blob: bb7888429be6b98dec9cd7fcfd13f9827a515d05 [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
Mike Rapoport20132882018-10-30 15:09:21 -070038#include <linux/memblock.h>
Christoph Hellwigea8c64a2018-01-10 16:21:13 +010039#include <linux/dma-direct.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
56static char *xen_io_tlb_start, *xen_io_tlb_end;
57static unsigned long xen_io_tlb_nslabs;
58/*
59 * Quick lookup value of the bus address of the IOTLB.
60 */
61
Konrad Rzeszutek Wilkb8b0f552012-08-21 14:49:34 -040062static u64 start_dma_addr;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -040063
Ian Campbelle17b2f12014-01-20 11:30:41 +000064/*
Julien Grall9435cce2015-09-09 15:18:45 +010065 * Both of these functions should avoid XEN_PFN_PHYS because phys_addr_t
Ian Campbelle17b2f12014-01-20 11:30:41 +000066 * can be 32bit when dma_addr_t is 64bit leading to a loss in
67 * information if the shift is done before casting to 64bit.
68 */
Stefano Stabellini6b42a7e2013-10-25 10:33:27 +000069static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -040070{
Julien Grall9435cce2015-09-09 15:18:45 +010071 unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
72 dma_addr_t dma = (dma_addr_t)bfn << XEN_PAGE_SHIFT;
Ian Campbelle17b2f12014-01-20 11:30:41 +000073
Julien Grall9435cce2015-09-09 15:18:45 +010074 dma |= paddr & ~XEN_PAGE_MASK;
Ian Campbelle17b2f12014-01-20 11:30:41 +000075
76 return dma;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -040077}
78
Stefano Stabellini6b42a7e2013-10-25 10:33:27 +000079static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -040080{
Julien Grall9435cce2015-09-09 15:18:45 +010081 unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
82 dma_addr_t dma = (dma_addr_t)xen_pfn << XEN_PAGE_SHIFT;
Ian Campbelle17b2f12014-01-20 11:30:41 +000083 phys_addr_t paddr = dma;
84
Julien Grall9435cce2015-09-09 15:18:45 +010085 paddr |= baddr & ~XEN_PAGE_MASK;
Ian Campbelle17b2f12014-01-20 11:30:41 +000086
87 return paddr;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -040088}
89
Stefano Stabellini6b42a7e2013-10-25 10:33:27 +000090static inline dma_addr_t xen_virt_to_bus(void *address)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -040091{
92 return xen_phys_to_bus(virt_to_phys(address));
93}
94
Julien Grall9435cce2015-09-09 15:18:45 +010095static int check_pages_physically_contiguous(unsigned long xen_pfn,
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -040096 unsigned int offset,
97 size_t length)
98{
Julien Grall32e09872015-08-07 17:34:35 +010099 unsigned long next_bfn;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400100 int i;
101 int nr_pages;
102
Julien Grall9435cce2015-09-09 15:18:45 +0100103 next_bfn = pfn_to_bfn(xen_pfn);
104 nr_pages = (offset + length + XEN_PAGE_SIZE-1) >> XEN_PAGE_SHIFT;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400105
106 for (i = 1; i < nr_pages; i++) {
Julien Grall9435cce2015-09-09 15:18:45 +0100107 if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400108 return 0;
109 }
110 return 1;
111}
112
Stefano Stabellini6b42a7e2013-10-25 10:33:27 +0000113static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400114{
Julien Grall9435cce2015-09-09 15:18:45 +0100115 unsigned long xen_pfn = XEN_PFN_DOWN(p);
116 unsigned int offset = p & ~XEN_PAGE_MASK;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400117
Julien Grall9435cce2015-09-09 15:18:45 +0100118 if (offset + size <= XEN_PAGE_SIZE)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400119 return 0;
Julien Grall9435cce2015-09-09 15:18:45 +0100120 if (check_pages_physically_contiguous(xen_pfn, offset, size))
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400121 return 0;
122 return 1;
123}
124
125static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
126{
Julien Grall9435cce2015-09-09 15:18:45 +0100127 unsigned long bfn = XEN_PFN_DOWN(dma_addr);
128 unsigned long xen_pfn = bfn_to_local_pfn(bfn);
129 phys_addr_t paddr = XEN_PFN_PHYS(xen_pfn);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400130
131 /* If the address is outside our domain, it CAN
132 * have the same virtual address as another address
133 * in our domain. Therefore _only_ check address within our domain.
