1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
3 #include <linux/errno.h>
5 #include <linux/file.h>
6 #include <linux/blk-mq.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>
15 #include <uapi/linux/io_uring.h>
24 /* NOTE: kiocb has the file as the first member, so don't do it here */
31 static inline bool io_file_supports_nowait(struct io_kiocb *req)
33 return req->flags & REQ_F_SUPPORT_NOWAIT;
37 static int io_iov_compat_buffer_select_prep(struct io_rw *rw)
39 struct compat_iovec __user *uiov;
42 uiov = u64_to_user_ptr(rw->addr);
43 if (!access_ok(uiov, sizeof(*uiov)))
45 if (__get_user(clen, &uiov->iov_len))
55 static int io_iov_buffer_select_prep(struct io_kiocb *req)
57 struct iovec __user *uiov;
59 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
66 return io_iov_compat_buffer_select_prep(rw);
69 uiov = u64_to_user_ptr(rw->addr);
70 if (copy_from_user(&iov, uiov, sizeof(*uiov)))
72 rw->len = iov.iov_len;
76 int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
78 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
82 rw->kiocb.ki_pos = READ_ONCE(sqe->off);
83 /* used for fixed read/write too - just read unconditionally */
84 req->buf_index = READ_ONCE(sqe->buf_index);
86 ioprio = READ_ONCE(sqe->ioprio);
88 ret = ioprio_check_cap(ioprio);
92 rw->kiocb.ki_ioprio = ioprio;
94 rw->kiocb.ki_ioprio = get_current_ioprio();
96 rw->kiocb.dio_complete = NULL;
98 rw->addr = READ_ONCE(sqe->addr);
99 rw->len = READ_ONCE(sqe->len);
100 rw->flags = READ_ONCE(sqe->rw_flags);
104 int io_prep_rwv(struct io_kiocb *req, const struct io_uring_sqe *sqe)
108 ret = io_prep_rw(req, sqe);
113 * Have to do this validation here, as this is in io_read() rw->len
114 * might have chanaged due to buffer selection
116 if (req->flags & REQ_F_BUFFER_SELECT)
117 return io_iov_buffer_select_prep(req);
122 int io_prep_rw_fixed(struct io_kiocb *req, const struct io_uring_sqe *sqe)
124 struct io_ring_ctx *ctx = req->ctx;
128 ret = io_prep_rw(req, sqe);
132 if (unlikely(req->buf_index >= ctx->nr_user_bufs))
134 index = array_index_nospec(req->buf_index, ctx->nr_user_bufs);
135 req->imu = ctx->user_bufs[index];
136 io_req_set_rsrc_node(req, ctx, 0);
141 * Multishot read is prepared just like a normal read/write request, only
142 * difference is that we set the MULTISHOT flag.
144 int io_read_mshot_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
146 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
149 /* must be used with provided buffers */
150 if (!(req->flags & REQ_F_BUFFER_SELECT))
153 ret = io_prep_rw(req, sqe);
157 if (rw->addr || rw->len)
160 req->flags |= REQ_F_APOLL_MULTISHOT;
164 void io_readv_writev_cleanup(struct io_kiocb *req)
166 struct io_async_rw *io = req->async_data;
168 kfree(io->free_iovec);
171 static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req)
173 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
175 if (rw->kiocb.ki_pos != -1)
176 return &rw->kiocb.ki_pos;
178 if (!(req->file->f_mode & FMODE_STREAM)) {
179 req->flags |= REQ_F_CUR_POS;
180 rw->kiocb.ki_pos = req->file->f_pos;
181 return &rw->kiocb.ki_pos;
184 rw->kiocb.ki_pos = 0;
188 static void io_req_task_queue_reissue(struct io_kiocb *req)
190 req->io_task_work.func = io_queue_iowq;
191 io_req_task_work_add(req);
195 static bool io_resubmit_prep(struct io_kiocb *req)
197 struct io_async_rw *io = req->async_data;
199 if (!req_has_async_data(req))
200 return !io_req_prep_async(req);
201 iov_iter_restore(&io->s.iter, &io->s.iter_state);
205 static bool io_rw_should_reissue(struct io_kiocb *req)
207 umode_t mode = file_inode(req->file)->i_mode;
208 struct io_ring_ctx *ctx = req->ctx;
210 if (!S_ISBLK(mode) && !S_ISREG(mode))
212 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
213 !(ctx->flags & IORING_SETUP_IOPOLL)))
216 * If ref is dying, we might be running poll reap from the exit work.
