blob: 877baf2a94f4070a345923e05467cf9e320a8b2c [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 */
Mike Rapoport8a7f97b2019-03-11 23:30:31 -0700217 if (early) {
Mike Rapoport15c3c112018-10-30 15:08:58 -0700218 xen_io_tlb_start = memblock_alloc(PAGE_ALIGN(bytes),
219 PAGE_SIZE);
Mike Rapoport8a7f97b2019-03-11 23:30:31 -0700220 if (!xen_io_tlb_start)
221 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
222 __func__, PAGE_ALIGN(bytes), PAGE_SIZE);
223 } else {
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400224#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
225#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
226 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
Stefano Stabellini87465152015-04-24 10:16:40 +0100227 xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400228 if (xen_io_tlb_start)
229 break;
230 order--;
231 }
232 if (order != get_order(bytes)) {
Joe Perches283c0972013-06-28 03:21:41 -0700233 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
234 (PAGE_SIZE << order) >> 20);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400235 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
236 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
237 }
238 }
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400239 if (!xen_io_tlb_start) {
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400240 m_ret = XEN_SWIOTLB_ENOMEM;
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400241 goto error;
242 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400243 xen_io_tlb_end = xen_io_tlb_start + bytes;
244 /*
245 * And replace that memory with pages under 4GB.
246 */
247 rc = xen_swiotlb_fixup(xen_io_tlb_start,
248 bytes,
249 xen_io_tlb_nslabs);
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400250 if (rc) {
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400251 if (early)
Mike Rapoport20132882018-10-30 15:09:21 -0700252 memblock_free(__pa(xen_io_tlb_start),
253 PAGE_ALIGN(bytes));
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400254 else {
255 free_pages((unsigned long)xen_io_tlb_start, order);
256 xen_io_tlb_start = NULL;
257 }
Konrad Rzeszutek Wilk5bab7862012-08-23 14:03:55 -0400258 m_ret = XEN_SWIOTLB_EFIXUP;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400259 goto error;
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400260 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400261 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
Konrad Rzeszutek Wilkc468bde2012-09-17 10:20:09 -0400262 if (early) {
Yinghai Luac2cbab2013-01-24 12:20:16 -0800263 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
264 verbose))
265 panic("Cannot allocate SWIOTLB buffer");
Konrad Rzeszutek Wilkc468bde2012-09-17 10:20:09 -0400266 rc = 0;
267 } else
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400268 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
Konrad Rzeszutek Wilk7453c542016-12-20 10:02:02 -0500269
270 if (!rc)
271 swiotlb_set_max_segment(PAGE_SIZE);
272
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400273 return rc;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400274error:
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400275 if (repeat--) {
276 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
277 (xen_io_tlb_nslabs >> 1));
Joe Perches283c0972013-06-28 03:21:41 -0700278 pr_info("Lowering to %luMB\n",
279 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
Konrad Rzeszutek Wilkf4b2f072011-07-22 12:46:43 -0400280 goto retry;
281 }
Joe Perches283c0972013-06-28 03:21:41 -0700282 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
Konrad Rzeszutek Wilkb8277602012-08-23 14:36:15 -0400283 if (early)
284 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
285 else
286 free_pages((unsigned long)xen_io_tlb_start, order);
287 return rc;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400288}
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200289
290static void *
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400291xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
Andrzej Pietrasiewiczbaa676f2012-03-27 14:28:18 +0200292 dma_addr_t *dma_handle, gfp_t flags,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700293 unsigned long attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400294{
295 void *ret;
296 int order = get_order(size);
297 u64 dma_mask = DMA_BIT_MASK(32);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400298 phys_addr_t phys;
299 dma_addr_t dev_addr;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400300
301 /*
302 * Ignore region specifiers - the kernel's ideas of
303 * pseudo-phys memory layout has nothing to do with the
304 * machine physical layout. We can't allocate highmem
305 * because we can't return a pointer to it.
306 */
307 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
308
Joe Jin7250f422018-10-16 15:21:16 -0700309 /* Convert the size to actually allocated. */
310 size = 1UL << (order + XEN_PAGE_SHIFT);
311
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000312 /* On ARM this function returns an ioremap'ped virtual address for
313 * which virt_to_phys doesn't return the corresponding physical
314 * address. In fact on ARM virt_to_phys only works for kernel direct
315 * mapped RAM memory. Also see comment below.
