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 void io_rw_done(struct kiocb *kiocb, ssize_t ret)
177 case -ERESTARTNOINTR:
178 case -ERESTARTNOHAND:
179 case -ERESTART_RESTARTBLOCK:
181 * We can't just restart the syscall, since previously
182 * submitted sqes may already be in progress. Just fail this
188 kiocb->ki_complete(kiocb, ret);
192 static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req)
194 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
196 if (rw->kiocb.ki_pos != -1)
197 return &rw->kiocb.ki_pos;
199 if (!(req->file->f_mode & FMODE_STREAM)) {
200 req->flags |= REQ_F_CUR_POS;
201 rw->kiocb.ki_pos = req->file->f_pos;
202 return &rw->kiocb.ki_pos;
205 rw->kiocb.ki_pos = 0;
209 static void io_req_task_queue_reissue(struct io_kiocb *req)
211 req->io_task_work.func = io_queue_iowq;
212 io_req_task_work_add(req);
216 static bool io_resubmit_prep(struct io_kiocb *req)
218 struct io_async_rw *io = req->async_data;
220 if (!req_has_async_data(req))
221 return !io_req_prep_async(req);
222 iov_iter_restore(&io->s.iter, &io->s.iter_state);
226 static bool io_rw_should_reissue(struct io_kiocb *req)
228 umode_t mode = file_inode(req->file)->i_mode;
229 struct io_ring_ctx *ctx = req->ctx;
231 if (!S_ISBLK(mode) && !S_ISREG(mode))
233 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
234 !(ctx->flags & IORING_SETUP_IOPOLL)))
237 * If ref is dying, we might be running poll reap from the exit work.
238 * Don't attempt to reissue from that path, just let it fail with
241 if (percpu_ref_is_dying(&ctx->refs))
244 * Play it safe and assume not safe to re-import and reissue if we're
245 * not in the original thread group (or in task context).
247 if (!same_thread_group(req->task, current) || !in_task())
252 static bool io_resubmit_prep(struct io_kiocb *req)
256 static bool io_rw_should_reissue(struct io_kiocb *req)
262 static void io_req_end_write(struct io_kiocb *req)
264 if (req->flags & REQ_F_ISREG) {
265 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
267 kiocb_end_write(&rw->kiocb);
272 * Trigger the notifications after having done some IO, and finish the write
273 * accounting, if any.
275 static void io_req_io_end(struct io_kiocb *req)
277 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
279 if (rw->kiocb.ki_flags & IOCB_WRITE) {
280 io_req_end_write(req);
281 fsnotify_modify(req->file);
283 fsnotify_access(req->file);
287 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
289 if (unlikely(res != req->cqe.res)) {
290 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
291 io_rw_should_reissue(req)) {
293 * Reissue will start accounting again, finish the
297 req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
306 static inline int io_fixup_rw_res(struct io_kiocb *req, long res)
308 struct io_async_rw *io = req->async_data;
310 /* add previously done IO, if any */
311 if (req_has_async_data(req) && io->bytes_done > 0) {
313 res = io->bytes_done;
315 res += io->bytes_done;
320 void io_req_rw_complete(struct io_kiocb *req, struct io_tw_state *ts)
322 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
323 struct kiocb *kiocb = &rw->kiocb;
325 if ((kiocb->ki_flags & IOCB_DIO_CALLER_COMP) && kiocb->dio_complete) {
326 long res = kiocb->dio_complete(rw->kiocb.private);
328 io_req_set_res(req, io_fixup_rw_res(req, res), 0);
333 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
334 unsigned issue_flags = ts->locked ? 0 : IO_URING_F_UNLOCKED;
336 req->cqe.flags |= io_put_kbuf(req, issue_flags);
338 io_req_task_complete(req, ts);
341 static void io_complete_rw(struct kiocb *kiocb, long res)
343 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
344 struct io_kiocb *req = cmd_to_io_kiocb(rw);
346 if (!kiocb->dio_complete || !(kiocb->ki_flags & IOCB_DIO_CALLER_COMP)) {
347 if (__io_complete_rw_common(req, res))
349 io_req_set_res(req, io_fixup_rw_res(req, res), 0);
351 req->io_task_work.func = io_req_rw_complete;
352 __io_req_task_work_add(req, IOU_F_TWQ_LAZY_WAKE);
355 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
357 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
358 struct io_kiocb *req = cmd_to_io_kiocb(rw);
360 if (kiocb->ki_flags & IOCB_WRITE)
361 io_req_end_write(req);
362 if (unlikely(res != req->cqe.res)) {
363 if (res == -EAGAIN && io_rw_should_reissue(req)) {
364 req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
370 /* order with io_iopoll_complete() checking ->iopoll_completed */
371 smp_store_release(&req->iopoll_completed, 1);
374 static int kiocb_done(struct io_kiocb *req, ssize_t ret,
375 unsigned int issue_flags)
377 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
378 unsigned final_ret = io_fixup_rw_res(req, ret);
380 if (ret >= 0 && req->flags & REQ_F_CUR_POS)
381 req->file->f_pos = rw->kiocb.ki_pos;
382 if (ret >= 0 && (rw->kiocb.ki_complete == io_complete_rw)) {
383 if (!__io_complete_rw_common(req, ret)) {
385 * Safe to call io_end from here as we're inline
386 * from the submission path.