134 */
Julien Grall9435cce2015-09-09 15:18:45 +0100135 if (pfn_valid(PFN_DOWN(paddr))) {
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400136 return paddr >= virt_to_phys(xen_io_tlb_start) &&
137 paddr < virt_to_phys(xen_io_tlb_end);
138 }
139 return 0;
140}
141
142static int max_dma_bits = 32;
143
144static int
145xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
146{
147 int i, rc;
148 int dma_bits;
Stefano Stabellini69908902013-10-09 16:56:32 +0000149 dma_addr_t dma_handle;
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000150 phys_addr_t p = virt_to_phys(buf);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400151
152 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
153
154 i = 0;
155 do {
156 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
157
158 do {
159 rc = xen_create_contiguous_region(
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000160 p + (i << IO_TLB_SHIFT),
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400161 get_order(slabs << IO_TLB_SHIFT),
Stefano Stabellini69908902013-10-09 16:56:32 +0000162 dma_bits, &dma_handle);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400163 } while (rc && dma_bits++ < max_dma_bits);
164 if (rc)
165 return rc;
166
167 i += slabs;
168 } while (i < nslabs);
169 return 0;
170}
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400171static unsigned long xen_set_nslabs(unsigned long nr_tbl)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400172{
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400173 if (!nr_tbl) {
FUJITA Tomonori5f98ecd2011-06-05 11:47:29 +0900174 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
175 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400176 } else
177 xen_io_tlb_nslabs = nr_tbl;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400178
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400179 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
180}
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400181
182enum xen_swiotlb_err {
183 XEN_SWIOTLB_UNKNOWN = 0,
184 XEN_SWIOTLB_ENOMEM,
185 XEN_SWIOTLB_EFIXUP
186};
187
188static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
189{
190 switch (err) {
191 case XEN_SWIOTLB_ENOMEM:
192 return "Cannot allocate Xen-SWIOTLB buffer\n";
193 case XEN_SWIOTLB_EFIXUP:
194 return "Failed to get contiguous memory for DMA from Xen!\n"\
195 "You either: don't have the permissions, do not have"\
196 " enough free memory under 4GB, or the hypervisor memory"\
197 " is too fragmented!";
198 default:
199 break;
200 }
201 return "";
202}
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400203int __ref xen_swiotlb_init(int verbose, bool early)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400204{
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400205 unsigned long bytes, order;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400206 int rc = -ENOMEM;
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400207 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400208 unsigned int repeat = 3;
209
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400210 xen_io_tlb_nslabs = swiotlb_nr_tbl();
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400211retry:
Konrad Rzeszutek Wilk1cef36a2012-08-23 13:55:26 -0400212 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400213 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400214 /*
215 * Get IO TLB memory from any location.
216 */
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400217 if (early)
Mike Rapoport15c3c112018-10-30 15:08:58 -0700218 xen_io_tlb_start = memblock_alloc(PAGE_ALIGN(bytes),
219 PAGE_SIZE);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400220 else {
221#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
222#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
223 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
Stefano Stabellini87465152015-04-24 10:16:40 +0100224 xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400225 if (xen_io_tlb_start)
226 break;
227 order--;
228 }
229 if (order != get_order(bytes)) {
Joe Perches283c0972013-06-28 03:21:41 -0700230 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
231 (PAGE_SIZE << order) >> 20);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400232 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
233 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
234 }
235 }
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400236 if (!xen_io_tlb_start) {
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400237 m_ret = XEN_SWIOTLB_ENOMEM;
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400238 goto error;
239 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400240 xen_io_tlb_end = xen_io_tlb_start + bytes;
241 /*
242 * And replace that memory with pages under 4GB.
