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, ctx->rsrc_node, 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 int offset = up->offset + done; 473 u64 tag = 0; 474 475 err = io_copy_iov(ctx, &iov, iovs, done); 476 if (err) 477 break; 478 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) { 479 err = -EFAULT; 480 break; 481 } 482 err = io_buffer_validate(&iov); 483 if (err) 484 break; 485 if (!iov.iov_base && tag) { 486 err = -EINVAL; 487 break; 488 } 489 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage); 490 if (err) 491 break; 492 493 i = array_index_nospec(offset, ctx->nr_user_bufs); 494 if (ctx->user_bufs[i] != ctx->dummy_ubuf) { 495 err = io_queue_rsrc_removal(ctx->buf_data, i, 496 ctx->rsrc_node, ctx->user_bufs[i]); 497 if (unlikely(err)) { 498 io_buffer_unmap(ctx, &imu); 499 break; 500 } 501 ctx->user_bufs[i] = ctx->dummy_ubuf; 502 needs_switch = true; 503 } 504 505 ctx->user_bufs[i] = imu; 506 *io_get_tag_slot(ctx->buf_data, i) = tag; 507 } 508 509 if (needs_switch) 510 io_rsrc_node_switch(ctx, ctx->buf_data); 511 return done ? done : err; 512 } 513 514 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type, 515 struct io_uring_rsrc_update2 *up, 516 unsigned nr_args) 517 { 518 __u32 tmp; 519 int err; 520 521 lockdep_assert_held(&ctx->uring_lock); 522 523 if (check_add_overflow(up->offset, nr_args, &tmp)) 524 return -EOVERFLOW; 525 err = io_rsrc_node_switch_start(ctx); 526 if (err) 527 return err; 528 529 switch (type) { 530 case IORING_RSRC_FILE: 531 return __io_sqe_files_update(ctx, up, nr_args); 532 case IORING_RSRC_BUFFER: 533 return __io_sqe_buffers_update(ctx, up, nr_args); 534 } 535 return -EINVAL; 536 } 537 538 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg, 539 unsigned nr_args) 540 { 541 struct io_uring_rsrc_update2 up; 542 543 if (!nr_args) 544 return -EINVAL; 545 memset(&up, 0, sizeof(up)); 546 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update))) 547 return -EFAULT; 548 if (up.resv || up.resv2) 549 return -EINVAL; 550 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args); 551 } 552 553 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg, 554 unsigned size, unsigned type) 555 { 556 struct io_uring_rsrc_update2 up; 557 558 if (size != sizeof(up)) 559 return -EINVAL; 560 if (copy_from_user(&up, arg, sizeof(up))) 561 return -EFAULT; 562 if (!up.nr || up.resv || up.resv2) 563 return -EINVAL; 564 return __io_register_rsrc_update(ctx, type, &up, up.nr); 565 } 566 567 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg, 568 unsigned int size, unsigned int type) 569 { 570 struct io_uring_rsrc_register rr; 571 572 /* keep it extendible */ 573 if (size != sizeof(rr)) 574 return -EINVAL; 575 576 memset(&rr, 0, sizeof(rr)); 577 if (copy_from_user(&rr, arg, size)) 578 return -EFAULT; 579 if (!rr.nr || rr.resv2) 580 return -EINVAL; 581 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE) 582 return -EINVAL; 583 584 switch (type) { 585 case IORING_RSRC_FILE: 586 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data) 587 break; 588 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data), 589 rr.nr, u64_to_user_ptr(rr.tags)); 590 case IORING_RSRC_BUFFER: 591 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data) 592 break; 593 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data), 594 rr.nr, u64_to_user_ptr(rr.tags)); 595 } 596 return -EINVAL; 597 } 598 599 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) 600 { 601 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 602 603 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) 604 return -EINVAL; 605 if (sqe->rw_flags || sqe->splice_fd_in) 606 return -EINVAL; 607 608 up->offset = READ_ONCE(sqe->off); 609 up->nr_args = READ_ONCE(sqe->len); 610 if (!up->nr_args) 611 return -EINVAL; 