xref: /openbmc/linux/io_uring/rw.c (revision a5d2bb06)
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 #ifdef CONFIG_COMPAT
37 static 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 
55 static 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(&iov, uiov, sizeof(*uiov)))
71 		return -EFAULT;
72 	rw->len = iov.iov_len;
73 	return 0;
74 }
75 
76 int 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, 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 void io_readv_writev_cleanup(struct io_kiocb *req)
127 {
128 	struct io_async_rw *io = req->async_data;
129 
130 	kfree(io->free_iovec);
131 }
132 
133 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
134 {
135 	switch (ret) {
136 	case -EIOCBQUEUED:
137 		break;
138 	case -ERESTARTSYS:
139 	case -ERESTARTNOINTR:
140 	case -ERESTARTNOHAND:
141 	case -ERESTART_RESTARTBLOCK:
142 		/*
143 		 * We can't just restart the syscall, since previously
144 		 * submitted sqes may already be in progress. Just fail this
145 		 * IO with EINTR.
146 		 */
147 		ret = -EINTR;
148 		fallthrough;
149 	default:
150 		kiocb->ki_complete(kiocb, ret);
151 	}
152 }
153 
154 static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req)
155 {
156 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
157 
158 	if (rw->kiocb.ki_pos != -1)
159 		return &rw->kiocb.ki_pos;
160 
161 	if (!(req->file->f_mode & FMODE_STREAM)) {
162 		req->flags |= REQ_F_CUR_POS;
163 		rw->kiocb.ki_pos = req->file->f_pos;
164 		return &rw->kiocb.ki_pos;
165 	}
166 
167 	rw->kiocb.ki_pos = 0;
168 	return NULL;
169 }
170 
171 static void io_req_task_queue_reissue(struct io_kiocb *req)
172 {
173 	req->io_task_work.func = io_queue_iowq;
174 	io_req_task_work_add(req);
175 }
176 
177 #ifdef CONFIG_BLOCK
178 static bool io_resubmit_prep(struct io_kiocb *req)
179 {
180 	struct io_async_rw *io = req->async_data;
181 
182 	if (!req_has_async_data(req))
183 		return !io_req_prep_async(req);
184 	iov_iter_restore(&io->s.iter, &io->s.iter_state);
185 	return true;
186 }
187 
188 static bool io_rw_should_reissue(struct io_kiocb *req)
189 {
190 	umode_t mode = file_inode(req->file)->i_mode;
191 	struct io_ring_ctx *ctx = req->ctx;
192 
193 	if (!S_ISBLK(mode) && !S_ISREG(mode))
194 		return false;
195 	if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
196 	    !(ctx->flags & IORING_SETUP_IOPOLL)))
197 		return false;
198 	/*
199 	 * If ref is dying, we might be running poll reap from the exit work.
200 	 * Don't attempt to reissue from that path, just let it fail with
201 	 * -EAGAIN.
202 	 */
203 	if (percpu_ref_is_dying(&ctx->refs))
204 		return false;
205 	/*
206 	 * Play it safe and assume not safe to re-import and reissue if we're
207 	 * not in the original thread group (or in task context).
208 	 */
209 	if (!same_thread_group(req->task, current) || !in_task())
210 		return false;
211 	return true;
212 }
213 #else
214 static bool io_resubmit_prep(struct io_kiocb *req)
215 {
216 	return false;
217 }
218 static bool io_rw_should_reissue(struct io_kiocb *req)
219 {
220 	return false;
221 }
222 #endif
223 
224 static void io_req_end_write(struct io_kiocb *req)
225 {
226 	if (req->flags & REQ_F_ISREG) {
227 		struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
228 
229 		kiocb_end_write(&rw->kiocb);
230 	}
231 }
232 
233 /*
234  * Trigger the notifications after having done some IO, and finish the write
235  * accounting, if any.
236  */
237 static void io_req_io_end(struct io_kiocb *req)
238 {
239 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
240 
241 	if (rw->kiocb.ki_flags & IOCB_WRITE) {
242 		io_req_end_write(req);
243 		fsnotify_modify(req->file);
244 	} else {
245 		fsnotify_access(req->file);
246 	}
247 }
248 
249 static bool __io_complete_rw_common(struct io_kiocb *req, long res)
250 {
251 	if (unlikely(res != req->cqe.res)) {
252 		if ((res == -EAGAIN || res == -EOPNOTSUPP) &&
253 		    io_rw_should_reissue(req)) {
254 			/*
255 			 * Reissue will start accounting again, finish the
256 			 * current cycle.
