]>
Commit | Line | Data |
---|---|---|
f3b44f92 JA |
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> | |
b66509b8 | 13 | #include <linux/io_uring/cmd.h> |
f3b44f92 JA |
14 | |
15 | #include <uapi/linux/io_uring.h> | |
16 | ||
f3b44f92 JA |
17 | #include "io_uring.h" |
18 | #include "opdef.h" | |
19 | #include "kbuf.h" | |
20 | #include "rsrc.h" | |
c79f52f0 | 21 | #include "poll.h" |
f3b44f92 JA |
22 | #include "rw.h" |
23 | ||
24 | struct io_rw { | |
25 | /* NOTE: kiocb has the file as the first member, so don't do it here */ | |
26 | struct kiocb kiocb; | |
27 | u64 addr; | |
28 | u32 len; | |
29 | rwf_t flags; | |
30 | }; | |
31 | ||
32 | static inline bool io_file_supports_nowait(struct io_kiocb *req) | |
33 | { | |
34 | return req->flags & REQ_F_SUPPORT_NOWAIT; | |
35 | } | |
36 | ||
4ab9d465 DY |
37 | #ifdef CONFIG_COMPAT |
38 | static int io_iov_compat_buffer_select_prep(struct io_rw *rw) | |
39 | { | |
40 | struct compat_iovec __user *uiov; | |
41 | compat_ssize_t clen; | |
42 | ||
43 | uiov = u64_to_user_ptr(rw->addr); | |
44 | if (!access_ok(uiov, sizeof(*uiov))) | |
45 | return -EFAULT; | |
46 | if (__get_user(clen, &uiov->iov_len)) | |
47 | return -EFAULT; | |
48 | if (clen < 0) | |
49 | return -EINVAL; | |
50 | ||
51 | rw->len = clen; | |
52 | return 0; | |
53 | } | |
54 | #endif | |
55 | ||
56 | static int io_iov_buffer_select_prep(struct io_kiocb *req) | |
57 | { | |
58 | struct iovec __user *uiov; | |
59 | struct iovec iov; | |
60 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); | |
61 | ||
62 | if (rw->len != 1) | |
63 | return -EINVAL; | |
64 | ||
65 | #ifdef CONFIG_COMPAT | |
66 | if (req->ctx->compat) | |
67 | return io_iov_compat_buffer_select_prep(rw); | |
68 | #endif | |
69 | ||
70 | uiov = u64_to_user_ptr(rw->addr); | |
71 | if (copy_from_user(&iov, uiov, sizeof(*uiov))) | |
72 | return -EFAULT; | |
73 | rw->len = iov.iov_len; | |
74 | return 0; | |
75 | } | |
76 | ||
f3b44f92 JA |
77 | int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
78 | { | |
f2ccb5ae | 79 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
f3b44f92 JA |
80 | unsigned ioprio; |
81 | int ret; | |
82 | ||
83 | rw->kiocb.ki_pos = READ_ONCE(sqe->off); | |
84 | /* used for fixed read/write too - just read unconditionally */ | |
85 | req->buf_index = READ_ONCE(sqe->buf_index); | |
86 | ||
f3b44f92 JA |
87 | ioprio = READ_ONCE(sqe->ioprio); |
88 | if (ioprio) { | |
89 | ret = ioprio_check_cap(ioprio); | |
90 | if (ret) | |
91 | return ret; | |
92 | ||
93 | rw->kiocb.ki_ioprio = ioprio; | |
94 | } else { | |
95 | rw->kiocb.ki_ioprio = get_current_ioprio(); | |
96 | } | |
099ada2c | 97 | rw->kiocb.dio_complete = NULL; |
f3b44f92 JA |
98 | |
99 | rw->addr = READ_ONCE(sqe->addr); | |
100 | rw->len = READ_ONCE(sqe->len); | |
101 | rw->flags = READ_ONCE(sqe->rw_flags); | |
0e984ec8 JA |
102 | return 0; |
103 | } | |
104 | ||
105 | int io_prep_rwv(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
106 | { | |
107 | int ret; | |
4ab9d465 | 108 | |
0e984ec8 JA |
109 | ret = io_prep_rw(req, sqe); |
110 | if (unlikely(ret)) | |
111 | return ret; | |
112 | ||
113 | /* | |
114 | * Have to do this validation here, as this is in io_read() rw->len | |
115 | * might have chanaged due to buffer selection | |
4ab9d465 | 116 | */ |
0e984ec8 JA |
117 | if (req->flags & REQ_F_BUFFER_SELECT) |
118 | return io_iov_buffer_select_prep(req); | |
4ab9d465 | 119 | |
f3b44f92 JA |
120 | return 0; |
121 | } | |
122 | ||
f688944c JA |
123 | int io_prep_rw_fixed(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
124 | { | |
125 | struct io_ring_ctx *ctx = req->ctx; | |
126 | u16 index; | |
127 | int ret; | |
128 | ||
129 | ret = io_prep_rw(req, sqe); | |
130 | if (unlikely(ret)) | |
131 | return ret; | |
132 | ||
133 | if (unlikely(req->buf_index >= ctx->nr_user_bufs)) | |
134 | return -EFAULT; | |
135 | index = array_index_nospec(req->buf_index, ctx->nr_user_bufs); | |
136 | req->imu = ctx->user_bufs[index]; | |
137 | io_req_set_rsrc_node(req, ctx, 0); | |
138 | return 0; | |
139 | } | |
140 | ||
fc68fcda JA |
141 | /* |
142 | * Multishot read is prepared just like a normal read/write request, only | |
143 | * difference is that we set the MULTISHOT flag. | |
144 | */ | |
145 | int io_read_mshot_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
146 | { | |
49fbe994 | 147 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
fc68fcda JA |
148 | int ret; |
149 | ||
0df96fb7 JA |
150 | /* must be used with provided buffers */ |
151 | if (!(req->flags & REQ_F_BUFFER_SELECT)) | |
152 | return -EINVAL; | |
153 | ||
fc68fcda JA |
154 | ret = io_prep_rw(req, sqe); |
155 | if (unlikely(ret)) | |
156 | return ret; | |
157 | ||
49fbe994 DY |
158 | if (rw->addr || rw->len) |
159 | return -EINVAL; | |
160 | ||
fc68fcda JA |
161 | req->flags |= REQ_F_APOLL_MULTISHOT; |
162 | return 0; | |
163 | } | |
164 | ||
f3b44f92 JA |
165 | void io_readv_writev_cleanup(struct io_kiocb *req) |
166 | { | |
167 | struct io_async_rw *io = req->async_data; | |
168 | ||
169 | kfree(io->free_iovec); | |
170 | } | |
171 | ||
f3b44f92 JA |
172 | static inline loff_t *io_kiocb_update_pos(struct io_kiocb *req) |
173 | { | |
f2ccb5ae | 174 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
f3b44f92 JA |
175 | |
176 | if (rw->kiocb.ki_pos != -1) | |
177 | return &rw->kiocb.ki_pos; | |
