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_one(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 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node) 154 { 155 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data; 156 157 if (likely(!ref_node->empty)) 158 io_rsrc_put_work_one(rsrc_data, &ref_node->item); 159 160 io_rsrc_node_destroy(rsrc_data->ctx, ref_node); 161 } 162 163 void io_rsrc_node_destroy(struct io_ring_ctx *ctx, struct io_rsrc_node *node) 164 { 165 if (!io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache)) 166 kfree(node); 167 } 168 169 void io_rsrc_node_ref_zero(struct io_rsrc_node *node) 170 __must_hold(&node->rsrc_data->ctx->uring_lock) 171 { 172 struct io_ring_ctx *ctx = node->rsrc_data->ctx; 173 174 while (!list_empty(&ctx->rsrc_ref_list)) { 175 node = list_first_entry(&ctx->rsrc_ref_list, 176 struct io_rsrc_node, node); 177 /* recycle ref nodes in order */ 178 if (node->refs) 179 break; 180 list_del(&node->node); 181 __io_rsrc_put_work(node); 182 } 183 if (list_empty(&ctx->rsrc_ref_list) && unlikely(ctx->rsrc_quiesce)) 184 wake_up_all(&ctx->rsrc_quiesce_wq); 185 } 186 187 struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx) 188 { 189 struct io_rsrc_node *ref_node; 190 struct io_cache_entry *entry; 191 192 entry = io_alloc_cache_get(&ctx->rsrc_node_cache); 193 if (entry) { 194 ref_node = container_of(entry, struct io_rsrc_node, cache); 195 } else { 196 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL); 197 if (!ref_node) 198 return NULL; 199 } 200 201 ref_node->rsrc_data = NULL; 202 ref_node->empty = 0; 203 ref_node->refs = 1; 204 return ref_node; 205 } 206 207 __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, 208 struct io_ring_ctx *ctx) 209 { 210 struct io_rsrc_node *backup; 211 DEFINE_WAIT(we); 212 int ret; 213 214 /* As We may drop ->uring_lock, other task may have started quiesce */ 215 if (data->quiesce) 216 return -ENXIO; 217 218 backup = io_rsrc_node_alloc(ctx); 219 if (!backup) 220 return -ENOMEM; 221 ctx->rsrc_node->empty = true; 222 ctx->rsrc_node->rsrc_data = data; 223 list_add_tail(&ctx->rsrc_node->node, &ctx->rsrc_ref_list); 224 io_put_rsrc_node(ctx, ctx->rsrc_node); 225 ctx->rsrc_node = backup; 226 227 if (list_empty(&ctx->rsrc_ref_list)) 228 return 0; 229 230 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { 231 atomic_set(&ctx->cq_wait_nr, 1); 232 smp_mb(); 233 } 234 235 ctx->rsrc_quiesce++; 236 data->quiesce = true; 237 do { 238 prepare_to_wait(&ctx->rsrc_quiesce_wq, &we, TASK_INTERRUPTIBLE); 239 mutex_unlock(&ctx->uring_lock); 240 241 ret = io_run_task_work_sig(ctx); 242 if (ret < 0) { 243 mutex_lock(&ctx->uring_lock); 244 if (list_empty(&ctx->rsrc_ref_list)) 245 ret = 0; 246 break; 247 } 248 249 schedule(); 250 __set_current_state(TASK_RUNNING); 251 mutex_lock(&ctx->uring_lock); 252 ret = 0; 253 } while (!list_empty(&ctx->rsrc_ref_list)); 254 255 finish_wait(&ctx->rsrc_quiesce_wq, &we); 256 data->quiesce = false; 257 ctx->rsrc_quiesce--; 258 259 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { 260 atomic_set(&ctx->cq_wait_nr, 0); 261 smp_mb(); 262 } 263 return ret; 264 } 265 266 static void io_free_page_table(void **table, size_t size) 267 { 268 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); 269 270 for (i = 0; i < nr_tables; i++) 271 kfree(table[i]); 272 kfree(table); 273 } 274 275 static void io_rsrc_data_free(struct io_rsrc_data *data) 276 { 277 size_t size = data->nr * sizeof(data->tags[0][0]); 278 279 if (data->tags) 280 io_free_page_table((void **)data->tags, size); 281 kfree(data); 282 } 283 284 static __cold void **io_alloc_page_table(size_t size) 285 { 286 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); 287 size_t init_size = size; 