389 io_req_set_res(req, final_ret,
390 io_put_kbuf(req, issue_flags));
394 io_rw_done(&rw->kiocb, ret);
397 if (req->flags & REQ_F_REISSUE) {
398 req->flags &= ~REQ_F_REISSUE;
399 if (io_resubmit_prep(req))
400 io_req_task_queue_reissue(req);
402 io_req_task_queue_fail(req, final_ret);
404 return IOU_ISSUE_SKIP_COMPLETE;
407 static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req,
408 struct io_rw_state *s,
409 unsigned int issue_flags)
411 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
412 struct iov_iter *iter = &s->iter;
413 u8 opcode = req->opcode;
419 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
420 ret = io_import_fixed(ddir, iter, req->imu, rw->addr, rw->len);
426 buf = u64_to_user_ptr(rw->addr);
429 if (!io_issue_defs[opcode].vectored || req->flags & REQ_F_BUFFER_SELECT) {
430 if (io_do_buffer_select(req)) {
431 buf = io_buffer_select(req, &sqe_len, issue_flags);
433 return ERR_PTR(-ENOBUFS);
434 rw->addr = (unsigned long) buf;
438 ret = import_ubuf(ddir, buf, sqe_len, iter);
445 ret = __import_iovec(ddir, buf, sqe_len, UIO_FASTIOV, &iovec, iter,
447 if (unlikely(ret < 0))
452 static inline int io_import_iovec(int rw, struct io_kiocb *req,
453 struct iovec **iovec, struct io_rw_state *s,
454 unsigned int issue_flags)
456 *iovec = __io_import_iovec(rw, req, s, issue_flags);
458 return PTR_ERR(*iovec);
460 iov_iter_save_state(&s->iter, &s->iter_state);
464 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
466 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
470 * For files that don't have ->read_iter() and ->write_iter(), handle them
471 * by looping over ->read() or ->write() manually.
473 static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter)
475 struct kiocb *kiocb = &rw->kiocb;
476 struct file *file = kiocb->ki_filp;
481 * Don't support polled IO through this interface, and we can't
482 * support non-blocking either. For the latter, this just causes
483 * the kiocb to be handled from an async context.