178 | ||
179 | if (!(req->file->f_mode & FMODE_STREAM)) { | |
180 | req->flags |= REQ_F_CUR_POS; | |
181 | rw->kiocb.ki_pos = req->file->f_pos; | |
182 | return &rw->kiocb.ki_pos; | |
183 | } | |
184 | ||
185 | rw->kiocb.ki_pos = 0; | |
186 | return NULL; | |
187 | } | |
188 | ||
189 | static void io_req_task_queue_reissue(struct io_kiocb *req) | |
190 | { | |
191 | req->io_task_work.func = io_queue_iowq; | |
192 | io_req_task_work_add(req); | |
193 | } | |
194 | ||
195 | #ifdef CONFIG_BLOCK | |
196 | static bool io_resubmit_prep(struct io_kiocb *req) | |
197 | { | |
198 | struct io_async_rw *io = req->async_data; | |
199 | ||
200 | if (!req_has_async_data(req)) | |
201 | return !io_req_prep_async(req); | |
202 | iov_iter_restore(&io->s.iter, &io->s.iter_state); | |
203 | return true; | |
204 | } | |
205 | ||
206 | static bool io_rw_should_reissue(struct io_kiocb *req) | |
207 | { | |
208 | umode_t mode = file_inode(req->file)->i_mode; | |
209 | struct io_ring_ctx *ctx = req->ctx; | |
210 | ||
211 | if (!S_ISBLK(mode) && !S_ISREG(mode)) | |
212 | return false; | |
213 | if ((req->flags & REQ_F_NOWAIT) || (io_wq_current_is_worker() && | |
214 | !(ctx->flags & IORING_SETUP_IOPOLL))) | |
215 | return false; | |
216 | /* | |
217 | * If ref is dying, we might be running poll reap from the exit work. | |
218 | * Don't attempt to reissue from that path, just let it fail with | |
219 | * -EAGAIN. | |
220 | */ | |
221 | if (percpu_ref_is_dying(&ctx->refs)) | |
222 | return false; | |
223 | /* | |
224 | * Play it safe and assume not safe to re-import and reissue if we're | |
225 | * not in the original thread group (or in task context). | |
226 | */ | |
227 | if (!same_thread_group(req->task, current) || !in_task()) | |
228 | return false; | |
229 | return true; | |
230 | } | |
231 | #else | |
232 | static bool io_resubmit_prep(struct io_kiocb *req) | |
233 | { | |
234 | return false; | |
235 | } | |
236 | static bool io_rw_should_reissue(struct io_kiocb *req) | |
237 | { | |
238 | return false; | |
239 | } | |
240 | #endif | |
241 | ||
a370167f | 242 | static void io_req_end_write(struct io_kiocb *req) |
f3b44f92 | 243 | { |
f3b44f92 | 244 | if (req->flags & REQ_F_ISREG) { |
e484fd73 | 245 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
f3b44f92 | 246 | |
e484fd73 | 247 | kiocb_end_write(&rw->kiocb); |
f3b44f92 JA |
248 | } |
249 | } | |
250 | ||
2ec33a6c JA |
251 | /* |
252 | * Trigger the notifications after having done some IO, and finish the write | |
253 | * accounting, if any. | |
254 | */ | |
255 | static void io_req_io_end(struct io_kiocb *req) | |
256 | { | |
257 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); | |
258 | ||
2ec33a6c | 259 | if (rw->kiocb.ki_flags & IOCB_WRITE) { |
a370167f | 260 | io_req_end_write(req); |
2ec33a6c JA |
261 | fsnotify_modify(req->file); |
262 | } else { | |
263 | fsnotify_access(req->file); | |
264 | } | |
265 | } | |
266 | ||
f3b44f92 JA |
267 | static bool __io_complete_rw_common(struct io_kiocb *req, long res) |
268 | { | |
f3b44f92 JA |
269 | if (unlikely(res != req->cqe.res)) { |
270 | if ((res == -EAGAIN || res == -EOPNOTSUPP) && | |
271 | io_rw_should_reissue(req)) { | |
2ec33a6c JA |
272 | /* |
273 | * Reissue will start accounting again, finish the | |
274 | * current cycle. | |
275 | */ | |
276 | io_req_io_end(req); | |
f3b44f92 JA |
277 | req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO; |
278 | return true; | |
279 | } | |
280 | req_set_fail(req); | |
281 | req->cqe.res = res; | |
282 | } | |
283 | return false; | |
284 | } | |
285 | ||
62bb0647 | 286 | static inline int io_fixup_rw_res(struct io_kiocb *req, long res) |
4d9cb92c PB |
287 | { |
288 | struct io_async_rw *io = req->async_data; | |
289 | ||
290 | /* add previously done IO, if any */ | |
291 | if (req_has_async_data(req) && io->bytes_done > 0) { | |
292 | if (res < 0) | |
293 | res = io->bytes_done; | |
294 | else | |
295 | res += io->bytes_done; | |
296 | } | |
297 | return res; | |
298 | } | |
299 | ||
c92fcfc2 | 300 | void io_req_rw_complete(struct io_kiocb *req, struct io_tw_state *ts) |
b000145e | 301 | { |
099ada2c JA |
302 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
303 | struct kiocb *kiocb = &rw->kiocb; | |
304 | ||
305 | if ((kiocb->ki_flags & IOCB_DIO_CALLER_COMP) && kiocb->dio_complete) { | |
306 | long res = kiocb->dio_complete(rw->kiocb.private); | |
307 | ||
308 | io_req_set_res(req, io_fixup_rw_res(req, res), 0); | |
309 | } | |
310 | ||
2ec33a6c | 311 | io_req_io_end(req); |
3671163b PB |
312 | |
313 | if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) { | |
a282967c | 314 | unsigned issue_flags = ts->locked ? 0 : IO_URING_F_UNLOCKED; |
3671163b PB |
315 | |
316 | req->cqe.flags |= io_put_kbuf(req, issue_flags); | |
317 | } | |
a282967c | 318 | io_req_task_complete(req, ts); |
b000145e JA |
319 | } |
320 | ||
f3b44f92 JA |
321 | static void io_complete_rw(struct kiocb *kiocb, long res) |
322 | { | |
323 | struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb); | |
324 | struct io_kiocb *req = cmd_to_io_kiocb(rw); | |
325 | ||
099ada2c JA |
326 | if (!kiocb->dio_complete || !(kiocb->ki_flags & IOCB_DIO_CALLER_COMP)) { |
327 | if (__io_complete_rw_common(req, res)) | |
328 | return; | |
329 | io_req_set_res(req, io_fixup_rw_res(req, res), 0); | |
330 | } | |