288 void **table; 289 290 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT); 291 if (!table) 292 return NULL; 293 294 for (i = 0; i < nr_tables; i++) { 295 unsigned int this_size = min_t(size_t, size, PAGE_SIZE); 296 297 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT); 298 if (!table[i]) { 299 io_free_page_table(table, init_size); 300 return NULL; 301 } 302 size -= this_size; 303 } 304 return table; 305 } 306 307 __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, 308 rsrc_put_fn *do_put, u64 __user *utags, 309 unsigned nr, struct io_rsrc_data **pdata) 310 { 311 struct io_rsrc_data *data; 312 int ret = 0; 313 unsigned i; 314 315 data = kzalloc(sizeof(*data), GFP_KERNEL); 316 if (!data) 317 return -ENOMEM; 318 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0])); 319 if (!data->tags) { 320 kfree(data); 321 return -ENOMEM; 322 } 323 324 data->nr = nr; 325 data->ctx = ctx; 326 data->do_put = do_put; 327 if (utags) { 328 ret = -EFAULT; 329 for (i = 0; i < nr; i++) { 330 u64 *tag_slot = io_get_tag_slot(data, i); 331 332 if (copy_from_user(tag_slot, &utags[i], 333 sizeof(*tag_slot))) 334 goto fail; 335 } 336 } 337 *pdata = data; 338 return 0; 339 fail: 340 io_rsrc_data_free(data); 341 return ret; 342 } 343 344 static int __io_sqe_files_update(struct io_ring_ctx *ctx, 345 struct io_uring_rsrc_update2 *up, 346 unsigned nr_args) 347 { 348 u64 __user *tags = u64_to_user_ptr(up->tags); 349 __s32 __user *fds = u64_to_user_ptr(up->data); 350 struct io_rsrc_data *data = ctx->file_data; 351 struct io_fixed_file *file_slot; 352 struct file *file; 353 int fd, i, err = 0; 354 unsigned int done; 355 356 if (!ctx->file_data) 357 return -ENXIO; 358 if (up->offset + nr_args > ctx->nr_user_files) 359 return -EINVAL; 360 361 for (done = 0; done < nr_args; done++) { 362 u64 tag = 0; 363 364 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) || 365 copy_from_user(&fd, &fds[done], sizeof(fd))) { 366 err = -EFAULT; 367 break; 368 } 369 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) { 370 err = -EINVAL; 371 break; 372 } 373 if (fd == IORING_REGISTER_FILES_SKIP) 374 continue; 375 376 i = array_index_nospec(up->offset + done, ctx->nr_user_files); 377 file_slot = io_fixed_file_slot(&ctx->file_table, i); 378 379 if (file_slot->file_ptr) { 380 file = (struct file *)(file_slot->file_ptr & FFS_MASK); 381 err = io_queue_rsrc_removal(data, i, file); 382 if (err) 383 break; 384 file_slot->file_ptr = 0; 385 io_file_bitmap_clear(&ctx->file_table, i); 386 } 387 if (fd != -1) { 388 file = fget(fd); 389 if (!file) { 390 err = -EBADF; 391 break; 392 } 393 /* 394 * Don't allow io_uring instances to be registered. If 395 * UNIX isn't enabled, then this causes a reference 396 * cycle and this instance can never get freed. If UNIX 397 * is enabled we'll handle it just fine, but there's 398 * still no point in allowing a ring fd as it doesn't 399 * support regular read/write anyway. 400 */ 401 if (io_is_uring_fops(file)) { 402 fput(file); 403 err = -EBADF; 404 break; 405 } 406 err = io_scm_file_account(ctx, file); 407 if (err) { 408 fput(file); 409 break; 410 } 411 *io_get_tag_slot(data, i) = tag; 412 io_fixed_file_set(file_slot, file); 413 io_file_bitmap_set(&ctx->file_table, i); 414 } 415 } 416 return done ? done : err; 417 } 418 419 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx, 420 struct io_uring_rsrc_update2 *up, 421 unsigned int nr_args) 422 { 423 u64 __user *tags = u64_to_user_ptr(up->tags); 424 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data); 425 struct page *last_hpage = NULL; 426 __u32 done; 427 int i, err; 428 429 if (!ctx->buf_data) 430 return -ENXIO; 431 if (up->offset + nr_args > ctx->nr_user_bufs) 432 return -EINVAL; 433 434 for (done = 