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 pgoff_t pgoff = vmf->pgoff; 36 37 if (pgoff >= ubuf->pagecount) 38 return VM_FAULT_SIGBUS; 39 vmf->page = ubuf->pages[pgoff]; 40 get_page(vmf->page); 41 return 0; 42 } 43 44 static const struct vm_operations_struct udmabuf_vm_ops = { 45 .fault = udmabuf_vm_fault, 46 }; 47 48 static int mmap_udmabuf(struct dma_buf *buf, struct vm_area_struct *vma) 49 { 50 struct udmabuf *ubuf = buf->priv; 51 52 if ((vma->vm_flags & (VM_SHARED | VM_MAYSHARE)) == 0) 53 return -EINVAL; 54 55 vma->vm_ops = &udmabuf_vm_ops; 56 vma->vm_private_data = ubuf; 57 return 0; 58 } 59 60 static struct sg_table *get_sg_table(struct device *dev, struct dma_buf *buf, 61 enum dma_data_direction direction) 62 { 63 struct udmabuf *ubuf = buf->priv; 64 struct sg_table *sg; 65 int ret; 66 67 sg = kzalloc(sizeof(*sg), GFP_KERNEL); 68 if (!sg) 69 return ERR_PTR(-ENOMEM); 70 ret = sg_alloc_table_from_pages(sg, ubuf->pages, ubuf->pagecount, 71 0, ubuf->pagecount << PAGE_SHIFT, 72 GFP_KERNEL); 73 if (ret < 0) 74 goto err; 75 ret = dma_map_sgtable(dev, sg, direction, 0); 76 if (ret < 0) 77 goto err; 78 return sg; 79 80 err: 81 sg_free_table(sg); 82 kfree(sg); 83 return ERR_PTR(ret); 84 } 85 86 static void put_sg_table(struct device *dev, struct sg_table *sg, 87 enum dma_data_direction direction) 88 { 89 dma_unmap_sgtable(dev, sg, direction, 0); 90 sg_free_table(sg); 91 kfree(sg); 92 } 93 94 static struct sg_table *map_udmabuf(struct dma_buf_attachment *at, 95 enum dma_data_direction direction) 96 { 97 return get_sg_table(at->dev, at->dmabuf, direction); 98 } 99 100 static void unmap_udmabuf(struct dma_buf_attachment *at, 101 struct sg_table *sg, 102 enum dma_data_direction direction) 103 { 104 return put_sg_table(at->dev, sg, direction); 105 } 106 107 static void release_udmabuf(struct dma_buf *buf) 108 { 109 struct udmabuf *ubuf = buf->priv; 110 struct device *dev = ubuf->device->this_device; 111 pgoff_t pg; 112 113 if (ubuf->sg) 114 put_sg_table(dev, ubuf->sg, DMA_BIDIRECTIONAL); 115 116 for (pg = 0; pg < ubuf->pagecount; pg++) 117 put_page(ubuf->pages[pg]); 118 kfree(ubuf->pages); 119 kfree(ubuf); 120 } 121 122 static int begin_cpu_udmabuf(struct dma_buf *buf, 123 enum dma_data_direction direction) 124 { 125 struct udmabuf *ubuf = buf->priv; 126 struct device *dev = ubuf->device->this_device; 127 128 if (!ubuf->sg) { 129 ubuf->sg = get_sg_table(dev, buf, direction); 130 if (IS_ERR(ubuf->sg)) 131 return PTR_ERR(ubuf->sg); 132 } else { 133 dma_sync_sg_for_cpu(dev, ubuf->sg->sgl, ubuf->sg->nents, 134 direction); 135 } 136 137 return 0; 138 } 139 140 static int end_cpu_udmabuf(struct dma_buf *buf, 141 enum dma_data_direction direction) 142 { 143 struct udmabuf *ubuf = buf->priv; 144 struct device *dev = ubuf->device->this_device; 145 146 if (!ubuf->sg) 147 return -EINVAL; 148 149 dma_sync_sg_for_device(dev, ubuf->sg->sgl, ubuf->sg->nents, direction); 150 return 0; 151 } 152 153 static const struct dma_buf_ops udmabuf_ops = { 154 .cache_sgt_mapping = true, 155 .map_dma_buf = map_udmabuf, 156 .unmap_dma_buf = unmap_udmabuf, 157 .release = release_udmabuf, 158 .mmap = mmap_udmabuf, 159 .begin_cpu_access = begin_cpu_udmabuf, 160 .end_cpu_access = end_cpu_udmabuf, 161 }; 162 163 #define SEALS_WANTED (F_SEAL_SHRINK) 164 #define SEALS_DENIED (F_SEAL_WRITE) 165 166 static long udmabuf_create(struct miscdevice *device, 167 struct udmabuf_create_list *head, 168 struct udmabuf_create_item *list) 169 { 170 DEFINE_DMA_BUF_EXPORT_INFO(exp_info); 171 struct file *memfd = NULL; 172 struct address_space *mapping = NULL; 173 struct udmabuf *ubuf; 174 struct dma_buf *buf; 175 pgoff_t pgoff, pgcnt, pgidx, pgbuf = 0, pglimit; 176 struct page *page, *hpage = NULL; 177 pgoff_t subpgoff, maxsubpgs; 178 struct hstate *hpstate; 179 int seals, ret = -EINVAL; 180 u32 i, flags; 181 182 ubuf = kzalloc(sizeof(*ubuf), GFP_KERNEL); 183 if (!ubuf) 184 return -ENOMEM; 185 186 pglimit = (size_limit_mb * 1024 * 1024) >> PAGE_SHIFT; 187 for (i = 0; i < head->count; i++) { 188 if (!IS_ALIGNED(list[i].offset, PAGE_SIZE)) 189 goto err; 190 if (!IS_ALIGNED(list[i].size, PAGE_SIZE)) 191 goto err; 192 ubuf->pagecount += list[i].size >> PAGE_SHIFT; 193 if (ubuf->pagecount > pglimit) 194 goto err; 195 } 196 197 if (!ubuf->pagecount) 198 goto err; 199 200 ubuf->pages = kmalloc_array(ubuf->pagecount, sizeof(*ubuf->pages), 201 GFP_KERNEL); 202 if (!ubuf->pages) { 203 ret = -ENOMEM; 204 goto err; 205 } 206 207 pgbuf = 0; 208 for (i = 0; i < head->count; i++) { 209 ret = -EBADFD; 210 memfd = fget(list[i].memfd); 211 if (!memfd) 212 goto err; 213 mapping = file_inode(memfd)->i_mapping; 214 if (!shmem_mapping(mapping) && !is_file_hugepages(memfd)) 215 goto err; 216 seals = memfd_fcntl(memfd, F_GET_SEALS, 0); 217 if (seals == -EINVAL) 218 goto err; 219 ret = -EINVAL; 220 if ((seals & SEALS_WANTED) != SEALS_WANTED || 221 (seals & SEALS_DENIED) != 0) 222 goto err; 223 pgoff = list[i].offset >> PAGE_SHIFT; 224 pgcnt = list[i].size >> PAGE_SHIFT; 225 if (is_file_hugepages(memfd)) { 226 hpstate = hstate_file(memfd); 227 pgoff = list[i].offset >> huge_page_shift(hpstate); 228 subpgoff = (list[i].offset & 229 ~huge_page_mask(hpstate)) >> PAGE_SHIFT; 230 maxsubpgs = huge_page_size(hpstate) >> PAGE_SHIFT; 231 } 232 for (pgidx = 0; pgidx < pgcnt; pgidx++) { 233 if (is_file_hugepages(memfd)) { 234 if (!hpage) { 235 hpage = find_get_page_flags(mapping, pgoff, 236 FGP_ACCESSED); 237 if (!hpage) { 238 ret = -EINVAL; 239 goto err; 240 } 241 } 242 page = hpage + subpgoff; 243 get_page(page); 244 subpgoff++; 245 if (subpgoff == maxsubpgs) { 246 put_page(hpage); 247 hpage = NULL; 248 subpgoff = 0; 249 pgoff++; 250 } 251 } else { 252 page = shmem_read_mapping_page(mapping, 253 pgoff + pgidx); 254 if (IS_ERR(page)) { 255 ret = PTR_ERR(page); 256 goto err; 257 } 258 } 259 ubuf->pages[pgbuf++] = page; 260 } 261 fput(memfd); 262 memfd = NULL; 263 if (hpage) { 264 put_page(hpage); 265 hpage = NULL; 266 } 267 } 268 269 exp_info.ops = &udmabuf_ops; 270 exp_info.size = ubuf->pagecount << PAGE_SHIFT; 271 exp_info.priv = ubuf; 