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.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 if (req->opcode == IORING_OP_READ_FIXED ||
87 req->opcode == IORING_OP_WRITE_FIXED) {
88 struct io_ring_ctx *ctx = req->ctx;
91 if (unlikely(req->buf_index >= ctx->nr_user_bufs))
93 index = array_index_nospec(req->buf_index, ctx->nr_user_bufs);
94 req->imu = ctx->user_bufs[index];
95 io_req_set_rsrc_node(req, ctx, 0);
98 ioprio = READ_ONCE(sqe->ioprio);
100 ret = ioprio_check_cap(ioprio);
104 rw->kiocb.ki_ioprio = ioprio;
106 rw->kiocb.ki_ioprio = get_current_ioprio();
109 rw->addr = READ_ONCE(sqe->addr);
110 rw->len = READ_ONCE(sqe->len);
111 rw->flags = READ_ONCE(sqe->rw_flags);
113 /* Have to do this validation here, as this is in io_read() rw->len might
114 * have chanaged due to buffer selection
116 if (req->opcode == IORING_OP_READV && req->flags & REQ_F_BUFFER_SELECT) {
117 ret = io_iov_buffer_select_prep(req);
125 void io_readv_writev_cleanup(struct io_kiocb *req)
127 struct io_async_rw *io = req->async_data;
129 kfree(io->free_iovec);
132 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
138 case -ERESTARTNOINTR:
139 case -ERESTARTNOHAND:
140 case -ERESTART_RESTARTBLOCK:
142 * We can't just restart the syscall, since previously
143 * submitted sqes may already be in progress. Just fail this
149 kiocb->ki_complete(kiocb, ret);
153 static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req)
155 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
157 if (rw->kiocb.ki_pos != -1)
158 return &rw->kiocb.ki_pos;
160 if (!(req->file->f_mode & FMODE_STREAM)) {
161 req->flags |= REQ_F_CUR_POS;
162 rw->kiocb.ki_pos = req->file->f_pos;
163 return &rw->kiocb.ki_pos;
166 rw->kiocb.ki_pos = 0;
170 static void io_req_task_queue_reissue(struct io_kiocb *req)
172 req->io_task_work.func = io_queue_iowq;
173 io_req_task_work_add(req);
177 static bool io_resubmit_prep(struct io_kiocb *req)
179 struct io_async_rw *io = req->async_data;
181 if (!req_has_async_data(req))
182 return !io_req_prep_async(req);
183 iov_iter_restore(&io->s.iter, &io->s.iter_state);
187 static bool io_rw_should_reissue(struct io_kiocb *req)
189 umode_t mode = file_inode(req->file)->i_mode;
190 struct io_ring_ctx *ctx = req->ctx;
192 if (!S_ISBLK(mode) && !S_ISREG(mode))
194 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
195 !(ctx->flags & IORING_SETUP_IOPOLL)))
198 * If ref is dying, we might be running poll reap from the exit work.
199 * Don't attempt to reissue from that path, just let it fail with
202 if (percpu_ref_is_dying(&ctx->refs))
205 * Play it safe and assume not safe to re-import and reissue if we're
206 * not in the original thread group (or in task context).
208 if (!same_thread_group(req->task, current) || !in_task())
213 static bool io_resubmit_prep(struct io_kiocb *req)
217 static bool io_rw_should_reissue(struct io_kiocb *req)
223 static void kiocb_end_write(struct io_kiocb *req)
226 * Tell lockdep we inherited freeze protection from submission
229 if (req->flags & REQ_F_ISREG) {
230 struct super_block *sb = file_inode(req->file)->i_sb;
232 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
238 * Trigger the notifications after having done some IO, and finish the write
239 * accounting, if any.
