xref: /linux/io_uring/rw.c (revision c65c3f3a2cbf21ed429d9b9c725bdb5dc6abf4cf)
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 
23 struct 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 
31 static inline bool io_file_supports_nowait(struct io_kiocb *req)
32 {
33 	return req->flags & REQ_F_SUPPORT_NOWAIT;
34 }
35 
36 int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe)
37 {
38 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
39 	unsigned ioprio;
40 	int ret;
41 
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);
45 
46 	if (req->opcode == IORING_OP_READ_FIXED ||
47 	    req->opcode == IORING_OP_WRITE_FIXED) {
48 		struct io_ring_ctx *ctx = req->ctx;
49 		u16 index;
50 
51 		if (unlikely(req->buf_index >= ctx->nr_user_bufs))
52 			return -EFAULT;
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);
56 	}
57 
58 	ioprio = READ_ONCE(sqe->ioprio);
59 	if (ioprio) {
60 		ret = ioprio_check_cap(ioprio);
61 		if (ret)
62 			return ret;
63 
64 		rw->kiocb.ki_ioprio = ioprio;
65 	} else {
66 		rw->kiocb.ki_ioprio = get_current_ioprio();
67 	}
68 
69 	rw->addr = READ_ONCE(sqe->addr);
70 	rw->len = READ_ONCE(sqe->len);
71 	rw->flags = READ_ONCE(sqe->rw_flags);
72 	return 0;
73 }
74 
75 void io_readv_writev_cleanup(struct io_kiocb *req)
76 {
77 	struct io_async_rw *io = req->async_data;
78 
79 	kfree(io->free_iovec);
80 }
81 
82 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
83 {
84 	switch (ret) {
85 	case -EIOCBQUEUED:
86 		break;
87 	case -ERESTARTSYS:
88 	case -ERESTARTNOINTR:
89 	case -ERESTARTNOHAND:
90 	case -ERESTART_RESTARTBLOCK:
91 		/*
92 		 * We can't just restart the syscall, since previously
93 		 * submitted sqes may already be in progress. Just fail this
94 		 * IO with EINTR.
95 		 */
96 		ret = -EINTR;
97 		fallthrough;
98 	default:
99 		kiocb->ki_complete(kiocb, ret);
100 	}
101 }
102 
103 static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req)
104 {
105 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
106 
107 	if (rw->kiocb.ki_pos != -1)
108 		return &rw->kiocb.ki_pos;
109 
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;
114 	}
115 
116 	rw->kiocb.ki_pos = 0;
117 	return NULL;
118 }
119 
120 static void io_req_task_queue_reissue(struct io_kiocb *req)
121 {
122 	req->io_task_work.func = io_queue_iowq;
123 	io_req_task_work_add(req);
124 }
125 
126 #ifdef CONFIG_BLOCK
127 static bool io_resubmit_prep(struct io_kiocb *req)
128 {
129 	struct io_async_rw *io = req->async_data;
130 
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);
134 	return true;
135 }
136 
137 static bool io_rw_should_reissue(struct io_kiocb *req)
138 {
139 	umode_t mode = file_inode(req->file)->i_mode;
140 	struct io_ring_ctx *ctx = req->ctx;
141 
142 	if (!S_ISBLK(mode) && !S_ISREG(mode))
143 		return false;
144 	if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
145 	    !(ctx->flags & IORING_SETUP_IOPOLL)))
146 		return false;
147 	/*
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
150 	 * -EAGAIN.
151 	 */
152 	if (percpu_ref_is_dying(&ctx->refs))
153 		return false;
154 	/*
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).
157 	 */
158 	if (!same_thread_group(req->task, current) || !in_task())
159 		return false;
160 	return true;
161 }
162 #else
163 static bool io_resubmit_prep(struct io_kiocb *req)
164 {
165 	return false;
166 }
167 static bool io_rw_should_reissue(struct io_kiocb *req)
168 {
169 	return false;
170 }
171 #endif
172 
173 static void kiocb_end_write(struct io_kiocb *req)
174 {
175 	/*
176 	 * Tell lockdep we inherited freeze protection from submission
177 	 * thread.