316 */
317 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400318
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400319 if (!ret)
320 return ret;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400321
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400322 if (hwdev && hwdev->coherent_dma_mask)
Christoph Hellwig038d07a2018-03-19 11:38:14 +0100323 dma_mask = hwdev->coherent_dma_mask;
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400324
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000325 /* At this point dma_handle is the physical address, next we are
326 * going to set it to the machine address.
327 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
328 * to *dma_handle. */
329 phys = *dma_handle;
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400330 dev_addr = xen_phys_to_bus(phys);
331 if (((dev_addr + size - 1 <= dma_mask)) &&
332 !range_straddles_page_boundary(phys, size))
333 *dma_handle = dev_addr;
334 else {
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000335 if (xen_create_contiguous_region(phys, order,
Stefano Stabellini69908902013-10-09 16:56:32 +0000336 fls64(dma_mask), dma_handle) != 0) {
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000337 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400338 return NULL;
339 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400340 }
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400341 memset(ret, 0, size);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400342 return ret;
343}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400344
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200345static void
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400346xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700347 dma_addr_t dev_addr, unsigned long attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400348{
349 int order = get_order(size);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400350 phys_addr_t phys;
351 u64 dma_mask = DMA_BIT_MASK(32);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400352
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400353 if (hwdev && hwdev->coherent_dma_mask)
354 dma_mask = hwdev->coherent_dma_mask;
355
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000356 /* do not use virt_to_phys because on ARM it doesn't return you the
357 * physical address */
358 phys = xen_bus_to_phys(dev_addr);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400359
Joe Jin7250f422018-10-16 15:21:16 -0700360 /* Convert the size to actually allocated. */
361 size = 1UL << (order + XEN_PAGE_SHIFT);
362
Joe Jin4855c922018-05-17 12:33:28 -0700363 if (((dev_addr + size - 1 <= dma_mask)) ||
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400364 range_straddles_page_boundary(phys, size))
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000365 xen_destroy_contiguous_region(phys, order);
Konrad Rzeszutek Wilk6810df82011-08-25 16:13:54 -0400366
Stefano Stabellini1b65c4e2013-10-10 13:41:10 +0000367 xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400368}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400369
370/*
371 * Map a single buffer of the indicated size for DMA in streaming mode. The
372 * physical address to use is returned.
373 *
374 * Once the device is given the dma address, the device owns this memory until
375 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
376 */
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200377static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400378 unsigned long offset, size_t size,
379 enum dma_data_direction dir,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700380 unsigned long attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400381{
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700382 phys_addr_t map, phys = page_to_phys(page) + offset;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400383 dma_addr_t dev_addr = xen_phys_to_bus(phys);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400384
385 BUG_ON(dir == DMA_NONE);
386 /*
387 * If the address happens to be in the device's DMA window,
388 * we can safely return the device addr and not worry about bounce
389 * buffering it.
390 */
391 if (dma_capable(dev, dev_addr, size) &&
Stefano Stabellinia4dba132014-11-21 11:07:39 +0000392 !range_straddles_page_boundary(phys, size) &&
Julien Grall291be102015-09-09 15:17:33 +0100393 !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
Geert Uytterhoevenae7871b2016-12-16 14:28:41 +0100394 (swiotlb_force != SWIOTLB_FORCE)) {
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000395 /* we are not interested in the dma_addr returned by
396 * xen_dma_map_page, only in the potential cache flushes executed
397 * by the function. */
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000398 xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400399 return dev_addr;
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000400 }
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400401
402 /*
403 * Oh well, have to allocate and map a bounce buffer.