0; done < nr_args; done++) { 435 struct io_mapped_ubuf *imu; 436 u64 tag = 0; 437 438 err = io_copy_iov(ctx, &iov, iovs, done); 439 if (err) 440 break; 441 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) { 442 err = -EFAULT; 443 break; 444 } 445 err = io_buffer_validate(&iov); 446 if (err) 447 break; 448 if (!iov.iov_base && tag) { 449 err = -EINVAL; 450 break; 451 } 452 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage); 453 if (err) 454 break; 455 456 i = array_index_nospec(up->offset + done, ctx->nr_user_bufs); 457 if (ctx->user_bufs[i] != ctx->dummy_ubuf) { 458 err = io_queue_rsrc_removal(ctx->buf_data, i, 459 ctx->user_bufs[i]); 460 if (unlikely(err)) { 461 io_buffer_unmap(ctx, &imu); 462 break; 463 } 464 ctx->user_bufs[i] = ctx->dummy_ubuf; 465 } 466 467 ctx->user_bufs[i] = imu; 468 *io_get_tag_slot(ctx->buf_data, i) = tag; 469 } 470 return done ? done : err; 471 } 472 473 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type, 474 struct io_uring_rsrc_update2 *up, 475 unsigned nr_args) 476 { 477 __u32 tmp; 478 479 lockdep_assert_held(&ctx->uring_lock); 480 481 if (check_add_overflow(up->offset, nr_args, &tmp)) 482 return -EOVERFLOW; 483 484 switch (type) { 485 case IORING_RSRC_FILE: 486 return __io_sqe_files_update(ctx, up, nr_args); 487 case IORING_RSRC_BUFFER: 488 return __io_sqe_buffers_update(ctx, up, nr_args); 489 } 490 return -EINVAL; 491 } 492 493 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg, 494 unsigned nr_args) 495 { 496 struct io_uring_rsrc_update2 up; 497 498 if (!nr_args) 499 return -EINVAL; 500 memset(&up, 0, sizeof(up)); 501 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update))) 502 return -EFAULT; 503 if (up.resv || up.resv2) 504 return -EINVAL; 505 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args); 506 } 507 508 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg, 509 unsigned size, unsigned type) 510 { 511 struct io_uring_rsrc_update2 up; 512 513 if (size != sizeof(up)) 514 return -EINVAL; 515 if (copy_from_user(&up, arg, sizeof(up))) 516 return -EFAULT; 517 if (!up.nr || up.resv || up.resv2) 518 return -EINVAL; 519 return __io_register_rsrc_update(ctx, type, &up, up.nr); 520 } 521 522 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg, 523 unsigned int size, unsigned int type) 524 { 525 struct io_uring_rsrc_register rr; 526 527 /* keep it extendible */ 528 if (size != sizeof(rr)) 529 return -EINVAL; 530 531 memset(&rr, 0, sizeof(rr)); 532 if (copy_from_user(&rr, arg, size)) 533 return -EFAULT; 534 if (!rr.nr || rr.resv2) 535 return -EINVAL; 536 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE) 537 return -EINVAL; 538 539 switch (type) { 540 case IORING_RSRC_FILE: 541 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data) 542 break; 543 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data), 544 rr.nr, u64_to_user_ptr(rr.tags)); 545 case IORING_RSRC_BUFFER: 546 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data) 547 break; 548 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data), 549 rr.nr, u64_to_user_ptr(rr.tags)); 550 } 551 return -EINVAL; 552 } 553 554 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) 555 { 556 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 557 558 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) 559 return -EINVAL; 560 if (sqe->rw_flags || sqe->splice_fd_in) 561 return -EINVAL; 562 563 up->offset = READ_ONCE(sqe->off); 564 up->nr_args = READ_ONCE(sqe->len); 565 if (!up->nr_args) 566 return -EINVAL; 567 up->arg = READ_ONCE(sqe->addr); 568 return 0; 569 } 570 571 static int io_files_update_with_index_alloc(struct io_kiocb *req, 572 unsigned int issue_flags) 573 { 574 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 