1// SPDX-License-Identifier: GPL-2.0
2#include <linux/kernel.h>
3#include <linux/errno.h>
4#include <linux/fs.h>
5#include <linux/file.h>
6#include <linux/blk-mq.h>
7#include <linux/mm.h>
8#include <linux/slab.h>
9#include <linux/fsnotify.h>
10#include <linux/poll.h>
11#include <linux/nospec.h>
12#include <linux/compat.h>
13#include <linux/io_uring.h>
14
15#include <uapi/linux/io_uring.h>
16
17#include "io_uring.h"
18#include "opdef.h"
19#include "kbuf.h"
20#include "rsrc.h"
21#include "rw.h"
22
23struct io_rw {
24 /* NOTE: kiocb has the file as the first member, so don't do it here */
25 struct kiocb kiocb;
26 u64 addr;
27 u32 len;
28 rwf_t flags;
29};
30
31static inline bool io_file_supports_nowait(struct io_kiocb *req)
32{
33 return req->flags & REQ_F_SUPPORT_NOWAIT;
34}
35
36#ifdef CONFIG_COMPAT
37static int io_iov_compat_buffer_select_prep(struct io_rw *rw)
38{
39 struct compat_iovec __user *uiov;
40 compat_ssize_t clen;
41
42 uiov = u64_to_user_ptr(rw->addr);
43 if (!access_ok(uiov, sizeof(*uiov)))
44 return -EFAULT;
45 if (__get_user(clen, &uiov->iov_len))
46 return -EFAULT;
47 if (clen < 0)
48 return -EINVAL;
49
50 rw->len = clen;
51 return 0;
52}
53#endif
54
55static int io_iov_buffer_select_prep(struct io_kiocb *req)
56{
57 struct iovec __user *uiov;
58 struct iovec iov;
59 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
60
61 if (rw->len != 1)
62 return -EINVAL;
63
64#ifdef CONFIG_COMPAT
65 if (req->ctx->compat)
66 return io_iov_compat_buffer_select_prep(rw);
67#endif
68
69 uiov = u64_to_user_ptr(rw->addr);
70 if (copy_from_user(to: &iov, from: uiov, n: sizeof(*uiov)))
71 return -EFAULT;
72 rw->len = iov.iov_len;
73 return 0;
74}
75
76int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
77{
78 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
79 unsigned ioprio;
80 int ret;
81
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);
85
86 if (req->opcode == IORING_OP_READ_FIXED ||
87 req->opcode == IORING_OP_WRITE_FIXED) {
88 struct io_ring_ctx *ctx = req->ctx;
89 u16 index;
90
91 if (unlikely(req->buf_index >= ctx->nr_user_bufs))
92 return -EFAULT;
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, issue_flags: 0);
96 }
97
98 ioprio = READ_ONCE(sqe->ioprio);
99 if (ioprio) {
100 ret = ioprio_check_cap(ioprio);
101 if (ret)
102 return ret;
103
104 rw->kiocb.ki_ioprio = ioprio;
105 } else {
106 rw->kiocb.ki_ioprio = get_current_ioprio();
107 }
108 rw->kiocb.dio_complete = NULL;
109
110 rw->addr = READ_ONCE(sqe->addr);
111 rw->len = READ_ONCE(sqe->len);
112 rw->flags = READ_ONCE(sqe->rw_flags);
113
114 /* Have to do this validation here, as this is in io_read() rw->len might
115 * have chanaged due to buffer selection
116 */
117 if (req->opcode == IORING_OP_READV && req->flags & REQ_F_BUFFER_SELECT) {
118 ret = io_iov_buffer_select_prep(req);
119 if (ret)
120 return ret;
121 }
122
123 return 0;
124}
125
126/*
127 * Multishot read is prepared just like a normal read/write request, only
128 * difference is that we set the MULTISHOT flag.
129 */
130int io_read_mshot_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
131{
132 int ret;
133
134 ret = io_prep_rw(req, sqe);
135 if (unlikely(ret))
136 return ret;
137
138 req->flags |= REQ_F_APOLL_MULTISHOT;
139 return 0;
140}
141
142void io_readv_writev_cleanup(struct io_kiocb *req)
143{
144 struct io_async_rw *io = req->async_data;
145
146 kfree(objp: io->free_iovec);
147}
148
149static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
150{
151 switch (ret) {
152 case -EIOCBQUEUED:
153 break;
154 case -ERESTARTSYS:
155 case -ERESTARTNOINTR:
156 case -ERESTARTNOHAND:
157 case -ERESTART_RESTARTBLOCK:
158 /*
159 * We can't just restart the syscall, since previously
160 * submitted sqes may already be in progress. Just fail this
161 * IO with EINTR.
