1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/kernel.h> 3 #include <linux/errno.h> 4 #include <linux/fs.h> 5 #include <linux/file.h> 6 #include <linux/mm.h> 7 #include <linux/slab.h> 8 #include <linux/nospec.h> 9 #include <linux/hugetlb.h> 10 #include <linux/compat.h> 11 #include <linux/io_uring.h> 12 13 #include <uapi/linux/io_uring.h> 14 15 #include "io_uring.h" 16 #include "openclose.h" 17 #include "rsrc.h" 18 19 struct io_rsrc_update { 20 struct file *file; 21 u64 arg; 22 u32 nr_args; 23 u32 offset; 24 }; 25 26 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov, 27 struct io_mapped_ubuf **pimu, 28 struct page **last_hpage); 29 30 /* only define max */ 31 #define IORING_MAX_FIXED_FILES (1U << 20) 32 #define IORING_MAX_REG_BUFFERS (1U << 14) 33 34 int __io_account_mem(struct user_struct *user, unsigned long nr_pages) 35 { 36 unsigned long page_limit, cur_pages, new_pages; 37 38 if (!nr_pages) 39 return 0; 40 41 /* Don't allow more pages than we can safely lock */ 42 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 43 44 cur_pages = atomic_long_read(&user->locked_vm); 45 do { 46 new_pages = cur_pages + nr_pages; 47 if (new_pages > page_limit) 48 return -ENOMEM; 49 } while (!atomic_long_try_cmpxchg(&user->locked_vm, 50 &cur_pages, new_pages)); 51 return 0; 52 } 53 54 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) 55 { 56 if (ctx->user) 57 __io_unaccount_mem(ctx->user, nr_pages); 58 59 if (ctx->mm_account) 60 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm); 61 } 62 63 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) 64 { 65 int ret; 66 67 if (ctx->user) { 68 ret = __io_account_mem(ctx->user, nr_pages); 69 if (ret) 70 return ret; 71 } 72 73 if (ctx->mm_account) 74 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm); 75 76 return 0; 77 } 78 79 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, 80 void __user *arg, unsigned index) 81 { 82 struct iovec __user *src; 83 84 #ifdef CONFIG_COMPAT 85 if (ctx->compat) { 86 struct compat_iovec __user *ciovs; 87 struct compat_iovec ciov; 88 89 ciovs = (struct compat_iovec __user *) arg; 90 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) 91 return -EFAULT; 92 93 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); 94 dst->iov_len = ciov.iov_len; 95 return 0; 96 } 97 #endif 98 src = (struct iovec __user *) arg; 99 if (copy_from_user(dst, &src[index], sizeof(*dst))) 100 return -EFAULT; 101 return 0; 102 } 103 104 static int io_buffer_validate(struct iovec *iov) 105 { 106 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1); 107 108 /* 109 * Don't impose further limits on the size and buffer 110 * constraints here, we'll -EINVAL later when IO is 111 * submitted if they are wrong. 112 */ 113 if (!iov->iov_base) 114 return iov->iov_len ? -EFAULT : 0; 115 if (!iov->iov_len) 116 return -EFAULT; 117 118 /* arbitrary limit, but we need something */ 119 if (iov->iov_len > SZ_1G) 120 return -EFAULT; 121 122 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp)) 123 return -EOVERFLOW; 124 125 return 0; 126 } 127 128 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot) 129 { 130 struct io_mapped_ubuf *imu = *slot; 131 unsigned int i; 132 133 if (imu != ctx->dummy_ubuf) { 134 for (i = 0; i < imu->nr_bvecs; i++) 135 unpin_user_page(imu->bvec[i].bv_page); 136 if (imu->acct_pages) 137 io_unaccount_mem(ctx, imu->acct_pages); 138 kvfree(imu); 139 } 140 *slot = NULL; 141 } 142 143 static void io_rsrc_put_work(struct io_rsrc_data *rsrc_data, 144 struct io_rsrc_put *prsrc) 145 { 146 struct io_ring_ctx *ctx = rsrc_data->ctx; 147 148 if (prsrc->tag) 149 io_post_aux_cqe(ctx, prsrc->tag, 0, 0); 150 rsrc_data->do_put(ctx, prsrc); 151 } 152 153 void io_rsrc_node_destroy(struct io_ring_ctx *ctx, struct io_rsrc_node *node) 154 { 155 if (!io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache)) 156 kfree(node); 157 } 158 159 void io_rsrc_node_ref_zero(struct io_rsrc_node *node) 160 __must_hold(&node->rsrc_data->ctx->uring_lock) 161 { 162 struct io_ring_ctx *ctx = node->rsrc_data->ctx; 163 164 while (!list_empty(&ctx->rsrc_ref_list)) { 165 node = list_first_entry(&ctx->rsrc_ref_list, 166 struct