]> git.ipfire.org Git - thirdparty/linux.git/blame - io_uring/io_uring.c
btrfs: remove refs_to_drop argument from __btrfs_free_extent()
[thirdparty/linux.git] / io_uring / io_uring.c
CommitLineData
2b188cc1
JA
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
5 *
6 * A note on the read/write ordering memory barriers that are matched between
1e84b97b
SB
7 * the application and kernel side.
8 *
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
d068b506 14 * through a control-dependency in io_get_cqe (smp_store_release to
1e84b97b
SB
15 * store head will do). Failure to do so could lead to reading invalid
16 * CQ entries.
17 *
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
23 * head will do).
24 *
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
28 * between.
2b188cc1
JA
29 *
30 * Also see the examples in the liburing library:
31 *
32 * git://git.kernel.dk/liburing
33 *
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
38 *
39 * Copyright (C) 2018-2019 Jens Axboe
c992fe29 40 * Copyright (c) 2018-2019 Christoph Hellwig
2b188cc1
JA
41 */
42#include <linux/kernel.h>
43#include <linux/init.h>
44#include <linux/errno.h>
45#include <linux/syscalls.h>
52de1fe1 46#include <net/compat.h>
2b188cc1
JA
47#include <linux/refcount.h>
48#include <linux/uio.h>
6b47ee6e 49#include <linux/bits.h>
2b188cc1
JA
50
51#include <linux/sched/signal.h>
52#include <linux/fs.h>
53#include <linux/file.h>
54#include <linux/fdtable.h>
55#include <linux/mm.h>
56#include <linux/mman.h>
2b188cc1
JA
57#include <linux/percpu.h>
58#include <linux/slab.h>
edafccee 59#include <linux/bvec.h>
2b188cc1
JA
60#include <linux/net.h>
61#include <net/sock.h>
62#include <net/af_unix.h>
6b06314c 63#include <net/scm.h>
2b188cc1
JA
64#include <linux/anon_inodes.h>
65#include <linux/sched/mm.h>
66#include <linux/uaccess.h>
67#include <linux/nospec.h>
aa4c3967 68#include <linux/highmem.h>
15b71abe 69#include <linux/fsnotify.h>
4840e418 70#include <linux/fadvise.h>
b41e9852 71#include <linux/task_work.h>
0f212204 72#include <linux/io_uring.h>
5bd2182d 73#include <linux/audit.h>
cdc1404a 74#include <linux/security.h>
d808459b 75#include <asm/shmparam.h>
2b188cc1 76
c826bd7a
DD
77#define CREATE_TRACE_POINTS
78#include <trace/events/io_uring.h>
79
2b188cc1
JA
80#include <uapi/linux/io_uring.h>
81
561fb04a 82#include "io-wq.h"
2b188cc1 83
de23077e 84#include "io_uring.h"
329061d3 85#include "opdef.h"
e418bbc9 86#include "refs.h"
c9f06aa7 87#include "tctx.h"
17437f31 88#include "sqpoll.h"
a4ad4f74 89#include "fdinfo.h"
3b77495a 90#include "kbuf.h"
73572984 91#include "rsrc.h"
38513c46 92#include "cancel.h"
43e0bbbd 93#include "net.h"
eb42cebb 94#include "notif.h"
e27f928e 95
59915143 96#include "timeout.h"
329061d3 97#include "poll.h"
c92fcfc2 98#include "rw.h"
9b797a37 99#include "alloc_cache.h"
5e2a18d9 100
5277deaa 101#define IORING_MAX_ENTRIES 32768
33a107f0 102#define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
65e19f54 103
21b55dbc
SG
104#define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
105 IORING_REGISTER_LAST + IORING_OP_LAST)
2b188cc1 106
68fe256a
PB
107#define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
108 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
109
5562a8d7
PB
110#define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
111 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
68fe256a 112
c854357b 113#define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
9cae36a0
JA
114 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
115 REQ_F_ASYNC_DATA)
b16fed66 116
a538be5b
PB
117#define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