162 */
163 ret = -EINTR;
164 fallthrough;
165 default:
166 kiocb->ki_complete(kiocb, ret);
167 }
168}
169
170static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req)
171{
172 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
173
174 if (rw->kiocb.ki_pos != -1)
175 return &rw->kiocb.ki_pos;
176
177 if (!(req->file->f_mode & FMODE_STREAM)) {
178 req->flags |= REQ_F_CUR_POS;
179 rw->kiocb.ki_pos = req->file->f_pos;
180 return &rw->kiocb.ki_pos;
181 }
182
183 rw->kiocb.ki_pos = 0;
184 return NULL;
185}
186
187static void io_req_task_queue_reissue(struct io_kiocb *req)
188{
189 req->io_task_work.func = io_queue_iowq;
190 io_req_task_work_add(req);
191}
192
193#ifdef CONFIG_BLOCK
194static bool io_resubmit_prep(struct io_kiocb *req)
195{
196 struct io_async_rw *io = req->async_data;
197
198 if (!req_has_async_data(req))
199 return !io_req_prep_async(req);
200 iov_iter_restore(i: &io->s.iter, state: &io->s.iter_state);
201 return true;
202}
203
204static bool io_rw_should_reissue(struct io_kiocb *req)
205{
206 umode_t mode = file_inode(f: req->file)->i_mode;
207 struct io_ring_ctx *ctx = req->ctx;
208
209 if (!S_ISBLK(mode) && !S_ISREG(mode))
210 return false;
211 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
212 !(ctx->flags & IORING_SETUP_IOPOLL)))
213 return false;
214 /*
215 * If ref is dying, we might be running poll reap from the exit work.
216 * Don't attempt to reissue from that path, just let it fail with
217 * -EAGAIN.
218 */
219 if (percpu_ref_is_dying(ref: &ctx->refs))
220 return false;
221 /*
222 * Play it safe and assume not safe to re-import and reissue if we're
223 * not in the original thread group (or in task context).
224 */
225 if (!same_thread_group(p1: req->task, current) || !in_task())
226 return false;
227 return true;
228}
229#else
230static bool io_resubmit_prep(struct io_kiocb *req)
231{
232 return false;
233}
234static bool io_rw_should_reissue(struct io_kiocb *req)
235{
236 return false;
237}
238#endif
239
240static void io_req_end_write(struct io_kiocb *req)
241{
242 if (req->flags & REQ_F_ISREG) {
243 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
244
245 kiocb_end_write(iocb: &rw->kiocb);
246 }
247}
248
249/*
250 * Trigger the notifications after having done some IO, and finish the write
251 * accounting, if any.
252 */
253static void io_req_io_end(struct io_kiocb *req)
254{
255 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
256
257 if (rw->kiocb.ki_flags & IOCB_WRITE) {
258 io_req_end_write(req);
259 fsnotify_modify(file: req->file);
260 } else {
261 fsnotify_access(file: req->file);
262 }
263}
264
265static bool __io_complete_rw_common(struct io_kiocb *req, long res)
266{
267 if (unlikely(res != req->cqe.res)) {
268 if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
269 io_rw_should_reissue(req)) {
270 /*
271 * Reissue will start accounting again, finish the
272 * current cycle.
273 */
274 io_req_io_end(req);
275 req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
276 return true;
277 }
278 req_set_fail(req);
279 req->cqe.res = res;
280 }
281 return false;
282}
283
284static inline int io_fixup_rw_res(struct io_kiocb *req, long res)
285{
286 struct io_async_rw *io = req->async_data;
287
288 /* add previously done IO, if any */
289 if (req_has_async_data(req) && io->bytes_done > 0) {
290 if (res < 0)
291 res = io->bytes_done;
292 else
293 res += io->bytes_done;
294 }
295 return res;
296}
297
298void io_req_rw_complete(struct io_kiocb *req, struct io_tw_state *ts)
299{
300 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
301 struct kiocb *kiocb = &rw->kiocb;
302
303 if ((kiocb->ki_flags & IOCB_DIO_CALLER_COMP) && kiocb->dio_complete) {
304 long res = kiocb->dio_complete(rw->kiocb.private);
305
306 io_req_set_res(req, res: io_fixup_rw_res(req, res), cflags: 0);
307 }
308
309 io_req_io_end(req);
310
311 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
312 unsigned issue_flags = ts->locked ? 0 : IO_URING_F_UNLOCKED;
313
314 req->cqe.flags |= io_put_kbuf(req, issue_flags);
315 }
316 io_req_task_complete(req, ts);
317}
318
319static void io_complete_rw(struct kiocb *kiocb, long res)
320{
321 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
322 struct io_kiocb *req = cmd_to_io_kiocb(rw);