217 * Don't attempt to reissue from that path, just let it fail with
220 if (percpu_ref_is_dying(&ctx->refs))
223 * Play it safe and assume not safe to re-import and reissue if we're
224 * not in the original thread group (or in task context).
226 if (!same_thread_group(req->task, current) || !in_task())
231 static bool io_resubmit_prep(struct io_kiocb *req)
235 static bool io_rw_should_reissue(struct io_kiocb *req)
241 static void io_req_end_write(struct io_kiocb *req)
243 if (req->flags & REQ_F_ISREG) {
244 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
246 kiocb_end_write(&rw->kiocb);
251 * Trigger the notifications after having done some IO, and finish the write
252 * accounting, if any.
254 static void io_req_io_end(struct io_kiocb *req)
256 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
258 if (rw->kiocb.ki_flags & IOCB_WRITE) {
259 io_req_end_write(req);
260 fsnotify_modify(req->file);
262 fsnotify_access(req->file);
266 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
268 if (unlikely(res != req->cqe.res)) {
269 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
270 io_rw_should_reissue(req)) {
272 * Reissue will start accounting again, finish the
276 req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
285 static inline int io_fixup_rw_res(struct io_kiocb *req, long res)
287 struct io_async_rw *io = req->async_data;
289 /* add previously done IO, if any */
290 if (req_has_async_data(req) && io->bytes_done > 0) {
292 res = io->bytes_done;
294 res += io->bytes_done;
299 void io_req_rw_complete(struct io_kiocb *req, struct io_tw_state *ts)
301 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
302 struct kiocb *kiocb = &rw->kiocb;
304 if ((kiocb->ki_flags & IOCB_DIO_CALLER_COMP) && kiocb->dio_complete) {
305 long res = kiocb->dio_complete(rw->kiocb.private);
307 io_req_set_res(req, io_fixup_rw_res(req, res), 0);
312 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
313 unsigned issue_flags = ts->locked ? 0 : IO_URING_F_UNLOCKED;
315 req->cqe.flags |= io_put_kbuf(req, issue_flags);
317 io_req_task_complete(req, ts);
320 static void io_complete_rw(struct kiocb *kiocb, long res)
322 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
323 struct io_kiocb *req = cmd_to_io_kiocb(rw);
325 if (!kiocb->dio_complete || !(kiocb->ki_flags & IOCB_DIO_CALLER_COMP)) {
326 if (__io_complete_rw_common(req, res))
328 io_req_set_res(req, io_fixup_rw_res(req, res), 0);
330 req->io_task_work.func = io_req_rw_complete;
331 __io_req_task_work_add(req, IOU_F_TWQ_LAZY_WAKE);
334 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
336 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
337 struct io_kiocb *req = cmd_to_io_kiocb(rw);
339 if (kiocb->ki_flags & IOCB_WRITE)
340 io_req_end_write(req);
341 if (unlikely(res != req->cqe.res)) {
342 if (res == -EAGAIN && io_rw_should_reissue(req)) {
343 req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
349 /* order with io_iopoll_complete() checking ->iopoll_completed */
350 smp_store_release(&req->iopoll_completed, 1);
353 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
355 /* IO was queued async, completion will happen later */
356 if (ret == -EIOCBQUEUED)
359 /* transform internal restart error codes */
360 if (unlikely(ret < 0)) {
363 case -ERESTARTNOINTR:
364 case -ERESTARTNOHAND:
365 case -ERESTART_RESTARTBLOCK:
367 * We can't just restart the syscall, since previously
368 * submitted sqes may already be in progress. Just fail
369 * this IO with EINTR.