241 static void io_req_io_end(struct io_kiocb *req)
243 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
245 if (rw->kiocb.ki_flags & IOCB_WRITE) {
246 kiocb_end_write(req);
247 fsnotify_modify(req->file);
249 fsnotify_access(req->file);
253 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
255 if (unlikely(res != req->cqe.res)) {
256 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
257 io_rw_should_reissue(req)) {
259 * Reissue will start accounting again, finish the
263 req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
272 static inline int io_fixup_rw_res(struct io_kiocb *req, long res)
274 struct io_async_rw *io = req->async_data;
276 /* add previously done IO, if any */
277 if (req_has_async_data(req) && io->bytes_done > 0) {
279 res = io->bytes_done;
281 res += io->bytes_done;
286 static void io_req_rw_complete(struct io_kiocb *req, bool *locked)
289 io_req_task_complete(req, locked);
292 static void io_complete_rw(struct kiocb *kiocb, long res)
294 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
295 struct io_kiocb *req = cmd_to_io_kiocb(rw);
297 if (__io_complete_rw_common(req, res))
299 io_req_set_res(req, io_fixup_rw_res(req, res), 0);
300 req->io_task_work.func = io_req_rw_complete;
301 io_req_task_work_add(req);
304 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
306 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
307 struct io_kiocb *req = cmd_to_io_kiocb(rw);
309 if (kiocb->ki_flags & IOCB_WRITE)
310 kiocb_end_write(req);
311 if (unlikely(res != req->cqe.res)) {
312 if (res == -EAGAIN && io_rw_should_reissue(req)) {
313 req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
319 /* order with io_iopoll_complete() checking ->iopoll_completed */
320 smp_store_release(&req->iopoll_completed, 1);
323 static int kiocb_done(struct io_kiocb *req, ssize_t ret,
324 unsigned int issue_flags)
326 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
327 unsigned final_ret = io_fixup_rw_res(req, ret);
329 if (ret >= 0 && req->flags & REQ_F_CUR_POS)
330 req->file->f_pos = rw->kiocb.ki_pos;
331 if (ret >= 0 && (rw->kiocb.ki_complete == io_complete_rw)) {
332 if (!__io_complete_rw_common(req, ret)) {
334 * Safe to call io_end from here as we're inline
335 * from the submission path.
338 io_req_set_res(req, final_ret,
339 io_put_kbuf(req, issue_flags));
343 io_rw_done(&rw->kiocb, ret);
346 if (req->flags & REQ_F_REISSUE) {
347 req->flags &= ~REQ_F_REISSUE;
348 if (io_resubmit_prep(req))
349 io_req_task_queue_reissue(req);
351 io_req_task_queue_fail(req, final_ret);
353 return IOU_ISSUE_SKIP_COMPLETE;
356 static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req,
357 struct io_rw_state *s,
358 unsigned int issue_flags)
360 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
361 struct iov_iter *iter = &s->iter;
362 u8 opcode = req->opcode;
368 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
369 ret = io_import_fixed(ddir, iter, req->imu, rw->addr, rw->len);
375 buf = u64_to_user_ptr(rw->addr);
378 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE ||
379 (req->flags & REQ_F_BUFFER_SELECT)) {
380 if (io_do_buffer_select(req)) {
381 buf = io_buffer_select(req, &sqe_len, issue_flags);
383 return ERR_PTR(-ENOBUFS);
384 rw->addr = (unsigned long) buf;
388 ret = import_single_range(ddir, buf, sqe_len, s->fast_iov, iter);
395 ret = __import_iovec(ddir, buf, sqe_len, UIO_FASTIOV, &iovec, iter,
397 if (unlikely(ret < 0))
402 static inline int io_import_iovec(int rw, struct io_kiocb *req,
403 struct iovec **iovec, struct io_rw_state *s,
404 unsigned int issue_flags)
406 *iovec = __io_import_iovec(rw, req, s, issue_flags);
407 if (unlikely(IS_ERR(*iovec)))
408 return PTR_ERR(*iovec);
410 iov_iter_save_state(&s->iter, &s->iter_state);
414 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
416 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
420 * For files that don't have ->read_iter() and ->write_iter(), handle them
421 * by looping over ->read() or ->write() manually.
423 static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter)
425 struct kiocb *kiocb = &rw->kiocb;
426 struct file *file = kiocb->ki_filp;
431 * Don't support polled IO through this interface, and we can't
432 * support non-blocking either. For the latter, this just causes
433 * the kiocb to be handled from an async context.