323
324 if (!kiocb->dio_complete || !(kiocb->ki_flags & IOCB_DIO_CALLER_COMP)) {
325 if (__io_complete_rw_common(req, res))
326 return;
327 io_req_set_res(req, res: io_fixup_rw_res(req, res), cflags: 0);
328 }
329 req->io_task_work.func = io_req_rw_complete;
330 __io_req_task_work_add(req, flags: IOU_F_TWQ_LAZY_WAKE);
331}
332
333static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
334{
335 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
336 struct io_kiocb *req = cmd_to_io_kiocb(rw);
337
338 if (kiocb->ki_flags & IOCB_WRITE)
339 io_req_end_write(req);
340 if (unlikely(res != req->cqe.res)) {
341 if (res == -EAGAIN && io_rw_should_reissue(req)) {
342 req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
343 return;
344 }
345 req->cqe.res = res;
346 }
347
348 /* order with io_iopoll_complete() checking ->iopoll_completed */
349 smp_store_release(&req->iopoll_completed, 1);
350}
351
352static int kiocb_done(struct io_kiocb *req, ssize_t ret,
353 unsigned int issue_flags)
354{
355 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
356 unsigned final_ret = io_fixup_rw_res(req, res: ret);
357
358 if (ret >= 0 && req->flags & REQ_F_CUR_POS)
359 req->file->f_pos = rw->kiocb.ki_pos;
360 if (ret >= 0 && (rw->kiocb.ki_complete == io_complete_rw)) {
361 if (!__io_complete_rw_common(req, res: ret)) {
362 /*
363 * Safe to call io_end from here as we're inline
364 * from the submission path.
365 */
366 io_req_io_end(req);
367 io_req_set_res(req, res: final_ret,
368 cflags: io_put_kbuf(req, issue_flags));
369 return IOU_OK;
370 }
371 } else {
372 io_rw_done(kiocb: &rw->kiocb, ret);
373 }
374
375 if (req->flags & REQ_F_REISSUE) {
376 req->flags &= ~REQ_F_REISSUE;
377 if (io_resubmit_prep(req))
378 io_req_task_queue_reissue(req);
379 else
380 io_req_task_queue_fail(req, ret: final_ret);
381 }
382 return IOU_ISSUE_SKIP_COMPLETE;
383}
384
385static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req,
386 struct io_rw_state *s,
387 unsigned int issue_flags)
388{
389 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
390 struct iov_iter *iter = &s->iter;
391 u8 opcode = req->opcode;
392 struct iovec *iovec;
393 void __user *buf;
394 size_t sqe_len;
395 ssize_t ret;
396
397 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
398 ret = io_import_fixed(ddir, iter, imu: req->imu, buf_addr: rw->addr, len: rw->len);
399 if (ret)
400 return ERR_PTR(error: ret);
401 return NULL;
402 }
403
404 buf = u64_to_user_ptr(rw->addr);
405 sqe_len = rw->len;
406
407 if (!io_issue_defs[opcode].vectored || req->flags & REQ_F_BUFFER_SELECT) {
408 if (io_do_buffer_select(req)) {
409 buf = io_buffer_select(req, len: &sqe_len, issue_flags);
410 if (!buf)
411 return ERR_PTR(error: -ENOBUFS);
412 rw->addr = (unsigned long) buf;
413 rw->len = sqe_len;
414 }
415
416 ret = import_ubuf(type: ddir, buf, len: sqe_len, i: iter);
417 if (ret)
418 return ERR_PTR(error: ret);
419 return NULL;
420 }
421
422 iovec = s->fast_iov;
423 ret = __import_iovec(type: ddir, uvec: buf, nr_segs: sqe_len, UIO_FASTIOV, iovp: &iovec, i: iter,
424 compat: req->ctx->compat);
425 if (unlikely(ret < 0))
426 return ERR_PTR(error: ret);
427 return iovec;
428}
429
430static inline int io_import_iovec(int rw, struct io_kiocb *req,
431 struct iovec **iovec, struct io_rw_state *s,
432 unsigned int issue_flags)
433{
434 *iovec = __io_import_iovec(ddir: rw, req, s, issue_flags);
435 if (IS_ERR(ptr: *iovec))
436 return PTR_ERR(ptr: *iovec);
437
438 iov_iter_save_state(iter: &s->iter, state: &s->iter_state);
439 return 0;
440}
441
442static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
443{
444 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
445}
446
447/*
448 * For files that don't have ->read_iter() and ->write_iter(), handle them
449 * by looping over ->read() or ->write() manually.
450 */
451static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter)
452{
453 struct kiocb *kiocb = &rw->kiocb;
454 struct file *file = kiocb->ki_filp;
455 ssize_t ret = 0;
456 loff_t *ppos;
457
458 /*
459 * Don't support polled IO through this interface, and we can't
460 * support non-blocking either. For the latter, this just causes
461 * the kiocb to be handled from an async context.
