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/cmd.h>
14#include <linux/indirect_call_wrapper.h>
15
16#include <uapi/linux/io_uring.h>
17
18#include "filetable.h"
19#include "io_uring.h"
20#include "opdef.h"
21#include "kbuf.h"
22#include "alloc_cache.h"
23#include "rsrc.h"
24#include "poll.h"
25#include "rw.h"
26
27static void io_complete_rw(struct kiocb *kiocb, long res);
28static void io_complete_rw_iopoll(struct kiocb *kiocb, long res);
29
30struct io_rw {
31 /* NOTE: kiocb has the file as the first member, so don't do it here */
32 struct kiocb kiocb;
33 u64 addr;
34 u32 len;
35 rwf_t flags;
36};
37
38static bool io_file_supports_nowait(struct io_kiocb *req, __poll_t mask)
39{
40 /* If FMODE_NOWAIT is set for a file, we're golden */
41 if (req->flags & REQ_F_SUPPORT_NOWAIT)
42 return true;
43 /* No FMODE_NOWAIT, if we can poll, check the status */
44 if (io_file_can_poll(req)) {
45 struct poll_table_struct pt = { ._key = mask };
46
47 return vfs_poll(file: req->file, pt: &pt) & mask;
48 }
49 /* No FMODE_NOWAIT support, and file isn't pollable. Tough luck. */
50 return false;
51}
52
53static int io_iov_compat_buffer_select_prep(struct io_rw *rw)
54{
55 struct compat_iovec __user *uiov = u64_to_user_ptr(rw->addr);
56 struct compat_iovec iov;
57
58 if (copy_from_user(to: &iov, from: uiov, n: sizeof(iov)))
59 return -EFAULT;
60 rw->len = iov.iov_len;
61 return 0;
62}
63
64static int io_iov_buffer_select_prep(struct io_kiocb *req)
65{
66 struct iovec __user *uiov;
67 struct iovec iov;
68 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
69
70 if (rw->len != 1)
71 return -EINVAL;
72
73 if (io_is_compat(ctx: req->ctx))
74 return io_iov_compat_buffer_select_prep(rw);
75
76 uiov = u64_to_user_ptr(rw->addr);
77 if (copy_from_user(to: &iov, from: uiov, n: sizeof(*uiov)))
78 return -EFAULT;
79 rw->len = iov.iov_len;
80 return 0;
81}
82
83static int io_import_vec(int ddir, struct io_kiocb *req,
84 struct io_async_rw *io,
85 const struct iovec __user *uvec,
86 size_t uvec_segs)
87{
88 int ret, nr_segs;
89 struct iovec *iov;
90
91 if (io->vec.iovec) {
92 nr_segs = io->vec.nr;
93 iov = io->vec.iovec;
94 } else {
95 nr_segs = 1;
96 iov = &io->fast_iov;
97 }
98
99 ret = __import_iovec(type: ddir, uvec, nr_segs: uvec_segs, fast_segs: nr_segs, iovp: &iov, i: &io->iter,
100 compat: io_is_compat(ctx: req->ctx));
101 if (unlikely(ret < 0))
102 return ret;
103 if (iov) {
104 req->flags |= REQ_F_NEED_CLEANUP;
105 io_vec_reset_iovec(iv: &io->vec, iovec: iov, nr: io->iter.nr_segs);
106 }
107 return 0;
108}
109
110static int __io_import_rw_buffer(int ddir, struct io_kiocb *req,
111 struct io_async_rw *io, struct io_br_sel *sel,
112 unsigned int issue_flags)
113{
114 const struct io_issue_def *def = &io_issue_defs[req->opcode];
115 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
116 size_t sqe_len = rw->len;
117
118 sel->addr = u64_to_user_ptr(rw->addr);
119 if (def->vectored && !(req->flags & REQ_F_BUFFER_SELECT))
120 return io_import_vec(ddir, req, io, uvec: sel->addr, uvec_segs: sqe_len);
121
122 if (io_do_buffer_select(req)) {
123 *sel = io_buffer_select(req, len: &sqe_len, buf_group: io->buf_group, issue_flags);
124 if (!sel->addr)
125 return -ENOBUFS;
126 rw->addr = (unsigned long) sel->addr;
127 rw->len = sqe_len;
128 }
129 return import_ubuf(type: ddir, buf: sel->addr, len: sqe_len, i: &io->iter);
130}
131
132static inline int io_import_rw_buffer(int rw, struct io_kiocb *req,
133 struct io_async_rw *io,
134 struct io_br_sel *sel,
135 unsigned int issue_flags)
136{
137 int ret;
138
139 ret = __io_import_rw_buffer(ddir: rw, req, io, sel, issue_flags);
140 if (unlikely(ret < 0))
141 return ret;
142
143 iov_iter_save_state(iter: &io->iter, state: &io->iter_state);
144 return 0;
145}
146
147static void io_rw_recycle(struct io_kiocb *req, unsigned int issue_flags)
148{
149 struct io_async_rw *rw = req->async_data;
150
151 if (unlikely(issue_flags & IO_URING_F_UNLOCKED))
152 return;
153
154 io_alloc_cache_vec_kasan(iv: &rw->vec);
155 if (rw->vec.nr > IO_VEC_CACHE_SOFT_CAP)
156 io_vec_free(iv: &rw->vec);
157
158 if (io_alloc_cache_put(cache: &req->ctx->rw_cache, entry: rw))
159 io_req_async_data_clear(req, extra_flags: 0);
160}
161
162static void io_req_rw_cleanup(struct io_kiocb *req, unsigned int issue_flags)
163{
164 /*
165 * Disable quick recycling for anything that's gone through io-wq.
166 * In theory, this should be fine to cleanup. However, some read or
167 * write iter handling touches the iovec AFTER having called into the
168 * handler, eg to reexpand or revert. This means we can have:
169 *
170 * task io-wq
171 * issue
172 * punt to io-wq
173 * issue
174 * blkdev_write_iter()
175 * ->ki_complete()
176 * io_complete_rw()
177 * queue tw complete
178 * run tw
179 * req_rw_cleanup
180 * iov_iter_count() <- look at iov_iter again
181 *
182 * which can lead to a UAF. This is only possible for io-wq offload
183 * as the cleanup can run in parallel. As io-wq is not the fast path,
184 * just leave cleanup to the end.
