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/blk-mq.h> 7#include <linux/mm.h> 8#include <linux/slab.h> 9#include <linux/fsnotify.h> 10#include <linux/poll.h> 11#include <linux/nospec.h> 12#include <linux/compat.h> 13#include <linux/io_uring/cmd.h> 14#include <linux/indirect_call_wrapper.h> 15 16#include <uapi/linux/io_uring.h> 17 18#include "io_uring.h" 19#include "opdef.h" 20#include "kbuf.h" 21#include "alloc_cache.h" 22#include "rsrc.h" 23#include "poll.h" 24#include "rw.h" 25 26struct io_rw { 27 /* NOTE: kiocb has the file as the first member, so don't do it here */ 28 struct kiocb kiocb; 29 u64 addr; 30 u32 len; 31 rwf_t flags; 32}; 33 34static inline bool io_file_supports_nowait(struct io_kiocb *req) 35{ 36 return req->flags & REQ_F_SUPPORT_NOWAIT; 37} 38 39#ifdef CONFIG_COMPAT 40static int io_iov_compat_buffer_select_prep(struct io_rw *rw) 41{ 42 struct compat_iovec __user *uiov; 43 compat_ssize_t clen; 44 45 uiov = u64_to_user_ptr(rw->addr); 46 if (!access_ok(uiov, sizeof(*uiov))) 47 return -EFAULT; 48 if (__get_user(clen, &uiov->iov_len)) 49 return -EFAULT; 50 if (clen < 0) 51 return -EINVAL; 52 53 rw->len = clen; 54 return 0; 55} 56#endif 57 58static int io_iov_buffer_select_prep(struct io_kiocb *req) 59{ 60 struct iovec __user *uiov; 61 struct iovec iov; 62 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 63 64 if (rw->len != 1) 65 return -EINVAL; 66 67#ifdef CONFIG_COMPAT 68 if (req->ctx->compat) 69 return io_iov_compat_buffer_select_prep(rw); 70#endif 71 72 uiov = u64_to_user_ptr(rw->addr); 73 if (copy_from_user(&iov, uiov, sizeof(*uiov))) 74 return -EFAULT; 75 rw->len = iov.iov_len; 76 return 0; 77} 78 79static int __io_import_iovec(int ddir, struct io_kiocb *req, 80 struct io_async_rw *io, 81 unsigned int issue_flags) 82{ 83 const struct io_issue_def *def = &io_issue_defs[req->opcode]; 84 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 85 struct iovec *iov; 86 void __user *buf; 87 int nr_segs, ret; 88 size_t sqe_len; 89 90 buf = u64_to_user_ptr(rw->addr); 91 sqe_len = rw->len; 92 93 if (!def->vectored || req->flags & REQ_F_BUFFER_SELECT) { 94 if (io_do_buffer_select(req)) { 95 buf = io_buffer_select(req, &sqe_len, issue_flags); 96 if (!buf) 97 return -ENOBUFS; 98 rw->addr = (unsigned long) buf; 99 rw->len = sqe_len; 100 } 101 102 return import_ubuf(ddir, buf, sqe_len, &io->iter); 103 } 104 105 if (io->free_iovec) { 106 nr_segs = io->free_iov_nr; 107 iov = io->free_iovec; 108 } else { 109 iov = &io->fast_iov; 110 nr_segs = 1; 111 } 112 ret = __import_iovec(ddir, buf, sqe_len, nr_segs, &iov, &io->iter, 113 req->ctx->compat); 114 if (unlikely(ret < 0)) 115 return ret; 116 if (iov) { 117 req->flags |= REQ_F_NEED_CLEANUP; 118 io->free_iov_nr = io->iter.nr_segs; 119 kfree(io->free_iovec); 120 io->free_iovec = iov; 121 } 122 return 0; 123} 124 125static inline int io_import_iovec(int rw, struct io_kiocb *req, 126 struct io_async_rw *io, 127 unsigned int issue_flags) 128{ 129 int ret; 130 131 ret = __io_import_iovec(rw, req, io, issue_flags); 132 if (unlikely(ret < 0)) 133 return ret; 134 135 iov_iter_save_state(&io->iter, &io->iter_state); 136 return 0; 137} 138 139static void io_rw_iovec_free(struct io_async_rw *rw) 140{ 141 if (rw->free_iovec) { 142 kfree(rw->free_iovec); 143 rw->free_iov_nr = 0; 144 rw->free_iovec = NULL; 145 } 146} 147 148static void io_rw_recycle(struct io_kiocb *req, unsigned int issue_flags) 149{ 150 struct io_async_rw *rw = req->async_data; 151 struct iovec *iov; 152 153 if (unlikely(issue_flags & IO_URING_F_UNLOCKED)) { 154 io_rw_iovec_free(rw); 155 return; 156 } 157 iov = rw->free_iovec; 158 if (io_alloc_cache_put(&req->ctx->rw_cache, rw)) { 159 if (iov) 160 kasan_mempool_poison_object(iov); 161 req->async_data = NULL; 162 req->flags &= ~REQ_F_ASYNC_DATA; 163 } 164} 165 166static void io_req_rw_cleanup(struct io_kiocb *req, unsigned int issue_flags) 167{ 168 /* 169 * Disable quick recycling for anything that's gone through io-wq. 170 * In theory, this should be fine to cleanup. However, some read or 171 * write iter handling touches the iovec AFTER having called into the 172 * handler, eg to reexpand or revert. This means we can have: 173 * 174 * task io-wq 175 * issue 176 * punt