257 			 */
258 			io_req_io_end(req);
259 			req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
260 			return true;
261 		}
262 		req_set_fail(req);
263 		req->cqe.res = res;
264 	}
265 	return false;
266 }
267 
268 static inline int io_fixup_rw_res(struct io_kiocb *req, long res)
269 {
270 	struct io_async_rw *io = req->async_data;
271 
272 	/* add previously done IO, if any */
273 	if (req_has_async_data(req) && io->bytes_done > 0) {
274 		if (res < 0)
275 			res = io->bytes_done;
276 		else
277 			res += io->bytes_done;
278 	}
279 	return res;
280 }
281 
282 void io_req_rw_complete(struct io_kiocb *req, struct io_tw_state *ts)
283 {
284 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
285 	struct kiocb *kiocb = &rw->kiocb;
286 
287 	if ((kiocb->ki_flags & IOCB_DIO_CALLER_COMP) && kiocb->dio_complete) {
288 		long res = kiocb->dio_complete(rw->kiocb.private);
289 
290 		io_req_set_res(req, io_fixup_rw_res(req, res), 0);
291 	}
292 
293 	io_req_io_end(req);
294 
295 	if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
296 		unsigned issue_flags = ts->locked ? 0 : IO_URING_F_UNLOCKED;
297 
298 		req->cqe.flags |= io_put_kbuf(req, issue_flags);
299 	}
300 	io_req_task_complete(req, ts);
301 }
302 
303 static void io_complete_rw(struct kiocb *kiocb, long res)
304 {
305 	struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
306 	struct io_kiocb *req = cmd_to_io_kiocb(rw);
307 
308 	if (!kiocb->dio_complete || !(kiocb->ki_flags & IOCB_DIO_CALLER_COMP)) {
309 		if (__io_complete_rw_common(req, res))
310 			return;
311 		io_req_set_res(req, io_fixup_rw_res(req, res), 0);
312 	}
313 	req->io_task_work.func = io_req_rw_complete;
314 	__io_req_task_work_add(req, IOU_F_TWQ_LAZY_WAKE);
315 }
316 
317 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
318 {
319 	struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
320 	struct io_kiocb *req = cmd_to_io_kiocb(rw);
321 
322 	if (kiocb->ki_flags & IOCB_WRITE)
323 		io_req_end_write(req);
324 	if (unlikely(res != req->cqe.res)) {
325 		if (res == -EAGAIN && io_rw_should_reissue(req)) {
326 			req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO;
327 			return;
328 		}
329 		req->cqe.res = res;
330 	}
331 
332 	/* order with io_iopoll_complete() checking ->iopoll_completed */
333 	smp_store_release(&req->iopoll_completed, 1);
334 }
335 
336 static int kiocb_done(struct io_kiocb *req, ssize_t ret,
337 		       unsigned int issue_flags)
338 {
339 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
340 	unsigned final_ret = io_fixup_rw_res(req, ret);
341 
342 	if (ret >= 0 && req->flags & REQ_F_CUR_POS)
343 		req->file->f_pos = rw->kiocb.ki_pos;
344 	if (ret >= 0 && (rw->kiocb.ki_complete == io_complete_rw)) {
345 		if (!__io_complete_rw_common(req, ret)) {
346 			/*
347 			 * Safe to call io_end from here as we're inline
348 			 * from the submission path.