404 */
Zoltan Kiss2b2b6142013-09-04 21:11:05 +0100405 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
406
Alexander Duyck0443fa02016-11-02 07:13:02 -0400407 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir,
408 attrs);
Christoph Hellwigb907e202018-12-03 11:42:52 +0100409 if (map == DMA_MAPPING_ERROR)
Christoph Hellwiga4abe0a2018-11-21 19:38:19 +0100410 return DMA_MAPPING_ERROR;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400411
Stefano Stabellinif1225ee2017-01-19 10:39:09 -0800412 dev_addr = xen_phys_to_bus(map);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000413 xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000414 dev_addr, map & ~PAGE_MASK, size, dir, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400415
416 /*
417 * Ensure that the address returned is DMA'ble
418 */
Alexander Duyck76418422016-11-02 07:12:47 -0400419 if (dma_capable(dev, dev_addr, size))
420 return dev_addr;
421
Alexander Duyckd29fa0c2016-11-10 07:05:31 -0500422 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
423 swiotlb_tbl_unmap_single(dev, map, size, dir, attrs);
Alexander Duyck76418422016-11-02 07:12:47 -0400424
Christoph Hellwiga4abe0a2018-11-21 19:38:19 +0100425 return DMA_MAPPING_ERROR;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400426}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400427
428/*
429 * Unmap a single streaming mode DMA translation. The dma_addr and size must
430 * match what was provided for in a previous xen_swiotlb_map_page call. All
431 * other usages are undefined.
432 *
433 * After this call, reads by the cpu to the buffer are guaranteed to see
434 * whatever the device wrote there.
435 */
436static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000437 size_t size, enum dma_data_direction dir,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700438 unsigned long attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400439{
440 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
441
442 BUG_ON(dir == DMA_NONE);
443
Stefano Stabellinid6883e62014-11-21 11:09:39 +0000444 xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000445
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400446 /* NOTE: We use dev_addr here, not paddr! */
Christoph Hellwig68c60832018-12-06 07:06:04 -0800447 if (is_xen_swiotlb_buffer(dev_addr))
Alexander Duyck0443fa02016-11-02 07:13:02 -0400448 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400449}
450
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200451static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400452 size_t size, enum dma_data_direction dir,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700453 unsigned long attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400454{
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000455 xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400456}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400457
458/*
459 * Make physical memory consistent for a single streaming mode DMA translation
460 * after a transfer.
461 *
462 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
463 * using the cpu, yet do not wish to teardown the dma mapping, you must
464 * call this function before doing so. At the next point you give the dma
465 * address back to the card, you must first perform a
466 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
467 */
468static void
469xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
470 size_t size, enum dma_data_direction dir,
471 enum dma_sync_target target)
472{
473 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
474
475 BUG_ON(dir == DMA_NONE);
476
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000477 if (target == SYNC_FOR_CPU)
Stefano Stabellinid6883e62014-11-21 11:09:39 +0000478 xen_dma_sync_single_for_cpu(hwdev, dev_addr, size, dir);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000479
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400480 /* NOTE: We use dev_addr here, not paddr! */
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000481 if (is_xen_swiotlb_buffer(dev_addr))
Alexander Duyckfbfda892012-10-15 10:19:49 -0700482 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000483
484 if (target == SYNC_FOR_DEVICE)
Stefano Stabellini9490c6c2014-11-21 16:55:12 +0000485 xen_dma_sync_single_for_device(hwdev, dev_addr, size, dir);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400486}
487
488void
489xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
490 size_t size, enum dma_data_direction dir)
491{
492 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
493}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400494
495void
496xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
497 size_t size, enum dma_data_direction dir)
498{
499 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
500}
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200501
502/*
503 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
504 * concerning calls here are the same as for swiotlb_unmap_page() above.
505 */
506static void
507xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
508 int nelems, enum dma_data_direction dir,
509 unsigned long attrs)
510{
511 struct scatterlist *sg;
512 int i;
513
514 BUG_ON(dir == DMA_NONE);
515
516 for_each_sg(sgl, sg, nelems, i)
517 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
518
519}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400520
521/*
522 * Map a set of buffers described by scatterlist in streaming mode for DMA.
523 * This is the scatter-gather version of the above xen_swiotlb_map_page
524 * interface. Here the scatter gather list elements are each tagged with the
525 * appropriate dma address and length. They are obtained via
526 * sg_dma_{address,length}(SG).
527 *
528 * NOTE: An implementation may be able to use a smaller number of
529 * DMA address/length pairs than there are SG table elements.
