blob: 810ad419e34ca76ef3f2d0444e789d579dea1b9e [file] [log] [blame]
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -04001/*
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
Joe Perches283c0972013-06-28 03:21:41 -070036#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
37
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -040038#include <linux/bootmem.h>
39#include <linux/dma-mapping.h>
Paul Gortmaker63c97442011-07-10 13:22:07 -040040#include <linux/export.h>
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -040041#include <xen/swiotlb-xen.h>
42#include <xen/page.h>
43#include <xen/xen-ops.h>
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -040044#include <xen/hvc-console.h>
Zoltan Kiss2b2b6142013-09-04 21:11:05 +010045
Stefano Stabellini83862cc2013-10-10 13:40:44 +000046#include <asm/dma-mapping.h>
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +000047#include <asm/xen/page-coherent.h>
Konrad Rzeszutek Wilke1d8f622013-11-08 15:36:09 -050048
Zoltan Kiss2b2b6142013-09-04 21:11:05 +010049#include <trace/events/swiotlb.h>
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -040050/*
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
Stefano Stabellini83862cc2013-10-10 13:40:44 +000056#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
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -040070static 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
Konrad Rzeszutek Wilkb8b0f552012-08-21 14:49:34 -040076static u64 start_dma_addr;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -040077
Ian Campbelle17b2f12014-01-20 11:30:41 +000078/*
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 */
Stefano Stabellini6b42a7e2013-10-25 10:33:27 +000083static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -040084{
Ian Campbelle17b2f12014-01-20 11:30:41 +000085 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;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -040091}
92
Stefano Stabellini6b42a7e2013-10-25 10:33:27 +000093static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -040094{
Ian Campbelle17b2f12014-01-20 11:30:41 +000095 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
Ian Campbelle17b2f12014-01-20 11:30:41 +000099 paddr |= baddr & ~PAGE_MASK;
100
101 return paddr;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400102}
103
Stefano Stabellini6b42a7e2013-10-25 10:33:27 +0000104static inline dma_addr_t xen_virt_to_bus(void *address)
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400105{
106 return xen_phys_to_bus(virt_to_phys(address));
107}
108
109static int check_pages_physically_contiguous(unsigned long pfn,
110 unsigned int offset,
111 size_t length)
112{
113 unsigned long next_mfn;
114 int i;
115 int nr_pages;
116
117 next_mfn = pfn_to_mfn(pfn);
118 nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
119
120 for (i = 1; i < nr_pages; i++) {
121 if (pfn_to_mfn(++pfn) != ++next_mfn)
122 return 0;
123 }
124 return 1;
125}
126
Stefano Stabellini6b42a7e2013-10-25 10:33:27 +0000127static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400128{
129 unsigned long pfn = PFN_DOWN(p);
130 unsigned int offset = p & ~PAGE_MASK;
131
132 if (offset + size <= PAGE_SIZE)
133 return 0;
134 if (check_pages_physically_contiguous(pfn, offset, size))
135 return 0;
136 return 1;
137}
138
139static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
140{
141 unsigned long mfn = PFN_DOWN(dma_addr);
142 unsigned long pfn = mfn_to_local_pfn(mfn);
143 phys_addr_t paddr;
144
145 /* If the address is outside our domain, it CAN
146 * have the same virtual address as another address
147 * in our domain. Therefore _only_ check address within our domain.