462 */
463 if (kiocb->ki_flags & IOCB_HIPRI)
464 return -EOPNOTSUPP;
465 if ((kiocb->ki_flags & IOCB_NOWAIT) &&
466 !(kiocb->ki_filp->f_flags & O_NONBLOCK))
467 return -EAGAIN;
468
469 ppos = io_kiocb_ppos(kiocb);
470
471 while (iov_iter_count(i: iter)) {
472 void __user *addr;
473 size_t len;
474 ssize_t nr;
475
476 if (iter_is_ubuf(i: iter)) {
477 addr = iter->ubuf + iter->iov_offset;
478 len = iov_iter_count(i: iter);
479 } else if (!iov_iter_is_bvec(i: iter)) {
480 addr = iter_iov_addr(iter);
481 len = iter_iov_len(iter);
482 } else {
483 addr = u64_to_user_ptr(rw->addr);
484 len = rw->len;
485 }
486
487 if (ddir == READ)
488 nr = file->f_op->read(file, addr, len, ppos);
489 else
490 nr = file->f_op->write(file, addr, len, ppos);
491
492 if (nr < 0) {
493 if (!ret)
494 ret = nr;
495 break;
496 }
497 ret += nr;
498 if (!iov_iter_is_bvec(i: iter)) {
499 iov_iter_advance(i: iter, bytes: nr);
500 } else {
501 rw->addr += nr;
502 rw->len -= nr;
503 if (!rw->len)
504 break;
505 }
506 if (nr != len)
507 break;
508 }
509
510 return ret;
511}
512
513static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
514 const struct iovec *fast_iov, struct iov_iter *iter)
515{
516 struct io_async_rw *io = req->async_data;
517
518 memcpy(&io->s.iter, iter, sizeof(*iter));
519 io->free_iovec = iovec;
520 io->bytes_done = 0;
521 /* can only be fixed buffers, no need to do anything */
522 if (iov_iter_is_bvec(i: iter) || iter_is_ubuf(i: iter))
523 return;
524 if (!iovec) {
525 unsigned iov_off = 0;
526
527 io->s.iter.__iov = io->s.fast_iov;
528 if (iter->__iov != fast_iov) {
529 iov_off = iter_iov(iter) - fast_iov;
530 io->s.iter.__iov += iov_off;
531 }
532 if (io->s.fast_iov != fast_iov)
533 memcpy(io->s.fast_iov + iov_off, fast_iov + iov_off,
534 sizeof(struct iovec) * iter->nr_segs);
535 } else {
536 req->flags |= REQ_F_NEED_CLEANUP;
537 }
538}
539
540static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
541 struct io_rw_state *s, bool force)
542{
543 if (!force && !io_cold_defs[req->opcode].prep_async)
544 return 0;
545 if (!req_has_async_data(req)) {
546 struct io_async_rw *iorw;
547
548 if (io_alloc_async_data(req)) {
549 kfree(objp: iovec);
550 return -ENOMEM;
551 }
552
553 io_req_map_rw(req, iovec, fast_iov: s->fast_iov, iter: &s->iter);
554 iorw = req->async_data;
555 /* we've copied and mapped the iter, ensure state is saved */
556 iov_iter_save_state(iter: &iorw->s.iter, state: &iorw->s.iter_state);
557 }
558 return 0;
559}
560
561static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
562{
563 struct io_async_rw *iorw = req->async_data;
564 struct iovec *iov;
565 int ret;
566
567 /* submission path, ->uring_lock should already be taken */
568 ret = io_import_iovec(rw, req, iovec: &iov, s: &iorw->s, issue_flags: 0);
569 if (unlikely(ret < 0))
570 return ret;
571
572 iorw->bytes_done = 0;
573 iorw->free_iovec = iov;
574 if (iov)
575 req->flags |= REQ_F_NEED_CLEANUP;
576 return 0;
577}
578
579int io_readv_prep_async(struct io_kiocb *req)
580{
581 return io_rw_prep_async(req, ITER_DEST);
582}
583
584int io_writev_prep_async(struct io_kiocb *req)
585{
586 return io_rw_prep_async(req, ITER_SOURCE);
587}
588
589/*
590 * This is our waitqueue callback handler, registered through __folio_lock_async()
591 * when we initially tried to do the IO with the iocb armed our waitqueue.
592 * This gets called when the page is unlocked, and we generally expect that to
593 * happen when the page IO is completed and the page is now uptodate. This will
594 * queue a task_work based retry of the operation, attempting to copy the data
595 * again. If the latter fails because the page was NOT uptodate, then we will
596 * do a thread based blocking retry of the operation. That's the unexpected
597 * slow path.