575 __s32 __user *fds = u64_to_user_ptr(up->arg); 576 unsigned int done; 577 struct file *file; 578 int ret, fd; 579 580 if (!req->ctx->file_data) 581 return -ENXIO; 582 583 for (done = 0; done < up->nr_args; done++) { 584 if (copy_from_user(&fd, &fds[done], sizeof(fd))) { 585 ret = -EFAULT; 586 break; 587 } 588 589 file = fget(fd); 590 if (!file) { 591 ret = -EBADF; 592 break; 593 } 594 ret = io_fixed_fd_install(req, issue_flags, file, 595 IORING_FILE_INDEX_ALLOC); 596 if (ret < 0) 597 break; 598 if (copy_to_user(&fds[done], &ret, sizeof(ret))) { 599 __io_close_fixed(req->ctx, issue_flags, ret); 600 ret = -EFAULT; 601 break; 602 } 603 } 604 605 if (done) 606 return done; 607 return ret; 608 } 609 610 int io_files_update(struct io_kiocb *req, unsigned int issue_flags) 611 { 612 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 613 struct io_ring_ctx *ctx = req->ctx; 614 struct io_uring_rsrc_update2 up2; 615 int ret; 616 617 up2.offset = up->offset; 618 up2.data = up->arg; 619 up2.nr = 0; 620 up2.tags = 0; 621 up2.resv = 0; 622 up2.resv2 = 0; 623 624 if (up->offset == IORING_FILE_INDEX_ALLOC) { 625 ret = io_files_update_with_index_alloc(req, issue_flags); 626 } else { 627 io_ring_submit_lock(ctx, issue_flags); 628 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE, 629 &up2, up->nr_args); 630 io_ring_submit_unlock(ctx, issue_flags); 631 } 632 633 if (ret < 0) 634 req_set_fail(req); 635 io_req_set_res(req, ret, 0); 636 return IOU_OK; 637 } 638 639 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx, void *rsrc) 640 { 641 struct io_ring_ctx *ctx = data->ctx; 642 struct io_rsrc_node *node = ctx->rsrc_node; 643 u64 *tag_slot = io_get_tag_slot(data, idx); 644 645 ctx->rsrc_node = io_rsrc_node_alloc(ctx); 646 if (unlikely(!ctx->rsrc_node)) { 647 ctx->rsrc_node = node; 648 return -ENOMEM; 649 } 650 651 node->item.rsrc = rsrc; 652 node->item.tag = *tag_slot; 653 *tag_slot = 0; 654 655 node->rsrc_data = data; 656 list_add_tail(&node->node, &ctx->rsrc_ref_list); 657 io_put_rsrc_node(ctx, node); 658 return 0; 659 } 660 661 void __io_sqe_files_unregister(struct io_ring_ctx *ctx) 662 { 663 int i; 664 665 for (i = 0; i < ctx->nr_user_files; i++) { 666 struct file *file = io_file_from_index(&ctx->file_table, i); 667 668 /* skip scm accounted files, they'll be freed by ->ring_sock */ 669 if (!file || io_file_need_scm(file)) 670 continue; 671 io_file_bitmap_clear(&ctx->file_table, i); 672 fput(file); 673 } 674 675 #if defined(CONFIG_UNIX) 676 if (ctx->ring_sock) { 677 struct sock *sock = ctx->ring_sock->sk; 678 struct sk_buff *skb; 679 680 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) 681 kfree_skb(skb); 682 } 683 #endif 684 io_free_file_tables(&ctx->file_table); 685 io_file_table_set_alloc_range(ctx, 0, 0); 686 io_rsrc_data_free(ctx->file_data); 687 ctx->file_data = NULL; 688 ctx->nr_user_files = 0; 689 } 690 691 int io_sqe_files_unregister(struct io_ring_ctx *ctx) 692 { 693 unsigned nr = ctx->nr_user_files; 694 int ret; 695 696 if (!ctx->file_data) 697 return -ENXIO; 698 699 /* 700 * Quiesce may unlock ->uring_lock, and while it's not held 701 * prevent new requests using the table. 702 */ 703 ctx->nr_user_files = 0; 704 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx); 705 ctx->nr_user_files = nr; 706 if (!ret) 707 __io_sqe_files_unregister(ctx); 708 return ret; 709 } 710 711 /* 712 * Ensure the UNIX gc is aware of our file set, so we are certain that 713 * the io_uring can be safely unregistered on process exit, even if we have 714 * loops in the file referencing. We account only files that can hold other 715 * files because otherwise they can't form a loop and so are not interesting 716 * for GC. 717 */ 718 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file) 