to io-wq 177 * issue 178 * blkdev_write_iter() 179 * ->ki_complete() 180 * io_complete_rw() 181 * queue tw complete 182 * run tw 183 * req_rw_cleanup 184 * iov_iter_count() <- look at iov_iter again 185 * 186 * which can lead to a UAF. This is only possible for io-wq offload 187 * as the cleanup can run in parallel. As io-wq is not the fast path, 188 * just leave cleanup to the end. 189 * 190 * This is really a bug in the core code that does this, any issue 191 * path should assume that a successful (or -EIOCBQUEUED) return can 192 * mean that the underlying data can be gone at any time. But that 193 * should be fixed seperately, and then this check could be killed. 194 */ 195 if (!(req->flags & REQ_F_REFCOUNT)) { 196 req->flags &= ~REQ_F_NEED_CLEANUP; 197 io_rw_recycle(req, issue_flags); 198 } 199} 200 201static int io_rw_alloc_async(struct io_kiocb *req) 202{ 203 struct io_ring_ctx *ctx = req->ctx; 204 struct io_async_rw *rw; 205 206 rw = io_alloc_cache_get(&ctx->rw_cache); 207 if (rw) { 208 if (rw->free_iovec) { 209 kasan_mempool_unpoison_object(rw->free_iovec, 210 rw->free_iov_nr * sizeof(struct iovec)); 211 req->flags |= REQ_F_NEED_CLEANUP; 212 } 213 req->flags |= REQ_F_ASYNC_DATA; 214 req->async_data = rw; 215 goto done; 216 } 217 218 if (!io_alloc_async_data(req)) { 219 rw = req->async_data; 220 rw->free_iovec = NULL; 221 rw->free_iov_nr = 0; 222done: 223 rw->bytes_done = 0; 224 return 0; 225 } 226 227 return -ENOMEM; 228} 229 230static int io_prep_rw_setup(struct io_kiocb *req, int ddir, bool do_import) 231{ 232 struct io_async_rw *rw; 233 int ret; 234 235 if (io_rw_alloc_async(req)) 236 return -ENOMEM; 237 238 if (!do_import || io_do_buffer_select(req)) 239 return 0; 240 241 rw = req->async_data; 242 ret = io_import_iovec(ddir, req, rw, 0); 243 if (unlikely(ret < 0)) 244 return ret; 245 246 iov_iter_save_state(&rw->iter, &rw->iter_state); 247 return 0; 248} 249 250static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe, 251 int ddir, bool do_import) 252{ 253 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 254 unsigned ioprio; 255 int ret; 256 257 rw->kiocb.ki_pos = READ_ONCE(sqe->off); 258 /* used for fixed read/write too - just read unconditionally */ 259 req->buf_index = READ_ONCE(sqe->buf_index); 260 261 ioprio = READ_ONCE(sqe->ioprio); 262 if (ioprio) { 263 ret = ioprio_check_cap(ioprio); 264 if (ret) 265 return ret; 266 267 rw->kiocb.ki_ioprio = ioprio; 268 } else { 269 rw->kiocb.ki_ioprio = get_current_ioprio(); 270 } 271 rw->kiocb.dio_complete = NULL; 272 273 rw->addr = READ_ONCE(sqe->addr); 274 rw->len = READ_ONCE(sqe->len); 275 rw->flags = READ_ONCE(sqe->rw_flags); 276 return io_prep_rw_setup(req, ddir, do_import); 277} 278 279int io_prep_read(struct io_kiocb *req, const struct io_uring_sqe *sqe) 280{ 281 return io_prep_rw(req, sqe, ITER_DEST, true); 282} 283 284int io_prep_write(struct io_kiocb *req, const struct io_uring_sqe *sqe) 285{ 286 return io_prep_rw(req, sqe, ITER_SOURCE, true); 287} 288 289static int io_prep_rwv(struct io_kiocb *req, const struct io_uring_sqe *sqe, 290 int ddir) 291{ 292 const bool do_import = !(req->flags & REQ_F_BUFFER_SELECT); 293 int ret; 294 295 ret = io_prep_rw(req, sqe, ddir, do_import); 296 if (unlikely(ret)) 297 return ret; 298 if (do_import) 299 return 0; 300 301 /* 302 * Have to do this validation here, as this is in io_read() rw->len 303 * might have chanaged due to buffer selection 304 */ 305 return io_iov_buffer_select_prep(req); 306} 307 308int io_prep_readv(struct io_kiocb *req, const struct io_uring_sqe *sqe) 309{ 310 return io_prep_rwv(req, sqe, ITER_DEST); 311} 312 313int io_prep_writev(struct io_kiocb *req, const struct io_uring_sqe *sqe) 314{ 315 return io_prep_rwv(req, sqe, ITER_SOURCE); 316} 317 318static int io_prep_rw_fixed(struct io_kiocb *req, const struct io_uring_sqe *sqe, 319 int ddir) 320{ 321 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 322 struct io_ring_ctx *ctx = req->ctx; 323 struct io_async_rw *io; 324 u16 index; 325 int ret; 326 327 ret = io_prep_rw(req, sqe, ddir, false); 328 if (unlikely(ret)) 329 return ret; 330 331 if (unlikely(req->buf_index >= ctx->nr_user_bufs)) 332 return -EFAULT; 333 index = array_index_nospec(req->buf_index, ctx->nr_user_bufs); 