2 * Dynamic DMA mapping support.
5 #include <linux/types.h>
7 #include <linux/string.h>
9 #include <linux/module.h>
11 #include <asm/proto.h>
12 #include <asm/calgary.h>
14 int iommu_merge __read_mostly = 0;
15 EXPORT_SYMBOL(iommu_merge);
17 dma_addr_t bad_dma_address __read_mostly;
18 EXPORT_SYMBOL(bad_dma_address);
20 /* This tells the BIO block layer to assume merging. Default to off
21 because we cannot guarantee merging later. */
22 int iommu_bio_merge __read_mostly = 0;
23 EXPORT_SYMBOL(iommu_bio_merge);
25 int iommu_sac_force __read_mostly = 0;
26 EXPORT_SYMBOL(iommu_sac_force);
28 int no_iommu __read_mostly;
29 #ifdef CONFIG_IOMMU_DEBUG
30 int panic_on_overflow __read_mostly = 1;
31 int force_iommu __read_mostly = 1;
33 int panic_on_overflow __read_mostly = 0;
34 int force_iommu __read_mostly= 0;
37 /* Set this to 1 if there is a HW IOMMU in the system */
38 int iommu_detected __read_mostly = 0;
40 /* Dummy device used for NULL arguments (normally ISA). Better would
41 be probably a smaller DMA mask, but this is bug-to-bug compatible
43 struct device fallback_dev = {
44 .bus_id = "fallback device",
45 .coherent_dma_mask = DMA_32BIT_MASK,
46 .dma_mask = &fallback_dev.coherent_dma_mask,
49 /* Allocate DMA memory on node near device */
50 noinline static void *
51 dma_alloc_pages(struct device *dev, gfp_t gfp, unsigned order)
56 if (dev->bus == &pci_bus_type)
57 node = pcibus_to_node(to_pci_dev(dev)->bus);
60 node = numa_node_id();
62 if (node < first_node(node_online_map))
63 node = first_node(node_online_map);
65 page = alloc_pages_node(node, gfp, order);
66 return page ? page_address(page) : NULL;
70 * Allocate memory for a coherent mapping.
73 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle,
77 unsigned long dma_mask = 0;
82 dma_mask = dev->coherent_dma_mask;
84 dma_mask = DMA_32BIT_MASK;
86 /* Don't invoke OOM killer */
89 /* Kludge to make it bug-to-bug compatible with i386. i386
90 uses the normal dma_mask for alloc_coherent. */
91 dma_mask &= *dev->dma_mask;
93 /* Why <=? Even when the mask is smaller than 4GB it is often
94 larger than 16MB and in this case we have a chance of
95 finding fitting memory in the next higher zone first. If
96 not retry with true GFP_DMA. -AK */
97 if (dma_mask <= DMA_32BIT_MASK)
101 memory = dma_alloc_pages(dev, gfp, get_order(size));
107 bus = virt_to_bus(memory);
108 high = (bus + size) >= dma_mask;
110 if (force_iommu && !(gfp & GFP_DMA))
113 free_pages((unsigned long)memory,
116 /* Don't use the 16MB ZONE_DMA unless absolutely
117 needed. It's better to use remapping first. */
118 if (dma_mask < DMA_32BIT_MASK && !(gfp & GFP_DMA)) {
119 gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
123 /* Let low level make its own zone decisions */
124 gfp &= ~(GFP_DMA32|GFP_DMA);
126 if (dma_ops->alloc_coherent)
127 return dma_ops->alloc_coherent(dev, size,
132 memset(memory, 0, size);
134 *dma_handle = virt_to_bus(memory);
139 if (dma_ops->alloc_coherent) {
140 free_pages((unsigned long)memory, get_order(size));
141 gfp &= ~(GFP_DMA|GFP_DMA32);
142 return dma_ops->alloc_coherent(dev, size, dma_handle, gfp);
145 if (dma_ops->map_simple) {
146 *dma_handle = dma_ops->map_simple(dev, memory,
148 PCI_DMA_BIDIRECTIONAL);
149 if (*dma_handle != bad_dma_address)
153 if (panic_on_overflow)
154 panic("dma_alloc_coherent: IOMMU overflow by %lu bytes\n",size);
155 free_pages((unsigned long)memory, get_order(size));
158 EXPORT_SYMBOL(dma_alloc_coherent);
161 * Unmap coherent memory.
