xref: /linux/arch/x86/kernel/pci-dma.c (revision 2fe05e1139a555ae91f00a812cb9520e7d3022ab)
1 #include <linux/dma-mapping.h>
2 #include <linux/dma-debug.h>
3 #include <linux/dmar.h>
4 #include <linux/export.h>
5 #include <linux/bootmem.h>
6 #include <linux/gfp.h>
7 #include <linux/pci.h>
8 #include <linux/kmemleak.h>
9 
10 #include <asm/proto.h>
11 #include <asm/dma.h>
12 #include <asm/iommu.h>
13 #include <asm/gart.h>
14 #include <asm/calgary.h>
15 #include <asm/x86_init.h>
16 #include <asm/iommu_table.h>
17 
18 static int forbid_dac __read_mostly;
19 
20 const struct dma_map_ops *dma_ops = &nommu_dma_ops;
21 EXPORT_SYMBOL(dma_ops);
22 
23 static int iommu_sac_force __read_mostly;
24 
25 #ifdef CONFIG_IOMMU_DEBUG
26 int panic_on_overflow __read_mostly = 1;
27 int force_iommu __read_mostly = 1;
28 #else
29 int panic_on_overflow __read_mostly = 0;
30 int force_iommu __read_mostly = 0;
31 #endif
32 
33 int iommu_merge __read_mostly = 0;
34 
35 int no_iommu __read_mostly;
36 /* Set this to 1 if there is a HW IOMMU in the system */
37 int iommu_detected __read_mostly = 0;
38 
39 /*
40  * This variable becomes 1 if iommu=pt is passed on the kernel command line.
41  * If this variable is 1, IOMMU implementations do no DMA translation for
42  * devices and allow every device to access to whole physical memory. This is
43  * useful if a user wants to use an IOMMU only for KVM device assignment to
44  * guests and not for driver dma translation.
45  */
46 int iommu_pass_through __read_mostly;
47 
48 extern struct iommu_table_entry __iommu_table[], __iommu_table_end[];
49 
50 /* Dummy device used for NULL arguments (normally ISA). */
51 struct device x86_dma_fallback_dev = {
52 	.init_name = "fallback device",
53 	.coherent_dma_mask = ISA_DMA_BIT_MASK,
54 	.dma_mask = &x86_dma_fallback_dev.coherent_dma_mask,
55 };
56 EXPORT_SYMBOL(x86_dma_fallback_dev);
57 
58 /* Number of entries preallocated for DMA-API debugging */
59 #define PREALLOC_DMA_DEBUG_ENTRIES       65536
60 
61 void __init pci_iommu_alloc(void)
62 {
63 	struct iommu_table_entry *p;
64 
65 	sort_iommu_table(__iommu_table, __iommu_table_end);
66 	check_iommu_entries(__iommu_table, __iommu_table_end);
67 
68 	for (p = __iommu_table; p < __iommu_table_end; p++) {
69 		if (p && p->detect && p->detect() > 0) {
70 			p->flags |= IOMMU_DETECTED;
71 			if (p->early_init)
72 				p->early_init();
73 			if (p->flags & IOMMU_FINISH_IF_DETECTED)
74 				break;
75 		}
76 	}
77 }
78 void *dma_generic_alloc_coherent(struct device *dev, size_t size,
79 				 dma_addr_t *dma_addr, gfp_t flag,
80 				 unsigned long attrs)
81 {
82 	unsigned long dma_mask;
83 	struct page *page;
84 	unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
85 	dma_addr_t addr;
86 
87 	dma_mask = dma_alloc_coherent_mask(dev, flag);
88 
89 	flag &= ~__GFP_ZERO;
90 again:
91 	page = NULL;
92 	/* CMA can be used only in the context which permits sleeping */
93 	if (gfpflags_allow_blocking(flag)) {
94 		page = dma_alloc_from_contiguous(dev, count, get_order(size),
95 						 flag);
96 		if (page && page_to_phys(page) + size > dma_mask) {
97 			dma_release_from_contiguous(dev, page, count);
98 			page = NULL;
99 		}
100 	}
101 	/* fallback */
102 	if (!page)
103 		page = alloc_pages_node(dev_to_node(dev), flag, get_order(size));
104 	if (!page)
105 		return NULL;
106 
107 	addr = page_to_phys(page);
108 	if (addr + size > dma_mask) {
109 		__free_pages(page, get_order(size));
110 
111 		if (dma_mask < DMA_BIT_MASK(32) && !(flag & GFP_DMA)) {
112 			flag = (flag & ~GFP_DMA32) | GFP_DMA;
113 			goto again;
114 		}
115 
116 		return NULL;
117 	}
118 	memset(page_address(page), 0, size);
119 	*dma_addr = addr;
120 	return page_address(page);
121 }
122 
123 void dma_generic_free_coherent(struct device *dev, size_t size, void *vaddr,
124 			       dma_addr_t dma_addr, unsigned long attrs)
125 {
126 	unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
127 	struct page *page = virt_to_page(vaddr);
128 
129 	if (!dma_release_from_contiguous(dev, page, count))
130 		free_pages((unsigned long)vaddr, get_order(size));
131 }
132 
133 bool arch_dma_alloc_attrs(struct device **dev, gfp_t *gfp)
134 {
135 	if (!*dev)
136 		*dev = &x86_dma_fallback_dev;
137 
138 	*gfp &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
139 	*gfp = dma_alloc_coherent_gfp_flags(*dev, *gfp);
140 
141 	if (!is_device_dma_capable(*dev))
142 		return false;
143 	return true;
144 
145 }
146 EXPORT_SYMBOL(arch_dma_alloc_attrs);
147 
148 /*
149  * See <Documentation/x86/x86_64/boot-options.txt> for the iommu kernel
150  * parameter documentation.