349 			 */
350 			io_req_io_end(req);
351 			io_req_set_res(req, final_ret,
352 				       io_put_kbuf(req, issue_flags));
353 			return IOU_OK;
354 		}
355 	} else {
356 		io_rw_done(&rw->kiocb, ret);
357 	}
358 
359 	if (req->flags & REQ_F_REISSUE) {
360 		req->flags &= ~REQ_F_REISSUE;
361 		if (io_resubmit_prep(req))
362 			io_req_task_queue_reissue(req);
363 		else
364 			io_req_task_queue_fail(req, final_ret);
365 	}
366 	return IOU_ISSUE_SKIP_COMPLETE;
367 }
368 
369 static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req,
370 				       struct io_rw_state *s,
371 				       unsigned int issue_flags)
372 {
373 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
374 	struct iov_iter *iter = &s->iter;
375 	u8 opcode = req->opcode;
376 	struct iovec *iovec;
377 	void __user *buf;
378 	size_t sqe_len;
379 	ssize_t ret;
380 
381 	if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
382 		ret = io_import_fixed(ddir, iter, req->imu, rw->addr, rw->len);
383 		if (ret)
384 			return ERR_PTR(ret);
385 		return NULL;
386 	}
387 
388 	buf = u64_to_user_ptr(rw->addr);
389 	sqe_len = rw->len;
390 
391 	if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE ||
392 	    (req->flags & REQ_F_BUFFER_SELECT)) {
393 		if (io_do_buffer_select(req)) {
394 			buf = io_buffer_select(req, &sqe_len, issue_flags);
395 			if (!buf)
396 				return ERR_PTR(-ENOBUFS);
397 			rw->addr = (unsigned long) buf;
398 			rw->len = sqe_len;
399 		}
400 
401 		ret = import_ubuf(ddir, buf, sqe_len, iter);
402 		if (ret)
403 			return ERR_PTR(ret);
404 		return NULL;
405 	}
406 
407 	iovec = s->fast_iov;
408 	ret = __import_iovec(ddir, buf, sqe_len, UIO_FASTIOV, &iovec, iter,
409 			      req->ctx->compat);
410 	if (unlikely(ret < 0))
411 		return ERR_PTR(ret);
412 	return iovec;
413 }
414 
415 static inline int io_import_iovec(int rw, struct io_kiocb *req,
416 				  struct iovec **iovec, struct io_rw_state *s,
417 				  unsigned int issue_flags)
418 {
419 	*iovec = __io_import_iovec(rw, req, s, issue_flags);
420 	if (IS_ERR(*iovec))
421 		return PTR_ERR(*iovec);
422 
423 	iov_iter_save_state(&s->iter, &s->iter_state);
424 	return 0;
425 }
426 
427 static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
428 {
429 	return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
430 }
431 
432 /*
433  * For files that don't have ->read_iter() and ->write_iter(), handle them
434  * by looping over ->read() or ->write() manually.
435  */
436 static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter)
437 {
438 	struct kiocb *kiocb = &rw->kiocb;
439 	struct file *file = kiocb->ki_filp;
440 	ssize_t ret = 0;
441 	loff_t *ppos;
442 
443 	/*
444 	 * Don't support polled IO through this interface, and we can't
445 	 * support non-blocking either. For the latter, this just causes
446 	 * the kiocb to be handled from an async context.
447 	 */
448 	if (kiocb->ki_flags & IOCB_HIPRI)
449 		return -EOPNOTSUPP;
450 	if ((kiocb->ki_flags & IOCB_NOWAIT) &&
451 	    !(kiocb->ki_filp->f_flags & O_NONBLOCK))
452 		return -EAGAIN;
453 
454 	ppos = io_kiocb_ppos(kiocb);
455 
456 	while (iov_iter_count(iter)) {
457 		void __user *addr;
458 		size_t len;
459 		ssize_t nr;
460 
461 		if (iter_is_ubuf(iter)) {
462 			addr = iter->ubuf + iter->iov_offset;
463 			len = iov_iter_count(iter);
464 		} else if (!iov_iter_is_bvec(iter)) {
465 			addr = iter_iov_addr(iter);
466 			len = iter_iov_len(iter);
467 		} else {
468 			addr = u64_to_user_ptr(rw->addr);
469 			len = rw->len;
470 		}
471 
472 		if (ddir == READ)
473 			nr = file->f_op->read(file, addr, len, ppos);
474 		else
475 			nr = file->f_op->write(file, addr, len, ppos);
476 
477 		if (nr < 0) {
478 			if (!ret)
479 				ret = nr;
480 			break;
481 		}
482 		ret += nr;
483 		if (!iov_iter_is_bvec(iter)) {
484 			iov_iter_advance(iter, nr);
485 		} else {
486 			rw->addr += nr;
487 			rw->len -= nr;
488 			if (!rw->len)
489 				break;
490 		}