b000145e | 331 | req->io_task_work.func = io_req_rw_complete; |
8751d154 | 332 | __io_req_task_work_add(req, IOU_F_TWQ_LAZY_WAKE); |
f3b44f92 JA |
333 | } |
334 | ||
335 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res) | |
336 | { | |
337 | struct io_rw *rw = container_of(kiocb, struct io_rw, kiocb); | |
338 | struct io_kiocb *req = cmd_to_io_kiocb(rw); | |
339 | ||
340 | if (kiocb->ki_flags & IOCB_WRITE) | |
a370167f | 341 | io_req_end_write(req); |
f3b44f92 JA |
342 | if (unlikely(res != req->cqe.res)) { |
343 | if (res == -EAGAIN && io_rw_should_reissue(req)) { | |
344 | req->flags |= REQ_F_REISSUE | REQ_F_PARTIAL_IO; | |
345 | return; | |
346 | } | |
347 | req->cqe.res = res; | |
348 | } | |
349 | ||
350 | /* order with io_iopoll_complete() checking ->iopoll_completed */ | |
351 | smp_store_release(&req->iopoll_completed, 1); | |
352 | } | |
353 | ||
fe80eb15 JA |
354 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) |
355 | { | |
356 | /* IO was queued async, completion will happen later */ | |
357 | if (ret == -EIOCBQUEUED) | |
358 | return; | |
359 | ||
360 | /* transform internal restart error codes */ | |
361 | if (unlikely(ret < 0)) { | |
362 | switch (ret) { | |
363 | case -ERESTARTSYS: | |
364 | case -ERESTARTNOINTR: | |
365 | case -ERESTARTNOHAND: | |
366 | case -ERESTART_RESTARTBLOCK: | |
367 | /* | |
368 | * We can't just restart the syscall, since previously | |
369 | * submitted sqes may already be in progress. Just fail | |
370 | * this IO with EINTR. | |
371 | */ | |
372 | ret = -EINTR; | |
373 | break; | |
374 | } | |
375 | } | |
376 | ||
377 | INDIRECT_CALL_2(kiocb->ki_complete, io_complete_rw_iopoll, | |
378 | io_complete_rw, kiocb, ret); | |
379 | } | |
380 | ||
df9830d8 | 381 | static int kiocb_done(struct io_kiocb *req, ssize_t ret, |
f3b44f92 JA |
382 | unsigned int issue_flags) |
383 | { | |
f2ccb5ae | 384 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
4d9cb92c | 385 | unsigned final_ret = io_fixup_rw_res(req, ret); |
f3b44f92 | 386 | |
1939316b | 387 | if (ret >= 0 && req->flags & REQ_F_CUR_POS) |
f3b44f92 | 388 | req->file->f_pos = rw->kiocb.ki_pos; |
df9830d8 PB |
389 | if (ret >= 0 && (rw->kiocb.ki_complete == io_complete_rw)) { |
390 | if (!__io_complete_rw_common(req, ret)) { | |
2ec33a6c JA |
391 | /* |
392 | * Safe to call io_end from here as we're inline | |
393 | * from the submission path. | |
394 | */ | |
395 | io_req_io_end(req); | |
4d9cb92c | 396 | io_req_set_res(req, final_ret, |
df9830d8 PB |
397 | io_put_kbuf(req, issue_flags)); |
398 | return IOU_OK; | |
399 | } | |
400 | } else { | |
f3b44f92 | 401 | io_rw_done(&rw->kiocb, ret); |
df9830d8 | 402 | } |
f3b44f92 JA |
403 | |
404 | if (req->flags & REQ_F_REISSUE) { | |
405 | req->flags &= ~REQ_F_REISSUE; | |
406 | if (io_resubmit_prep(req)) | |
407 | io_req_task_queue_reissue(req); | |
408 | else | |
4d9cb92c | 409 | io_req_task_queue_fail(req, final_ret); |
f3b44f92 | 410 | } |
df9830d8 | 411 | return IOU_ISSUE_SKIP_COMPLETE; |
f3b44f92 JA |
412 | } |
413 | ||
f3b44f92 JA |
414 | static struct iovec *__io_import_iovec(int ddir, struct io_kiocb *req, |
415 | struct io_rw_state *s, | |
416 | unsigned int issue_flags) | |
417 | { | |
f2ccb5ae | 418 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
f3b44f92 JA |
419 | struct iov_iter *iter = &s->iter; |
420 | u8 opcode = req->opcode; | |
421 | struct iovec *iovec; | |
422 | void __user *buf; | |
423 | size_t sqe_len; | |
424 | ssize_t ret; | |
425 | ||
426 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { | |
f337a84d | 427 | ret = io_import_fixed(ddir, iter, req->imu, rw->addr, rw->len); |
f3b44f92 JA |
428 | if (ret) |
429 | return ERR_PTR(ret); | |
430 | return NULL; | |
431 | } | |
432 | ||
433 | buf = u64_to_user_ptr(rw->addr); | |
434 | sqe_len = rw->len; | |
435 | ||
d2d778fb | 436 | if (!io_issue_defs[opcode].vectored || req->flags & REQ_F_BUFFER_SELECT) { |
f3b44f92 JA |
437 | if (io_do_buffer_select(req)) { |
438 | buf = io_buffer_select(req, &sqe_len, issue_flags); | |
439 | if (!buf) | |
440 | return ERR_PTR(-ENOBUFS); | |
441 | rw->addr = (unsigned long) buf; | |
442 | rw->len = sqe_len; | |
443 | } | |
444 | ||
1e23db45 | 445 | ret = import_ubuf(ddir, buf, sqe_len, iter); |
f3b44f92 JA |
446 | if (ret) |
447 | return ERR_PTR(ret); | |
448 | return NULL; | |
449 | } | |
450 | ||
451 | iovec = s->fast_iov; | |
f3b44f92 JA |
452 | ret = __import_iovec(ddir, buf, sqe_len, UIO_FASTIOV, &iovec, iter, |
453 | req->ctx->compat); | |
454 | if (unlikely(ret < 0)) | |
455 | return ERR_PTR(ret); | |
456 | return iovec; | |
457 | } | |
458 | ||
459 | static inline int io_import_iovec(int rw, struct io_kiocb *req, | |
460 | struct iovec **iovec, struct io_rw_state *s, | |
461 | unsigned int issue_flags) | |
462 | { | |
463 | *iovec = __io_import_iovec(rw, req, s, issue_flags); | |
81594e7e | 464 | if (IS_ERR(*iovec)) |
f3b44f92 JA |
465 | return PTR_ERR(*iovec); |
466 | ||
467 | iov_iter_save_state(&s->iter, &s->iter_state); | |
468 | return 0; | |
469 | } | |
470 | ||
471 | static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb) | |
472 | { | |
473 | return (kiocb->ki_filp->f_mode & FMODE_STREAM) ? NULL : &kiocb->ki_pos; | |
474 | } | |
475 | ||
476 | /* | |
477 | * For files that don't have ->read_iter() and ->write_iter(), handle them | |
478 | * by looping over ->read() or ->write() manually. | |