485 if (kiocb->ki_flags & IOCB_HIPRI)
487 if ((kiocb->ki_flags & IOCB_NOWAIT) &&
488 !(kiocb->ki_filp->f_flags & O_NONBLOCK))
491 ppos = io_kiocb_ppos(kiocb);
493 while (iov_iter_count(iter)) {
498 if (iter_is_ubuf(iter)) {
499 addr = iter->ubuf + iter->iov_offset;
500 len = iov_iter_count(iter);
501 } else if (!iov_iter_is_bvec(iter)) {
502 addr = iter_iov_addr(iter);
503 len = iter_iov_len(iter);
505 addr = u64_to_user_ptr(rw->addr);
510 nr = file->f_op->read(file, addr, len, ppos);
512 nr = file->f_op->write(file, addr, len, ppos);
520 if (!iov_iter_is_bvec(iter)) {
521 iov_iter_advance(iter, nr);
535 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
536 const struct iovec *fast_iov, struct iov_iter *iter)
538 struct io_async_rw *io = req->async_data;
540 memcpy(&io->s.iter, iter, sizeof(*iter));
541 io->free_iovec = iovec;
543 /* can only be fixed buffers, no need to do anything */
544 if (iov_iter_is_bvec(iter) || iter_is_ubuf(iter))
547 unsigned iov_off = 0;
549 io->s.iter.__iov = io->s.fast_iov;
550 if (iter->__iov != fast_iov) {
551 iov_off = iter_iov(iter) - fast_iov;
552 io->s.iter.__iov += iov_off;
554 if (io->s.fast_iov != fast_iov)
555 memcpy(io->s.fast_iov + iov_off, fast_iov + iov_off,
556 sizeof(struct iovec) * iter->nr_segs);
558 req->flags |= REQ_F_NEED_CLEANUP;
562 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
563 struct io_rw_state *s, bool force)
565 if (!force && !io_cold_defs[req->opcode].prep_async)
567 /* opcode type doesn't need async data */
568 if (!io_cold_defs[req->opcode].async_size)
570 if (!req_has_async_data(req)) {
571 struct io_async_rw *iorw;
573 if (io_alloc_async_data(req)) {
578 io_req_map_rw(req, iovec, s->fast_iov, &s->iter);
579 iorw = req->async_data;
580 /* we've copied and mapped the iter, ensure state is saved */
581 iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state);
586 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
588 struct io_async_rw *iorw = req->async_data;
592 iorw->bytes_done = 0;
593 iorw->free_iovec = NULL;
595 /* submission path, ->uring_lock should already be taken */
596 ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
597 if (unlikely(ret < 0))
601 iorw->free_iovec = iov;
602 req->flags |= REQ_F_NEED_CLEANUP;
608 int io_readv_prep_async(struct io_kiocb *req)
610 return io_rw_prep_async(req, ITER_DEST);
613 int io_writev_prep_async(struct io_kiocb *req)
615 return io_rw_prep_async(req, ITER_SOURCE);
619 * This is our waitqueue callback handler, registered through __folio_lock_async()
620 * when we initially tried to do the IO with the iocb armed our waitqueue.
621 * This gets called when the page is unlocked, and we generally expect that to
622 * happen when the page IO is completed and the page is now uptodate. This will
623 * queue a task_work based retry of the operation, attempting to copy the data
624 * again. If the latter fails because the page was NOT uptodate, then we will
625 * do a thread based blocking retry of the operation. That's the unexpected
628 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
631 struct wait_page_queue *wpq;
632 struct io_kiocb *req = wait->private;
633 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
634 struct wait_page_key *key = arg;
636 wpq = container_of(wait, struct wait_page_queue, wait);
638 if (!wake_page_match(wpq, key))
641 rw->kiocb.ki_flags &= ~IOCB_WAITQ;
642 list_del_init(&wait->entry);
643 io_req_task_queue(req);
648 * This controls whether a given IO request should be armed for async page
649 * based retry. If we return false here, the request is handed to the async
650 * worker threads for retry. If we're doing buffered reads on a regular file,
651 * we prepare a private wait_page_queue entry and retry the operation. This
652 * will either succeed because the page is now uptodate and unlocked, or it
653 * will register a callback when the page is unlocked at IO completion. Through
654 * that callback, io_uring uses task_work to setup a retry of the operation.
655 * That retry will attempt the buffered read again. The retry will generally
656 * succeed, or in rare cases where it fails, we then fall back to using the
657 * async worker threads for a blocking retry.
