xref: /linux/io_uring/rw.c (revision a4a755c422242c27cb0f7900ac00cf33ac17b1ce)
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/cmd.h>
14 #include <linux/indirect_call_wrapper.h>
15 
16 #include <uapi/linux/io_uring.h>
17 
18 #include "io_uring.h"
19 #include "opdef.h"
20 #include "kbuf.h"
21 #include "rsrc.h"
22 #include "poll.h"
23 #include "rw.h"
24 
25 struct io_rw {
26 	/* NOTE: kiocb has the file as the first member, so don't do it here */
27 	struct kiocb			kiocb;
28 	u64				addr;
29 	u32				len;
30 	rwf_t				flags;
31 };
32 
33 static inline bool io_file_supports_nowait(struct io_kiocb *req)
34 {
35 	return req->flags & REQ_F_SUPPORT_NOWAIT;
36 }
37 
38 #ifdef CONFIG_COMPAT
39 static int io_iov_compat_buffer_select_prep(struct io_rw *rw)
40 {
41 	struct compat_iovec __user *uiov;
42 	compat_ssize_t clen;
43 
44 	uiov = u64_to_user_ptr(rw->addr);
45 	if (!access_ok(uiov, sizeof(*uiov)))
46 		return -EFAULT;
47 	if (__get_user(clen, &uiov->iov_len))
48 		return -EFAULT;
49 	if (clen < 0)
50 		return -EINVAL;
51 
52 	rw->len = clen;
53 	return 0;
54 }
55 #endif
56 
57 static int io_iov_buffer_select_prep(struct io_kiocb *req)
58 {
59 	struct iovec __user *uiov;
60 	struct iovec iov;
61 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
62 
63 	if (rw->len != 1)
64 		return -EINVAL;
65 
66 #ifdef CONFIG_COMPAT
67 	if (req->ctx->compat)
68 		return io_iov_compat_buffer_select_prep(rw);
69 #endif
70 
71 	uiov = u64_to_user_ptr(rw->addr);
72 	if (copy_from_user(&iov, uiov, sizeof(*uiov)))
73 		return -EFAULT;
74 	rw->len = iov.iov_len;
75 	return 0;
76 }
77 
78 int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
79 {
80 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
81 	unsigned ioprio;
82 	int ret;
83 
84 	rw->kiocb.ki_pos = READ_ONCE(sqe->off);
85 	/* used for fixed read/write too - just read unconditionally */
86 	req->buf_index = READ_ONCE(sqe->buf_index);
87 
88 	ioprio = READ_ONCE(sqe->ioprio);
89 	if (ioprio) {
90 		ret = ioprio_check_cap(ioprio);
91 		if (ret)
92 			return ret;
93 
94 		rw->kiocb.ki_ioprio = ioprio;
95 	} else {
96 		rw->kiocb.ki_ioprio = get_current_ioprio();
97 	}
98 	rw->kiocb.dio_complete = NULL;
99 
100 	rw->addr = READ_ONCE(sqe->addr);
101 	rw->len = READ_ONCE(sqe->len);
102 	rw->flags = READ_ONCE(sqe->rw_flags);
103 	return 0;
104 }
105 
106 int io_prep_rwv(struct io_kiocb *req, const struct io_uring_sqe *sqe)
107 {
108 	int ret;
109 
110 	ret = io_prep_rw(req, sqe);
111 	if (unlikely(ret))
112 		return ret;
113 
114 	/*
115 	 * Have to do this validation here, as this is in io_read() rw->len
116 	 * might have chanaged due to buffer selection
117 	 */
118 	if (req->flags & REQ_F_BUFFER_SELECT)
119 		return io_iov_buffer_select_prep(req);
120 
121 	return 0;
122 }
123 
124 int io_prep_rw_fixed(struct io_kiocb *req, const struct io_uring_sqe *sqe)
125 {
126 	struct io_ring_ctx *ctx = req->ctx;
127 	u16 index;
128 	int ret;
129 
130 	ret = io_prep_rw(req, sqe);
131 	if (unlikely(ret))
132 		return ret;
133 
134 	if (unlikely(req->buf_index >= ctx->nr_user_bufs))
135 		return -EFAULT;
136 	index = array_index_nospec(req->buf_index, ctx->nr_user_bufs);
137 	req->imu = ctx->user_bufs[index];
138 	io_req_set_rsrc_node(req, ctx, 0);
139 	return 0;
140 }
141 
142 /*
143  * Multishot read is prepared just like a normal read/write request, only
144  * difference is that we set the MULTISHOT flag.