185 *
186 * This is really a bug in the core code that does this, any issue
187 * path should assume that a successful (or -EIOCBQUEUED) return can
188 * mean that the underlying data can be gone at any time. But that
189 * should be fixed seperately, and then this check could be killed.
190 */
191 if (!(req->flags & (REQ_F_REISSUE | REQ_F_REFCOUNT))) {
192 req->flags &= ~REQ_F_NEED_CLEANUP;
193 io_rw_recycle(req, issue_flags);
194 }
195}
196
197static int io_rw_alloc_async(struct io_kiocb *req)
198{
199 struct io_ring_ctx *ctx = req->ctx;
200 struct io_async_rw *rw;
201
202 rw = io_uring_alloc_async_data(cache: &ctx->rw_cache, req);
203 if (!rw)
204 return -ENOMEM;
205 if (rw->vec.iovec)
206 req->flags |= REQ_F_NEED_CLEANUP;
207 rw->bytes_done = 0;
208 return 0;
209}
210
211static inline void io_meta_save_state(struct io_async_rw *io)
212{
213 io->meta_state.seed = io->meta.seed;
214 iov_iter_save_state(iter: &io->meta.iter, state: &io->meta_state.iter_meta);
215}
216
217static inline void io_meta_restore(struct io_async_rw *io, struct kiocb *kiocb)
218{
219 if (kiocb->ki_flags & IOCB_HAS_METADATA) {
220 io->meta.seed = io->meta_state.seed;
221 iov_iter_restore(i: &io->meta.iter, state: &io->meta_state.iter_meta);
222 }
223}
224
225static int io_prep_rw_pi(struct io_kiocb *req, struct io_rw *rw, int ddir,
226 u64 attr_ptr, u64 attr_type_mask)
227{
228 struct io_uring_attr_pi pi_attr;
229 struct io_async_rw *io;
230 int ret;
231
232 if (copy_from_user(to: &pi_attr, u64_to_user_ptr(attr_ptr),
233 n: sizeof(pi_attr)))
234 return -EFAULT;
235
236 if (pi_attr.rsvd)
237 return -EINVAL;
238
239 io = req->async_data;
240 io->meta.flags = pi_attr.flags;
241 io->meta.app_tag = pi_attr.app_tag;
242 io->meta.seed = pi_attr.seed;
243 ret = import_ubuf(type: ddir, u64_to_user_ptr(pi_attr.addr),
244 len: pi_attr.len, i: &io->meta.iter);
245 if (unlikely(ret < 0))
246 return ret;
247 req->flags |= REQ_F_HAS_METADATA;
248 io_meta_save_state(io);
249 return ret;
250}
251
252static int __io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
253 int ddir)
254{
255 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
256 struct io_async_rw *io;
257 unsigned ioprio;
258 u64 attr_type_mask;
259 int ret;
260
261 if (io_rw_alloc_async(req))
262 return -ENOMEM;
263 io = req->async_data;
264
265 rw->kiocb.ki_pos = READ_ONCE(sqe->off);
266 /* used for fixed read/write too - just read unconditionally */
267 req->buf_index = READ_ONCE(sqe->buf_index);
268 io->buf_group = req->buf_index;
269
270 ioprio = READ_ONCE(sqe->ioprio);
271 if (ioprio) {
272 ret = ioprio_check_cap(ioprio);
273 if (ret)
274 return ret;
275
276 rw->kiocb.ki_ioprio = ioprio;
277 } else {
278 rw->kiocb.ki_ioprio = get_current_ioprio();
279 }
280 rw->kiocb.dio_complete = NULL;
281 rw->kiocb.ki_flags = 0;
282 rw->kiocb.ki_write_stream = READ_ONCE(sqe->write_stream);
283
284 if (req->ctx->flags & IORING_SETUP_IOPOLL)
285 rw->kiocb.ki_complete = io_complete_rw_iopoll;
286 else
287 rw->kiocb.ki_complete = io_complete_rw;
288
289 rw->addr = READ_ONCE(sqe->addr);
290 rw->len = READ_ONCE(sqe->len);
291 rw->flags = (__force rwf_t) READ_ONCE(sqe->rw_flags);
292
293 attr_type_mask = READ_ONCE(sqe->attr_type_mask);
294 if (attr_type_mask) {
295 u64 attr_ptr;
296
297 /* only PI attribute is supported currently */
298 if (attr_type_mask != IORING_RW_ATTR_FLAG_PI)
299 return -EINVAL;
300
301 attr_ptr = READ_ONCE(sqe->attr_ptr);
302 return io_prep_rw_pi(req, rw, ddir, attr_ptr, attr_type_mask);
303 }
304 return 0;
305}
306
307static int io_rw_do_import(struct io_kiocb *req, int ddir)
308{
309 struct io_br_sel sel = { };
310
311 if (io_do_buffer_select(req))
312 return 0;
313
314 return io_import_rw_buffer(rw: ddir, req, io: req->async_data, sel: &sel, issue_flags: 0);
315}
316
317static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
318 int ddir)
319{
320 int ret;
321
322 ret = __io_prep_rw(req, sqe, ddir);
323 if (unlikely(ret))
324 return ret;
325
326 return io_rw_do_import(req, ddir);
327}
328
329int io_prep_read(struct io_kiocb *req, const struct io_uring_sqe *sqe)
330{
331 return io_prep_rw(req, sqe, ITER_DEST);
332}
333
334int io_prep_write(struct io_kiocb *req, const struct io_uring_sqe *sqe)
335{
336 return io_prep_rw(req, sqe, ITER_SOURCE);
337}
338
339static int io_prep_rwv(struct io_kiocb *req, const struct io_uring_sqe *sqe,
340 int ddir)
341{
342 int ret;
343
344 ret = io_prep_rw(req, sqe, ddir);
345 if (unlikely(ret))
346 return ret;
347 if (!(req->flags & REQ_F_BUFFER_SELECT))
348 return 0;
349
350 /*
351 * Have to do this validation here, as this is in io_read() rw->len
352 * might have chanaged due to buffer selection
353 */
354 return io_iov_buffer_select_prep(req);
355}
