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 static inline bool io_put_rsrc_data_ref(struct io_rsrc_data *rsrc_data) 35 { 36 return !--rsrc_data->refs; 37 } 38 39 int __io_account_mem(struct user_struct *user, unsigned long nr_pages) 40 { 41 unsigned long page_limit, cur_pages, new_pages; 42 43 if (!nr_pages) 44 return 0; 45 46 /* Don't allow more pages than we can safely lock */ 47 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 48 49 cur_pages = atomic_long_read(&user->locked_vm); 50 do { 51 new_pages = cur_pages + nr_pages; 52 if (new_pages > page_limit) 53 return -ENOMEM; 54 } while (!atomic_long_try_cmpxchg(&user->locked_vm, 55 &cur_pages, new_pages)); 56 return 0; 57 } 58 59 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) 60 { 61 if (ctx->user) 62 __io_unaccount_mem(ctx->user, nr_pages); 63 64 if (ctx->mm_account) 65 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm); 66 } 67 68 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) 69 { 70 int ret; 71 72 if (ctx->user) { 73 ret = __io_account_mem(ctx->user, nr_pages); 74 if (ret) 75 return ret; 76 } 77 78 if (ctx->mm_account) 79 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm); 80 81 return 0; 82 } 83 84 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, 85 void __user *arg, unsigned index) 86 { 87 struct iovec __user *src; 88 89 #ifdef CONFIG_COMPAT 90 if (ctx->compat) { 91 struct compat_iovec __user *ciovs; 92 struct compat_iovec ciov; 93 94 ciovs = (struct compat_iovec __user *) arg; 95 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) 96 return -EFAULT; 97 98 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); 99 dst->iov_len = ciov.iov_len; 100 return 0; 101 } 102 #endif 103 src = (struct iovec __user *) arg; 104 if (copy_from_user(dst, &src[index], sizeof(*dst))) 105 return -EFAULT; 106 return 0; 107 } 108 109 static int io_buffer_validate(struct iovec *iov) 110 { 111 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1); 112 113 /* 114 * Don't impose further limits on the size and buffer 115 * constraints here, we'll -EINVAL later when IO is 116 * submitted if they are wrong. 117 */ 118 if (!iov->iov_base) 119 return iov->iov_len ? -EFAULT : 0; 120 if (!iov->iov_len) 121 return -EFAULT; 122 123 /* arbitrary limit, but we need something */ 124 if (iov->iov_len > SZ_1G) 125 return -EFAULT; 126 127 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp)) 128 return -EOVERFLOW; 129 130 return 0; 131 } 132 133 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot) 134 { 135 struct io_mapped_ubuf *imu = *slot; 136 unsigned int i; 137 138 if (imu != ctx->dummy_ubuf) { 139 for (i = 0; i < imu->nr_bvecs; i++) 140 unpin_user_page(imu->bvec[i].bv_page); 141 if (imu->acct_pages) 142 io_unaccount_mem(ctx, imu->acct_pages); 143 kvfree(imu); 144 } 145 *slot = NULL; 146 } 147 148 static void io_rsrc_put_work_one(struct io_rsrc_data *rsrc_data, 149 struct io_rsrc_put *prsrc) 150 { 151 struct io_ring_ctx *ctx = rsrc_data->ctx; 152 153 if (prsrc->tag) 154 io_post_aux_cqe(ctx, prsrc->tag, 0, 0); 155 rsrc_data->do_put(ctx, prsrc); 156 } 157 158 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node) 159 { 160 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data; 161 struct io_rsrc_put *prsrc, *tmp; 162 163 if (ref_node->inline_items) 164 io_rsrc_put_work_one(rsrc_data, &ref_node->item); 165 166 list_for_each_entry_safe(prsrc, tmp, &ref_node->item_list, list) { 167 list_del(&prsrc->list); 168 io_rsrc_put_work_one(rsrc_data, prsrc); 169 kfree(prsrc); 170 } 171 172 io_rsrc_node_destroy(rsrc_data->ctx, ref_node); 173 if (io_put_rsrc_data_ref(rsrc_data)) 174 complete(&rsrc_data->done); 175 } 176 177 void io_wait_rsrc_data(struct io_rsrc_data *data) 178 { 179 if (data && !io_put_rsrc_data_ref(data)) 180 wait_for_completion(&data->done); 181 } 182 183 void io_rsrc_node_destroy(struct io_ring_ctx *ctx, struct io_rsrc_node *node) 184 { 185 if (!io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache)) 186 kfree(node); 187 } 188 189 void io_rsrc_node_ref_zero(struct io_rsrc_node *node) 190 __must_hold(&node->rsrc_data->ctx->uring_lock) 191 { 192 struct io_ring_ctx *ctx = node->rsrc_data->ctx; 193 194 while (!list_empty(&ctx->rsrc_ref_list)) { 195 node = list_first_entry(&ctx->rsrc_ref_list, 