719 { 720 #if defined(CONFIG_UNIX) 721 struct sock *sk = ctx->ring_sock->sk; 722 struct sk_buff_head *head = &sk->sk_receive_queue; 723 struct scm_fp_list *fpl; 724 struct sk_buff *skb; 725 726 if (likely(!io_file_need_scm(file))) 727 return 0; 728 729 /* 730 * See if we can merge this file into an existing skb SCM_RIGHTS 731 * file set. If there's no room, fall back to allocating a new skb 732 * and filling it in. 733 */ 734 spin_lock_irq(&head->lock); 735 skb = skb_peek(head); 736 if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD) 737 __skb_unlink(skb, head); 738 else 739 skb = NULL; 740 spin_unlock_irq(&head->lock); 741 742 if (!skb) { 743 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); 744 if (!fpl) 745 return -ENOMEM; 746 747 skb = alloc_skb(0, GFP_KERNEL); 748 if (!skb) { 749 kfree(fpl); 750 return -ENOMEM; 751 } 752 753 fpl->user = get_uid(current_user()); 754 fpl->max = SCM_MAX_FD; 755 fpl->count = 0; 756 757 UNIXCB(skb).fp = fpl; 758 skb->sk = sk; 759 skb->scm_io_uring = 1; 760 skb->destructor = unix_destruct_scm; 761 refcount_add(skb->truesize, &sk->sk_wmem_alloc); 762 } 763 764 fpl = UNIXCB(skb).fp; 765 fpl->fp[fpl->count++] = get_file(file); 766 unix_inflight(fpl->user, file); 767 skb_queue_head(head, skb); 768 fput(file); 769 #endif 770 return 0; 771 } 772 773 static __cold void io_rsrc_file_scm_put(struct io_ring_ctx *ctx, struct file *file) 774 { 775 #if defined(CONFIG_UNIX) 776 struct sock *sock = ctx->ring_sock->sk; 777 struct sk_buff_head list, *head = &sock->sk_receive_queue; 778 struct sk_buff *skb; 779 int i; 780 781 __skb_queue_head_init(&list); 782 783 /* 784 * Find the skb that holds this file in its SCM_RIGHTS. When found, 785 * remove this entry and rearrange the file array. 786 */ 787 skb = skb_dequeue(head); 788 while (skb) { 789 struct scm_fp_list *fp; 790 791 fp = UNIXCB(skb).fp; 792 for (i = 0; i < fp->count; i++) { 793 int left; 794 795 if (fp->fp[i] != file) 796 continue; 797 798 unix_notinflight(fp->user, fp->fp[i]); 799 left = fp->count - 1 - i; 800 if (left) { 801 memmove(&fp->fp[i], &fp->fp[i + 1], 802 left * sizeof(struct file *)); 803 } 804 fp->count--; 805 if (!fp->count) { 806 kfree_skb(skb); 807 skb = NULL; 808 } else { 809 __skb_queue_tail(&list, skb); 810 } 811 fput(file); 812 file = NULL; 813 break; 814 } 815 816 if (!file) 817 break; 818 819 __skb_queue_tail(&list, skb); 820 821 skb = skb_dequeue(head); 822 } 823 824 if (skb_peek(&list)) { 825 spin_lock_irq(&head->lock); 826 while ((skb = __skb_dequeue(&list)) != NULL) 827 __skb_queue_tail(head, skb); 828 spin_unlock_irq(&head->lock); 829 } 830 #endif 831 } 832 833 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) 834 { 835 struct file *file = prsrc->file; 836 837 if (likely(!io_file_need_scm(file))) 838 fput(file); 839 else 840 io_rsrc_file_scm_put(ctx, file); 841 } 842 843 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, 844 unsigned nr_args, u64 __user *tags) 845 { 846 __s32 __user *fds = (__s32 __user *) arg; 847 struct file *file; 848 int fd, ret; 849 unsigned i; 850 851 if (ctx->file_data) 852 return -EBUSY; 853 if (!nr_args) 854 return -EINVAL; 855 if (nr_args > IORING_MAX_FIXED_FILES) 856 return -EMFILE; 857 if (nr_args > rlimit(RLIMIT_NOFILE)) 858 return -EMFILE; 859 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args, 860 &ctx->file_data); 861 if (ret) 862 return ret; 863 864 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) { 865 io_rsrc_data_free(ctx->file_data); 866 ctx->file_data = NULL; 867 return -ENOMEM; 868 } 869 870 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { 871 struct io_fixed_file *file_slot; 872 873 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) { 874 ret = -EFAULT; 875 goto fail; 876 } 877 /* allow sparse sets */ 878 if (!fds || fd == -1) { 879 ret = -EINVAL; 880 if (unlikely(*io_get_tag_slot(ctx->file_data, i))) 881 goto fail; 882 continue; 883 } 884 885 file = fget(fd); 886 ret = -EBADF; 887 if (unlikely(!file)) 888 goto fail; 889 890 /* 891 * Don't allow io_uring instances to be registered. If UNIX 892 * isn't enabled, then this causes a reference cycle and this 893 * instance can never get freed. If UNIX is enabled we'll 894 * handle it just fine, but there's still no point in allowing 895 * a ring fd as it doesn't support regular read/write anyway. 896 */ 897 if (io_is_uring_fops(file)) { 898 fput(file); 899 goto fail; 900 } 901 ret = io_scm_file_account(ctx, file); 902 if (ret) { 903 fput(file); 904 goto fail; 905 } 906 file_slot = io_fixed_file_slot(&ctx->file_table, i); 907 io_fixed_file_set(file_slot, file); 908 io_file_bitmap_set(&ctx->file_table, i); 909 } 910 911 /* default it to the whole table */ 912 io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files); 913 return 0; 914 fail: 915 __io_sqe_files_unregister(ctx); 916 return ret; 917 } 918 919 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) 920 { 921 io_buffer_unmap(ctx, &prsrc->buf); 922 prsrc->buf = NULL; 923 } 924 925 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx) 926 { 927 unsigned int i; 928 929 for (i = 0; i < ctx->nr_user_bufs; i++) 930 io_buffer_unmap(ctx, &ctx->user_bufs[i]); 931 kfree(ctx->user_bufs); 932 io_rsrc_data_free(ctx->buf_data); 933 ctx->user_bufs = NULL; 934 ctx->buf_data = NULL; 935 ctx->nr_user_bufs = 0; 936 } 937 938 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx) 939 { 940 unsigned nr = ctx->nr_user_bufs; 941 int ret; 942 943 if (!ctx->buf_data) 944 return -ENXIO; 945 946 /* 947 * Quiesce may unlock ->uring_lock, and while it's not held 948 * prevent new requests using the table. 949 */ 950 ctx->nr_user_bufs = 0; 951 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx); 952 ctx->nr_user_bufs = nr; 953 if (!ret) 954 __io_sqe_buffers_unregister(ctx); 955 return ret; 956 } 957 958 /* 959 * Not super efficient, but this is just a registration time. And we do cache 960 * the last compound head, so generally we'll only do a full search if we don't 961 * match that one. 962 * 963 * We check if the given compound head page has already been accounted, to 964 * avoid double accounting it. This allows us to account the full size of the 965 * page, not just the constituent pages of a huge page. 966 */ 967 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages, 968 int nr_pages, struct page *hpage) 969 { 970 int i, j; 971 972 /* check current page array */ 973 for (i = 0; i < nr_pages; i++) { 974 if (!PageCompound(pages[i])) 975 continue; 976 if (compound_head(pages[i]) == hpage) 977 return true; 978 } 979 980 /* check previously registered pages */ 981 for (i = 0; i < ctx->nr_user_bufs; i++) { 982 struct io_mapped_ubuf *imu = ctx->user_bufs[i]; 983 984 for (j = 0; j < imu->nr_bvecs; j++) { 985 if (!PageCompound(imu->bvec[j].bv_page)) 986 continue; 987 if (compound_head(imu->bvec[j].bv_page) == hpage) 988 return true; 989 } 990 } 991 992 return false; 993 } 994 995 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages, 996 int nr_pages, struct io_mapped_ubuf *imu, 997 struct page **last_hpage) 998 { 999 int i, ret; 1000 1001 imu->acct_pages = 0; 1002 for (i = 0; i < nr_pages; i++) { 1003 if (!PageCompound(pages[i])) { 1004 imu->acct_pages++; 1005 } else { 1006 struct page *hpage; 1007 1008 hpage = compound_head(pages[i]); 1009 if (hpage == *last_hpage) 1010 continue; 1011 *last_hpage = hpage; 1012 if (headpage_already_acct(ctx, pages, i, hpage)) 1013 continue; 1014 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT; 