io_rsrc_node, node); 167 /* recycle ref nodes in order */ 168 if (node->refs) 169 break; 170 list_del(&node->node); 171 172 if (likely(!node->empty)) 173 io_rsrc_put_work(node->rsrc_data, &node->item); 174 io_rsrc_node_destroy(ctx, node); 175 } 176 if (list_empty(&ctx->rsrc_ref_list) && unlikely(ctx->rsrc_quiesce)) 177 wake_up_all(&ctx->rsrc_quiesce_wq); 178 } 179 180 struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx) 181 { 182 struct io_rsrc_node *ref_node; 183 struct io_cache_entry *entry; 184 185 entry = io_alloc_cache_get(&ctx->rsrc_node_cache); 186 if (entry) { 187 ref_node = container_of(entry, struct io_rsrc_node, cache); 188 } else { 189 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL); 190 if (!ref_node) 191 return NULL; 192 } 193 194 ref_node->rsrc_data = NULL; 195 ref_node->empty = 0; 196 ref_node->refs = 1; 197 return ref_node; 198 } 199 200 __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, 201 struct io_ring_ctx *ctx) 202 { 203 struct io_rsrc_node *backup; 204 DEFINE_WAIT(we); 205 int ret; 206 207 /* As We may drop ->uring_lock, other task may have started quiesce */ 208 if (data->quiesce) 209 return -ENXIO; 210 211 backup = io_rsrc_node_alloc(ctx); 212 if (!backup) 213 return -ENOMEM; 214 ctx->rsrc_node->empty = true; 215 ctx->rsrc_node->rsrc_data = data; 216 list_add_tail(&ctx->rsrc_node->node, &ctx->rsrc_ref_list); 217 io_put_rsrc_node(ctx, ctx->rsrc_node); 218 ctx->rsrc_node = backup; 219 220 if (list_empty(&ctx->rsrc_ref_list)) 221 return 0; 222 223 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { 224 atomic_set(&ctx->cq_wait_nr, 1); 225 smp_mb(); 226 } 227 228 ctx->rsrc_quiesce++; 229 data->quiesce = true; 230 do { 231 prepare_to_wait(&ctx->rsrc_quiesce_wq, &we, TASK_INTERRUPTIBLE); 232 mutex_unlock(&ctx->uring_lock); 233 234 ret = io_run_task_work_sig(ctx); 235 if (ret < 0) { 236 mutex_lock(&ctx->uring_lock); 237 if (list_empty(&ctx->rsrc_ref_list)) 238 ret = 0; 239 break; 240 } 241 242 schedule(); 243 __set_current_state(TASK_RUNNING); 244 mutex_lock(&ctx->uring_lock); 245 ret = 0; 246 } while (!list_empty(&ctx->rsrc_ref_list)); 247 248 finish_wait(&ctx->rsrc_quiesce_wq, &we); 249 data->quiesce = false; 250 ctx->rsrc_quiesce--; 251 252 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { 253 atomic_set(&ctx->cq_wait_nr, 0); 254 smp_mb(); 255 } 256 return ret; 257 } 258 259 static void io_free_page_table(void **table, size_t size) 260 { 261 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); 262 263 for (i = 0; i < nr_tables; i++) 264 kfree(table[i]); 265 kfree(table); 266 } 267 268 static void io_rsrc_data_free(struct io_rsrc_data *data) 269 { 270 size_t size = data->nr * sizeof(data->tags[0][0]); 271 272 if (data->tags) 273 io_free_page_table((void **)data->tags, size); 274 kfree(data); 275 } 276 277 static __cold void **io_alloc_page_table(size_t size) 278 { 279 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); 280 size_t init_size = size; 281 void **table; 282 283 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT); 284 if (!table) 285 return NULL; 286 287 for (i = 0; i < nr_tables; i++) { 288 unsigned int this_size = min_t(size_t, size, PAGE_SIZE); 289 290 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT); 291 if (!table[i]) { 292 io_free_page_table(table, init_size); 293 return NULL; 294 } 295 size -= this_size; 296 } 297 return table; 298 } 299 300 __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, 301 rsrc_put_fn *do_put, u64 __user *utags, 302 unsigned nr, struct io_rsrc_data **pdata) 303 { 304 struct io_rsrc_data *data; 305 int ret = 0; 306 unsigned i; 307 308 data = kzalloc(sizeof(*data), GFP_KERNEL); 309 if (!data) 310 return -ENOMEM; 311 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0])); 312 if (!data->tags) { 313 kfree(data); 314 return -ENOMEM; 315 } 316 317 data->nr = nr; 318 data->ctx = ctx; 319 data->do_put = do_put; 320 if (utags) { 321 ret = -EFAULT; 322 for (i = 0; i < nr; i++) { 323 u64 *tag_slot = io_get_tag_slot(data, i); 324 325 if (copy_from_user(tag_slot, &utags[i], 326 sizeof(*tag_slot))) 327 goto fail; 328 } 329 } 330 *pdata = data; 331 return 0; 332 fail: 