376 INDIRECT_CALL_2(kiocb->ki_complete, io_complete_rw_iopoll,
377 io_complete_rw, kiocb, ret);
380 static int kiocb_done(struct io_kiocb *req, ssize_t ret,
381 unsigned int issue_flags)
383 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
384 unsigned final_ret = io_fixup_rw_res(req, ret);
386 if (ret >= 0 && req->flags & REQ_F_CUR_POS)
387 req->file->f_pos = rw->kiocb.ki_pos;
388 if (ret >= 0 && (rw->kiocb.ki_complete == io_complete_rw)) {
389 if (!__io_complete_rw_common(req, ret)) {
391 * Safe to call io_end from here as we're inline
392 * from the submission path.
395 io_req_set_res(req, final_ret,
396 io_put_kbuf(req, issue_flags));
400 io_rw_done(&rw->kiocb, ret);
403 if (req->flags & REQ_F_REISSUE) {
404 req->flags &= ~REQ_F_REISSUE;
405 if (io_resubmit_prep(req))
406 io_req_task_queue_reissue(req);
408 io_req_task_queue_fail(req, final_ret);
410 return IOU_ISSUE_SKIP_COMPLETE;
413 static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req,
414 struct io_rw_state *s,
415 unsigned int issue_flags)
417 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
418 struct iov_iter *iter = &s->iter;
419 u8 opcode = req->opcode;
425 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
426 ret = io_import_fixed(ddir, iter, req->imu, rw->addr, rw->len);
432 buf = u64_to_user_ptr(rw->addr);
435 if (!io_issue_defs[opcode].vectored || req->flags & REQ_F_BUFFER_SELECT) {
436 if (io_do_buffer_select(req)) {
437 buf = io_buffer_select(req, &sqe_len, issue_flags);
439 return ERR_PTR(-ENOBUFS);
440 rw->addr = (unsigned long) buf;
444 ret = import_ubuf(ddir, buf, sqe_len, iter);
451 ret = __import_iovec(ddir, buf, sqe_len, UIO_FASTIOV, &iovec, iter,
453 if (unlikely(ret < 0))
458 static inline int io_import_iovec(int rw, struct io_kiocb *req,
459 struct iovec **iovec, struct io_rw_state *s,
460 unsigned int issue_flags)
462 *iovec = __io_import_iovec(rw, req, s, issue_flags);
464 return PTR_ERR(*iovec);
466 iov_iter_save_state(&s->iter, &s->iter_state);
470 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
472 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
476 * For files that don't have ->read_iter() and ->write_iter(), handle them
477 * by looping over ->read() or ->write() manually.
479 static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter)
481 struct kiocb *kiocb = &rw->kiocb;
482 struct file *file = kiocb->ki_filp;
487 * Don't support polled IO through this interface, and we can't
488 * support non-blocking either. For the latter, this just causes
489 * the kiocb to be handled from an async context.
491 if (kiocb->ki_flags & IOCB_HIPRI)
493 if ((kiocb->ki_flags & IOCB_NOWAIT) &&
494 !(kiocb->ki_filp->f_flags & O_NONBLOCK))
497 ppos = io_kiocb_ppos(kiocb);
499 while (iov_iter_count(iter)) {
504 if (iter_is_ubuf(iter)) {
505 addr = iter->ubuf + iter->iov_offset;
506 len = iov_iter_count(iter);
507 } else if (!iov_iter_is_bvec(iter)) {
508 addr = iter_iov_addr(iter);
509 len = iter_iov_len(iter);
511 addr = u64_to_user_ptr(rw->addr);
516 nr = file->f_op->read(file, addr, len, ppos);
518 nr = file->f_op->write(file, addr, len, ppos);
526 if (!iov_iter_is_bvec(iter)) {
527 iov_iter_advance(iter, nr);
541 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
542 const struct iovec *fast_iov, struct iov_iter *iter)
544 struct io_async_rw *io = req->async_data;
546 memcpy(&io->s.iter, iter, sizeof(*iter));