356
357int io_prep_readv(struct io_kiocb *req, const struct io_uring_sqe *sqe)
358{
359 return io_prep_rwv(req, sqe, ITER_DEST);
360}
361
362int io_prep_writev(struct io_kiocb *req, const struct io_uring_sqe *sqe)
363{
364 return io_prep_rwv(req, sqe, ITER_SOURCE);
365}
366
367static int io_init_rw_fixed(struct io_kiocb *req, unsigned int issue_flags,
368 int ddir)
369{
370 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
371 struct io_async_rw *io = req->async_data;
372 int ret;
373
374 if (io->bytes_done)
375 return 0;
376
377 ret = io_import_reg_buf(req, iter: &io->iter, buf_addr: rw->addr, len: rw->len, ddir,
378 issue_flags);
379 iov_iter_save_state(iter: &io->iter, state: &io->iter_state);
380 return ret;
381}
382
383int io_prep_read_fixed(struct io_kiocb *req, const struct io_uring_sqe *sqe)
384{
385 return __io_prep_rw(req, sqe, ITER_DEST);
386}
387
388int io_prep_write_fixed(struct io_kiocb *req, const struct io_uring_sqe *sqe)
389{
390 return __io_prep_rw(req, sqe, ITER_SOURCE);
391}
392
393static int io_rw_import_reg_vec(struct io_kiocb *req,
394 struct io_async_rw *io,
395 int ddir, unsigned int issue_flags)
396{
397 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
398 unsigned uvec_segs = rw->len;
399 int ret;
400
401 ret = io_import_reg_vec(ddir, iter: &io->iter, req, vec: &io->vec,
402 nr_iovs: uvec_segs, issue_flags);
403 if (unlikely(ret))
404 return ret;
405 iov_iter_save_state(iter: &io->iter, state: &io->iter_state);
406 req->flags &= ~REQ_F_IMPORT_BUFFER;
407 return 0;
408}
409
410static int io_rw_prep_reg_vec(struct io_kiocb *req)
411{
412 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
413 struct io_async_rw *io = req->async_data;
414 const struct iovec __user *uvec;
415
416 uvec = u64_to_user_ptr(rw->addr);
417 return io_prep_reg_iovec(req, iv: &io->vec, uvec, uvec_segs: rw->len);
418}
419
420int io_prep_readv_fixed(struct io_kiocb *req, const struct io_uring_sqe *sqe)
421{
422 int ret;
423
424 ret = __io_prep_rw(req, sqe, ITER_DEST);
425 if (unlikely(ret))
426 return ret;
427 return io_rw_prep_reg_vec(req);
428}
429
430int io_prep_writev_fixed(struct io_kiocb *req, const struct io_uring_sqe *sqe)
431{
432 int ret;
433
434 ret = __io_prep_rw(req, sqe, ITER_SOURCE);
435 if (unlikely(ret))
436 return ret;
437 return io_rw_prep_reg_vec(req);
438}
439
440/*
441 * Multishot read is prepared just like a normal read/write request, only
442 * difference is that we set the MULTISHOT flag.
443 */
444int io_read_mshot_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
445{
446 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
447 int ret;
448
449 /* must be used with provided buffers */
450 if (!(req->flags & REQ_F_BUFFER_SELECT))
451 return -EINVAL;
452
453 ret = __io_prep_rw(req, sqe, ITER_DEST);
454 if (unlikely(ret))
455 return ret;
456
457 if (rw->addr || rw->len)
458 return -EINVAL;
459
460 req->flags |= REQ_F_APOLL_MULTISHOT;
461 return 0;
462}
463
464void io_readv_writev_cleanup(struct io_kiocb *req)
465{
466 lockdep_assert_held(&req->ctx->uring_lock);
467 io_rw_recycle(req, issue_flags: 0);
468}
469
470static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req)
471{
472 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
473
474 if (rw->kiocb.ki_pos != -1)
475 return &rw->kiocb.ki_pos;
476
477 if (!(req->file->f_mode & FMODE_STREAM)) {
478 req->flags |= REQ_F_CUR_POS;
479 rw->kiocb.ki_pos = req->file->f_pos;
480 return &rw->kiocb.ki_pos;
481 }
482
483 rw->kiocb.ki_pos = 0;
484 return NULL;
485}
486
487static bool io_rw_should_reissue(struct io_kiocb *req)
488{
489#ifdef CONFIG_BLOCK
490 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
491 umode_t mode = file_inode(f: req->file)->i_mode;
492 struct io_async_rw *io = req->async_data;
493 struct io_ring_ctx *ctx = req->ctx;
494
495 if (!S_ISBLK(mode) && !S_ISREG(mode))
496 return false;
497 if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() &&
498 !(ctx->flags & IORING_SETUP_IOPOLL)))
499 return false;
500 /*
501 * If ref is dying, we might be running poll reap from the exit work.
502 * Don't attempt to reissue from that path, just let it fail with
503 * -EAGAIN.
504 */
505 if (percpu_ref_is_dying(ref: &ctx->refs))
506 return false;
507
508 io_meta_restore(io, kiocb: &rw->kiocb);
509 iov_iter_restore(i: &io->iter, state: &io->iter_state);
510 return true;
511#else
512 return false;
513#endif
514}
515
516static void io_req_end_write(struct io_kiocb *req)
517{
518 if (req->flags & REQ_F_ISREG) {
519 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
520
521 kiocb_end_write(iocb: &rw->kiocb);
522 }
523}
524
525/*
526 * Trigger the notifications after having done some IO, and finish the write
527 * accounting, if any.
