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