612 up->arg = READ_ONCE(sqe->addr); 613 return 0; 614 } 615 616 static int io_files_update_with_index_alloc(struct io_kiocb *req, 617 unsigned int issue_flags) 618 { 619 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 620 __s32 __user *fds = u64_to_user_ptr(up->arg); 621 unsigned int done; 622 struct file *file; 623 int ret, fd; 624 625 if (!req->ctx->file_data) 626 return -ENXIO; 627 628 for (done = 0; done < up->nr_args; done++) { 629 if (copy_from_user(&fd, &fds[done], sizeof(fd))) { 630 ret = -EFAULT; 631 break; 632 } 633 634 file = fget(fd); 635 if (!file) { 636 ret = -EBADF; 637 break; 638 } 639 ret = io_fixed_fd_install(req, issue_flags, file, 640 IORING_FILE_INDEX_ALLOC); 641 if (ret < 0) 642 break; 643 if (copy_to_user(&fds[done], &ret, sizeof(ret))) { 644 __io_close_fixed(req->ctx, issue_flags, ret); 645 ret = -EFAULT; 646 break; 647 } 648 } 649 650 if (done) 651 return done; 652 return ret; 653 } 654 655 int io_files_update(struct io_kiocb *req, unsigned int issue_flags) 656 { 657 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 658 struct io_ring_ctx *ctx = req->ctx; 659 struct io_uring_rsrc_update2 up2; 660 int ret; 661 662 up2.offset = up->offset; 663 up2.data = up->arg; 664 up2.nr = 0; 665 up2.tags = 0; 666 up2.resv = 0; 667 up2.resv2 = 0; 668 669 if (up->offset == IORING_FILE_INDEX_ALLOC) { 670 ret = io_files_update_with_index_alloc(req, issue_flags); 671 } else { 672 io_ring_submit_lock(ctx, issue_flags); 673 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE, 674 &up2, up->nr_args); 675 io_ring_submit_unlock(ctx, issue_flags); 676 } 677 678 if (ret < 0) 679 req_set_fail(req); 680 io_req_set_res(req, ret, 0); 681 return IOU_OK; 682 } 683 684 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx, 685 struct io_rsrc_node *node, void *rsrc) 686 { 687 u64 *tag_slot = io_get_tag_slot(data, idx); 688 struct io_rsrc_put *prsrc; 689 bool inline_item = true; 690 691 if (!node->inline_items) { 692 prsrc = &node->item; 693 node->inline_items++; 694 } else { 695 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL); 696 if (!prsrc) 697 return -ENOMEM; 698 inline_item = false; 699 } 700 701 prsrc->tag = *tag_slot; 702 *tag_slot = 0; 703 prsrc->rsrc = rsrc; 704 if (!inline_item) 705 list_add(&prsrc->list, &node->item_list); 706 return 0; 707 } 708 709 void __io_sqe_files_unregister(struct io_ring_ctx *ctx) 710 { 711 int i; 712 713 for (i = 0; i < ctx->nr_user_files; i++) { 714 struct file *file = io_file_from_index(&ctx->file_table, i); 715 716 /* skip scm accounted files, they'll be freed by ->ring_sock */ 717 if (!file || io_file_need_scm(file)) 718 continue; 719 io_file_bitmap_clear(&ctx->file_table, i); 720 fput(file); 721 } 722 723 #if defined(CONFIG_UNIX) 724 if (ctx->ring_sock) { 725 struct sock *sock = ctx->ring_sock->sk; 726 struct sk_buff *skb; 727 728 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) 729 kfree_skb(skb); 730 } 731 #endif 732 io_free_file_tables(&ctx->file_table); 733 io_file_table_set_alloc_range(ctx, 0, 0); 734 io_rsrc_data_free(ctx->file_data); 735 ctx->file_data = NULL; 736 ctx->nr_user_files = 0; 737 } 738 739 int io_sqe_files_unregister(struct io_ring_ctx *ctx) 740 { 741 unsigned nr = ctx->nr_user_files; 742 int ret; 743 744 if (!ctx->file_data) 745 return -ENXIO; 746 747 /* 748 * Quiesce may unlock ->uring_lock, and while it's not held 749 * prevent new requests using the table. 