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