659 static bool io_rw_should_retry(struct io_kiocb *req)
661 struct io_async_rw *io = req->async_data;
662 struct wait_page_queue *wait = &io->wpq;
663 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
664 struct kiocb *kiocb = &rw->kiocb;
666 /* never retry for NOWAIT, we just complete with -EAGAIN */
667 if (req->flags & REQ_F_NOWAIT)
670 /* Only for buffered IO */
671 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
675 * just use poll if we can, and don't attempt if the fs doesn't
676 * support callback based unlocks
678 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
681 wait->wait.func = io_async_buf_func;
682 wait->wait.private = req;
683 wait->wait.flags = 0;
684 INIT_LIST_HEAD(&wait->wait.entry);
685 kiocb->ki_flags |= IOCB_WAITQ;
686 kiocb->ki_flags &= ~IOCB_NOWAIT;
687 kiocb->ki_waitq = wait;
691 static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
693 struct file *file = rw->kiocb.ki_filp;
695 if (likely(file->f_op->read_iter))
696 return call_read_iter(file, &rw->kiocb, iter);
697 else if (file->f_op->read)
698 return loop_rw_iter(READ, rw, iter);
703 static bool need_complete_io(struct io_kiocb *req)
705 return req->flags & REQ_F_ISREG ||
706 S_ISBLK(file_inode(req->file)->i_mode);
709 static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
711 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
712 struct kiocb *kiocb = &rw->kiocb;
713 struct io_ring_ctx *ctx = req->ctx;
714 struct file *file = req->file;
717 if (unlikely(!file || !(file->f_mode & mode)))
720 if (!(req->flags & REQ_F_FIXED_FILE))
721 req->flags |= io_file_get_flags(file);
723 kiocb->ki_flags = file->f_iocb_flags;
724 ret = kiocb_set_rw_flags(kiocb, rw->flags);
727 kiocb->ki_flags |= IOCB_ALLOC_CACHE;
730 * If the file is marked O_NONBLOCK, still allow retry for it if it
731 * supports async. Otherwise it's impossible to use O_NONBLOCK files
732 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
734 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
735 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
736 req->flags |= REQ_F_NOWAIT;
738 if (ctx->flags & IORING_SETUP_IOPOLL) {
739 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
742 kiocb->private = NULL;
743 kiocb->ki_flags |= IOCB_HIPRI;
744 kiocb->ki_complete = io_complete_rw_iopoll;
745 req->iopoll_completed = 0;
747 if (kiocb->ki_flags & IOCB_HIPRI)
749 kiocb->ki_complete = io_complete_rw;
755 static int __io_read(struct io_kiocb *req, unsigned int issue_flags)
757 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
758 struct io_rw_state __s, *s = &__s;
760 struct kiocb *kiocb = &rw->kiocb;
761 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
762 struct io_async_rw *io;
766 if (!req_has_async_data(req)) {
767 ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
768 if (unlikely(ret < 0))
771 io = req->async_data;
775 * Safe and required to re-import if we're using provided
776 * buffers, as we dropped the selected one before retry.
778 if (io_do_buffer_select(req)) {
779 ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
780 if (unlikely(ret < 0))
785 * We come here from an earlier attempt, restore our state to
786 * match in case it doesn't. It's cheap enough that we don't
787 * need to make this conditional.
789 iov_iter_restore(&s->iter, &s->iter_state);
792 ret = io_rw_init_file(req, FMODE_READ);
797 req->cqe.res = iov_iter_count(&s->iter);
799 if (force_nonblock) {
800 /* If the file doesn't support async, just async punt */
801 if (unlikely(!io_file_supports_nowait(req))) {
802 ret = io_setup_async_rw(req, iovec, s, true);
803 return ret ?: -EAGAIN;
805 kiocb->ki_flags |= IOCB_NOWAIT;
807 /* Ensure we clear previously set non-block flag */
808 kiocb->ki_flags &= ~IOCB_NOWAIT;
811 ppos = io_kiocb_update_pos(req);
813 ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
819 ret = io_iter_do_read(rw, &s->iter);
821 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
822 req->flags &= ~REQ_F_REISSUE;
824 * If we can poll, just do that. For a vectored read, we'll
825 * need to copy state first.
827 if (file_can_poll(req->file) && !io_issue_defs[req->opcode].vectored)
829 /* IOPOLL retry should happen for io-wq threads */
830 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
832 /* no retry on NONBLOCK nor RWF_NOWAIT */
833 if (req->flags & REQ_F_NOWAIT)
836 } else if (ret == -EIOCBQUEUED) {
839 return IOU_ISSUE_SKIP_COMPLETE;
840 } else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
841 (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) {
842 /* read all, failed, already did sync or don't want to retry */
847 * Don't depend on the iter state matching what was consumed, or being
848 * untouched in case of error. Restore it and we'll advance it
849 * manually if we need to.
851 iov_iter_restore(&s->iter, &s->iter_state);
853 ret2 = io_setup_async_rw(req, iovec, s, true);
856 ret = ret > 0 ? ret : ret2;
860 io = req->async_data;
863 * Now use our persistent iterator and state, if we aren't already.