435 if (kiocb->ki_flags & IOCB_HIPRI)
437 if ((kiocb->ki_flags & IOCB_NOWAIT) &&
438 !(kiocb->ki_filp->f_flags & O_NONBLOCK))
441 ppos = io_kiocb_ppos(kiocb);
443 while (iov_iter_count(iter)) {
447 if (!iov_iter_is_bvec(iter)) {
448 iovec = iov_iter_iovec(iter);
450 iovec.iov_base = u64_to_user_ptr(rw->addr);
451 iovec.iov_len = rw->len;
455 nr = file->f_op->read(file, iovec.iov_base,
456 iovec.iov_len, ppos);
458 nr = file->f_op->write(file, iovec.iov_base,
459 iovec.iov_len, ppos);
468 if (!iov_iter_is_bvec(iter)) {
469 iov_iter_advance(iter, nr);
476 if (nr != iovec.iov_len)
483 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
484 const struct iovec *fast_iov, struct iov_iter *iter)
486 struct io_async_rw *io = req->async_data;
488 memcpy(&io->s.iter, iter, sizeof(*iter));
489 io->free_iovec = iovec;
491 /* can only be fixed buffers, no need to do anything */
492 if (iov_iter_is_bvec(iter))
495 unsigned iov_off = 0;
497 io->s.iter.iov = io->s.fast_iov;
498 if (iter->iov != fast_iov) {
499 iov_off = iter->iov - fast_iov;
500 io->s.iter.iov += iov_off;
502 if (io->s.fast_iov != fast_iov)
503 memcpy(io->s.fast_iov + iov_off, fast_iov + iov_off,
504 sizeof(struct iovec) * iter->nr_segs);
506 req->flags |= REQ_F_NEED_CLEANUP;
510 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
511 struct io_rw_state *s, bool force)
513 if (!force && !io_op_defs[req->opcode].prep_async)
515 if (!req_has_async_data(req)) {
516 struct io_async_rw *iorw;
518 if (io_alloc_async_data(req)) {
523 io_req_map_rw(req, iovec, s->fast_iov, &s->iter);
524 iorw = req->async_data;
525 /* we've copied and mapped the iter, ensure state is saved */
526 iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state);
531 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
533 struct io_async_rw *iorw = req->async_data;
537 iorw->bytes_done = 0;
538 iorw->free_iovec = NULL;
540 /* submission path, ->uring_lock should already be taken */
541 ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
542 if (unlikely(ret < 0))
546 iorw->free_iovec = iov;
547 req->flags |= REQ_F_NEED_CLEANUP;
553 int io_readv_prep_async(struct io_kiocb *req)
555 return io_rw_prep_async(req, ITER_DEST);
558 int io_writev_prep_async(struct io_kiocb *req)
560 return io_rw_prep_async(req, ITER_SOURCE);
564 * This is our waitqueue callback handler, registered through __folio_lock_async()
565 * when we initially tried to do the IO with the iocb armed our waitqueue.
566 * This gets called when the page is unlocked, and we generally expect that to
567 * happen when the page IO is completed and the page is now uptodate. This will
568 * queue a task_work based retry of the operation, attempting to copy the data
569 * again. If the latter fails because the page was NOT uptodate, then we will
570 * do a thread based blocking retry of the operation. That's the unexpected
573 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
576 struct wait_page_queue *wpq;
577 struct io_kiocb *req = wait->private;
578 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
579 struct wait_page_key *key = arg;
581 wpq = container_of(wait, struct wait_page_queue, wait);
583 if (!wake_page_match(wpq, key))
586 rw->kiocb.ki_flags &= ~IOCB_WAITQ;
587 list_del_init(&wait->entry);
588 io_req_task_queue(req);
593 * This controls whether a given IO request should be armed for async page
594 * based retry. If we return false here, the request is handed to the async
595 * worker threads for retry. If we're doing buffered reads on a regular file,
596 * we prepare a private wait_page_queue entry and retry the operation. This
597 * will either succeed because the page is now uptodate and unlocked, or it
598 * will register a callback when the page is unlocked at IO completion. Through
599 * that callback, io_uring uses task_work to setup a retry of the operation.
600 * That retry will attempt the buffered read again. The retry will generally
601 * succeed, or in rare cases where it fails, we then fall back to using the
602 * async worker threads for a blocking retry.