196 struct io_rsrc_node, node); 197 /* recycle ref nodes in order */ 198 if (node->refs) 199 break; 200 list_del(&node->node); 201 __io_rsrc_put_work(node); 202 } 203 } 204 205 struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx) 206 { 207 struct io_rsrc_node *ref_node; 208 struct io_cache_entry *entry; 209 210 entry = io_alloc_cache_get(&ctx->rsrc_node_cache); 211 if (entry) { 212 ref_node = container_of(entry, struct io_rsrc_node, cache); 213 } else { 214 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL); 215 if (!ref_node) 216 return NULL; 217 } 218 219 ref_node->rsrc_data = NULL; 220 ref_node->refs = 1; 221 INIT_LIST_HEAD(&ref_node->node); 222 INIT_LIST_HEAD(&ref_node->item_list); 223 ref_node->inline_items = 0; 224 return ref_node; 225 } 226 227 void io_rsrc_node_switch(struct io_ring_ctx *ctx, 228 struct io_rsrc_data *data_to_kill) 229 __must_hold(&ctx->uring_lock) 230 { 231 struct io_rsrc_node *node = ctx->rsrc_node; 232 struct io_rsrc_node *backup = io_rsrc_node_alloc(ctx); 233 234 if (WARN_ON_ONCE(!backup)) 235 return; 236 237 data_to_kill->refs++; 238 node->rsrc_data = data_to_kill; 239 list_add_tail(&node->node, &ctx->rsrc_ref_list); 240 /* put master ref */ 241 io_put_rsrc_node(ctx, node); 242 ctx->rsrc_node = backup; 243 } 244 245 int io_rsrc_node_switch_start(struct io_ring_ctx *ctx) 246 { 247 if (io_alloc_cache_empty(&ctx->rsrc_node_cache)) { 248 struct io_rsrc_node *node = kzalloc(sizeof(*node), GFP_KERNEL); 249 250 if (!node) 251 return -ENOMEM; 252 io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache); 253 } 254 return 0; 255 } 256 257 __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, 258 struct io_ring_ctx *ctx) 259 { 260 int ret; 261 262 /* As we may drop ->uring_lock, other task may have started quiesce */ 263 if (data->quiesce) 264 return -ENXIO; 265 ret = io_rsrc_node_switch_start(ctx); 266 if (ret) 267 return ret; 268 io_rsrc_node_switch(ctx, data); 269 270 /* kill initial ref */ 271 if (io_put_rsrc_data_ref(data)) 272 return 0; 273 274 data->quiesce = true; 275 mutex_unlock(&ctx->uring_lock); 276 do { 277 ret = io_run_task_work_sig(ctx); 278 if (ret < 0) { 279 mutex_lock(&ctx->uring_lock); 280 if (!data->refs) { 281 ret = 0; 282 } else { 283 /* restore the master reference */ 284 data->refs++; 285 } 286 break; 287 } 288 ret = wait_for_completion_interruptible(&data->done); 289 if (!ret) { 290 mutex_lock(&ctx->uring_lock); 291 if (!data->refs) 292 break; 293 /* 294 * it has been revived by another thread while 295 * we were unlocked 296 */ 297 mutex_unlock(&ctx->uring_lock); 298 } 299 } while (1); 300 data->quiesce = false; 301 302 return ret; 303 } 304 305 static void io_free_page_table(void **table, size_t size) 306 { 307 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); 308 309 for (i = 0; i < nr_tables; i++) 310 kfree(table[i]); 311 kfree(table); 312 } 313 314 static void io_rsrc_data_free(struct io_rsrc_data *data) 315 { 316 size_t size = data->nr * sizeof(data->tags[0][0]); 317 318 if (data->tags) 319 io_free_page_table((void **)data->tags, size); 320 kfree(data); 321 } 322 323 static __cold void **io_alloc_page_table(size_t size) 324 { 325 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); 326 size_t init_size = size; 327 void **table; 328 329 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT); 330 if (!table) 331 return NULL; 332 333 for (i = 0; i < nr_tables; i++) { 334 unsigned int this_size = min_t(size_t, size, PAGE_SIZE); 335 336 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT); 337 if (!table[i]) { 338 io_free_page_table(table, init_size); 339 return NULL; 340 } 341 size -= this_size; 342 } 343 return table; 344 } 345 346 __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, 347 rsrc_put_fn *do_put, u64 __user *utags, 348 unsigned nr, struct io_rsrc_data **pdata) 349 { 350 struct io_rsrc_data *data; 351 int ret = 0; 352 unsigned i; 353 354 data = kzalloc(sizeof(*data), GFP_KERNEL); 355 if (!data) 356 return -ENOMEM; 357 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0])); 358 if (!data->tags) { 359 kfree(data); 360 return -ENOMEM; 361 } 362 363 data->nr = nr; 364 data->ctx = ctx; 365 data->do_put = do_put; 366 data->refs = 1; 367 if (utags) { 368 ret = -EFAULT; 369 for (i = 0; i < nr; i++) { 