491 		if (nr != len)
492 			break;
493 	}
494 
495 	return ret;
496 }
497 
498 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
499 			  const struct iovec *fast_iov, struct iov_iter *iter)
500 {
501 	struct io_async_rw *io = req->async_data;
502 
503 	memcpy(&io->s.iter, iter, sizeof(*iter));
504 	io->free_iovec = iovec;
505 	io->bytes_done = 0;
506 	/* can only be fixed buffers, no need to do anything */
507 	if (iov_iter_is_bvec(iter) || iter_is_ubuf(iter))
508 		return;
509 	if (!iovec) {
510 		unsigned iov_off = 0;
511 
512 		io->s.iter.__iov = io->s.fast_iov;
513 		if (iter->__iov != fast_iov) {
514 			iov_off = iter_iov(iter) - fast_iov;
515 			io->s.iter.__iov += iov_off;
516 		}
517 		if (io->s.fast_iov != fast_iov)
518 			memcpy(io->s.fast_iov + iov_off, fast_iov + iov_off,
519 			       sizeof(struct iovec) * iter->nr_segs);
520 	} else {
521 		req->flags |= REQ_F_NEED_CLEANUP;
522 	}
523 }
524 
525 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
526 			     struct io_rw_state *s, bool force)
527 {
528 	if (!force && !io_cold_defs[req->opcode].prep_async)
529 		return 0;
530 	if (!req_has_async_data(req)) {
531 		struct io_async_rw *iorw;
532 
533 		if (io_alloc_async_data(req)) {
534 			kfree(iovec);
535 			return -ENOMEM;
536 		}
537 
538 		io_req_map_rw(req, iovec, s->fast_iov, &s->iter);
539 		iorw = req->async_data;
540 		/* we've copied and mapped the iter, ensure state is saved */
541 		iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state);
542 	}
543 	return 0;
544 }
545 
546 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
547 {
548 	struct io_async_rw *iorw = req->async_data;
549 	struct iovec *iov;
550 	int ret;
551 
552 	/* submission path, ->uring_lock should already be taken */
553 	ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
554 	if (unlikely(ret < 0))
555 		return ret;
556 
557 	iorw->bytes_done = 0;
558 	iorw->free_iovec = iov;
559 	if (iov)
560 		req->flags |= REQ_F_NEED_CLEANUP;
561 	return 0;
562 }
563 
564 int io_readv_prep_async(struct io_kiocb *req)
565 {
566 	return io_rw_prep_async(req, ITER_DEST);
567 }
568 
569 int io_writev_prep_async(struct io_kiocb *req)
570 {
571 	return io_rw_prep_async(req, ITER_SOURCE);
572 }
573 
574 /*
575  * This is our waitqueue callback handler, registered through __folio_lock_async()
576  * when we initially tried to do the IO with the iocb armed our waitqueue.
577  * This gets called when the page is unlocked, and we generally expect that to
578  * happen when the page IO is completed and the page is now uptodate. This will
579  * queue a task_work based retry of the operation, attempting to copy the data
580  * again. If the latter fails because the page was NOT uptodate, then we will
581  * do a thread based blocking retry of the operation. That's the unexpected
582  * slow path.
583  */
584 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
585 			     int sync, void *arg)
586 {
587 	struct wait_page_queue *wpq;
588 	struct io_kiocb *req = wait->private;
589 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
590 	struct wait_page_key *key = arg;
591 
592 	wpq = container_of(wait, struct wait_page_queue, wait);
593 
594 	if (!wake_page_match(wpq, key))
595 		return 0;
596 
597 	rw->kiocb.ki_flags &= ~IOCB_WAITQ;
598 	list_del_init(&wait->entry);
599 	io_req_task_queue(req);
600 	return 1;
601 }
602 
603 /*
604  * This controls whether a given IO request should be armed for async page
605  * based retry. If we return false here, the request is handed to the async
606  * worker threads for retry. If we're doing buffered reads on a regular file,
607  * we prepare a private wait_page_queue entry and retry the operation. This
608  * will either succeed because the page is now uptodate and unlocked, or it
609  * will register a callback when the page is unlocked at IO completion. Through
610  * that callback, io_uring uses task_work to setup a retry of the operation.