604 static bool io_rw_should_retry(struct io_kiocb *req)
606 struct io_async_rw *io = req->async_data;
607 struct wait_page_queue *wait = &io->wpq;
608 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
609 struct kiocb *kiocb = &rw->kiocb;
611 /* never retry for NOWAIT, we just complete with -EAGAIN */
612 if (req->flags & REQ_F_NOWAIT)
615 /* Only for buffered IO */
616 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
620 * just use poll if we can, and don't attempt if the fs doesn't
621 * support callback based unlocks
623 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
626 wait->wait.func = io_async_buf_func;
627 wait->wait.private = req;
628 wait->wait.flags = 0;
629 INIT_LIST_HEAD(&wait->wait.entry);
630 kiocb->ki_flags |= IOCB_WAITQ;
631 kiocb->ki_flags &= ~IOCB_NOWAIT;
632 kiocb->ki_waitq = wait;
636 static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
638 struct file *file = rw->kiocb.ki_filp;
640 if (likely(file->f_op->read_iter))
641 return call_read_iter(file, &rw->kiocb, iter);
642 else if (file->f_op->read)
643 return loop_rw_iter(READ, rw, iter);
648 static bool need_complete_io(struct io_kiocb *req)
650 return req->flags & REQ_F_ISREG ||
651 S_ISBLK(file_inode(req->file)->i_mode);
654 static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
656 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
657 struct kiocb *kiocb = &rw->kiocb;
658 struct io_ring_ctx *ctx = req->ctx;
659 struct file *file = req->file;
662 if (unlikely(!file || !(file->f_mode & mode)))
665 if (!io_req_ffs_set(req))
666 req->flags |= io_file_get_flags(file) << REQ_F_SUPPORT_NOWAIT_BIT;
668 kiocb->ki_flags = file->f_iocb_flags;
669 ret = kiocb_set_rw_flags(kiocb, rw->flags);
674 * If the file is marked O_NONBLOCK, still allow retry for it if it
675 * supports async. Otherwise it's impossible to use O_NONBLOCK files
676 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
678 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
679 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
680 req->flags |= REQ_F_NOWAIT;
682 if (ctx->flags & IORING_SETUP_IOPOLL) {
683 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
686 kiocb->private = NULL;
687 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
688 kiocb->ki_complete = io_complete_rw_iopoll;
689 req->iopoll_completed = 0;
691 if (kiocb->ki_flags & IOCB_HIPRI)
693 kiocb->ki_complete = io_complete_rw;
699 int io_read(struct io_kiocb *req, unsigned int issue_flags)
701 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
702 struct io_rw_state __s, *s = &__s;
704 struct kiocb *kiocb = &rw->kiocb;
705 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
706 struct io_async_rw *io;
710 if (!req_has_async_data(req)) {
711 ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
712 if (unlikely(ret < 0))
715 io = req->async_data;
719 * Safe and required to re-import if we're using provided
720 * buffers, as we dropped the selected one before retry.
722 if (io_do_buffer_select(req)) {
723 ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
724 if (unlikely(ret < 0))
729 * We come here from an earlier attempt, restore our state to
730 * match in case it doesn't. It's cheap enough that we don't
731 * need to make this conditional.
733 iov_iter_restore(&s->iter, &s->iter_state);
736 ret = io_rw_init_file(req, FMODE_READ);
741 req->cqe.res = iov_iter_count(&s->iter);
743 if (force_nonblock) {
744 /* If the file doesn't support async, just async punt */
745 if (unlikely(!io_file_supports_nowait(req))) {
746 ret = io_setup_async_rw(req, iovec, s, true);
747 return ret ?: -EAGAIN;
749 kiocb->ki_flags |= IOCB_NOWAIT;
751 /* Ensure we clear previously set non-block flag */
752 kiocb->ki_flags &= ~IOCB_NOWAIT;
755 ppos = io_kiocb_update_pos(req);
757 ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
763 ret = io_iter_do_read(rw, &s->iter);
765 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
766 req->flags &= ~REQ_F_REISSUE;
767 /* if we can poll, just do that */
768 if (req->opcode == IORING_OP_READ && file_can_poll(req->file))
770 /* IOPOLL retry should happen for io-wq threads */
771 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
773 /* no retry on NONBLOCK nor RWF_NOWAIT */
774 if (req->flags & REQ_F_NOWAIT)
777 } else if (ret == -EIOCBQUEUED) {
780 return IOU_ISSUE_SKIP_COMPLETE;
781 } else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
782 (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) {
783 /* read all, failed, already did sync or don't want to retry */
788 * Don't depend on the iter state matching what was consumed, or being
789 * untouched in case of error. Restore it and we'll advance it
790 * manually if we need to.