334 req->imu = ctx->user_bufs[index]; 335 io_req_set_rsrc_node(req, ctx, 0); 336 337 io = req->async_data; 338 ret = io_import_fixed(ddir, &io->iter, req->imu, rw->addr, rw->len); 339 iov_iter_save_state(&io->iter, &io->iter_state); 340 return ret; 341} 342 343int io_prep_read_fixed(struct io_kiocb *req, const struct io_uring_sqe *sqe) 344{ 345 return io_prep_rw_fixed(req, sqe, ITER_DEST); 346} 347 348int io_prep_write_fixed(struct io_kiocb *req, const struct io_uring_sqe *sqe) 349{ 350 return io_prep_rw_fixed(req, sqe, ITER_SOURCE); 351} 352 353/* 354 * Multishot read is prepared just like a normal read/write request, only 355 * difference is that we set the MULTISHOT flag. 356 */ 357int io_read_mshot_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) 358{ 359 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 360 int ret; 361 362 /* must be used with provided buffers */ 363 if (!(req->flags & REQ_F_BUFFER_SELECT)) 364 return -EINVAL; 365 366 ret = io_prep_rw(req, sqe, ITER_DEST, false); 367 if (unlikely(ret)) 368 return ret; 369 370 if (rw->addr || rw->len) 371 return -EINVAL; 372 373 req->flags |= REQ_F_APOLL_MULTISHOT; 374 return 0; 375} 376 377void io_readv_writev_cleanup(struct io_kiocb *req) 378{ 379 io_rw_iovec_free(req->async_data); 380} 381 382static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req) 383{ 384 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 385 386 if (rw->kiocb.ki_pos != -1) 387 return &rw->kiocb.ki_pos; 388 389 if (!(req->file->f_mode & FMODE_STREAM)) { 390 req->flags |= REQ_F_CUR_POS; 391 rw->kiocb.ki_pos = req->file->f_pos; 392 return &rw->kiocb.ki_pos; 393 } 394 395 rw->kiocb.ki_pos = 0; 396 return NULL; 397} 398 399#ifdef CONFIG_BLOCK 400static void io_resubmit_prep(struct io_kiocb *req) 401{ 402 struct io_async_rw *io = req->async_data; 403 404 iov_iter_restore(&io->iter, &io->iter_state); 405} 406 407static bool io_rw_should_reissue(struct io_kiocb *req) 408{ 409 umode_t mode = file_inode(req->file)->i_mode; 410 struct io_ring_ctx *ctx = req->ctx; 411 412 if (!S_ISBLK(mode) && !S_ISREG(mode)) 413 return false; 414 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() && 415 !(ctx->flags & IORING_SETUP_IOPOLL))) 416 return false; 417 /* 418 * If ref is dying, we might be running poll reap from the exit work. 419 * Don't attempt to reissue from that path, just let it fail with 420 * -EAGAIN. 421 */ 422 if (percpu_ref_is_dying(&ctx->refs)) 423 return false; 424 /* 425 * Play it safe and assume not safe to re-import and reissue if we're 426 * not in the original thread group (or in task context). 427 */ 428 if (!same_thread_group(req->task, current) || !in_task()) 429 return false; 430 return true; 431} 432#else 433static void io_resubmit_prep(struct io_kiocb *req) 434{ 435} 436static bool io_rw_should_reissue(struct io_kiocb *req) 437{ 438 return false; 439} 440#endif 441 442static void io_req_end_write(struct io_kiocb *req) 443{ 444 if (req->flags & REQ_F_ISREG) { 445 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 446 447 kiocb_end_write(&rw->kiocb); 448 } 449} 450 451/* 452 * Trigger the notifications after having done some IO, and finish the write 453 * accounting, if any. 454 */ 455static void io_req_io_end(struct io_kiocb *req) 456{ 457 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 458 459 if (rw->kiocb.ki_flags & IOCB_WRITE) { 460 io_req_end_write(req); 461 fsnotify_modify(req->file); 462 } else { 463 fsnotify_access(req->file); 464 } 465} 466 467static bool __io_complete_rw_common(struct io_kiocb *req, long res) 468{ 469 if (unlikely(res != req->cqe.res)) { 470 if ((res == -EAGAIN || res == -EOPNOTSUPP) && 471 io_rw_should_reissue(req)) { 472 /* 473 * Reissue will start accounting again, finish the 474 * current cycle. 