611  * That retry will attempt the buffered read again. The retry will generally
612  * succeed, or in rare cases where it fails, we then fall back to using the
613  * async worker threads for a blocking retry.
614  */
615 static bool io_rw_should_retry(struct io_kiocb *req)
616 {
617 	struct io_async_rw *io = req->async_data;
618 	struct wait_page_queue *wait = &io->wpq;
619 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
620 	struct kiocb *kiocb = &rw->kiocb;
621 
622 	/* never retry for NOWAIT, we just complete with -EAGAIN */
623 	if (req->flags & REQ_F_NOWAIT)
624 		return false;
625 
626 	/* Only for buffered IO */
627 	if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
628 		return false;
629 
630 	/*
631 	 * just use poll if we can, and don't attempt if the fs doesn't
632 	 * support callback based unlocks
633 	 */
634 	if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
635 		return false;
636 
637 	wait->wait.func = io_async_buf_func;
638 	wait->wait.private = req;
639 	wait->wait.flags = 0;
640 	INIT_LIST_HEAD(&wait->wait.entry);
641 	kiocb->ki_flags |= IOCB_WAITQ;
642 	kiocb->ki_flags &= ~IOCB_NOWAIT;
643 	kiocb->ki_waitq = wait;
644 	return true;
645 }
646 
647 static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
648 {
649 	struct file *file = rw->kiocb.ki_filp;
650 
651 	if (likely(file->f_op->read_iter))
652 		return call_read_iter(file, &rw->kiocb, iter);
653 	else if (file->f_op->read)
654 		return loop_rw_iter(READ, rw, iter);
655 	else
656 		return -EINVAL;
657 }
658 
659 static bool need_complete_io(struct io_kiocb *req)
660 {
661 	return req->flags & REQ_F_ISREG ||
662 		S_ISBLK(file_inode(req->file)->i_mode);
663 }
664 
665 static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
666 {
667 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
668 	struct kiocb *kiocb = &rw->kiocb;
669 	struct io_ring_ctx *ctx = req->ctx;
670 	struct file *file = req->file;
671 	int ret;
672 
673 	if (unlikely(!file || !(file->f_mode & mode)))
674 		return -EBADF;
675 
676 	if (!(req->flags & REQ_F_FIXED_FILE))
677 		req->flags |= io_file_get_flags(file);
678 
679 	kiocb->ki_flags = file->f_iocb_flags;
680 	ret = kiocb_set_rw_flags(kiocb, rw->flags);
681 	if (unlikely(ret))
682 		return ret;
683 	kiocb->ki_flags |= IOCB_ALLOC_CACHE;
684 
685 	/*
686 	 * If the file is marked O_NONBLOCK, still allow retry for it if it
687 	 * supports async. Otherwise it's impossible to use O_NONBLOCK files
688 	 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
689 	 */
690 	if ((kiocb->ki_flags & IOCB_NOWAIT) ||
691 	    ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
692 		req->flags |= REQ_F_NOWAIT;
693 
694 	if (ctx->flags & IORING_SETUP_IOPOLL) {
695 		if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
696 			return -EOPNOTSUPP;
697 
698 		kiocb->private = NULL;
699 		kiocb->ki_flags |= IOCB_HIPRI;
700 		kiocb->ki_complete = io_complete_rw_iopoll;
701 		req->iopoll_completed = 0;
702 	} else {
703 		if (kiocb->ki_flags & IOCB_HIPRI)
704 			return -EINVAL;
705 		kiocb->ki_complete = io_complete_rw;
706 	}
707 
708 	return 0;
709 }
710 
711 int io_read(struct io_kiocb *req, unsigned int issue_flags)
712 {
713 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
714 	struct io_rw_state __s, *s = &__s;
715 	struct iovec *iovec;
716 	struct kiocb *kiocb = &rw->kiocb;
717 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
718 	struct io_async_rw *io;
719 	ssize_t ret, ret2;
720 	loff_t *ppos;
721 
722 	if (!req_has_async_data(req)) {
723 		ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
724 		if (unlikely(ret < 0))
725 			return ret;
726 	} else {
727 		io = req->async_data;
728 		s = &io->s;
729 
730 		/*
731 		 * Safe and required to re-import if we're using provided
732 		 * buffers, as we dropped the selected one before retry.
