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