547 io->free_iovec = iovec;
549 /* can only be fixed buffers, no need to do anything */
550 if (iov_iter_is_bvec(iter) || iter_is_ubuf(iter))
553 unsigned iov_off = 0;
555 io->s.iter.__iov = io->s.fast_iov;
556 if (iter->__iov != fast_iov) {
557 iov_off = iter_iov(iter) - fast_iov;
558 io->s.iter.__iov += iov_off;
560 if (io->s.fast_iov != fast_iov)
561 memcpy(io->s.fast_iov + iov_off, fast_iov + iov_off,
562 sizeof(struct iovec) * iter->nr_segs);
564 req->flags |= REQ_F_NEED_CLEANUP;
568 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
569 struct io_rw_state *s, bool force)
571 if (!force && !io_cold_defs[req->opcode].prep_async)
573 /* opcode type doesn't need async data */
574 if (!io_cold_defs[req->opcode].async_size)
576 if (!req_has_async_data(req)) {
577 struct io_async_rw *iorw;
579 if (io_alloc_async_data(req)) {
584 io_req_map_rw(req, iovec, s->fast_iov, &s->iter);
585 iorw = req->async_data;
586 /* we've copied and mapped the iter, ensure state is saved */
587 iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state);
592 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
594 struct io_async_rw *iorw = req->async_data;
598 iorw->bytes_done = 0;
599 iorw->free_iovec = NULL;
601 /* submission path, ->uring_lock should already be taken */
602 ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
603 if (unlikely(ret < 0))
607 iorw->free_iovec = iov;
608 req->flags |= REQ_F_NEED_CLEANUP;
614 int io_readv_prep_async(struct io_kiocb *req)
616 return io_rw_prep_async(req, ITER_DEST);
619 int io_writev_prep_async(struct io_kiocb *req)
621 return io_rw_prep_async(req, ITER_SOURCE);
625 * This is our waitqueue callback handler, registered through __folio_lock_async()
626 * when we initially tried to do the IO with the iocb armed our waitqueue.
627 * This gets called when the page is unlocked, and we generally expect that to
628 * happen when the page IO is completed and the page is now uptodate. This will
629 * queue a task_work based retry of the operation, attempting to copy the data
630 * again. If the latter fails because the page was NOT uptodate, then we will
631 * do a thread based blocking retry of the operation. That's the unexpected
634 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
637 struct wait_page_queue *wpq;
638 struct io_kiocb *req = wait->private;
639 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
640 struct wait_page_key *key = arg;
642 wpq = container_of(wait, struct wait_page_queue, wait);
644 if (!wake_page_match(wpq, key))
647 rw->kiocb.ki_flags &= ~IOCB_WAITQ;
648 list_del_init(&wait->entry);
649 io_req_task_queue(req);
654 * This controls whether a given IO request should be armed for async page
655 * based retry. If we return false here, the request is handed to the async
656 * worker threads for retry. If we're doing buffered reads on a regular file,
657 * we prepare a private wait_page_queue entry and retry the operation. This
658 * will either succeed because the page is now uptodate and unlocked, or it
659 * will register a callback when the page is unlocked at IO completion. Through
660 * that callback, io_uring uses task_work to setup a retry of the operation.
661 * That retry will attempt the buffered read again. The retry will generally
662 * succeed, or in rare cases where it fails, we then fall back to using the
663 * async worker threads for a blocking retry.