145  */
146 int io_read_mshot_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
147 {
148 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
149 	int ret;
150 
151 	/* must be used with provided buffers */
152 	if (!(req->flags & REQ_F_BUFFER_SELECT))
153 		return -EINVAL;
154 
155 	ret = io_prep_rw(req, sqe);
156 	if (unlikely(ret))
157 		return ret;
158 
159 	if (rw->addr || rw->len)
160 		return -EINVAL;
161 
162 	req->flags |= REQ_F_APOLL_MULTISHOT;
163 	return 0;
164 }
165 
166 void io_readv_writev_cleanup(struct io_kiocb *req)
167 {
168 	struct io_async_rw *io = req->async_data;
169 
170 	kfree(io->free_iovec);
171 }
172 
173 static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req)
174 {
175 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
176 
177 	if (rw->kiocb.ki_pos != -1)
178 		return &rw->kiocb.ki_pos;
179 
180 	if (!(req->file->f_mode & FMODE_STREAM)) {
181 		req->flags |= REQ_F_CUR_POS;
182 		rw->kiocb.ki_pos = req->file->f_pos;
183 		return &rw->kiocb.ki_pos;
184 	}
185 
186 	rw->kiocb.ki_pos = 0;
187 	return NULL;
188 }
189 
190 static void io_req_task_queue_reissue(struct io_kiocb *req)
191 {
192 	req->io_task_work.func = io_queue_iowq;
193 	io_req_task_work_add(req);
194 }
195 
196 #ifdef CONFIG_BLOCK
197 static bool io_resubmit_prep(struct io_kiocb *req)
198 {
199 	struct io_async_rw *io = req->async_data;
200 
201 	if (!req_has_async_data(req))
202 		return !io_req_prep_async(req);
203 	iov_iter_restore(&io->s.iter, &io->s.iter_state);
204 	return true;
205 }
206 
207 static bool io_rw_should_reissue(struct io_kiocb *req)
208 {
209 	umode_t mode = file_inode(req->file)->i_mode;
210 	struct io_ring_ctx *ctx = req->ctx;
211 
212 	if (!S_ISBLK(mode) && !S_ISREG(mode))
213 		return false;
214 	if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
215 	    !(ctx->flags & IORING_SETUP_IOPOLL)))
216 		return false;
217 	/*
218 	 * If ref is dying, we might be running poll reap from the exit work.
219 	 * Don't attempt to reissue from that path, just let it fail with
220 	 * -EAGAIN.
221 	 */
222 	if (percpu_ref_is_dying(&ctx->refs))
223 		return false;
224 	/*
225 	 * Play it safe and assume not safe to re-import and reissue if we're
226 	 * not in the original thread group (or in task context).
227 	 */
228 	if (!same_thread_group(req->task, current) || !in_task())
229 		return false;
230 	return true;
231 }
232 #else
233 static bool io_resubmit_prep(struct io_kiocb *req)
234 {
235 	return false;
236 }
237 static bool io_rw_should_reissue(struct io_kiocb *req)
238 {
239 	return false;
240 }
241 #endif
242 
243 static void io_req_end_write(struct io_kiocb *req)
244 {
245 	if (req->flags & REQ_F_ISREG) {
246 		struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
247 
248 		kiocb_end_write(&rw->kiocb);
249 	}
250 }
251 
252 /*
253  * Trigger the notifications after having done some IO, and finish the write
254  * accounting, if any.
255  */
256 static void io_req_io_end(struct io_kiocb *req)
257 {
258 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
259 
260 	if (rw->kiocb.ki_flags & IOCB_WRITE) {
261 		io_req_end_write(req);
262 		fsnotify_modify(req->file);
263 	} else {
264 		fsnotify_access(req->file);
265 	}
266 }
267 
268 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
269 {
270 	if (unlikely(res != req->cqe.res)) {
271 		if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
272 		    io_rw_should_reissue(req)) {
273 			/*
274 			 * Reissue will start accounting again, finish the
275 			 * current cycle.
276 			 */
277 			io_req_io_end(req);
278 			req->flags |= REQ_F_REISSUE | REQ_F_BL_NO_RECYCLE;
279 			return true;
280 		}
281 		req_set_fail(req);
282 		req->cqe.res = res;
283 	}
284 	return false;
285 }
286 
287 static inline int io_fixup_rw_res(struct io_kiocb *req, long res)
288 {
289 	struct io_async_rw *io = req->async_data;
290 
291 	/* add previously done IO, if any */
292 	if (req_has_async_data(req) && io->bytes_done > 0) {
293 		if (res < 0)
294 			res = io->bytes_done;
295 		else
296 			res += io->bytes_done;
297 	}
298 	return res;
299 }
300 
301 void io_req_rw_complete(struct io_kiocb *req, struct io_tw_state *ts)
302 {
303 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
304 	struct kiocb *kiocb = &rw->kiocb;
305 
306 	if ((kiocb->ki_flags & IOCB_DIO_CALLER_COMP) && kiocb->dio_complete) {
307 		long res = kiocb->dio_complete(rw->kiocb.private);
308 
309 		io_req_set_res(req, io_fixup_rw_res(req, res), 0);
310 	}
311 
312 	io_req_io_end(req);
313 
314 	if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
315 		unsigned issue_flags = ts->locked ? 0 : IO_URING_F_UNLOCKED;
316 
317 		req->cqe.flags |= io_put_kbuf(req, issue_flags);
318 	}
319 	io_req_task_complete(req, ts);
320 }
321 