333 io_rsrc_data_free(data); 334 return ret; 335 } 336 337 static int __io_sqe_files_update(struct io_ring_ctx *ctx, 338 struct io_uring_rsrc_update2 *up, 339 unsigned nr_args) 340 { 341 u64 __user *tags = u64_to_user_ptr(up->tags); 342 __s32 __user *fds = u64_to_user_ptr(up->data); 343 struct io_rsrc_data *data = ctx->file_data; 344 struct io_fixed_file *file_slot; 345 struct file *file; 346 int fd, i, err = 0; 347 unsigned int done; 348 349 if (!ctx->file_data) 350 return -ENXIO; 351 if (up->offset + nr_args > ctx->nr_user_files) 352 return -EINVAL; 353 354 for (done = 0; done < nr_args; done++) { 355 u64 tag = 0; 356 357 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) || 358 copy_from_user(&fd, &fds[done], sizeof(fd))) { 359 err = -EFAULT; 360 break; 361 } 362 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) { 363 err = -EINVAL; 364 break; 365 } 366 if (fd == IORING_REGISTER_FILES_SKIP) 367 continue; 368 369 i = array_index_nospec(up->offset + done, ctx->nr_user_files); 370 file_slot = io_fixed_file_slot(&ctx->file_table, i); 371 372 if (file_slot->file_ptr) { 373 file = (struct file *)(file_slot->file_ptr & FFS_MASK); 374 err = io_queue_rsrc_removal(data, i, file); 375 if (err) 376 break; 377 file_slot->file_ptr = 0; 378 io_file_bitmap_clear(&ctx->file_table, i); 379 } 380 if (fd != -1) { 381 file = fget(fd); 382 if (!file) { 383 err = -EBADF; 384 break; 385 } 386 /* 387 * Don't allow io_uring instances to be registered. If 388 * UNIX isn't enabled, then this causes a reference 389 * cycle and this instance can never get freed. If UNIX 390 * is enabled we'll handle it just fine, but there's 391 * still no point in allowing a ring fd as it doesn't 392 * support regular read/write anyway. 393 */ 394 if (io_is_uring_fops(file)) { 395 fput(file); 396 err = -EBADF; 397 break; 398 } 399 err = io_scm_file_account(ctx, file); 400 if (err) { 401 fput(file); 402 break; 403 } 404 *io_get_tag_slot(data, i) = tag; 405 io_fixed_file_set(file_slot, file); 406 io_file_bitmap_set(&ctx->file_table, i); 407 } 408 } 409 return done ? done : err; 410 } 411 412 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx, 413 struct io_uring_rsrc_update2 *up, 414 unsigned int nr_args) 415 { 416 u64 __user *tags = u64_to_user_ptr(up->tags); 417 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data); 418 struct page *last_hpage = NULL; 419 __u32 done; 420 int i, err; 421 422 if (!ctx->buf_data) 423 return -ENXIO; 424 if (up->offset + nr_args > ctx->nr_user_bufs) 425 return -EINVAL; 426 427 for (done = 0; done < nr_args; done++) { 428 struct io_mapped_ubuf *imu; 429 u64 tag = 0; 430 431 err = io_copy_iov(ctx, &iov, iovs, done); 432 if (err) 433 break; 434 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) { 435 err = -EFAULT; 436 break; 437 } 438 err = io_buffer_validate(&iov); 439 if (err) 440 break; 441 if (!iov.iov_base && tag) { 442 err = -EINVAL; 443 break; 444 } 445 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage); 446 if (err) 447 break; 448 449 i = array_index_nospec(up->offset + done, ctx->nr_user_bufs); 450 if (ctx->user_bufs[i] != ctx->dummy_ubuf) { 451 err = io_queue_rsrc_removal(ctx->buf_data, i, 452 ctx->user_bufs[i]); 453 if (unlikely(err)) { 454 io_buffer_unmap(ctx, &imu); 455 break; 456 } 457 ctx->user_bufs[i] = ctx->dummy_ubuf; 458 } 459 460 ctx->user_bufs[i] = imu; 461 *io_get_tag_slot(ctx->buf_data, i) = tag; 462 } 463 return done ? done : err; 464 } 465 466 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type, 467 struct io_uring_rsrc_update2 *up, 468 unsigned nr_args) 469 { 470 __u32 tmp; 471 472 lockdep_assert_held(&ctx->uring_lock); 473 474 if (check_add_overflow(up->offset, nr_args, &tmp)) 475 return -EOVERFLOW; 476 477 switch (type) { 478 case IORING_RSRC_FILE: 479 return __io_sqe_files_update(ctx, up, nr_args); 480 case IORING_RSRC_BUFFER: 481 return __io_sqe_buffers_update(ctx, up, nr_args); 482 } 483 return -EINVAL; 484 } 485 486 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg, 487 unsigned nr_args) 488 { 489 struct io_uring_rsrc_update2 up; 490 491 if (!nr_args) 492 return -EINVAL; 493 memset(&up, 0, sizeof(up)); 494 