blob: 3725ee4ff43c0f2ed66391dd79f09b42e4a26e62 [file] [log] [blame]
Konrad Rzeszutek Wilkb097186f2010-05-11 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 Wilkb097186f2010-05-11 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 Wilkb097186f2010-05-11 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 Wilkb097186f2010-05-11 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 Wilkb097186f2010-05-11 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 Wilkb097186f2010-05-11 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 Wilkb097186f2010-05-11 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 Wilkb097186f2010-05-11 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 Wilkb097186f2010-05-11 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
99 BUG_ON(paddr != dma); /* truncation has occurred, should never happen */
100
101 paddr |= baddr & ~PAGE_MASK;
102
103 return paddr;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400104}
105
Stefano Stabellini6b42a7e2013-10-25 10:33:27 +0000106static inline dma_addr_t xen_virt_to_bus(void *address)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400107{
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
Stefano Stabellini6b42a7e2013-10-25 10:33:27 +0000129static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400130{
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;
Stefano Stabellini69908902013-10-09 16:56:32 +0000166 dma_addr_t dma_handle;
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000167 phys_addr_t p = virt_to_phys(buf);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400168
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(
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000177 p + (i << IO_TLB_SHIFT),
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400178 get_order(slabs << IO_TLB_SHIFT),
Stefano Stabellini69908902013-10-09 16:56:32 +0000179 dma_bits, &dma_handle);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400180 } 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}
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400188static unsigned long xen_set_nslabs(unsigned long nr_tbl)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400189{
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400190 if (!nr_tbl) {
FUJITA Tomonori5f98ecd2011-06-05 11:47:29 +0900191 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
192 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400193 } else
194 xen_io_tlb_nslabs = nr_tbl;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400195
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400196 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
197}
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400198
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}
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400220int __ref xen_swiotlb_init(int verbose, bool early)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400221{
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400222 unsigned long bytes, order;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400223 int rc = -ENOMEM;
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400224 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400225 unsigned int repeat = 3;
226
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400227 xen_io_tlb_nslabs = swiotlb_nr_tbl();
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400228retry:
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400229 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400230 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400231 /*
232 * Get IO TLB memory from any location.
233 */
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400234 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)) {
Joe Perches283c0972013-06-28 03:21:41 -0700246 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
247 (PAGE_SIZE << order) >> 20);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400248 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
249 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
250 }
251 }
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400252 if (!xen_io_tlb_start) {
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400253 m_ret = XEN_SWIOTLB_ENOMEM;
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400254 goto error;
255 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400256 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);
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400263 if (rc) {
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400264 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 }
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400270 m_ret = XEN_SWIOTLB_EFIXUP;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400271 goto error;
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400272 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400273 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
Konrad Rzeszutek Wilkc468bde2012-09-17 10:20:09 -0400274 if (early) {
Yinghai Luac2cbab2013-01-24 12:20:16 -0800275 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
276 verbose))
277 panic("Cannot allocate SWIOTLB buffer");
Konrad Rzeszutek Wilkc468bde2012-09-17 10:20:09 -0400278 rc = 0;
279 } else
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400280 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
281 return rc;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400282error:
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400283 if (repeat--) {
284 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
285 (xen_io_tlb_nslabs >> 1));
Joe Perches283c0972013-06-28 03:21:41 -0700286 pr_info("Lowering to %luMB\n",
287 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400288 goto retry;
289 }
Joe Perches283c0972013-06-28 03:21:41 -0700290 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400291 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;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400296}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400297void *
298xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
Andrzej Pietrasiewiczbaa676f2012-03-27 14:28:18 +0200299 dma_addr_t *dma_handle, gfp_t flags,
300 struct dma_attrs *attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400301{
302 void *ret;
303 int order = get_order(size);
304 u64 dma_mask = DMA_BIT_MASK(32);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400305 phys_addr_t phys;
306 dma_addr_t dev_addr;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400307
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
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000319 /* 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);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400325
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400326 if (!ret)
327 return ret;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400328
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400329 if (hwdev && hwdev->coherent_dma_mask)
Ronny Hegewaldb5031ed2012-08-31 09:57:52 +0000330 dma_mask = dma_alloc_coherent_mask(hwdev, flags);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400331
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000332 /* 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;
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400337 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 {
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000342 if (xen_create_contiguous_region(phys, order,
Stefano Stabellini69908902013-10-09 16:56:32 +0000343 fls64(dma_mask), dma_handle) != 0) {
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000344 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400345 return NULL;
346 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400347 }
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400348 memset(ret, 0, size);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400349 return ret;
350}
351EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
352
353void
354xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
Andrzej Pietrasiewiczbaa676f2012-03-27 14:28:18 +0200355 dma_addr_t dev_addr, struct dma_attrs *attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400356{
357 int order = get_order(size);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400358 phys_addr_t phys;
359 u64 dma_mask = DMA_BIT_MASK(32);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400360
361 if (dma_release_from_coherent(hwdev, order, vaddr))
362 return;
363
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400364 if (hwdev && hwdev->coherent_dma_mask)
365 dma_mask = hwdev->coherent_dma_mask;
366
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000367 /* 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);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400370
371 if (((dev_addr + size - 1 > dma_mask)) ||
372 range_straddles_page_boundary(phys, size))
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000373 xen_destroy_contiguous_region(phys, order);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400374
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000375 xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400376}
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{
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700392 phys_addr_t map, phys = page_to_phys(page) + offset;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400393 dma_addr_t dev_addr = xen_phys_to_bus(phys);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400394
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) &&
Stefano Stabellinia4dba132014-11-21 11:07:39 +0000402 !range_straddles_page_boundary(phys, size) &&
403 !xen_arch_need_swiotlb(dev, PFN_DOWN(phys), PFN_DOWN(dev_addr)) &&
404 !swiotlb_force) {
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000405 /* we are not interested in the dma_addr returned by
406 * xen_dma_map_page, only in the potential cache flushes executed
407 * by the function. */
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000408 xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400409 return dev_addr;
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000410 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400411
412 /*
413 * Oh well, have to allocate and map a bounce buffer.