178 	 */
179 	if (req->flags & REQ_F_ISREG) {
180 		struct super_block *sb = file_inode(req->file)->i_sb;
181 
182 		__sb_writers_acquired(sb, SB_FREEZE_WRITE);
183 		sb_end_write(sb);
184 	}
185 }
186 
187 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
188 {
189 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
190 
191 	if (rw->kiocb.ki_flags & IOCB_WRITE) {
192 		kiocb_end_write(req);
193 		fsnotify_modify(req->file);
194 	} else {
195 		fsnotify_access(req->file);
196 	}
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;
201 			return true;
202 		}
203 		req_set_fail(req);
204 		req->cqe.res = res;
205 	}
206 	return false;
207 }
208 
209 static void io_complete_rw(struct kiocb *kiocb, long res)
210 {
211 	struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
212 	struct io_kiocb *req = cmd_to_io_kiocb(rw);
213 
214 	if (__io_complete_rw_common(req, res))
215 		return;
216 	io_req_set_res(req, res, 0);
217 	req->io_task_work.func = io_req_task_complete;
218 	io_req_task_work_add(req);
219 }
220 
221 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
222 {
223 	struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
224 	struct io_kiocb *req = cmd_to_io_kiocb(rw);
225 
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;
231 			return;
232 		}
233 		req->cqe.res = res;
234 	}
235 
236 	/* order with io_iopoll_complete() checking ->iopoll_completed */
237 	smp_store_release(&req->iopoll_completed, 1);
238 }
239 
240 static int kiocb_done(struct io_kiocb *req, ssize_t ret,
241 		       unsigned int issue_flags)
242 {
243 	struct io_async_rw *io = req->async_data;
244 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
245 
246 	/* add previously done IO, if any */
247 	if (req_has_async_data(req) && io->bytes_done > 0) {
248 		if (ret < 0)
249 			ret = io->bytes_done;
250 		else
251 			ret += io->bytes_done;
252 	}
253 
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));
260 			return IOU_OK;
261 		}
262 	} else {
263 		io_rw_done(&rw->kiocb, ret);
264 	}
265 
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);
270 		else
271 			io_req_task_queue_fail(req, ret);
272 	}
273 	return IOU_ISSUE_SKIP_COMPLETE;
274 }
275 
276 #ifdef CONFIG_COMPAT
277 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
278 				unsigned int issue_flags)
279 {
280 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
281 	struct compat_iovec __user *uiov;
282 	compat_ssize_t clen;
283 	void __user *buf;
284 	size_t len;
285 
286 	uiov = u64_to_user_ptr(rw->addr);
287 	if (!access_ok(uiov, sizeof(*uiov)))
288 		return -EFAULT;
289 	if (__get_user(clen, &uiov->iov_len))
290 		return -EFAULT;
291 	if (clen < 0)
292 		return -EINVAL;
293 
294 	len = clen;
295 	buf = io_buffer_select(req, &len, issue_flags);
296 	if (!buf)
297 		return -ENOBUFS;
298 	rw->addr = (unsigned long) buf;
299 	iov[0].iov_base = buf;
300 	rw->len = iov[0].iov_len = (compat_size_t) len;
301 	return 0;
302 }
303 #endif
304 
305 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
306 				      unsigned int issue_flags)
307 {
308 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
309 	struct iovec __user *uiov = u64_to_user_ptr(rw->addr);
310 	void __user *buf;
311 	ssize_t len;
312 
313 	if (copy_from_user(iov, uiov, sizeof(*uiov)))
314 		return -EFAULT;
315 
316 	len = iov[0].iov_len;
317 	if (len < 0)
318 		return -EINVAL;