243 */
244 rc = xen_swiotlb_fixup(xen_io_tlb_start,
245 bytes,
246 xen_io_tlb_nslabs);
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400247 if (rc) {
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400248 if (early)
Mike Rapoport20132882018-10-30 15:09:21 -0700249 memblock_free(__pa(xen_io_tlb_start),
250 PAGE_ALIGN(bytes));
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400251 else {
252 free_pages((unsigned long)xen_io_tlb_start, order);
253 xen_io_tlb_start = NULL;
254 }
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400255 m_ret = XEN_SWIOTLB_EFIXUP;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400256 goto error;
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400257 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400258 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
Konrad Rzeszutek Wilkc468bde2012-09-17 10:20:09 -0400259 if (early) {
Yinghai Luac2cbab2013-01-24 12:20:16 -0800260 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
261 verbose))
262 panic("Cannot allocate SWIOTLB buffer");
Konrad Rzeszutek Wilkc468bde2012-09-17 10:20:09 -0400263 rc = 0;
264 } else
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400265 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
Konrad Rzeszutek Wilk7453c542016-12-20 10:02:02 -0500266
267 if (!rc)
268 swiotlb_set_max_segment(PAGE_SIZE);
269
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400270 return rc;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400271error:
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400272 if (repeat--) {
273 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
274 (xen_io_tlb_nslabs >> 1));
Joe Perches283c0972013-06-28 03:21:41 -0700275 pr_info("Lowering to %luMB\n",
276 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400277 goto retry;
278 }
Joe Perches283c0972013-06-28 03:21:41 -0700279 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400280 if (early)
281 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
282 else
283 free_pages((unsigned long)xen_io_tlb_start, order);
284 return rc;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400285}
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200286
287static void *
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400288xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
Andrzej Pietrasiewiczbaa676f2012-03-27 14:28:18 +0200289 dma_addr_t *dma_handle, gfp_t flags,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700290 unsigned long attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400291{
292 void *ret;
293 int order = get_order(size);
294 u64 dma_mask = DMA_BIT_MASK(32);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400295 phys_addr_t phys;
296 dma_addr_t dev_addr;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400297
298 /*
299 * Ignore region specifiers - the kernel's ideas of
300 * pseudo-phys memory layout has nothing to do with the
301 * machine physical layout. We can't allocate highmem
302 * because we can't return a pointer to it.
303 */
304 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
305
Joe Jin7250f422018-10-16 15:21:16 -0700306 /* Convert the size to actually allocated. */
307 size = 1UL << (order + XEN_PAGE_SHIFT);
308
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000309 /* On ARM this function returns an ioremap'ped virtual address for
310 * which virt_to_phys doesn't return the corresponding physical
311 * address. In fact on ARM virt_to_phys only works for kernel direct
312 * mapped RAM memory. Also see comment below.
313 */
314 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400315
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400316 if (!ret)
317 return ret;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400318
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400319 if (hwdev && hwdev->coherent_dma_mask)
Christoph Hellwig038d07a2018-03-19 11:38:14 +0100320 dma_mask = hwdev->coherent_dma_mask;
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400321
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000322 /* At this point dma_handle is the physical address, next we are
323 * going to set it to the machine address.
324 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
325 * to *dma_handle. */
326 phys = *dma_handle;
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400327 dev_addr = xen_phys_to_bus(phys);
328 if (((dev_addr + size - 1 <= dma_mask)) &&
329 !range_straddles_page_boundary(phys, size))
330 *dma_handle = dev_addr;
331 else {
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000332 if (xen_create_contiguous_region(phys, order,
Stefano Stabellini69908902013-10-09 16:56:32 +0000333 fls64(dma_mask), dma_handle) != 0) {
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000334 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400335 return NULL;
336 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400337 }
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400338 memset(ret, 0, size);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400339 return ret;
340}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400341
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200342static void
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400343xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700344 dma_addr_t dev_addr, unsigned long attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400345{
346 int order = get_order(size);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400347 phys_addr_t phys;
348 u64 dma_mask = DMA_BIT_MASK(32);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400349
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400350 if (hwdev && hwdev->coherent_dma_mask)
351 dma_mask = hwdev->coherent_dma_mask;
352
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000353 /* do not use virt_to_phys because on ARM it doesn't return you the
354 * physical address */
355 phys = xen_bus_to_phys(dev_addr);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400356
Joe Jin7250f422018-10-16 15:21:16 -0700357 /* Convert the size to actually allocated. */
358 size = 1UL << (order + XEN_PAGE_SHIFT);
359
Joe Jin4855c922018-05-17 12:33:28 -0700360 if (((dev_addr + size - 1 <= dma_mask)) ||
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400361 range_straddles_page_boundary(phys, size))
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000362 xen_destroy_contiguous_region(phys, order);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400363
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000364 xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400365}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400366
367/*
368 * Map a single buffer of the indicated size for DMA in streaming mode. The
369 * physical address to use is returned.
