hwdev, paddr, size, dir, attrs);
 
 
 
 
 
451
452	/* NOTE: We use dev_addr here, not paddr! */
453	if (is_xen_swiotlb_buffer(dev_addr)) {
454		swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
455		return;
456	}
457
458	if (dir != DMA_FROM_DEVICE)
459		return;
460
461	/*
462	 * phys_to_virt doesn't work with hihgmem page but we could
463	 * call dma_mark_clean() with hihgmem page here. However, we
464	 * are fine since dma_mark_clean() is null on POWERPC. We can
465	 * make dma_mark_clean() take a physical address if necessary.
466	 */
467	dma_mark_clean(phys_to_virt(paddr), size);
468}
469
470void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
471			    size_t size, enum dma_data_direction dir,
472			    struct dma_attrs *attrs)
473{
474	xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
475}
476EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
477
478/*
479 * Make physical memory consistent for a single streaming mode DMA translation
480 * after a transfer.
481 *
482 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
483 * using the cpu, yet do not wish to teardown the dma mapping, you must
484 * call this function before doing so.  At the next point you give the dma
485 * address back to the card, you must first perform a
486 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
487 */
488static void
489xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
490			size_t size, enum dma_data_direction dir,
491			enum dma_sync_target target)
492{
493	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
494
495	BUG_ON(dir == DMA_NONE);
496
497	if (target == SYNC_FOR_CPU)
498		xen_dma_sync_single_for_cpu(hwdev, paddr, size, dir);
499
500	/* NOTE: We use dev_addr here, not paddr! */
501	if (is_xen_swiotlb_buffer(dev_addr))
502		swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
503
504	if (target == SYNC_FOR_DEVICE)
505		xen_dma_sync_single_for_cpu(hwdev, paddr, size, dir);
506
507	if (dir != DMA_FROM_DEVICE)
508		return;
509
510	dma_mark_clean(phys_to_virt(paddr), size);
 
511}
512
513void
514xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
515				size_t size, enum dma_data_direction dir)
516{
517	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
518}
519EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
 
520
521void
522xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
523				   size_t size, enum dma_data_direction dir)
524{
525	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
 
526}
527EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
528
529/*
530 * Map a set of buffers described by scatterlist in streaming mode for DMA.
531 * This is the scatter-gather version of the above xen_swiotlb_map_page
532 * interface.  Here the scatter gather list elements are each tagged with the
533 * appropriate dma address and length.  They are obtained via
534 * sg_dma_{address,length}(SG).
535 *
536 * NOTE: An implementation may be able to use a smaller number of
537 *       DMA address/length pairs than there are SG table elements.
538 *       (for example via virtual mapping capabilities)
539 *       The routine returns the number of addr/length pairs actually
540 *       used, at most nents.
541 *
542 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
543 * same here.
544 */
545int
546xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
547			 int nelems, enum dma_data_direction dir,
548			 struct dma_attrs *attrs)
549{
550	struct scatterlist *sg;
551	int i;
552
553	BUG_ON(dir == DMA_NONE);
554
555	for_each_sg(sgl, sg, nelems, i) {
556		phys_addr_t paddr = sg_phys(sg);
557		dma_addr_t dev_addr = xen_phys_to_bus(paddr);
558
559		if (swiotlb_force ||
560		    !dma_capable(hwdev, dev_addr, sg->length) ||
561		    range_straddles_page_boundary(paddr, sg->length)) {
562			phys_addr_t map = swiotlb_tbl_map_single(hwdev,
563								 start_dma_addr,
564								 sg_phys(sg),
565								 sg->length,
566								 dir);
567			if (map == SWIOTLB_MAP_ERROR) {
568				dev_warn(hwdev, "swiotlb buffer is full\n");
569				/* Don't panic here, we expect map_sg users
570				   to do proper error handling. */
571				xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
572							   attrs);
573				sg_dma_len(sgl) = 0;
574				return 0;
575			}
576			xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
577						map & ~PAGE_MASK,
578						sg->length,
579						dir,
580						attrs);
581			sg->dma_address = xen_phys_to_bus(map);
582		} else {
583			/* we are not interested in the dma_addr returned by
584			 * xen_dma_map_page, only in the potential cache flushes executed
585			 * by the function. */
586			xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
587						paddr & ~PAGE_MASK,
588						sg->length,
589						dir,
590						attrs);
591			sg->dma_address = dev_addr;
592		}
593		sg_dma_len(sg) = sg->length;
594	}
595	return nelems;
596}
597EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
598
599/*
600 * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
601 * concerning calls here are the same as for swiotlb_unmap_page() above.
602 */
603void
604xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
605			   int nelems, enum dma_data_direction dir,
606			   struct dma_attrs *attrs)
607{
608	struct scatterlist *sg;
609	int i;
610
611	BUG_ON(dir == DMA_NONE);
612
613	for_each_sg(sgl, sg, nelems, i)
614		xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
 
 
 
 
 
615
 
 
 
 
 
616}
617EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
618
619/*
620 * Make physical memory consistent for a set of streaming mode DMA translations
621 * after a transfer.
622 *
623 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
624 * and usage.
625 */
626static void
627xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
628		    int nelems, enum dma_data_direction dir,
629		    enum dma_sync_target target)
630{
631	struct scatterlist *sg;
632	int i;
633
634	for_each_sg(sgl, sg, nelems, i)
635		xen_swiotlb_sync_single(hwdev, sg->dma_address,
636					sg_dma_len(sg), dir, target);
 
637}
638
639void
640xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
641			    int nelems, enum dma_data_direction dir)
642{
643	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
644}
645EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
646
647void
648xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
649			       int nelems, enum dma_data_direction dir)
650{
651	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
652}
653EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
654
655int
656xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
657{
658	return !dma_addr;
659}
660EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
661
662/*
663 * Return whether the given device DMA address mask can be supported
664 * properly.  For example, if your device can only drive the low 24-bits
665 * during bus mastering, then you would pass 0x00ffffff as the mask to
666 * this function.
667 */
668int
669xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
670{
671	return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
672}
673EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);
674
675int
676xen_swiotlb_set_dma_mask(struct device *dev, u64 dma_mask)
677{
678	if (!dev->dma_mask || !xen_swiotlb_dma_supported(dev, dma_mask))
679		return -EIO;
680
681	*dev->dma_mask = dma_mask;
682
683	return 0;
684}
685EXPORT_SYMBOL_GPL(xen_swiotlb_set_dma_mask);