475 */ 476 io_req_io_end(req); 477 req->flags |= REQ_F_REISSUE | REQ_F_BL_NO_RECYCLE; 478 return true; 479 } 480 req_set_fail(req); 481 req->cqe.res = res; 482 } 483 return false; 484} 485 486static inline int io_fixup_rw_res(struct io_kiocb *req, long res) 487{ 488 struct io_async_rw *io = req->async_data; 489 490 /* add previously done IO, if any */ 491 if (req_has_async_data(req) && io->bytes_done > 0) { 492 if (res < 0) 493 res = io->bytes_done; 494 else 495 res += io->bytes_done; 496 } 497 return res; 498} 499 500void io_req_rw_complete(struct io_kiocb *req, struct io_tw_state *ts) 501{ 502 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 503 struct kiocb *kiocb = &rw->kiocb; 504 505 if ((kiocb->ki_flags & IOCB_DIO_CALLER_COMP) && kiocb->dio_complete) { 506 long res = kiocb->dio_complete(rw->kiocb.private); 507 508 io_req_set_res(req, io_fixup_rw_res(req, res), 0); 509 } 510 511 io_req_io_end(req); 512 513 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) 514 req->cqe.flags |= io_put_kbuf(req, 0); 515 516 io_req_rw_cleanup(req, 0); 517 io_req_task_complete(req, ts); 518} 519 520static void io_complete_rw(struct kiocb *kiocb, long res) 521{ 522 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb); 523 struct io_kiocb *req = cmd_to_io_kiocb(rw); 524 525 if (!kiocb->dio_complete || !(kiocb->ki_flags & IOCB_DIO_CALLER_COMP)) { 526 if (__io_complete_rw_common(req, res)) 527 return; 528 io_req_set_res(req, io_fixup_rw_res(req, res), 0); 529 } 530 req->io_task_work.func = io_req_rw_complete; 531 __io_req_task_work_add(req, IOU_F_TWQ_LAZY_WAKE); 532} 533 534static void io_complete_rw_iopoll(struct kiocb *kiocb, long res) 535{ 536 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb); 537 struct io_kiocb *req = cmd_to_io_kiocb(rw); 538 539 if (kiocb->ki_flags & IOCB_WRITE) 540 io_req_end_write(req); 541 if (unlikely(res != req->cqe.res)) { 542 if (res == -EAGAIN && io_rw_should_reissue(req)) { 543 req->flags |= REQ_F_REISSUE | REQ_F_BL_NO_RECYCLE; 544 return; 545 } 546 req->cqe.res = res; 547 } 548 549 /* order with io_iopoll_complete() checking ->iopoll_completed */ 550 smp_store_release(&req->iopoll_completed, 1); 551} 552 553static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) 554{ 555 /* IO was queued async, completion will happen later */ 556 if (ret == -EIOCBQUEUED) 557 return; 558 559 /* transform internal restart error codes */ 560 if (unlikely(ret < 0)) { 561 switch (ret) { 562 case -ERESTARTSYS: 563 case -ERESTARTNOINTR: 564 case -ERESTARTNOHAND: 565 case -ERESTART_RESTARTBLOCK: 566 /* 567 * We can't just restart the syscall, since previously 568 * submitted sqes may already be in progress. Just fail 569 * this IO with EINTR. 570 */ 571 ret = -EINTR; 572 break; 573 } 574 } 575 576 INDIRECT_CALL_2(kiocb->ki_complete, io_complete_rw_iopoll, 577 io_complete_rw, kiocb, ret); 578} 579 580static int kiocb_done(struct io_kiocb *req, ssize_t ret, 581 unsigned int issue_flags) 582{ 583 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 584 unsigned final_ret = io_fixup_rw_res(req, ret); 585 586 if (ret >= 0 && req->flags & REQ_F_CUR_POS) 587 req->file->f_pos = rw->kiocb.ki_pos; 588 if (ret >= 0 && (rw->kiocb.ki_complete == io_complete_rw)) { 589 if (!__io_complete_rw_common(req, ret)) { 590 /* 591 * Safe to call io_end from here as we're inline 592 * from the submission path. 593 */ 594 io_req_io_end(req); 595 io_req_set_res(req, final_ret, 596 io_put_kbuf(req, issue_flags)); 597 io_req_rw_cleanup(req, issue_flags); 598 return IOU_OK; 599 } 600 } else { 601 io_rw_done(&rw->kiocb, ret); 602 } 603 604 if (req->flags & REQ_F_REISSUE) { 605 req->flags &= ~REQ_F_REISSUE; 606 io_resubmit_prep(req); 607 return -EAGAIN; 608 } 609 return IOU_ISSUE_SKIP_COMPLETE; 610} 611 612static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb) 613{ 614 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos; 615} 616 617/* 618 * For files that don't have ->read_iter() and ->write_iter(), handle them 619 * by looping over ->read() or ->write() manually. 620 */ 621static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter) 622{ 623 struct kiocb *kiocb = &rw->kiocb; 624 struct file *file = kiocb->ki_filp; 625 ssize_t ret = 0; 626 loff_t *ppos; 627 628 /* 629 * Don't support polled IO through this interface, and we can't 630 * support non-blocking either. For the latter, this just causes 631 * the kiocb to be handled from an async context. 632 */ 633 if (kiocb->ki_flags & IOCB_HIPRI) 634 return -EOPNOTSUPP; 635 if ((kiocb->ki_flags & IOCB_NOWAIT) && 636 !