1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/kernel.h> 3 #include <linux/init.h> 4 #include <linux/errno.h> 5 #include <linux/mm.h> 6 #include <linux/mman.h> 7 #include <linux/slab.h> 8 #include <linux/vmalloc.h> 9 #include <linux/io_uring.h> 10 #include <linux/io_uring_types.h> 11 #include <asm/shmparam.h> 12 13 #include "memmap.h" 14 #include "kbuf.h" 15 #include "rsrc.h" 16 17 static void *io_mem_alloc_compound(struct page **pages, int nr_pages, 18 size_t size, gfp_t gfp) 19 { 20 struct page *page; 21 int i, order; 22 23 order = get_order(size); 24 if (order > MAX_PAGE_ORDER) 25 return ERR_PTR(-ENOMEM); 26 else if (order) 27 gfp |= __GFP_COMP; 28 29 page = alloc_pages(gfp, order); 30 if (!page) 31 return ERR_PTR(-ENOMEM); 32 33 for (i = 0; i < nr_pages; i++) 34 pages[i] = page + i; 35 36 return page_address(page); 37 } 38 39 static void *io_mem_alloc_single(struct page **pages, int nr_pages, size_t size, 40 gfp_t gfp) 41 { 42 void *ret; 43 int i; 44 45 for (i = 0; i < nr_pages; i++) { 46 pages[i] = alloc_page(gfp); 47 if (!pages[i]) 48 goto err; 49 } 50 51 ret = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL); 52 if (ret) 53 return ret; 54 err: 55 while (i--) 56 put_page(pages[i]); 57 return ERR_PTR(-ENOMEM); 58 } 59 60 void *io_pages_map(struct page ***out_pages, unsigned short *npages, 61 size_t size) 62 { 63 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN; 64 struct page **pages; 65 int nr_pages; 66 void *ret; 67 68 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; 69 pages = kvmalloc_array(nr_pages, sizeof(struct page *), gfp); 70 if (!pages) 71 return ERR_PTR(-ENOMEM); 72 73 ret = io_mem_alloc_compound(pages, nr_pages, size, gfp); 74 if (!IS_ERR(ret)) 75 goto done; 76 77 ret = io_mem_alloc_single(pages, nr_pages, size, gfp); 78 if (!IS_ERR(ret)) { 79 done: 80 *out_pages = pages; 81 *npages = nr_pages; 82 return ret; 83 } 84 85 kvfree(pages); 86 *out_pages = NULL; 87 *npages = 0; 88 return ret; 89 } 90 91 void io_pages_unmap(void *ptr, struct page ***pages, unsigned short *npages, 92 bool put_pages) 93 { 94 bool do_vunmap = false; 95 96 if (!ptr) 97 return; 98 99 if (put_pages && *npages) { 100 struct page **to_free = *pages; 101 int i; 102 103 /* 104 * Only did vmap for the non-compound multiple page case. 105 * For the compound page, we just need to put the head. 106 */ 107 if (PageCompound(to_free[0])) 108 *npages = 1; 109 else if (*npages > 1) 110 do_vunmap = true; 111 for (i = 0; i < *npages; i++) 112 put_page(to_free[i]); 113 } 114 if (do_vunmap) 115 vunmap(ptr); 116 kvfree(*pages); 117 *pages = NULL; 118 *npages = 0; 119 } 120 121 void io_pages_free(struct page ***pages, int npages) 122 { 123 struct page **page_array = *pages; 124 125 if (!page_array) 126 return; 127 128 unpin_user_pages(page_array, npages); 129 kvfree(page_array); 130 *pages = NULL; 131 } 132 133 struct page **io_pin_pages(unsigned long uaddr, unsigned long len, int *npages) 134 { 135 unsigned long start, end, nr_pages; 136 struct page **pages; 137 int ret; 138 139 if (check_add_overflow(uaddr, len, &end)) 140 return ERR_PTR(-EOVERFLOW); 141 if (check_add_overflow(end, PAGE_SIZE - 1, &end)) 142 return ERR_PTR(-EOVERFLOW); 143 144 end = end >> PAGE_SHIFT; 145 start = uaddr >> PAGE_SHIFT; 146 nr_pages = end - start; 147 if (WARN_ON_ONCE(!nr_pages)) 148 return ERR_PTR(-EINVAL); 149 if (WARN_ON_ONCE(nr_pages > INT_MAX)) 150 return ERR_PTR(-EOVERFLOW); 151 152 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); 153 if (!pages) 154 return ERR_PTR(-ENOMEM); 155 156 ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM, 157 pages); 158 /* success, mapped all pages */ 159 if (ret == nr_pages) { 160 *npages = nr_pages; 