792 iov_iter_restore(&s->iter, &s->iter_state);
794 ret2 = io_setup_async_rw(req, iovec, s, true);
797 ret = ret > 0 ? ret : ret2;
801 io = req->async_data;
804 * Now use our persistent iterator and state, if we aren't already.
805 * We've restored and mapped the iter to match.
810 * We end up here because of a partial read, either from
811 * above or inside this loop. Advance the iter by the bytes
812 * that were consumed.
814 iov_iter_advance(&s->iter, ret);
815 if (!iov_iter_count(&s->iter))
817 io->bytes_done += ret;
818 iov_iter_save_state(&s->iter, &s->iter_state);
820 /* if we can retry, do so with the callbacks armed */
821 if (!io_rw_should_retry(req)) {
822 kiocb->ki_flags &= ~IOCB_WAITQ;
826 req->cqe.res = iov_iter_count(&s->iter);
828 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
829 * we get -EIOCBQUEUED, then we'll get a notification when the
830 * desired page gets unlocked. We can also get a partial read
831 * here, and if we do, then just retry at the new offset.
833 ret = io_iter_do_read(rw, &s->iter);
834 if (ret == -EIOCBQUEUED)
835 return IOU_ISSUE_SKIP_COMPLETE;
836 /* we got some bytes, but not all. retry. */
837 kiocb->ki_flags &= ~IOCB_WAITQ;
838 iov_iter_restore(&s->iter, &s->iter_state);
841 /* it's faster to check here then delegate to kfree */
844 return kiocb_done(req, ret, issue_flags);
847 int io_write(struct io_kiocb *req, unsigned int issue_flags)
849 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
850 struct io_rw_state __s, *s = &__s;
852 struct kiocb *kiocb = &rw->kiocb;
853 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
857 if (!req_has_async_data(req)) {
858 ret = io_import_iovec(ITER_SOURCE, req, &iovec, s, issue_flags);
859 if (unlikely(ret < 0))
862 struct io_async_rw *io = req->async_data;
865 iov_iter_restore(&s->iter, &s->iter_state);
868 ret = io_rw_init_file(req, FMODE_WRITE);
873 req->cqe.res = iov_iter_count(&s->iter);
875 if (force_nonblock) {
876 /* If the file doesn't support async, just async punt */
877 if (unlikely(!io_file_supports_nowait(req)))
880 /* File path supports NOWAIT for non-direct_IO only for block devices. */
881 if (!(kiocb->ki_flags & IOCB_DIRECT) &&
882 !(kiocb->ki_filp->f_mode & FMODE_BUF_WASYNC) &&
883 (req->flags & REQ_F_ISREG))
886 kiocb->ki_flags |= IOCB_NOWAIT;
888 /* Ensure we clear previously set non-block flag */
889 kiocb->ki_flags &= ~IOCB_NOWAIT;
892 ppos = io_kiocb_update_pos(req);
894 ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
901 * Open-code file_start_write here to grab freeze protection,
902 * which will be released by another thread in
903 * io_complete_rw(). Fool lockdep by telling it the lock got
904 * released so that it doesn't complain about the held lock when
905 * we return to userspace.
907 if (req->flags & REQ_F_ISREG) {
908 sb_start_write(file_inode(req->file)->i_sb);
909 __sb_writers_release(file_inode(req->file)->i_sb,
912 kiocb->ki_flags |= IOCB_WRITE;
914 if (likely(req->file->f_op->write_iter))
915 ret2 = call_write_iter(req->file, kiocb, &s->iter);
916 else if (req->file->f_op->write)
917 ret2 = loop_rw_iter(WRITE, rw, &s->iter);
921 if (req->flags & REQ_F_REISSUE) {
922 req->flags &= ~REQ_F_REISSUE;
927 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
928 * retry them without IOCB_NOWAIT.