864 * We've restored and mapped the iter to match.
869 * We end up here because of a partial read, either from
870 * above or inside this loop. Advance the iter by the bytes
871 * that were consumed.
873 iov_iter_advance(&s->iter, ret);
874 if (!iov_iter_count(&s->iter))
876 io->bytes_done += ret;
877 iov_iter_save_state(&s->iter, &s->iter_state);
879 /* if we can retry, do so with the callbacks armed */
880 if (!io_rw_should_retry(req)) {
881 kiocb->ki_flags &= ~IOCB_WAITQ;
885 req->cqe.res = iov_iter_count(&s->iter);
887 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
888 * we get -EIOCBQUEUED, then we'll get a notification when the
889 * desired page gets unlocked. We can also get a partial read
890 * here, and if we do, then just retry at the new offset.
892 ret = io_iter_do_read(rw, &s->iter);
893 if (ret == -EIOCBQUEUED)
894 return IOU_ISSUE_SKIP_COMPLETE;
895 /* we got some bytes, but not all. retry. */
896 kiocb->ki_flags &= ~IOCB_WAITQ;
897 iov_iter_restore(&s->iter, &s->iter_state);
900 /* it's faster to check here then delegate to kfree */
906 int io_read(struct io_kiocb *req, unsigned int issue_flags)
910 ret = __io_read(req, issue_flags);
912 return kiocb_done(req, ret, issue_flags);
917 int io_read_mshot(struct io_kiocb *req, unsigned int issue_flags)
919 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
920 unsigned int cflags = 0;
924 * Multishot MUST be used on a pollable file
926 if (!file_can_poll(req->file))
929 ret = __io_read(req, issue_flags);
932 * If we get -EAGAIN, recycle our buffer and just let normal poll
935 if (ret == -EAGAIN) {
937 * Reset rw->len to 0 again to avoid clamping future mshot
938 * reads, in case the buffer size varies.
940 if (io_kbuf_recycle(req, issue_flags))
946 * Any successful return value will keep the multishot read armed.
950 * Put our buffer and post a CQE. If we fail to post a CQE, then
951 * jump to the termination path. This request is then done.
953 cflags = io_put_kbuf(req, issue_flags);
954 rw->len = 0; /* similarly to above, reset len to 0 */
956 if (io_fill_cqe_req_aux(req,
957 issue_flags & IO_URING_F_COMPLETE_DEFER,
958 ret, cflags | IORING_CQE_F_MORE)) {
959 if (issue_flags & IO_URING_F_MULTISHOT)
960 return IOU_ISSUE_SKIP_COMPLETE;
966 * Either an error, or we've hit overflow posting the CQE. For any
967 * multishot request, hitting overflow will terminate it.
969 io_req_set_res(req, ret, cflags);
970 if (issue_flags & IO_URING_F_MULTISHOT)
971 return IOU_STOP_MULTISHOT;
975 int io_write(struct io_kiocb *req, unsigned int issue_flags)
977 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
978 struct io_rw_state __s, *s = &__s;
980 struct kiocb *kiocb = &rw->kiocb;
981 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
985 if (!req_has_async_data(req)) {
986 ret = io_import_iovec(ITER_SOURCE, req, &iovec, s, issue_flags);
987 if (unlikely(ret < 0))
990 struct io_async_rw *io = req->async_data;
993 iov_iter_restore(&s->iter, &s->iter_state);
996 ret = io_rw_init_file(req, FMODE_WRITE);
1001 req->cqe.res = iov_iter_count(&s->iter);
1003 if (force_nonblock) {
1004 /* If the file doesn't support async, just async punt */
1005 if (unlikely(!io_file_supports_nowait(req)))
1008 /* File path supports NOWAIT for non-direct_IO only for block devices. */
1009 if (!(kiocb->ki_flags & IOCB_DIRECT) &&
1010 !(kiocb->ki_filp->f_mode & FMODE_BUF_WASYNC) &&
1011 (req->flags & REQ_F_ISREG))
1014 kiocb->ki_flags |= IOCB_NOWAIT;
1016 /* Ensure we clear previously set non-block flag */
1017 kiocb->ki_flags &= ~IOCB_NOWAIT;
1020 ppos = io_kiocb_update_pos(req);
1022 ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
1023 if (unlikely(ret)) {
1028 if (req->flags & REQ_F_ISREG)
1029 kiocb_start_write(kiocb);
1030 kiocb->ki_flags |= IOCB_WRITE;
1032 if (likely(req->file->f_op->write_iter))
1033 ret2 = call_write_iter(req->file, kiocb, &s->iter);
1034 else if (req->file->f_op->write)
1035 ret2 = loop_rw_iter(WRITE, rw, &s->iter);
1039 if (req->flags & REQ_F_REISSUE) {
1040 req->flags &= ~REQ_F_REISSUE;
1045 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
1046 * retry them without IOCB_NOWAIT.
