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
io_file_supports_nowait(struct io_kiocb * req)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
io_iov_compat_buffer_select_prep(struct io_rw * rw)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
io_iov_buffer_select_prep(struct io_kiocb * req)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
io_prep_rw(struct io_kiocb * req,const struct io_uring_sqe * sqe)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
io_readv_writev_cleanup(struct io_kiocb * req)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
io_rw_done(struct kiocb * kiocb,ssize_t ret)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
io_kiocb_update_pos(struct io_kiocb * req)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
io_req_task_queue_reissue(struct io_kiocb * req)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
io_resubmit_prep(struct io_kiocb * req)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
io_rw_should_reissue(struct io_kiocb * req)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
io_resubmit_prep(struct io_kiocb * req)214 static bool io_resubmit_prep(struct io_kiocb *req)
215 {
216 return false;
217 }
io_rw_should_reissue(struct io_kiocb * req)218 static bool io_rw_should_reissue(struct io_kiocb *req)
219 {
220 return false;
221 }
222 #endif
223
io_req_end_write(struct io_kiocb * req)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 */
io_req_io_end(struct io_kiocb * req)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
__io_complete_rw_common(struct io_kiocb * req,long res)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
io_fixup_rw_res(struct io_kiocb * req,long res)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
io_req_rw_complete(struct io_kiocb * req,struct io_tw_state * ts)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
io_complete_rw(struct kiocb * kiocb,long res)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
io_complete_rw_iopoll(struct kiocb * kiocb,long res)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
kiocb_done(struct io_kiocb * req,ssize_t ret,unsigned int issue_flags)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
__io_import_iovec(int ddir,struct io_kiocb * req,struct io_rw_state * s,unsigned int issue_flags)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
io_import_iovec(int rw,struct io_kiocb * req,struct iovec ** iovec,struct io_rw_state * s,unsigned int issue_flags)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
io_kiocb_ppos(struct kiocb * kiocb)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 */
loop_rw_iter(int ddir,struct io_rw * rw,struct iov_iter * iter)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
io_req_map_rw(struct io_kiocb * req,const struct iovec * iovec,const struct iovec * fast_iov,struct iov_iter * iter)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
io_setup_async_rw(struct io_kiocb * req,const struct iovec * iovec,struct io_rw_state * s,bool force)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
io_rw_prep_async(struct io_kiocb * req,int rw)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 iorw->bytes_done = 0;
553 iorw->free_iovec = NULL;
554
555 /* submission path, ->uring_lock should already be taken */
556 ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
557 if (unlikely(ret < 0))
558 return ret;
559
560 if (iov) {
561 iorw->free_iovec = iov;
562 req->flags |= REQ_F_NEED_CLEANUP;
563 }
564
565 return 0;
566 }
567
io_readv_prep_async(struct io_kiocb * req)568 int io_readv_prep_async(struct io_kiocb *req)
569 {
570 return io_rw_prep_async(req, ITER_DEST);
571 }
572
io_writev_prep_async(struct io_kiocb * req)573 int io_writev_prep_async(struct io_kiocb *req)
574 {
575 return io_rw_prep_async(req, ITER_SOURCE);
576 }
577
578 /*
579 * This is our waitqueue callback handler, registered through __folio_lock_async()
580 * when we initially tried to do the IO with the iocb armed our waitqueue.
581 * This gets called when the page is unlocked, and we generally expect that to
582 * happen when the page IO is completed and the page is now uptodate. This will
583 * queue a task_work based retry of the operation, attempting to copy the data
584 * again. If the latter fails because the page was NOT uptodate, then we will
585 * do a thread based blocking retry of the operation. That's the unexpected
586 * slow path.