1015 } 1016 } 1017 1018 if (!imu->acct_pages) 1019 return 0; 1020 1021 ret = io_account_mem(ctx, imu->acct_pages); 1022 if (ret) 1023 imu->acct_pages = 0; 1024 return ret; 1025 } 1026 1027 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages) 1028 { 1029 unsigned long start, end, nr_pages; 1030 struct vm_area_struct **vmas = NULL; 1031 struct page **pages = NULL; 1032 int i, pret, ret = -ENOMEM; 1033 1034 end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT; 1035 start = ubuf >> PAGE_SHIFT; 1036 nr_pages = end - start; 1037 1038 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); 1039 if (!pages) 1040 goto done; 1041 1042 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *), 1043 GFP_KERNEL); 1044 if (!vmas) 1045 goto done; 1046 1047 ret = 0; 1048 mmap_read_lock(current->mm); 1049 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM, 1050 pages, vmas); 1051 if (pret == nr_pages) { 1052 struct file *file = vmas[0]->vm_file; 1053 1054 /* don't support file backed memory */ 1055 for (i = 0; i < nr_pages; i++) { 1056 if (vmas[i]->vm_file != file) { 1057 ret = -EINVAL; 1058 break; 1059 } 1060 if (!file) 1061 continue; 1062 if (!vma_is_shmem(vmas[i]) && !is_file_hugepages(file)) { 1063 ret = -EOPNOTSUPP; 1064 break; 1065 } 1066 } 1067 *npages = nr_pages; 1068 } else { 1069 ret = pret < 0 ? pret : -EFAULT; 1070 } 1071 mmap_read_unlock(current->mm); 1072 if (ret) { 1073 /* 1074 * if we did partial map, or found file backed vmas, 1075 * release any pages we did get 1076 */ 1077 if (pret > 0) 1078 unpin_user_pages(pages, pret); 1079 goto done; 1080 } 1081 ret = 0; 1082 done: 1083 kvfree(vmas); 1084 if (ret < 0) { 1085 kvfree(pages); 1086 pages = ERR_PTR(ret); 1087 } 1088 return pages; 1089 } 1090 1091 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov, 1092 struct io_mapped_ubuf **pimu, 1093 struct page **last_hpage) 1094 { 1095 struct io_mapped_ubuf *imu = NULL; 1096 struct page **pages = NULL; 1097 unsigned long off; 1098 size_t size; 1099 int ret, nr_pages, i; 1100 struct folio *folio = NULL; 1101 1102 *pimu = ctx->dummy_ubuf; 1103 if (!iov->iov_base) 1104 return 0; 1105 1106 ret = -ENOMEM; 1107 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len, 1108 &nr_pages); 1109 if (IS_ERR(pages)) { 1110 ret = PTR_ERR(pages); 1111 pages = NULL; 1112 goto done; 1113 } 1114 1115 /* If it's a huge page, try to coalesce them into a single bvec entry */ 1116 if (nr_pages > 1) { 1117 folio = page_folio(pages[0]); 1118 for (i = 1; i < nr_pages; i++) { 1119 if (page_folio(pages[i]) != folio) { 1120 folio = NULL; 1121 break; 1122 } 1123 } 1124 if (folio) { 1125 /* 1126 * The pages are bound to the folio, it doesn't 1127 * actually unpin them but drops all but one reference, 1128 * which is usually put down by io_buffer_unmap(). 1129 * Note, needs a better helper. 1130 */ 1131 unpin_user_pages(&pages[1], nr_pages - 1); 1132 nr_pages = 1; 1133 } 1134 } 1135 1136 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL); 1137 if (!imu) 1138 goto done; 1139 1140 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage); 1141 if (ret) { 1142 unpin_user_pages(pages, nr_pages); 1143 goto done; 1144 } 1145 1146 off = (unsigned long) iov->iov_base & ~PAGE_MASK; 1147 size = iov->iov_len; 1148 /* store original address for later verification */ 1149 imu->ubuf = (unsigned long) iov->iov_base; 1150 imu->ubuf_end = imu->ubuf + iov->iov_len; 1151 imu->nr_bvecs = nr_pages; 1152 *pimu = imu; 1153 ret = 0; 1154 1155 if (folio) { 1156 bvec_set_page(&imu->bvec[0], pages[0], size, off); 1157 goto done; 1158 } 1159 for (i = 0; i < nr_pages; i++) { 1160 size_t vec_len; 1161 1162 vec_len = min_t(size_t, size, PAGE_SIZE - off); 1163 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off); 1164 off = 0; 1165 size -= vec_len; 1166 } 1167 done: 1168 if (ret) 1169 kvfree(imu); 1170 kvfree(pages); 1171 return ret; 1172 } 1173 1174 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args) 1175 { 1176 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL); 1177 return ctx->user_bufs ? 