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;
36 int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
38 struct io_rw *rw = io_kiocb_to_cmd(req);
42 rw->kiocb.ki_pos = READ_ONCE(sqe->off);
43 /* used for fixed read/write too - just read unconditionally */
44 req->buf_index = READ_ONCE(sqe->buf_index);
46 if (req->opcode == IORING_OP_READ_FIXED ||
47 req->opcode == IORING_OP_WRITE_FIXED) {
48 struct io_ring_ctx *ctx = req->ctx;
51 if (unlikely(req->buf_index >= ctx->nr_user_bufs))
53 index = array_index_nospec(req->buf_index, ctx->nr_user_bufs);
54 req->imu = ctx->user_bufs[index];
55 io_req_set_rsrc_node(req, ctx, 0);
58 ioprio = READ_ONCE(sqe->ioprio);
60 ret = ioprio_check_cap(ioprio);
64 rw->kiocb.ki_ioprio = ioprio;
66 rw->kiocb.ki_ioprio = get_current_ioprio();
69 rw->addr = READ_ONCE(sqe->addr);
70 rw->len = READ_ONCE(sqe->len);
71 rw->flags = READ_ONCE(sqe->rw_flags);
75 void io_readv_writev_cleanup(struct io_kiocb *req)
77 struct io_async_rw *io = req->async_data;
79 kfree(io->free_iovec);
82 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
90 case -ERESTART_RESTARTBLOCK:
92 * We can't just restart the syscall, since previously
93 * submitted sqes may already be in progress. Just fail this
99 kiocb->ki_complete(kiocb, ret);
103 static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req)
105 struct io_rw *rw = io_kiocb_to_cmd(req);
107 if (rw->kiocb.ki_pos != -1)
108 return &rw->kiocb.ki_pos;
110 if (!(req->file->f_mode & FMODE_STREAM)) {
111 req->flags |= REQ_F_CUR_POS;
112 rw->kiocb.ki_pos = req->file->f_pos;
113 return &rw->kiocb.ki_pos;
116 rw->kiocb.ki_pos = 0;
120 static void io_req_task_queue_reissue(struct io_kiocb *req)
122 req->io_task_work.func = io_queue_iowq;
123 io_req_task_work_add(req);
127 static bool io_resubmit_prep(struct io_kiocb *req)
129 struct io_async_rw *io = req->async_data;
131 if (!req_has_async_data(req))
132 return !io_req_prep_async(req);
133 iov_iter_restore(&io->s.iter, &io->s.iter_state);
137 static bool io_rw_should_reissue(struct io_kiocb *req)
139 umode_t mode = file_inode(req->file)->i_mode;
140 struct io_ring_ctx *ctx = req->ctx;
142 if (!S_ISBLK(mode) && !S_ISREG(mode))
144 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
145 !(ctx->flags & IORING_SETUP_IOPOLL)))
148 * If ref is dying, we might be running poll reap from the exit work.
149 * Don't attempt to reissue from that path, just let it fail with
152 if (percpu_ref_is_dying(&ctx->refs))
155 * Play it safe and assume not safe to re-import and reissue if we're
156 * not in the original thread group (or in task context).