319 	buf = io_buffer_select(req, &len, issue_flags);
320 	if (!buf)
321 		return -ENOBUFS;
322 	rw->addr = (unsigned long) buf;
323 	iov[0].iov_base = buf;
324 	rw->len = iov[0].iov_len = len;
325 	return 0;
326 }
327 
328 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
329 				    unsigned int issue_flags)
330 {
331 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
332 
333 	if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
334 		iov[0].iov_base = u64_to_user_ptr(rw->addr);
335 		iov[0].iov_len = rw->len;
336 		return 0;
337 	}
338 	if (rw->len != 1)
339 		return -EINVAL;
340 
341 #ifdef CONFIG_COMPAT
342 	if (req->ctx->compat)
343 		return io_compat_import(req, iov, issue_flags);
344 #endif
345 
346 	return __io_iov_buffer_select(req, iov, issue_flags);
347 }
348 
349 static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req,
350 				       struct io_rw_state *s,
351 				       unsigned int issue_flags)
352 {
353 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
354 	struct iov_iter *iter = &s->iter;
355 	u8 opcode = req->opcode;
356 	struct iovec *iovec;
357 	void __user *buf;
358 	size_t sqe_len;
359 	ssize_t ret;
360 
361 	if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
362 		ret = io_import_fixed(ddir, iter, req->imu, rw->addr, rw->len);
363 		if (ret)
364 			return ERR_PTR(ret);
365 		return NULL;
366 	}
367 
368 	buf = u64_to_user_ptr(rw->addr);
369 	sqe_len = rw->len;
370 
371 	if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
372 		if (io_do_buffer_select(req)) {
373 			buf = io_buffer_select(req, &sqe_len, issue_flags);
374 			if (!buf)
375 				return ERR_PTR(-ENOBUFS);
376 			rw->addr = (unsigned long) buf;
377 			rw->len = sqe_len;
378 		}
379 
380 		ret = import_single_range(ddir, buf, sqe_len, s->fast_iov, iter);
381 		if (ret)
382 			return ERR_PTR(ret);
383 		return NULL;
384 	}
385 
386 	iovec = s->fast_iov;
387 	if (req->flags & REQ_F_BUFFER_SELECT) {
388 		ret = io_iov_buffer_select(req, iovec, issue_flags);
389 		if (ret)
390 			return ERR_PTR(ret);
391 		iov_iter_init(iter, ddir, iovec, 1, iovec->iov_len);
392 		return NULL;
393 	}
394 
395 	ret = __import_iovec(ddir, buf, sqe_len, UIO_FASTIOV, &iovec, iter,
396 			      req->ctx->compat);
397 	if (unlikely(ret < 0))
398 		return ERR_PTR(ret);
399 	return iovec;
400 }
401 
402 static inline int io_import_iovec(int rw, struct io_kiocb *req,
403 				  struct iovec **iovec, struct io_rw_state *s,
404 				  unsigned int issue_flags)
405 {
406 	*iovec = __io_import_iovec(rw, req, s, issue_flags);
407 	if (unlikely(IS_ERR(*iovec)))
408 		return PTR_ERR(*iovec);
409 
410 	iov_iter_save_state(&s->iter, &s->iter_state);
411 	return 0;
412 }
413 
414 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
415 {
416 	return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
417 }
418 
419 /*
420  * For files that don't have ->read_iter() and ->write_iter(), handle them
421  * by looping over ->read() or ->write() manually.
422  */
423 static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter)
424 {
425 	struct kiocb *kiocb = &rw->kiocb;
426 	struct file *file = kiocb->ki_filp;
427 	ssize_t ret = 0;
428 	loff_t *ppos;
429 
430 	/*
431 	 * Don't support polled IO through this interface, and we can't
432 	 * support non-blocking either. For the latter, this just causes
433 	 * the kiocb to be handled from an async context.