528 */
529static void io_req_io_end(struct io_kiocb *req)
530{
531 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
532
533 if (rw->kiocb.ki_flags & IOCB_WRITE) {
534 io_req_end_write(req);
535 fsnotify_modify(file: req->file);
536 } else {
537 fsnotify_access(file: req->file);
538 }
539}
540
541static void __io_complete_rw_common(struct io_kiocb *req, long res)
542{
543 if (res == req->cqe.res)
544 return;
545 if (res == -EAGAIN && io_rw_should_reissue(req)) {
546 req->flags |= REQ_F_REISSUE | REQ_F_BL_NO_RECYCLE;
547 } else {
548 req_set_fail(req);
549 req->cqe.res = res;
550 }
551}
552
553static inline int io_fixup_rw_res(struct io_kiocb *req, long res)
554{
555 struct io_async_rw *io = req->async_data;
556
557 /* add previously done IO, if any */
558 if (req_has_async_data(req) && io->bytes_done > 0) {
559 if (res < 0)
560 res = io->bytes_done;
561 else
562 res += io->bytes_done;
563 }
564 return res;
565}
566
567void io_req_rw_complete(struct io_kiocb *req, io_tw_token_t tw)
568{
569 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
570 struct kiocb *kiocb = &rw->kiocb;
571
572 if ((kiocb->ki_flags & IOCB_DIO_CALLER_COMP) && kiocb->dio_complete) {
573 long res = kiocb->dio_complete(rw->kiocb.private);
574
575 io_req_set_res(req, res: io_fixup_rw_res(req, res), cflags: 0);
576 }
577
578 io_req_io_end(req);
579
580 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING))
581 req->cqe.flags |= io_put_kbuf(req, len: req->cqe.res, NULL);
582
583 io_req_rw_cleanup(req, issue_flags: 0);
584 io_req_task_complete(req, tw);
585}
586
587static void io_complete_rw(struct kiocb *kiocb, long res)
588{
589 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
590 struct io_kiocb *req = cmd_to_io_kiocb(ptr: rw);
591
592 if (!kiocb->dio_complete || !(kiocb->ki_flags & IOCB_DIO_CALLER_COMP)) {
593 __io_complete_rw_common(req, res);
594 io_req_set_res(req, res: io_fixup_rw_res(req, res), cflags: 0);
595 }
596 req->io_task_work.func = io_req_rw_complete;
597 __io_req_task_work_add(req, flags: IOU_F_TWQ_LAZY_WAKE);
598}
599
600static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)
601{
602 struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb);
603 struct io_kiocb *req = cmd_to_io_kiocb(ptr: rw);
604
605 if (kiocb->ki_flags & IOCB_WRITE)
606 io_req_end_write(req);
607 if (unlikely(res != req->cqe.res)) {
608 if (res == -EAGAIN && io_rw_should_reissue(req))
609 req->flags |= REQ_F_REISSUE | REQ_F_BL_NO_RECYCLE;
610 else
611 req->cqe.res = res;
612 }
613
614 /* order with io_iopoll_complete() checking ->iopoll_completed */
615 smp_store_release(&req->iopoll_completed, 1);
616}
617
618static inline void io_rw_done(struct io_kiocb *req, ssize_t ret)
619{
620 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
621
622 /* IO was queued async, completion will happen later */
623 if (ret == -EIOCBQUEUED)
624 return;
625
626 /* transform internal restart error codes */
627 if (unlikely(ret < 0)) {
628 switch (ret) {
629 case -ERESTARTSYS:
630 case -ERESTARTNOINTR:
631 case -ERESTARTNOHAND:
632 case -ERESTART_RESTARTBLOCK:
633 /*
634 * We can't just restart the syscall, since previously
635 * submitted sqes may already be in progress. Just fail
636 * this IO with EINTR.
637 */
638 ret = -EINTR;
639 break;
640 }
641 }
642
643 if (req->ctx->flags & IORING_SETUP_IOPOLL)
644 io_complete_rw_iopoll(kiocb: &rw->kiocb, res: ret);
645 else
646 io_complete_rw(kiocb: &rw->kiocb, res: ret);
647}
648
649static int kiocb_done(struct io_kiocb *req, ssize_t ret,
650 struct io_br_sel *sel, unsigned int issue_flags)
651{
652 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
653 unsigned final_ret = io_fixup_rw_res(req, res: ret);
654
655 if (ret >= 0 && req->flags & REQ_F_CUR_POS)
656 req->file->f_pos = rw->kiocb.ki_pos;
657 if (ret >= 0 && !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
658 __io_complete_rw_common(req, res: ret);
659 /*
660 * Safe to call io_end from here as we're inline
661 * from the submission path.
662 */
663 io_req_io_end(req);
664 io_req_set_res(req, res: final_ret, cflags: io_put_kbuf(req, len: ret, bl: sel->buf_list));
665 io_req_rw_cleanup(req, issue_flags);
666 return IOU_COMPLETE;
667 } else {
668 io_rw_done(req, ret);
669 }
670
671 return IOU_ISSUE_SKIP_COMPLETE;
672}
673
674static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb)
675{
676 return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos;
677}
678
679/*
680 * For files that don't have ->read_iter() and ->write_iter(), handle them
681 * by looping over ->read() or ->write() manually.
682 */
683static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter)
684{
685 struct io_kiocb *req = cmd_to_io_kiocb(ptr: rw);
686 struct kiocb *kiocb = &rw->kiocb;
687 struct file *file = kiocb->ki_filp;
688 ssize_t ret = 0;
689 loff_t *ppos;
690
691 /*
692 * Don't support polled IO through this interface, and we can't
693 * support non-blocking either. For the latter, this just causes
694 * the kiocb to be handled from an async context.
695 */
696 if (kiocb->ki_flags & IOCB_HIPRI)
697 return -EOPNOTSUPP;
698 if ((kiocb->ki_flags & IOCB_NOWAIT) &&
699 !(kiocb->ki_filp->f_flags & O_NONBLOCK))
700 return -EAGAIN;
701 if ((req->flags & REQ_F_BUF_NODE) && req->buf_node->buf->is_kbuf)
702 return -EFAULT;
703
704 ppos = io_kiocb_ppos(kiocb);
705
706 while (iov_iter_count(i: iter)) {
707 void __user *addr;
708 size_t len;
709 ssize_t nr;
710
711 if (iter_is_ubuf(i: iter)) {
712 addr = iter->ubuf + iter->iov_offset;
713 len = iov_iter_count(i: iter);
714 } else if (!iov_iter_is_bvec(i: iter)) {
715 addr = iter_iov_addr(iter);
716 len = iter_iov_len(i: iter);
717 } else {
718 addr = u64_to_user_ptr(rw->addr);
719 len = rw->len;
720 }
721
722 if (ddir == READ)
723 nr = file->f_op->read(file, addr, len, ppos);
724 else
725 nr = file->f_op->write(file, addr, len, ppos);
726
727 if (nr < 0) {
728 if (!ret)
729 ret = nr;
730 break;
731 }
732 ret += nr;
733 if (!iov_iter_is_bvec(i: iter)) {
734 iov_iter_advance(i: iter, bytes: nr);
735 } else {
736 rw->addr += nr;
737 rw->len -= nr;
738 if (!rw->len)
739 break;
740 }
741 if (nr != len)
742 break;
743 }
744
745 return ret;
746}
747
748/*
749 * This is our waitqueue callback handler, registered through __folio_lock_async()
750 * when we initially tried to do the IO with the iocb armed our waitqueue.