370 u64 *tag_slot = io_get_tag_slot(data, i); 371 372 if (copy_from_user(tag_slot, &utags[i], 373 sizeof(*tag_slot))) 374 goto fail; 375 } 376 } 377 init_completion(&data->done); 378 *pdata = data; 379 return 0; 380 fail: 381 io_rsrc_data_free(data); 382 return ret; 383 } 384 385 static int __io_sqe_files_update(struct io_ring_ctx *ctx, 386 struct io_uring_rsrc_update2 *up, 387 unsigned nr_args) 388 { 389 u64 __user *tags = u64_to_user_ptr(up->tags); 390 __s32 __user *fds = u64_to_user_ptr(up->data); 391 struct io_rsrc_data *data = ctx->file_data; 392 struct io_fixed_file *file_slot; 393 struct file *file; 394 int fd, i, err = 0; 395 unsigned int done; 396 bool needs_switch = false; 397 398 if (!ctx->file_data) 399 return -ENXIO; 400 if (up->offset + nr_args > ctx->nr_user_files) 401 return -EINVAL; 402 403 for (done = 0; done < nr_args; done++) { 404 u64 tag = 0; 405 406 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) || 407 copy_from_user(&fd, &fds[done], sizeof(fd))) { 408 err = -EFAULT; 409 break; 410 } 411 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) { 412 err = -EINVAL; 413 break; 414 } 415 if (fd == IORING_REGISTER_FILES_SKIP) 416 continue; 417 418 i = array_index_nospec(up->offset + done, ctx->nr_user_files); 419 file_slot = io_fixed_file_slot(&ctx->file_table, i); 420 421 if (file_slot->file_ptr) { 422 file = (struct file *)(file_slot->file_ptr & FFS_MASK); 423 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file); 424 if (err) 425 break; 426 file_slot->file_ptr = 0; 427 io_file_bitmap_clear(&ctx->file_table, i); 428 needs_switch = true; 429 } 430 if (fd != -1) { 431 file = fget(fd); 432 if (!file) { 433 err = -EBADF; 434 break; 435 } 436 /* 437 * Don't allow io_uring instances to be registered. If 438 * UNIX isn't enabled, then this causes a reference 439 * cycle and this instance can never get freed. If UNIX 440 * is enabled we'll handle it just fine, but there's 441 * still no point in allowing a ring fd as it doesn't 442 * support regular read/write anyway. 443 */ 444 if (io_is_uring_fops(file)) { 445 fput(file); 446 err = -EBADF; 447 break; 448 } 449 err = io_scm_file_account(ctx, file); 450 if (err) { 451 fput(file); 452 break; 453 } 454 *io_get_tag_slot(data, i) = tag; 455 io_fixed_file_set(file_slot, file); 456 io_file_bitmap_set(&ctx->file_table, i); 457 } 458 } 459 460 if (needs_switch) 461 io_rsrc_node_switch(ctx, data); 462 return done ? done : err; 463 } 464 465 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx, 466 struct io_uring_rsrc_update2 *up, 467 unsigned int nr_args) 468 { 469 u64 __user *tags = u64_to_user_ptr(up->tags); 470 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data); 471 struct page *last_hpage = NULL; 472 bool needs_switch = false; 473 __u32 done; 474 int i, err; 475 476 if (!ctx->buf_data) 477 return -ENXIO; 478 if (up->offset + nr_args > ctx->nr_user_bufs) 479 return -EINVAL; 480 481 for (done = 0; done < nr_args; done++) { 482 struct io_mapped_ubuf *imu; 483 int offset = up->offset + done; 484 u64 tag = 0; 485 486 err = io_copy_iov(ctx, &iov, iovs, done); 487 if (err) 488 break; 489 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) { 490 err = -EFAULT; 491 break; 492 } 493 err = io_buffer_validate(&iov); 494 if (err) 495 break; 496 if (!iov.iov_base && tag) { 497 err = -EINVAL; 498 break; 499 } 500 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage); 501 if (err) 502 break; 503 504 i = array_index_nospec(offset, ctx->nr_user_bufs); 505 if (ctx->user_bufs[i] != ctx->dummy_ubuf) { 506 err = io_queue_rsrc_removal(ctx->buf_data, i, 507 ctx->rsrc_node, ctx->user_bufs[i]); 508 if (unlikely(err)) { 509 io_buffer_unmap(ctx, &imu); 510 break; 511 } 512 ctx->user_bufs[i] = ctx->dummy_ubuf; 513 needs_switch = true; 514 } 515 516 ctx->user_bufs[i] = imu; 517 *io_get_tag_slot(ctx->buf_data, i) = tag; 518 } 519 520 if (needs_switch) 521 io_rsrc_node_switch(ctx, ctx->buf_data); 522 return done ? done : err; 523 } 524 525 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type, 526 struct io_uring_rsrc_update2 *up, 527 unsigned nr_args) 528 { 529 __u32 tmp; 530 int err; 531 532 lockdep_assert_held(&ctx->uring_lock); 533 534 if (check_add_overflow(up->offset, nr_args, &tmp)) 535 return -EOVERFLOW; 536 err = io_rsrc_node_switch_start(ctx); 537 if (err) 538 return err; 539 540 switch (type) { 541 case IORING_RSRC_FILE: 542 return __io_sqe_files_update(ctx, up, nr_args); 543 case IORING_RSRC_BUFFER: 544 return __io_sqe_buffers_update(ctx, up, nr_args); 545 } 546 return -EINVAL; 547 } 548 549 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg, 550 unsigned nr_args) 551 { 552 struct io_uring_rsrc_update2 up; 553 554 if (!nr_args) 555 return -EINVAL; 556 memset(&up, 0, sizeof(up)); 557 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update))) 558 return -EFAULT; 559 if (up.resv || up.resv2) 560 return -EINVAL; 561 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args); 562 } 563 564 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg, 565 unsigned size, unsigned type) 566 { 567 struct io_uring_rsrc_update2 up; 568 569 if (size != sizeof(up)) 570 return -EINVAL; 571 if (copy_from_user(&up, arg, sizeof(up))) 572 return -EFAULT; 573 if (!up.nr || up.resv || up.resv2) 574 return -EINVAL; 575 return __io_register_rsrc_update(ctx, type, &up, up.nr); 576 } 577 578 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg, 579 unsigned int size, unsigned int type) 580 { 581 struct io_uring_rsrc_register rr; 582 583 /* keep it extendible */ 584 if (size != sizeof(rr)) 585 return -EINVAL; 586 587 memset(&rr, 0, sizeof(rr)); 588 if (copy_from_user(&rr, arg, size)) 589 return -EFAULT; 590 if (!rr.nr || rr.resv2) 591 return -EINVAL; 592 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE) 593 return -EINVAL; 594 595 switch (type) { 596 case IORING_RSRC_FILE: 597 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data) 598 break; 599 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data), 600 rr.nr, u64_to_user_ptr(rr.tags)); 601 case IORING_RSRC_BUFFER: 602 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data) 603 break; 604 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data), 605 rr.nr, u64_to_user_ptr(rr.tags)); 606 } 607 return -EINVAL; 608 } 609 610 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) 611 { 612 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 613 614 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) 615 return -EINVAL; 616 if (sqe->rw_flags || sqe->splice_fd_in) 617 return -EINVAL; 618 619 up->offset = READ_ONCE(sqe->off); 620 up->nr_args = READ_ONCE(sqe->len); 621 if (!up->nr_args) 622 return -EINVAL; 623 up->arg = READ_ONCE(sqe->addr); 624 return 0; 625 } 626 627 static int io_files_update_with_index_alloc(struct io_kiocb *req, 628 unsigned int issue_flags) 629 { 630 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 631 __s32 __user *fds = u64_to_user_ptr(up->arg); 632 unsigned int done; 633 struct file *file; 634 int ret, fd; 635 636 if (!req->ctx->file_data) 637 return -ENXIO; 638 639 for (done = 0; done < up->nr_args; done++) { 640 if (copy_from_user(&fd, &fds[done], sizeof(fd))) { 641 ret = -EFAULT; 642 break; 643 } 644 645 file = fget(fd); 646 if (!file) { 647 ret = -EBADF; 648 break; 649 } 650 ret = io_fixed_fd_install(req, issue_flags, file, 651 IORING_FILE_INDEX_ALLOC); 652 if (ret < 0) 653 break; 654 if (copy_to_user(&fds[done], &ret, sizeof(ret))) { 655 __io_close_fixed(req->ctx, issue_flags, ret); 656 ret = -EFAULT; 657 break; 658 } 659 } 660 661 if (done) 662 return done; 663 return ret; 664 } 665 666 int io_files_update(struct io_kiocb *req, unsigned int issue_flags) 667 { 668 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 669 struct io_ring_ctx *ctx = req->ctx; 670 struct io_uring_rsrc_update2 up2; 671 int ret; 672 673 up2.offset = up->offset; 674 up2.data = up->arg; 675 up2.nr = 0; 676 up2.tags = 0; 677 up2.resv = 0; 678 up2.resv2 = 0; 679 680 if (up->offset == IORING_FILE_INDEX_ALLOC) { 681 ret = io_files_update_with_index_alloc(req, issue_flags); 682 } else { 683 io_ring_submit_lock(ctx, issue_flags); 684 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE, 685 &up2, up->nr_args); 686 io_ring_submit_unlock(ctx, issue_flags); 687 } 688 689 if (ret < 0) 690 req_set_fail(req); 691 io_req_set_res(req, ret, 0); 692 return IOU_OK; 693 } 694 695 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx, 696 struct io_rsrc_node *node, void *rsrc) 697 { 698 u64 *tag_slot = io_get_tag_slot(data, idx); 699 struct io_rsrc_put *prsrc; 700 bool inline_item = true; 701 702 if (!node->inline_items) { 703 prsrc = &node->item; 704 node->inline_items++; 705 } else { 706 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL); 707 if (!prsrc) 708 return -ENOMEM; 709 inline_item = false; 710 } 711 712 prsrc->tag = *tag_slot; 713 *tag_slot = 0; 714 prsrc->rsrc = rsrc; 715 if (!inline_item) 716 list_add(&prsrc->list, &node->item_list); 717 return 0; 718 } 719 720 void __io_sqe_files_unregister(struct io_ring_ctx *ctx) 721 { 722 int i; 723 724 for (i = 0; i < ctx->nr_user_files; i++) { 725 struct file *file = io_file_from_index(&ctx->file_table, i); 726 727 /* skip scm accounted files, they'll be freed by ->ring_sock */ 728 if (!file || io_file_need_scm(file)) 729 continue; 730 io_file_bitmap_clear(&ctx->file_table, i); 731 fput(file); 732 } 733 734 #if defined(CONFIG_UNIX) 735 if (ctx->ring_sock) { 736 struct sock *sock = ctx->ring_sock->sk; 737 struct sk_buff *skb; 738 739 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) 740 kfree_skb(skb); 741 } 742 #endif 743 io_free_file_tables(&ctx->file_table); 744 io_file_table_set_alloc_range(ctx, 0, 0); 745 io_rsrc_data_free(ctx->file_data); 746 ctx->file_data = NULL; 747 ctx->nr_user_files = 0; 748 } 749 750 int io_sqe_files_unregister(struct io_ring_ctx *ctx) 751 { 752 unsigned nr = ctx->nr_user_files; 753 int ret; 754 755 if (!ctx->file_data) 756 return -ENXIO; 757 758 /* 759 * Quiesce may unlock ->uring_lock, and while it's not held 760 * prevent new requests using the table. 761 */ 762 ctx->nr_user_files = 0; 763 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx); 764 ctx->nr_user_files = nr; 765 if (!ret) 766 __io_sqe_files_unregister(ctx); 767 return ret; 768 } 769 770 /* 771 * Ensure the UNIX gc is aware of our file set, so we are certain that 772 * the io_uring can be safely unregistered on process exit, even if we have 773 * loops in the file referencing. We account only files that can hold other 774 * files because otherwise they can't form a loop and so are not interesting 775 * for GC. 776 */ 777 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file) 778 { 779 #if defined(CONFIG_UNIX) 780 struct sock *sk = ctx->ring_sock->sk; 781 struct sk_buff_head *head = &sk->sk_receive_queue; 782 struct scm_fp_list *fpl; 783 struct sk_buff *skb; 784 785 if (likely(!io_file_need_scm(file))) 786 return 0; 787 788 /* 789 * See if we can merge this file into an existing skb SCM_RIGHTS 790 * file set. If there's no room, fall back to allocating a new skb 791 * and filling it in. 792 */ 793 spin_lock_irq(&head->lock); 794 skb = skb_peek(head); 795 if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD) 796 __skb_unlink(skb, head); 797 else 798 skb = NULL; 799 spin_unlock_irq(&head->lock); 800 801 if (!skb) { 802 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); 803 if (!fpl) 804 return -ENOMEM; 805 806 skb = alloc_skb(0, GFP_KERNEL); 807 if (!skb) { 808 kfree(fpl); 809 return -ENOMEM; 810 } 811 812 fpl->user = get_uid(current_user()); 813 fpl->max = SCM_MAX_FD; 814 fpl->count = 0; 815 816 UNIXCB(skb).fp = fpl; 817 skb->sk = sk; 818 skb->scm_io_uring = 1; 819 skb->destructor = unix_destruct_scm; 820 refcount_add(skb->truesize, &sk->sk_wmem_alloc); 821 } 822 823 fpl = UNIXCB(skb).fp; 824 fpl->fp[fpl->count++] = get_file(file); 825 unix_inflight(fpl->user, file); 826 skb_queue_head(head, skb); 827 fput(file); 828 #endif 829 return 0; 830 } 831 832 static __cold void io_rsrc_file_scm_put(struct io_ring_ctx *ctx, struct file *file) 833 { 834 #if defined(CONFIG_UNIX) 835 struct sock *sock = ctx->ring_sock->sk; 836 struct sk_buff_head list, *head = &sock->sk_receive_queue; 837 struct sk_buff *skb; 838 int i; 839 840 __skb_queue_head_init(&list); 841 842 /* 843 * Find the skb that holds this file in its SCM_RIGHTS. When found, 844 * remove this entry and rearrange the file array. 