Merge branch 'md-fixes' of https://git.kernel.org/pub/scm/linux/kernel/git/song/md...
[platform/kernel/linux-starfive.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_ubuf(ddir, buf, sqe_len, 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 (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 (iter_is_ubuf(iter)) {
454                         iovec.iov_base = iter->ubuf + iter->iov_offset;
455                         iovec.iov_len = iov_iter_count(iter);
456                 } else if (!iov_iter_is_bvec(iter)) {
457                         iovec = iov_iter_iovec(iter);
458                 } else {
459                         iovec.iov_base = u64_to_user_ptr(rw->addr);
460                         iovec.iov_len = rw->len;
461                 }
462
463                 if (ddir == READ) {
464                         nr = file->f_op->read(file, iovec.iov_base,
465                                               iovec.iov_len, ppos);
466                 } else {
467                         nr = file->f_op->write(file, iovec.iov_base,
468                                                iovec.iov_len, ppos);
469                 }
470
471                 if (nr < 0) {
472                         if (!ret)
473                                 ret = nr;
474                         break;
475                 }
476                 ret += nr;
477                 if (!iov_iter_is_bvec(iter)) {
478                         iov_iter_advance(iter, nr);
479                 } else {
480                         rw->addr += nr;
481                         rw->len -= nr;
482                         if (!rw->len)
483                                 break;
484                 }
485                 if (nr != iovec.iov_len)
486                         break;
487         }
488
489         return ret;
490 }
491
492 static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec,
493                           const struct iovec *fast_iov, struct iov_iter *iter)
494 {
495         struct io_async_rw *io = req->async_data;
496
497         memcpy(&io->s.iter, iter, sizeof(*iter));
498         io->free_iovec = iovec;
499         io->bytes_done = 0;
500         /* can only be fixed buffers, no need to do anything */
501         if (iov_iter_is_bvec(iter) || iter_is_ubuf(iter))
502                 return;
503         if (!iovec) {
504                 unsigned iov_off = 0;
505
506                 io->s.iter.iov = io->s.fast_iov;
507                 if (iter->iov != fast_iov) {
508                         iov_off = iter->iov - fast_iov;
509                         io->s.iter.iov += iov_off;
510                 }
511                 if (io->s.fast_iov != fast_iov)
512                         memcpy(io->s.fast_iov + iov_off, fast_iov + iov_off,
513                                sizeof(struct iovec) * iter->nr_segs);
514         } else {
515                 req->flags |= REQ_F_NEED_CLEANUP;
516         }
517 }
518
519 static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec,
520                              struct io_rw_state *s, bool force)
521 {
522         if (!force && !io_cold_defs[req->opcode].prep_async)
523                 return 0;
524         if (!req_has_async_data(req)) {
525                 struct io_async_rw *iorw;
526
527                 if (io_alloc_async_data(req)) {
528                         kfree(iovec);
529                         return -ENOMEM;
530                 }
531
532                 io_req_map_rw(req, iovec, s->fast_iov, &s->iter);
533                 iorw = req->async_data;
534                 /* we've copied and mapped the iter, ensure state is saved */
535                 iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state);
536         }
537         return 0;
538 }
539
540 static inline int io_rw_prep_async(struct io_kiocb *req, int rw)
541 {
542         struct io_async_rw *iorw = req->async_data;
543         struct iovec *iov;
544         int ret;
545
546         /* submission path, ->uring_lock should already be taken */
547         ret = io_import_iovec(rw, req, &iov, &iorw->s, 0);
548         if (unlikely(ret < 0))
549                 return ret;
550
551         iorw->bytes_done = 0;
552         iorw->free_iovec = iov;
553         if (iov)
554                 req->flags |= REQ_F_NEED_CLEANUP;
555         return 0;
556 }
557
558 int io_readv_prep_async(struct io_kiocb *req)
559 {
560         return io_rw_prep_async(req, ITER_DEST);
561 }
562
563 int io_writev_prep_async(struct io_kiocb *req)
564 {
565         return io_rw_prep_async(req, ITER_SOURCE);
566 }
567
568 /*
569  * This is our waitqueue callback handler, registered through __folio_lock_async()
570  * when we initially tried to do the IO with the iocb armed our waitqueue.