751 * This gets called when the page is unlocked, and we generally expect that to
752 * happen when the page IO is completed and the page is now uptodate. This will
753 * queue a task_work based retry of the operation, attempting to copy the data
754 * again. If the latter fails because the page was NOT uptodate, then we will
755 * do a thread based blocking retry of the operation. That's the unexpected
756 * slow path.
757 */
758static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode,
759 int sync, void *arg)
760{
761 struct wait_page_queue *wpq;
762 struct io_kiocb *req = wait->private;
763 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
764 struct wait_page_key *key = arg;
765
766 wpq = container_of(wait, struct wait_page_queue, wait);
767
768 if (!wake_page_match(wait_page: wpq, key))
769 return 0;
770
771 rw->kiocb.ki_flags &= ~IOCB_WAITQ;
772 list_del_init(entry: &wait->entry);
773 io_req_task_queue(req);
774 return 1;
775}
776
777/*
778 * This controls whether a given IO request should be armed for async page
779 * based retry. If we return false here, the request is handed to the async
780 * worker threads for retry. If we're doing buffered reads on a regular file,
781 * we prepare a private wait_page_queue entry and retry the operation. This
782 * will either succeed because the page is now uptodate and unlocked, or it
783 * will register a callback when the page is unlocked at IO completion. Through
784 * that callback, io_uring uses task_work to setup a retry of the operation.
785 * That retry will attempt the buffered read again. The retry will generally
786 * succeed, or in rare cases where it fails, we then fall back to using the
787 * async worker threads for a blocking retry.
788 */
789static bool io_rw_should_retry(struct io_kiocb *req)
790{
791 struct io_async_rw *io = req->async_data;
792 struct wait_page_queue *wait = &io->wpq;
793 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
794 struct kiocb *kiocb = &rw->kiocb;
795
796 /*
797 * Never retry for NOWAIT or a request with metadata, we just complete
798 * with -EAGAIN.
799 */
800 if (req->flags & (REQ_F_NOWAIT | REQ_F_HAS_METADATA))
801 return false;
802
803 /* Only for buffered IO */
804 if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI))
805 return false;
806
807 /*
808 * just use poll if we can, and don't attempt if the fs doesn't
809 * support callback based unlocks
810 */
811 if (io_file_can_poll(req) ||
812 !(req->file->f_op->fop_flags & FOP_BUFFER_RASYNC))
813 return false;
814
815 wait->wait.func = io_async_buf_func;
816 wait->wait.private = req;
817 wait->wait.flags = 0;
818 INIT_LIST_HEAD(list: &wait->wait.entry);
819 kiocb->ki_flags |= IOCB_WAITQ;
820 kiocb->ki_flags &= ~IOCB_NOWAIT;
821 kiocb->ki_waitq = wait;
822 return true;
823}
824
825static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter)
826{
827 struct file *file = rw->kiocb.ki_filp;
828
829 if (likely(file->f_op->read_iter))
830 return file->f_op->read_iter(&rw->kiocb, iter);
831 else if (file->f_op->read)
832 return loop_rw_iter(READ, rw, iter);
833 else
834 return -EINVAL;
835}
836
837static bool need_complete_io(struct io_kiocb *req)
838{
839 return req->flags & REQ_F_ISREG ||
840 S_ISBLK(file_inode(req->file)->i_mode);
841}
842
843static int io_rw_init_file(struct io_kiocb *req, fmode_t mode, int rw_type)
844{
845 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
846 struct kiocb *kiocb = &rw->kiocb;
847 struct io_ring_ctx *ctx = req->ctx;
848 struct file *file = req->file;
849 int ret;
850
851 if (unlikely(!(file->f_mode & mode)))
852 return -EBADF;
853
854 if (!(req->flags & REQ_F_FIXED_FILE))
855 req->flags |= io_file_get_flags(file);
856
857 kiocb->ki_flags = file->f_iocb_flags;
858 ret = kiocb_set_rw_flags(ki: kiocb, flags: rw->flags, rw_type);
859 if (unlikely(ret))
860 return ret;
861 kiocb->ki_flags |= IOCB_ALLOC_CACHE;
862
863 /*
864 * If the file is marked O_NONBLOCK, still allow retry for it if it
865 * supports async. Otherwise it's impossible to use O_NONBLOCK files
866 * reliably. If not, or it IOCB_NOWAIT is set, don't retry.
867 */
868 if (kiocb->ki_flags & IOCB_NOWAIT ||
869 ((file->f_flags & O_NONBLOCK && !(req->flags & REQ_F_SUPPORT_NOWAIT))))
870 req->flags |= REQ_F_NOWAIT;
871
872 if (ctx->flags & IORING_SETUP_IOPOLL) {
873 if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll)
874 return -EOPNOTSUPP;
875 kiocb->private = NULL;
876 kiocb->ki_flags |= IOCB_HIPRI;
877 req->iopoll_completed = 0;
878 if (ctx->flags & IORING_SETUP_HYBRID_IOPOLL) {
879 /* make sure every req only blocks once*/
880 req->flags &= ~REQ_F_IOPOLL_STATE;
881 req->iopoll_start = ktime_get_ns();
882 }
883 } else {
884 if (kiocb->ki_flags & IOCB_HIPRI)
885 return -EINVAL;
886 }
887
888 if (req->flags & REQ_F_HAS_METADATA) {
889 struct io_async_rw *io = req->async_data;
890
891 if (!(file->f_mode & FMODE_HAS_METADATA))
892 return -EINVAL;
893
894 /*
895 * We have a union of meta fields with wpq used for buffered-io
896 * in io_async_rw, so fail it here.