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update))) 495 return -EFAULT; 496 if (up.resv || up.resv2) 497 return -EINVAL; 498 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args); 499 } 500 501 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg, 502 unsigned size, unsigned type) 503 { 504 struct io_uring_rsrc_update2 up; 505 506 if (size != sizeof(up)) 507 return -EINVAL; 508 if (copy_from_user(&up, arg, sizeof(up))) 509 return -EFAULT; 510 if (!up.nr || up.resv || up.resv2) 511 return -EINVAL; 512 return __io_register_rsrc_update(ctx, type, &up, up.nr); 513 } 514 515 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg, 516 unsigned int size, unsigned int type) 517 { 518 struct io_uring_rsrc_register rr; 519 520 /* keep it extendible */ 521 if (size != sizeof(rr)) 522 return -EINVAL; 523 524 memset(&rr, 0, sizeof(rr)); 525 if (copy_from_user(&rr, arg, size)) 526 return -EFAULT; 527 if (!rr.nr || rr.resv2) 528 return -EINVAL; 529 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE) 530 return -EINVAL; 531 532 switch (type) { 533 case IORING_RSRC_FILE: 534 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data) 535 break; 536 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data), 537 rr.nr, u64_to_user_ptr(rr.tags)); 538 case IORING_RSRC_BUFFER: 539 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data) 540 break; 541 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data), 542 rr.nr, u64_to_user_ptr(rr.tags)); 543 } 544 return -EINVAL; 545 } 546 547 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) 548 { 549 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 550 551 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) 552 return -EINVAL; 553 if (sqe->rw_flags || sqe->splice_fd_in) 554 return -EINVAL; 555 556 up->offset = READ_ONCE(sqe->off); 557 up->nr_args = READ_ONCE(sqe->len); 558 if (!up->nr_args) 559 return -EINVAL; 560 up->arg = READ_ONCE(sqe->addr); 561 return 0; 562 } 563 564 static int io_files_update_with_index_alloc(struct io_kiocb *req, 565 unsigned int issue_flags) 566 { 567 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 568 __s32 __user *fds = u64_to_user_ptr(up->arg); 569 unsigned int done; 570 struct file *file; 571 int ret, fd; 572 573 if (!req->ctx->file_data) 574 return -ENXIO; 575 576 for (done = 0; done < up->nr_args; done++) { 577 if (copy_from_user(&fd, &fds[done], sizeof(fd))) { 578 ret = -EFAULT; 579 break; 580 } 581 582 file = fget(fd); 583 if (!file) { 584 ret = -EBADF; 585 break; 586 } 587 ret = io_fixed_fd_install(req, issue_flags, file, 588 IORING_FILE_INDEX_ALLOC); 589 if (ret < 0) 590 break; 591 if (copy_to_user(&fds[done], &ret, sizeof(ret))) { 592 __io_close_fixed(req->ctx, issue_flags, ret); 593 ret = -EFAULT; 594 break; 595 } 596 } 597 598 if (done) 599 return done; 600 return ret; 601 } 602 603 int io_files_update(struct io_kiocb *req, unsigned int issue_flags) 604 { 605 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 606 struct io_ring_ctx *ctx = req->ctx; 607 struct io_uring_rsrc_update2 up2; 608 int ret; 609 610 up2.offset = up->offset; 611 up2.data = up->arg; 612 up2.nr = 0; 613 up2.tags = 0; 614 up2.resv = 0; 615 up2.resv2 = 0; 616 617 if (up->offset == IORING_FILE_INDEX_ALLOC) { 618 ret = io_files_update_with_index_alloc(req, issue_flags); 619 } else { 620 io_ring_submit_lock(ctx, issue_flags); 621 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE, 622 &up2, up->nr_args); 623 io_ring_submit_unlock(ctx, issue_flags); 624 } 625 626 if (ret < 0) 627 req_set_fail(req); 628 io_req_set_res(req, ret, 0); 629 return IOU_OK; 630 } 631 632 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx, void *rsrc) 633 { 634 struct io_ring_ctx *ctx = data->ctx; 635 struct io_rsrc_node *node = ctx->rsrc_node; 636 u64 *tag_slot = io_get_tag_slot(data, idx); 637 638 ctx->rsrc_node = io_rsrc_node_alloc(ctx); 639 if (unlikely(!ctx->rsrc_node)) { 640 ctx->rsrc_node = node; 641 return -ENOMEM; 642 } 643 644 node->item.rsrc = rsrc; 645 node->item.tag = *tag_slot; 646 *tag_slot = 0; 647 648 node->rsrc_data = data; 649 list_add_tail(&node->node, &ctx->rsrc_ref_list); 650 io_put_rsrc_node(ctx, node); 