665 static bool io_rw_should_retry(struct io_kiocb *req)
667 struct io_async_rw *io = req->async_data;
668 struct wait_page_queue *wait = &io->wpq;
669 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
670 struct kiocb *kiocb = &rw->kiocb;
672 /* never retry for NOWAIT, we just complete with -EAGAIN */
673 if (req->flags & REQ_F_NOWAIT)
676 /* Only for buffered IO */
677 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
681 * just use poll if we can, and don't attempt if the fs doesn't
682 * support callback based unlocks
684 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
687 wait->wait.func = io_async_buf_func;
688 wait->wait.private = req;
689 wait->wait.flags = 0;
690 INIT_LIST_HEAD(&wait->wait.entry);
691 kiocb->ki_flags |= IOCB_WAITQ;
692 kiocb->ki_flags &= ~IOCB_NOWAIT;
693 kiocb->ki_waitq = wait;
697 static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
699 struct file *file = rw->kiocb.ki_filp;
701 if (likely(file->f_op->read_iter))
702 return call_read_iter(file, &rw->kiocb, iter);
703 else if (file->f_op->read)
704 return loop_rw_iter(READ, rw, iter);
709 static bool need_complete_io(struct io_kiocb *req)
711 return req->flags & REQ_F_ISREG ||
712 S_ISBLK(file_inode(req->file)->i_mode);
715 static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
717 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
718 struct kiocb *kiocb = &rw->kiocb;
719 struct io_ring_ctx *ctx = req->ctx;
720 struct file *file = req->file;
723 if (unlikely(!file || !(file->f_mode & mode)))
726 if (!(req->flags & REQ_F_FIXED_FILE))
727 req->flags |= io_file_get_flags(file);
729 kiocb->ki_flags = file->f_iocb_flags;
730 ret = kiocb_set_rw_flags(kiocb, rw->flags);
733 kiocb->ki_flags |= IOCB_ALLOC_CACHE;
736 * If the file is marked O_NONBLOCK, still allow retry for it if it
737 * supports async. Otherwise it's impossible to use O_NONBLOCK files
738 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
740 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
741 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
742 req->flags |= REQ_F_NOWAIT;
744 if (ctx->flags & IORING_SETUP_IOPOLL) {
745 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
748 kiocb->private = NULL;
749 kiocb->ki_flags |= IOCB_HIPRI;
750 kiocb->ki_complete = io_complete_rw_iopoll;
751 req->iopoll_completed = 0;
753 if (kiocb->ki_flags & IOCB_HIPRI)
755 kiocb->ki_complete = io_complete_rw;
761 static int __io_read(struct io_kiocb *req, unsigned int issue_flags)
763 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
764 struct io_rw_state __s, *s = &__s;
766 struct kiocb *kiocb = &rw->kiocb;
767 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
768 struct io_async_rw *io;
772 if (!req_has_async_data(req)) {
773 ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
774 if (unlikely(ret < 0))
777 io = req->async_data;
781 * Safe and required to re-import if we're using provided
782 * buffers, as we dropped the selected one before retry.
784 if (io_do_buffer_select(req)) {
785 ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
786 if (unlikely(ret < 0))
791 * We come here from an earlier attempt, restore our state to
792 * match in case it doesn't. It's cheap enough that we don't
793 * need to make this conditional.
795 iov_iter_restore(&s->iter, &s->iter_state);
798 ret = io_rw_init_file(req, FMODE_READ);
803 req->cqe.res = iov_iter_count(&s->iter);
805 if (force_nonblock) {
806 /* If the file doesn't support async, just async punt */
807 if (unlikely(!io_file_supports_nowait(req))) {
808 ret = io_setup_async_rw(req, iovec, s, true);
809 return ret ?: -EAGAIN;
811 kiocb->ki_flags |= IOCB_NOWAIT;
813 /* Ensure we clear previously set non-block flag */
814 kiocb->ki_flags &= ~IOCB_NOWAIT;
817 ppos = io_kiocb_update_pos(req);
819 ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
825 ret = io_iter_do_read(rw, &s->iter);
827 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
828 req->flags &= ~REQ_F_REISSUE;
830 * If we can poll, just do that. For a vectored read, we'll
831 * need to copy state first.
833 if (file_can_poll(req->file) && !io_issue_defs[req->opcode].vectored)
835 /* IOPOLL retry should happen for io-wq threads */
836 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
838 /* no retry on NONBLOCK nor RWF_NOWAIT */
839 if (req->flags & REQ_F_NOWAIT)
842 } else if (ret == -EIOCBQUEUED) {
845 return IOU_ISSUE_SKIP_COMPLETE;
846 } else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
847 (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) {
848 /* read all, failed, already did sync or don't want to retry */
853 * Don't depend on the iter state matching what was consumed, or being
854 * untouched in case of error. Restore it and we'll advance it
855 * manually if we need to.
857 iov_iter_restore(&s->iter, &s->iter_state);
859 ret2 = io_setup_async_rw(req, iovec, s, true);
862 ret = ret > 0 ? ret : ret2;
866 io = req->async_data;
869 * Now use our persistent iterator and state, if we aren't already.
870 * We've restored and mapped the iter to match.
875 * We end up here because of a partial read, either from
876 * above or inside this loop. Advance the iter by the bytes
877 * that were consumed.