845 */ 846 skb = skb_dequeue(head); 847 while (skb) { 848 struct scm_fp_list *fp; 849 850 fp = UNIXCB(skb).fp; 851 for (i = 0; i < fp->count; i++) { 852 int left; 853 854 if (fp->fp[i] != file) 855 continue; 856 857 unix_notinflight(fp->user, fp->fp[i]); 858 left = fp->count - 1 - i; 859 if (left) { 860 memmove(&fp->fp[i], &fp->fp[i + 1], 861 left * sizeof(struct file *)); 862 } 863 fp->count--; 864 if (!fp->count) { 865 kfree_skb(skb); 866 skb = NULL; 867 } else { 868 __skb_queue_tail(&list, skb); 869 } 870 fput(file); 871 file = NULL; 872 break; 873 } 874 875 if (!file) 876 break; 877 878 __skb_queue_tail(&list, skb); 879 880 skb = skb_dequeue(head); 881 } 882 883 if (skb_peek(&list)) { 884 spin_lock_irq(&head->lock); 885 while ((skb = __skb_dequeue(&list)) != NULL) 886 __skb_queue_tail(head, skb); 887 spin_unlock_irq(&head->lock); 888 } 889 #endif 890 } 891 892 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) 893 { 894 struct file *file = prsrc->file; 895 896 if (likely(!io_file_need_scm(file))) 897 fput(file); 898 else 899 io_rsrc_file_scm_put(ctx, file); 900 } 901 902 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, 903 unsigned nr_args, u64 __user *tags) 904 { 905 __s32 __user *fds = (__s32 __user *) arg; 906 struct file *file; 907 int fd, ret; 908 unsigned i; 909 910 if (ctx->file_data) 911 return -EBUSY; 912 if (!nr_args) 913 return -EINVAL; 914 if (nr_args > IORING_MAX_FIXED_FILES) 915 return -EMFILE; 916 if (nr_args > rlimit(RLIMIT_NOFILE)) 917 return -EMFILE; 918 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args, 919 &ctx->file_data); 920 if (ret) 921 return ret; 922 923 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) { 924 io_rsrc_data_free(ctx->file_data); 925 ctx->file_data = NULL; 926 return -ENOMEM; 927 } 928 929 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { 930 struct io_fixed_file *file_slot; 931 932 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) { 933 ret = -EFAULT; 934 goto fail; 935 } 936 /* allow sparse sets */ 937 if (!fds || fd == -1) { 938 ret = -EINVAL; 939 if (unlikely(*io_get_tag_slot(ctx->file_data, i))) 940 goto fail; 941 continue; 942 } 943 944 file = fget(fd); 945 ret = -EBADF; 946 if (unlikely(!file)) 947 goto fail; 948 949 /* 950 * Don't allow io_uring instances to be registered. If UNIX 951 * isn't enabled, then this causes a reference cycle and this 952 * instance can never get freed. If UNIX is enabled we'll 953 * handle it just fine, but there's still no point in allowing 954 * a ring fd as it doesn't support regular read/write anyway. 955 */ 956 if (io_is_uring_fops(file)) { 957 fput(file); 958 goto fail; 959 } 960 ret = io_scm_file_account(ctx, file); 961 if (ret) { 962 fput(file); 963 goto fail; 964 } 965 file_slot = io_fixed_file_slot(&ctx->file_table, i); 966 io_fixed_file_set(file_slot, file); 967 io_file_bitmap_set(&ctx->file_table, i); 968 } 969 970 /* default it to the whole table */ 971 io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files); 972 return 0; 973 fail: 974 __io_sqe_files_unregister(ctx); 975 return ret; 976 } 977 978 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) 979 { 980 io_buffer_unmap(ctx, &prsrc->buf); 981 prsrc->buf = NULL; 982 } 983 984 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx) 985 { 986 unsigned int i; 987 988 for (i = 0; i < ctx->nr_user_bufs; i++) 989 io_buffer_unmap(ctx, &ctx->user_bufs[i]); 990 kfree(ctx->user_bufs); 991 io_rsrc_data_free(ctx->buf_data); 992 ctx->user_bufs = NULL; 993 ctx->buf_data = NULL; 994 ctx->nr_user_bufs = 0; 995 } 996 997 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx) 998 { 999 unsigned nr = ctx->nr_user_bufs; 1000 int ret; 1001 1002 if (!ctx->buf_data) 1003 return -ENXIO; 1004 1005 /* 1006 * Quiesce may unlock ->uring_lock, and while it's not held 1007 * prevent new requests using the table. 1008 */ 1009 ctx->nr_user_bufs = 0; 1010 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx); 1011 ctx->nr_user_bufs = nr; 1012 if (!ret) 1013 __io_sqe_buffers_unregister(ctx); 1014 return ret; 1015 } 1016 1017 /* 1018 * Not super efficient, but this is just a registration time. And we do cache 1019 * the last compound head, so generally we'll only do a full search if we don't 1020 * match that one. 1021 * 1022 * We check if the given compound head page has already been accounted, to 1023 * avoid double accounting it. This allows us to account the full size of the 1024 * page, not just the constituent pages of a huge page. 1025 */ 1026 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages, 1027 int nr_pages, struct page *hpage) 1028 { 1029 int i, j; 1030 1031 /* check current page array */ 1032 for (i = 0; i < nr_pages; i++) { 1033 if (!PageCompound(pages[i])) 1034 continue; 1035 if (compound_head(pages[i]) == hpage) 1036 return true; 1037 } 1038 1039 /* check previously registered pages */ 1040 for (i = 0; i < ctx->nr_user_bufs; i++) { 1041 struct io_mapped_ubuf *imu = ctx->user_bufs[i]; 1042 1043 for (j = 0; j < imu->nr_bvecs; j++) { 1044 if (!PageCompound(imu->bvec[j].bv_page)) 1045 continue; 1046 if (compound_head(imu->bvec[j].bv_page) == hpage) 1047 return true; 1048 } 1049 } 1050 1051 return false; 1052 } 1053 1054 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages, 1055 int nr_pages, struct io_mapped_ubuf *imu, 1056 struct page **last_hpage) 1057 { 1058 int i, ret; 1059 1060 imu->acct_pages = 0; 1061 for (i = 0; i < nr_pages; i++) { 1062 if (!PageCompound(pages[i])) { 1063 imu->acct_pages++; 1064 } else { 1065 struct page *hpage; 1066 1067 hpage = compound_head(pages[i]); 1068 if (hpage == *last_hpage) 1069 continue; 1070 *last_hpage = hpage; 1071 if (headpage_already_acct(ctx, pages, i, hpage)) 1072 continue; 1073 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT; 1074 } 1075 } 1076 1077 if (!imu->acct_pages) 1078 return 0; 1079 1080 ret = io_account_mem(ctx, imu->acct_pages); 1081 if (ret) 1082 imu->acct_pages = 0; 1083 return ret; 1084 } 1085 1086 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages) 1087 { 1088 unsigned long start, end, nr_pages; 1089 struct vm_area_struct **vmas = NULL; 1090 struct page **pages = NULL; 1091 int i, pret, ret = -ENOMEM; 1092 1093 end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT; 1094 start = ubuf >> PAGE_SHIFT; 1095 nr_pages = end - start; 1096 1097 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); 1098 if (!pages) 1099 goto done; 1100 1101 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *), 1102 GFP_KERNEL); 1103 if (!vmas) 1104 goto done; 1105 1106 ret = 0; 1107 mmap_read_lock(current->mm); 1108 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM, 1109 pages, vmas); 1110 if (pret == nr_pages) { 1111 struct file *file = vmas[0]->vm_file; 1112 1113 /* don't support file backed memory */ 1114 for (i = 0; i < nr_pages; i++) { 1115 if (vmas[i]->vm_file != file) { 1116 ret = -EINVAL; 1117 break; 1118 } 1119 if (!file) 1120 continue; 1121 if (!vma_is_shmem(vmas[i]) && !is_file_hugepages(file)) { 1122 ret = -EOPNOTSUPP; 1123 break; 1124 } 1125 } 1126 *npages = nr_pages; 1127 } else { 1128 ret = pret < 0 ? pret : -EFAULT; 1129 } 1130 mmap_read_unlock(current->mm); 1131 if (ret) { 1132 /* 1133 * if we did partial map, or found file backed vmas, 1134 * release any pages we did get 1135 */ 1136 if (pret > 0) 1137 unpin_user_pages(pages, pret); 1138 goto done; 1139 } 1140 ret = 0; 1141 done: 1142 kvfree(vmas); 1143 if (ret < 0) { 1144 kvfree(pages); 1145 pages = ERR_PTR(ret); 1146 } 1147 return pages; 1148 } 1149 1150 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov, 1151 struct io_mapped_ubuf **pimu, 1152 struct page **last_hpage) 1153 { 1154 struct io_mapped_ubuf *imu = NULL; 1155 struct page **pages = NULL; 1156 unsigned long off; 1157 size_t size; 1158 int ret, nr_pages, i; 1159 struct folio *folio = NULL; 1160 1161 *pimu = ctx->dummy_ubuf; 1162 if (!iov->iov_base) 1163 return 0; 1164 1165 ret = -ENOMEM; 1166 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len, 1167 &nr_pages); 1168 if (IS_ERR(pages)) { 1169 ret = PTR_ERR(pages); 1170 pages = NULL; 1171 goto done; 1172 } 1173 1174 /* If it's a huge page, try to coalesce them into a single bvec entry */ 1175 if (nr_pages > 1) { 1176 folio = page_folio(pages[0]); 1177 for (i = 1; i < nr_pages; i++) { 1178 if (page_folio(pages[i]) != folio) { 1179 folio = NULL; 1180 break; 1181 } 1182 } 1183 if (folio) { 1184 /* 1185 * The pages are bound to the folio, it doesn't 1186 * actually unpin them but drops all but one reference, 1187 * which is usually put down by io_buffer_unmap(). 1188 * Note, needs a better helper. 