750 */ 751 ctx->nr_user_files = 0; 752 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx); 753 ctx->nr_user_files = nr; 754 if (!ret) 755 __io_sqe_files_unregister(ctx); 756 return ret; 757 } 758 759 /* 760 * Ensure the UNIX gc is aware of our file set, so we are certain that 761 * the io_uring can be safely unregistered on process exit, even if we have 762 * loops in the file referencing. We account only files that can hold other 763 * files because otherwise they can't form a loop and so are not interesting 764 * for GC. 765 */ 766 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file) 767 { 768 #if defined(CONFIG_UNIX) 769 struct sock *sk = ctx->ring_sock->sk; 770 struct sk_buff_head *head = &sk->sk_receive_queue; 771 struct scm_fp_list *fpl; 772 struct sk_buff *skb; 773 774 if (likely(!io_file_need_scm(file))) 775 return 0; 776 777 /* 778 * See if we can merge this file into an existing skb SCM_RIGHTS 779 * file set. If there's no room, fall back to allocating a new skb 780 * and filling it in. 781 */ 782 spin_lock_irq(&head->lock); 783 skb = skb_peek(head); 784 if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD) 785 __skb_unlink(skb, head); 786 else 787 skb = NULL; 788 spin_unlock_irq(&head->lock); 789 790 if (!skb) { 791 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); 792 if (!fpl) 793 return -ENOMEM; 794 795 skb = alloc_skb(0, GFP_KERNEL); 796 if (!skb) { 797 kfree(fpl); 798 return -ENOMEM; 799 } 800 801 fpl->user = get_uid(current_user()); 802 fpl->max = SCM_MAX_FD; 803 fpl->count = 0; 804 805 UNIXCB(skb).fp = fpl; 806 skb->sk = sk; 807 skb->scm_io_uring = 1; 808 skb->destructor = unix_destruct_scm; 809 refcount_add(skb->truesize, &sk->sk_wmem_alloc); 810 } 811 812 fpl = UNIXCB(skb).fp; 813 fpl->fp[fpl->count++] = get_file(file); 814 unix_inflight(fpl->user, file); 815 skb_queue_head(head, skb); 816 fput(file); 817 #endif 818 return 0; 819 } 820 821 static __cold void io_rsrc_file_scm_put(struct io_ring_ctx *ctx, struct file *file) 822 { 823 #if defined(CONFIG_UNIX) 824 struct sock *sock = ctx->ring_sock->sk; 825 struct sk_buff_head list, *head = &sock->sk_receive_queue; 826 struct sk_buff *skb; 827 int i; 828 829 __skb_queue_head_init(&list); 830 831 /* 832 * Find the skb that holds this file in its SCM_RIGHTS. When found, 833 * remove this entry and rearrange the file array. 834 */ 835 skb = skb_dequeue(head); 836 while (skb) { 837 struct scm_fp_list *fp; 838 839 fp = UNIXCB(skb).fp; 840 for (i = 0; i < fp->count; i++) { 841 int left; 842 843 if (fp->fp[i] != file) 844 continue; 845 846 unix_notinflight(fp->user, fp->fp[i]); 847 left = fp->count - 1 - i; 848 if (left) { 849 memmove(&fp->fp[i], &fp->fp[i + 1], 850 left * sizeof(struct file *)); 851 } 852 fp->count--; 853 if (!fp->count) { 854 kfree_skb(skb); 855 skb = NULL; 856 } else { 857 __skb_queue_tail(&list, skb); 858 } 859 fput(file); 860 file = NULL; 861 break; 862 } 863 864 if (!file) 865 break; 866 867 __skb_queue_tail(&list, skb); 868 869 skb = skb_dequeue(head); 870 } 871 872 if (skb_peek(&list)) { 873 spin_lock_irq(&head->lock); 874 while ((skb = __skb_dequeue(&list)) != NULL) 875 __skb_queue_tail(head, skb); 876 spin_unlock_irq(&head->lock); 877 } 878 #endif 879 } 880 881 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) 882 { 883 struct file *file = prsrc->file; 884 885 if (likely(!io_file_need_scm(file))) 886 fput(file); 887 else 888 io_rsrc_file_scm_put(ctx, file); 889 } 890 891 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, 892 unsigned nr_args, u64 __user *tags) 893 { 894 __s32 __user *fds = (__s32 __user *) arg; 895 struct file *file; 896 int fd, ret; 897 unsigned i; 898 899 if (ctx->file_data) 900 return -EBUSY; 901 if (!nr_args) 902 return -EINVAL; 903 if (nr_args > IORING_MAX_FIXED_FILES) 904 return -EMFILE; 905 if (nr_args > rlimit(RLIMIT_NOFILE)) 906 return -EMFILE; 907 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args, 908 &ctx->file_data); 909 if (ret) 910 return ret; 911 912 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) { 913 io_rsrc_data_free(ctx->file_data); 914 ctx->file_data = NULL; 915 return -ENOMEM; 916 } 917 918 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { 919 struct io_fixed_file *file_slot; 920 921 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) { 922 ret = -EFAULT; 923 goto fail; 924 } 925 /* allow sparse sets */ 926 if (!fds || fd == -1) { 927 ret = -EINVAL; 928 if (unlikely(*io_get_tag_slot(ctx->file_data, i))) 929 goto fail; 930 continue; 931 } 932 933 file = fget(fd); 934 ret = -EBADF; 935 if (unlikely(!file)) 936 goto fail; 937 938 /* 939 * Don't allow io_uring instances to be registered. If UNIX 940 * isn't enabled, then this causes a reference cycle and this 941 * instance can never get freed. If UNIX is enabled we'll 942 * handle it just fine, but there's still no point in allowing 943 * a ring fd as it doesn't support regular read/write anyway. 944 */ 945 if (io_is_uring_fops(file)) { 946 fput(file); 947 goto fail; 948 } 949 ret = io_scm_file_account(ctx, file); 950 if (ret) { 951 fput(file); 952 goto fail; 953 } 954 file_slot = io_fixed_file_slot(&ctx->file_table, i); 955 io_fixed_file_set(file_slot, file); 956 io_file_bitmap_set(&ctx->file_table, i); 957 } 958 959 /* default it to the whole table */ 960 io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files); 961 return 0; 962 fail: 963 __io_sqe_files_unregister(ctx); 964 return ret; 965 } 966 967 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) 968 { 969 io_buffer_unmap(ctx, &prsrc->buf); 970 prsrc->buf = NULL; 971 } 972 973 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx) 974 { 975 unsigned int i; 976 977 for (i = 0; i < ctx->nr_user_bufs; i++) 978 io_buffer_unmap(ctx, &ctx->user_bufs[i]); 979 kfree(ctx->user_bufs); 980 io_rsrc_data_free(ctx->buf_data); 981 ctx->user_bufs = NULL; 982 ctx->buf_data = NULL; 983 ctx->nr_user_bufs = 0; 984 } 985 986 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx) 987 { 988 unsigned nr = ctx->nr_user_bufs; 989 int ret; 990 991 if (!ctx->buf_data) 992 return -ENXIO; 993 994 /* 995 * Quiesce may unlock ->uring_lock, and while it's not held 996 * prevent new requests using the table. 997 */ 998 ctx->nr_user_bufs = 0; 999 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx); 1000 ctx->nr_user_bufs = nr; 1001 if (!ret) 1002 __io_sqe_buffers_unregister(ctx); 1003 return ret; 1004 } 1005 1006 /* 1007 * Not super efficient, but this is just a registration time. And we do cache 1008 * the last compound head, so generally we'll only do a full search if we don't 1009 * match that one. 1010 * 1011 * We check if the given compound head page has already been accounted, to 1012 * avoid double accounting it. This allows us to account the full size of the 1013 * page, not just the constituent pages of a huge page. 1014 */ 1015 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages, 1016 int nr_pages, struct page *hpage) 1017 { 1018 int i, j; 1019 1020 /* check current page array */ 1021 for (i = 0; i < nr_pages; i++) { 1022 if (!PageCompound(pages[i])) 1023 continue; 1024 if (compound_head(pages[i]) == hpage) 1025 return true; 1026 } 1027 1028 /* check previously registered pages */ 1029 for (i = 0; i < ctx->nr_user_bufs; i++) { 1030 struct io_mapped_ubuf *imu = ctx->user_bufs[i]; 1031 1032 for (j = 0; j < imu->nr_bvecs; j++) { 1033 if (!PageCompound(imu->bvec[j].bv_page)) 1034 continue; 1035 if (compound_head(imu->bvec[j].bv_page) == hpage) 1036 return true; 1037 } 1038 } 1039 1040 return false; 1041 } 1042 1043 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages, 1044 int nr_pages, struct io_mapped_ubuf *imu, 1045 struct page **last_hpage) 1046 { 1047 int i, ret; 1048 1049 imu->acct_pages = 0; 1050 for (i = 0; i < nr_pages; i++) { 1051 if (!PageCompound(pages[i])) { 1052 imu->acct_pages++; 1053 } else { 1054 struct page *hpage; 1055 1056 hpage = compound_head(pages[i]); 1057 if (hpage == *last_hpage) 1058 continue; 1059 *last_hpage = hpage; 1060 if (headpage_already_acct(ctx, pages, i, hpage)) 1061 continue; 1062 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT; 1063 } 1064 } 1065 1066 if (!imu->acct_pages) 1067 return 0; 1068 1069 ret = io_account_mem(ctx, imu->acct_pages); 1070 if (ret) 1071 imu->acct_pages = 0; 1072 return ret; 1073 } 1074 1075 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages) 1076 { 1077 unsigned long start, end, nr_pages; 1078 struct vm_area_struct **vmas = NULL; 1079 struct page **pages = NULL; 1080 int i, pret, ret = -ENOMEM; 1081 1082 end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT; 1083 start = ubuf >> PAGE_SHIFT; 1084 nr_pages = end - start; 1085 1086 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); 1087 if (!pages) 1088 goto done; 1089 1090 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *), 1091 GFP_KERNEL); 1092 if (!vmas) 1093 goto done; 1094 1095 ret = 0; 1096 mmap_read_lock(current->mm); 1097 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM, 1098 pages, vmas); 1099 if (pret == nr_pages) { 1100 struct file *file = vmas[0]->vm_file; 1101 1102 /* don't support file backed memory */ 1103 for (i = 0; i < nr_pages; i++) { 1104 if (vmas[i]->vm_file != file) { 1105 ret = -EINVAL; 1106 break; 1107 } 1108 if (!file) 1109 continue; 1110 if (!vma_is_shmem(vmas[i]) && !is_file_hugepages(file)) { 1111 ret = -EOPNOTSUPP; 1112 break; 1113 } 1114 } 1115 *npages = nr_pages; 1116 } else { 1117 ret = pret < 0 ? pret : -EFAULT; 1118 } 1119 mmap_read_unlock(current->mm); 1120 if (ret) { 1121 /* 1122 * if we did partial map, or found file backed vmas, 1123 * release any pages we did get 1124 */ 1125 if (pret > 0) 1126 unpin_user_pages(pages, pret); 1127 goto done; 1128 } 1129 ret = 0; 1130 done: 1131 kvfree(vmas); 1132 if (ret < 0) { 1133 kvfree(pages); 1134 pages = ERR_PTR(ret); 1135 } 1136 return pages; 1137 } 1138 1139 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov, 1140 struct io_mapped_ubuf **pimu, 1141 struct page **last_hpage) 1142 { 1143 struct io_mapped_ubuf *imu = NULL; 1144 struct page **pages = NULL; 1145 unsigned long off; 1146 size_t size; 1147 int ret, nr_pages, i; 1148 struct folio *folio = NULL; 1149 1150 *pimu = ctx->dummy_ubuf; 1151 if (!iov->iov_base) 1152 return 0; 1153 1154 ret = -ENOMEM; 1155 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len, 1156 &nr_pages); 1157 if (IS_ERR(pages)) { 1158 ret = PTR_ERR(pages); 1159 pages = NULL; 1160 goto done; 1161 } 1162 1163 /* If it's a huge page, try to coalesce them into a single bvec entry */ 1164 if (nr_pages > 1) { 1165 folio = page_folio(pages[0]); 1166 for (i = 1; i < nr_pages; i++) { 1167 if (page_folio(pages[i]) != folio) { 1168 folio = NULL; 1169 break; 1170 } 1171 } 1172 if (folio) { 1173 /* 1174 * The pages are bound to the folio, it doesn't 1175 * actually unpin them but drops all but one reference, 1176 * which is usually put down by io_buffer_unmap(). 1177 * Note, needs a better helper. 1178 */ 1179 unpin_user_pages(&pages[1], nr_pages - 1); 1180 nr_pages = 1; 1181 } 1182 } 1183 1184 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL); 1185 if (!imu) 1186 goto done; 1187 1188 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage); 1189 if (ret) { 1190 unpin_user_pages(pages, nr_pages); 1191 goto done; 1192 } 1193 1194 off = (unsigned long) iov->iov_base & ~PAGE_MASK; 1195 size = iov->iov_len; 1196 /* store original address for later verification */ 1197 imu->ubuf = (unsigned long) iov->iov_base; 1198 imu->ubuf_end = imu->ubuf + iov->iov_len; 1199 imu->nr_bvecs = nr_pages; 1200 *pimu = imu; 1201 ret = 0; 1202 1203 if (folio) { 1204 bvec_set_page(&imu->bvec[0], pages[0], size, off); 1205 goto done; 1206 } 1207 for (i = 0; i < nr_pages; i++) { 1208 size_t vec_len; 1209 1210 vec_len = min_t(size_t, size, PAGE_SIZE - off); 1211 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off); 1212 off = 0; 1213 size -= vec_len; 1214 } 1215 done: 1216 if (ret) 1217 kvfree(imu); 1218 kvfree(pages); 1219 return ret; 1220 } 1221 1222 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args) 1223 { 1224 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL); 1225 return ctx->user_bufs ? 