272 exp_info.flags = O_RDWR; 273 274 ubuf->device = device; 275 buf = dma_buf_export(&exp_info); 276 if (IS_ERR(buf)) { 277 ret = PTR_ERR(buf); 278 goto err; 279 } 280 281 flags = 0; 282 if (head->flags & UDMABUF_FLAGS_CLOEXEC) 283 flags |= O_CLOEXEC; 284 return dma_buf_fd(buf, flags); 285 286 err: 287 while (pgbuf > 0) 288 put_page(ubuf->pages[--pgbuf]); 289 if (memfd) 290 fput(memfd); 291 kfree(ubuf->pages); 292 kfree(ubuf); 293 return ret; 294 } 295 296 static long udmabuf_ioctl_create(struct file *filp, unsigned long arg) 297 { 298 struct udmabuf_create create; 299 struct udmabuf_create_list head; 300 struct udmabuf_create_item list; 301 302 if (copy_from_user(&create, (void __user *)arg, 303 sizeof(create))) 304 return -EFAULT; 305 306 head.flags = create.flags; 307 head.count = 1; 308 list.memfd = create.memfd; 309 list.offset = create.offset; 310 list.size = create.size; 311 312 return udmabuf_create(filp->private_data, &head, &list); 313 } 314 315 static long udmabuf_ioctl_create_list(struct file *filp, unsigned long arg) 316 { 317 struct udmabuf_create_list head; 318 struct udmabuf_create_item *list; 319 int ret = -EINVAL; 320 u32 lsize; 321 322 if (copy_from_user(&head, (void __user *)arg, sizeof(head))) 323 return -EFAULT; 324 if (head.count > list_limit) 325 return -EINVAL; 326 lsize = sizeof(struct udmabuf_create_item) * head.count; 327 list = memdup_user((void __user *)(arg + sizeof(head)), lsize); 328 if (IS_ERR(list)) 329 return PTR_ERR(list); 330 331 ret = udmabuf_create(filp->private_data, &head, list); 332 kfree(list); 333 return ret; 334 } 335 336 static long udmabuf_ioctl(struct file *filp, unsigned int ioctl, 337 unsigned long arg) 338 { 339 long ret; 340 341 switch (ioctl) { 342 case UDMABUF_CREATE: 343 ret = udmabuf_ioctl_create(filp, arg); 344 break; 345 case UDMABUF_CREATE_LIST: 346 ret = udmabuf_ioctl_create_list(filp, arg); 347 break; 348 default: 349 ret = -ENOTTY; 350 break; 351 } 352 return ret; 353 } 354 355 static const struct file_operations udmabuf_fops = { 356 .owner = THIS_MODULE, 357 .unlocked_ioctl = udmabuf_ioctl, 358 #ifdef CONFIG_COMPAT 359 .compat_ioctl = udmabuf_ioctl, 360 #endif 361 }; 362 363 static struct miscdevice udmabuf_misc = { 364 .minor = MISC_DYNAMIC_MINOR, 365 .name = "udmabuf", 366 .fops = &udmabuf_fops, 367 }; 368 369 static int __init udmabuf_dev_init(void) 370 { 371 int ret; 372 373 ret = misc_register(&udmabuf_misc); 374 if (ret < 0) { 375 pr_err("Could not initialize udmabuf device\n"); 376 return ret; 377 } 378 379 ret = dma_coerce_mask_and_coherent(udmabuf_misc.this_device, 380 DMA_BIT_MASK(64)); 381 if (ret < 0) { 382 pr_err("Could not setup DMA mask for udmabuf device\n"); 383 misc_deregister(&udmabuf_misc); 384 return ret; 385 } 386 387 return 0; 388 } 389 390 static void __exit udmabuf_dev_exit(void) 391 { 392 misc_deregister(&udmabuf_misc); 393 } 394 395 module_init(udmabuf_dev_init) 396 module_exit(udmabuf_dev_exit) 397 398 MODULE_AUTHOR("Gerd Hoffmann <kraxel@redhat.com>"); 399 MODULE_LICENSE("GPL v2"); 400