434 	 */
435 	if (kiocb->ki_flags & IOCB_HIPRI)
436 		return -EOPNOTSUPP;
437 	if ((kiocb->ki_flags & IOCB_NOWAIT) &&
438 	    !(kiocb->ki_filp->f_flags & O_NONBLOCK))
439 		return -EAGAIN;
440 
441 	ppos = io_kiocb_ppos(kiocb);
442 
443 	while (iov_iter_count(iter)) {
444 		struct iovec iovec;
445 		ssize_t nr;
446 
447 		if (!iov_iter_is_bvec(iter)) {
448 			iovec = iov_iter_iovec(iter);
449 		} else {
450 			iovec.iov_base = u64_to_user_ptr(rw->addr);
451 			iovec.iov_len = rw->len;
452 		}
453 
454 		if (ddir == READ) {
455 			nr = file->f_op->read(file, iovec.iov_base,
456 					      iovec.iov_len, ppos);
457 		} else {
458 			nr = file->f_op->write(file, iovec.iov_base,
459 					       iovec.iov_len, ppos);
460 		}
461 
462 		if (nr < 0) {
463 			if (!ret)
464 				ret = nr;
465 			break;
466 		}
467 		ret += nr;
468 		if (!iov_iter_is_bvec(iter)) {
469 			iov_iter_advance(iter, nr);
470 		} else {
471 			rw->addr += nr;
472 			rw->len -= nr;
473 			if (!rw->len)
474 				break;
475 		}
476 		if (nr != iovec.iov_len)
477 			break;
478 	}
479 
480 	return ret;
481 }
482 
483 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
484 			  const struct iovec *fast_iov, struct iov_iter *iter)
485 {
486 	struct io_async_rw *io = req->async_data;
487 
488 	memcpy(&io->s.iter, iter, sizeof(*iter));
489 	io->free_iovec = iovec;
490 	io->bytes_done = 0;
491 	/* can only be fixed buffers, no need to do anything */
492 	if (iov_iter_is_bvec(iter))
493 		return;
494 	if (!iovec) {
495 		unsigned iov_off = 0;
496 
497 		io->s.iter.iov = io->s.fast_iov;
498 		if (iter->iov != fast_iov) {
499 			iov_off = iter->iov - fast_iov;
500 			io->s.iter.iov += iov_off;
501 		}
502 		if (io->s.fast_iov != fast_iov)
503 			memcpy(io->s.fast_iov + iov_off, fast_iov + iov_off,
504 			       sizeof(struct iovec) * iter->nr_segs);
505 	} else {
506 		req->flags |= REQ_F_NEED_CLEANUP;
507 	}
508 }
509 
510 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
511 			     struct io_rw_state *s, bool force)
512 {
513 	if (!force && !io_op_defs[req->opcode].prep_async)
514 		return 0;
515 	if (!req_has_async_data(req)) {
516 		struct io_async_rw *iorw;
517 
518 		if (io_alloc_async_data(req)) {
519 			kfree(iovec);
520 			return -ENOMEM;
521 		}
522 
523 		io_req_map_rw(req, iovec, s->fast_iov, &s->iter);
524 		iorw = req->async_data;
525 		/* we've copied and mapped the iter, ensure state is saved */
526 		iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state);
527 	}
528 	return 0;
529 }
530 
531 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
532 {
533 	struct io_async_rw *iorw = req->async_data;
534 	struct iovec *iov;
535 	int ret;
536 
537 	/* submission path, ->uring_lock should already be taken */
538 	ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
539 	if (unlikely(ret < 0))
540 		return ret;
541 
542 	iorw->bytes_done = 0;
543 	iorw->free_iovec = iov;
544 	if (iov)
545 		req->flags |= REQ_F_NEED_CLEANUP;
546 	return 0;
547 }
548 
549 int io_readv_prep_async(struct io_kiocb *req)
550 {
551 	return io_rw_prep_async(req, READ);
552 }
553 
554 int io_writev_prep_async(struct io_kiocb *req)
555 {
556 	return io_rw_prep_async(req, WRITE);
557 }
558 
559 /*
560  * This is our waitqueue callback handler, registered through __folio_lock_async()
561  * when we initially tried to do the IO with the iocb armed our waitqueue.
562  * This gets called when the page is unlocked, and we generally expect that to
563  * happen when the page IO is completed and the page is now uptodate. This will
564  * queue a task_work based retry of the operation, attempting to copy the data
565  * again. If the latter fails because the page was NOT uptodate, then we will
566  * do a thread based blocking retry of the operation. That's the unexpected
567  * slow path.
568  */
569 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
570 			     int sync, void *arg)
571 {
572 	struct wait_page_queue *wpq;
573 	struct io_kiocb *req = wait->private;
574 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
575 	struct wait_page_key *key = arg;
576 
577 	wpq = container_of(wait, struct wait_page_queue, wait);
578 
579 	if (!wake_page_match(wpq, key))
580 		return 0;
581 
582 	rw->kiocb.ki_flags &= ~IOCB_WAITQ;
583 	list_del_init(&wait->entry);
584 	io_req_task_queue(req);
585 	return 1;
586 }
587 
588 /*
589  * This controls whether a given IO request should be armed for async page
590  * based retry. If we return false here, the request is handed to the async
591  * worker threads for retry. If we're doing buffered reads on a regular file,
592  * we prepare a private wait_page_queue entry and retry the operation. This
593  * will either succeed because the page is now uptodate and unlocked, or it
594  * will register a callback when the page is unlocked at IO completion. Through
595  * that callback, io_uring uses task_work to setup a retry of the operation.