587 */
io_async_buf_func(struct wait_queue_entry * wait,unsigned mode,int sync,void * arg)588 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
589 int sync, void *arg)
590 {
591 struct wait_page_queue *wpq;
592 struct io_kiocb *req = wait->private;
593 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
594 struct wait_page_key *key = arg;
595
596 wpq = container_of(wait, struct wait_page_queue, wait);
597
598 if (!wake_page_match(wpq, key))
599 return 0;
600
601 rw->kiocb.ki_flags &= ~IOCB_WAITQ;
602 list_del_init(&wait->entry);
603 io_req_task_queue(req);
604 return 1;
605 }
606
607 /*
608 * This controls whether a given IO request should be armed for async page
609 * based retry. If we return false here, the request is handed to the async
610 * worker threads for retry. If we're doing buffered reads on a regular file,
611 * we prepare a private wait_page_queue entry and retry the operation. This
612 * will either succeed because the page is now uptodate and unlocked, or it
613 * will register a callback when the page is unlocked at IO completion. Through
614 * that callback, io_uring uses task_work to setup a retry of the operation.
615 * That retry will attempt the buffered read again. The retry will generally
616 * succeed, or in rare cases where it fails, we then fall back to using the
617 * async worker threads for a blocking retry.
618 */
io_rw_should_retry(struct io_kiocb * req)619 static bool io_rw_should_retry(struct io_kiocb *req)
620 {
621 struct io_async_rw *io = req->async_data;
622 struct wait_page_queue *wait = &io->wpq;
623 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
624 struct kiocb *kiocb = &rw->kiocb;
625
626 /* never retry for NOWAIT, we just complete with -EAGAIN */
627 if (req->flags & REQ_F_NOWAIT)
628 return false;
629
630 /* Only for buffered IO */
631 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
632 return false;
633
634 /*
635 * just use poll if we can, and don't attempt if the fs doesn't
636 * support callback based unlocks
637 */
638 if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
639 return false;
640
641 wait->wait.func = io_async_buf_func;
642 wait->wait.private = req;
643 wait->wait.flags = 0;
644 INIT_LIST_HEAD(&wait->wait.entry);
645 kiocb->ki_flags |= IOCB_WAITQ;
646 kiocb->ki_flags &= ~IOCB_NOWAIT;
647 kiocb->ki_waitq = wait;
648 return true;
649 }
650
io_iter_do_read(struct io_rw * rw,struct iov_iter * iter)651 static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
652 {
653 struct file *file = rw->kiocb.ki_filp;
654
655 if (likely(file->f_op->read_iter))
656 return call_read_iter(file, &rw->kiocb, iter);
657 else if (file->f_op->read)
658 return loop_rw_iter(READ, rw, iter);
659 else
660 return -EINVAL;
661 }
662
need_complete_io(struct io_kiocb * req)663 static bool need_complete_io(struct io_kiocb *req)
664 {
665 return req->flags & REQ_F_ISREG ||
666 S_ISBLK(file_inode(req->file)->i_mode);
667 }
668
io_rw_init_file(struct io_kiocb * req,fmode_t mode)669 static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
670 {
671 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
672 struct kiocb *kiocb = &rw->kiocb;
673 struct io_ring_ctx *ctx = req->ctx;
674 struct file *file = req->file;
675 int ret;
676
677 if (unlikely(!file || !(file->f_mode & mode)))
678 return -EBADF;
679
680 if (!(req->flags & REQ_F_FIXED_FILE))
681 req->flags |= io_file_get_flags(file);
682
683 kiocb->ki_flags = file->f_iocb_flags;
684 ret = kiocb_set_rw_flags(kiocb, rw->flags);
685 if (unlikely(ret))
686 return ret;
687 kiocb->ki_flags |= IOCB_ALLOC_CACHE;
688
689 /*
690 * If the file is marked O_NONBLOCK, still allow retry for it if it
691 * supports async. Otherwise it's impossible to use O_NONBLOCK files
692 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
693 */
694 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
695 ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
696 req->flags |= REQ_F_NOWAIT;
697
698 if (ctx->flags & IORING_SETUP_IOPOLL) {
699 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
700 return -EOPNOTSUPP;
701
702 kiocb->private = NULL;
703 kiocb->ki_flags |= IOCB_HIPRI;
704 kiocb->ki_complete = io_complete_rw_iopoll;
705 req->iopoll_completed = 0;
706 } else {
707 if (kiocb->ki_flags & IOCB_HIPRI)
708 return -EINVAL;
709 kiocb->ki_complete = io_complete_rw;
710 }
711
712 return 0;
713 }
714
io_read(struct io_kiocb * req,unsigned int issue_flags)715 int io_read(struct io_kiocb *req, unsigned int issue_flags)
716 {
717 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
718 struct io_rw_state __s, *s = &__s;
719 struct iovec *iovec;
720 struct kiocb *kiocb = &rw->kiocb;
721 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
722 struct io_async_rw *io;
723 ssize_t ret, ret2;
724 loff_t *ppos;
725
726 if (!req_has_async_data(req)) {
727 ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
728 if (unlikely(ret < 0))
729 return ret;
730 } else {
731 io = req->async_data;
732 s = &io->s;
733
734 /*
735 * Safe and required to re-import if we're using provided
736 * buffers, as we dropped the selected one before retry.