651 return 0; 652 } 653 654 void __io_sqe_files_unregister(struct io_ring_ctx *ctx) 655 { 656 int i; 657 658 for (i = 0; i < ctx->nr_user_files; i++) { 659 struct file *file = io_file_from_index(&ctx->file_table, i); 660 661 /* skip scm accounted files, they'll be freed by ->ring_sock */ 662 if (!file || io_file_need_scm(file)) 663 continue; 664 io_file_bitmap_clear(&ctx->file_table, i); 665 fput(file); 666 } 667 668 #if defined(CONFIG_UNIX) 669 if (ctx->ring_sock) { 670 struct sock *sock = ctx->ring_sock->sk; 671 struct sk_buff *skb; 672 673 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) 674 kfree_skb(skb); 675 } 676 #endif 677 io_free_file_tables(&ctx->file_table); 678 io_file_table_set_alloc_range(ctx, 0, 0); 679 io_rsrc_data_free(ctx->file_data); 680 ctx->file_data = NULL; 681 ctx->nr_user_files = 0; 682 } 683 684 int io_sqe_files_unregister(struct io_ring_ctx *ctx) 685 { 686 unsigned nr = ctx->nr_user_files; 687 int ret; 688 689 if (!ctx->file_data) 690 return -ENXIO; 691 692 /* 693 * Quiesce may unlock ->uring_lock, and while it's not held 694 * prevent new requests using the table. 695 */ 696 ctx->nr_user_files = 0; 697 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx); 698 ctx->nr_user_files = nr; 699 if (!ret) 700 __io_sqe_files_unregister(ctx); 701 return ret; 702 } 703 704 /* 705 * Ensure the UNIX gc is aware of our file set, so we are certain that 706 * the io_uring can be safely unregistered on process exit, even if we have 707 * loops in the file referencing. We account only files that can hold other 708 * files because otherwise they can't form a loop and so are not interesting 709 * for GC. 710 */ 711 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file) 712 { 713 #if defined(CONFIG_UNIX) 714 struct sock *sk = ctx->ring_sock->sk; 715 struct sk_buff_head *head = &sk->sk_receive_queue; 716 struct scm_fp_list *fpl; 717 struct sk_buff *skb; 718 719 if (likely(!io_file_need_scm(file))) 720 return 0; 721 722 /* 723 * See if we can merge this file into an existing skb SCM_RIGHTS 724 * file set. If there's no room, fall back to allocating a new skb 725 * and filling it in. 726 */ 727 spin_lock_irq(&head->lock); 728 skb = skb_peek(head); 729 if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD) 730 __skb_unlink(skb, head); 731 else 732 skb = NULL; 733 spin_unlock_irq(&head->lock); 734 735 if (!skb) { 736 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); 737 if (!fpl) 738 return -ENOMEM; 739 740 skb = alloc_skb(0, GFP_KERNEL); 741 if (!skb) { 742 kfree(fpl); 743 return -ENOMEM; 744 } 745 746 fpl->user = get_uid(current_user()); 747 fpl->max = SCM_MAX_FD; 748 fpl->count = 0; 749 750 UNIXCB(skb).fp = fpl; 751 skb->sk = sk; 752 skb->scm_io_uring = 1; 753 skb->destructor = unix_destruct_scm; 754 refcount_add(skb->truesize, &sk->sk_wmem_alloc); 755 } 756 757 fpl = UNIXCB(skb).fp; 758 fpl->fp[fpl->count++] = get_file(file); 759 unix_inflight(fpl->user, file); 760 skb_queue_head(head, skb); 761 fput(file); 762 #endif 763 return 0; 764 } 765 766 static __cold void io_rsrc_file_scm_put(struct io_ring_ctx *ctx, struct file *file) 767 { 768 #if defined(CONFIG_UNIX) 769 struct sock *sock = ctx->ring_sock->sk; 770 struct sk_buff_head list, *head = &sock->sk_receive_queue; 771 struct sk_buff *skb; 772 int i; 773 774 __skb_queue_head_init(&list); 775 776 /* 777 * Find the skb that holds this file in its SCM_RIGHTS. When found, 778 * remove this entry and rearrange the file array. 779 */ 780 skb = skb_dequeue(head); 781 while (skb) { 782 struct scm_fp_list *fp; 783 784 fp = UNIXCB(skb).fp; 785 for (i = 0; i < fp->count; i++) { 786 int left; 787 788 if (fp->fp[i] != file) 789 continue; 790 791 unix_notinflight(fp->user, fp->fp[i]); 792 left = fp->count - 1 - i; 793 if (left) { 794 memmove(&fp->fp[i], &fp->fp[i + 1], 795 left * sizeof(struct file *)); 796 } 797 fp->count--; 798 if (!fp->count) { 799 kfree_skb(skb); 800 skb = NULL; 801 } else { 802 __skb_queue_tail(&list, skb); 803 } 804 fput(file); 805 file = NULL; 806 break; 807 } 808 809 if (!file) 810 break; 811 812 __skb_queue_tail(&list, skb); 813 814 skb = skb_dequeue(head); 815 } 816 817 if (skb_peek(&list)) { 818 spin_lock_irq(&head->lock); 819 while ((skb = __skb_dequeue(&list)) != NULL) 820 __skb_queue_tail(head, skb); 821 spin_unlock_irq(&head->lock); 822 } 823 #endif 824 } 825 826 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) 827 { 828 struct file *file = prsrc->file; 829 830 if (likely(!io_file_need_scm(file))) 831 fput(file); 832 else 833 io_rsrc_file_scm_put(ctx, file); 834 } 835 836 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, 837 unsigned nr_args, u64 __user *tags) 838 { 839 __s32 __user *fds = (__s32 __user *) arg; 840 struct file *file; 841 int fd, ret; 842 unsigned i; 843 844 if (ctx->file_data) 845 return -EBUSY; 846 if (!nr_args) 847 return -EINVAL; 848 if (nr_args > IORING_MAX_FIXED_FILES) 849 return -EMFILE; 850 if (nr_args > rlimit(RLIMIT_NOFILE)) 851 return -EMFILE; 852 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args, 853 &ctx->file_data); 854 if (ret) 855 return ret; 856 857 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) { 858 io_rsrc_data_free(ctx->file_data); 859 ctx->file_data = NULL; 860 return -ENOMEM; 861 } 862 863 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { 864 struct io_fixed_file *file_slot; 865 866 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) { 867 ret = -EFAULT; 868 goto fail; 869 } 870 /* allow sparse sets */ 871 if (!fds || fd == -1) { 872 ret = -EINVAL; 873 if (unlikely(*io_get_tag_slot(ctx->file_data, i))) 874 goto fail; 875 continue; 876 } 877 878 file = fget(fd); 879 ret = -EBADF; 880 if (unlikely(!file)) 881 goto fail; 882 883 /* 884 * Don't allow io_uring instances to be registered. If UNIX 885 * isn't enabled, then this causes a reference cycle and this 886 * instance can never get freed. If UNIX is enabled we'll 887 * handle it just fine, but there's still no point in allowing 888 * a ring fd as it doesn't support regular read/write anyway. 889 */ 890 if (io_is_uring_fops(file)) { 891 fput(file); 892 goto fail; 893 } 894 ret = io_scm_file_account(ctx, file); 895 if (ret) { 896 fput(file); 897 goto fail; 898 } 899 file_slot = io_fixed_file_slot(&ctx->file_table, i); 900 io_fixed_file_set(file_slot, file); 901 io_file_bitmap_set(&ctx->file_table, i); 902 } 903 904 /* default it to the whole table */ 905 io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files); 906 return 0; 907 fail: 908 __io_sqe_files_unregister(ctx); 909 return ret; 910 } 911 912 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) 913 { 914 io_buffer_unmap(ctx, &prsrc->buf); 915 prsrc->buf = NULL; 916 } 917 918 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx) 919 { 920 unsigned int i; 921 922 for (i = 0; i < ctx->nr_user_bufs; i++) 923 io_buffer_unmap(ctx, &ctx->user_bufs[i]); 924 kfree(ctx->user_bufs); 925 io_rsrc_data_free(ctx->buf_data); 926 ctx->user_bufs = NULL; 927 ctx->buf_data = NULL; 928 ctx->nr_user_bufs = 0; 929 } 930 931 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx) 932 { 933 unsigned nr = ctx->nr_user_bufs; 934 int ret; 935 936 if (!ctx->buf_data) 937 return -ENXIO; 938 939 /* 940 * Quiesce may unlock ->uring_lock, and while it's not held 941 * prevent new requests using the table. 942 */ 943 ctx->nr_user_bufs = 0; 944 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx); 945 ctx->nr_user_bufs = nr; 946 if (!ret) 947 __io_sqe_buffers_unregister(ctx); 948 return ret; 949 } 950 951 /* 952 * Not super efficient, but this is just a registration time. And we do cache 953 * the last compound head, so generally we'll only do a full search if we don't 954 * match that one. 955 * 956 * We check if the given compound head page has already been accounted, to 957 * avoid double accounting it. This allows us to account the full size of the 958 * page, not just the constituent pages of a huge page. 959 */ 960 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages, 961 int nr_pages, struct page *hpage) 962 { 963 int i, j; 964 965 /* check current page array */ 966 for (i = 0; i < nr_pages; i++) { 967 if (!PageCompound(pages[i])) 968 continue; 969 if (compound_head(pages[i]) == hpage) 970 return true; 971 } 972 973 /* check previously registered pages */ 974 for (i = 0; i < ctx->nr_user_bufs; i++) { 975 struct io_mapped_ubuf *imu = ctx->user_bufs[i]; 976 977 for (j = 0; j < imu->nr_bvecs; j++) { 978 if (!PageCompound(imu->bvec[j].bv_page)) 979 continue; 980 if (compound_head(imu->bvec[j].bv_page) == hpage) 981 return true; 982 } 983 } 984 985 return false; 986 } 987 988 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages, 989 int nr_pages, struct io_mapped_ubuf *imu, 990 struct page **last_hpage) 991 { 992 int i, ret; 993 994 imu->acct_pages = 0; 995 for (i = 0; i < nr_pages; i++) { 996 if (!PageCompound(pages[i])) { 997 imu->acct_pages++; 998 } else { 999 struct page *hpage; 1000 1001 hpage = compound_head(pages[i]); 1002 if (hpage == *last_hpage) 1003 continue; 1004 *last_hpage = hpage; 1005 if (headpage_already_acct(ctx, pages, i, hpage)) 1006 continue; 1007 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT; 1008 } 1009 } 1010 1011 if (!imu->acct_pages) 1012 return 0; 1013 1014 ret = io_account_mem(ctx, imu->acct_pages); 1015 if (ret) 1016 imu->acct_pages = 0; 1017 return ret; 1018 } 1019 1020 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages) 1021 { 1022 unsigned long start, end, nr_pages; 1023 struct vm_area_struct **vmas = NULL; 1024 struct page **pages = NULL; 1025 int i, pret, ret = -ENOMEM; 1026 1027 end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT; 1028 start = ubuf >> PAGE_SHIFT; 1029 nr_pages = end - start; 1030 1031 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); 1032 if (!pages) 1033 goto done; 1034 1035 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *), 1036 GFP_KERNEL); 1037 if (!vmas) 1038 goto done; 1039 1040 ret = 0; 1041 mmap_read_lock(current->mm); 1042 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM, 1043 pages, vmas); 1044 if (pret == nr_pages) { 1045 struct file *file = vmas[0]->vm_file; 1046 1047 /* don't support file backed memory */ 1048 for (i = 0; i < nr_pages; i++) { 1049 if (vmas[i]->vm_file != file) { 1050 ret = -EINVAL; 1051 break; 1052 } 1053 if (!file) 1054 continue; 1055 if (!vma_is_shmem(vmas[i]) && !is_file_hugepages(file)) { 1056 ret = -EOPNOTSUPP; 1057 break; 1058 } 1059 } 1060 *npages = nr_pages; 1061 } else { 1062 ret = pret < 0 ? pret : -EFAULT; 1063 } 1064 mmap_read_unlock(current->mm); 1065 if (ret) { 1066 /* 1067 * if we did partial map, or found file backed vmas, 1068 * release any pages we did get 1069 */ 1070 if (pret > 0) 1071 unpin_user_pages(pages, pret); 1072 goto done; 1073 } 1074 ret = 0; 1075 done: 1076 kvfree(vmas); 1077 if (ret < 0) { 1078 kvfree(pages); 1079 pages = ERR_PTR(ret); 1080 } 1081 return pages; 1082 } 1083 1084 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov, 1085 struct io_mapped_ubuf **pimu, 1086 struct page **last_hpage) 1087 { 1088 struct io_mapped_ubuf *imu = NULL; 1089 struct page **pages = NULL; 1090 unsigned long off; 1091 size_t size; 1092 int ret, nr_pages, i; 1093 struct folio *folio = NULL; 1094 1095 *pimu = ctx->dummy_ubuf; 1096 if (!iov->iov_base) 1097 return 0; 1098 1099 ret = -ENOMEM; 1100 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len, 1101 &nr_pages); 1102 if (IS_ERR(pages)) { 1103 ret = PTR_ERR(pages); 1104 pages = NULL; 1105 goto done; 1106 } 1107 1108 /* If it's a huge page, try to coalesce them into a single bvec entry */ 1109 if (nr_pages > 1) { 1110 folio = page_folio(pages[0]); 1111 for (i = 1; i < nr_pages; i++) { 1112 if (page_folio(pages[i]) != folio) { 1113 folio = NULL; 1114 break; 1115 } 1116 } 1117 if (folio) { 1118 /* 1119 * The pages are bound to the folio, it doesn't 1120 * actually unpin them but drops all but one reference, 1121 * which is usually put down by io_buffer_unmap(). 1122 * Note, needs a better helper. 