158 if (!same_thread_group(req->task, current) || !in_task())
163 static bool io_resubmit_prep(struct io_kiocb *req)
167 static bool io_rw_should_reissue(struct io_kiocb *req)
173 static void kiocb_end_write(struct io_kiocb *req)
176 * Tell lockdep we inherited freeze protection from submission
179 if (req->flags & REQ_F_ISREG) {
180 struct super_block *sb = file_inode(req->file)->i_sb;
182 __sb_writers_acquired(sb, SB_FREEZE_WRITE);
187 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
189 struct io_rw *rw = io_kiocb_to_cmd(req);
191 if (rw->kiocb.ki_flags & IOCB_WRITE) {
192 kiocb_end_write(req);
193 fsnotify_modify(req->file);
195 fsnotify_access(req->file);
197 if (unlikely(res != req->cqe.res)) {
198 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
199 io_rw_should_reissue(req)) {
200 req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
209 static void io_complete_rw(struct kiocb *kiocb, long res)
211 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
212 struct io_kiocb *req = cmd_to_io_kiocb(rw);
214 if (__io_complete_rw_common(req, res))
216 io_req_set_res(req, res, 0);
217 req->io_task_work.func = io_req_task_complete;
218 io_req_task_prio_work_add(req);
221 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
223 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
224 struct io_kiocb *req = cmd_to_io_kiocb(rw);
226 if (kiocb->ki_flags & IOCB_WRITE)
227 kiocb_end_write(req);
228 if (unlikely(res != req->cqe.res)) {
229 if (res == -EAGAIN && io_rw_should_reissue(req)) {
230 req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
236 /* order with io_iopoll_complete() checking ->iopoll_completed */
237 smp_store_release(&req->iopoll_completed, 1);
240 static int kiocb_done(struct io_kiocb *req, ssize_t ret,
241 unsigned int issue_flags)
243 struct io_async_rw *io = req->async_data;
244 struct io_rw *rw = io_kiocb_to_cmd(req);
246 /* add previously done IO, if any */
247 if (req_has_async_data(req) && io->bytes_done > 0) {
249 ret = io->bytes_done;
251 ret += io->bytes_done;
254 if (req->flags & REQ_F_CUR_POS)
255 req->file->f_pos = rw->kiocb.ki_pos;
256 if (ret >= 0 && (rw->kiocb.ki_complete == io_complete_rw)) {
257 if (!__io_complete_rw_common(req, ret)) {
258 io_req_set_res(req, req->cqe.res,
259 io_put_kbuf(req, issue_flags));
263 io_rw_done(&rw->kiocb, ret);
266 if (req->flags & REQ_F_REISSUE) {
267 req->flags &= ~REQ_F_REISSUE;
268 if (io_resubmit_prep(req))
269 io_req_task_queue_reissue(req);
271 io_req_task_queue_fail(req, ret);
273 return IOU_ISSUE_SKIP_COMPLETE;
276 static int __io_import_fixed(struct io_kiocb *req, int ddir,
277 struct iov_iter *iter, struct io_mapped_ubuf *imu)
279 struct io_rw *rw = io_kiocb_to_cmd(req);
280 size_t len = rw->len;
281 u64 buf_end, buf_addr = rw->addr;
284 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end)))
286 /* not inside the mapped region */
287 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end))
291 * May not be a start of buffer, set size appropriately
292 * and advance us to the beginning.
294 offset = buf_addr - imu->ubuf;
295 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len);
299 * Don't use iov_iter_advance() here, as it's really slow for
300 * using the latter parts of a big fixed buffer - it iterates
301 * over each segment manually. We can cheat a bit here, because
304 * 1) it's a BVEC iter, we set it up
305 * 2) all bvecs are PAGE_SIZE in size, except potentially the
306 * first and last bvec
308 * So just find our index, and adjust the iterator afterwards.
309 * If the offset is within the first bvec (or the whole first
310 * bvec, just use iov_iter_advance(). This makes it easier
311 * since we can just skip the first segment, which may not
312 * be PAGE_SIZE aligned.
314 const struct bio_vec *bvec = imu->bvec;
316 if (offset <= bvec->bv_len) {
317 iov_iter_advance(iter, offset);
319 unsigned long seg_skip;
322 offset -= bvec->bv_len;
323 seg_skip = 1 + (offset >> PAGE_SHIFT);
325 iter->bvec = bvec + seg_skip;
326 iter->nr_segs -= seg_skip;
327 iter->count -= bvec->bv_len + offset;
328 iter->iov_offset = offset & ~PAGE_MASK;
335 static int io_import_fixed(struct io_kiocb *req, int rw, struct iov_iter *iter,
336 unsigned int issue_flags)
338 if (WARN_ON_ONCE(!req->imu))
340 return __io_import_fixed(req, rw, iter, req->imu);
344 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
345 unsigned int issue_flags)
347 struct io_rw *rw = io_kiocb_to_cmd(req);
348 struct compat_iovec __user *uiov;
353 uiov = u64_to_user_ptr(rw->addr);
354 if (!access_ok(uiov, sizeof(*uiov)))
356 if (__get_user(clen, &uiov->iov_len))
362 buf = io_buffer_select(req, &len, issue_flags);
365 rw->addr = (unsigned long) buf;
366 iov[0].iov_base = buf;
367 rw->len = iov[0].iov_len = (compat_size_t) len;
372 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
373 unsigned int issue_flags)
375 struct io_rw *rw = io_kiocb_to_cmd(req);