737 */
738 if (io_do_buffer_select(req)) {
739 ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
740 if (unlikely(ret < 0))
741 return ret;
742 }
743
744 /*
745 * We come here from an earlier attempt, restore our state to
746 * match in case it doesn't. It's cheap enough that we don't
747 * need to make this conditional.
748 */
749 iov_iter_restore(&s->iter, &s->iter_state);
750 iovec = NULL;
751 }
752 ret = io_rw_init_file(req, FMODE_READ);
753 if (unlikely(ret)) {
754 kfree(iovec);
755 return ret;
756 }
757 req->cqe.res = iov_iter_count(&s->iter);
758
759 if (force_nonblock) {
760 /* If the file doesn't support async, just async punt */
761 if (unlikely(!io_file_supports_nowait(req))) {
762 ret = io_setup_async_rw(req, iovec, s, true);
763 return ret ?: -EAGAIN;
764 }
765 kiocb->ki_flags |= IOCB_NOWAIT;
766 } else {
767 /* Ensure we clear previously set non-block flag */
768 kiocb->ki_flags &= ~IOCB_NOWAIT;
769 }
770
771 ppos = io_kiocb_update_pos(req);
772
773 ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
774 if (unlikely(ret)) {
775 kfree(iovec);
776 return ret;
777 }
778
779 ret = io_iter_do_read(rw, &s->iter);
780
781 if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
782 req->flags &= ~REQ_F_REISSUE;
783 /* if we can poll, just do that */
784 if (req->opcode == IORING_OP_READ && file_can_poll(req->file))
785 return -EAGAIN;
786 /* IOPOLL retry should happen for io-wq threads */
787 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
788 goto done;
789 /* no retry on NONBLOCK nor RWF_NOWAIT */
790 if (req->flags & REQ_F_NOWAIT)
791 goto done;
792 ret = 0;
793 } else if (ret == -EIOCBQUEUED) {
794 if (iovec)
795 kfree(iovec);
796 return IOU_ISSUE_SKIP_COMPLETE;
797 } else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
798 (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) {
799 /* read all, failed, already did sync or don't want to retry */
800 goto done;
801 }
802
803 /*
804 * Don't depend on the iter state matching what was consumed, or being
805 * untouched in case of error. Restore it and we'll advance it
806 * manually if we need to.
807 */
808 iov_iter_restore(&s->iter, &s->iter_state);
809
810 ret2 = io_setup_async_rw(req, iovec, s, true);
811 iovec = NULL;
812 if (ret2) {
813 ret = ret > 0 ? ret : ret2;
814 goto done;
815 }
816
817 io = req->async_data;
818 s = &io->s;
819 /*
820 * Now use our persistent iterator and state, if we aren't already.
821 * We've restored and mapped the iter to match.
822 */
823
824 do {
825 /*
826 * We end up here because of a partial read, either from
827 * above or inside this loop. Advance the iter by the bytes
828 * that were consumed.
829 */
830 iov_iter_advance(&s->iter, ret);
831 if (!iov_iter_count(&s->iter))
832 break;
833 io->bytes_done += ret;
834 iov_iter_save_state(&s->iter, &s->iter_state);
835
836 /* if we can retry, do so with the callbacks armed */
837 if (!io_rw_should_retry(req)) {
838 kiocb->ki_flags &= ~IOCB_WAITQ;
839 return -EAGAIN;
840 }
841
842 req->cqe.res = iov_iter_count(&s->iter);
843 /*
844 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
845 * we get -EIOCBQUEUED, then we'll get a notification when the
846 * desired page gets unlocked. We can also get a partial read
847 * here, and if we do, then just retry at the new offset.