148 */
149 if (pfn_valid(pfn)) {
150 paddr = PFN_PHYS(pfn);
151 return paddr >= virt_to_phys(xen_io_tlb_start) &&
152 paddr < virt_to_phys(xen_io_tlb_end);
153 }
154 return 0;
155}
156
157static int max_dma_bits = 32;
158
159static int
160xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
161{
162 int i, rc;
163 int dma_bits;
Stefano Stabellini69908902013-10-09 16:56:32 +0000164 dma_addr_t dma_handle;
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000165 phys_addr_t p = virt_to_phys(buf);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400166
167 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
168
169 i = 0;
170 do {
171 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
172
173 do {
174 rc = xen_create_contiguous_region(
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000175 p + (i << IO_TLB_SHIFT),
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400176 get_order(slabs << IO_TLB_SHIFT),
Stefano Stabellini69908902013-10-09 16:56:32 +0000177 dma_bits, &dma_handle);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400178 } while (rc && dma_bits++ < max_dma_bits);
179 if (rc)
180 return rc;
181
182 i += slabs;
183 } while (i < nslabs);
184 return 0;
185}
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400186static unsigned long xen_set_nslabs(unsigned long nr_tbl)
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400187{
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400188 if (!nr_tbl) {
FUJITA Tomonori5f98ecd2011-06-05 11:47:29 +0900189 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
190 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400191 } else
192 xen_io_tlb_nslabs = nr_tbl;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400193
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400194 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
195}
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400196
197enum xen_swiotlb_err {
198 XEN_SWIOTLB_UNKNOWN = 0,
199 XEN_SWIOTLB_ENOMEM,
200 XEN_SWIOTLB_EFIXUP
201};
202
203static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
204{
205 switch (err) {
206 case XEN_SWIOTLB_ENOMEM:
207 return "Cannot allocate Xen-SWIOTLB buffer\n";
208 case XEN_SWIOTLB_EFIXUP:
209 return "Failed to get contiguous memory for DMA from Xen!\n"\
210 "You either: don't have the permissions, do not have"\
211 " enough free memory under 4GB, or the hypervisor memory"\
212 " is too fragmented!";
213 default:
214 break;
215 }
216 return "";
217}
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400218int __ref xen_swiotlb_init(int verbose, bool early)
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400219{
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400220 unsigned long bytes, order;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400221 int rc = -ENOMEM;
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400222 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400223 unsigned int repeat = 3;
224
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400225 xen_io_tlb_nslabs = swiotlb_nr_tbl();
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400226retry:
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400227 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400228 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400229 /*
230 * Get IO TLB memory from any location.
231 */
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400232 if (early)
233 xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes));
234 else {
235#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
236#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
237 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
238 xen_io_tlb_start = (void *)__get_free_pages(__GFP_NOWARN, order);
239 if (xen_io_tlb_start)
240 break;
241 order--;
242 }
243 if (order != get_order(bytes)) {
Joe Perches283c0972013-06-28 03:21:41 -0700244 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
245 (PAGE_SIZE << order) >> 20);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400246 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
247 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
248 }
249 }
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400250 if (!xen_io_tlb_start) {
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400251 m_ret = XEN_SWIOTLB_ENOMEM;
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400252 goto error;
253 }
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400254 xen_io_tlb_end = xen_io_tlb_start + bytes;
255 /*
256 * And replace that memory with pages under 4GB.
257 */
258 rc = xen_swiotlb_fixup(xen_io_tlb_start,
259 bytes,
260 xen_io_tlb_nslabs);
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400261 if (rc) {
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400262 if (early)
263 free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes));
264 else {
265 free_pages((unsigned long)xen_io_tlb_start, order);
266 xen_io_tlb_start = NULL;
267 }
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400268 m_ret = XEN_SWIOTLB_EFIXUP;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400269 goto error;
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400270 }
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400271 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
Konrad Rzeszutek Wilkc468bde2012-09-17 10:20:09 -0400272 if (early) {
Yinghai Luac2cbab2013-01-24 12:20:16 -0800273 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
274 verbose))
275 panic("Cannot allocate SWIOTLB buffer");
Konrad Rzeszutek Wilkc468bde2012-09-17 10:20:09 -0400276 rc = 0;
277 } else
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400278 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
279 return rc;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400280error:
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400281 if (repeat--) {
282 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
283 (xen_io_tlb_nslabs >> 1));
Joe Perches283c0972013-06-28 03:21:41 -0700284 pr_info("Lowering to %luMB\n",
285 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400286 goto retry;
287 }
Joe Perches283c0972013-06-28 03:21:41 -0700288 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400289 if (early)
290 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
291 else
292 free_pages((unsigned long)xen_io_tlb_start, order);
293 return rc;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400294}
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400295void *
296xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
Andrzej Pietrasiewiczbaa676f2012-03-27 14:28:18 +0200297 dma_addr_t *dma_handle, gfp_t flags,
298 struct dma_attrs *attrs)
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400299{
300 void *ret;
301 int order = get_order(size);
302 u64 dma_mask = DMA_BIT_MASK(32);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400303 phys_addr_t phys;
304 dma_addr_t dev_addr;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400305
306 /*
307 * Ignore region specifiers - the kernel's ideas of
308 * pseudo-phys memory layout has nothing to do with the
309 * machine physical layout. We can't allocate highmem
310 * because we can't return a pointer to it.
311 */
312 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
313
314 if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
315 return ret;
316
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000317 /* On ARM this function returns an ioremap'ped virtual address for
318 * which virt_to_phys doesn't return the corresponding physical
319 * address. In fact on ARM virt_to_phys only works for kernel direct
320 * mapped RAM memory. Also see comment below.