733 		 */
734 		if (io_do_buffer_select(req)) {
735 			ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
736 			if (unlikely(ret < 0))
737 				return ret;
738 		}
739 
740 		/*
741 		 * We come here from an earlier attempt, restore our state to
742 		 * match in case it doesn't. It's cheap enough that we don't
743 		 * need to make this conditional.
744 		 */
745 		iov_iter_restore(&s->iter, &s->iter_state);
746 		iovec = NULL;
747 	}
748 	ret = io_rw_init_file(req, FMODE_READ);
749 	if (unlikely(ret)) {
750 		kfree(iovec);
751 		return ret;
752 	}
753 	req->cqe.res = iov_iter_count(&s->iter);
754 
755 	if (force_nonblock) {
756 		/* If the file doesn't support async, just async punt */
757 		if (unlikely(!io_file_supports_nowait(req))) {
758 			ret = io_setup_async_rw(req, iovec, s, true);
759 			return ret ?: -EAGAIN;
760 		}
761 		kiocb->ki_flags |= IOCB_NOWAIT;
762 	} else {
763 		/* Ensure we clear previously set non-block flag */
764 		kiocb->ki_flags &= ~IOCB_NOWAIT;
765 	}
766 
767 	ppos = io_kiocb_update_pos(req);
768 
769 	ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
770 	if (unlikely(ret)) {
771 		kfree(iovec);
772 		return ret;
773 	}
774 
775 	ret = io_iter_do_read(rw, &s->iter);
776 
777 	if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
778 		req->flags &= ~REQ_F_REISSUE;
779 		/* if we can poll, just do that */
780 		if (req->opcode == IORING_OP_READ && file_can_poll(req->file))
781 			return -EAGAIN;
782 		/* IOPOLL retry should happen for io-wq threads */
783 		if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
784 			goto done;
785 		/* no retry on NONBLOCK nor RWF_NOWAIT */
786 		if (req->flags & REQ_F_NOWAIT)
787 			goto done;
788 		ret = 0;
789 	} else if (ret == -EIOCBQUEUED) {
790 		if (iovec)
791 			kfree(iovec);
792 		return IOU_ISSUE_SKIP_COMPLETE;
793 	} else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
794 		   (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) {
795 		/* read all, failed, already did sync or don't want to retry */
796 		goto done;
797 	}
798 
799 	/*
800 	 * Don't depend on the iter state matching what was consumed, or being
801 	 * untouched in case of error. Restore it and we'll advance it
802 	 * manually if we need to.
803 	 */
804 	iov_iter_restore(&s->iter, &s->iter_state);
805 
806 	ret2 = io_setup_async_rw(req, iovec, s, true);
807 	iovec = NULL;
808 	if (ret2) {
809 		ret = ret > 0 ? ret : ret2;
810 		goto done;
811 	}
812 
813 	io = req->async_data;
814 	s = &io->s;
815 	/*
816 	 * Now use our persistent iterator and state, if we aren't already.
817 	 * We've restored and mapped the iter to match.
818 	 */
819 
820 	do {
821 		/*
822 		 * We end up here because of a partial read, either from
823 		 * above or inside this loop. Advance the iter by the bytes
824 		 * that were consumed.
825 		 */
826 		iov_iter_advance(&s->iter, ret);
827 		if (!iov_iter_count(&s->iter))
828 			break;
829 		io->bytes_done += ret;
830 		iov_iter_save_state(&s->iter, &s->iter_state);
831 
832 		/* if we can retry, do so with the callbacks armed */
833 		if (!io_rw_should_retry(req)) {
834 			kiocb->ki_flags &= ~IOCB_WAITQ;
835 			return -EAGAIN;
836 		}
837 
838 		req->cqe.res = iov_iter_count(&s->iter);
839 		/*
840 		 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
841 		 * we get -EIOCBQUEUED, then we'll get a notification when the
842 		 * desired page gets unlocked. We can also get a partial read
843 		 * here, and if we do, then just retry at the new offset.