479 | */ | |
480 | static ssize_t loop_rw_iter(int ddir, struct io_rw *rw, struct iov_iter *iter) | |
481 | { | |
482 | struct kiocb *kiocb = &rw->kiocb; | |
483 | struct file *file = kiocb->ki_filp; | |
484 | ssize_t ret = 0; | |
485 | loff_t *ppos; | |
486 | ||
487 | /* | |
488 | * Don't support polled IO through this interface, and we can't | |
489 | * support non-blocking either. For the latter, this just causes | |
490 | * the kiocb to be handled from an async context. | |
491 | */ | |
492 | if (kiocb->ki_flags & IOCB_HIPRI) | |
493 | return -EOPNOTSUPP; | |
494 | if ((kiocb->ki_flags & IOCB_NOWAIT) && | |
495 | !(kiocb->ki_filp->f_flags & O_NONBLOCK)) | |
496 | return -EAGAIN; | |
497 | ||
498 | ppos = io_kiocb_ppos(kiocb); | |
499 | ||
500 | while (iov_iter_count(iter)) { | |
95e49cf8 JA |
501 | void __user *addr; |
502 | size_t len; | |
f3b44f92 JA |
503 | ssize_t nr; |
504 | ||
1e23db45 | 505 | if (iter_is_ubuf(iter)) { |
95e49cf8 JA |
506 | addr = iter->ubuf + iter->iov_offset; |
507 | len = iov_iter_count(iter); | |
1e23db45 | 508 | } else if (!iov_iter_is_bvec(iter)) { |
95e49cf8 JA |
509 | addr = iter_iov_addr(iter); |
510 | len = iter_iov_len(iter); | |
f3b44f92 | 511 | } else { |
95e49cf8 JA |
512 | addr = u64_to_user_ptr(rw->addr); |
513 | len = rw->len; | |
f3b44f92 JA |
514 | } |
515 | ||
95e49cf8 JA |
516 | if (ddir == READ) |
517 | nr = file->f_op->read(file, addr, len, ppos); | |
518 | else | |
519 | nr = file->f_op->write(file, addr, len, ppos); | |
f3b44f92 JA |
520 | |
521 | if (nr < 0) { | |
522 | if (!ret) | |
523 | ret = nr; | |
524 | break; | |
525 | } | |
526 | ret += nr; | |
527 | if (!iov_iter_is_bvec(iter)) { | |
528 | iov_iter_advance(iter, nr); | |
529 | } else { | |
530 | rw->addr += nr; | |
531 | rw->len -= nr; | |
532 | if (!rw->len) | |
533 | break; | |
534 | } | |
95e49cf8 | 535 | if (nr != len) |
f3b44f92 JA |
536 | break; |
537 | } | |
538 | ||
539 | return ret; | |
540 | } | |
541 | ||
542 | static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec, | |
543 | const struct iovec *fast_iov, struct iov_iter *iter) | |
544 | { | |
545 | struct io_async_rw *io = req->async_data; | |
546 | ||
547 | memcpy(&io->s.iter, iter, sizeof(*iter)); | |
548 | io->free_iovec = iovec; | |
549 | io->bytes_done = 0; | |
550 | /* can only be fixed buffers, no need to do anything */ | |
1e23db45 | 551 | if (iov_iter_is_bvec(iter) || iter_is_ubuf(iter)) |
f3b44f92 JA |
552 | return; |
553 | if (!iovec) { | |
554 | unsigned iov_off = 0; | |
555 | ||
de4f5fed JA |
556 | io->s.iter.__iov = io->s.fast_iov; |
557 | if (iter->__iov != fast_iov) { | |
558 | iov_off = iter_iov(iter) - fast_iov; | |
559 | io->s.iter.__iov += iov_off; | |
f3b44f92 JA |
560 | } |
561 | if (io->s.fast_iov != fast_iov) | |
562 | memcpy(io->s.fast_iov + iov_off, fast_iov + iov_off, | |
563 | sizeof(struct iovec) * iter->nr_segs); | |
564 | } else { | |
565 | req->flags |= REQ_F_NEED_CLEANUP; | |
566 | } | |
567 | } | |
568 | ||
569 | static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec, | |
570 | struct io_rw_state *s, bool force) | |
571 | { | |
f30bd4d0 | 572 | if (!force && !io_cold_defs[req->opcode].prep_async) |
f3b44f92 | 573 | return 0; |
0df96fb7 JA |
574 | /* opcode type doesn't need async data */ |
575 | if (!io_cold_defs[req->opcode].async_size) | |
576 | return 0; | |
f3b44f92 JA |
577 | if (!req_has_async_data(req)) { |
578 | struct io_async_rw *iorw; | |
579 | ||
580 | if (io_alloc_async_data(req)) { | |
581 | kfree(iovec); | |
582 | return -ENOMEM; | |
583 | } | |
584 | ||
585 | io_req_map_rw(req, iovec, s->fast_iov, &s->iter); | |
586 | iorw = req->async_data; | |
587 | /* we've copied and mapped the iter, ensure state is saved */ | |
588 | iov_iter_save_state(&iorw->s.iter, &iorw->s.iter_state); | |
589 | } | |
590 | return 0; | |
591 | } | |
592 | ||
593 | static inline int io_rw_prep_async(struct io_kiocb *req, int rw) | |
594 | { | |
595 | struct io_async_rw *iorw = req->async_data; | |
596 | struct iovec *iov; | |
597 | int ret; | |
598 | ||
0a535edd JA |
599 | iorw->bytes_done = 0; |
600 | iorw->free_iovec = NULL; | |
601 | ||
f3b44f92 JA |
602 | /* submission path, ->uring_lock should already be taken */ |
603 | ret = io_import_iovec(rw, req, &iov, &iorw->s, 0); | |
604 | if (unlikely(ret < 0)) | |
605 | return ret; | |
606 | ||
0a535edd JA |
607 | if (iov) { |
608 | iorw->free_iovec = iov; | |
f3b44f92 | 609 | req->flags |= REQ_F_NEED_CLEANUP; |
0a535edd JA |
610 | } |
611 | ||
f3b44f92 JA |
612 | return 0; |
613 | } | |
614 | ||
615 | int io_readv_prep_async(struct io_kiocb *req) | |
616 | { | |
de4eda9d | 617 | return io_rw_prep_async(req, ITER_DEST); |
f3b44f92 JA |
618 | } |
619 | ||
620 | int io_writev_prep_async(struct io_kiocb *req) | |
621 | { | |
de4eda9d | 622 | return io_rw_prep_async(req, ITER_SOURCE); |
f3b44f92 JA |
623 | } |
624 | ||
625 | /* | |
626 | * This is our waitqueue callback handler, registered through __folio_lock_async() | |
627 | * when we initially tried to do the IO with the iocb armed our waitqueue. | |
628 | * This gets called when the page is unlocked, and we generally expect that to | |
629 | * happen when the page IO is completed and the page is now uptodate. This will | |
630 | * queue a task_work based retry of the operation, attempting to copy the data | |