376 struct iovec __user *uiov = u64_to_user_ptr(rw->addr);
380 if (copy_from_user(iov, uiov, sizeof(*uiov)))
383 len = iov[0].iov_len;
386 buf = io_buffer_select(req, &len, issue_flags);
389 rw->addr = (unsigned long) buf;
390 iov[0].iov_base = buf;
391 rw->len = iov[0].iov_len = len;
395 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
396 unsigned int issue_flags)
398 struct io_rw *rw = io_kiocb_to_cmd(req);
400 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
401 iov[0].iov_base = u64_to_user_ptr(rw->addr);
402 iov[0].iov_len = rw->len;
409 if (req->ctx->compat)
410 return io_compat_import(req, iov, issue_flags);
413 return __io_iov_buffer_select(req, iov, issue_flags);
416 static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req,
417 struct io_rw_state *s,
418 unsigned int issue_flags)
420 struct io_rw *rw = io_kiocb_to_cmd(req);
421 struct iov_iter *iter = &s->iter;
422 u8 opcode = req->opcode;
428 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
429 ret = io_import_fixed(req, ddir, iter, issue_flags);
435 buf = u64_to_user_ptr(rw->addr);
438 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
439 if (io_do_buffer_select(req)) {
440 buf = io_buffer_select(req, &sqe_len, issue_flags);
442 return ERR_PTR(-ENOBUFS);
443 rw->addr = (unsigned long) buf;
447 ret = import_single_range(ddir, buf, sqe_len, s->fast_iov, iter);
454 if (req->flags & REQ_F_BUFFER_SELECT) {
455 ret = io_iov_buffer_select(req, iovec, issue_flags);
458 iov_iter_init(iter, ddir, iovec, 1, iovec->iov_len);
462 ret = __import_iovec(ddir, buf, sqe_len, UIO_FASTIOV, &iovec, iter,
464 if (unlikely(ret < 0))
469 static inline int io_import_iovec(int rw, struct io_kiocb *req,
470 struct iovec **iovec, struct io_rw_state *s,
471 unsigned int issue_flags)
473 *iovec = __io_import_iovec(rw, req, s, issue_flags);
474 if (unlikely(IS_ERR(*iovec)))
475 return PTR_ERR(*iovec);
477 iov_iter_save_state(&s->iter, &s->iter_state);
481 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
483 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
487 * For files that don't have ->read_iter() and ->write_iter(), handle them
488 * by looping over ->read() or ->write() manually.
490 static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter)
492 struct kiocb *kiocb = &rw->kiocb;
493 struct file *file = kiocb->ki_filp;
498 * Don't support polled IO through this interface, and we can't
499 * support non-blocking either. For the latter, this just causes
500 * the kiocb to be handled from an async context.
502 if (kiocb->ki_flags & IOCB_HIPRI)
504 if ((kiocb->ki_flags & IOCB_NOWAIT) &&
505 !(kiocb->ki_filp->f_flags & O_NONBLOCK))
508 ppos = io_kiocb_ppos(kiocb);
510 while (iov_iter_count(iter)) {
514 if (!iov_iter_is_bvec(iter)) {
515 iovec = iov_iter_iovec(iter);
517 iovec.iov_base = u64_to_user_ptr(rw->addr);
518 iovec.iov_len = rw->len;
522 nr = file->f_op->read(file, iovec.iov_base,
523 iovec.iov_len, ppos);
525 nr = file->f_op->write(file, iovec.iov_base,
526 iovec.iov_len, ppos);
535 if (!iov_iter_is_bvec(iter)) {
536 iov_iter_advance(iter, nr);
543 if (nr != iovec.iov_len)
550 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
551 const struct iovec *fast_iov, struct iov_iter *iter)
553 struct io_async_rw *io = req->async_data;
555 memcpy(&io->s.iter, iter, sizeof(*iter));
556 io->free_iovec = iovec;
558 /* can only be fixed buffers, no need to do anything */
559 if (iov_iter_is_bvec(iter))
562 unsigned iov_off = 0;
564 io->s.iter.iov = io->s.fast_iov;
565 if (iter->iov != fast_iov) {
566 iov_off = iter->iov - fast_iov;
567 io->s.iter.iov += iov_off;
569 if (io->s.fast_iov != fast_iov)
570 memcpy(io->s.fast_iov + iov_off, fast_iov + iov_off,
571 sizeof(struct iovec) * iter->nr_segs);
573 req->flags |= REQ_F_NEED_CLEANUP;
577 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
578 struct io_rw_state *s, bool force)
580 if (!force && !io_op_defs[req->opcode].prep_async)
582 if (!req_has_async_data(req)) {
583 struct io_async_rw *iorw;
585 if (io_alloc_async_data(req)) {
590 io_req_map_rw(req, iovec, s->fast_iov, &s->iter);
591 iorw = req->async_data;
592 /* we've copied and mapped the iter, ensure state is saved */
593 iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state);
598 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
600 struct io_async_rw *iorw = req->async_data;
604 /* submission path, ->uring_lock should already be taken */
605 ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
606 if (unlikely(ret < 0))
609 iorw->bytes_done = 0;
610 iorw->free_iovec = iov;
612 req->flags |= REQ_F_NEED_CLEANUP;
616 int io_readv_prep_async(struct io_kiocb *req)
618 return io_rw_prep_async(req, READ);
621 int io_writev_prep_async(struct io_kiocb *req)
623 return io_rw_prep_async(req, WRITE);
627 * This is our waitqueue callback handler, registered through __folio_lock_async()
628 * when we initially tried to do the IO with the iocb armed our waitqueue.