844 		 */
845 		ret = io_iter_do_read(rw, &s->iter);
846 		if (ret == -EIOCBQUEUED)
847 			return IOU_ISSUE_SKIP_COMPLETE;
848 		/* we got some bytes, but not all. retry. */
849 		kiocb->ki_flags &= ~IOCB_WAITQ;
850 		iov_iter_restore(&s->iter, &s->iter_state);
851 	} while (ret > 0);
852 done:
853 	/* it's faster to check here then delegate to kfree */
854 	if (iovec)
855 		kfree(iovec);
856 	return kiocb_done(req, ret, issue_flags);
857 }
858 
859 int io_write(struct io_kiocb *req, unsigned int issue_flags)
860 {
861 	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
862 	struct io_rw_state __s, *s = &__s;
863 	struct iovec *iovec;
864 	struct kiocb *kiocb = &rw->kiocb;
865 	bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
866 	ssize_t ret, ret2;
867 	loff_t *ppos;
868 
869 	if (!req_has_async_data(req)) {
870 		ret = io_import_iovec(ITER_SOURCE, req, &iovec, s, issue_flags);
871 		if (unlikely(ret < 0))
872 			return ret;
873 	} else {
874 		struct io_async_rw *io = req->async_data;
875 
876 		s = &io->s;
877 		iov_iter_restore(&s->iter, &s->iter_state);
878 		iovec = NULL;
879 	}
880 	ret = io_rw_init_file(req, FMODE_WRITE);
881 	if (unlikely(ret)) {
882 		kfree(iovec);
883 		return ret;
884 	}
885 	req->cqe.res = iov_iter_count(&s->iter);
886 
887 	if (force_nonblock) {
888 		/* If the file doesn't support async, just async punt */
889 		if (unlikely(!io_file_supports_nowait(req)))
890 			goto copy_iov;
891 
892 		/* File path supports NOWAIT for non-direct_IO only for block devices. */
893 		if (!(kiocb->ki_flags & IOCB_DIRECT) &&
894 			!(kiocb->ki_filp->f_mode & FMODE_BUF_WASYNC) &&
895 			(req->flags & REQ_F_ISREG))
896 			goto copy_iov;
897 
898 		kiocb->ki_flags |= IOCB_NOWAIT;
899 	} else {
900 		/* Ensure we clear previously set non-block flag */
901 		kiocb->ki_flags &= ~IOCB_NOWAIT;
902 	}
903 
904 	ppos = io_kiocb_update_pos(req);
905 
906 	ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
907 	if (unlikely(ret)) {
908 		kfree(iovec);
909 		return ret;
910 	}
911 
912 	if (req->flags & REQ_F_ISREG)
913 		kiocb_start_write(kiocb);
914 	kiocb->ki_flags |= IOCB_WRITE;
915 
916 	if (likely(req->file->f_op->write_iter))
917 		ret2 = call_write_iter(req->file, kiocb, &s->iter);
918 	else if (req->file->f_op->write)
919 		ret2 = loop_rw_iter(WRITE, rw, &s->iter);
920 	else
921 		ret2 = -EINVAL;
922 
923 	if (req->flags & REQ_F_REISSUE) {
924 		req->flags &= ~REQ_F_REISSUE;
925 		ret2 = -EAGAIN;
926 	}
927 
928 	/*
929 	 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
930 	 * retry them without IOCB_NOWAIT.
931 	 */
932 	if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
933 		ret2 = -EAGAIN;
934 	/* no retry on NONBLOCK nor RWF_NOWAIT */
935 	if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
936 		goto done;
937 	if (!force_nonblock || ret2 != -EAGAIN) {
938 		/* IOPOLL retry should happen for io-wq threads */
939 		if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
940 			goto copy_iov;
941 
942 		if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) {
943 			struct io_async_rw *io;
944 
945 			trace_io_uring_short_write(req->ctx, kiocb->ki_pos - ret2,
946 						req->cqe.res, ret2);
947 
948 			/* This is a partial write. The file pos has already been
949 			 * updated, setup the async struct to complete the request
950 			 * in the worker. Also update bytes_done to account for
951 			 * the bytes already written.