596  * That retry will attempt the buffered read again. The retry will generally
597  * succeed, or in rare cases where it fails, we then fall back to using the
598  * async worker threads for a blocking retry.
599  */
600 static bool io_rw_should_retry(struct io_kiocb *req)
601 {
602 	struct io_async_rw *io = req->async_data;
603 	struct wait_page_queue *wait = &io->wpq;
604 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
605 	struct kiocb *kiocb = &rw->kiocb;
606 
607 	/* never retry for NOWAIT, we just complete with -EAGAIN */
608 	if (req->flags & REQ_F_NOWAIT)
609 		return false;
610 
611 	/* Only for buffered IO */
612 	if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
613 		return false;
614 
615 	/*
616 	 * just use poll if we can, and don't attempt if the fs doesn't
617 	 * support callback based unlocks
618 	 */
619 	if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
620 		return false;
621 
622 	wait->wait.func = io_async_buf_func;
623 	wait->wait.private = req;
624 	wait->wait.flags = 0;
625 	INIT_LIST_HEAD(&wait->wait.entry);
626 	kiocb->ki_flags |= IOCB_WAITQ;
627 	kiocb->ki_flags &= ~IOCB_NOWAIT;
628 	kiocb->ki_waitq = wait;
629 	return true;
630 }
631 
632 static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
633 {
634 	struct file *file = rw->kiocb.ki_filp;
635 
636 	if (likely(file->f_op->read_iter))
637 		return call_read_iter(file, &rw->kiocb, iter);
638 	else if (file->f_op->read)
639 		return loop_rw_iter(READ, rw, iter);
640 	else
641 		return -EINVAL;
642 }
643 
644 static bool need_complete_io(struct io_kiocb *req)
645 {
646 	return req->flags & REQ_F_ISREG ||
647 		S_ISBLK(file_inode(req->file)->i_mode);
648 }
649 
650 static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
651 {
652 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
653 	struct kiocb *kiocb = &rw->kiocb;
654 	struct io_ring_ctx *ctx = req->ctx;
655 	struct file *file = req->file;
656 	int ret;
657 
658 	if (unlikely(!file || !(file->f_mode & mode)))
659 		return -EBADF;
660 
661 	if (!io_req_ffs_set(req))
662 		req->flags |= io_file_get_flags(file) << REQ_F_SUPPORT_NOWAIT_BIT;
663 
664 	kiocb->ki_flags = file->f_iocb_flags;
665 	ret = kiocb_set_rw_flags(kiocb, rw->flags);
666 	if (unlikely(ret))
667 		return ret;
668 
669 	/*
670 	 * If the file is marked O_NONBLOCK, still allow retry for it if it
671 	 * supports async. Otherwise it's impossible to use O_NONBLOCK files
672 	 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
673 	 */
674 	if ((kiocb->ki_flags & IOCB_NOWAIT) ||
675 	    ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
676 		req->flags |= REQ_F_NOWAIT;
677 
678 	if (ctx->flags & IORING_SETUP_IOPOLL) {
679 		if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
680 			return -EOPNOTSUPP;
681 
682 		kiocb->private = NULL;
683 		kiocb->ki_flags |= IOCB_HIPRI | IOCB_ALLOC_CACHE;
684 		kiocb->ki_complete = io_complete_rw_iopoll;
685 		req->iopoll_completed = 0;
686 	} else {
687 		if (kiocb->ki_flags & IOCB_HIPRI)
688 			return -EINVAL;
689 		kiocb->ki_complete = io_complete_rw;
690 	}
691 
692 	return 0;
693 }
694 
695 int io_read(struct io_kiocb *req, unsigned int issue_flags)
696 {
697 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
698 	struct io_rw_state __s, *s = &__s;
699 	struct iovec *iovec;
700 	struct kiocb *kiocb = &rw->kiocb;
701 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
702 	struct io_async_rw *io;
703 	ssize_t ret, ret2;
704 	loff_t *ppos;
705 
706 	if (!req_has_async_data(req)) {
707 		ret = io_import_iovec(READ, req, &iovec, s, issue_flags);
708 		if (unlikely(ret < 0))
709 			return ret;
710 	} else {
711 		io = req->async_data;
712 		s = &io->s;
713 
714 		/*
715 		 * Safe and required to re-import if we're using provided
716 		 * buffers, as we dropped the selected one before retry.