414 */
Zoltan Kiss2b2b6142013-09-04 21:11:05 +0100415 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
416
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400417 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700418 if (map == SWIOTLB_MAP_ERROR)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400419 return DMA_ERROR_CODE;
420
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000421 xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000422 dev_addr, map & ~PAGE_MASK, size, dir, attrs);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700423 dev_addr = xen_phys_to_bus(map);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400424
425 /*
426 * Ensure that the address returned is DMA'ble
427 */
Konrad Rzeszutek Wilkab2a47b2011-07-22 12:51:48 -0400428 if (!dma_capable(dev, dev_addr, size)) {
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700429 swiotlb_tbl_unmap_single(dev, map, size, dir);
Konrad Rzeszutek Wilkab2a47b2011-07-22 12:51:48 -0400430 dev_addr = 0;
431 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400432 return dev_addr;
433}
434EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
435
436/*
437 * Unmap a single streaming mode DMA translation. The dma_addr and size must
438 * match what was provided for in a previous xen_swiotlb_map_page call. All
439 * other usages are undefined.
440 *
441 * After this call, reads by the cpu to the buffer are guaranteed to see
442 * whatever the device wrote there.
443 */
444static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000445 size_t size, enum dma_data_direction dir,
446 struct dma_attrs *attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400447{
448 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
449
450 BUG_ON(dir == DMA_NONE);
451
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000452 xen_dma_unmap_page(hwdev, paddr, size, dir, attrs);
453
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400454 /* NOTE: We use dev_addr here, not paddr! */
455 if (is_xen_swiotlb_buffer(dev_addr)) {
Alexander Duyck61ca08c2012-10-15 10:19:44 -0700456 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400457 return;
458 }
459
460 if (dir != DMA_FROM_DEVICE)
461 return;
462
463 /*
464 * phys_to_virt doesn't work with hihgmem page but we could
465 * call dma_mark_clean() with hihgmem page here. However, we
466 * are fine since dma_mark_clean() is null on POWERPC. We can
467 * make dma_mark_clean() take a physical address if necessary.
468 */
469 dma_mark_clean(phys_to_virt(paddr), size);
470}
471
472void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
473 size_t size, enum dma_data_direction dir,
474 struct dma_attrs *attrs)
475{
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000476 xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400477}
478EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
479
480/*
481 * Make physical memory consistent for a single streaming mode DMA translation
482 * after a transfer.
483 *
484 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
485 * using the cpu, yet do not wish to teardown the dma mapping, you must
486 * call this function before doing so. At the next point you give the dma
487 * address back to the card, you must first perform a
488 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
489 */
490static void
491xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
492 size_t size, enum dma_data_direction dir,
493 enum dma_sync_target target)
494{
495 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
496
497 BUG_ON(dir == DMA_NONE);
498
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000499 if (target == SYNC_FOR_CPU)
500 xen_dma_sync_single_for_cpu(hwdev, paddr, size, dir);
501
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400502 /* NOTE: We use dev_addr here, not paddr! */
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000503 if (is_xen_swiotlb_buffer(dev_addr))
Alexander Duyckfbfda892012-10-15 10:19:49 -0700504 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000505
506 if (target == SYNC_FOR_DEVICE)
507 xen_dma_sync_single_for_cpu(hwdev, paddr, size, dir);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400508
509 if (dir != DMA_FROM_DEVICE)
510 return;
511
512 dma_mark_clean(phys_to_virt(paddr), size);
513}
514
515void
516xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
517 size_t size, enum dma_data_direction dir)
518{
519 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
520}
521EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
522
523void
524xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
525 size_t size, enum dma_data_direction dir)
526{
527 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
528}
529EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
530
531/*
532 * Map a set of buffers described by scatterlist in streaming mode for DMA.