370 *
371 * Once the device is given the dma address, the device owns this memory until
372 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
373 */
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200374static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400375 unsigned long offset, size_t size,
376 enum dma_data_direction dir,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700377 unsigned long attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400378{
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700379 phys_addr_t map, phys = page_to_phys(page) + offset;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400380 dma_addr_t dev_addr = xen_phys_to_bus(phys);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400381
382 BUG_ON(dir == DMA_NONE);
383 /*
384 * If the address happens to be in the device's DMA window,
385 * we can safely return the device addr and not worry about bounce
386 * buffering it.
387 */
388 if (dma_capable(dev, dev_addr, size) &&
Stefano Stabellinia4dba132014-11-21 11:07:39 +0000389 !range_straddles_page_boundary(phys, size) &&
Julien Grall291be102015-09-09 15:17:33 +0100390 !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
Geert Uytterhoevenae7871b2016-12-16 14:28:41 +0100391 (swiotlb_force != SWIOTLB_FORCE)) {
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000392 /* we are not interested in the dma_addr returned by
393 * xen_dma_map_page, only in the potential cache flushes executed
394 * by the function. */
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000395 xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400396 return dev_addr;
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000397 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400398
399 /*
400 * Oh well, have to allocate and map a bounce buffer.
401 */
Zoltan Kiss2b2b6142013-09-04 21:11:05 +0100402 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
403
Alexander Duyck0443fa02016-11-02 07:13:02 -0400404 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir,
405 attrs);
Christoph Hellwigb907e202018-12-03 11:42:52 +0100406 if (map == DMA_MAPPING_ERROR)
Christoph Hellwiga4abe0a2018-11-21 19:38:19 +0100407 return DMA_MAPPING_ERROR;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400408
Stefano Stabellinif1225ee2017-01-19 10:39:09 -0800409 dev_addr = xen_phys_to_bus(map);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000410 xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000411 dev_addr, map & ~PAGE_MASK, size, dir, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400412
413 /*
414 * Ensure that the address returned is DMA'ble
415 */
Alexander Duyck76418422016-11-02 07:12:47 -0400416 if (dma_capable(dev, dev_addr, size))
417 return dev_addr;
418
Alexander Duyckd29fa0c2016-11-10 07:05:31 -0500419 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
420 swiotlb_tbl_unmap_single(dev, map, size, dir, attrs);
Alexander Duyck76418422016-11-02 07:12:47 -0400421
Christoph Hellwiga4abe0a2018-11-21 19:38:19 +0100422 return DMA_MAPPING_ERROR;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400423}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400424
425/*
426 * Unmap a single streaming mode DMA translation. The dma_addr and size must
427 * match what was provided for in a previous xen_swiotlb_map_page call. All
428 * other usages are undefined.
429 *
430 * After this call, reads by the cpu to the buffer are guaranteed to see
431 * whatever the device wrote there.
432 */
433static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000434 size_t size, enum dma_data_direction dir,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700435 unsigned long attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400436{
437 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
438
439 BUG_ON(dir == DMA_NONE);
440
Stefano Stabellinid6883e62014-11-21 11:09:39 +0000441 xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000442
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400443 /* NOTE: We use dev_addr here, not paddr! */
Christoph Hellwig68c60832018-12-06 07:06:04 -0800444 if (is_xen_swiotlb_buffer(dev_addr))
Alexander Duyck0443fa02016-11-02 07:13:02 -0400445 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400446}
447
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200448static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400449 size_t size, enum dma_data_direction dir,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700450 unsigned long attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400451{
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000452 xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400453}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400454
455/*
456 * Make physical memory consistent for a single streaming mode DMA translation
457 * after a transfer.
458 *
459 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
460 * using the cpu, yet do not wish to teardown the dma mapping, you must
461 * call this function before doing so. At the next point you give the dma
462 * address back to the card, you must first perform a
463 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
464 */
465static void
466xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
467 size_t size, enum dma_data_direction dir,
468 enum dma_sync_target target)
469{
470 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
471
472 BUG_ON(dir == DMA_NONE);
473
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000474 if (target == SYNC_FOR_CPU)
Stefano Stabellinid6883e62014-11-21 11:09:39 +0000475 xen_dma_sync_single_for_cpu(hwdev, dev_addr, size, dir);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000476
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400477 /* NOTE: We use dev_addr here, not paddr! */
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000478 if (is_xen_swiotlb_buffer(dev_addr))
Alexander Duyckfbfda892012-10-15 10:19:49 -0700479 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000480
481 if (target == SYNC_FOR_DEVICE)
Stefano Stabellini9490c6c2014-11-21 16:55:12 +0000482 xen_dma_sync_single_for_device(hwdev, dev_addr, size, dir);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400483}
484
485void
486xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
487 size_t size, enum dma_data_direction dir)
488{
489 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
490}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400491
492void
493xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
494 size_t size, enum dma_data_direction dir)
495{
496 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
497}
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200498
499/*
500 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
501 * concerning calls here are the same as for swiotlb_unmap_page() above.