1189 */ 1190 unpin_user_pages(&pages[1], nr_pages - 1); 1191 nr_pages = 1; 1192 } 1193 } 1194 1195 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL); 1196 if (!imu) 1197 goto done; 1198 1199 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage); 1200 if (ret) { 1201 unpin_user_pages(pages, nr_pages); 1202 goto done; 1203 } 1204 1205 off = (unsigned long) iov->iov_base & ~PAGE_MASK; 1206 size = iov->iov_len; 1207 /* store original address for later verification */ 1208 imu->ubuf = (unsigned long) iov->iov_base; 1209 imu->ubuf_end = imu->ubuf + iov->iov_len; 1210 imu->nr_bvecs = nr_pages; 1211 *pimu = imu; 1212 ret = 0; 1213 1214 if (folio) { 1215 bvec_set_page(&imu->bvec[0], pages[0], size, off); 1216 goto done; 1217 } 1218 for (i = 0; i < nr_pages; i++) { 1219 size_t vec_len; 1220 1221 vec_len = min_t(size_t, size, PAGE_SIZE - off); 1222 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off); 1223 off = 0; 1224 size -= vec_len; 1225 } 1226 done: 1227 if (ret) 1228 kvfree(imu); 1229 kvfree(pages); 1230 return ret; 1231 } 1232 1233 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args) 1234 { 1235 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL); 1236 return ctx->user_bufs ? 0 : -ENOMEM; 1237 } 1238 1239 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg, 1240 unsigned int nr_args, u64 __user *tags) 1241 { 1242 struct page *last_hpage = NULL; 1243 struct io_rsrc_data *data; 1244 int i, ret; 1245 struct iovec iov; 1246 1247 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16)); 1248 1249 if (ctx->user_bufs) 1250 return -EBUSY; 1251 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS) 1252 return -EINVAL; 1253 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data); 1254 if (ret) 1255 return ret; 1256 ret = io_buffers_map_alloc(ctx, nr_args); 1257 if (ret) { 1258 io_rsrc_data_free(data); 1259 return ret; 1260 } 1261 1262 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) { 1263 if (arg) { 1264 ret = io_copy_iov(ctx, &iov, arg, i); 1265 if (ret) 1266 break; 1267 ret = io_buffer_validate(&iov); 1268 if (ret) 1269 break; 1270 } else { 1271 memset(&iov, 0, sizeof(iov)); 1272 } 1273 1274 if (!iov.iov_base && *io_get_tag_slot(data, i)) { 1275 ret = -EINVAL; 1276 break; 1277 } 1278 1279 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i], 1280 &last_hpage); 1281 if (ret) 1282 break; 1283 } 1284 1285 WARN_ON_ONCE(ctx->buf_data); 1286 1287 ctx->buf_data = data; 1288 if (ret) 1289 __io_sqe_buffers_unregister(ctx); 1290 return ret; 1291 } 1292 1293 int io_import_fixed(int ddir, struct iov_iter *iter, 1294 struct io_mapped_ubuf *imu, 1295 u64 buf_addr, size_t len) 1296 { 1297 u64 buf_end; 1298 size_t offset; 1299 1300 if (WARN_ON_ONCE(!imu)) 1301 return -EFAULT; 1302 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end))) 1303 return -EFAULT; 1304 /* not inside the mapped region */ 1305 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end)) 1306 return -EFAULT; 1307 1308 /* 1309 * Might not be a start of buffer, set size appropriately 1310 * and advance us to the beginning. 1311 */ 1312 offset = buf_addr - imu->ubuf; 1313 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len); 1314 1315 if (offset) { 1316 /* 1317 * Don't use iov_iter_advance() here, as it's really slow for 1318 * using the latter parts of a big fixed buffer - it iterates 1319 * over each segment manually. We can cheat a bit here, because 1320 * we know that: 1321 * 1322 * 1) it's a BVEC iter, we set it up 1323 * 2) all bvecs are PAGE_SIZE in size, except potentially the 1324 * first and last bvec 1325 * 1326 * So just find our index, and adjust the iterator afterwards. 1327 * If the offset is within the first bvec (or the whole first 1328 * bvec, just use iov_iter_advance(). This makes it easier 1329 * since we can just skip the first segment, which may not 1330 * be PAGE_SIZE aligned. 1331 */ 1332 const struct bio_vec *bvec = imu->bvec; 1333 1334 if (offset <= bvec->bv_len) { 1335 /* 1336 * Note, huge pages buffers consists of one large 1337 * bvec entry and should always go this way. The other 1338 * branch doesn't expect non PAGE_SIZE'd chunks. 1339 */ 1340 iter->bvec = bvec; 1341 iter->nr_segs = bvec->bv_len; 1342 iter->count -= offset; 1343 iter->iov_offset = offset; 1344 } else { 1345 unsigned long seg_skip; 1346 1347 /* skip first vec */ 1348 offset -= bvec->bv_len; 1349 seg_skip = 1 + (offset >> PAGE_SHIFT); 1350 1351 iter->bvec = bvec + seg_skip; 1352 iter->nr_segs -= seg_skip; 1353 iter->count -= bvec->bv_len + offset; 1354 iter->iov_offset = offset & ~PAGE_MASK; 1355 } 1356 } 1357 1358 return 0; 1359 } 1360