631 | * again. If the latter fails because the page was NOT uptodate, then we will | |
632 | * do a thread based blocking retry of the operation. That's the unexpected | |
633 | * slow path. | |
634 | */ | |
635 | static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode, | |
636 | int sync, void *arg) | |
637 | { | |
638 | struct wait_page_queue *wpq; | |
639 | struct io_kiocb *req = wait->private; | |
f2ccb5ae | 640 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
f3b44f92 JA |
641 | struct wait_page_key *key = arg; |
642 | ||
643 | wpq = container_of(wait, struct wait_page_queue, wait); | |
644 | ||
645 | if (!wake_page_match(wpq, key)) | |
646 | return 0; | |
647 | ||
648 | rw->kiocb.ki_flags &= ~IOCB_WAITQ; | |
649 | list_del_init(&wait->entry); | |
650 | io_req_task_queue(req); | |
651 | return 1; | |
652 | } | |
653 | ||
654 | /* | |
655 | * This controls whether a given IO request should be armed for async page | |
656 | * based retry. If we return false here, the request is handed to the async | |
657 | * worker threads for retry. If we're doing buffered reads on a regular file, | |
658 | * we prepare a private wait_page_queue entry and retry the operation. This | |
659 | * will either succeed because the page is now uptodate and unlocked, or it | |
660 | * will register a callback when the page is unlocked at IO completion. Through | |
661 | * that callback, io_uring uses task_work to setup a retry of the operation. | |
662 | * That retry will attempt the buffered read again. The retry will generally | |
663 | * succeed, or in rare cases where it fails, we then fall back to using the | |
664 | * async worker threads for a blocking retry. | |
665 | */ | |
666 | static bool io_rw_should_retry(struct io_kiocb *req) | |
667 | { | |
668 | struct io_async_rw *io = req->async_data; | |
669 | struct wait_page_queue *wait = &io->wpq; | |
f2ccb5ae | 670 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
f3b44f92 JA |
671 | struct kiocb *kiocb = &rw->kiocb; |
672 | ||
673 | /* never retry for NOWAIT, we just complete with -EAGAIN */ | |
674 | if (req->flags & REQ_F_NOWAIT) | |
675 | return false; | |
676 | ||
677 | /* Only for buffered IO */ | |
678 | if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI)) | |
679 | return false; | |
680 | ||
681 | /* | |
682 | * just use poll if we can, and don't attempt if the fs doesn't | |
683 | * support callback based unlocks | |
684 | */ | |
685 | if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC)) | |
686 | return false; | |
687 | ||
688 | wait->wait.func = io_async_buf_func; | |
689 | wait->wait.private = req; | |
690 | wait->wait.flags = 0; | |
691 | INIT_LIST_HEAD(&wait->wait.entry); | |
692 | kiocb->ki_flags |= IOCB_WAITQ; | |
693 | kiocb->ki_flags &= ~IOCB_NOWAIT; | |
694 | kiocb->ki_waitq = wait; | |
695 | return true; | |
696 | } | |
697 | ||
698 | static inline int io_iter_do_read(struct io_rw *rw, struct iov_iter *iter) | |
699 | { | |
700 | struct file *file = rw->kiocb.ki_filp; | |
701 | ||
702 | if (likely(file->f_op->read_iter)) | |
703 | return call_read_iter(file, &rw->kiocb, iter); | |
704 | else if (file->f_op->read) | |
705 | return loop_rw_iter(READ, rw, iter); | |
706 | else | |
707 | return -EINVAL; | |
708 | } | |
709 | ||
4e17aaab | 710 | static bool need_complete_io(struct io_kiocb *req) |
f3b44f92 JA |
711 | { |
712 | return req->flags & REQ_F_ISREG || | |
713 | S_ISBLK(file_inode(req->file)->i_mode); | |
714 | } | |
715 | ||
f3b44f92 JA |
716 | static int io_rw_init_file(struct io_kiocb *req, fmode_t mode) |
717 | { | |
f2ccb5ae | 718 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
f3b44f92 JA |
719 | struct kiocb *kiocb = &rw->kiocb; |
720 | struct io_ring_ctx *ctx = req->ctx; | |
721 | struct file *file = req->file; | |
722 | int ret; | |
723 | ||
724 | if (unlikely(!file || !(file->f_mode & mode))) | |
725 | return -EBADF; | |
726 | ||
3beed235 | 727 | if (!(req->flags & REQ_F_FIXED_FILE)) |
8487f083 | 728 | req->flags |= io_file_get_flags(file); |
f3b44f92 | 729 | |
5264406c | 730 | kiocb->ki_flags = file->f_iocb_flags; |
f3b44f92 JA |
731 | ret = kiocb_set_rw_flags(kiocb, rw->flags); |
732 | if (unlikely(ret)) | |
733 | return ret; | |
12e4e8c7 | 734 | kiocb->ki_flags |= IOCB_ALLOC_CACHE; |
f3b44f92 JA |
735 | |
736 | /* | |
737 | * If the file is marked O_NONBLOCK, still allow retry for it if it | |
738 | * supports async. Otherwise it's impossible to use O_NONBLOCK files | |
739 | * reliably. If not, or it IOCB_NOWAIT is set, don't retry. | |
740 | */ | |
741 | if ((kiocb->ki_flags & IOCB_NOWAIT) || | |
742 | ((file->f_flags & O_NONBLOCK) && !io_file_supports_nowait(req))) | |
743 | req->flags |= REQ_F_NOWAIT; | |
744 | ||
745 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
746 | if (!(kiocb->ki_flags & IOCB_DIRECT) || !file->f_op->iopoll) | |
747 | return -EOPNOTSUPP; | |
748 | ||
749 | kiocb->private = NULL; | |
12e4e8c7 | 750 | kiocb->ki_flags |= IOCB_HIPRI; |
f3b44f92 JA |
751 | kiocb->ki_complete = io_complete_rw_iopoll; |
752 | req->iopoll_completed = 0; | |
753 | } else { | |
754 | if (kiocb->ki_flags & IOCB_HIPRI) | |
755 | return -EINVAL; | |
756 | kiocb->ki_complete = io_complete_rw; | |
757 | } | |
758 | ||
759 | return 0; | |
760 | } | |
761 | ||
a08d195b | 762 | static int __io_read(struct io_kiocb *req, unsigned int issue_flags) |