879 iov_iter_advance(&s->iter, ret);
880 if (!iov_iter_count(&s->iter))
882 io->bytes_done += ret;
883 iov_iter_save_state(&s->iter, &s->iter_state);
885 /* if we can retry, do so with the callbacks armed */
886 if (!io_rw_should_retry(req)) {
887 kiocb->ki_flags &= ~IOCB_WAITQ;
891 req->cqe.res = iov_iter_count(&s->iter);
893 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
894 * we get -EIOCBQUEUED, then we'll get a notification when the
895 * desired page gets unlocked. We can also get a partial read
896 * here, and if we do, then just retry at the new offset.
898 ret = io_iter_do_read(rw, &s->iter);
899 if (ret == -EIOCBQUEUED)
900 return IOU_ISSUE_SKIP_COMPLETE;
901 /* we got some bytes, but not all. retry. */
902 kiocb->ki_flags &= ~IOCB_WAITQ;
903 iov_iter_restore(&s->iter, &s->iter_state);
906 /* it's faster to check here then delegate to kfree */
912 int io_read(struct io_kiocb *req, unsigned int issue_flags)
916 ret = __io_read(req, issue_flags);
918 return kiocb_done(req, ret, issue_flags);
923 int io_read_mshot(struct io_kiocb *req, unsigned int issue_flags)
925 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
926 unsigned int cflags = 0;
930 * Multishot MUST be used on a pollable file
932 if (!file_can_poll(req->file))
935 ret = __io_read(req, issue_flags);
938 * If we get -EAGAIN, recycle our buffer and just let normal poll
941 if (ret == -EAGAIN) {
943 * Reset rw->len to 0 again to avoid clamping future mshot
944 * reads, in case the buffer size varies.
946 if (io_kbuf_recycle(req, issue_flags))
952 * Any successful return value will keep the multishot read armed.
956 * Put our buffer and post a CQE. If we fail to post a CQE, then
957 * jump to the termination path. This request is then done.
959 cflags = io_put_kbuf(req, issue_flags);
960 rw->len = 0; /* similarly to above, reset len to 0 */
962 if (io_fill_cqe_req_aux(req,
963 issue_flags & IO_URING_F_COMPLETE_DEFER,
964 ret, cflags | IORING_CQE_F_MORE)) {
965 if (issue_flags & IO_URING_F_MULTISHOT)
966 return IOU_ISSUE_SKIP_COMPLETE;
972 * Either an error, or we've hit overflow posting the CQE. For any
973 * multishot request, hitting overflow will terminate it.
975 io_req_set_res(req, ret, cflags);
976 if (issue_flags & IO_URING_F_MULTISHOT)
977 return IOU_STOP_MULTISHOT;
981 int io_write(struct io_kiocb *req, unsigned int issue_flags)
983 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
984 struct io_rw_state __s, *s = &__s;
986 struct kiocb *kiocb = &rw->kiocb;
987 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
991 if (!req_has_async_data(req)) {
992 ret = io_import_iovec(ITER_SOURCE, req, &iovec, s, issue_flags);
993 if (unlikely(ret < 0))
996 struct io_async_rw *io = req->async_data;
999 iov_iter_restore(&s->iter, &s->iter_state);
1002 ret = io_rw_init_file(req, FMODE_WRITE);
1003 if (unlikely(ret)) {
1007 req->cqe.res = iov_iter_count(&s->iter);
1009 if (force_nonblock) {
1010 /* If the file doesn't support async, just async punt */
1011 if (unlikely(!io_file_supports_nowait(req)))
1014 /* File path supports NOWAIT for non-direct_IO only for block devices. */
1015 if (!(kiocb->ki_flags & IOCB_DIRECT) &&
1016 !(kiocb->ki_filp->f_mode & FMODE_BUF_WASYNC) &&
1017 (req->flags & REQ_F_ISREG))
1020 kiocb->ki_flags |= IOCB_NOWAIT;
1022 /* Ensure we clear previously set non-block flag */
1023 kiocb->ki_flags &= ~IOCB_NOWAIT;
1026 ppos = io_kiocb_update_pos(req);
1028 ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
1029 if (unlikely(ret)) {
1034 if (req->flags & REQ_F_ISREG)
1035 kiocb_start_write(kiocb);
1036 kiocb->ki_flags |= IOCB_WRITE;
1038 if (likely(req->file->f_op->write_iter))
1039 ret2 = call_write_iter(req->file, kiocb, &s->iter);
1040 else if (req->file->f_op->write)
1041 ret2 = loop_rw_iter(WRITE, rw, &s->iter);
1045 if (req->flags & REQ_F_REISSUE) {
1046 req->flags &= ~REQ_F_REISSUE;
1051 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
1052 * retry them without IOCB_NOWAIT.