717 		 */
718 		if (io_do_buffer_select(req)) {
719 			ret = io_import_iovec(READ, req, &iovec, s, issue_flags);
720 			if (unlikely(ret < 0))
721 				return ret;
722 		}
723 
724 		/*
725 		 * We come here from an earlier attempt, restore our state to
726 		 * match in case it doesn't. It's cheap enough that we don't
727 		 * need to make this conditional.
728 		 */
729 		iov_iter_restore(&s->iter, &s->iter_state);
730 		iovec = NULL;
731 	}
732 	ret = io_rw_init_file(req, FMODE_READ);
733 	if (unlikely(ret)) {
734 		kfree(iovec);
735 		return ret;
736 	}
737 	req->cqe.res = iov_iter_count(&s->iter);
738 
739 	if (force_nonblock) {
740 		/* If the file doesn't support async, just async punt */
741 		if (unlikely(!io_file_supports_nowait(req))) {
742 			ret = io_setup_async_rw(req, iovec, s, true);
743 			return ret ?: -EAGAIN;
744 		}
745 		kiocb->ki_flags |= IOCB_NOWAIT;
746 	} else {
747 		/* Ensure we clear previously set non-block flag */
748 		kiocb->ki_flags &= ~IOCB_NOWAIT;
749 	}
750 
751 	ppos = io_kiocb_update_pos(req);
752 
753 	ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
754 	if (unlikely(ret)) {
755 		kfree(iovec);
756 		return ret;
757 	}
758 
759 	ret = io_iter_do_read(rw, &s->iter);
760 
761 	if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
762 		req->flags &= ~REQ_F_REISSUE;
763 		/* if we can poll, just do that */
764 		if (req->opcode == IORING_OP_READ && file_can_poll(req->file))
765 			return -EAGAIN;
766 		/* IOPOLL retry should happen for io-wq threads */
767 		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
768 			goto done;
769 		/* no retry on NONBLOCK nor RWF_NOWAIT */
770 		if (req->flags & REQ_F_NOWAIT)
771 			goto done;
772 		ret = 0;
773 	} else if (ret == -EIOCBQUEUED) {
774 		if (iovec)
775 			kfree(iovec);
776 		return IOU_ISSUE_SKIP_COMPLETE;
777 	} else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
778 		   (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) {
779 		/* read all, failed, already did sync or don't want to retry */
780 		goto done;
781 	}
782 
783 	/*
784 	 * Don't depend on the iter state matching what was consumed, or being
785 	 * untouched in case of error. Restore it and we'll advance it
786 	 * manually if we need to.
787 	 */
788 	iov_iter_restore(&s->iter, &s->iter_state);
789 
790 	ret2 = io_setup_async_rw(req, iovec, s, true);
791 	if (ret2)
792 		return ret2;
793 
794 	iovec = NULL;
795 	io = req->async_data;
796 	s = &io->s;
797 	/*
798 	 * Now use our persistent iterator and state, if we aren't already.
799 	 * We've restored and mapped the iter to match.
800 	 */
801 
802 	do {
803 		/*
804 		 * We end up here because of a partial read, either from
805 		 * above or inside this loop. Advance the iter by the bytes
806 		 * that were consumed.
807 		 */
808 		iov_iter_advance(&s->iter, ret);
809 		if (!iov_iter_count(&s->iter))
810 			break;
811 		io->bytes_done += ret;
812 		iov_iter_save_state(&s->iter, &s->iter_state);
813 
814 		/* if we can retry, do so with the callbacks armed */
815 		if (!io_rw_should_retry(req)) {
816 			kiocb->ki_flags &= ~IOCB_WAITQ;
817 			return -EAGAIN;
818 		}
819 
820 		/*
821 		 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
822 		 * we get -EIOCBQUEUED, then we'll get a notification when the
823 		 * desired page gets unlocked. We can also get a partial read
824 		 * here, and if we do, then just retry at the new offset.