930 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
932 /* no retry on NONBLOCK nor RWF_NOWAIT */
933 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
935 if (!force_nonblock || ret2 != -EAGAIN) {
936 /* IOPOLL retry should happen for io-wq threads */
937 if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
940 if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) {
941 struct io_async_rw *io;
943 trace_io_uring_short_write(req->ctx, kiocb->ki_pos - ret2,
946 /* This is a partial write. The file pos has already been
947 * updated, setup the async struct to complete the request
948 * in the worker. Also update bytes_done to account for
949 * the bytes already written.
951 iov_iter_save_state(&s->iter, &s->iter_state);
952 ret = io_setup_async_rw(req, iovec, s, true);
954 io = req->async_data;
956 io->bytes_done += ret2;
958 if (kiocb->ki_flags & IOCB_WRITE)
959 kiocb_end_write(req);
960 return ret ? ret : -EAGAIN;
963 ret = kiocb_done(req, ret2, issue_flags);
966 iov_iter_restore(&s->iter, &s->iter_state);
967 ret = io_setup_async_rw(req, iovec, s, false);
969 if (kiocb->ki_flags & IOCB_WRITE)
970 kiocb_end_write(req);
975 /* it's reportedly faster than delegating the null check to kfree() */
981 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
983 io_commit_cqring_flush(ctx);
984 if (ctx->flags & IORING_SETUP_SQPOLL)
988 void io_rw_fail(struct io_kiocb *req)
992 res = io_fixup_rw_res(req, req->cqe.res);
993 io_req_set_res(req, res, req->cqe.flags);
996 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
998 struct io_wq_work_node *pos, *start, *prev;
999 unsigned int poll_flags = BLK_POLL_NOSLEEP;
1000 DEFINE_IO_COMP_BATCH(iob);
1004 * Only spin for completions if we don't have multiple devices hanging
1005 * off our complete list.
1007 if (ctx->poll_multi_queue || force_nonspin)
1008 poll_flags |= BLK_POLL_ONESHOT;
1010 wq_list_for_each(pos, start, &ctx->iopoll_list) {
1011 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1012 struct file *file = req->file;
1016 * Move completed and retryable entries to our local lists.
1017 * If we find a request that requires polling, break out
1018 * and complete those lists first, if we have entries there.
1020 if (READ_ONCE(req->iopoll_completed))
1023 if (req->opcode == IORING_OP_URING_CMD) {
1024 struct io_uring_cmd *ioucmd;
1026 ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
1027 ret = file->f_op->uring_cmd_iopoll(ioucmd, &iob,
1030 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1032 ret = file->f_op->iopoll(&rw->kiocb, &iob, poll_flags);
1034 if (unlikely(ret < 0))
1037 poll_flags |= BLK_POLL_ONESHOT;
1039 /* iopoll may have completed current req */
1040 if (!rq_list_empty(iob.req_list) ||
1041 READ_ONCE(req->iopoll_completed))
1045 if (!rq_list_empty(iob.req_list))
1051 wq_list_for_each_resume(pos, prev) {
1052 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1054 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
1055 if (!smp_load_acquire(&req->iopoll_completed))
1058 if (unlikely(req->flags & REQ_F_CQE_SKIP))
1061 req->cqe.flags = io_put_kbuf(req, 0);
1062 if (unlikely(!__io_fill_cqe_req(ctx, req))) {
1063 spin_lock(&ctx->completion_lock);
1064 io_req_cqe_overflow(req);
1065 spin_unlock(&ctx->completion_lock);
1069 if (unlikely(!nr_events))
1072 io_commit_cqring(ctx);
1073 io_cqring_ev_posted_iopoll(ctx);
1074 pos = start ? start->next : ctx->iopoll_list.first;
1075 wq_list_cut(&ctx->iopoll_list, prev, start);
1076 io_free_batch_list(ctx, pos);