571  * This gets called when the page is unlocked, and we generally expect that to
572  * happen when the page IO is completed and the page is now uptodate. This will
573  * queue a task_work based retry of the operation, attempting to copy the data
574  * again. If the latter fails because the page was NOT uptodate, then we will
575  * do a thread based blocking retry of the operation. That's the unexpected
576  * slow path.
577  */
578 static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
579                              int sync, void *arg)
580 {
581         struct wait_page_queue *wpq;
582         struct io_kiocb *req = wait->private;
583         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
584         struct wait_page_key *key = arg;
585
586         wpq = container_of(wait, struct wait_page_queue, wait);
587
588         if (!wake_page_match(wpq, key))
589                 return 0;
590
591         rw->kiocb.ki_flags &= ~IOCB_WAITQ;
592         list_del_init(&wait->entry);
593         io_req_task_queue(req);
594         return 1;
595 }
596
597 /*
598  * This controls whether a given IO request should be armed for async page
599  * based retry. If we return false here, the request is handed to the async
600  * worker threads for retry. If we're doing buffered reads on a regular file,
601  * we prepare a private wait_page_queue entry and retry the operation. This
602  * will either succeed because the page is now uptodate and unlocked, or it
603  * will register a callback when the page is unlocked at IO completion. Through
604  * that callback, io_uring uses task_work to setup a retry of the operation.
605  * That retry will attempt the buffered read again. The retry will generally
606  * succeed, or in rare cases where it fails, we then fall back to using the
607  * async worker threads for a blocking retry.
608  */
609 static bool io_rw_should_retry(struct io_kiocb *req)
610 {
611         struct io_async_rw *io = req->async_data;
612         struct wait_page_queue *wait = &io->wpq;
613         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
614         struct kiocb *kiocb = &rw->kiocb;
615
616         /* never retry for NOWAIT, we just complete with -EAGAIN */
617         if (req->flags & REQ_F_NOWAIT)
618                 return false;
619
620         /* Only for buffered IO */
621         if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
622                 return false;
623
624         /*
625          * just use poll if we can, and don't attempt if the fs doesn't
626          * support callback based unlocks
627          */
628         if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC))
629                 return false;
630
631         wait->wait.func = io_async_buf_func;
632         wait->wait.private = req;
633         wait->wait.flags = 0;
634         INIT_LIST_HEAD(&wait->wait.entry);
635         kiocb->ki_flags |= IOCB_WAITQ;
636         kiocb->ki_flags &= ~IOCB_NOWAIT;
637         kiocb->ki_waitq = wait;
638         return true;
639 }
640
641 static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
642 {
643         struct file *file = rw->kiocb.ki_filp;
644
645         if (likely(file->f_op->read_iter))
646                 return call_read_iter(file, &rw->kiocb, iter);
647         else if (file->f_op->read)
648                 return loop_rw_iter(READ, rw, iter);
649         else
650                 return -EINVAL;
651 }
652
653 static bool need_complete_io(struct io_kiocb *req)
654 {
655         return req->flags & REQ_F_ISREG ||
656                 S_ISBLK(file_inode(req->file)->i_mode);
657 }
658
659 static int io_rw_init_file(struct io_kiocb *req, fmode_t mode)
660 {
661         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
662         struct kiocb *kiocb = &rw->kiocb;
663         struct io_ring_ctx *ctx = req->ctx;
664         struct file *file = req->file;
665         int ret;
666
667         if (unlikely(!file || !(file->f_mode & mode)))
668                 return -EBADF;
669
670         if (!io_req_ffs_set(req))
671                 req->flags |= io_file_get_flags(file) << REQ_F_SUPPORT_NOWAIT_BIT;
672
673         kiocb->ki_flags = file->f_iocb_flags;
674         ret = kiocb_set_rw_flags(kiocb, rw->flags);
675         if (unlikely(ret))
676                 return ret;
677         kiocb->ki_flags |= IOCB_ALLOC_CACHE;
678
679         /*
680          * If the file is marked O_NONBLOCK, still allow retry for it if it
681          * supports async. Otherwise it's impossible to use O_NONBLOCK files
682          * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
683          */
684         if ((kiocb->ki_flags & IOCB_NOWAIT) ||