1123 */ 1124 unpin_user_pages(&pages[1], nr_pages - 1); 1125 nr_pages = 1; 1126 } 1127 } 1128 1129 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL); 1130 if (!imu) 1131 goto done; 1132 1133 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage); 1134 if (ret) { 1135 unpin_user_pages(pages, nr_pages); 1136 goto done; 1137 } 1138 1139 off = (unsigned long) iov->iov_base & ~PAGE_MASK; 1140 size = iov->iov_len; 1141 /* store original address for later verification */ 1142 imu->ubuf = (unsigned long) iov->iov_base; 1143 imu->ubuf_end = imu->ubuf + iov->iov_len; 1144 imu->nr_bvecs = nr_pages; 1145 *pimu = imu; 1146 ret = 0; 1147 1148 if (folio) { 1149 bvec_set_page(&imu->bvec[0], pages[0], size, off); 1150 goto done; 1151 } 1152 for (i = 0; i < nr_pages; i++) { 1153 size_t vec_len; 1154 1155 vec_len = min_t(size_t, size, PAGE_SIZE - off); 1156 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off); 1157 off = 0; 1158 size -= vec_len; 1159 } 1160 done: 1161 if (ret) 1162 kvfree(imu); 1163 kvfree(pages); 1164 return ret; 1165 } 1166 1167 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args) 1168 { 1169 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL); 1170 return ctx->user_bufs ? 0 : -ENOMEM; 1171 } 1172 1173 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg, 1174 unsigned int nr_args, u64 __user *tags) 1175 { 1176 struct page *last_hpage = NULL; 1177 struct io_rsrc_data *data; 1178 int i, ret; 1179 struct iovec iov; 1180 1181 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16)); 1182 1183 if (ctx->user_bufs) 1184 return -EBUSY; 1185 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS) 1186 return -EINVAL; 1187 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data); 1188 if (ret) 1189 return ret; 1190 ret = io_buffers_map_alloc(ctx, nr_args); 1191 if (ret) { 1192 io_rsrc_data_free(data); 1193 return ret; 1194 } 1195 1196 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) { 1197 if (arg) { 1198 ret = io_copy_iov(ctx, &iov, arg, i); 1199 if (ret) 1200 break; 1201 ret = io_buffer_validate(&iov); 1202 if (ret) 1203 break; 1204 } else { 1205 memset(&iov, 0, sizeof(iov)); 1206 } 1207 1208 if (!iov.iov_base && *io_get_tag_slot(data, i)) { 1209 ret = -EINVAL; 1210 break; 1211 } 1212 1213 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i], 1214 &last_hpage); 1215 if (ret) 1216 break; 1217 } 1218 1219 WARN_ON_ONCE(ctx->buf_data); 1220 1221 ctx->buf_data = data; 1222 if (ret) 1223 __io_sqe_buffers_unregister(ctx); 1224 return ret; 1225 } 1226 1227 int io_import_fixed(int ddir, struct iov_iter *iter, 1228 struct io_mapped_ubuf *imu, 1229 u64 buf_addr, size_t len) 1230 { 1231 u64 buf_end; 1232 size_t offset; 1233 1234 if (WARN_ON_ONCE(!imu)) 1235 return -EFAULT; 1236 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end))) 1237 return -EFAULT; 1238 /* not inside the mapped region */ 1239 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end)) 1240 return -EFAULT; 1241 1242 /* 1243 * Might not be a start of buffer, set size appropriately 1244 * and advance us to the beginning. 1245 */ 1246 offset = buf_addr - imu->ubuf; 1247 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len); 1248 1249 if (offset) { 1250 /* 1251 * Don't use iov_iter_advance() here, as it's really slow for 1252 * using the latter parts of a big fixed buffer - it iterates 1253 * over each segment manually. We can cheat a bit here, because 1254 * we know that: 1255 * 1256 * 1) it's a BVEC iter, we set it up 1257 * 2) all bvecs are PAGE_SIZE in size, except potentially the 1258 * first and last bvec 1259 * 1260 * So just find our index, and adjust the iterator afterwards. 1261 * If the offset is within the first bvec (or the whole first 1262 * bvec, just use iov_iter_advance(). This makes it easier 1263 * since we can just skip the first segment, which may not 1264 * be PAGE_SIZE aligned. 1265 */ 1266 const struct bio_vec *bvec = imu->bvec; 1267 1268 if (offset <= bvec->bv_len) { 1269 /* 1270 * Note, huge pages buffers consists of one large 1271 * bvec entry and should always go this way. The other 1272 * branch doesn't expect non PAGE_SIZE'd chunks. 1273 */ 1274 iter->bvec = bvec; 1275 iter->nr_segs = bvec->bv_len; 1276 iter->count -= offset; 1277 iter->iov_offset = offset; 1278 } else { 1279 unsigned long seg_skip; 1280 1281 /* skip first vec */ 1282 offset -= bvec->bv_len; 1283 seg_skip = 1 + (offset >> PAGE_SHIFT); 1284 1285 iter->bvec = bvec + seg_skip; 1286 iter->nr_segs -= seg_skip; 1287 iter->count -= bvec->bv_len + offset; 1288 iter->iov_offset = offset & ~PAGE_MASK; 1289 } 1290 } 1291 1292 return 0; 1293 } 1294