1054 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
1056 /* no retry on NONBLOCK nor RWF_NOWAIT */
1057 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
1059 if (!force_nonblock || ret2 != -EAGAIN) {
1060 /* IOPOLL retry should happen for io-wq threads */
1061 if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
1064 if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) {
1065 struct io_async_rw *io;
1067 trace_io_uring_short_write(req->ctx, kiocb->ki_pos - ret2,
1068 req->cqe.res, ret2);
1070 /* This is a partial write. The file pos has already been
1071 * updated, setup the async struct to complete the request
1072 * in the worker. Also update bytes_done to account for
1073 * the bytes already written.
1075 iov_iter_save_state(&s->iter, &s->iter_state);
1076 ret = io_setup_async_rw(req, iovec, s, true);
1078 io = req->async_data;
1080 io->bytes_done += ret2;
1082 if (kiocb->ki_flags & IOCB_WRITE)
1083 io_req_end_write(req);
1084 return ret ? ret : -EAGAIN;
1087 ret = kiocb_done(req, ret2, issue_flags);
1090 iov_iter_restore(&s->iter, &s->iter_state);
1091 ret = io_setup_async_rw(req, iovec, s, false);
1093 if (kiocb->ki_flags & IOCB_WRITE)
1094 io_req_end_write(req);
1099 /* it's reportedly faster than delegating the null check to kfree() */
1105 void io_rw_fail(struct io_kiocb *req)
1109 res = io_fixup_rw_res(req, req->cqe.res);
1110 io_req_set_res(req, res, req->cqe.flags);
1113 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
1115 struct io_wq_work_node *pos, *start, *prev;
1116 unsigned int poll_flags = 0;
1117 DEFINE_IO_COMP_BATCH(iob);
1121 * Only spin for completions if we don't have multiple devices hanging
1122 * off our complete list.
1124 if (ctx->poll_multi_queue || force_nonspin)
1125 poll_flags |= BLK_POLL_ONESHOT;
1127 wq_list_for_each(pos, start, &ctx->iopoll_list) {
1128 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1129 struct file *file = req->file;
1133 * Move completed and retryable entries to our local lists.
1134 * If we find a request that requires polling, break out
1135 * and complete those lists first, if we have entries there.
1137 if (READ_ONCE(req->iopoll_completed))
1140 if (req->opcode == IORING_OP_URING_CMD) {
1141 struct io_uring_cmd *ioucmd;
1143 ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
1144 ret = file->f_op->uring_cmd_iopoll(ioucmd, &iob,
1147 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1149 ret = file->f_op->iopoll(&rw->kiocb, &iob, poll_flags);
1151 if (unlikely(ret < 0))
1154 poll_flags |= BLK_POLL_ONESHOT;
1156 /* iopoll may have completed current req */
1157 if (!rq_list_empty(iob.req_list) ||
1158 READ_ONCE(req->iopoll_completed))
1162 if (!rq_list_empty(iob.req_list))
1168 wq_list_for_each_resume(pos, prev) {
1169 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1171 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
1172 if (!smp_load_acquire(&req->iopoll_completed))
1175 req->cqe.flags = io_put_kbuf(req, 0);
1177 if (unlikely(!nr_events))
1180 pos = start ? start->next : ctx->iopoll_list.first;
1181 wq_list_cut(&ctx->iopoll_list, prev, start);
1183 if (WARN_ON_ONCE(!wq_list_empty(&ctx->submit_state.compl_reqs)))
1185 ctx->submit_state.compl_reqs.first = pos;
1186 __io_submit_flush_completions(ctx);