530 * (for example via virtual mapping capabilities)
531 * The routine returns the number of addr/length pairs actually
532 * used, at most nents.
533 *
534 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
535 * same here.
536 */
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200537static int
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400538xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
539 int nelems, enum dma_data_direction dir,
Krzysztof Kozlowski00085f12016-08-03 13:46:00 -0700540 unsigned long attrs)
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400541{
542 struct scatterlist *sg;
543 int i;
544
545 BUG_ON(dir == DMA_NONE);
546
547 for_each_sg(sgl, sg, nelems, i) {
548 phys_addr_t paddr = sg_phys(sg);
549 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
550
Geert Uytterhoevenae7871b2016-12-16 14:28:41 +0100551 if (swiotlb_force == SWIOTLB_FORCE ||
Julien Grall291be102015-09-09 15:17:33 +0100552 xen_arch_need_swiotlb(hwdev, paddr, dev_addr) ||
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400553 !dma_capable(hwdev, dev_addr, sg->length) ||
554 range_straddles_page_boundary(paddr, sg->length)) {
Alexander Duycke05ed4d2012-10-15 10:19:39 -0700555 phys_addr_t map = swiotlb_tbl_map_single(hwdev,
556 start_dma_addr,
557 sg_phys(sg),
558 sg->length,
Alexander Duyck0443fa02016-11-02 07:13:02 -0400559 dir, attrs);
Christoph Hellwigb907e202018-12-03 11:42:52 +0100560 if (map == DMA_MAPPING_ERROR) {
Stefano Stabellini783d0282013-10-25 10:33:26 +0000561 dev_warn(hwdev, "swiotlb buffer is full\n");
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400562 /* Don't panic here, we expect map_sg users
563 to do proper error handling. */
Alexander Duyck0443fa02016-11-02 07:13:02 -0400564 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400565 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
566 attrs);
Stefano Stabellini781575c2013-08-05 17:30:48 +0100567 sg_dma_len(sgl) = 0;
Stefano Stabellini15177602013-10-29 00:37:37 +0000568 return 0;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400569 }
Stefano Stabellinif1225ee2017-01-19 10:39:09 -0800570 dev_addr = xen_phys_to_bus(map);
Stefano Stabellini71bfae92013-11-15 14:22:15 +0000571 xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000572 dev_addr,
Stefano Stabellini71bfae92013-11-15 14:22:15 +0000573 map & ~PAGE_MASK,
574 sg->length,
575 dir,
576 attrs);
Stefano Stabellinif1225ee2017-01-19 10:39:09 -0800577 sg->dma_address = dev_addr;
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000578 } else {
579 /* we are not interested in the dma_addr returned by
580 * xen_dma_map_page, only in the potential cache flushes executed
581 * by the function. */
582 xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
Stefano Stabellinia0f2dee2014-11-21 11:04:39 +0000583 dev_addr,
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000584 paddr & ~PAGE_MASK,
585 sg->length,
586 dir,
587 attrs);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400588 sg->dma_address = dev_addr;
Stefano Stabellini6cf05462013-10-25 10:33:25 +0000589 }
Stefano Stabellini781575c2013-08-05 17:30:48 +0100590 sg_dma_len(sg) = sg->length;
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400591 }
592 return nelems;
593}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400594
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400595/*
596 * Make physical memory consistent for a set of streaming mode DMA translations
597 * after a transfer.
598 *
599 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
600 * and usage.
601 */
602static void
603xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
604 int nelems, enum dma_data_direction dir,
605 enum dma_sync_target target)
606{
607 struct scatterlist *sg;
608 int i;
609
610 for_each_sg(sgl, sg, nelems, i)
611 xen_swiotlb_sync_single(hwdev, sg->dma_address,
Stefano Stabellini781575c2013-08-05 17:30:48 +0100612 sg_dma_len(sg), dir, target);
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400613}
614
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200615static void
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400616xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
617 int nelems, enum dma_data_direction dir)
618{
619 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
620}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400621
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200622static void
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400623xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
624 int nelems, enum dma_data_direction dir)
625{
626 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
627}
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400628
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400629/*
630 * Return whether the given device DMA address mask can be supported
631 * properly. For example, if your device can only drive the low 24-bits
632 * during bus mastering, then you would pass 0x00ffffff as the mask to
633 * this function.