897 */
898 if (!(req->file->f_flags & O_DIRECT))
899 return -EOPNOTSUPP;
900 kiocb->ki_flags |= IOCB_HAS_METADATA;
901 kiocb->private = &io->meta;
902 }
903
904 return 0;
905}
906
907static int __io_read(struct io_kiocb *req, struct io_br_sel *sel,
908 unsigned int issue_flags)
909{
910 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
911 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
912 struct io_async_rw *io = req->async_data;
913 struct kiocb *kiocb = &rw->kiocb;
914 ssize_t ret;
915 loff_t *ppos;
916
917 if (req->flags & REQ_F_IMPORT_BUFFER) {
918 ret = io_rw_import_reg_vec(req, io, ITER_DEST, issue_flags);
919 if (unlikely(ret))
920 return ret;
921 } else if (io_do_buffer_select(req)) {
922 ret = io_import_rw_buffer(ITER_DEST, req, io, sel, issue_flags);
923 if (unlikely(ret < 0))
924 return ret;
925 }
926 ret = io_rw_init_file(req, FMODE_READ, READ);
927 if (unlikely(ret))
928 return ret;
929 req->cqe.res = iov_iter_count(i: &io->iter);
930
931 if (force_nonblock) {
932 /* If the file doesn't support async, just async punt */
933 if (unlikely(!io_file_supports_nowait(req, EPOLLIN)))
934 return -EAGAIN;
935 kiocb->ki_flags |= IOCB_NOWAIT;
936 } else {
937 /* Ensure we clear previously set non-block flag */
938 kiocb->ki_flags &= ~IOCB_NOWAIT;
939 }
940
941 ppos = io_kiocb_update_pos(req);
942
943 ret = rw_verify_area(READ, req->file, ppos, req->cqe.res);
944 if (unlikely(ret))
945 return ret;
946
947 ret = io_iter_do_read(rw, iter: &io->iter);
948
949 /*
950 * Some file systems like to return -EOPNOTSUPP for an IOCB_NOWAIT
951 * issue, even though they should be returning -EAGAIN. To be safe,
952 * retry from blocking context for either.
953 */
954 if (ret == -EOPNOTSUPP && force_nonblock)
955 ret = -EAGAIN;
956
957 if (ret == -EAGAIN) {
958 /* If we can poll, just do that. */
959 if (io_file_can_poll(req))
960 return -EAGAIN;
961 /* IOPOLL retry should happen for io-wq threads */
962 if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL))
963 goto done;
964 /* no retry on NONBLOCK nor RWF_NOWAIT */
965 if (req->flags & REQ_F_NOWAIT)
966 goto done;
967 ret = 0;
968 } else if (ret == -EIOCBQUEUED) {
969 return IOU_ISSUE_SKIP_COMPLETE;
970 } else if (ret == req->cqe.res || ret <= 0 || !force_nonblock ||
971 (req->flags & REQ_F_NOWAIT) || !need_complete_io(req) ||
972 (issue_flags & IO_URING_F_MULTISHOT)) {
973 /* read all, failed, already did sync or don't want to retry */
974 goto done;
975 }
976
977 /*
978 * Don't depend on the iter state matching what was consumed, or being
979 * untouched in case of error. Restore it and we'll advance it
980 * manually if we need to.
981 */
982 iov_iter_restore(i: &io->iter, state: &io->iter_state);
983 io_meta_restore(io, kiocb);
984
985 do {
986 /*
987 * We end up here because of a partial read, either from
988 * above or inside this loop. Advance the iter by the bytes
989 * that were consumed.
990 */
991 iov_iter_advance(i: &io->iter, bytes: ret);
992 if (!iov_iter_count(i: &io->iter))
993 break;
994 io->bytes_done += ret;
995 iov_iter_save_state(iter: &io->iter, state: &io->iter_state);
996
997 /* if we can retry, do so with the callbacks armed */
998 if (!io_rw_should_retry(req)) {
999 kiocb->ki_flags &= ~IOCB_WAITQ;
1000 return -EAGAIN;
1001 }
1002
1003 req->cqe.res = iov_iter_count(i: &io->iter);
1004 /*
1005 * Now retry read with the IOCB_WAITQ parts set in the iocb. If
1006 * we get -EIOCBQUEUED, then we'll get a notification when the
1007 * desired page gets unlocked. We can also get a partial read
1008 * here, and if we do, then just retry at the new offset.
1009 */
1010 ret = io_iter_do_read(rw, iter: &io->iter);
1011 if (ret == -EIOCBQUEUED)
1012 return IOU_ISSUE_SKIP_COMPLETE;
1013 /* we got some bytes, but not all. retry. */
1014 kiocb->ki_flags &= ~IOCB_WAITQ;
1015 iov_iter_restore(i: &io->iter, state: &io->iter_state);
1016 } while (ret > 0);
1017done:
1018 /* it's faster to check here then delegate to kfree */
1019 return ret;
1020}
1021
1022int io_read(struct io_kiocb *req, unsigned int issue_flags)
1023{
1024 struct io_br_sel sel = { };
1025 int ret;
1026
1027 ret = __io_read(req, sel: &sel, issue_flags);
1028 if (ret >= 0)
1029 return kiocb_done(req, ret, sel: &sel, issue_flags);
1030
1031 if (req->flags & REQ_F_BUFFERS_COMMIT)
1032 io_kbuf_recycle(req, bl: sel.buf_list, issue_flags);
1033 return ret;
1034}
1035
1036int io_read_mshot(struct io_kiocb *req, unsigned int issue_flags)
1037{
1038 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1039 struct io_br_sel sel = { };
1040 unsigned int cflags = 0;
1041 int ret;
1042
1043 /*
1044 * Multishot MUST be used on a pollable file
1045 */
1046 if (!io_file_can_poll(req))
1047 return -EBADFD;
1048
1049 /* make it sync, multishot doesn't support async execution */
1050 rw->kiocb.ki_complete = NULL;
1051 ret = __io_read(req, sel: &sel, issue_flags);
1052
1053 /*
1054 * If we get -EAGAIN, recycle our buffer and just let normal poll
1055 * handling arm it.
1056 */
1057 if (ret == -EAGAIN) {
1058 /*
1059 * Reset rw->len to 0 again to avoid clamping future mshot
1060 * reads, in case the buffer size varies.