848 */
849 ret = io_iter_do_read(rw, &s->iter);
850 if (ret == -EIOCBQUEUED)
851 return IOU_ISSUE_SKIP_COMPLETE;
852 /* we got some bytes, but not all. retry. */
853 kiocb->ki_flags &= ~IOCB_WAITQ;
854 iov_iter_restore(&s->iter, &s->iter_state);
855 } while (ret > 0);
856 done:
857 /* it's faster to check here then delegate to kfree */
858 if (iovec)
859 kfree(iovec);
860 return kiocb_done(req, ret, issue_flags);
861 }
862
io_write(struct io_kiocb * req,unsigned int issue_flags)863 int io_write(struct io_kiocb *req, unsigned int issue_flags)
864 {
865 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
866 struct io_rw_state __s, *s = &__s;
867 struct iovec *iovec;
868 struct kiocb *kiocb = &rw->kiocb;
869 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
870 ssize_t ret, ret2;
871 loff_t *ppos;
872
873 if (!req_has_async_data(req)) {
874 ret = io_import_iovec(ITER_SOURCE, req, &iovec, s, issue_flags);
875 if (unlikely(ret < 0))
876 return ret;
877 } else {
878 struct io_async_rw *io = req->async_data;
879
880 s = &io->s;
881 iov_iter_restore(&s->iter, &s->iter_state);
882 iovec = NULL;
883 }
884 ret = io_rw_init_file(req, FMODE_WRITE);
885 if (unlikely(ret)) {
886 kfree(iovec);
887 return ret;
888 }
889 req->cqe.res = iov_iter_count(&s->iter);
890
891 if (force_nonblock) {
892 /* If the file doesn't support async, just async punt */
893 if (unlikely(!io_file_supports_nowait(req)))
894 goto copy_iov;
895
896 /* File path supports NOWAIT for non-direct_IO only for block devices. */
897 if (!(kiocb->ki_flags & IOCB_DIRECT) &&
898 !(kiocb->ki_filp->f_mode & FMODE_BUF_WASYNC) &&
899 (req->flags & REQ_F_ISREG))
900 goto copy_iov;
901
902 kiocb->ki_flags |= IOCB_NOWAIT;
903 } else {
904 /* Ensure we clear previously set non-block flag */
905 kiocb->ki_flags &= ~IOCB_NOWAIT;
906 }
907
908 ppos = io_kiocb_update_pos(req);
909
910 ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
911 if (unlikely(ret)) {
912 kfree(iovec);
913 return ret;
914 }
915
916 if (req->flags & REQ_F_ISREG)
917 kiocb_start_write(kiocb);
918 kiocb->ki_flags |= IOCB_WRITE;
919
920 if (likely(req->file->f_op->write_iter))
921 ret2 = call_write_iter(req->file, kiocb, &s->iter);
922 else if (req->file->f_op->write)
923 ret2 = loop_rw_iter(WRITE, rw, &s->iter);
924 else
925 ret2 = -EINVAL;
926
927 if (req->flags & REQ_F_REISSUE) {
928 req->flags &= ~REQ_F_REISSUE;
929 ret2 = -EAGAIN;
930 }
931
932 /*
933 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
934 * retry them without IOCB_NOWAIT.
935 */
936 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
937 ret2 = -EAGAIN;
938 /* no retry on NONBLOCK nor RWF_NOWAIT */
939 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
940 goto done;
941 if (!force_nonblock || ret2 != -EAGAIN) {
942 /* IOPOLL retry should happen for io-wq threads */
943 if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
944 goto copy_iov;
945
946 if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) {
947 struct io_async_rw *io;
948
949 trace_io_uring_short_write(req->ctx, kiocb->ki_pos - ret2,
950 req->cqe.res, ret2);
951
952 /* This is a partial write. The file pos has already been
953 * updated, setup the async struct to complete the request
954 * in the worker. Also update bytes_done to account for
955 * the bytes already written.