533 * This is the scatter-gather version of the above xen_swiotlb_map_page
534 * interface. Here the scatter gather list elements are each tagged with the
535 * appropriate dma address and length. They are obtained via
536 * sg_dma_{address,length}(SG).
537 *
538 * NOTE: An implementation may be able to use a smaller number of
539 * DMA address/length pairs than there are SG table elements.
540 * (for example via virtual mapping capabilities)
541 * The routine returns the number of addr/length pairs actually
542 * used, at most nents.
543 *
544 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
545 * same here.
546 */
547int
548xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
549 int nelems, enum dma_data_direction dir,
550 struct dma_attrs *attrs)
551{
552 struct scatterlist *sg;
553 int i;
554
555 BUG_ON(dir == DMA_NONE);
556
557 for_each_sg(sgl, sg, nelems, i) {
558 phys_addr_t paddr = sg_phys(sg);
559 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
560
561 if (swiotlb_force ||
Stefano Stabellinia4dba132014-11-21 11:07:39 +0000562 xen_arch_need_swiotlb(hwdev, PFN_DOWN(paddr), PFN_DOWN(dev_addr)) ||
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400563 !dma_capable(hwdev, dev_addr, sg->length) ||
564 range_straddles_page_boundary(paddr, sg->length)) {
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700565 phys_addr_t map = swiotlb_tbl_map_single(hwdev,
566 start_dma_addr,
567 sg_phys(sg),
568 sg->length,
569 dir);
570 if (map == SWIOTLB_MAP_ERROR) {
Stefano Stabellini783d0282013-10-25 10:33:26 +0000571 dev_warn(hwdev, "swiotlb buffer is full\n");
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400572 /* Don't panic here, we expect map_sg users
573 to do proper error handling. */
574 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
575 attrs);
Stefano Stabellini781575c2013-08-05 17:30:48 +0100576 sg_dma_len(sgl) = 0;
Stefano Stabellini15177602013-10-29 00:37:37 +0000577 return 0;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400578 }
Stefano Stabellini71bfae92013-11-15 14:22:15 +0000579 xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000580 dev_addr,
Stefano Stabellini71bfae92013-11-15 14:22:15 +0000581 map & ~PAGE_MASK,
582 sg->length,
583 dir,
584 attrs);
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700585 sg->dma_address = xen_phys_to_bus(map);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000586 } else {
587 /* we are not interested in the dma_addr returned by
588 * xen_dma_map_page, only in the potential cache flushes executed
589 * by the function. */
590 xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000591 dev_addr,
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000592 paddr & ~PAGE_MASK,
593 sg->length,
594 dir,
595 attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400596 sg->dma_address = dev_addr;
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000597 }
Stefano Stabellini781575c2013-08-05 17:30:48 +0100598 sg_dma_len(sg) = sg->length;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400599 }
600 return nelems;
601}
602EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
603
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400604/*
605 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
606 * concerning calls here are the same as for swiotlb_unmap_page() above.
607 */
608void
609xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
610 int nelems, enum dma_data_direction dir,
611 struct dma_attrs *attrs)
612{
613 struct scatterlist *sg;
614 int i;
615
616 BUG_ON(dir == DMA_NONE);
617
618 for_each_sg(sgl, sg, nelems, i)
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000619 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400620
621}
622EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
623
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400624/*
625 * Make physical memory consistent for a set of streaming mode DMA translations
626 * after a transfer.
627 *
628 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
629 * and usage.
630 */
631static void
632xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
633 int nelems, enum dma_data_direction dir,
634 enum dma_sync_target target)
635{
636 struct scatterlist *sg;
637 int i;
638
639 for_each_sg(sgl, sg, nelems, i)
640 xen_swiotlb_sync_single(hwdev, sg->dma_address,
Stefano Stabellini781575c2013-08-05 17:30:48 +0100641 sg_dma_len(sg), dir, target);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400642}
643
644void
645xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
646 int nelems, enum dma_data_direction dir)
647{
648 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
649}
650EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
651
652void
653xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
654 int nelems, enum dma_data_direction dir)
655{
656 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
657}
658EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
659
660int
661xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
662{
663 return !dma_addr;
664}
665EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
666
667/*
668 * Return whether the given device DMA address mask can be supported
669 * properly. For example, if your device can only drive the low 24-bits
670 * during bus mastering, then you would pass 0x00ffffff as the mask to
671 * this function.
672 */
673int
674xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
675{
676 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
677}
678EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);
Stefano Stabellinieb1ddc02013-10-09 16:56:33 +0000679
680int
681xen_swiotlb_set_dma_mask(struct device *dev, u64 dma_mask)
682{
683 if (!dev->dma_mask || !xen_swiotlb_dma_supported(dev, dma_mask))
684 return -EIO;
685
686 *dev->dma_mask = dma_mask;
687
688 return 0;
689}
690EXPORT_SYMBOL_GPL(xen_swiotlb_set_dma_mask);