(kiocb->ki_filp->f_flags & O_NONBLOCK)) 637 return -EAGAIN; 638 639 ppos = io_kiocb_ppos(kiocb); 640 641 while (iov_iter_count(iter)) { 642 void __user *addr; 643 size_t len; 644 ssize_t nr; 645 646 if (iter_is_ubuf(iter)) { 647 addr = iter->ubuf + iter->iov_offset; 648 len = iov_iter_count(iter); 649 } else if (!iov_iter_is_bvec(iter)) { 650 addr = iter_iov_addr(iter); 651 len = iter_iov_len(iter); 652 } else { 653 addr = u64_to_user_ptr(rw->addr); 654 len = rw->len; 655 } 656 657 if (ddir == READ) 658 nr = file->f_op->read(file, addr, len, ppos); 659 else 660 nr = file->f_op->write(file, addr, len, ppos); 661 662 if (nr < 0) { 663 if (!ret) 664 ret = nr; 665 break; 666 } 667 ret += nr; 668 if (!iov_iter_is_bvec(iter)) { 669 iov_iter_advance(iter, nr); 670 } else { 671 rw->addr += nr; 672 rw->len -= nr; 673 if (!rw->len) 674 break; 675 } 676 if (nr != len) 677 break; 678 } 679 680 return ret; 681} 682 683/* 684 * This is our waitqueue callback handler, registered through __folio_lock_async() 685 * when we initially tried to do the IO with the iocb armed our waitqueue. 686 * This gets called when the page is unlocked, and we generally expect that to 687 * happen when the page IO is completed and the page is now uptodate. This will 688 * queue a task_work based retry of the operation, attempting to copy the data 689 * again. If the latter fails because the page was NOT uptodate, then we will 690 * do a thread based blocking retry of the operation. That's the unexpected 691 * slow path. 692 */ 693static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode, 694 int sync, void *arg) 695{ 696 struct wait_page_queue *wpq; 697 struct io_kiocb *req = wait->private; 698 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 699 struct wait_page_key *key = arg; 700 701 wpq = container_of(wait, struct wait_page_queue, wait); 702 703 if (!wake_page_match(wpq, key)) 704 return 0; 705 706 rw->kiocb.ki_flags &= ~IOCB_WAITQ; 707 list_del_init(&wait->entry); 708 io_req_task_queue(req); 709 return 1; 710} 711 712/* 713 * This controls whether a given IO request should be armed for async page 714 * based retry. If we return false here, the request is handed to the async 715 * worker threads for retry. If we're doing buffered reads on a regular file, 716 * we prepare a private wait_page_queue entry and retry the operation. This 717 * will either succeed because the page is now uptodate and unlocked, or it 718 * will register a callback when the page is unlocked at IO completion. Through 719 * that callback, io_uring uses task_work to setup a retry of the operation. 720 * That retry will attempt the buffered read again. The retry will generally 721 * succeed, or in rare cases where it fails, we then fall back to using the 722 * async worker threads for a blocking retry. 723 */ 724static bool io_rw_should_retry(struct io_kiocb *req) 725{ 726 struct io_async_rw *io = req->async_data; 727 struct wait_page_queue *wait = &io->wpq; 728 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 729 struct kiocb *kiocb = &rw->kiocb; 730 731 /* never retry for NOWAIT, we just complete with -EAGAIN */ 732 if (req->flags & REQ_F_NOWAIT) 733 return false; 734 735 /* Only for buffered IO */ 736 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI)) 737 return false; 738 739 /* 740 * just use poll if we can, and don't attempt if the fs doesn't 741 * support callback based unlocks 742 */ 743 if (io_file_can_poll(req) || 744 !(req->file->f_op->fop_flags & FOP_BUFFER_RASYNC)) 745 return false; 746 747 wait->wait.func = io_async_buf_func; 748 wait->wait.private = req; 749 wait->wait.flags = 0; 750 INIT_LIST_HEAD(&wait->wait.entry); 751 kiocb->ki_flags |= IOCB_WAITQ; 752 kiocb->ki_flags &= ~IOCB_NOWAIT; 753 kiocb->ki_waitq = wait; 754 return true; 755} 756 757static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter) 758{ 759 struct file *file = rw->kiocb.ki_filp; 760 761 if (likely(file->f_op->read_iter)) 762 return call_read_iter(file, &rw->kiocb, iter); 763 else if (file->f_op->read) 764 return loop_rw_iter(READ, rw, iter); 765 else 766 return -EINVAL; 767} 768 769static bool need_complete_io(struct io_kiocb *req) 770{ 771 return req->flags & REQ_F_ISREG || 772 S_ISBLK(file_inode(req->file)->i_mode); 773} 774 775static int io_rw_init_file(struct io_kiocb *req, fmode_t mode) 776{ 777 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 778 struct kiocb *kiocb = &rw->kiocb; 779 struct io_ring_ctx *ctx = req->ctx; 780 struct file *file = req->file; 781 int ret; 782 783 if (unlikely(!