956 */
957 iov_iter_save_state(&s->iter, &s->iter_state);
958 ret = io_setup_async_rw(req, iovec, s, true);
959
960 io = req->async_data;
961 if (io)
962 io->bytes_done += ret2;
963
964 if (kiocb->ki_flags & IOCB_WRITE)
965 io_req_end_write(req);
966 return ret ? ret : -EAGAIN;
967 }
968 done:
969 ret = kiocb_done(req, ret2, issue_flags);
970 } else {
971 copy_iov:
972 iov_iter_restore(&s->iter, &s->iter_state);
973 ret = io_setup_async_rw(req, iovec, s, false);
974 if (!ret) {
975 if (kiocb->ki_flags & IOCB_WRITE)
976 io_req_end_write(req);
977 return -EAGAIN;
978 }
979 return ret;
980 }
981 /* it's reportedly faster than delegating the null check to kfree() */
982 if (iovec)
983 kfree(iovec);
984 return ret;
985 }
986
io_rw_fail(struct io_kiocb * req)987 void io_rw_fail(struct io_kiocb *req)
988 {
989 int res;
990
991 res = io_fixup_rw_res(req, req->cqe.res);
992 io_req_set_res(req, res, req->cqe.flags);
993 }
994
io_do_iopoll(struct io_ring_ctx * ctx,bool force_nonspin)995 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
996 {
997 struct io_wq_work_node *pos, *start, *prev;
998 unsigned int poll_flags = 0;
999 DEFINE_IO_COMP_BATCH(iob);
1000 int nr_events = 0;
1001
1002 /*
1003 * Only spin for completions if we don't have multiple devices hanging
1004 * off our complete list.
1005 */
1006 if (ctx->poll_multi_queue || force_nonspin)
1007 poll_flags |= BLK_POLL_ONESHOT;
1008
1009 wq_list_for_each(pos, start, &ctx->iopoll_list) {
1010 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1011 struct file *file = req->file;
1012 int ret;
1013
1014 /*
1015 * Move completed and retryable entries to our local lists.
1016 * If we find a request that requires polling, break out
1017 * and complete those lists first, if we have entries there.
1018 */
1019 if (READ_ONCE(req->iopoll_completed))
1020 break;
1021
1022 if (req->opcode == IORING_OP_URING_CMD) {
1023 struct io_uring_cmd *ioucmd;
1024
1025 ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
1026 ret = file->f_op->uring_cmd_iopoll(ioucmd, &iob,
1027 poll_flags);
1028 } else {
1029 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1030
1031 ret = file->f_op->iopoll(&rw->kiocb, &iob, poll_flags);
1032 }
1033 if (unlikely(ret < 0))
1034 return ret;
1035 else if (ret)
1036 poll_flags |= BLK_POLL_ONESHOT;
1037
1038 /* iopoll may have completed current req */
1039 if (!rq_list_empty(iob.req_list) ||
1040 READ_ONCE(req->iopoll_completed))
1041 break;
1042 }
1043
1044 if (!rq_list_empty(iob.req_list))
1045 iob.complete(&iob);
1046 else if (!pos)
1047 return 0;
1048
1049 prev = start;
1050 wq_list_for_each_resume(pos, prev) {
1051 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1052
1053 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
1054 if (!smp_load_acquire(&req->iopoll_completed))
1055 break;
1056 nr_events++;
1057 req->cqe.flags = io_put_kbuf(req, 0);
1058 }
1059 if (unlikely(!nr_events))
1060 return 0;
1061
1062 pos = start ? start->next : ctx->iopoll_list.first;
1063 wq_list_cut(&ctx->iopoll_list, prev, start);
1064
1065 if (WARN_ON_ONCE(!wq_list_empty(&ctx->submit_state.compl_reqs)))
1066 return 0;
1067 ctx->submit_state.compl_reqs.first = pos;
1068 __io_submit_flush_completions(ctx);
1069 return nr_events;
1070 }
1071