0 : -ENOMEM; 1226 } 1227 1228 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg, 1229 unsigned int nr_args, u64 __user *tags) 1230 { 1231 struct page *last_hpage = NULL; 1232 struct io_rsrc_data *data; 1233 int i, ret; 1234 struct iovec iov; 1235 1236 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16)); 1237 1238 if (ctx->user_bufs) 1239 return -EBUSY; 1240 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS) 1241 return -EINVAL; 1242 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data); 1243 if (ret) 1244 return ret; 1245 ret = io_buffers_map_alloc(ctx, nr_args); 1246 if (ret) { 1247 io_rsrc_data_free(data); 1248 return ret; 1249 } 1250 1251 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) { 1252 if (arg) { 1253 ret = io_copy_iov(ctx, &iov, arg, i); 1254 if (ret) 1255 break; 1256 ret = io_buffer_validate(&iov); 1257 if (ret) 1258 break; 1259 } else { 1260 memset(&iov, 0, sizeof(iov)); 1261 } 1262 1263 if (!iov.iov_base && *io_get_tag_slot(data, i)) { 1264 ret = -EINVAL; 1265 break; 1266 } 1267 1268 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i], 1269 &last_hpage); 1270 if (ret) 1271 break; 1272 } 1273 1274 WARN_ON_ONCE(ctx->buf_data); 1275 1276 ctx->buf_data = data; 1277 if (ret) 1278 __io_sqe_buffers_unregister(ctx); 1279 return ret; 1280 } 1281 1282 int io_import_fixed(int ddir, struct iov_iter *iter, 1283 struct io_mapped_ubuf *imu, 1284 u64 buf_addr, size_t len) 1285 { 1286 u64 buf_end; 1287 size_t offset; 1288 1289 if (WARN_ON_ONCE(!imu)) 1290 return -EFAULT; 1291 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end))) 1292 return -EFAULT; 1293 /* not inside the mapped region */ 1294 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end)) 1295 return -EFAULT; 1296 1297 /* 1298 * Might not be a start of buffer, set size appropriately 1299 * and advance us to the beginning. 1300 */ 1301 offset = buf_addr - imu->ubuf; 1302 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len); 1303 1304 if (offset) { 1305 /* 1306 * Don't use iov_iter_advance() here, as it's really slow for 1307 * using the latter parts of a big fixed buffer - it iterates 1308 * over each segment manually. We can cheat a bit here, because 1309 * we know that: 1310 * 1311 * 1) it's a BVEC iter, we set it up 1312 * 2) all bvecs are PAGE_SIZE in size, except potentially the 1313 * first and last bvec 1314 * 1315 * So just find our index, and adjust the iterator afterwards. 1316 * If the offset is within the first bvec (or the whole first 1317 * bvec, just use iov_iter_advance(). This makes it easier 1318 * since we can just skip the first segment, which may not 1319 * be PAGE_SIZE aligned. 1320 */ 1321 const struct bio_vec *bvec = imu->bvec; 1322 1323 if (offset <= bvec->bv_len) { 1324 /* 1325 * Note, huge pages buffers consists of one large 1326 * bvec entry and should always go this way. The other 1327 * branch doesn't expect non PAGE_SIZE'd chunks. 1328 */ 1329 iter->bvec = bvec; 1330 iter->nr_segs = bvec->bv_len; 1331 iter->count -= offset; 1332 iter->iov_offset = offset; 1333 } else { 1334 unsigned long seg_skip; 1335 1336 /* skip first vec */ 1337 offset -= bvec->bv_len; 1338 seg_skip = 1 + (offset >> PAGE_SHIFT); 1339 1340 iter->bvec = bvec + seg_skip; 1341 iter->nr_segs -= seg_skip; 1342 iter->count -= bvec->bv_len + offset; 1343 iter->iov_offset = offset & ~PAGE_MASK; 1344 } 1345 } 1346 1347 return 0; 1348 } 1349