f3b44f92 | 763 | { |
f2ccb5ae | 764 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
f3b44f92 JA |
765 | struct io_rw_state __s, *s = &__s; |
766 | struct iovec *iovec; | |
767 | struct kiocb *kiocb = &rw->kiocb; | |
768 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; | |
769 | struct io_async_rw *io; | |
770 | ssize_t ret, ret2; | |
771 | loff_t *ppos; | |
772 | ||
773 | if (!req_has_async_data(req)) { | |
de4eda9d | 774 | ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags); |
f3b44f92 JA |
775 | if (unlikely(ret < 0)) |
776 | return ret; | |
777 | } else { | |
778 | io = req->async_data; | |
779 | s = &io->s; | |
780 | ||
781 | /* | |
782 | * Safe and required to re-import if we're using provided | |
783 | * buffers, as we dropped the selected one before retry. | |
784 | */ | |
785 | if (io_do_buffer_select(req)) { | |
de4eda9d | 786 | ret = io_import_iovec(ITER_DEST, req, &iovec, s, issue_flags); |
f3b44f92 JA |
787 | if (unlikely(ret < 0)) |
788 | return ret; | |
789 | } | |
790 | ||
791 | /* | |
792 | * We come here from an earlier attempt, restore our state to | |
793 | * match in case it doesn't. It's cheap enough that we don't | |
794 | * need to make this conditional. | |
795 | */ | |
796 | iov_iter_restore(&s->iter, &s->iter_state); | |
797 | iovec = NULL; | |
798 | } | |
799 | ret = io_rw_init_file(req, FMODE_READ); | |
800 | if (unlikely(ret)) { | |
801 | kfree(iovec); | |
802 | return ret; | |
803 | } | |
804 | req->cqe.res = iov_iter_count(&s->iter); | |
805 | ||
806 | if (force_nonblock) { | |
807 | /* If the file doesn't support async, just async punt */ | |
808 | if (unlikely(!io_file_supports_nowait(req))) { | |
809 | ret = io_setup_async_rw(req, iovec, s, true); | |
810 | return ret ?: -EAGAIN; | |
811 | } | |
812 | kiocb->ki_flags |= IOCB_NOWAIT; | |
813 | } else { | |
814 | /* Ensure we clear previously set non-block flag */ | |
815 | kiocb->ki_flags &= ~IOCB_NOWAIT; | |
816 | } | |
817 | ||
818 | ppos = io_kiocb_update_pos(req); | |
819 | ||
820 | ret = rw_verify_area(READ, req->file, ppos, req->cqe.res); | |
821 | if (unlikely(ret)) { | |
822 | kfree(iovec); | |
823 | return ret; | |
824 | } | |
825 | ||
826 | ret = io_iter_do_read(rw, &s->iter); | |
827 | ||
828 | if (ret == -EAGAIN || (req->flags & REQ_F_REISSUE)) { | |
829 | req->flags &= ~REQ_F_REISSUE; | |
d2d778fb JA |
830 | /* |
831 | * If we can poll, just do that. For a vectored read, we'll | |
832 | * need to copy state first. | |
833 | */ | |
834 | if (file_can_poll(req->file) && !io_issue_defs[req->opcode].vectored) | |
f3b44f92 JA |
835 | return -EAGAIN; |
836 | /* IOPOLL retry should happen for io-wq threads */ | |
837 | if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
838 | goto done; | |
839 | /* no retry on NONBLOCK nor RWF_NOWAIT */ | |
840 | if (req->flags & REQ_F_NOWAIT) | |
841 | goto done; | |
842 | ret = 0; | |
843 | } else if (ret == -EIOCBQUEUED) { | |
df9830d8 PB |
844 | if (iovec) |
845 | kfree(iovec); | |
846 | return IOU_ISSUE_SKIP_COMPLETE; | |
f3b44f92 | 847 | } else if (ret == req->cqe.res || ret <= 0 || !force_nonblock || |
4e17aaab | 848 | (req->flags & REQ_F_NOWAIT) || !need_complete_io(req)) { |
f3b44f92 JA |
849 | /* read all, failed, already did sync or don't want to retry */ |
850 | goto done; | |
851 | } | |
852 | ||
853 | /* | |
854 | * Don't depend on the iter state matching what was consumed, or being | |
855 | * untouched in case of error. Restore it and we'll advance it | |
856 | * manually if we need to. | |
857 | */ | |
858 | iov_iter_restore(&s->iter, &s->iter_state); | |
859 | ||
860 | ret2 = io_setup_async_rw(req, iovec, s, true); | |
f3b44f92 | 861 | iovec = NULL; |
c278d9f8 PB |
862 | if (ret2) { |
863 | ret = ret > 0 ? ret : ret2; | |
864 | goto done; | |
865 | } | |
866 | ||
f3b44f92 JA |
867 | io = req->async_data; |
868 | s = &io->s; | |
869 | /* | |
870 | * Now use our persistent iterator and state, if we aren't already. | |
871 | * We've restored and mapped the iter to match. | |
872 | */ | |
873 | ||
874 | do { | |
875 | /* | |
876 | * We end up here because of a partial read, either from | |
877 | * above or inside this loop. Advance the iter by the bytes | |
878 | * that were consumed. | |
879 | */ | |
880 | iov_iter_advance(&s->iter, ret); | |
881 | if (!iov_iter_count(&s->iter)) | |
882 | break; | |
883 | io->bytes_done += ret; | |
884 | iov_iter_save_state(&s->iter, &s->iter_state); | |
885 | ||
886 | /* if we can retry, do so with the callbacks armed */ | |
887 | if (!io_rw_should_retry(req)) { | |
888 | kiocb->ki_flags &= ~IOCB_WAITQ; | |
889 | return -EAGAIN; | |
890 | } | |
891 | ||
bf68b5b3 | 892 | req->cqe.res = iov_iter_count(&s->iter); |
f3b44f92 JA |
893 | /* |
894 | * Now retry read with the IOCB_WAITQ parts set in the iocb. If | |
895 | * we get -EIOCBQUEUED, then we'll get a notification when the | |
896 | * desired page gets unlocked. We can also get a partial read | |
897 | * here, and if we do, then just retry at the new offset. | |
898 | */ | |
899 | ret = io_iter_do_read(rw, &s->iter); | |
900 | if (ret == -EIOCBQUEUED) | |
901 | return IOU_ISSUE_SKIP_COMPLETE; | |
902 | /* we got some bytes, but not all. retry. */ | |
903 | kiocb->ki_flags &= ~IOCB_WAITQ; | |
904 | iov_iter_restore(&s->iter, &s->iter_state); | |
905 | } while (ret > 0); | |