(file->f_mode & mode))) 784 return -EBADF; 785 786 if (!(req->flags & REQ_F_FIXED_FILE)) 787 req->flags |= io_file_get_flags(file); 788 789 kiocb->ki_flags = file->f_iocb_flags; 790 ret = kiocb_set_rw_flags(kiocb, rw->flags); 791 if (unlikely(ret)) 792 return ret; 793 kiocb->ki_flags |= IOCB_ALLOC_CACHE; 794 795 /* 796 * If the file is marked O_NONBLOCK, still allow retry for it if it 797 * supports async. Otherwise it's impossible to use O_NONBLOCK files 798 * reliably. If not, or it IOCB_NOWAIT is set, don't retry. 799 */ 800 if ((kiocb->ki_flags & IOCB_NOWAIT) || 801 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req))) 802 req->flags |= REQ_F_NOWAIT; 803 804 if (ctx->flags & IORING_SETUP_IOPOLL) { 805 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll) 806 return -EOPNOTSUPP; 807 808 kiocb->private = NULL; 809 kiocb->ki_flags |= IOCB_HIPRI; 810 kiocb->ki_complete = io_complete_rw_iopoll; 811 req->iopoll_completed = 0; 812 } else { 813 if (kiocb->ki_flags & IOCB_HIPRI) 814 return -EINVAL; 815 kiocb->ki_complete = io_complete_rw; 816 } 817 818 return 0; 819} 820 821static int __io_read(struct io_kiocb *req, unsigned int issue_flags) 822{ 823 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; 824 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 825 struct io_async_rw *io = req->async_data; 826 struct kiocb *kiocb = &rw->kiocb; 827 ssize_t ret; 828 loff_t *ppos; 829 830 if (io_do_buffer_select(req)) { 831 ret = io_import_iovec(ITER_DEST, req, io, issue_flags); 832 if (unlikely(ret < 0)) 833 return ret; 834 } 835 836 ret = io_rw_init_file(req, FMODE_READ); 837 if (unlikely(ret)) 838 return ret; 839 req->cqe.res = iov_iter_count(&io->iter); 840 841 if (force_nonblock) { 842 /* If the file doesn't support async, just async punt */ 843 if (unlikely(!io_file_supports_nowait(req))) 844 return -EAGAIN; 845 kiocb->ki_flags |= IOCB_NOWAIT; 846 } else { 847 /* Ensure we clear previously set non-block flag */ 848 kiocb->ki_flags &= ~IOCB_NOWAIT; 849 } 850 851 ppos = io_kiocb_update_pos(req); 852 853 ret = rw_verify_area(READ, req->file, ppos, req->cqe.res); 854 if (unlikely(ret)) 855 return ret; 856 857 ret = io_iter_do_read(rw, &io->iter); 858 859 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) { 860 req->flags &= ~REQ_F_REISSUE; 861 /* If we can poll, just do that. */ 862 if (io_file_can_poll(req)) 863 return -EAGAIN; 864 /* IOPOLL retry should happen for io-wq threads */ 865 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL)) 866 goto done; 867 /* no retry on NONBLOCK nor RWF_NOWAIT */ 868 if (req->flags & REQ_F_NOWAIT) 869 goto done; 870 ret = 0; 871 } else if (ret == -EIOCBQUEUED) { 872 return IOU_ISSUE_SKIP_COMPLETE; 873 } else if (ret == req->cqe.res || ret <= 0 || !force_nonblock || 874 (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) { 875 /* read all, failed, already did sync or don't want to retry */ 876 goto done; 877 } 878 879 /* 880 * Don't depend on the iter state matching what was consumed, or being 881 * untouched in case of error. Restore it and we'll advance it 882 * manually if we need to. 883 */ 884 iov_iter_restore(&io->iter, &io->iter_state); 885 886 do { 887 /* 888 * We end up here because of a partial read, either from 889 * above or inside this loop. Advance the iter by the bytes 890 * that were consumed. 891 */ 892 iov_iter_advance(&io->iter, ret); 893 if (!iov_iter_count(&io->iter)) 894 break; 895 io->bytes_done += ret; 896 iov_iter_save_state(&io->iter, &io->iter_state); 897 898 /* if we can retry, do so with the callbacks armed */ 899 if (!io_rw_should_retry(req)) { 900 kiocb->ki_flags &= ~IOCB_WAITQ; 901 return -EAGAIN; 902 } 903 904 req->cqe.res = iov_iter_count(&io->iter); 905 /* 906 * Now retry read with the IOCB_WAITQ parts set in the iocb. If 907 * we get -EIOCBQUEUED, then we'll get a notification when the 908 * desired page gets unlocked. We can also get a partial read 909 * here, and if we do, then just retry at the new offset. 