322 static void io_complete_rw(struct kiocb *kiocb, long res)
323 {
324 	struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
325 	struct io_kiocb *req = cmd_to_io_kiocb(rw);
326 
327 	if (!kiocb->dio_complete || !(kiocb->ki_flags & IOCB_DIO_CALLER_COMP)) {
328 		if (__io_complete_rw_common(req, res))
329 			return;
330 		io_req_set_res(req, io_fixup_rw_res(req, res), 0);
331 	}
332 	req->io_task_work.func = io_req_rw_complete;
333 	__io_req_task_work_add(req, IOU_F_TWQ_LAZY_WAKE);
334 }
335 
336 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
337 {
338 	struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
339 	struct io_kiocb *req = cmd_to_io_kiocb(rw);
340 
341 	if (kiocb->ki_flags & IOCB_WRITE)
342 		io_req_end_write(req);
343 	if (unlikely(res != req->cqe.res)) {
344 		if (res == -EAGAIN && io_rw_should_reissue(req)) {
345 			req->flags |= REQ_F_REISSUE | REQ_F_BL_NO_RECYCLE;
346 			return;
347 		}
348 		req->cqe.res = res;
349 	}
350 
351 	/* order with io_iopoll_complete() checking ->iopoll_completed */
352 	smp_store_release(&req->iopoll_completed, 1);
353 }
354 
355 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
356 {
357 	/* IO was queued async, completion will happen later */
358 	if (ret == -EIOCBQUEUED)
359 		return;
360 
361 	/* transform internal restart error codes */
362 	if (unlikely(ret < 0)) {
363 		switch (ret) {
364 		case -ERESTARTSYS:
365 		case -ERESTARTNOINTR:
366 		case -ERESTARTNOHAND:
367 		case -ERESTART_RESTARTBLOCK:
368 			/*
369 			 * We can't just restart the syscall, since previously
370 			 * submitted sqes may already be in progress. Just fail
371 			 * this IO with EINTR.
372 			 */
373 			ret = -EINTR;
374 			break;
375 		}
376 	}
377 
378 	INDIRECT_CALL_2(kiocb->ki_complete, io_complete_rw_iopoll,
379 			io_complete_rw, kiocb, ret);
380 }
381 
382 static int kiocb_done(struct io_kiocb *req, ssize_t ret,
383 		       unsigned int issue_flags)
384 {
385 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
386 	unsigned final_ret = io_fixup_rw_res(req, ret);
387 
388 	if (ret >= 0 && req->flags & REQ_F_CUR_POS)
389 		req->file->f_pos = rw->kiocb.ki_pos;
390 	if (ret >= 0 && (rw->kiocb.ki_complete == io_complete_rw)) {
391 		if (!__io_complete_rw_common(req, ret)) {
392 			/*
393 			 * Safe to call io_end from here as we're inline
394 			 * from the submission path.
395 			 */
396 			io_req_io_end(req);
397 			io_req_set_res(req, final_ret,
398 				       io_put_kbuf(req, issue_flags));
399 			return IOU_OK;
400 		}
401 	} else {
402 		io_rw_done(&rw->kiocb, ret);
403 	}
404 
405 	if (req->flags & REQ_F_REISSUE) {
406 		req->flags &= ~REQ_F_REISSUE;
407 		if (io_resubmit_prep(req))
408 			io_req_task_queue_reissue(req);
409 		else
410 			io_req_task_queue_fail(req, final_ret);
411 	}
412 	return IOU_ISSUE_SKIP_COMPLETE;
413 }
414 
415 static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req,
416 				       struct io_rw_state *s,
417 				       unsigned int issue_flags)
418 {
419 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
420 	struct iov_iter *iter = &s->iter;
421 	u8 opcode = req->opcode;
422 	struct iovec *iovec;
423 	void __user *buf;
424 	size_t sqe_len;
425 	ssize_t ret;
426 
427 	if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
428 		ret = io_import_fixed(ddir, iter, req->imu, rw->addr, rw->len);
429 		if (ret)
430 			return ERR_PTR(ret);
431 		return NULL;
432 	}
433 
434 	buf = u64_to_user_ptr(rw->addr);
435 	sqe_len = rw->len;
436 
437 	if (!io_issue_defs[opcode].vectored || req->flags & REQ_F_BUFFER_SELECT) {
438 		if (io_do_buffer_select(req)) {
439 			buf = io_buffer_select(req, &sqe_len, issue_flags);
440 			if (!buf)
441 				return ERR_PTR(-ENOBUFS);
442 			rw->addr = (unsigned long) buf;
443 			rw->len = sqe_len;
444 		}
445 
446 		ret = import_ubuf(ddir, buf, sqe_len, iter);
447 		if (ret)
448 			return ERR_PTR(ret);
449 		return NULL;
450 	}
451 
452 	iovec = s->fast_iov;
453 	ret = __import_iovec(ddir, buf, sqe_len, UIO_FASTIOV, &iovec, iter,
454 			      req->ctx->compat);
455 	if (unlikely(ret < 0))
456 		return ERR_PTR(ret);
457 	return iovec;
458 }
459 
460 static inline int io_import_iovec(int rw, struct io_kiocb *req,
461 				  struct iovec **iovec, struct io_rw_state *s,
462 				  unsigned int issue_flags)
463 {
464 	*iovec = __io_import_iovec(rw, req, s, issue_flags);
465 	if (IS_ERR(*iovec))
466 		return PTR_ERR(*iovec);
467 
468 	iov_iter_save_state(&s->iter, &s->iter_state);
469 	return 0;
470 }
471 
472 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
473 {
474 	return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
475 }
476 
477 /*
478  * For files that don't have ->read_iter() and ->write_iter(), handle them
479  * by looping over ->read() or ->write() manually.