161 return pages; 162 } 163 164 /* partial map, or didn't map anything */ 165 if (ret >= 0) { 166 /* if we did partial map, release any pages we did get */ 167 if (ret) 168 unpin_user_pages(pages, ret); 169 ret = -EFAULT; 170 } 171 kvfree(pages); 172 return ERR_PTR(ret); 173 } 174 175 void *__io_uaddr_map(struct page ***pages, unsigned short *npages, 176 unsigned long uaddr, size_t size) 177 { 178 struct page **page_array; 179 unsigned int nr_pages; 180 void *page_addr; 181 182 *npages = 0; 183 184 if (uaddr & (PAGE_SIZE - 1) || !size) 185 return ERR_PTR(-EINVAL); 186 187 nr_pages = 0; 188 page_array = io_pin_pages(uaddr, size, &nr_pages); 189 if (IS_ERR(page_array)) 190 return page_array; 191 192 page_addr = vmap(page_array, nr_pages, VM_MAP, PAGE_KERNEL); 193 if (page_addr) { 194 *pages = page_array; 195 *npages = nr_pages; 196 return page_addr; 197 } 198 199 io_pages_free(&page_array, nr_pages); 200 return ERR_PTR(-ENOMEM); 201 } 202 203 void io_free_region(struct io_ring_ctx *ctx, struct io_mapped_region *mr) 204 { 205 if (mr->pages) { 206 unpin_user_pages(mr->pages, mr->nr_pages); 207 kvfree(mr->pages); 208 } 209 if (mr->vmap_ptr) 210 vunmap(mr->vmap_ptr); 211 if (mr->nr_pages && ctx->user) 212 __io_unaccount_mem(ctx->user, mr->nr_pages); 213 214 memset(mr, 0, sizeof(*mr)); 215 } 216 217 int io_create_region(struct io_ring_ctx *ctx, struct io_mapped_region *mr, 218 struct io_uring_region_desc *reg) 219 { 220 int pages_accounted = 0; 221 struct page **pages; 222 int nr_pages, ret; 223 void *vptr; 224 u64 end; 225 226 if (WARN_ON_ONCE(mr->pages || mr->vmap_ptr || mr->nr_pages)) 227 return -EFAULT; 228 if (memchr_inv(®->__resv, 0, sizeof(reg->__resv))) 229 return -EINVAL; 230 if (reg->flags != IORING_MEM_REGION_TYPE_USER) 231 return -EINVAL; 232 if (!reg->user_addr) 233 return -EFAULT; 234 if (!reg->size || reg->mmap_offset || reg->id) 235 return -EINVAL; 236 if ((reg->size >> PAGE_SHIFT) > INT_MAX) 237 return -E2BIG; 238 if ((reg->user_addr | reg->size) & ~PAGE_MASK) 239 return -EINVAL; 240 if (check_add_overflow(reg->user_addr, reg->size, &end)) 241 return -EOVERFLOW; 242 243 pages = io_pin_pages(reg->user_addr, reg->size, &nr_pages); 244 if (IS_ERR(pages)) 245 return PTR_ERR(pages); 246 247 if (ctx->user) { 248 ret = __io_account_mem(ctx->user, nr_pages); 249 if (ret) 250 goto out_free; 251 pages_accounted = nr_pages; 252 } 253 254 vptr = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL); 255 if (!vptr) { 256 ret = -ENOMEM; 257 goto out_free; 258 } 259 260 mr->pages = pages; 261 mr->vmap_ptr = vptr; 262 mr->nr_pages = nr_pages; 263 return 0; 264 out_free: 265 if (pages_accounted) 266 __io_unaccount_mem(ctx->user, pages_accounted); 267 io_pages_free(&pages, nr_pages); 268 return ret; 269 } 270 271 static void *io_uring_validate_mmap_request(struct file *file, loff_t pgoff, 272 size_t sz) 273 { 274 struct io_ring_ctx *ctx = file->private_data; 275 loff_t offset = pgoff << PAGE_SHIFT; 276 277 switch ((pgoff << PAGE_SHIFT) & IORING_OFF_MMAP_MASK) { 278 case IORING_OFF_SQ_RING: 279 case IORING_OFF_CQ_RING: 280 /* Don't allow mmap if the ring was setup without it */ 281 if (ctx->flags & IORING_SETUP_NO_MMAP) 282 return ERR_PTR(-EINVAL); 283 if (!ctx->rings) 284 return ERR_PTR(-EFAULT); 285 return ctx->rings; 286 case IORING_OFF_SQES: 287 /* Don't allow mmap if the ring was setup without it */ 288 if (ctx->flags & IORING_SETUP_NO_MMAP) 289 return ERR_PTR(-EINVAL); 290 if (!ctx->sq_sqes) 291 return ERR_PTR(-EFAULT); 292 return ctx->sq_sqes; 293 case IORING_OFF_PBUF_RING: { 294 struct io_buffer_list *bl; 295 unsigned int bgid; 296 void *ptr; 297 298 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT; 299 bl = io_pbuf_get_bl(ctx, bgid); 300 if (IS_ERR(bl)) 301 return bl; 302 ptr = bl->buf_ring; 303 io_put_bl(ctx, bl); 304 return ptr; 305 } 306 } 307 308 return ERR_PTR(-EINVAL); 309 } 310 311 int io_uring_mmap_pages(struct io_ring_ctx *ctx, struct vm_area_struct *vma, 312 struct page **pages, int npages) 313 { 314 unsigned long nr_pages = npages; 315 316 vm_flags_set(vma, VM_DONTEXPAND); 317 return vm_insert_pages(vma, vma->vm_start, pages, &nr_pages); 318 } 319 320 #ifdef CONFIG_MMU 321 322 __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) 323 { 324 struct io_ring_ctx *ctx = file->private_data; 325 size_t sz = vma->vm_end - vma->vm_start; 326 long offset = vma->vm_pgoff << PAGE_SHIFT; 327 unsigned int npages; 328 void *ptr; 329 330 guard(mutex)(&ctx->resize_lock); 331 332 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); 333 if (IS_ERR(ptr)) 334 return PTR_ERR(ptr); 335 336 switch (offset & IORING_OFF_MMAP_MASK) { 337 case IORING_OFF_SQ_RING: 338 case IORING_OFF_CQ_RING: 339 npages = min(ctx->n_ring_pages, (sz + PAGE_SIZE - 1) >> PAGE_SHIFT); 340 return io_uring_mmap_pages(ctx, vma, ctx->ring_pages, npages); 341 case IORING_OFF_SQES: 342 return io_uring_mmap_pages(ctx, vma, ctx->sqe_pages, 343 ctx->n_sqe_pages); 344 case IORING_OFF_PBUF_RING: 345 return io_pbuf_mmap(file, vma); 346 } 347 348 return -EINVAL; 349 } 350 351 unsigned long io_uring_get_unmapped_area(struct file *filp, unsigned long addr, 352 unsigned long len, unsigned long pgoff, 353 unsigned long flags) 354 { 355 struct io_ring_ctx *ctx = filp->private_data; 356 void *ptr; 357 358 /* 359 * Do not allow to map to user-provided address to avoid breaking the 360 * aliasing rules. Userspace is not able to guess the offset address of 361 * kernel kmalloc()ed memory area. 362 */ 363 if (addr) 364 return -EINVAL; 365 366 guard(mutex)(&ctx->resize_lock); 367 368 ptr = io_uring_validate_mmap_request(filp, pgoff, len); 369 if (IS_ERR(ptr)) 370 return -ENOMEM; 371 372 /* 373 * Some architectures have strong cache aliasing requirements. 374 * For such architectures we need a coherent mapping which aliases 375 * kernel memory *and* userspace memory. To achieve that: 376 * - use a NULL file pointer to reference physical memory, and 377 * - use the kernel virtual address of the shared io_uring context 378 * (instead of the userspace-provided address, which has to be 0UL 379 * anyway). 380 * - use the same pgoff which the get_unmapped_area() uses to 381 * calculate the page colouring. 382 * For architectures without such aliasing requirements, the 383 * architecture will return any suitable mapping because addr is 0. 384 */ 385 filp = NULL; 386 flags |= MAP_SHARED; 387 pgoff = 0; /* has been translated to ptr above */ 388 #ifdef SHM_COLOUR 389 addr = (uintptr_t) ptr; 390 pgoff = addr >> PAGE_SHIFT; 391 #else 392 addr = 0UL; 393 #endif 394 return mm_get_unmapped_area(current->mm, filp, addr, len, pgoff, flags); 395 } 396 397 #else /* !CONFIG_MMU */ 398 399 int io_uring_mmap(struct file *file, struct vm_area_struct *vma) 400 { 401 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL; 402 } 403 404 unsigned int io_uring_nommu_mmap_capabilities(struct file *file) 405 { 406 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; 407 } 408 409 unsigned long io_uring_get_unmapped_area(struct file *file, unsigned long addr, 410 unsigned long len, unsigned long pgoff, 411 unsigned long flags) 412 { 413 struct io_ring_ctx *ctx = file->private_data; 414 void *ptr; 415 416 guard(mutex)(&ctx->resize_lock); 417 418 ptr = io_uring_validate_mmap_request(file, pgoff, len); 419 if (IS_ERR(ptr)) 420 return PTR_ERR(ptr); 421 422 return (unsigned long) ptr; 423 } 424 425 #endif /* !CONFIG_MMU */ 426