1061 */
1062 if (io_kbuf_recycle(req, bl: sel.buf_list, issue_flags))
1063 rw->len = 0;
1064 return IOU_RETRY;
1065 } else if (ret <= 0) {
1066 io_kbuf_recycle(req, bl: sel.buf_list, issue_flags);
1067 if (ret < 0)
1068 req_set_fail(req);
1069 } else if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
1070 cflags = io_put_kbuf(req, len: ret, bl: sel.buf_list);
1071 } else {
1072 /*
1073 * Any successful return value will keep the multishot read
1074 * armed, if it's still set. Put our buffer and post a CQE. If
1075 * we fail to post a CQE, or multishot is no longer set, then
1076 * jump to the termination path. This request is then done.
1077 */
1078 cflags = io_put_kbuf(req, len: ret, bl: sel.buf_list);
1079 rw->len = 0; /* similarly to above, reset len to 0 */
1080
1081 if (io_req_post_cqe(req, res: ret, cflags: cflags | IORING_CQE_F_MORE)) {
1082 if (issue_flags & IO_URING_F_MULTISHOT)
1083 /*
1084 * Force retry, as we might have more data to
1085 * be read and otherwise it won't get retried
1086 * until (if ever) another poll is triggered.
1087 */
1088 io_poll_multishot_retry(req);
1089
1090 return IOU_RETRY;
1091 }
1092 }
1093
1094 /*
1095 * Either an error, or we've hit overflow posting the CQE. For any
1096 * multishot request, hitting overflow will terminate it.
1097 */
1098 io_req_set_res(req, res: ret, cflags);
1099 io_req_rw_cleanup(req, issue_flags);
1100 return IOU_COMPLETE;
1101}
1102
1103static bool io_kiocb_start_write(struct io_kiocb *req, struct kiocb *kiocb)
1104{
1105 struct inode *inode;
1106 bool ret;
1107
1108 if (!(req->flags & REQ_F_ISREG))
1109 return true;
1110 if (!(kiocb->ki_flags & IOCB_NOWAIT)) {
1111 kiocb_start_write(iocb: kiocb);
1112 return true;
1113 }
1114
1115 inode = file_inode(f: kiocb->ki_filp);
1116 ret = sb_start_write_trylock(sb: inode->i_sb);
1117 if (ret)
1118 __sb_writers_release(inode->i_sb, SB_FREEZE_WRITE);
1119 return ret;
1120}
1121
1122int io_write(struct io_kiocb *req, unsigned int issue_flags)
1123{
1124 bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
1125 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1126 struct io_async_rw *io = req->async_data;
1127 struct kiocb *kiocb = &rw->kiocb;
1128 ssize_t ret, ret2;
1129 loff_t *ppos;
1130
1131 if (req->flags & REQ_F_IMPORT_BUFFER) {
1132 ret = io_rw_import_reg_vec(req, io, ITER_SOURCE, issue_flags);
1133 if (unlikely(ret))
1134 return ret;
1135 }
1136
1137 ret = io_rw_init_file(req, FMODE_WRITE, WRITE);
1138 if (unlikely(ret))
1139 return ret;
1140 req->cqe.res = iov_iter_count(i: &io->iter);
1141
1142 if (force_nonblock) {
1143 /* If the file doesn't support async, just async punt */
1144 if (unlikely(!io_file_supports_nowait(req, EPOLLOUT)))
1145 goto ret_eagain;
1146
1147 /* Check if we can support NOWAIT. */
1148 if (!(kiocb->ki_flags & IOCB_DIRECT) &&
1149 !(req->file->f_op->fop_flags & FOP_BUFFER_WASYNC) &&
1150 (req->flags & REQ_F_ISREG))
1151 goto ret_eagain;
1152
1153 kiocb->ki_flags |= IOCB_NOWAIT;
1154 } else {
1155 /* Ensure we clear previously set non-block flag */
1156 kiocb->ki_flags &= ~IOCB_NOWAIT;
1157 }
1158
1159 ppos = io_kiocb_update_pos(req);
1160
1161 ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res);
1162 if (unlikely(ret))
1163 return ret;
1164
1165 if (unlikely(!io_kiocb_start_write(req, kiocb)))
1166 return -EAGAIN;
1167 kiocb->ki_flags |= IOCB_WRITE;
1168
1169 if (likely(req->file->f_op->write_iter))
1170 ret2 = req->file->f_op->write_iter(kiocb, &io->iter);
1171 else if (req->file->f_op->write)
1172 ret2 = loop_rw_iter(WRITE, rw, iter: &io->iter);
1173 else
1174 ret2 = -EINVAL;
1175
1176 /*
1177 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
1178 * retry them without IOCB_NOWAIT.
1179 */
1180 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
1181 ret2 = -EAGAIN;
1182 /* no retry on NONBLOCK nor RWF_NOWAIT */
1183 if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT))
1184 goto done;
1185 if (!force_nonblock || ret2 != -EAGAIN) {
1186 /* IOPOLL retry should happen for io-wq threads */
1187 if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL))
1188 goto ret_eagain;
1189
1190 if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) {
1191 trace_io_uring_short_write(ctx: req->ctx, fpos: kiocb->ki_pos - ret2,
1192 wanted: req->cqe.res, got: ret2);
1193
1194 /* This is a partial write. The file pos has already been
1195 * updated, setup the async struct to complete the request
1196 * in the worker. Also update bytes_done to account for
1197 * the bytes already written.