906 | done: | |
f3b44f92 JA |
907 | /* it's faster to check here then delegate to kfree */ |
908 | if (iovec) | |
909 | kfree(iovec); | |
a08d195b JA |
910 | return ret; |
911 | } | |
912 | ||
913 | int io_read(struct io_kiocb *req, unsigned int issue_flags) | |
914 | { | |
915 | int ret; | |
916 | ||
917 | ret = __io_read(req, issue_flags); | |
918 | if (ret >= 0) | |
919 | return kiocb_done(req, ret, issue_flags); | |
920 | ||
921 | return ret; | |
f3b44f92 JA |
922 | } |
923 | ||
fc68fcda JA |
924 | int io_read_mshot(struct io_kiocb *req, unsigned int issue_flags) |
925 | { | |
e5375929 | 926 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
fc68fcda JA |
927 | unsigned int cflags = 0; |
928 | int ret; | |
929 | ||
930 | /* | |
931 | * Multishot MUST be used on a pollable file | |
932 | */ | |
933 | if (!file_can_poll(req->file)) | |
934 | return -EBADFD; | |
935 | ||
936 | ret = __io_read(req, issue_flags); | |
937 | ||
938 | /* | |
939 | * If we get -EAGAIN, recycle our buffer and just let normal poll | |
940 | * handling arm it. | |
941 | */ | |
942 | if (ret == -EAGAIN) { | |
e5375929 DY |
943 | /* |
944 | * Reset rw->len to 0 again to avoid clamping future mshot | |
945 | * reads, in case the buffer size varies. | |
946 | */ | |
947 | if (io_kbuf_recycle(req, issue_flags)) | |
948 | rw->len = 0; | |
fc68fcda JA |
949 | return -EAGAIN; |
950 | } | |
951 | ||
952 | /* | |
953 | * Any successful return value will keep the multishot read armed. | |
954 | */ | |
955 | if (ret > 0) { | |
956 | /* | |
957 | * Put our buffer and post a CQE. If we fail to post a CQE, then | |
958 | * jump to the termination path. This request is then done. | |
959 | */ | |
960 | cflags = io_put_kbuf(req, issue_flags); | |
e5375929 | 961 | rw->len = 0; /* similarly to above, reset len to 0 */ |
fc68fcda JA |
962 | |
963 | if (io_fill_cqe_req_aux(req, | |
964 | issue_flags & IO_URING_F_COMPLETE_DEFER, | |
965 | ret, cflags | IORING_CQE_F_MORE)) { | |
c79f52f0 JA |
966 | if (issue_flags & IO_URING_F_MULTISHOT) { |
967 | /* | |
968 | * Force retry, as we might have more data to | |
969 | * be read and otherwise it won't get retried | |
970 | * until (if ever) another poll is triggered. | |
971 | */ | |
972 | io_poll_multishot_retry(req); | |
fc68fcda | 973 | return IOU_ISSUE_SKIP_COMPLETE; |
c79f52f0 | 974 | } |
fc68fcda JA |
975 | return -EAGAIN; |
976 | } | |
977 | } | |
978 | ||
979 | /* | |
980 | * Either an error, or we've hit overflow posting the CQE. For any | |
981 | * multishot request, hitting overflow will terminate it. | |
982 | */ | |
983 | io_req_set_res(req, ret, cflags); | |
984 | if (issue_flags & IO_URING_F_MULTISHOT) | |
985 | return IOU_STOP_MULTISHOT; | |
986 | return IOU_OK; | |
f3b44f92 JA |
987 | } |
988 | ||
989 | int io_write(struct io_kiocb *req, unsigned int issue_flags) | |
990 | { | |
f2ccb5ae | 991 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); |
f3b44f92 JA |
992 | struct io_rw_state __s, *s = &__s; |
993 | struct iovec *iovec; | |
994 | struct kiocb *kiocb = &rw->kiocb; | |
995 | bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK; | |
996 | ssize_t ret, ret2; | |
997 | loff_t *ppos; | |
998 | ||
999 | if (!req_has_async_data(req)) { | |
de4eda9d | 1000 | ret = io_import_iovec(ITER_SOURCE, req, &iovec, s, issue_flags); |
f3b44f92 JA |
1001 | if (unlikely(ret < 0)) |
1002 | return ret; | |
1003 | } else { | |
1004 | struct io_async_rw *io = req->async_data; | |
1005 | ||
1006 | s = &io->s; | |
1007 | iov_iter_restore(&s->iter, &s->iter_state); | |
1008 | iovec = NULL; | |
1009 | } | |
1010 | ret = io_rw_init_file(req, FMODE_WRITE); | |
1011 | if (unlikely(ret)) { | |
1012 | kfree(iovec); | |
1013 | return ret; | |
1014 | } | |
1015 | req->cqe.res = iov_iter_count(&s->iter); | |
1016 | ||
1017 | if (force_nonblock) { | |
1018 | /* If the file doesn't support async, just async punt */ | |
1019 | if (unlikely(!io_file_supports_nowait(req))) | |
1020 | goto copy_iov; | |
1021 | ||
4e17aaab SR |
1022 | /* File path supports NOWAIT for non-direct_IO only for block devices. */ |
1023 | if (!(kiocb->ki_flags & IOCB_DIRECT) && | |
1024 | !(kiocb->ki_filp->f_mode & FMODE_BUF_WASYNC) && | |
1025 | (req->flags & REQ_F_ISREG)) | |
f3b44f92 JA |
1026 | goto copy_iov; |
1027 | ||
1028 | kiocb->ki_flags |= IOCB_NOWAIT; | |
1029 | } else { | |
1030 | /* Ensure we clear previously set non-block flag */ | |
1031 | kiocb->ki_flags &= ~IOCB_NOWAIT; | |
1032 | } | |
1033 | ||
1034 | ppos = io_kiocb_update_pos(req); | |
1035 | ||
1036 | ret = rw_verify_area(WRITE, req->file, ppos, req->cqe.res); | |
df9830d8 PB |
1037 | if (unlikely(ret)) { |
1038 | kfree(iovec); | |
1039 | return ret; | |
1040 | } | |
f3b44f92 | 1041 | |
e484fd73 AG |
1042 | if (req->flags & REQ_F_ISREG) |
1043 | kiocb_start_write(kiocb); | |
f3b44f92 JA |
1044 | kiocb->ki_flags |= IOCB_WRITE; |
1045 | ||
1046 | if (likely(req->file->f_op->write_iter)) | |
1047 | ret2 = call_write_iter(req->file, kiocb, &s->iter); | |
1048 | else if (req->file->f_op->write) | |
1049 | ret2 = loop_rw_iter(WRITE, rw, &s->iter); | |
1050 | else | |
1051 | ret2 = -EINVAL; | |
1052 | ||
1053 | if (req->flags & REQ_F_REISSUE) { | |
1054 | req->flags &= ~REQ_F_REISSUE; | |
1055 | ret2 = -EAGAIN; | |
1056 | } | |
1057 | ||
1058 | /* | |
1059 | * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just | |