952 			 */
953 			iov_iter_save_state(&s->iter, &s->iter_state);
954 			ret = io_setup_async_rw(req, iovec, s, true);
955 
956 			io = req->async_data;
957 			if (io)
958 				io->bytes_done += ret2;
959 
960 			if (kiocb->ki_flags & IOCB_WRITE)
961 				io_req_end_write(req);
962 			return ret ? ret : -EAGAIN;
963 		}
964 done:
965 		ret = kiocb_done(req, ret2, issue_flags);
966 	} else {
967 copy_iov:
968 		iov_iter_restore(&s->iter, &s->iter_state);
969 		ret = io_setup_async_rw(req, iovec, s, false);
970 		if (!ret) {
971 			if (kiocb->ki_flags & IOCB_WRITE)
972 				io_req_end_write(req);
973 			return -EAGAIN;
974 		}
975 		return ret;
976 	}
977 	/* it's reportedly faster than delegating the null check to kfree() */
978 	if (iovec)
979 		kfree(iovec);
980 	return ret;
981 }
982 
983 void io_rw_fail(struct io_kiocb *req)
984 {
985 	int res;
986 
987 	res = io_fixup_rw_res(req, req->cqe.res);
988 	io_req_set_res(req, res, req->cqe.flags);
989 }
990 
991 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
992 {
993 	struct io_wq_work_node *pos, *start, *prev;
994 	unsigned int poll_flags = 0;
995 	DEFINE_IO_COMP_BATCH(iob);
996 	int nr_events = 0;
997 
998 	/*
999 	 * Only spin for completions if we don't have multiple devices hanging
1000 	 * off our complete list.
1001 	 */
1002 	if (ctx->poll_multi_queue || force_nonspin)
1003 		poll_flags |= BLK_POLL_ONESHOT;
1004 
1005 	wq_list_for_each(pos, start, &ctx->iopoll_list) {
1006 		struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1007 		struct file *file = req->file;
1008 		int ret;
1009 
1010 		/*
1011 		 * Move completed and retryable entries to our local lists.
1012 		 * If we find a request that requires polling, break out
1013 		 * and complete those lists first, if we have entries there.
1014 		 */
1015 		if (READ_ONCE(req->iopoll_completed))
1016 			break;
1017 
1018 		if (req->opcode == IORING_OP_URING_CMD) {
1019 			struct io_uring_cmd *ioucmd;
1020 
1021 			ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
1022 			ret = file->f_op->uring_cmd_iopoll(ioucmd, &iob,
1023 								poll_flags);
1024 		} else {
1025 			struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1026 
1027 			ret = file->f_op->iopoll(&rw->kiocb, &iob, poll_flags);
1028 		}
1029 		if (unlikely(ret < 0))
1030 			return ret;
1031 		else if (ret)
1032 			poll_flags |= BLK_POLL_ONESHOT;
1033 
1034 		/* iopoll may have completed current req */
1035 		if (!rq_list_empty(iob.req_list) ||
1036 		    READ_ONCE(req->iopoll_completed))
1037 			break;
1038 	}
1039 
1040 	if (!rq_list_empty(iob.req_list))
1041 		iob.complete(&iob);
1042 	else if (!pos)
1043 		return 0;
1044 
1045 	prev = start;
1046 	wq_list_for_each_resume(pos, prev) {
1047 		struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1048 
1049 		/* order with io_complete_rw_iopoll(), e.g. ->result updates */
1050 		if (!smp_load_acquire(&req->iopoll_completed))
1051 			break;
1052 		nr_events++;
1053 		req->cqe.flags = io_put_kbuf(req, 0);
1054 	}
1055 	if (unlikely(!nr_events))
1056 		return 0;
1057 
1058 	pos = start ? start->next : ctx->iopoll_list.first;
1059 	wq_list_cut(&ctx->iopoll_list, prev, start);
1060 
1061 	if (WARN_ON_ONCE(!wq_list_empty(&ctx->submit_state.compl_reqs)))
1062 		return 0;
1063 	ctx->submit_state.compl_reqs.first = pos;
1064 	__io_submit_flush_completions(ctx);
1065 	return nr_events;
1066 }
1067