480  */
481 static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter)
482 {
483 	struct kiocb *kiocb = &rw->kiocb;
484 	struct file *file = kiocb->ki_filp;
485 	ssize_t ret = 0;
486 	loff_t *ppos;
487 
488 	/*
489 	 * Don't support polled IO through this interface, and we can't
490 	 * support non-blocking either. For the latter, this just causes
491 	 * the kiocb to be handled from an async context.
492 	 */
493 	if (kiocb->ki_flags & IOCB_HIPRI)
494 		return -EOPNOTSUPP;
495 	if ((kiocb->ki_flags & IOCB_NOWAIT) &&
496 	    !(kiocb->ki_filp->f_flags & O_NONBLOCK))
497 		return -EAGAIN;
498 
499 	ppos = io_kiocb_ppos(kiocb);
500 
501 	while (iov_iter_count(iter)) {
502 		void __user *addr;
503 		size_t len;
504 		ssize_t nr;
505 
506 		if (iter_is_ubuf(iter)) {
507 			addr = iter->ubuf + iter->iov_offset;
508 			len = iov_iter_count(iter);
509 		} else if (!iov_iter_is_bvec(iter)) {
510 			addr = iter_iov_addr(iter);
511 			len = iter_iov_len(iter);
512 		} else {
513 			addr = u64_to_user_ptr(rw->addr);
514 			len = rw->len;
515 		}
516 
517 		if (ddir == READ)
518 			nr = file->f_op->read(file, addr, len, ppos);
519 		else
520 			nr = file->f_op->write(file, addr, len, ppos);
521 
522 		if (nr < 0) {
523 			if (!ret)
524 				ret = nr;
525 			break;
526 		}
527 		ret += nr;
528 		if (!iov_iter_is_bvec(iter)) {
529 			iov_iter_advance(iter, nr);
530 		} else {
531 			rw->addr += nr;
532 			rw->len -= nr;
533 			if (!rw->len)
534 				break;
535 		}
536 		if (nr != len)
537 			break;
538 	}
539 
540 	return ret;
541 }
542 
543 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
544 			  const struct iovec *fast_iov, struct iov_iter *iter)
545 {
546 	struct io_async_rw *io = req->async_data;
547 
548 	memcpy(&io->s.iter, iter, sizeof(*iter));
549 	io->free_iovec = iovec;
550 	io->bytes_done = 0;
551 	/* can only be fixed buffers, no need to do anything */
552 	if (iov_iter_is_bvec(iter) || iter_is_ubuf(iter))
553 		return;
554 	if (!iovec) {
555 		unsigned iov_off = 0;
556 
557 		io->s.iter.__iov = io->s.fast_iov;
558 		if (iter->__iov != fast_iov) {
559 			iov_off = iter_iov(iter) - fast_iov;
560 			io->s.iter.__iov += iov_off;
561 		}
562 		if (io->s.fast_iov != fast_iov)
563 			memcpy(io->s.fast_iov + iov_off, fast_iov + iov_off,
564 			       sizeof(struct iovec) * iter->nr_segs);
565 	} else {
566 		req->flags |= REQ_F_NEED_CLEANUP;
567 	}
568 }
569 
570 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
571 			     struct io_rw_state *s, bool force)
572 {
573 	if (!force && !io_cold_defs[req->opcode].prep_async)
574 		return 0;
575 	/* opcode type doesn't need async data */
576 	if (!io_cold_defs[req->opcode].async_size)
577 		return 0;
578 	if (!req_has_async_data(req)) {
579 		struct io_async_rw *iorw;
580 
581 		if (io_alloc_async_data(req)) {
582 			kfree(iovec);
583 			return -ENOMEM;
584 		}
585 
586 		io_req_map_rw(req, iovec, s->fast_iov, &s->iter);
587 		iorw = req->async_data;
588 		/* we've copied and mapped the iter, ensure state is saved */
589 		iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state);
590 	}
591 	return 0;
592 }
593 
594 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
595 {
596 	struct io_async_rw *iorw = req->async_data;
597 	struct iovec *iov;
598 	int ret;
599 
600 	iorw->bytes_done = 0;
601 	iorw->free_iovec = NULL;
602 
603 	/* submission path, ->uring_lock should already be taken */
604 	ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
605 	if (unlikely(ret < 0))
606 		return ret;
607 
608 	if (iov) {
609 		iorw->free_iovec = iov;
610 		req->flags |= REQ_F_NEED_CLEANUP;
611 	}
612 
613 	return 0;
614 }
615 
616 int io_readv_prep_async(struct io_kiocb *req)
617 {
618 	return io_rw_prep_async(req, ITER_DEST);
619 }
620 
621 int io_writev_prep_async(struct io_kiocb *req)
622 {
623 	return io_rw_prep_async(req, ITER_SOURCE);
624 }
625 
626 /*
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
634  * slow path.