685             ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req)))
686                 req->flags |= REQ_F_NOWAIT;
687
688         if (ctx->flags & IORING_SETUP_IOPOLL) {
689                 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
690                         return -EOPNOTSUPP;
691
692                 kiocb->private = NULL;
693                 kiocb->ki_flags |= IOCB_HIPRI;
694                 kiocb->ki_complete = io_complete_rw_iopoll;
695                 req->iopoll_completed = 0;
696         } else {
697                 if (kiocb->ki_flags & IOCB_HIPRI)
698                         return -EINVAL;
699                 kiocb->ki_complete = io_complete_rw;
700         }
701
702         return 0;
703 }
704
705 int io_read(struct io_kiocb *req, unsigned int issue_flags)
706 {
707         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
708         struct io_rw_state __s, *s = &__s;
709         struct iovec *iovec;
710         struct kiocb *kiocb = &rw->kiocb;
711         bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
712         struct io_async_rw *io;
713         ssize_t ret, ret2;
714         loff_t *ppos;
715
716         if (!req_has_async_data(req)) {
717                 ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
718                 if (unlikely(ret < 0))
719                         return ret;
720         } else {
721                 io = req->async_data;
722                 s = &io->s;
723
724                 /*
725                  * Safe and required to re-import if we're using provided
726                  * buffers, as we dropped the selected one before retry.
727                  */
728                 if (io_do_buffer_select(req)) {
729                         ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags);
730                         if (unlikely(ret < 0))
731                                 return ret;
732                 }
733
734                 /*
735                  * We come here from an earlier attempt, restore our state to
736                  * match in case it doesn't. It's cheap enough that we don't
737                  * need to make this conditional.
738                  */
739                 iov_iter_restore(&s->iter, &s->iter_state);
740                 iovec = NULL;
741         }
742         ret = io_rw_init_file(req, FMODE_READ);
743         if (unlikely(ret)) {
744                 kfree(iovec);
745                 return ret;
746         }
747         req->cqe.res = iov_iter_count(&s->iter);
748
749         if (force_nonblock) {
750                 /* If the file doesn't support async, just async punt */
751                 if (unlikely(!io_file_supports_nowait(req))) {
752                         ret = io_setup_async_rw(req, iovec, s, true);
753                         return ret ?: -EAGAIN;
754                 }
755                 kiocb->ki_flags |= IOCB_NOWAIT;
756         } else {
757                 /* Ensure we clear previously set non-block flag */
758                 kiocb->ki_flags &= ~IOCB_NOWAIT;
759         }
760
761         ppos = io_kiocb_update_pos(req);
762
763         ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
764         if (unlikely(ret)) {
765                 kfree(iovec);
766                 return ret;
767         }
768
769         ret = io_iter_do_read(rw, &s->iter);
770
771         if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) {
772                 req->flags &= ~REQ_F_REISSUE;
773                 /* if we can poll, just do that */
774                 if (req->opcode == IORING_OP_READ && file_can_poll(req->file))
775                         return -EAGAIN;
776                 /* IOPOLL retry should happen for io-wq threads */
777                 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
778                         goto done;
779                 /* no retry on NONBLOCK nor RWF_NOWAIT */
780                 if (req->flags & REQ_F_NOWAIT)
781                         goto done;
782                 ret = 0;
783         } else if (ret == -EIOCBQUEUED) {
784                 if (iovec)
785                         kfree(iovec);
786                 return IOU_ISSUE_SKIP_COMPLETE;
787         } else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
788                    (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) {
789                 /* read all, failed, already did sync or don't want to retry */
790                 goto done;
791         }
792
793         /*
794          * Don't depend on the iter state matching what was consumed, or being
795          * untouched in case of error. Restore it and we'll advance it
796          * manually if we need to.