910 */ 911 ret = io_iter_do_read(rw, &io->iter); 912 if (ret == -EIOCBQUEUED) 913 return IOU_ISSUE_SKIP_COMPLETE; 914 /* we got some bytes, but not all. retry. */ 915 kiocb->ki_flags &= ~IOCB_WAITQ; 916 iov_iter_restore(&io->iter, &io->iter_state); 917 } while (ret > 0); 918done: 919 /* it's faster to check here then delegate to kfree */ 920 return ret; 921} 922 923int io_read(struct io_kiocb *req, unsigned int issue_flags) 924{ 925 int ret; 926 927 ret = __io_read(req, issue_flags); 928 if (ret >= 0) 929 return kiocb_done(req, ret, issue_flags); 930 931 return ret; 932} 933 934int io_read_mshot(struct io_kiocb *req, unsigned int issue_flags) 935{ 936 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 937 unsigned int cflags = 0; 938 int ret; 939 940 /* 941 * Multishot MUST be used on a pollable file 942 */ 943 if (!io_file_can_poll(req)) 944 return -EBADFD; 945 946 ret = __io_read(req, issue_flags); 947 948 /* 949 * If the file doesn't support proper NOWAIT, then disable multishot 950 * and stay in single shot mode. 951 */ 952 if (!io_file_supports_nowait(req)) 953 req->flags &= ~REQ_F_APOLL_MULTISHOT; 954 955 /* 956 * If we get -EAGAIN, recycle our buffer and just let normal poll 957 * handling arm it. 958 */ 959 if (ret == -EAGAIN) { 960 /* 961 * Reset rw->len to 0 again to avoid clamping future mshot 962 * reads, in case the buffer size varies. 963 */ 964 if (io_kbuf_recycle(req, issue_flags)) 965 rw->len = 0; 966 if (issue_flags & IO_URING_F_MULTISHOT) 967 return IOU_ISSUE_SKIP_COMPLETE; 968 return -EAGAIN; 969 } 970 971 /* 972 * Any successful return value will keep the multishot read armed. 973 */ 974 if (ret > 0 && req->flags & REQ_F_APOLL_MULTISHOT) { 975 /* 976 * Put our buffer and post a CQE. If we fail to post a CQE, then 977 * jump to the termination path. This request is then done. 978 */ 979 cflags = io_put_kbuf(req, issue_flags); 980 rw->len = 0; /* similarly to above, reset len to 0 */ 981 982 if (io_req_post_cqe(req, ret, cflags | IORING_CQE_F_MORE)) { 983 if (issue_flags & IO_URING_F_MULTISHOT) { 984 /* 985 * Force retry, as we might have more data to 986 * be read and otherwise it won't get retried 987 * until (if ever) another poll is triggered. 988 */ 989 io_poll_multishot_retry(req); 990 return IOU_ISSUE_SKIP_COMPLETE; 991 } 992 return -EAGAIN; 993 } 994 } 995 996 /* 997 * Either an error, or we've hit overflow posting the CQE. For any 998 * multishot request, hitting overflow will terminate it. 999 */ 1000 io_req_set_res(req, ret, cflags); 1001 io_req_rw_cleanup(req, issue_flags); 1002 if (issue_flags & IO_URING_F_MULTISHOT) 1003 return IOU_STOP_MULTISHOT; 1004 return IOU_OK; 1005} 1006 1007int io_write(struct io_kiocb *req, unsigned int issue_flags) 1008{ 1009 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; 1010 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 1011 struct io_async_rw *io = req->async_data; 1012 struct kiocb *kiocb = &rw->kiocb; 1013 ssize_t ret, ret2; 1014 loff_t *ppos; 1015 1016 ret = io_rw_init_file(req, FMODE_WRITE); 1017 if (unlikely(ret)) 1018 return ret; 1019 req->cqe.res = iov_iter_count(&io->iter); 1020 1021 if (force_nonblock) { 1022 /* If the file doesn't support async, just async punt */ 1023 if (unlikely(!io_file_supports_nowait(req))) 1024 goto ret_eagain; 1025 1026 /* Check if we can support NOWAIT. */ 1027 if (!(kiocb->ki_flags & IOCB_DIRECT) && 1028 !(req->file->f_op->fop_flags & FOP_BUFFER_WASYNC) && 1029 (req->flags & REQ_F_ISREG)) 1030 goto ret_eagain; 1031 1032 kiocb->ki_flags |= IOCB_NOWAIT; 1033 } else { 1034 /* Ensure we clear previously set non-block flag */ 1035 kiocb->ki_flags &= ~IOCB_NOWAIT; 1036 } 1037 1038 ppos = io_kiocb_update_pos(req); 1039 1040 ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res); 1041 if (unlikely(ret)) 1042 return ret; 1043 1044 if (req->flags & REQ_F_ISREG) 1045 kiocb_start_write(kiocb); 1046 kiocb->ki_flags |= IOCB_WRITE; 1047 1048 if (likely(req->file->f_op->write_iter)) 1049 ret2 = call_write_iter(req->file, kiocb, &io->iter); 1050 else if (req->file->f_op->write) 1051 ret2 = loop_rw_iter(WRITE, rw, &io->iter); 1052 else 1053 ret2 = -EINVAL; 1054 1055 if (req->flags & REQ_F_REISSUE) { 1056 req->flags &= ~REQ_F_REISSUE; 1057 ret2 = -EAGAIN; 1058 } 1059 1060 /* 1061 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just 1062 * retry them without IOCB_NOWAIT. 