825 		 */
826 		ret = io_iter_do_read(rw, &s->iter);
827 		if (ret == -EIOCBQUEUED)
828 			return IOU_ISSUE_SKIP_COMPLETE;
829 		/* we got some bytes, but not all. retry. */
830 		kiocb->ki_flags &= ~IOCB_WAITQ;
831 		iov_iter_restore(&s->iter, &s->iter_state);
832 	} while (ret > 0);
833 done:
834 	/* it's faster to check here then delegate to kfree */
835 	if (iovec)
836 		kfree(iovec);
837 	return kiocb_done(req, ret, issue_flags);
838 }
839 
840 int io_write(struct io_kiocb *req, unsigned int issue_flags)
841 {
842 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
843 	struct io_rw_state __s, *s = &__s;
844 	struct iovec *iovec;
845 	struct kiocb *kiocb = &rw->kiocb;
846 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
847 	ssize_t ret, ret2;
848 	loff_t *ppos;
849 
850 	if (!req_has_async_data(req)) {
851 		ret = io_import_iovec(WRITE, req, &iovec, s, issue_flags);
852 		if (unlikely(ret < 0))
853 			return ret;
854 	} else {
855 		struct io_async_rw *io = req->async_data;
856 
857 		s = &io->s;
858 		iov_iter_restore(&s->iter, &s->iter_state);
859 		iovec = NULL;
860 	}
861 	ret = io_rw_init_file(req, FMODE_WRITE);
862 	if (unlikely(ret)) {
863 		kfree(iovec);
864 		return ret;
865 	}
866 	req->cqe.res = iov_iter_count(&s->iter);
867 
868 	if (force_nonblock) {
869 		/* If the file doesn't support async, just async punt */
870 		if (unlikely(!io_file_supports_nowait(req)))
871 			goto copy_iov;
872 
873 		/* File path supports NOWAIT for non-direct_IO only for block devices. */
874 		if (!(kiocb->ki_flags & IOCB_DIRECT) &&
875 			!(kiocb->ki_filp->f_mode & FMODE_BUF_WASYNC) &&
876 			(req->flags & REQ_F_ISREG))
877 			goto copy_iov;
878 
879 		kiocb->ki_flags |= IOCB_NOWAIT;
880 	} else {
881 		/* Ensure we clear previously set non-block flag */
882 		kiocb->ki_flags &= ~IOCB_NOWAIT;
883 	}
884 
885 	ppos = io_kiocb_update_pos(req);
886 
887 	ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
888 	if (unlikely(ret)) {
889 		kfree(iovec);
890 		return ret;
891 	}
892 
893 	/*
894 	 * Open-code file_start_write here to grab freeze protection,
895 	 * which will be released by another thread in
896 	 * io_complete_rw().  Fool lockdep by telling it the lock got
897 	 * released so that it doesn't complain about the held lock when
898 	 * we return to userspace.
899 	 */
900 	if (req->flags & REQ_F_ISREG) {
901 		sb_start_write(file_inode(req->file)->i_sb);
902 		__sb_writers_release(file_inode(req->file)->i_sb,
903 					SB_FREEZE_WRITE);
904 	}
905 	kiocb->ki_flags |= IOCB_WRITE;
906 
907 	if (likely(req->file->f_op->write_iter))
908 		ret2 = call_write_iter(req->file, kiocb, &s->iter);
909 	else if (req->file->f_op->write)
910 		ret2 = loop_rw_iter(WRITE, rw, &s->iter);
911 	else
912 		ret2 = -EINVAL;
913 
914 	if (req->flags & REQ_F_REISSUE) {
915 		req->flags &= ~REQ_F_REISSUE;
916 		ret2 = -EAGAIN;
917 	}
918 
919 	/*
920 	 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
921 	 * retry them without IOCB_NOWAIT.
922 	 */
923 	if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
924 		ret2 = -EAGAIN;
925 	/* no retry on NONBLOCK nor RWF_NOWAIT */
926 	if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
927 		goto done;
928 	if (!force_nonblock || ret2 != -EAGAIN) {
929 		/* IOPOLL retry should happen for io-wq threads */
930 		if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
931 			goto copy_iov;
932 
933 		if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) {
934 			struct io_async_rw *rw;
935 
936 			trace_io_uring_short_write(req->ctx, kiocb->ki_pos - ret2,
937 						req->cqe.res, ret2);
938 
939 			/* This is a partial write. The file pos has already been
940 			 * updated, setup the async struct to complete the request
941 			 * in the worker. Also update bytes_done to account for
942 			 * the bytes already written.