162 * The caller must ensure that the device has finished accessing the mapping.
164 void dma_free_coherent(struct device *dev, size_t size,
165 void *vaddr, dma_addr_t bus)
167 if (dma_ops->unmap_single)
168 dma_ops->unmap_single(dev, bus, size, 0);
169 free_pages((unsigned long)vaddr, get_order(size));
171 EXPORT_SYMBOL(dma_free_coherent);
173 int dma_supported(struct device *dev, u64 mask)
175 if (dma_ops->dma_supported)
176 return dma_ops->dma_supported(dev, mask);
178 /* Copied from i386. Doesn't make much sense, because it will
179 only work for pci_alloc_coherent.
180 The caller just has to use GFP_DMA in this case. */
181 if (mask < DMA_24BIT_MASK)
184 /* Tell the device to use SAC when IOMMU force is on. This
185 allows the driver to use cheaper accesses in some cases.
187 Problem with this is that if we overflow the IOMMU area and
188 return DAC as fallback address the device may not handle it
191 As a special case some controllers have a 39bit address
192 mode that is as efficient as 32bit (aic79xx). Don't force
193 SAC for these. Assume all masks <= 40 bits are of this
194 type. Normally this doesn't make any difference, but gives
195 more gentle handling of IOMMU overflow. */
196 if (iommu_sac_force && (mask >= DMA_40BIT_MASK)) {
197 printk(KERN_INFO "%s: Force SAC with mask %Lx\n", dev->bus_id,mask);
203 EXPORT_SYMBOL(dma_supported);
205 int dma_set_mask(struct device *dev, u64 mask)
207 if (!dev->dma_mask || !dma_supported(dev, mask))
209 *dev->dma_mask = mask;
212 EXPORT_SYMBOL(dma_set_mask);
214 /* iommu=[size][,noagp][,off][,force][,noforce][,leak][,memaper[=order]][,merge]
215 [,forcesac][,fullflush][,nomerge][,biomerge]
216 size set size of iommu (in bytes)
217 noagp don't initialize the AGP driver and use full aperture.
218 off don't use the IOMMU
219 leak turn on simple iommu leak tracing (only when CONFIG_IOMMU_LEAK is on)
220 memaper[=order] allocate an own aperture over RAM with size 32MB^order.
221 noforce don't force IOMMU usage. Default.
223 merge Do lazy merging. This may improve performance on some block devices.
224 Implies force (experimental)
225 biomerge Do merging at the BIO layer. This is more efficient than merge,
226 but should be only done with very big IOMMUs. Implies merge,force.
227 nomerge Don't do SG merging.
228 forcesac For SAC mode for masks <40bits (experimental)
229 fullflush Flush IOMMU on each allocation (default)
230 nofullflush Don't use IOMMU fullflush
231 allowed overwrite iommu off workarounds for specific chipsets.
232 soft Use software bounce buffering (default for Intel machines)
233 noaperture Don't touch the aperture for AGP.
235 __init int iommu_setup(char *p)
243 if (!strncmp(p,"off",3))
245 /* gart_parse_options has more force support */
246 if (!strncmp(p,"force",5))
248 if (!strncmp(p,"noforce",7)) {
253 if (!strncmp(p, "biomerge",8)) {
254 iommu_bio_merge = 4096;
258 if (!strncmp(p, "panic",5))
259 panic_on_overflow = 1;
260 if (!strncmp(p, "nopanic",7))
261 panic_on_overflow = 0;
262 if (!strncmp(p, "merge",5)) {
266 if (!strncmp(p, "nomerge",7))
268 if (!strncmp(p, "forcesac",8))
271 #ifdef CONFIG_SWIOTLB
272 if (!strncmp(p, "soft",4))
277 gart_parse_options(p);
280 p += strcspn(p, ",");
286 early_param("iommu", iommu_setup);
288 void __init pci_iommu_alloc(void)
291 * The order of these functions is important for
292 * fall-back/fail-over reasons
298 #ifdef CONFIG_CALGARY_IOMMU
302 #ifdef CONFIG_SWIOTLB
307 static int __init pci_iommu_init(void)
309 #ifdef CONFIG_CALGARY_IOMMU
310 calgary_iommu_init();
321 /* Must execute after PCI subsystem */
322 fs_initcall(pci_iommu_init);