0 : -ENOMEM; 1178 } 1179 1180 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg, 1181 unsigned int nr_args, u64 __user *tags) 1182 { 1183 struct page *last_hpage = NULL; 1184 struct io_rsrc_data *data; 1185 int i, ret; 1186 struct iovec iov; 1187 1188 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16)); 1189 1190 if (ctx->user_bufs) 1191 return -EBUSY; 1192 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS) 1193 return -EINVAL; 1194 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data); 1195 if (ret) 1196 return ret; 1197 ret = io_buffers_map_alloc(ctx, nr_args); 1198 if (ret) { 1199 io_rsrc_data_free(data); 1200 return ret; 1201 } 1202 1203 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) { 1204 if (arg) { 1205 ret = io_copy_iov(ctx, &iov, arg, i); 1206 if (ret) 1207 break; 1208 ret = io_buffer_validate(&iov); 1209 if (ret) 1210 break; 1211 } else { 1212 memset(&iov, 0, sizeof(iov)); 1213 } 1214 1215 if (!iov.iov_base && *io_get_tag_slot(data, i)) { 1216 ret = -EINVAL; 1217 break; 1218 } 1219 1220 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i], 1221 &last_hpage); 1222 if (ret) 1223 break; 1224 } 1225 1226 WARN_ON_ONCE(ctx->buf_data); 1227 1228 ctx->buf_data = data; 1229 if (ret) 1230 __io_sqe_buffers_unregister(ctx); 1231 return ret; 1232 } 1233 1234 int io_import_fixed(int ddir, struct iov_iter *iter, 1235 struct io_mapped_ubuf *imu, 1236 u64 buf_addr, size_t len) 1237 { 1238 u64 buf_end; 1239 size_t offset; 1240 1241 if (WARN_ON_ONCE(!imu)) 1242 return -EFAULT; 1243 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end))) 1244 return -EFAULT; 1245 /* not inside the mapped region */ 1246 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end)) 1247 return -EFAULT; 1248 1249 /* 1250 * Might not be a start of buffer, set size appropriately 1251 * and advance us to the beginning. 1252 */ 1253 offset = buf_addr - imu->ubuf; 1254 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len); 1255 1256 if (offset) { 1257 /* 1258 * Don't use iov_iter_advance() here, as it's really slow for 1259 * using the latter parts of a big fixed buffer - it iterates 1260 * over each segment manually. We can cheat a bit here, because 1261 * we know that: 1262 * 1263 * 1) it's a BVEC iter, we set it up 1264 * 2) all bvecs are PAGE_SIZE in size, except potentially the 1265 * first and last bvec 1266 * 1267 * So just find our index, and adjust the iterator afterwards. 1268 * If the offset is within the first bvec (or the whole first 1269 * bvec, just use iov_iter_advance(). This makes it easier 1270 * since we can just skip the first segment, which may not 1271 * be PAGE_SIZE aligned. 1272 */ 1273 const struct bio_vec *bvec = imu->bvec; 1274 1275 if (offset <= bvec->bv_len) { 1276 /* 1277 * Note, huge pages buffers consists of one large 1278 * bvec entry and should always go this way. The other 1279 * branch doesn't expect non PAGE_SIZE'd chunks. 1280 */ 1281 iter->bvec = bvec; 1282 iter->nr_segs = bvec->bv_len; 1283 iter->count -= offset; 1284 iter->iov_offset = offset; 1285 } else { 1286 unsigned long seg_skip; 1287 1288 /* skip first vec */ 1289 offset -= bvec->bv_len; 1290 seg_skip = 1 + (offset >> PAGE_SHIFT); 1291 1292 iter->bvec = bvec + seg_skip; 1293 iter->nr_segs -= seg_skip; 1294 iter->count -= bvec->bv_len + offset; 1295 iter->iov_offset = offset & ~PAGE_MASK; 1296 } 1297 } 1298 1299 return 0; 1300 } 1301