629 * This gets called when the page is unlocked, and we generally expect that to
630 * happen when the page IO is completed and the page is now uptodate. This will
631 * queue a task_work based retry of the operation, attempting to copy the data
632 * again. If the latter fails because the page was NOT uptodate, then we will
633 * do a thread based blocking retry of the operation. That's the unexpected
636 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
639 struct wait_page_queue *wpq;
640 struct io_kiocb *req = wait->private;
641 struct io_rw *rw = io_kiocb_to_cmd(req);
642 struct wait_page_key *key = arg;
644 wpq = container_of(wait, struct wait_page_queue, wait);
646 if (!wake_page_match(wpq, key))
649 rw->kiocb.ki_flags &= ~IOCB_WAITQ;
650 list_del_init(&wait->entry);
651 io_req_task_queue(req);
656 * This controls whether a given IO request should be armed for async page
657 * based retry. If we return false here, the request is handed to the async
658 * worker threads for retry. If we're doing buffered reads on a regular file,
659 * we prepare a private wait_page_queue entry and retry the operation. This
660 * will either succeed because the page is now uptodate and unlocked, or it
661 * will register a callback when the page is unlocked at IO completion. Through
662 * that callback, io_uring uses task_work to setup a retry of the operation.
663 * That retry will attempt the buffered read again. The retry will generally
664 * succeed, or in rare cases where it fails, we then fall back to using the
665 * async worker threads for a blocking retry.
667 static bool io_rw_should_retry(struct io_kiocb *req)
669 struct io_async_rw *io = req->async_data;
670 struct wait_page_queue *wait = &io->wpq;
671 struct io_rw *rw = io_kiocb_to_cmd(req);
672 struct kiocb *kiocb = &rw->kiocb;
674 /* never retry for NOWAIT, we just complete with -EAGAIN */
675 if (req->flags & REQ_F_NOWAIT)
678 /* Only for buffered IO */
679 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
683 * just use poll if we can, and don't attempt if the fs doesn't
684 * support callback based unlocks
686 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
689 wait->wait.func = io_async_buf_func;
690 wait->wait.private = req;
691 wait->wait.flags = 0;
692 INIT_LIST_HEAD(&wait->wait.entry);
693 kiocb->ki_flags |= IOCB_WAITQ;
694 kiocb->ki_flags &= ~IOCB_NOWAIT;
695 kiocb->ki_waitq = wait;
699 static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
701 struct file *file = rw->kiocb.ki_filp;
703 if (likely(file->f_op->read_iter))
704 return call_read_iter(file, &rw->kiocb, iter);
705 else if (file->f_op->read)
706 return loop_rw_iter(READ, rw, iter);
711 static bool need_read_all(struct io_kiocb *req)
713 return req->flags & REQ_F_ISREG ||
714 S_ISBLK(file_inode(req->file)->i_mode);
717 static inline bool io_req_ffs_set(struct io_kiocb *req)
719 return req->flags & REQ_F_FIXED_FILE;
722 static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
724 struct io_rw *rw = io_kiocb_to_cmd(req);
725 struct kiocb *kiocb = &rw->kiocb;
726 struct io_ring_ctx *ctx = req->ctx;
727 struct file *file = req->file;
730 if (unlikely(!file || !(file->f_mode & mode)))
733 if (!io_req_ffs_set(req))
734 req->flags |= io_file_get_flags(file) << REQ_F_SUPPORT_NOWAIT_BIT;
736 kiocb->ki_flags = iocb_flags(file);
737 ret = kiocb_set_rw_flags(kiocb, rw->flags);
742 * If the file is marked O_NONBLOCK, still allow retry for it if it
743 * supports async. Otherwise it's impossible to use O_NONBLOCK files
744 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
746 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
747 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
748 req->flags |= REQ_F_NOWAIT;
750 if (ctx->flags & IORING_SETUP_IOPOLL) {
751 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
754 kiocb->private = NULL;
755 kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
756 kiocb->ki_complete = io_complete_rw_iopoll;
757 req->iopoll_completed = 0;
759 if (kiocb->ki_flags & IOCB_HIPRI)
761 kiocb->ki_complete = io_complete_rw;
767 int io_read(struct io_kiocb *req, unsigned int issue_flags)
769 struct io_rw *rw = io_kiocb_to_cmd(req);
770 struct io_rw_state __s, *s = &__s;
772 struct kiocb *kiocb = &rw->kiocb;
773 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
774 struct io_async_rw *io;
778 if (!req_has_async_data(req)) {
779 ret = io_import_iovec(READ, req, &iovec, s, issue_flags);
780 if (unlikely(ret < 0))
783 io = req->async_data;
787 * Safe and required to re-import if we're using provided
788 * buffers, as we dropped the selected one before retry.