634 */
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200635static int
Konrad Rzeszutek Wilkb097186f2010-05-11 10:05:49 -0400636xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
637{
638 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
639}
Stefano Stabellinieb1ddc02013-10-09 16:56:33 +0000640
Stefano Stabellini7e91c7d2017-02-07 19:58:02 +0200641/*
642 * Create userspace mapping for the DMA-coherent memory.
643 * This function should be called with the pages from the current domain only,
644 * passing pages mapped from other domains would lead to memory corruption.
645 */
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200646static int
Stefano Stabellini7e91c7d2017-02-07 19:58:02 +0200647xen_swiotlb_dma_mmap(struct device *dev, struct vm_area_struct *vma,
648 void *cpu_addr, dma_addr_t dma_addr, size_t size,
649 unsigned long attrs)
650{
Christoph Hellwig60d8cd52019-01-16 19:01:48 +0100651#ifdef CONFIG_ARM
Stefano Stabellinid5ff5062017-04-13 14:04:22 -0700652 if (xen_get_dma_ops(dev)->mmap)
653 return xen_get_dma_ops(dev)->mmap(dev, vma, cpu_addr,
Stefano Stabellini7e91c7d2017-02-07 19:58:02 +0200654 dma_addr, size, attrs);
655#endif
Christoph Hellwig58b04402018-09-11 08:55:28 +0200656 return dma_common_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
Stefano Stabellini7e91c7d2017-02-07 19:58:02 +0200657}
Andrii Anisov69369f52017-02-07 19:58:03 +0200658
659/*
660 * This function should be called with the pages from the current domain only,
661 * passing pages mapped from other domains would lead to memory corruption.
662 */
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200663static int
Andrii Anisov69369f52017-02-07 19:58:03 +0200664xen_swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
665 void *cpu_addr, dma_addr_t handle, size_t size,
666 unsigned long attrs)
667{
Christoph Hellwig60d8cd52019-01-16 19:01:48 +0100668#ifdef CONFIG_ARM
Stefano Stabellinid5ff5062017-04-13 14:04:22 -0700669 if (xen_get_dma_ops(dev)->get_sgtable) {
Andrii Anisov69369f52017-02-07 19:58:03 +0200670#if 0
671 /*
672 * This check verifies that the page belongs to the current domain and
673 * is not one mapped from another domain.
674 * This check is for debug only, and should not go to production build
675 */
676 unsigned long bfn = PHYS_PFN(dma_to_phys(dev, handle));
677 BUG_ON (!page_is_ram(bfn));
678#endif
Stefano Stabellinid5ff5062017-04-13 14:04:22 -0700679 return xen_get_dma_ops(dev)->get_sgtable(dev, sgt, cpu_addr,
Andrii Anisov69369f52017-02-07 19:58:03 +0200680 handle, size, attrs);
681 }
682#endif
Christoph Hellwig9406a492018-08-23 09:39:38 +0200683 return dma_common_get_sgtable(dev, sgt, cpu_addr, handle, size, attrs);
Andrii Anisov69369f52017-02-07 19:58:03 +0200684}
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200685
686const struct dma_map_ops xen_swiotlb_dma_ops = {
687 .alloc = xen_swiotlb_alloc_coherent,
688 .free = xen_swiotlb_free_coherent,
689 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
690 .sync_single_for_device = xen_swiotlb_sync_single_for_device,
691 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
692 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
693 .map_sg = xen_swiotlb_map_sg_attrs,
694 .unmap_sg = xen_swiotlb_unmap_sg_attrs,
695 .map_page = xen_swiotlb_map_page,
696 .unmap_page = xen_swiotlb_unmap_page,
697 .dma_supported = xen_swiotlb_dma_supported,
Christoph Hellwigdceb1a62017-05-21 13:15:13 +0200698 .mmap = xen_swiotlb_dma_mmap,
699 .get_sgtable = xen_swiotlb_get_sgtable,
700};