1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * DMA operations that map physical memory directly without using an IOMMU or 4 * flushing caches. 5 */ 6 #include <linux/export.h> 7 #include <linux/mm.h> 8 #include <linux/dma-direct.h> 9 #include <linux/scatterlist.h> 10 #include <linux/dma-contiguous.h> 11 #include <linux/pfn.h> 12 #include <linux/set_memory.h> 13 14 #define DIRECT_MAPPING_ERROR 0 15 16 /* 17 * Most architectures use ZONE_DMA for the first 16 Megabytes, but 18 * some use it for entirely different regions: 19 */ 20 #ifndef ARCH_ZONE_DMA_BITS 21 #define ARCH_ZONE_DMA_BITS 24 22 #endif 23 24 /* 25 * For AMD SEV all DMA must be to unencrypted addresses. 26 */ 27 static inline bool force_dma_unencrypted(void) 28 { 29 return sev_active(); 30 } 31 32 static bool 33 check_addr(struct device *dev, dma_addr_t dma_addr, size_t size, 34 const char *caller) 35 { 36 if (unlikely(dev && !dma_capable(dev, dma_addr, size))) { 37 if (!dev->dma_mask) { 38 dev_err(dev, 39 "%s: call on device without dma_mask\n", 40 caller); 41 return false; 42 } 43 44 if (*dev->dma_mask >= DMA_BIT_MASK(32)) { 45 dev_err(dev, 46 "%s: overflow %pad+%zu of device mask %llx\n", 47 caller, &dma_addr, size, *dev->dma_mask); 48 } 49 return false; 50 } 51 return true; 52 } 53 54 static bool dma_coherent_ok(struct device *dev, phys_addr_t phys, size_t size) 55 { 56 dma_addr_t addr = force_dma_unencrypted() ? 57 __phys_to_dma(dev, phys) : phys_to_dma(dev, phys); 58 return addr + size - 1 <= dev->coherent_dma_mask; 59 } 60 61 void *dma_direct_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle, 62 gfp_t gfp, unsigned long attrs) 63 { 64 unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT; 65 int page_order = get_order(size); 66 struct page *page = NULL; 67 void *ret; 68 69 /* we always manually zero the memory once we are done: */ 70 gfp &= ~__GFP_ZERO; 71 72 /* GFP_DMA32 and GFP_DMA are no ops without the corresponding zones: */ 73 if (dev->coherent_dma_mask <= DMA_BIT_MASK(ARCH_ZONE_DMA_BITS)) 74 gfp |= GFP_DMA; 75 if (dev->coherent_dma_mask <= DMA_BIT_MASK(32) && !(gfp & GFP_DMA)) 76 gfp |= GFP_DMA32; 77 78 again: 79 /* CMA can be used only in the context which permits sleeping */ 80 if (gfpflags_allow_blocking(gfp)) { 81 page = dma_alloc_from_contiguous(dev, count, page_order, 82 gfp & __GFP_NOWARN); 83 if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) { 84 dma_release_from_contiguous(dev, page, count); 85 page = NULL; 86 } 87 } 88 if (!page) 89 page = alloc_pages_node(dev_to_node(dev), gfp, page_order); 90 91 if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) { 92 __free_pages(page, page_order); 93 page = NULL; 94 95 if (IS_ENABLED(CONFIG_ZONE_DMA32) && 96 dev->coherent_dma_mask < DMA_BIT_MASK(64) && 97 !(gfp & (GFP_DMA32 | GFP_DMA))) { 98 gfp |= GFP_DMA32; 99 goto again; 100 } 101 102 if (IS_ENABLED(CONFIG_ZONE_DMA) && 103 dev->coherent_dma_mask < DMA_BIT_MASK(32) && 104 !(gfp & GFP_DMA)) { 105 gfp = (gfp & ~GFP_DMA32) | GFP_DMA; 106 goto again; 107 } 108 } 109 110 if (!page) 111 return NULL; 112 ret = page_address(page); 113 if (force_dma_unencrypted()) { 114 set_memory_decrypted((unsigned long)ret, 1 << page_order); 115 *dma_handle = __phys_to_dma(dev, page_to_phys(page)); 116 } else { 117 *dma_handle = phys_to_dma(dev, page_to_phys(page)); 118 } 119 memset(ret, 0, size); 120 return ret; 121 } 122 123 /* 124 * NOTE: this function must never look at the dma_addr argument, because we want 125 * to be able to use it as a helper for iommu implementations as well. 126 */ 127 void dma_direct_free(struct device *dev, size_t size, void *cpu_addr, 128 dma_addr_t dma_addr, unsigned long attrs) 129 { 130 unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT; 131 unsigned int page_order = get_order(size); 132 133 if (force_dma_unencrypted()) 134 set_memory_encrypted((unsigned long)cpu_addr, 1 << page_order); 135 if (!dma_release_from_contiguous(dev, virt_to_page(cpu_addr), count)) 136 free_pages((unsigned long)cpu_addr, page_order); 137 } 138 139 dma_addr_t dma_direct_map_page(struct device *dev, struct page *page, 140 unsigned long offset, size_t size, enum dma_data_direction dir, 141 unsigned long attrs) 142 { 143 dma_addr_t dma_addr = phys_to_dma(dev, page_to_phys(page)) + offset; 144 145 if (!check_addr(dev, dma_addr, size, __func__)) 146 return DIRECT_MAPPING_ERROR; 147 return dma_addr; 148 } 149 150 int dma_direct_map_sg(struct device *dev, struct scatterlist *sgl, int nents, 151 enum dma_data_direction dir, unsigned long attrs) 152 { 153 int i; 154 struct scatterlist *sg; 155 156 for_each_sg(sgl, sg, nents, i) { 157 BUG_ON(!sg_page(sg)); 158 159 sg_dma_address(sg) = phys_to_dma(dev, sg_phys(sg)); 160 if (!check_addr(dev, sg_dma_address(sg), sg->length, __func__)) 161 return 0; 162 sg_dma_len(sg) = sg->length; 163 } 164 165 return nents; 166 } 167 168 int dma_direct_supported(struct device *dev, u64 mask) 169 { 170 #ifdef CONFIG_ZONE_DMA 171 if (mask < phys_to_dma(dev, DMA_BIT_MASK(ARCH_ZONE_DMA_BITS))) 172 return 0; 173 #else 174 /* 175 * Because 32-bit DMA masks are so common we expect every architecture 176 * to be able to satisfy them - either by not supporting more physical 177 * memory, or by providing a ZONE_DMA32. If neither is the case, the 178 * architecture needs to use an IOMMU instead of the direct mapping. 179 */ 180 if (mask < phys_to_dma(dev, DMA_BIT_MASK(32))) 181 return 0; 182 #endif 183 /* 184 * Upstream PCI/PCIe bridges or SoC interconnects may not carry 185 * as many DMA address bits as the device itself supports. 186 */ 187 if (dev->bus_dma_mask && mask > dev->bus_dma_mask) 188 return 0; 189 return 1; 190 } 191 192 int dma_direct_mapping_error(struct device *dev, dma_addr_t dma_addr) 193 { 194 return dma_addr == DIRECT_MAPPING_ERROR; 195 } 196 197 const struct dma_map_ops dma_direct_ops = { 198 .alloc = dma_direct_alloc, 199 .free = dma_direct_free, 200 .map_page = dma_direct_map_page, 201 .map_sg = dma_direct_map_sg, 202 .dma_supported = dma_direct_supported, 203 .mapping_error = dma_direct_mapping_error, 204 }; 205 EXPORT_SYMBOL(dma_direct_ops); 206