1063 */ 1064 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT)) 1065 ret2 = -EAGAIN; 1066 /* no retry on NONBLOCK nor RWF_NOWAIT */ 1067 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT)) 1068 goto done; 1069 if (!force_nonblock || ret2 != -EAGAIN) { 1070 /* IOPOLL retry should happen for io-wq threads */ 1071 if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL)) 1072 goto ret_eagain; 1073 1074 if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) { 1075 trace_io_uring_short_write(req->ctx, kiocb->ki_pos - ret2, 1076 req->cqe.res, ret2); 1077 1078 /* This is a partial write. The file pos has already been 1079 * updated, setup the async struct to complete the request 1080 * in the worker. Also update bytes_done to account for 1081 * the bytes already written. 1082 */ 1083 iov_iter_save_state(&io->iter, &io->iter_state); 1084 io->bytes_done += ret2; 1085 1086 if (kiocb->ki_flags & IOCB_WRITE) 1087 io_req_end_write(req); 1088 return -EAGAIN; 1089 } 1090done: 1091 return kiocb_done(req, ret2, issue_flags); 1092 } else { 1093ret_eagain: 1094 iov_iter_restore(&io->iter, &io->iter_state); 1095 if (kiocb->ki_flags & IOCB_WRITE) 1096 io_req_end_write(req); 1097 return -EAGAIN; 1098 } 1099} 1100 1101void io_rw_fail(struct io_kiocb *req) 1102{ 1103 int res; 1104 1105 res = io_fixup_rw_res(req, req->cqe.res); 1106 io_req_set_res(req, res, req->cqe.flags); 1107} 1108 1109int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin) 1110{ 1111 struct io_wq_work_node *pos, *start, *prev; 1112 unsigned int poll_flags = 0; 1113 DEFINE_IO_COMP_BATCH(iob); 1114 int nr_events = 0; 1115 1116 /* 1117 * Only spin for completions if we don't have multiple devices hanging 1118 * off our complete list. 1119 */ 1120 if (ctx->poll_multi_queue || force_nonspin) 1121 poll_flags |= BLK_POLL_ONESHOT; 1122 1123 wq_list_for_each(pos, start, &ctx->iopoll_list) { 1124 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list); 1125 struct file *file = req->file; 1126 int ret; 1127 1128 /* 1129 * Move completed and retryable entries to our local lists. 1130 * If we find a request that requires polling, break out 1131 * and complete those lists first, if we have entries there. 1132 */ 1133 if (READ_ONCE(req->iopoll_completed)) 1134 break; 1135 1136 if (req->opcode == IORING_OP_URING_CMD) { 1137 struct io_uring_cmd *ioucmd; 1138 1139 ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd); 1140 ret = file->f_op->uring_cmd_iopoll(ioucmd, &iob, 1141 poll_flags); 1142 } else { 1143 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); 1144 1145 ret = file->f_op->iopoll(&rw->kiocb, &iob, poll_flags); 1146 } 1147 if (unlikely(ret < 0)) 1148 return ret; 1149 else if (ret) 1150 poll_flags |= BLK_POLL_ONESHOT; 1151 1152 /* iopoll may have completed current req */ 1153 if (!rq_list_empty(iob.req_list) || 1154 READ_ONCE(req->iopoll_completed)) 1155 break; 1156 } 1157 1158 if (!rq_list_empty(iob.req_list)) 1159 iob.complete(&iob); 1160 else if (!pos) 1161 return 0; 1162 1163 prev = start; 1164 wq_list_for_each_resume(pos, prev) { 1165 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list); 1166 1167 /* order with io_complete_rw_iopoll(), e.g. ->result updates */ 1168 if (!smp_load_acquire(&req->iopoll_completed)) 1169 break; 1170 nr_events++; 1171 req->cqe.flags = io_put_kbuf(req, 0); 1172 if (req->opcode != IORING_OP_URING_CMD) 1173 io_req_rw_cleanup(req, 0); 1174 } 1175 if (unlikely(!nr_events)) 1176 return 0; 1177 1178 pos = start ? start->next : ctx->iopoll_list.first; 1179 wq_list_cut(&ctx->iopoll_list, prev, start); 1180 1181 if (WARN_ON_ONCE(!wq_list_empty(&ctx->submit_state.compl_reqs))) 1182 return 0; 1183 ctx->submit_state.compl_reqs.first = pos; 1184 __io_submit_flush_completions(ctx); 1185 return nr_events; 1186} 1187 1188void io_rw_cache_free(const void *entry) 1189{ 1190 struct io_async_rw *rw = (struct io_async_rw *) entry; 1191 1192 if (rw->free_iovec) { 1193 kasan_mempool_unpoison_object(rw->free_iovec, 1194 rw->free_iov_nr * sizeof(struct iovec)); 1195 io_rw_iovec_free(rw); 1196 } 1197 kfree(rw); 1198} 1199