1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/cred.h> 3 #include <linux/device.h> 4 #include <linux/dma-buf.h> 5 #include <linux/highmem.h> 6 #include <linux/init.h> 7 #include <linux/kernel.h> 8 #include <linux/memfd.h> 9 #include <linux/miscdevice.h> 10 #include <linux/module.h> 11 #include <linux/shmem_fs.h> 12 #include <linux/slab.h> 13 #include <linux/udmabuf.h> 14 #include <linux/hugetlb.h> 15 16 static int list_limit = 1024; 17 module_param(list_limit, int, 0644); 18 MODULE_PARM_DESC(list_limit, "udmabuf_create_list->count limit. Default is 1024."); 19 20 static int size_limit_mb = 64; 21 module_param(size_limit_mb, int, 0644); 22 MODULE_PARM_DESC(size_limit_mb, "Max size of a dmabuf, in megabytes. Default is 64."); 23 24 struct udmabuf { 25 pgoff_t pagecount; 26 struct page **pages; 27 struct sg_table *sg; 28 struct miscdevice *device; 29 }; 30 31 static vm_fault_t udmabuf_vm_fault(struct vm_fault *vmf) 32 { 33 struct vm_area_struct *vma = vmf->vma; 34 struct udmabuf *ubuf = vma->vm_private_data; 35 36 vmf->page = ubuf->pages[vmf->pgoff]; 37 get_page(vmf->page); 38 return 0; 39 } 40 41 static const struct vm_operations_struct udmabuf_vm_ops = { 42 .fault = udmabuf_vm_fault, 43 }; 44 45 static int mmap_udmabuf(struct dma_buf *buf, struct vm_area_struct *vma) 46 { 47 struct udmabuf *ubuf = buf->priv; 48 49 if ((vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) == 0) 50 return -EINVAL; 51 52 vma->vm_ops = &udmabuf_vm_ops; 53 vma->vm_private_data = ubuf; 54 return 0; 55 } 56 57 static struct sg_table *get_sg_table(struct device *dev, struct dma_buf *buf, 58 enum dma_data_direction direction) 59 { 60 struct udmabuf *ubuf = buf->priv; 61 struct sg_table *sg; 62 int ret; 63 64 sg = kzalloc(sizeof(*sg), GFP_KERNEL); 65 if (!sg) 66 return ERR_PTR(-ENOMEM); 67 ret = sg_alloc_table_from_pages(sg, ubuf->pages, ubuf->pagecount, 68 0, ubuf->pagecount << PAGE_SHIFT, 69 GFP_KERNEL); 70 if (ret < 0) 71 goto err; 72 ret = dma_map_sgtable(dev, sg, direction, 0); 73 if (ret < 0) 74 goto err; 75 return sg; 76 77 err: 78 sg_free_table(sg); 79 kfree(sg); 80 return ERR_PTR(ret); 81 } 82 83 static void put_sg_table(struct device *dev, struct sg_table *sg, 84 enum dma_data_direction direction) 85 { 86 dma_unmap_sgtable(dev, sg, direction, 0); 87 sg_free_table(sg); 88 kfree(sg); 89 } 90 91 static struct sg_table *map_udmabuf(struct dma_buf_attachment *at, 92 enum dma_data_direction direction) 93 { 94 return get_sg_table(at->dev, at->dmabuf, direction); 95 } 96 97 static void unmap_udmabuf(struct dma_buf_attachment *at, 98 struct sg_table *sg, 99 enum dma_data_direction direction) 100 { 101 return put_sg_table(at->dev, sg, direction); 102 } 103 104 static void release_udmabuf(struct dma_buf *buf) 105 { 106 struct udmabuf *ubuf = buf->priv; 107 struct device *dev = ubuf->device->this_device; 108 pgoff_t pg; 109 110 if (ubuf->sg) 111 put_sg_table(dev, ubuf->sg, DMA_BIDIRECTIONAL); 112 113 for (pg = 0; pg < ubuf->pagecount; pg++) 114 put_page(ubuf->pages[pg]); 115 kfree(ubuf->pages); 116 kfree(ubuf); 117 } 118 119 static int begin_cpu_udmabuf(struct dma_buf *buf, 120 enum dma_data_direction direction) 121 { 122 struct udmabuf *ubuf = buf->priv; 123 struct device *dev = ubuf->device->this_device; 124 125 if (!ubuf->sg) { 126 ubuf->sg = get_sg_table(dev, buf, direction); 127 if (IS_ERR(ubuf->sg)) 128 return PTR_ERR(ubuf->sg); 129 } else { 130 dma_sync_sg_for_cpu(dev, ubuf->sg->sgl, ubuf->sg->nents, 131 direction); 132 } 133 134 return 0; 135 } 136 137 static int end_cpu_udmabuf(struct dma_buf *buf, 138 enum dma_data_direction direction) 139 { 140 struct udmabuf *ubuf = buf->priv; 141 struct device *dev = ubuf->device->this_device; 142 143 if (!ubuf->sg) 144 return -EINVAL; 145 146 dma_sync_sg_for_device(dev, ubuf->sg->sgl, ubuf->sg->nents, direction); 147 return 0; 148 } 149 150 static const struct dma_buf_ops udmabuf_ops = { 151 .cache_sgt_mapping = true, 152 .map_dma_buf = map_udmabuf, 153 .unmap_dma_buf = unmap_udmabuf, 154 .release = release_udmabuf, 155 .mmap = mmap_udmabuf, 156 .begin_cpu_access = begin_cpu_udmabuf, 157 .end_cpu_access = end_cpu_udmabuf, 158 }; 159 160 #define SEALS_WANTED (F_SEAL_SHRINK) 161 #define SEALS_DENIED (F_SEAL_WRITE) 162 163 static long udmabuf_create(struct miscdevice *device, 164 struct udmabuf_create_list *head, 165 struct udmabuf_create_item *list) 166 { 167 DEFINE_DMA_BUF_EXPORT_INFO(exp_info); 168 struct file *memfd = NULL; 169 struct address_space *mapping = NULL; 170 struct udmabuf *ubuf; 171 struct dma_buf *buf; 172 pgoff_t pgoff, pgcnt, pgidx, pgbuf = 0, pglimit; 173 struct page *page, *hpage = NULL; 174 pgoff_t subpgoff, maxsubpgs; 175 struct hstate *hpstate; 176 int seals, ret = -EINVAL; 177 u32 i, flags; 178 179 ubuf = kzalloc(sizeof(*ubuf), GFP_KERNEL); 180 if (!ubuf) 181 return -ENOMEM; 182 183 pglimit = (size_limit_mb * 1024 * 1024) >> PAGE_SHIFT; 184 for (i = 0; i < head->count; i++) { 185 if (!IS_ALIGNED(list[i].offset, PAGE_SIZE)) 186 goto err; 187 if (!IS_ALIGNED(list[i].size, PAGE_SIZE)) 188 goto err; 189 ubuf->pagecount += list[i].size >> PAGE_SHIFT; 190 if (ubuf->pagecount > pglimit) 191 goto err; 192 } 193 194 if (!ubuf->pagecount) 195 goto err; 196 197 ubuf->pages = kmalloc_array(ubuf->pagecount, sizeof(*ubuf->pages), 198 GFP_KERNEL); 199 if (!ubuf->pages) { 200 ret = -ENOMEM; 201 goto err; 202 } 203 204 pgbuf = 0; 205 for (i = 0; i < head->count; i++) { 206 ret = -EBADFD; 207 memfd = fget(list[i].memfd); 208 if (!memfd) 209 goto err; 210 mapping = file_inode(memfd)->i_mapping; 211 if (!shmem_mapping(mapping) && !is_file_hugepages(memfd)) 212 goto err; 213 seals = memfd_fcntl(memfd, F_GET_SEALS, 0); 214 if (seals == -EINVAL) 215 goto err; 216 ret = -EINVAL; 217 if ((seals & SEALS_WANTED) != SEALS_WANTED || 218 (seals & SEALS_DENIED) != 0) 219 goto err; 220 pgoff = list[i].offset >> PAGE_SHIFT; 221 pgcnt = list[i].size >> PAGE_SHIFT; 222 if (is_file_hugepages(memfd)) { 223 hpstate = hstate_file(memfd); 224 pgoff = list[i].offset >> huge_page_shift(hpstate); 225 subpgoff = (list[i].offset & 226 ~huge_page_mask(hpstate)) >> PAGE_SHIFT; 227 maxsubpgs = huge_page_size(hpstate) >> PAGE_SHIFT; 228 } 229 for (pgidx = 0; pgidx < pgcnt; pgidx++) { 230 if (is_file_hugepages(memfd)) { 231 if (!hpage) { 232 hpage = find_get_page_flags(mapping, pgoff, 233 FGP_ACCESSED); 234 if (!hpage) { 235 ret = -EINVAL; 236 goto err; 237 } 238 } 239 page = hpage + subpgoff; 240 get_page(page); 241 subpgoff++; 242 if (subpgoff == maxsubpgs) { 243 put_page(hpage); 244 hpage = NULL; 245 subpgoff = 0; 246 pgoff++; 247 } 248 } else { 249 page = shmem_read_mapping_page(mapping, 250 pgoff + pgidx); 251 if (IS_ERR(page)) { 252 