797          */
798         iov_iter_restore(&s->iter, &s->iter_state);
799
800         ret2 = io_setup_async_rw(req, iovec, s, true);
801         iovec = NULL;
802         if (ret2) {
803                 ret = ret > 0 ? ret : ret2;
804                 goto done;
805         }
806
807         io = req->async_data;
808         s = &io->s;
809         /*
810          * Now use our persistent iterator and state, if we aren't already.
811          * We've restored and mapped the iter to match.
812          */
813
814         do {
815                 /*
816                  * We end up here because of a partial read, either from
817                  * above or inside this loop. Advance the iter by the bytes
818                  * that were consumed.
819                  */
820                 iov_iter_advance(&s->iter, ret);
821                 if (!iov_iter_count(&s->iter))
822                         break;
823                 io->bytes_done += ret;
824                 iov_iter_save_state(&s->iter, &s->iter_state);
825
826                 /* if we can retry, do so with the callbacks armed */
827                 if (!io_rw_should_retry(req)) {
828                         kiocb->ki_flags &= ~IOCB_WAITQ;
829                         return -EAGAIN;
830                 }
831
832                 req->cqe.res = iov_iter_count(&s->iter);
833                 /*
834                  * Now retry read with the IOCB_WAITQ parts set in the iocb. If
835                  * we get -EIOCBQUEUED, then we'll get a notification when the
836                  * desired page gets unlocked. We can also get a partial read
837                  * here, and if we do, then just retry at the new offset.
838                  */
839                 ret = io_iter_do_read(rw, &s->iter);
840                 if (ret == -EIOCBQUEUED)
841                         return IOU_ISSUE_SKIP_COMPLETE;
842                 /* we got some bytes, but not all. retry. */
843                 kiocb->ki_flags &= ~IOCB_WAITQ;
844                 iov_iter_restore(&s->iter, &s->iter_state);
845         } while (ret > 0);
846 done:
847         /* it's faster to check here then delegate to kfree */
848         if (iovec)
849                 kfree(iovec);
850         return kiocb_done(req, ret, issue_flags);
851 }
852
853 int io_write(struct io_kiocb *req, unsigned int issue_flags)
854 {
855         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
856         struct io_rw_state __s, *s = &__s;
857         struct iovec *iovec;
858         struct kiocb *kiocb = &rw->kiocb;
859         bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
860         ssize_t ret, ret2;
861         loff_t *ppos;
862
863         if (!req_has_async_data(req)) {
864                 ret = io_import_iovec(ITER_SOURCE, req, &iovec, s, issue_flags);
865                 if (unlikely(ret < 0))
866                         return ret;
867         } else {
868                 struct io_async_rw *io = req->async_data;
869
870                 s = &io->s;
871                 iov_iter_restore(&s->iter, &s->iter_state);
872                 iovec = NULL;
873         }
874         ret = io_rw_init_file(req, FMODE_WRITE);
875         if (unlikely(ret)) {
876                 kfree(iovec);
877                 return ret;
878         }
879         req->cqe.res = iov_iter_count(&s->iter);
880
881         if (force_nonblock) {
882                 /* If the file doesn't support async, just async punt */
883                 if (unlikely(!io_file_supports_nowait(req)))
884                         goto copy_iov;
885
886                 /* File path supports NOWAIT for non-direct_IO only for block devices. */
887                 if (!(kiocb->ki_flags & IOCB_DIRECT) &&
888                         !(kiocb->ki_filp->f_mode & FMODE_BUF_WASYNC) &&
889                         (req->flags & REQ_F_ISREG))
890                         goto copy_iov;
891
892                 kiocb->ki_flags |= IOCB_NOWAIT;
893         } else {
894                 /* Ensure we clear previously set non-block flag */
895                 kiocb->ki_flags &= ~IOCB_NOWAIT;
896         }
897
898         ppos = io_kiocb_update_pos(req);
899
900         ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
901         if (unlikely(ret)) {
902                 kfree(iovec);
903                 return ret;
904         }
905
906         /*
907          * Open-code file_start_write here to grab freeze protection,
908          * which will be released by another thread in
909          * io_complete_rw().  Fool lockdep by telling it the lock got
910          * released so that it doesn't complain about the held lock when
911          * we return to userspace.