943 			 */
944 			iov_iter_save_state(&s->iter, &s->iter_state);
945 			ret = io_setup_async_rw(req, iovec, s, true);
946 
947 			rw = req->async_data;
948 			if (rw)
949 				rw->bytes_done += ret2;
950 
951 			if (kiocb->ki_flags & IOCB_WRITE)
952 				kiocb_end_write(req);
953 			return ret ? ret : -EAGAIN;
954 		}
955 done:
956 		ret = kiocb_done(req, ret2, issue_flags);
957 	} else {
958 copy_iov:
959 		iov_iter_restore(&s->iter, &s->iter_state);
960 		ret = io_setup_async_rw(req, iovec, s, false);
961 		if (!ret) {
962 			if (kiocb->ki_flags & IOCB_WRITE)
963 				kiocb_end_write(req);
964 			return -EAGAIN;
965 		}
966 		return ret;
967 	}
968 	/* it's reportedly faster than delegating the null check to kfree() */
969 	if (iovec)
970 		kfree(iovec);
971 	return ret;
972 }
973 
974 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
975 {
976 	io_commit_cqring_flush(ctx);
977 	if (ctx->flags & IORING_SETUP_SQPOLL)
978 		io_cqring_wake(ctx);
979 }
980 
981 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
982 {
983 	struct io_wq_work_node *pos, *start, *prev;
984 	unsigned int poll_flags = BLK_POLL_NOSLEEP;
985 	DEFINE_IO_COMP_BATCH(iob);
986 	int nr_events = 0;
987 
988 	/*
989 	 * Only spin for completions if we don't have multiple devices hanging
990 	 * off our complete list.
991 	 */
992 	if (ctx->poll_multi_queue || force_nonspin)
993 		poll_flags |= BLK_POLL_ONESHOT;
994 
995 	wq_list_for_each(pos, start, &ctx->iopoll_list) {
996 		struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
997 		struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
998 		int ret;
999 
1000 		/*
1001 		 * Move completed and retryable entries to our local lists.
1002 		 * If we find a request that requires polling, break out
1003 		 * and complete those lists first, if we have entries there.
1004 		 */
1005 		if (READ_ONCE(req->iopoll_completed))
1006 			break;
1007 
1008 		ret = rw->kiocb.ki_filp->f_op->iopoll(&rw->kiocb, &iob, poll_flags);
1009 		if (unlikely(ret < 0))
1010 			return ret;
1011 		else if (ret)
1012 			poll_flags |= BLK_POLL_ONESHOT;
1013 
1014 		/* iopoll may have completed current req */
1015 		if (!rq_list_empty(iob.req_list) ||
1016 		    READ_ONCE(req->iopoll_completed))
1017 			break;
1018 	}
1019 
1020 	if (!rq_list_empty(iob.req_list))
1021 		iob.complete(&iob);
1022 	else if (!pos)
1023 		return 0;
1024 
1025 	prev = start;
1026 	wq_list_for_each_resume(pos, prev) {
1027 		struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1028 
1029 		/* order with io_complete_rw_iopoll(), e.g. ->result updates */
1030 		if (!smp_load_acquire(&req->iopoll_completed))
1031 			break;
1032 		nr_events++;
1033 		if (unlikely(req->flags & REQ_F_CQE_SKIP))
1034 			continue;
1035 
1036 		req->cqe.flags = io_put_kbuf(req, 0);
1037 		__io_fill_cqe_req(req->ctx, req);
1038 	}
1039 
1040 	if (unlikely(!nr_events))
1041 		return 0;
1042 
1043 	io_commit_cqring(ctx);
1044 	io_cqring_ev_posted_iopoll(ctx);
1045 	pos = start ? start->next : ctx->iopoll_list.first;
1046 	wq_list_cut(&ctx->iopoll_list, prev, start);
1047 	io_free_batch_list(ctx, pos);
1048 	return nr_events;
1049 }
1050