635  */
636 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
637 			     int sync, void *arg)
638 {
639 	struct wait_page_queue *wpq;
640 	struct io_kiocb *req = wait->private;
641 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
642 	struct wait_page_key *key = arg;
643 
644 	wpq = container_of(wait, struct wait_page_queue, wait);
645 
646 	if (!wake_page_match(wpq, key))
647 		return 0;
648 
649 	rw->kiocb.ki_flags &= ~IOCB_WAITQ;
650 	list_del_init(&wait->entry);
651 	io_req_task_queue(req);
652 	return 1;
653 }
654 
655 /*
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.
666  */
667 static bool io_rw_should_retry(struct io_kiocb *req)
668 {
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, struct io_rw);
672 	struct kiocb *kiocb = &rw->kiocb;
673 
674 	/* never retry for NOWAIT, we just complete with -EAGAIN */
675 	if (req->flags & REQ_F_NOWAIT)
676 		return false;
677 
678 	/* Only for buffered IO */
679 	if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
680 		return false;
681 
682 	/*
683 	 * just use poll if we can, and don't attempt if the fs doesn't
684 	 * support callback based unlocks
685 	 */
686 	if (io_file_can_poll(req) || !(req->file->f_mode & FMODE_BUF_RASYNC))
687 		return false;
688 
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;
696 	return true;
697 }
698 
699 static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
700 {
701 	struct file *file = rw->kiocb.ki_filp;
702 
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);
707 	else
708 		return -EINVAL;
709 }
710 
711 static bool need_complete_io(struct io_kiocb *req)
712 {
713 	return req->flags & REQ_F_ISREG ||
714 		S_ISBLK(file_inode(req->file)->i_mode);
715 }
716 
717 static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
718 {
719 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
720 	struct kiocb *kiocb = &rw->kiocb;
721 	struct io_ring_ctx *ctx = req->ctx;
722 	struct file *file = req->file;
723 	int ret;
724 
725 	if (unlikely(!(file->f_mode & mode)))
726 		return -EBADF;
727 
728 	if (!(req->flags & REQ_F_FIXED_FILE))
729 		req->flags |= io_file_get_flags(file);
730 
731 	kiocb->ki_flags = file->f_iocb_flags;
732 	ret = kiocb_set_rw_flags(kiocb, rw->flags);
733 	if (unlikely(ret))
734 		return ret;
735 	kiocb->ki_flags |= IOCB_ALLOC_CACHE;
736 
737 	/*
738 	 * If the file is marked O_NONBLOCK, still allow retry for it if it
739 	 * supports async. Otherwise it's impossible to use O_NONBLOCK files
740 	 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
741 	 */
742 	if ((kiocb->ki_flags & IOCB_NOWAIT) ||
743 	    ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
744 		req->flags |= REQ_F_NOWAIT;
745 
746 	if (ctx->flags & IORING_SETUP_IOPOLL) {
747 		if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
748 			return -EOPNOTSUPP;
749 
750 		kiocb->private = NULL;
751 		kiocb->ki_flags |= IOCB_HIPRI;
752 		kiocb->ki_complete = io_complete_rw_iopoll;
753 		req->iopoll_completed = 0;
754 	} else {
755 		if (kiocb->ki_flags & IOCB_HIPRI)
756 			return -EINVAL;
757 		kiocb->ki_complete = io_complete_rw;
758 	}
759 
760 	return 0;
761 }
762 
763 static int __io_read(struct io_kiocb *req, unsigned int issue_flags)
764 {
765 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
766 	struct io_rw_state __s, *s = &__s;
767 	struct iovec *iovec;
768 	struct kiocb *kiocb = &rw->kiocb;
769 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
770 	struct io_async_rw *io;
771 	ssize_t ret, ret2;
772 	loff_t *ppos;
773 
774 	if (!req_has_async_data(req)) {
775 		ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
776 		if (unlikely(ret < 0))
777 			return ret;
778 	} else {
779 		io = req->async_data;
780 		s = &io->s;
781 
782 		/*
783 		 * Safe and required to re-import if we're using provided
784 		 * buffers, as we dropped the selected one before retry.
785 		 */
786 		if (io_do_buffer_select(req)) {
787 			ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
788 			if (unlikely(ret < 0))
789 				return ret;
790 		}
791 
792 		/*
793 		 * We come here from an earlier attempt, restore our state to
794 		 * match in case it doesn't. It's cheap enough that we don't
795 		 * need to make this conditional.