1060 | * retry them without IOCB_NOWAIT. | |
1061 | */ | |
1062 | if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT)) | |
1063 | ret2 = -EAGAIN; | |
1064 | /* no retry on NONBLOCK nor RWF_NOWAIT */ | |
1065 | if (ret2 == -EAGAIN && (req->flags & REQ_F_NOWAIT)) | |
1066 | goto done; | |
1067 | if (!force_nonblock || ret2 != -EAGAIN) { | |
1068 | /* IOPOLL retry should happen for io-wq threads */ | |
1069 | if (ret2 == -EAGAIN && (req->ctx->flags & IORING_SETUP_IOPOLL)) | |
1070 | goto copy_iov; | |
4e17aaab SR |
1071 | |
1072 | if (ret2 != req->cqe.res && ret2 >= 0 && need_complete_io(req)) { | |
c86416c6 | 1073 | struct io_async_rw *io; |
4e17aaab | 1074 | |
1c849b48 SR |
1075 | trace_io_uring_short_write(req->ctx, kiocb->ki_pos - ret2, |
1076 | req->cqe.res, ret2); | |
1077 | ||
4e17aaab SR |
1078 | /* This is a partial write. The file pos has already been |
1079 | * updated, setup the async struct to complete the request | |
1080 | * in the worker. Also update bytes_done to account for | |
1081 | * the bytes already written. | |
1082 | */ | |
1083 | iov_iter_save_state(&s->iter, &s->iter_state); | |
1084 | ret = io_setup_async_rw(req, iovec, s, true); | |
1085 | ||
c86416c6 SR |
1086 | io = req->async_data; |
1087 | if (io) | |
1088 | io->bytes_done += ret2; | |
4e17aaab | 1089 | |
e053aaf4 | 1090 | if (kiocb->ki_flags & IOCB_WRITE) |
a370167f | 1091 | io_req_end_write(req); |
4e17aaab SR |
1092 | return ret ? ret : -EAGAIN; |
1093 | } | |
f3b44f92 | 1094 | done: |
df9830d8 | 1095 | ret = kiocb_done(req, ret2, issue_flags); |
f3b44f92 JA |
1096 | } else { |
1097 | copy_iov: | |
1098 | iov_iter_restore(&s->iter, &s->iter_state); | |
1099 | ret = io_setup_async_rw(req, iovec, s, false); | |
e053aaf4 JA |
1100 | if (!ret) { |
1101 | if (kiocb->ki_flags & IOCB_WRITE) | |
a370167f | 1102 | io_req_end_write(req); |
e053aaf4 JA |
1103 | return -EAGAIN; |
1104 | } | |
1105 | return ret; | |
f3b44f92 | 1106 | } |
f3b44f92 JA |
1107 | /* it's reportedly faster than delegating the null check to kfree() */ |
1108 | if (iovec) | |
1109 | kfree(iovec); | |
1110 | return ret; | |
1111 | } | |
1112 | ||
47b4c686 PB |
1113 | void io_rw_fail(struct io_kiocb *req) |
1114 | { | |
1115 | int res; | |
1116 | ||
1117 | res = io_fixup_rw_res(req, req->cqe.res); | |
1118 | io_req_set_res(req, res, req->cqe.flags); | |
1119 | } | |
1120 | ||
f3b44f92 JA |
1121 | int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin) |
1122 | { | |
1123 | struct io_wq_work_node *pos, *start, *prev; | |
54bdd67d | 1124 | unsigned int poll_flags = 0; |
f3b44f92 JA |
1125 | DEFINE_IO_COMP_BATCH(iob); |
1126 | int nr_events = 0; | |
1127 | ||
1128 | /* | |
1129 | * Only spin for completions if we don't have multiple devices hanging | |
1130 | * off our complete list. | |
1131 | */ | |
1132 | if (ctx->poll_multi_queue || force_nonspin) | |
1133 | poll_flags |= BLK_POLL_ONESHOT; | |
1134 | ||
1135 | wq_list_for_each(pos, start, &ctx->iopoll_list) { | |
1136 | struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list); | |
a1119fb0 | 1137 | struct file *file = req->file; |
f3b44f92 JA |
1138 | int ret; |
1139 | ||
1140 | /* | |
1141 | * Move completed and retryable entries to our local lists. | |
1142 | * If we find a request that requires polling, break out | |
1143 | * and complete those lists first, if we have entries there. | |
1144 | */ | |
1145 | if (READ_ONCE(req->iopoll_completed)) | |
1146 | break; | |
1147 | ||
5756a3a7 | 1148 | if (req->opcode == IORING_OP_URING_CMD) { |
a1119fb0 | 1149 | struct io_uring_cmd *ioucmd; |
5756a3a7 | 1150 | |
a1119fb0 | 1151 | ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd); |
de97fcb3 JA |
1152 | ret = file->f_op->uring_cmd_iopoll(ioucmd, &iob, |
1153 | poll_flags); | |
a1119fb0 JA |
1154 | } else { |
1155 | struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw); | |
1156 | ||
1157 | ret = file->f_op->iopoll(&rw->kiocb, &iob, poll_flags); | |
1158 | } | |
f3b44f92 JA |
1159 | if (unlikely(ret < 0)) |
1160 | return ret; | |
1161 | else if (ret) | |
1162 | poll_flags |= BLK_POLL_ONESHOT; | |
1163 | ||
1164 | /* iopoll may have completed current req */ | |
1165 | if (!rq_list_empty(iob.req_list) || | |
1166 | READ_ONCE(req->iopoll_completed)) | |
1167 | break; | |
1168 | } | |
1169 | ||
1170 | if (!rq_list_empty(iob.req_list)) | |
1171 | iob.complete(&iob); | |
1172 | else if (!pos) | |
1173 | return 0; | |
1174 | ||
1175 | prev = start; | |
1176 | wq_list_for_each_resume(pos, prev) { | |
1177 | struct io_kiocb *req = container_of(pos, struct io_kiocb, comp_list); | |
1178 | ||
1179 | /* order with io_complete_rw_iopoll(), e.g. ->result updates */ | |
1180 | if (!smp_load_acquire(&req->iopoll_completed)) | |
1181 | break; | |
1182 | nr_events++; | |
f3b44f92 | 1183 | req->cqe.flags = io_put_kbuf(req, 0); |
f3b44f92 | 1184 | } |
f3b44f92 JA |
1185 | if (unlikely(!nr_events)) |
1186 | return 0; | |
1187 | ||
f3b44f92 JA |
1188 | pos = start ? start->next : ctx->iopoll_list.first; |
1189 | wq_list_cut(&ctx->iopoll_list, prev, start); | |
ec26c225 PB |
1190 | |
1191 | if (WARN_ON_ONCE(!wq_list_empty(&ctx->submit_state.compl_reqs))) | |
1192 | return 0; | |
1193 | ctx->submit_state.compl_reqs.first = pos; | |
1194 | __io_submit_flush_completions(ctx); | |
f3b44f92 JA |
1195 | return nr_events; |
1196 | } |