ret = PTR_ERR(page); 253 goto err; 254 } 255 } 256 ubuf->pages[pgbuf++] = page; 257 } 258 fput(memfd); 259 memfd = NULL; 260 if (hpage) { 261 put_page(hpage); 262 hpage = NULL; 263 } 264 } 265 266 exp_info.ops = &udmabuf_ops; 267 exp_info.size = ubuf->pagecount << PAGE_SHIFT; 268 exp_info.priv = ubuf; 269 exp_info.flags = O_RDWR; 270 271 ubuf->device = device; 272 buf = dma_buf_export(&exp_info); 273 if (IS_ERR(buf)) { 274 ret = PTR_ERR(buf); 275 goto err; 276 } 277 278 flags = 0; 279 if (head->flags & UDMABUF_FLAGS_CLOEXEC) 280 flags |= O_CLOEXEC; 281 return dma_buf_fd(buf, flags); 282 283 err: 284 while (pgbuf > 0) 285 put_page(ubuf->pages[--pgbuf]); 286 if (memfd) 287 fput(memfd); 288 kfree(ubuf->pages); 289 kfree(ubuf); 290 return ret; 291 } 292 293 static long udmabuf_ioctl_create(struct file *filp, unsigned long arg) 294 { 295 struct udmabuf_create create; 296 struct udmabuf_create_list head; 297 struct udmabuf_create_item list; 298 299 if (copy_from_user(&create, (void __user *)arg, 300 sizeof(create))) 301 return -EFAULT; 302 303 head.flags = create.flags; 304 head.count = 1; 305 list.memfd = create.memfd; 306 list.offset = create.offset; 307 list.size = create.size; 308 309 return udmabuf_create(filp->private_data, &head, &list); 310 } 311 312 static long udmabuf_ioctl_create_list(struct file *filp, unsigned long arg) 313 { 314 struct udmabuf_create_list head; 315 struct udmabuf_create_item *list; 316 int ret = -EINVAL; 317 u32 lsize; 318 319 if (copy_from_user(&head, (void __user *)arg, sizeof(head))) 320 return -EFAULT; 321 if (head.count > list_limit) 322 return -EINVAL; 323 lsize = sizeof(struct udmabuf_create_item) * head.count; 324 list = memdup_user((void __user *)(arg + sizeof(head)), lsize); 325 if (IS_ERR(list)) 326 return PTR_ERR(list); 327 328 ret = udmabuf_create(filp->private_data, &head, list); 329 kfree(list); 330 return ret; 331 } 332 333 static long udmabuf_ioctl(struct file *filp, unsigned int ioctl, 334 unsigned long arg) 335 { 336 long ret; 337 338 switch (ioctl) { 339 case UDMABUF_CREATE: 340 ret = udmabuf_ioctl_create(filp, arg); 341 break; 342 case UDMABUF_CREATE_LIST: 343 ret = udmabuf_ioctl_create_list(filp, arg); 344 break; 345 default: 346 ret = -ENOTTY; 347 break; 348 } 349 return ret; 350 } 351 352 static const struct file_operations udmabuf_fops = { 353 .owner = THIS_MODULE, 354 .unlocked_ioctl = udmabuf_ioctl, 355 #ifdef CONFIG_COMPAT 356 .compat_ioctl = udmabuf_ioctl, 357 #endif 358 }; 359 360 static struct miscdevice udmabuf_misc = { 361 .minor = MISC_DYNAMIC_MINOR, 362 .name = "udmabuf", 363 .fops = &udmabuf_fops, 364 }; 365 366 static int __init udmabuf_dev_init(void) 367 { 368 int ret; 369 370 ret = misc_register(&udmabuf_misc); 371 if (ret < 0) { 372 pr_err("Could not initialize udmabuf device\n"); 373 return ret; 374 } 375 376 ret = dma_coerce_mask_and_coherent(udmabuf_misc.this_device, 377 DMA_BIT_MASK(64)); 378 if (ret < 0) { 379 pr_err("Could not setup DMA mask for udmabuf device\n"); 380 misc_deregister(&udmabuf_misc); 381 return ret; 382 } 383 384 return 0; 385 } 386 387 static void __exit udmabuf_dev_exit(void) 388 { 389 misc_deregister(&udmabuf_misc); 390 } 391 392 module_init(udmabuf_dev_init) 393 module_exit(udmabuf_dev_exit) 394 395 MODULE_AUTHOR("Gerd Hoffmann <kraxel@redhat.com>"); 396 MODULE_LICENSE("GPL v2"); 397