790 if (io_do_buffer_select(req)) {
791 ret = io_import_iovec(READ, req, &iovec, s, issue_flags);
792 if (unlikely(ret < 0))
797 * We come here from an earlier attempt, restore our state to
798 * match in case it doesn't. It's cheap enough that we don't
799 * need to make this conditional.
801 iov_iter_restore(&s->iter, &s->iter_state);
804 ret = io_rw_init_file(req, FMODE_READ);
809 req->cqe.res = iov_iter_count(&s->iter);
811 if (force_nonblock) {
812 /* If the file doesn't support async, just async punt */
813 if (unlikely(!io_file_supports_nowait(req))) {
814 ret = io_setup_async_rw(req, iovec, s, true);
815 return ret ?: -EAGAIN;
817 kiocb->ki_flags |= IOCB_NOWAIT;
819 /* Ensure we clear previously set non-block flag */
820 kiocb->ki_flags &= ~IOCB_NOWAIT;
823 ppos = io_kiocb_update_pos(req);
825 ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
831 ret = io_iter_do_read(rw, &s->iter);
833 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
834 req->flags &= ~REQ_F_REISSUE;
835 /* if we can poll, just do that */
836 if (req->opcode == IORING_OP_READ && file_can_poll(req->file))
838 /* IOPOLL retry should happen for io-wq threads */
839 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
841 /* no retry on NONBLOCK nor RWF_NOWAIT */
842 if (req->flags & REQ_F_NOWAIT)
845 } else if (ret == -EIOCBQUEUED) {
848 return IOU_ISSUE_SKIP_COMPLETE;
849 } else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
850 (req->flags & REQ_F_NOWAIT) || !need_read_all(req)) {
851 /* read all, failed, already did sync or don't want to retry */
856 * Don't depend on the iter state matching what was consumed, or being
857 * untouched in case of error. Restore it and we'll advance it
858 * manually if we need to.
860 iov_iter_restore(&s->iter, &s->iter_state);
862 ret2 = io_setup_async_rw(req, iovec, s, true);
867 io = req->async_data;
870 * Now use our persistent iterator and state, if we aren't already.
871 * We've restored and mapped the iter to match.
876 * We end up here because of a partial read, either from
877 * above or inside this loop. Advance the iter by the bytes
878 * that were consumed.