151  */
152 static __init int iommu_setup(char *p)
153 {
154 	iommu_merge = 1;
155 
156 	if (!p)
157 		return -EINVAL;
158 
159 	while (*p) {
160 		if (!strncmp(p, "off", 3))
161 			no_iommu = 1;
162 		/* gart_parse_options has more force support */
163 		if (!strncmp(p, "force", 5))
164 			force_iommu = 1;
165 		if (!strncmp(p, "noforce", 7)) {
166 			iommu_merge = 0;
167 			force_iommu = 0;
168 		}
169 
170 		if (!strncmp(p, "biomerge", 8)) {
171 			iommu_merge = 1;
172 			force_iommu = 1;
173 		}
174 		if (!strncmp(p, "panic", 5))
175 			panic_on_overflow = 1;
176 		if (!strncmp(p, "nopanic", 7))
177 			panic_on_overflow = 0;
178 		if (!strncmp(p, "merge", 5)) {
179 			iommu_merge = 1;
180 			force_iommu = 1;
181 		}
182 		if (!strncmp(p, "nomerge", 7))
183 			iommu_merge = 0;
184 		if (!strncmp(p, "forcesac", 8))
185 			iommu_sac_force = 1;
186 		if (!strncmp(p, "allowdac", 8))
187 			forbid_dac = 0;
188 		if (!strncmp(p, "nodac", 5))
189 			forbid_dac = 1;
190 		if (!strncmp(p, "usedac", 6)) {
191 			forbid_dac = -1;
192 			return 1;
193 		}
194 #ifdef CONFIG_SWIOTLB
195 		if (!strncmp(p, "soft", 4))
196 			swiotlb = 1;
197 #endif
198 		if (!strncmp(p, "pt", 2))
199 			iommu_pass_through = 1;
200 
201 		gart_parse_options(p);
202 
203 #ifdef CONFIG_CALGARY_IOMMU
204 		if (!strncmp(p, "calgary", 7))
205 			use_calgary = 1;
206 #endif /* CONFIG_CALGARY_IOMMU */
207 
208 		p += strcspn(p, ",");
209 		if (*p == ',')
210 			++p;
211 	}
212 	return 0;
213 }
214 early_param("iommu", iommu_setup);
215 
216 int x86_dma_supported(struct device *dev, u64 mask)
217 {
218 #ifdef CONFIG_PCI
219 	if (mask > 0xffffffff && forbid_dac > 0) {
220 		dev_info(dev, "PCI: Disallowing DAC for device\n");
221 		return 0;
222 	}
223 #endif
224 
225 	/* Copied from i386. Doesn't make much sense, because it will
226 	   only work for pci_alloc_coherent.
227 	   The caller just has to use GFP_DMA in this case. */
228 	if (mask < DMA_BIT_MASK(24))
229 		return 0;
230 
231 	/* Tell the device to use SAC when IOMMU force is on.  This
232 	   allows the driver to use cheaper accesses in some cases.
233 
234 	   Problem with this is that if we overflow the IOMMU area and
235 	   return DAC as fallback address the device may not handle it
236 	   correctly.
237 
238 	   As a special case some controllers have a 39bit address
239 	   mode that is as efficient as 32bit (aic79xx). Don't force
240 	   SAC for these.  Assume all masks <= 40 bits are of this
241 	   type. Normally this doesn't make any difference, but gives
242 	   more gentle handling of IOMMU overflow. */
243 	if (iommu_sac_force && (mask >= DMA_BIT_MASK(40))) {
244 		dev_info(dev, "Force SAC with mask %Lx\n", mask);
245 		return 0;
246 	}
247 
248 	return 1;
249 }
250 
251 static int __init pci_iommu_init(void)
252 {
253 	struct iommu_table_entry *p;
254 	dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
255 
256 #ifdef CONFIG_PCI
257 	dma_debug_add_bus(&pci_bus_type);
258 #endif
259 	x86_init.iommu.iommu_init();
260 
261 	for (p = __iommu_table; p < __iommu_table_end; p++) {
262 		if (p && (p->flags & IOMMU_DETECTED) && p->late_init)
263 			p->late_init();
264 	}
265 
266 	return 0;
267 }
268 /* Must execute after PCI subsystem */
269 rootfs_initcall(pci_iommu_init);
270 
271 #ifdef CONFIG_PCI
272 /* Many VIA bridges seem to corrupt data for DAC. Disable it here */
273 
274 static void via_no_dac(struct pci_dev *dev)
275 {
276 	if (forbid_dac == 0) {
277 		dev_info(&dev->dev, "disabling DAC on VIA PCI bridge\n");
278 		forbid_dac = 1;
279 	}
280 }
281 DECLARE_PCI_FIXUP_CLASS_FINAL(PCI_VENDOR_ID_VIA, PCI_ANY_ID,
282 				PCI_CLASS_BRIDGE_PCI, 8, via_no_dac);
283 #endif
284