321 */
322 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400323
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400324 if (!ret)
325 return ret;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400326
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400327 if (hwdev && hwdev->coherent_dma_mask)
Ronny Hegewaldb5031ed2012-08-31 09:57:52 +0000328 dma_mask = dma_alloc_coherent_mask(hwdev, flags);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400329
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000330 /* At this point dma_handle is the physical address, next we are
331 * going to set it to the machine address.
332 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
333 * to *dma_handle. */
334 phys = *dma_handle;
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400335 dev_addr = xen_phys_to_bus(phys);
336 if (((dev_addr + size - 1 <= dma_mask)) &&
337 !range_straddles_page_boundary(phys, size))
338 *dma_handle = dev_addr;
339 else {
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000340 if (xen_create_contiguous_region(phys, order,
Stefano Stabellini69908902013-10-09 16:56:32 +0000341 fls64(dma_mask), dma_handle) != 0) {
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000342 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400343 return NULL;
344 }
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400345 }
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400346 memset(ret, 0, size);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400347 return ret;
348}
349EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
350
351void
352xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
Andrzej Pietrasiewiczbaa676f2012-03-27 14:28:18 +0200353 dma_addr_t dev_addr, struct dma_attrs *attrs)
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400354{
355 int order = get_order(size);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400356 phys_addr_t phys;
357 u64 dma_mask = DMA_BIT_MASK(32);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400358
359 if (dma_release_from_coherent(hwdev, order, vaddr))
360 return;
361
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400362 if (hwdev && hwdev->coherent_dma_mask)
363 dma_mask = hwdev->coherent_dma_mask;
364
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000365 /* do not use virt_to_phys because on ARM it doesn't return you the
366 * physical address */
367 phys = xen_bus_to_phys(dev_addr);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400368
369 if (((dev_addr + size - 1 > dma_mask)) ||
370 range_straddles_page_boundary(phys, size))
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000371 xen_destroy_contiguous_region(phys, order);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400372
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000373 xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400374}
375EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
376
377
378/*
379 * Map a single buffer of the indicated size for DMA in streaming mode. The
380 * physical address to use is returned.
381 *
382 * Once the device is given the dma address, the device owns this memory until
383 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
384 */
385dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
386 unsigned long offset, size_t size,
387 enum dma_data_direction dir,
388 struct dma_attrs *attrs)
389{
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700390 phys_addr_t map, phys = page_to_phys(page) + offset;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400391 dma_addr_t dev_addr = xen_phys_to_bus(phys);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400392
393 BUG_ON(dir == DMA_NONE);
394 /*
395 * If the address happens to be in the device's DMA window,
396 * we can safely return the device addr and not worry about bounce
397 * buffering it.
398 */
399 if (dma_capable(dev, dev_addr, size) &&
Stefano Stabellinia4dba132014-11-21 11:07:39 +0000400 !range_straddles_page_boundary(phys, size) &&
401 !xen_arch_need_swiotlb(dev, PFN_DOWN(phys), PFN_DOWN(dev_addr)) &&
402 !swiotlb_force) {
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000403 /* 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. */
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000406 xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400407 return dev_addr;
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000408 }
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400409
410 /*
411 * Oh well, have to allocate and map a bounce buffer.
412 */
Zoltan Kiss2b2b6142013-09-04 21:11:05 +0100413 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
414
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400415 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700416 if (map == SWIOTLB_MAP_ERROR)
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400417 return DMA_ERROR_CODE;
418
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000419 xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000420 dev_addr, map & ~PAGE_MASK, size, dir, attrs);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700421 dev_addr = xen_phys_to_bus(map);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400422
423 /*
424 * Ensure that the address returned is DMA'ble
425 */
Konrad Rzeszutek Wilkab2a47b2011-07-22 12:51:48 -0400426 if (!dma_capable(dev, dev_addr, size)) {
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700427 swiotlb_tbl_unmap_single(dev, map, size, dir);
Konrad Rzeszutek Wilkab2a47b2011-07-22 12:51:48 -0400428 dev_addr = 0;
429 }
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400430 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,
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000443 size_t size, enum dma_data_direction dir,
444 struct dma_attrs *attrs)
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400445{
446 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
447
448 BUG_ON(dir == DMA_NONE);
449
Stefano Stabellinid6883e62014-11-21 11:09:39 +0000450 xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000451
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400452 /* NOTE: We use dev_addr here, not paddr! */