796 		 */
797 		iov_iter_restore(&s->iter, &s->iter_state);
798 		iovec = NULL;
799 	}
800 	ret = io_rw_init_file(req, FMODE_READ);
801 	if (unlikely(ret)) {
802 		kfree(iovec);
803 		return ret;
804 	}
805 	req->cqe.res = iov_iter_count(&s->iter);
806 
807 	if (force_nonblock) {
808 		/* If the file doesn't support async, just async punt */
809 		if (unlikely(!io_file_supports_nowait(req))) {
810 			ret = io_setup_async_rw(req, iovec, s, true);
811 			return ret ?: -EAGAIN;
812 		}
813 		kiocb->ki_flags |= IOCB_NOWAIT;
814 	} else {
815 		/* Ensure we clear previously set non-block flag */
816 		kiocb->ki_flags &= ~IOCB_NOWAIT;
817 	}
818 
819 	ppos = io_kiocb_update_pos(req);
820 
821 	ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
822 	if (unlikely(ret)) {
823 		kfree(iovec);
824 		return ret;
825 	}
826 
827 	ret = io_iter_do_read(rw, &s->iter);
828 
829 	if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
830 		req->flags &= ~REQ_F_REISSUE;
831 		/*
832 		 * If we can poll, just do that. For a vectored read, we'll
833 		 * need to copy state first.
834 		 */
835 		if (io_file_can_poll(req) && !io_issue_defs[req->opcode].vectored)
836 			return -EAGAIN;
837 		/* IOPOLL retry should happen for io-wq threads */
838 		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
839 			goto done;
840 		/* no retry on NONBLOCK nor RWF_NOWAIT */
841 		if (req->flags & REQ_F_NOWAIT)
842 			goto done;
843 		ret = 0;
844 	} else if (ret == -EIOCBQUEUED) {
845 		if (iovec)
846 			kfree(iovec);
847 		return IOU_ISSUE_SKIP_COMPLETE;
848 	} else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
849 		   (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) {
850 		/* read all, failed, already did sync or don't want to retry */
851 		goto done;
852 	}
853 
854 	/*
855 	 * Don't depend on the iter state matching what was consumed, or being
856 	 * untouched in case of error. Restore it and we'll advance it
857 	 * manually if we need to.
858 	 */
859 	iov_iter_restore(&s->iter, &s->iter_state);
860 
861 	ret2 = io_setup_async_rw(req, iovec, s, true);
862 	iovec = NULL;
863 	if (ret2) {
864 		ret = ret > 0 ? ret : ret2;
865 		goto done;
866 	}
867 
868 	io = req->async_data;
869 	s = &io->s;
870 	/*
871 	 * Now use our persistent iterator and state, if we aren't already.
872 	 * We've restored and mapped the iter to match.
873 	 */
874 
875 	do {
876 		/*
877 		 * We end up here because of a partial read, either from
878 		 * above or inside this loop. Advance the iter by the bytes
879 		 * that were consumed.
880 		 */
881 		iov_iter_advance(&s->iter, ret);
882 		if (!iov_iter_count(&s->iter))
883 			break;
884 		io->bytes_done += ret;
885 		iov_iter_save_state(&s->iter, &s->iter_state);
886 
887 		/* if we can retry, do so with the callbacks armed */
888 		if (!io_rw_should_retry(req)) {
889 			kiocb->ki_flags &= ~IOCB_WAITQ;
890 			return -EAGAIN;
891 		}
892 
893 		req->cqe.res = iov_iter_count(&s->iter);
894 		/*
895 		 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
896 		 * we get -EIOCBQUEUED, then we'll get a notification when the
897 		 * desired page gets unlocked. We can also get a partial read
898 		 * here, and if we do, then just retry at the new offset.
899 		 */
900 		ret = io_iter_do_read(rw, &s->iter);
901 		if (ret == -EIOCBQUEUED)
902 			return IOU_ISSUE_SKIP_COMPLETE;
903 		/* we got some bytes, but not all. retry. */
904 		kiocb->ki_flags &= ~IOCB_WAITQ;
905 		iov_iter_restore(&s->iter, &s->iter_state);
906 	} while (ret > 0);
907 done:
908 	/* it's faster to check here then delegate to kfree */
909 	if (iovec)
910 		kfree(iovec);
911 	return ret;
912 }
913 
914 int io_read(struct io_kiocb *req, unsigned int issue_flags)
915 {
916 	int ret;
917 
918 	ret = __io_read(req, issue_flags);
919 	if (ret >= 0)
920 		return kiocb_done(req, ret, issue_flags);
921 
922 	return ret;
923 }
924 
925 int io_read_mshot(struct io_kiocb *req, unsigned int issue_flags)
926 {
927 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
928 	unsigned int cflags = 0;
929 	int ret;
930 
931 	/*
932 	 * Multishot MUST be used on a pollable file
933 	 */
934 	if (!io_file_can_poll(req))
935 		return -EBADFD;
936 
937 	ret = __io_read(req, issue_flags);
938 
939 	/*
940 	 * If the file doesn't support proper NOWAIT, then disable multishot
941 	 * and stay in single shot mode.