598 */
599static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
600 int sync, void *arg)
601{
602 struct wait_page_queue *wpq;
603 struct io_kiocb *req = wait->private;
604 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
605 struct wait_page_key *key = arg;
606
607 wpq = container_of(wait, struct wait_page_queue, wait);
608
609 if (!wake_page_match(wait_page: wpq, key))
610 return 0;
611
612 rw->kiocb.ki_flags &= ~IOCB_WAITQ;
613 list_del_init(entry: &wait->entry);
614 io_req_task_queue(req);
615 return 1;
616}
617
618/*
619 * This controls whether a given IO request should be armed for async page
620 * based retry. If we return false here, the request is handed to the async
621 * worker threads for retry. If we're doing buffered reads on a regular file,
622 * we prepare a private wait_page_queue entry and retry the operation. This
623 * will either succeed because the page is now uptodate and unlocked, or it
624 * will register a callback when the page is unlocked at IO completion. Through
625 * that callback, io_uring uses task_work to setup a retry of the operation.
626 * That retry will attempt the buffered read again. The retry will generally
627 * succeed, or in rare cases where it fails, we then fall back to using the
628 * async worker threads for a blocking retry.
629 */
630static bool io_rw_should_retry(struct io_kiocb *req)
631{
632 struct io_async_rw *io = req->async_data;
633 struct wait_page_queue *wait = &io->wpq;
634 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
635 struct kiocb *kiocb = &rw->kiocb;
636
637 /* never retry for NOWAIT, we just complete with -EAGAIN */
638 if (req->flags & REQ_F_NOWAIT)
639 return false;
640
641 /* Only for buffered IO */
642 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
643 return false;
644
645 /*
646 * just use poll if we can, and don't attempt if the fs doesn't
647 * support callback based unlocks
648 */
649 if (file_can_poll(file: req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
650 return false;
651
652 wait->wait.func = io_async_buf_func;
653 wait->wait.private = req;
654 wait->wait.flags = 0;
655 INIT_LIST_HEAD(list: &wait->wait.entry);
656 kiocb->ki_flags |= IOCB_WAITQ;
657 kiocb->ki_flags &= ~IOCB_NOWAIT;
658 kiocb->ki_waitq = wait;
659 return true;
660}
661
662static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
663{
664 struct file *file = rw->kiocb.ki_filp;
665
666 if (likely(file->f_op->read_iter))
667 return call_read_iter(file, kio: &rw->kiocb, iter);
668 else if (file->f_op->read)
669 return loop_rw_iter(READ, rw, iter);
670 else
671 return -EINVAL;
672}
673
674static bool need_complete_io(struct io_kiocb *req)
675{
676 return req->flags & REQ_F_ISREG ||
677 S_ISBLK(file_inode(req->file)->i_mode);
678}
679
680static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
681{
682 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
683 struct kiocb *kiocb = &rw->kiocb;
684 struct io_ring_ctx *ctx = req->ctx;
685 struct file *file = req->file;
686 int ret;
687
688 if (unlikely(!file || !(file->f_mode & mode)))
689 return -EBADF;
690
691 if (!(req->flags & REQ_F_FIXED_FILE))
692 req->flags |= io_file_get_flags(file);
693
694 kiocb->ki_flags = file->f_iocb_flags;
695 ret = kiocb_set_rw_flags(ki: kiocb, flags: rw->flags);
696 if (unlikely(ret))
697 return ret;
698 kiocb->ki_flags |= IOCB_ALLOC_CACHE;
699
700 /*
701 * If the file is marked O_NONBLOCK, still allow retry for it if it
702 * supports async. Otherwise it's impossible to use O_NONBLOCK files
703 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
704 */
705 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
706 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
707 req->flags |= REQ_F_NOWAIT;
708
709 if (ctx->flags & IORING_SETUP_IOPOLL) {
710 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
711 return -EOPNOTSUPP;
712
713 kiocb->private = NULL;
714 kiocb->ki_flags |= IOCB_HIPRI;
715 kiocb->ki_complete = io_complete_rw_iopoll;
716 req->iopoll_completed = 0;
717 } else {
718 if (kiocb->ki_flags & IOCB_HIPRI)
719 return -EINVAL;
720 kiocb->ki_complete = io_complete_rw;
721 }
722
723 return 0;
724}
725
726static int __io_read(struct io_kiocb *req, unsigned int issue_flags)
727{
728 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
729 struct io_rw_state __s, *s = &__s;
730 struct iovec *iovec;
731 struct kiocb *kiocb = &rw->kiocb;
732 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
733 struct io_async_rw *io;
734 ssize_t ret, ret2;
735 loff_t *ppos;
736
737 if (!req_has_async_data(req)) {
738 ret = io_import_iovec(ITER_DEST, req, iovec: &iovec, s, issue_flags);
739 if (unlikely(ret < 0))
740 return ret;
741 } else {
742 io = req->async_data;
743 s = &io->s;
744
745 /*
746 * Safe and required to re-import if we're using provided
747 * buffers, as we dropped the selected one before retry.