502 */
503static void
504xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
505 int nelems, enum dma_data_direction dir,
506 unsigned long attrs)
507{
508 struct scatterlist *sg;
509 int i;
510
511 BUG_ON(dir == DMA_NONE);
512
513 for_each_sg(sgl, sg, nelems, i)
514 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
515
516}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400517
518/*
519 * Map a set of buffers described by scatterlist in streaming mode for DMA.
520 * This is the scatter-gather version of the above xen_swiotlb_map_page
521 * interface. Here the scatter gather list elements are each tagged with the
522 * appropriate dma address and length. They are obtained via
523 * sg_dma_{address,length}(SG).
524 *
525 * NOTE: An implementation may be able to use a smaller number of
526 * DMA address/length pairs than there are SG table elements.
527 * (for example via virtual mapping capabilities)
528 * The routine returns the number of addr/length pairs actually
529 * used, at most nents.
530 *
531 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
532 * same here.
533 */
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200534static int
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400535xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
536 int nelems, enum dma_data_direction dir,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700537 unsigned long attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400538{
539 struct scatterlist *sg;
540 int i;
541
542 BUG_ON(dir == DMA_NONE);
543
544 for_each_sg(sgl, sg, nelems, i) {
545 phys_addr_t paddr = sg_phys(sg);
546 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
547
Geert Uytterhoevenae7871b2016-12-16 14:28:41 +0100548 if (swiotlb_force == SWIOTLB_FORCE ||
Julien Grall291be102015-09-09 15:17:33 +0100549 xen_arch_need_swiotlb(hwdev, paddr, dev_addr) ||
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400550 !dma_capable(hwdev, dev_addr, sg->length) ||
551 range_straddles_page_boundary(paddr, sg->length)) {
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700552 phys_addr_t map = swiotlb_tbl_map_single(hwdev,
553 start_dma_addr,
554 sg_phys(sg),
555 sg->length,
Alexander Duyck0443fa02016-11-02 07:13:02 -0400556 dir, attrs);
Christoph Hellwigb907e202018-12-03 11:42:52 +0100557 if (map == DMA_MAPPING_ERROR) {
Stefano Stabellini783d0282013-10-25 10:33:26 +0000558 dev_warn(hwdev, "swiotlb buffer is full\n");
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400559 /* Don't panic here, we expect map_sg users
560 to do proper error handling. */
Alexander Duyck0443fa02016-11-02 07:13:02 -0400561 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400562 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
563 attrs);
Stefano Stabellini781575c2013-08-05 17:30:48 +0100564 sg_dma_len(sgl) = 0;
Stefano Stabellini15177602013-10-29 00:37:37 +0000565 return 0;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400566 }
Stefano Stabellinif1225ee2017-01-19 10:39:09 -0800567 dev_addr = xen_phys_to_bus(map);
Stefano Stabellini71bfae92013-11-15 14:22:15 +0000568 xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000569 dev_addr,
Stefano Stabellini71bfae92013-11-15 14:22:15 +0000570 map & ~PAGE_MASK,
571 sg->length,
572 dir,
573 attrs);
Stefano Stabellinif1225ee2017-01-19 10:39:09 -0800574 sg->dma_address = dev_addr;
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000575 } else {
576 /* we are not interested in the dma_addr returned by
577 * xen_dma_map_page, only in the potential cache flushes executed
578 * by the function. */
579 xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000580 dev_addr,
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000581 paddr & ~PAGE_MASK,
582 sg->length,
583 dir,
584 attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400585 sg->dma_address = dev_addr;
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000586 }
Stefano Stabellini781575c2013-08-05 17:30:48 +0100587 sg_dma_len(sg) = sg->length;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400588 }
589 return nelems;
590}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400591
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400592/*
593 * Make physical memory consistent for a set of streaming mode DMA translations
594 * after a transfer.
595 *
596 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
597 * and usage.