453 if (is_xen_swiotlb_buffer(dev_addr)) {
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700454 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400455 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{
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000474 xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400475}
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
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000497 if (target == SYNC_FOR_CPU)
Stefano Stabellinid6883e62014-11-21 11:09:39 +0000498 xen_dma_sync_single_for_cpu(hwdev, dev_addr, size, dir);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000499
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400500 /* NOTE: We use dev_addr here, not paddr! */
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000501 if (is_xen_swiotlb_buffer(dev_addr))
Alexander Duyckfbfda892012-10-15 10:19:49 -0700502 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000503
504 if (target == SYNC_FOR_DEVICE)
Stefano Stabellini9490c6c2014-11-21 16:55:12 +0000505 xen_dma_sync_single_for_device(hwdev, dev_addr, size, dir);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400506
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 ||
Stefano Stabellinia4dba132014-11-21 11:07:39 +0000560 xen_arch_need_swiotlb(hwdev, PFN_DOWN(paddr), PFN_DOWN(dev_addr)) ||
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400561 !dma_capable(hwdev, dev_addr, sg->length) ||
562 range_straddles_page_boundary(paddr, sg->length)) {
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700563 phys_addr_t map = swiotlb_tbl_map_single(hwdev,
564 start_dma_addr,
565 sg_phys(sg),
566 sg->length,
567 dir);
568 if (map == SWIOTLB_MAP_ERROR) {
Stefano Stabellini783d0282013-10-25 10:33:26 +0000569 dev_warn(hwdev, "swiotlb buffer is full\n");
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400570 /* Don't panic here, we expect map_sg users
571 to do proper error handling. */
572 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
573 attrs);
Stefano Stabellini781575c2013-08-05 17:30:48 +0100574 sg_dma_len(sgl) = 0;
Stefano Stabellini15177602013-10-29 00:37:37 +0000575 return 0;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400576 }
Stefano Stabellini71bfae92013-11-15 14:22:15 +0000577 xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000578 dev_addr,
Stefano Stabellini71bfae92013-11-15 14:22:15 +0000579 map & ~PAGE_MASK,
580 sg->length,
581 dir,
582 attrs);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700583 sg->dma_address = xen_phys_to_bus(map);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000584 } else {
585 /* we are not interested in the dma_addr returned by
586 * xen_dma_map_page, only in the potential cache flushes executed
587 * by the function. */
588 xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000589 dev_addr,
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000590 paddr & ~PAGE_MASK,
591 sg->length,
592 dir,
593 attrs);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400594 sg->dma_address = dev_addr;
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000595 }
Stefano Stabellini781575c2013-08-05 17:30:48 +0100596 sg_dma_len(sg) = sg->length;
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400597 }
598 return nelems;
599}
600EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
601
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400602/*
603 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
604 * concerning calls here are the same as for swiotlb_unmap_page() above.
605 */
606void
607xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
608 int nelems, enum dma_data_direction dir,
609 struct dma_attrs *attrs)
610{
611 struct scatterlist *sg;
612 int i;
613
614 BUG_ON(dir == DMA_NONE);
615
616 for_each_sg(sgl, sg, nelems, i)
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000617 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400618
619}
620EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
621
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400622/*
623 * Make physical memory consistent for a set of streaming mode DMA translations
624 * after a transfer.
625 *
626 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
627 * and usage.
628 */
629static void
630xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
631 int nelems, enum dma_data_direction dir,
632 enum dma_sync_target target)
633{
634 struct scatterlist *sg;
635 int i;
636
637 for_each_sg(sgl, sg, nelems, i)
638 xen_swiotlb_sync_single(hwdev, sg->dma_address,
Stefano Stabellini781575c2013-08-05 17:30:48 +0100639 sg_dma_len(sg), dir, target);
Konrad Rzeszutek Wilkb0971862010-05-11 10:05:49 -0400640}
641
642void
643xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
644 int nelems, enum dma_data_direction dir)
645{
646 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
647}
648EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
649
650void
651xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
652 int nelems, enum dma_data_direction dir)
653{
654 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
655}
656EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
657
658int
659xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
660{
661 return !dma_addr;
662}
663EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
664
665/*
666 * Return whether the given device DMA address mask can be supported
667 * properly. For example, if your device can only drive the low 24-bits
668 * during bus mastering, then you would pass 0x00ffffff as the mask to
669 * this function.
670 */
671int
672xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
673{
674 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
675}
676EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);
Stefano Stabellinieb1ddc02013-10-09 16:56:33 +0000677
678int
679xen_swiotlb_set_dma_mask(struct device *dev, u64 dma_mask)
680{
681 if (!dev->dma_mask || !xen_swiotlb_dma_supported(dev, dma_mask))
682 return -EIO;
683
684 *dev->dma_mask = dma_mask;
685
686 return 0;
687}
688EXPORT_SYMBOL_GPL(xen_swiotlb_set_dma_mask);