942 	 */
943 	if (!io_file_supports_nowait(req))
944 		req->flags &= ~REQ_F_APOLL_MULTISHOT;
945 
946 	/*
947 	 * If we get -EAGAIN, recycle our buffer and just let normal poll
948 	 * handling arm it.
949 	 */
950 	if (ret == -EAGAIN) {
951 		/*
952 		 * Reset rw->len to 0 again to avoid clamping future mshot
953 		 * reads, in case the buffer size varies.
954 		 */
955 		if (io_kbuf_recycle(req, issue_flags))
956 			rw->len = 0;
957 		if (issue_flags & IO_URING_F_MULTISHOT)
958 			return IOU_ISSUE_SKIP_COMPLETE;
959 		return -EAGAIN;
960 	}
961 
962 	/*
963 	 * Any successful return value will keep the multishot read armed.
964 	 */
965 	if (ret > 0 && req->flags & REQ_F_APOLL_MULTISHOT) {
966 		/*
967 		 * Put our buffer and post a CQE. If we fail to post a CQE, then
968 		 * jump to the termination path. This request is then done.
969 		 */
970 		cflags = io_put_kbuf(req, issue_flags);
971 		rw->len = 0; /* similarly to above, reset len to 0 */
972 
973 		if (io_fill_cqe_req_aux(req,
974 					issue_flags & IO_URING_F_COMPLETE_DEFER,
975 					ret, cflags | IORING_CQE_F_MORE)) {
976 			if (issue_flags & IO_URING_F_MULTISHOT) {
977 				/*
978 				 * Force retry, as we might have more data to
979 				 * be read and otherwise it won't get retried
980 				 * until (if ever) another poll is triggered.
981 				 */
982 				io_poll_multishot_retry(req);
983 				return IOU_ISSUE_SKIP_COMPLETE;
984 			}
985 			return -EAGAIN;
986 		}
987 	}
988 
989 	/*
990 	 * Either an error, or we've hit overflow posting the CQE. For any
991 	 * multishot request, hitting overflow will terminate it.
992 	 */
993 	io_req_set_res(req, ret, cflags);
994 	if (issue_flags & IO_URING_F_MULTISHOT)
995 		return IOU_STOP_MULTISHOT;
996 	return IOU_OK;
997 }
998 
999 int io_write(struct io_kiocb *req, unsigned int issue_flags)
1000 {
1001 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1002 	struct io_rw_state __s, *s = &__s;
1003 	struct iovec *iovec;
1004 	struct kiocb *kiocb = &rw->kiocb;
1005 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1006 	ssize_t ret, ret2;
1007 	loff_t *ppos;
1008 
1009 	if (!req_has_async_data(req)) {
1010 		ret = io_import_iovec(ITER_SOURCE, req, &iovec, s, issue_flags);
1011 		if (unlikely(ret < 0))
1012 			return ret;
1013 	} else {
1014 		struct io_async_rw *io = req->async_data;
1015 
1016 		s = &io->s;
1017 		iov_iter_restore(&s->iter, &s->iter_state);
1018 		iovec = NULL;
1019 	}
1020 	ret = io_rw_init_file(req, FMODE_WRITE);
1021 	if (unlikely(ret)) {
1022 		kfree(iovec);
1023 		return ret;
1024 	}
1025 	req->cqe.res = iov_iter_count(&s->iter);
1026 
1027 	if (force_nonblock) {
1028 		/* If the file doesn't support async, just async punt */
1029 		if (unlikely(!io_file_supports_nowait(req)))
1030 			goto copy_iov;
1031 
1032 		/* File path supports NOWAIT for non-direct_IO only for block devices. */
1033 		if (!(kiocb->ki_flags & IOCB_DIRECT) &&
1034 			!(kiocb->ki_filp->f_mode & FMODE_BUF_WASYNC) &&
1035 			(req->flags & REQ_F_ISREG))
1036 			goto copy_iov;
1037 
1038 		kiocb->ki_flags |= IOCB_NOWAIT;
1039 	} else {
1040 		/* Ensure we clear previously set non-block flag */
1041 		kiocb->ki_flags &= ~IOCB_NOWAIT;
1042 	}
1043 
1044 	ppos = io_kiocb_update_pos(req);
1045 
1046 	ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
1047 	if (unlikely(ret)) {
1048 		kfree(iovec);
1049 		return ret;
1050 	}
1051 
1052 	if (req->flags & REQ_F_ISREG)
1053 		kiocb_start_write(kiocb);
1054 	kiocb->ki_flags |= IOCB_WRITE;
1055 
1056 	if (likely(req->file->f_op->write_iter))
1057 		ret2 = call_write_iter(req->file, kiocb, &s->iter);
1058 	else if (req->file->f_op->write)
1059 		ret2 = loop_rw_iter(WRITE, rw, &s->iter);
1060 	else
1061 		ret2 = -EINVAL;
1062 
1063 	if (req->flags & REQ_F_REISSUE) {
1064 		req->flags &= ~REQ_F_REISSUE;
1065 		ret2 = -EAGAIN;
1066 	}
1067 
1068 	/*
1069 	 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
1070 	 * retry them without IOCB_NOWAIT.