880 iov_iter_advance(&s->iter, ret);
881 if (!iov_iter_count(&s->iter))
883 io->bytes_done += ret;
884 iov_iter_save_state(&s->iter, &s->iter_state);
886 /* if we can retry, do so with the callbacks armed */
887 if (!io_rw_should_retry(req)) {
888 kiocb->ki_flags &= ~IOCB_WAITQ;
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 */
909 return kiocb_done(req, ret, issue_flags);
912 int io_write(struct io_kiocb *req, unsigned int issue_flags)
914 struct io_rw *rw = io_kiocb_to_cmd(req);
915 struct io_rw_state __s, *s = &__s;
917 struct kiocb *kiocb = &rw->kiocb;
918 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
922 if (!req_has_async_data(req)) {
923 ret = io_import_iovec(WRITE, req, &iovec, s, issue_flags);
924 if (unlikely(ret < 0))
927 struct io_async_rw *io = req->async_data;
930 iov_iter_restore(&s->iter, &s->iter_state);
933 ret = io_rw_init_file(req, FMODE_WRITE);
938 req->cqe.res = iov_iter_count(&s->iter);
940 if (force_nonblock) {
941 /* If the file doesn't support async, just async punt */
942 if (unlikely(!io_file_supports_nowait(req)))
945 /* file path doesn't support NOWAIT for non-direct_IO */
946 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
947 (req->flags & REQ_F_ISREG))
950 kiocb->ki_flags |= IOCB_NOWAIT;
952 /* Ensure we clear previously set non-block flag */
953 kiocb->ki_flags &= ~IOCB_NOWAIT;
956 ppos = io_kiocb_update_pos(req);
958 ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
965 * Open-code file_start_write here to grab freeze protection,
966 * which will be released by another thread in
967 * io_complete_rw(). Fool lockdep by telling it the lock got
968 * released so that it doesn't complain about the held lock when
969 * we return to userspace.
971 if (req->flags & REQ_F_ISREG) {
972 sb_start_write(file_inode(req->file)->i_sb);
973 __sb_writers_release(file_inode(req->file)->i_sb,
976 kiocb->ki_flags |= IOCB_WRITE;
978 if (likely(req->file->f_op->write_iter))
979 ret2 = call_write_iter(req->file, kiocb, &s->iter);
980 else if (req->file->f_op->write)
981 ret2 = loop_rw_iter(WRITE, rw, &s->iter);
985 if (req->flags & REQ_F_REISSUE) {
986 req->flags &= ~REQ_F_REISSUE;
991 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
992 * retry them without IOCB_NOWAIT.
994 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
996 /* no retry on NONBLOCK nor RWF_NOWAIT */
997 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
999 if (!force_nonblock || ret2 != -EAGAIN) {
1000 /* IOPOLL retry should happen for io-wq threads */
1001 if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
1004 ret = kiocb_done(req, ret2, issue_flags);
1007 iov_iter_restore(&s->iter, &s->iter_state);
1008 ret = io_setup_async_rw(req, iovec, s, false);
1009 return ret ?: -EAGAIN;
1011 /* it's reportedly faster than delegating the null check to kfree() */
1017 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
1019 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
1021 __io_commit_cqring_flush(ctx);
1023 if (ctx->flags & IORING_SETUP_SQPOLL)
1024 io_cqring_wake(ctx);
1027 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
1029 struct io_wq_work_node *pos, *start, *prev;
1030 unsigned int poll_flags = BLK_POLL_NOSLEEP;
1031 DEFINE_IO_COMP_BATCH(iob);
1035 * Only spin for completions if we don't have multiple devices hanging
1036 * off our complete list.
1038 if (ctx->poll_multi_queue || force_nonspin)
1039 poll_flags |= BLK_POLL_ONESHOT;
1041 wq_list_for_each(pos, start, &ctx->iopoll_list) {
1042 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1043 struct io_rw *rw = io_kiocb_to_cmd(req);
1047 * Move completed and retryable entries to our local lists.
1048 * If we find a request that requires polling, break out
1049 * and complete those lists first, if we have entries there.
1051 if (READ_ONCE(req->iopoll_completed))
1054 ret = rw->kiocb.ki_filp->f_op->iopoll(&rw->kiocb, &iob, poll_flags);
1055 if (unlikely(ret < 0))
1058 poll_flags |= BLK_POLL_ONESHOT;
1060 /* iopoll may have completed current req */
1061 if (!rq_list_empty(iob.req_list) ||
1062 READ_ONCE(req->iopoll_completed))
1066 if (!rq_list_empty(iob.req_list))
1072 wq_list_for_each_resume(pos, prev) {
1073 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1075 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
1076 if (!smp_load_acquire(&req->iopoll_completed))
1079 if (unlikely(req->flags & REQ_F_CQE_SKIP))
1082 req->cqe.flags = io_put_kbuf(req, 0);
1083 __io_fill_cqe_req(req->ctx, req);
1086 if (unlikely(!nr_events))
1089 io_commit_cqring(ctx);
1090 io_cqring_ev_posted_iopoll(ctx);
1091 pos = start ? start->next : ctx->iopoll_list.first;
1092 wq_list_cut(&ctx->iopoll_list, prev, start);
1093 io_free_batch_list(ctx, pos);