1198 */
1199 iov_iter_save_state(iter: &io->iter, state: &io->iter_state);
1200 io->bytes_done += ret2;
1201
1202 if (kiocb->ki_flags & IOCB_WRITE)
1203 io_req_end_write(req);
1204 return -EAGAIN;
1205 }
1206done:
1207 return kiocb_done(req, ret: ret2, NULL, issue_flags);
1208 } else {
1209ret_eagain:
1210 iov_iter_restore(i: &io->iter, state: &io->iter_state);
1211 io_meta_restore(io, kiocb);
1212 if (kiocb->ki_flags & IOCB_WRITE)
1213 io_req_end_write(req);
1214 return -EAGAIN;
1215 }
1216}
1217
1218int io_read_fixed(struct io_kiocb *req, unsigned int issue_flags)
1219{
1220 int ret;
1221
1222 ret = io_init_rw_fixed(req, issue_flags, ITER_DEST);
1223 if (unlikely(ret))
1224 return ret;
1225
1226 return io_read(req, issue_flags);
1227}
1228
1229int io_write_fixed(struct io_kiocb *req, unsigned int issue_flags)
1230{
1231 int ret;
1232
1233 ret = io_init_rw_fixed(req, issue_flags, ITER_SOURCE);
1234 if (unlikely(ret))
1235 return ret;
1236
1237 return io_write(req, issue_flags);
1238}
1239
1240void io_rw_fail(struct io_kiocb *req)
1241{
1242 int res;
1243
1244 res = io_fixup_rw_res(req, res: req->cqe.res);
1245 io_req_set_res(req, res, cflags: req->cqe.flags);
1246}
1247
1248static int io_uring_classic_poll(struct io_kiocb *req, struct io_comp_batch *iob,
1249 unsigned int poll_flags)
1250{
1251 struct file *file = req->file;
1252
1253 if (req->opcode == IORING_OP_URING_CMD) {
1254 struct io_uring_cmd *ioucmd;
1255
1256 ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);
1257 return file->f_op->uring_cmd_iopoll(ioucmd, iob, poll_flags);
1258 } else {
1259 struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
1260
1261 return file->f_op->iopoll(&rw->kiocb, iob, poll_flags);
1262 }
1263}
1264
1265static u64 io_hybrid_iopoll_delay(struct io_ring_ctx *ctx, struct io_kiocb *req)
1266{
1267 struct hrtimer_sleeper timer;
1268 enum hrtimer_mode mode;
1269 ktime_t kt;
1270 u64 sleep_time;
1271
1272 if (req->flags & REQ_F_IOPOLL_STATE)
1273 return 0;
1274
1275 if (ctx->hybrid_poll_time == LLONG_MAX)
1276 return 0;
1277
1278 /* Using half the running time to do schedule */
1279 sleep_time = ctx->hybrid_poll_time / 2;
1280
1281 kt = ktime_set(secs: 0, nsecs: sleep_time);
1282 req->flags |= REQ_F_IOPOLL_STATE;
1283
1284 mode = HRTIMER_MODE_REL;
1285 hrtimer_setup_sleeper_on_stack(sl: &timer, CLOCK_MONOTONIC, mode);
1286 hrtimer_set_expires(timer: &timer.timer, time: kt);
1287 set_current_state(TASK_INTERRUPTIBLE);
1288 hrtimer_sleeper_start_expires(sl: &timer, mode);
1289
1290 if (timer.task)
1291 io_schedule();
1292
1293 hrtimer_cancel(timer: &timer.timer);
1294 __set_current_state(TASK_RUNNING);
1295 destroy_hrtimer_on_stack(timer: &timer.timer);
1296 return sleep_time;
1297}
1298
1299static int io_uring_hybrid_poll(struct io_kiocb *req,
1300 struct io_comp_batch *iob, unsigned int poll_flags)
1301{
1302 struct io_ring_ctx *ctx = req->ctx;
1303 u64 runtime, sleep_time;
1304 int ret;
1305
1306 sleep_time = io_hybrid_iopoll_delay(ctx, req);
1307 ret = io_uring_classic_poll(req, iob, poll_flags);
1308 runtime = ktime_get_ns() - req->iopoll_start - sleep_time;
1309
1310 /*
1311 * Use minimum sleep time if we're polling devices with different
1312 * latencies. We could get more completions from the faster ones.
1313 */
1314 if (ctx->hybrid_poll_time > runtime)
1315 ctx->hybrid_poll_time = runtime;
1316
1317 return ret;
1318}
1319
1320int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
1321{
1322 struct io_wq_work_node *pos, *start, *prev;
1323 unsigned int poll_flags = 0;
1324 DEFINE_IO_COMP_BATCH(iob);
1325 int nr_events = 0;
1326
1327 /*
1328 * Only spin for completions if we don't have multiple devices hanging
1329 * off our complete list.
1330 */
1331 if (ctx->poll_multi_queue || force_nonspin)
1332 poll_flags |= BLK_POLL_ONESHOT;
1333
1334 wq_list_for_each(pos, start, &ctx->iopoll_list) {
1335 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1336 int ret;
1337
1338 /*
1339 * Move completed and retryable entries to our local lists.
1340 * If we find a request that requires polling, break out
1341 * and complete those lists first, if we have entries there.
1342 */
1343 if (READ_ONCE(req->iopoll_completed))
1344 break;
1345
1346 if (ctx->flags & IORING_SETUP_HYBRID_IOPOLL)
1347 ret = io_uring_hybrid_poll(req, iob: &iob, poll_flags);
1348 else
1349 ret = io_uring_classic_poll(req, iob: &iob, poll_flags);
1350
1351 if (unlikely(ret < 0))
1352 return ret;
1353 else if (ret)
1354 poll_flags |= BLK_POLL_ONESHOT;
1355
1356 /* iopoll may have completed current req */
1357 if (!rq_list_empty(rl: &iob.req_list) ||
1358 READ_ONCE(req->iopoll_completed))
1359 break;
1360 }
1361
1362 if (!rq_list_empty(rl: &iob.req_list))
1363 iob.complete(&iob);
1364 else if (!pos)
1365 return 0;
1366
1367 prev = start;
1368 wq_list_for_each_resume(pos, prev) {
1369 struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list);
1370
1371 /* order with io_complete_rw_iopoll(), e.g. ->result updates */
1372 if (!smp_load_acquire(&req->iopoll_completed))
1373 break;
1374 nr_events++;
1375 req->cqe.flags = io_put_kbuf(req, len: req->cqe.res, NULL);
1376 if (req->opcode != IORING_OP_URING_CMD)
1377 io_req_rw_cleanup(req, issue_flags: 0);
1378 }
1379 if (unlikely(!nr_events))
1380 return 0;
1381
1382 pos = start ? start->next : ctx->iopoll_list.first;
1383 wq_list_cut(list: &ctx->iopoll_list, last: prev, prev: start);
1384
1385 if (WARN_ON_ONCE(!wq_list_empty(&ctx->submit_state.compl_reqs)))
1386 return 0;
1387 ctx->submit_state.compl_reqs.first = pos;
1388 __io_submit_flush_completions(ctx);
1389 return nr_events;
1390}
1391
1392void io_rw_cache_free(const void *entry)
1393{
1394 struct io_async_rw *rw = (struct io_async_rw *) entry;
1395
1396 io_vec_free(iv: &rw->vec);
1397 kfree(objp: rw);
1398}
1399