748 */
749 if (io_do_buffer_select(req)) {
750 ret = io_import_iovec(ITER_DEST, req, iovec: &iovec, s, issue_flags);
751 if (unlikely(ret < 0))
752 return ret;
753 }
754
755 /*
756 * We come here from an earlier attempt, restore our state to
757 * match in case it doesn't. It's cheap enough that we don't
758 * need to make this conditional.
759 */
760 iov_iter_restore(i: &s->iter, state: &s->iter_state);
761 iovec = NULL;
762 }
763 ret = io_rw_init_file(req, FMODE_READ);
764 if (unlikely(ret)) {
765 kfree(objp: iovec);
766 return ret;
767 }
768 req->cqe.res = iov_iter_count(i: &s->iter);
769
770 if (force_nonblock) {
771 /* If the file doesn't support async, just async punt */
772 if (unlikely(!io_file_supports_nowait(req))) {
773 ret = io_setup_async_rw(req, iovec, s, force: true);
774 return ret ?: -EAGAIN;
775 }
776 kiocb->ki_flags |= IOCB_NOWAIT;
777 } else {
778 /* Ensure we clear previously set non-block flag */
779 kiocb->ki_flags &= ~IOCB_NOWAIT;
780 }
781
782 ppos = io_kiocb_update_pos(req);
783
784 ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
785 if (unlikely(ret)) {
786 kfree(objp: iovec);
787 return ret;
788 }
789
790 ret = io_iter_do_read(rw, iter: &s->iter);
791
792 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
793 req->flags &= ~REQ_F_REISSUE;
794 /*
795 * If we can poll, just do that. For a vectored read, we'll
796 * need to copy state first.
797 */
798 if (file_can_poll(file: req->file) && !io_issue_defs[req->opcode].vectored)
799 return -EAGAIN;
800 /* IOPOLL retry should happen for io-wq threads */
801 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
802 goto done;
803 /* no retry on NONBLOCK nor RWF_NOWAIT */
804 if (req->flags & REQ_F_NOWAIT)
805 goto done;
806 ret = 0;
807 } else if (ret == -EIOCBQUEUED) {
808 if (iovec)
809 kfree(objp: iovec);
810 return IOU_ISSUE_SKIP_COMPLETE;
811 } else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
812 (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) {
813 /* read all, failed, already did sync or don't want to retry */
814 goto done;
815 }
816
817 /*
818 * Don't depend on the iter state matching what was consumed, or being
819 * untouched in case of error. Restore it and we'll advance it
820 * manually if we need to.
821 */
822 iov_iter_restore(i: &s->iter, state: &s->iter_state);
823
824 ret2 = io_setup_async_rw(req, iovec, s, force: true);
825 iovec = NULL;
826 if (ret2) {
827 ret = ret > 0 ? ret : ret2;
828 goto done;
829 }
830
831 io = req->async_data;
832 s = &io->s;
833 /*
834 * Now use our persistent iterator and state, if we aren't already.
835 * We've restored and mapped the iter to match.
836 */
837
838 do {
839 /*
840 * We end up here because of a partial read, either from
841 * above or inside this loop. Advance the iter by the bytes
842 * that were consumed.
843 */
844 iov_iter_advance(i: &s->iter, bytes: ret);
845 if (!iov_iter_count(i: &s->iter))
846 break;
847 io->bytes_done += ret;
848 iov_iter_save_state(iter: &s->iter, state: &s->iter_state);
849
850 /* if we can retry, do so with the callbacks armed */
851 if (!io_rw_should_retry(req)) {
852 kiocb->ki_flags &= ~IOCB_WAITQ;
853 return -EAGAIN;
854 }
855
856 req->cqe.res = iov_iter_count(i: &s->iter);
857 /*
858 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
859 * we get -EIOCBQUEUED, then we'll get a notification when the
860 * desired page gets unlocked. We can also get a partial read
861 * here, and if we do, then just retry at the new offset.