1048 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
1050 /* no retry on NONBLOCK nor RWF_NOWAIT */
1051 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
1053 if (!force_nonblock || ret2 != -EAGAIN) {
1054 /* IOPOLL retry should happen for io-wq threads */
1055 if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
1058 if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) {
1059 struct io_async_rw *io;
1061 trace_io_uring_short_write(req->ctx, kiocb->ki_pos - ret2,
1062 req->cqe.res, ret2);
1064 /* This is a partial write. The file pos has already been
1065 * updated, setup the async struct to complete the request
1066 * in the worker. Also update bytes_done to account for
1067 * the bytes already written.
1069 iov_iter_save_state(&s->iter, &s->iter_state);
1070 ret = io_setup_async_rw(req, iovec, s, true);
1072 io = req->async_data;
1074 io->bytes_done += ret2;
1076 if (kiocb->ki_flags & IOCB_WRITE)
1077 io_req_end_write(req);
1078 return ret ? ret : -EAGAIN;
1081 ret = kiocb_done(req, ret2, issue_flags);
1084 iov_iter_restore(&s->iter, &s->iter_state);
1085 ret = io_setup_async_rw(req, iovec, s, false);
1087 if (kiocb->ki_flags & IOCB_WRITE)
1088 io_req_end_write(req);
1093 /* it's reportedly faster than delegating the null check to kfree() */
1099 void io_rw_fail(struct io_kiocb *req)
1103 res = io_fixup_rw_res(req, req->cqe.res);
1104 io_req_set_res(req, res, req->cqe.flags);
1107 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
1109 struct io_wq_work_node *pos, *start, *prev;
1110 unsigned int poll_flags = 0;
1111 DEFINE_IO_COMP_BATCH(iob);
1115 * Only spin for completions if we don't have multiple devices hanging
1116 * off our complete list.
1118 if (ctx->poll_multi_queue || force_nonspin)
1119 poll_flags |= BLK_POLL_ONESHOT;
1121 wq_list_for_each(pos, start, &ctx->iopoll_list) {
1122 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1123 struct file *file = req->file;
1127 * Move completed and retryable entries to our local lists.
1128 * If we find a request that requires polling, break out
1129 * and complete those lists first, if we have entries there.
1131 if (READ_ONCE(req->iopoll_completed))
1134 if (req->opcode == IORING_OP_URING_CMD) {
1135 struct io_uring_cmd *ioucmd;
1137 ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
1138 ret = file->f_op->uring_cmd_iopoll(ioucmd, &iob,
1141 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1143 ret = file->f_op->iopoll(&rw->kiocb, &iob, poll_flags);
1145 if (unlikely(ret < 0))
1148 poll_flags |= BLK_POLL_ONESHOT;
1150 /* iopoll may have completed current req */
1151 if (!rq_list_empty(iob.req_list) ||
1152 READ_ONCE(req->iopoll_completed))
1156 if (!rq_list_empty(iob.req_list))
1162 wq_list_for_each_resume(pos, prev) {
1163 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1165 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
1166 if (!smp_load_acquire(&req->iopoll_completed))
1169 req->cqe.flags = io_put_kbuf(req, 0);
1171 if (unlikely(!nr_events))
1174 pos = start ? start->next : ctx->iopoll_list.first;
1175 wq_list_cut(&ctx->iopoll_list, prev, start);
1177 if (WARN_ON_ONCE(!wq_list_empty(&ctx->submit_state.compl_reqs)))
1179 ctx->submit_state.compl_reqs.first = pos;
1180 __io_submit_flush_completions(ctx);