598 */
599static void
600xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
601 int nelems, enum dma_data_direction dir,
602 enum dma_sync_target target)
603{
604 struct scatterlist *sg;
605 int i;
606
607 for_each_sg(sgl, sg, nelems, i)
608 xen_swiotlb_sync_single(hwdev, sg->dma_address,
Stefano Stabellini781575c2013-08-05 17:30:48 +0100609 sg_dma_len(sg), dir, target);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400610}
611
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200612static void
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400613xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
614 int nelems, enum dma_data_direction dir)
615{
616 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
617}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400618
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200619static void
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400620xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
621 int nelems, enum dma_data_direction dir)
622{
623 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
624}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400625
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400626/*
627 * Return whether the given device DMA address mask can be supported
628 * properly. For example, if your device can only drive the low 24-bits
629 * during bus mastering, then you would pass 0x00ffffff as the mask to
630 * this function.
631 */
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200632static int
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400633xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
634{
635 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
636}
Stefano Stabellinieb1ddc02013-10-09 16:56:33 +0000637
Stefano Stabellini7e91c7d2017-02-07 19:58:02 +0200638/*
639 * Create userspace mapping for the DMA-coherent memory.
640 * This function should be called with the pages from the current domain only,
641 * passing pages mapped from other domains would lead to memory corruption.
642 */
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200643static int
Stefano Stabellini7e91c7d2017-02-07 19:58:02 +0200644xen_swiotlb_dma_mmap(struct device *dev, struct vm_area_struct *vma,
645 void *cpu_addr, dma_addr_t dma_addr, size_t size,
646 unsigned long attrs)
647{
Christoph Hellwig60d8cd52019-01-16 19:01:48 +0100648#ifdef CONFIG_ARM
Stefano Stabellinid5ff5062017-04-13 14:04:22 -0700649 if (xen_get_dma_ops(dev)->mmap)
650 return xen_get_dma_ops(dev)->mmap(dev, vma, cpu_addr,
Stefano Stabellini7e91c7d2017-02-07 19:58:02 +0200651 dma_addr, size, attrs);
652#endif
Christoph Hellwig58b04402018-09-11 08:55:28 +0200653 return dma_common_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
Stefano Stabellini7e91c7d2017-02-07 19:58:02 +0200654}
Andrii Anisov69369f52017-02-07 19:58:03 +0200655
656/*
657 * This function should be called with the pages from the current domain only,
658 * passing pages mapped from other domains would lead to memory corruption.
659 */
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200660static int
Andrii Anisov69369f52017-02-07 19:58:03 +0200661xen_swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
662 void *cpu_addr, dma_addr_t handle, size_t size,
663 unsigned long attrs)
664{
Christoph Hellwig60d8cd52019-01-16 19:01:48 +0100665#ifdef CONFIG_ARM
Stefano Stabellinid5ff5062017-04-13 14:04:22 -0700666 if (xen_get_dma_ops(dev)->get_sgtable) {
Andrii Anisov69369f52017-02-07 19:58:03 +0200667#if 0
668 /*
669 * This check verifies that the page belongs to the current domain and
670 * is not one mapped from another domain.
671 * This check is for debug only, and should not go to production build
672 */
673 unsigned long bfn = PHYS_PFN(dma_to_phys(dev, handle));
674 BUG_ON (!page_is_ram(bfn));
675#endif
Stefano Stabellinid5ff5062017-04-13 14:04:22 -0700676 return xen_get_dma_ops(dev)->get_sgtable(dev, sgt, cpu_addr,
Andrii Anisov69369f52017-02-07 19:58:03 +0200677 handle, size, attrs);
678 }
679#endif
Christoph Hellwig9406a492018-08-23 09:39:38 +0200680 return dma_common_get_sgtable(dev, sgt, cpu_addr, handle, size, attrs);
Andrii Anisov69369f52017-02-07 19:58:03 +0200681}
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200682
683const struct dma_map_ops xen_swiotlb_dma_ops = {
684 .alloc = xen_swiotlb_alloc_coherent,
685 .free = xen_swiotlb_free_coherent,
686 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
687 .sync_single_for_device = xen_swiotlb_sync_single_for_device,
688 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
689 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
690 .map_sg = xen_swiotlb_map_sg_attrs,
691 .unmap_sg = xen_swiotlb_unmap_sg_attrs,
692 .map_page = xen_swiotlb_map_page,
693 .unmap_page = xen_swiotlb_unmap_page,
694 .dma_supported = xen_swiotlb_dma_supported,
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200695 .mmap = xen_swiotlb_dma_mmap,
696 .get_sgtable = xen_swiotlb_get_sgtable,
697};