912          */
913         if (req->flags & REQ_F_ISREG) {
914                 sb_start_write(file_inode(req->file)->i_sb);
915                 __sb_writers_release(file_inode(req->file)->i_sb,
916                                         SB_FREEZE_WRITE);
917         }
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                                 kiocb_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                                 kiocb_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
987 static void io_cqring_ev_posted_iopoll(struct io_ring_ctx *ctx)
988 {
989         io_commit_cqring_flush(ctx);
990         if (ctx->flags & IORING_SETUP_SQPOLL)
991                 io_cqring_wake(ctx);
992 }
993
994 void io_rw_fail(struct io_kiocb *req)
995 {
996         int res;
997
998         res = io_fixup_rw_res(req, req->cqe.res);
999         io_req_set_res(req, res, req->cqe.flags);
1000 }
1001
1002 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
1003 {
1004         struct io_wq_work_node *pos, *start, *prev;
1005         unsigned int poll_flags = BLK_POLL_NOSLEEP;
1006         DEFINE_IO_COMP_BATCH(iob);
1007         int nr_events = 0;
1008
1009         /*
1010          * Only spin for completions if we don't have multiple devices hanging
1011          * off our complete list.
1012          */
1013         if (ctx->poll_multi_queue || force_nonspin)
1014                 poll_flags |= BLK_POLL_ONESHOT;
1015
1016         wq_list_for_each(pos, start, &ctx->iopoll_list) {
1017                 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1018                 struct file *file = req->file;
1019                 int ret;
1020
1021                 /*
1022                  * Move completed and retryable entries to our local lists.
1023                  * If we find a request that requires polling, break out
1024                  * and complete those lists first, if we have entries there.
1025                  */
1026                 if (READ_ONCE(req->iopoll_completed))
1027                         break;
1028
1029                 if (req->opcode == IORING_OP_URING_CMD) {
1030                         struct io_uring_cmd *ioucmd;
1031
1032                         ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
1033                         ret = file->f_op->uring_cmd_iopoll(ioucmd, &iob,
1034                                                                 poll_flags);
1035                 } else {
1036                         struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1037
1038                         ret = file->f_op->iopoll(&rw->kiocb, &iob, poll_flags);
1039                 }
1040                 if (unlikely(ret < 0))
1041                         return ret;
1042                 else if (ret)
1043                         poll_flags |= BLK_POLL_ONESHOT;
1044
1045                 /* iopoll may have completed current req */
1046                 if (!rq_list_empty(iob.req_list) ||
1047                     READ_ONCE(req->iopoll_completed))
1048                         break;
1049         }
1050
1051         if (!rq_list_empty(iob.req_list))
1052                 iob.complete(&iob);
1053         else if (!pos)
1054                 return 0;
1055
1056         prev = start;
1057         wq_list_for_each_resume(pos, prev) {
1058                 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1059
1060                 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
1061                 if (!smp_load_acquire(&req->iopoll_completed))
1062                         break;
1063                 nr_events++;
1064                 if (unlikely(req->flags & REQ_F_CQE_SKIP))
1065                         continue;
1066
1067                 req->cqe.flags = io_put_kbuf(req, 0);
1068                 if (unlikely(!__io_fill_cqe_req(ctx, req))) {
1069                         spin_lock(&ctx->completion_lock);
1070                         io_req_cqe_overflow(req);
1071                         spin_unlock(&ctx->completion_lock);
1072                 }
1073         }
1074
1075         if (unlikely(!nr_events))
1076                 return 0;
1077
1078         io_commit_cqring(ctx);
1079         io_cqring_ev_posted_iopoll(ctx);
1080         pos = start ? start->next : ctx->iopoll_list.first;
1081         wq_list_cut(&ctx->iopoll_list, prev, start);
1082         io_free_batch_list(ctx, pos);
1083         return nr_events;
1084 }