862 */
863 ret = io_iter_do_read(rw, iter: &s->iter);
864 if (ret == -EIOCBQUEUED)
865 return IOU_ISSUE_SKIP_COMPLETE;
866 /* we got some bytes, but not all. retry. */
867 kiocb->ki_flags &= ~IOCB_WAITQ;
868 iov_iter_restore(i: &s->iter, state: &s->iter_state);
869 } while (ret > 0);
870done:
871 /* it's faster to check here then delegate to kfree */
872 if (iovec)
873 kfree(objp: iovec);
874 return ret;
875}
876
877int io_read(struct io_kiocb *req, unsigned int issue_flags)
878{
879 int ret;
880
881 ret = __io_read(req, issue_flags);
882 if (ret >= 0)
883 return kiocb_done(req, ret, issue_flags);
884
885 return ret;
886}
887
888int io_read_mshot(struct io_kiocb *req, unsigned int issue_flags)
889{
890 unsigned int cflags = 0;
891 int ret;
892
893 /*
894 * Multishot MUST be used on a pollable file
895 */
896 if (!file_can_poll(file: req->file))
897 return -EBADFD;
898
899 ret = __io_read(req, issue_flags);
900
901 /*
902 * If we get -EAGAIN, recycle our buffer and just let normal poll
903 * handling arm it.
904 */
905 if (ret == -EAGAIN) {
906 io_kbuf_recycle(req, issue_flags);
907 return -EAGAIN;
908 }
909
910 /*
911 * Any successful return value will keep the multishot read armed.
912 */
913 if (ret > 0) {
914 /*
915 * Put our buffer and post a CQE. If we fail to post a CQE, then
916 * jump to the termination path. This request is then done.
917 */
918 cflags = io_put_kbuf(req, issue_flags);
919
920 if (io_fill_cqe_req_aux(req,
921 defer: issue_flags & IO_URING_F_COMPLETE_DEFER,
922 res: ret, cflags: cflags | IORING_CQE_F_MORE)) {
923 if (issue_flags & IO_URING_F_MULTISHOT)
924 return IOU_ISSUE_SKIP_COMPLETE;
925 return -EAGAIN;
926 }
927 }
928
929 /*
930 * Either an error, or we've hit overflow posting the CQE. For any
931 * multishot request, hitting overflow will terminate it.
932 */
933 io_req_set_res(req, res: ret, cflags);
934 if (issue_flags & IO_URING_F_MULTISHOT)
935 return IOU_STOP_MULTISHOT;
936 return IOU_OK;
937}
938
939int io_write(struct io_kiocb *req, unsigned int issue_flags)
940{
941 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
942 struct io_rw_state __s, *s = &__s;
943 struct iovec *iovec;
944 struct kiocb *kiocb = &rw->kiocb;
945 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
946 ssize_t ret, ret2;
947 loff_t *ppos;
948
949 if (!req_has_async_data(req)) {
950 ret = io_import_iovec(ITER_SOURCE, req, iovec: &iovec, s, issue_flags);
951 if (unlikely(ret < 0))
952 return ret;
953 } else {
954 struct io_async_rw *io = req->async_data;
955
956 s = &io->s;
957 iov_iter_restore(i: &s->iter, state: &s->iter_state);
958 iovec = NULL;
959 }
960 ret = io_rw_init_file(req, FMODE_WRITE);
961 if (unlikely(ret)) {
962 kfree(objp: iovec);
963 return ret;
964 }
965 req->cqe.res = iov_iter_count(i: &s->iter);
966
967 if (force_nonblock) {
968 /* If the file doesn't support async, just async punt */
969 if (unlikely(!io_file_supports_nowait(req)))
970 goto copy_iov;
971
972 /* File path supports NOWAIT for non-direct_IO only for block devices. */
973 if (!(kiocb->ki_flags & IOCB_DIRECT) &&
974 !(kiocb->ki_filp->f_mode & FMODE_BUF_WASYNC) &&
975 (req->flags & REQ_F_ISREG))
976 goto copy_iov;
977
978 kiocb->ki_flags |= IOCB_NOWAIT;
979 } else {
980 /* Ensure we clear previously set non-block flag */
981 kiocb->ki_flags &= ~IOCB_NOWAIT;
982 }
983
984 ppos = io_kiocb_update_pos(req);
985
986 ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
987 if (unlikely(ret)) {
988 kfree(objp: iovec);
989 return ret;
990 }
991
992 if (req->flags & REQ_F_ISREG)
993 kiocb_start_write(iocb: kiocb);
994 kiocb->ki_flags |= IOCB_WRITE;
995
996 if (likely(req->file->f_op->write_iter))
997 ret2 = call_write_iter(file: req->file, kio: kiocb, iter: &s->iter);
998 else if (req->file->f_op->write)
999 ret2 = loop_rw_iter(WRITE, rw, iter: &s->iter);
1000 else
1001 ret2 = -EINVAL;
1002
1003 if (req->flags & REQ_F_REISSUE) {
1004 req->flags &= ~REQ_F_REISSUE;
1005 ret2 = -EAGAIN;
1006 }
1007
1008 /*
1009 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
1010 * retry them without IOCB_NOWAIT.