1071 	 */
1072 	if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
1073 		ret2 = -EAGAIN;
1074 	/* no retry on NONBLOCK nor RWF_NOWAIT */
1075 	if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
1076 		goto done;
1077 	if (!force_nonblock || ret2 != -EAGAIN) {
1078 		/* IOPOLL retry should happen for io-wq threads */
1079 		if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
1080 			goto copy_iov;
1081 
1082 		if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) {
1083 			struct io_async_rw *io;
1084 
1085 			trace_io_uring_short_write(req->ctx, kiocb->ki_pos - ret2,
1086 						req->cqe.res, ret2);
1087 
1088 			/* This is a partial write. The file pos has already been
1089 			 * updated, setup the async struct to complete the request
1090 			 * in the worker. Also update bytes_done to account for
1091 			 * the bytes already written.
1092 			 */
1093 			iov_iter_save_state(&s->iter, &s->iter_state);
1094 			ret = io_setup_async_rw(req, iovec, s, true);
1095 
1096 			io = req->async_data;
1097 			if (io)
1098 				io->bytes_done += ret2;
1099 
1100 			if (kiocb->ki_flags & IOCB_WRITE)
1101 				io_req_end_write(req);
1102 			return ret ? ret : -EAGAIN;
1103 		}
1104 done:
1105 		ret = kiocb_done(req, ret2, issue_flags);
1106 	} else {
1107 copy_iov:
1108 		iov_iter_restore(&s->iter, &s->iter_state);
1109 		ret = io_setup_async_rw(req, iovec, s, false);
1110 		if (!ret) {
1111 			if (kiocb->ki_flags & IOCB_WRITE)
1112 				io_req_end_write(req);
1113 			return -EAGAIN;
1114 		}
1115 		return ret;
1116 	}
1117 	/* it's reportedly faster than delegating the null check to kfree() */
1118 	if (iovec)
1119 		kfree(iovec);
1120 	return ret;
1121 }
1122 
1123 void io_rw_fail(struct io_kiocb *req)
1124 {
1125 	int res;
1126 
1127 	res = io_fixup_rw_res(req, req->cqe.res);
1128 	io_req_set_res(req, res, req->cqe.flags);
1129 }
1130 
1131 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
1132 {
1133 	struct io_wq_work_node *pos, *start, *prev;
1134 	unsigned int poll_flags = 0;
1135 	DEFINE_IO_COMP_BATCH(iob);
1136 	int nr_events = 0;
1137 
1138 	/*
1139 	 * Only spin for completions if we don't have multiple devices hanging
1140 	 * off our complete list.
1141 	 */
1142 	if (ctx->poll_multi_queue || force_nonspin)
1143 		poll_flags |= BLK_POLL_ONESHOT;
1144 
1145 	wq_list_for_each(pos, start, &ctx->iopoll_list) {
1146 		struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1147 		struct file *file = req->file;
1148 		int ret;
1149 
1150 		/*
1151 		 * Move completed and retryable entries to our local lists.
1152 		 * If we find a request that requires polling, break out
1153 		 * and complete those lists first, if we have entries there.
1154 		 */
1155 		if (READ_ONCE(req->iopoll_completed))
1156 			break;
1157 
1158 		if (req->opcode == IORING_OP_URING_CMD) {
1159 			struct io_uring_cmd *ioucmd;
1160 
1161 			ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
1162 			ret = file->f_op->uring_cmd_iopoll(ioucmd, &iob,
1163 								poll_flags);
1164 		} else {
1165 			struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1166 
1167 			ret = file->f_op->iopoll(&rw->kiocb, &iob, poll_flags);
1168 		}
1169 		if (unlikely(ret < 0))
1170 			return ret;
1171 		else if (ret)
1172 			poll_flags |= BLK_POLL_ONESHOT;
1173 
1174 		/* iopoll may have completed current req */
1175 		if (!rq_list_empty(iob.req_list) ||
1176 		    READ_ONCE(req->iopoll_completed))
1177 			break;
1178 	}
1179 
1180 	if (!rq_list_empty(iob.req_list))
1181 		iob.complete(&iob);
1182 	else if (!pos)
1183 		return 0;
1184 
1185 	prev = start;
1186 	wq_list_for_each_resume(pos, prev) {
1187 		struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1188 
1189 		/* order with io_complete_rw_iopoll(), e.g. ->result updates */
1190 		if (!smp_load_acquire(&req->iopoll_completed))
1191 			break;
1192 		nr_events++;
1193 		req->cqe.flags = io_put_kbuf(req, 0);
1194 	}
1195 	if (unlikely(!nr_events))
1196 		return 0;
1197 
1198 	pos = start ? start->next : ctx->iopoll_list.first;
1199 	wq_list_cut(&ctx->iopoll_list, prev, start);
1200 
1201 	if (WARN_ON_ONCE(!wq_list_empty(&ctx->submit_state.compl_reqs)))
1202 		return 0;
1203 	ctx->submit_state.compl_reqs.first = pos;
1204 	__io_submit_flush_completions(ctx);
1205 	return nr_events;
1206 }
1207