1011 */
1012 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
1013 ret2 = -EAGAIN;
1014 /* no retry on NONBLOCK nor RWF_NOWAIT */
1015 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
1016 goto done;
1017 if (!force_nonblock || ret2 != -EAGAIN) {
1018 /* IOPOLL retry should happen for io-wq threads */
1019 if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
1020 goto copy_iov;
1021
1022 if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) {
1023 struct io_async_rw *io;
1024
1025 trace_io_uring_short_write(ctx: req->ctx, fpos: kiocb->ki_pos - ret2,
1026 wanted: req->cqe.res, got: ret2);
1027
1028 /* This is a partial write. The file pos has already been
1029 * updated, setup the async struct to complete the request
1030 * in the worker. Also update bytes_done to account for
1031 * the bytes already written.
1032 */
1033 iov_iter_save_state(iter: &s->iter, state: &s->iter_state);
1034 ret = io_setup_async_rw(req, iovec, s, force: true);
1035
1036 io = req->async_data;
1037 if (io)
1038 io->bytes_done += ret2;
1039
1040 if (kiocb->ki_flags & IOCB_WRITE)
1041 io_req_end_write(req);
1042 return ret ? ret : -EAGAIN;
1043 }
1044done:
1045 ret = kiocb_done(req, ret: ret2, issue_flags);
1046 } else {
1047copy_iov:
1048 iov_iter_restore(i: &s->iter, state: &s->iter_state);
1049 ret = io_setup_async_rw(req, iovec, s, force: false);
1050 if (!ret) {
1051 if (kiocb->ki_flags & IOCB_WRITE)
1052 io_req_end_write(req);
1053 return -EAGAIN;
1054 }
1055 return ret;
1056 }
1057 /* it's reportedly faster than delegating the null check to kfree() */
1058 if (iovec)
1059 kfree(objp: iovec);
1060 return ret;
1061}
1062
1063void io_rw_fail(struct io_kiocb *req)
1064{
1065 int res;
1066
1067 res = io_fixup_rw_res(req, res: req->cqe.res);
1068 io_req_set_res(req, res, cflags: req->cqe.flags);
1069}
1070
1071int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
1072{
1073 struct io_wq_work_node *pos, *start, *prev;
1074 unsigned int poll_flags = 0;
1075 DEFINE_IO_COMP_BATCH(iob);
1076 int nr_events = 0;
1077
1078 /*
1079 * Only spin for completions if we don't have multiple devices hanging
1080 * off our complete list.
1081 */
1082 if (ctx->poll_multi_queue || force_nonspin)
1083 poll_flags |= BLK_POLL_ONESHOT;
1084
1085 wq_list_for_each(pos, start, &ctx->iopoll_list) {
1086 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1087 struct file *file = req->file;
1088 int ret;
1089
1090 /*
1091 * Move completed and retryable entries to our local lists.
1092 * If we find a request that requires polling, break out
1093 * and complete those lists first, if we have entries there.
1094 */
1095 if (READ_ONCE(req->iopoll_completed))
1096 break;
1097
1098 if (req->opcode == IORING_OP_URING_CMD) {
1099 struct io_uring_cmd *ioucmd;
1100
1101 ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
1102 ret = file->f_op->uring_cmd_iopoll(ioucmd, &iob,
1103 poll_flags);
1104 } else {
1105 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1106
1107 ret = file->f_op->iopoll(&rw->kiocb, &iob, poll_flags);
1108 }
1109 if (unlikely(ret < 0))
1110 return ret;
1111 else if (ret)
1112 poll_flags |= BLK_POLL_ONESHOT;
1113
1114 /* iopoll may have completed current req */
1115 if (!rq_list_empty(iob.req_list) ||
1116 READ_ONCE(req->iopoll_completed))
1117 break;
1118 }
1119
1120 if (!rq_list_empty(iob.req_list))
1121 iob.complete(&iob);
1122 else if (!pos)
1123 return 0;
1124
1125 prev = start;
1126 wq_list_for_each_resume(pos, prev) {
1127 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1128
1129 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
1130 if (!smp_load_acquire(&req->iopoll_completed))
1131 break;
1132 nr_events++;
1133 req->cqe.flags = io_put_kbuf(req, issue_flags: 0);
1134 }
1135 if (unlikely(!nr_events))
1136 return 0;
1137
1138 pos = start ? start->next : ctx->iopoll_list.first;
1139 wq_list_cut(list: &ctx->iopoll_list, last: prev, prev: start);
1140
1141 if (WARN_ON_ONCE(!wq_list_empty(&ctx->submit_state.compl_reqs)))
1142 return 0;
1143 ctx->submit_state.compl_reqs.first = pos;
1144 __io_submit_flush_completions(ctx);
1145 return nr_events;
1146}
1147

source code of linux/io_uring/rw.c