118 IO_REQ_CLEAN_FLAGS)
119
09899b19
PB
120#define IO_TCTX_REFS_CACHE_NR (1U << 10)
121
6dd0be1e 122#define IO_COMPL_BATCH 32
bf019da7 123#define IO_REQ_ALLOC_BATCH 8
258b29a9 124
10988a0a
DY
125enum {
126 IO_CHECK_CQ_OVERFLOW_BIT,
155bc950 127 IO_CHECK_CQ_DROPPED_BIT,
10988a0a
DY
128};
129
21a091b9
DY
130enum {
131 IO_EVENTFD_OP_SIGNAL_BIT,
132 IO_EVENTFD_OP_FREE_BIT,
133};
134
27dc8338
PB
135struct io_defer_entry {
136 struct list_head list;
137 struct io_kiocb *req;
9cf7c104 138 u32 seq;
2b188cc1
JA
139};
140
0756a869
PB
141/* requests with any of those set should undergo io_disarm_next() */
142#define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
da1a08c5 143#define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
0756a869 144
affa87db 145static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9936c7c2 146 struct task_struct *task,
3dd0c97a 147 bool cancel_all);
1ffc5422 148
cbc2e203 149static void io_queue_sqe(struct io_kiocb *req);
de0617e4 150
c1755c25 151struct kmem_cache *req_cachep;
2b188cc1 152
76d3ccec
MR
153static int __read_mostly sysctl_io_uring_disabled;
154static int __read_mostly sysctl_io_uring_group = -1;
155
156#ifdef CONFIG_SYSCTL
157static struct ctl_table kernel_io_uring_disabled_table[] = {
158 {
159 .procname = "io_uring_disabled",
160 .data = &sysctl_io_uring_disabled,
161 .maxlen = sizeof(sysctl_io_uring_disabled),
162 .mode = 0644,
163 .proc_handler = proc_dointvec_minmax,
164 .extra1 = SYSCTL_ZERO,
165 .extra2 = SYSCTL_TWO,
166 },
167 {
168 .procname = "io_uring_group",
169 .data = &sysctl_io_uring_group,
170 .maxlen = sizeof(gid_t),
171 .mode = 0644,
172 .proc_handler = proc_dointvec,
173 },
174 {},
175};
176#endif
177
2b188cc1
JA
178struct sock *io_uring_get_socket(struct file *file)
179{
180#if defined(CONFIG_UNIX)
cd40cae2 181 if (io_is_uring_fops(file)) {
2b188cc1
JA
182 struct io_ring_ctx *ctx = file->private_data;
183
184 return ctx->ring_sock->sk;
185 }
186#endif
187 return NULL;
188}
189EXPORT_SYMBOL(io_uring_get_socket);
190
c450178d
PB
191static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
192{
931147dd
DY
193 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
194 ctx->submit_state.cqes_count)
c450178d
PB
195 __io_submit_flush_completions(ctx);
196}
197
faf88dde
PB
198static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
199{
200 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
201}
202
0fc8c2ac
DY
203static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
204{
205 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
206}
207
9cae36a0
JA
208static bool io_match_linked(struct io_kiocb *head)
209{
210 struct io_kiocb *req;
211
212 io_for_each_link(req, head) {
213 if (req->flags & REQ_F_INFLIGHT)
214 return true;
215 }
216 return false;
6af3f48b
PB
217}
218
219/*
220 * As io_match_task() but protected against racing with linked timeouts.
221 * User must not hold timeout_lock.
222 */
329061d3
JA
223bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
224 bool cancel_all)
6af3f48b 225{
9cae36a0
JA
226 bool matched;
227
6af3f48b
PB
228 if (task && head->task != task)
229 return false;
9cae36a0
JA
230 if (cancel_all)
231 return true;
232
233 if (head->flags & REQ_F_LINK_TIMEOUT) {
234 struct io_ring_ctx *ctx = head->ctx;
235
236 /* protect against races with linked timeouts */
237 spin_lock_irq(&ctx->timeout_lock);
238 matched = io_match_linked(head);
239 spin_unlock_irq(&ctx->timeout_lock);
240 } else {
241 matched = io_match_linked(head);
242 }
243 return matched;
6af3f48b
PB
244}
245
a8295b98
HX
246static inline void req_fail_link_node(struct io_kiocb *req, int res)
247{
248 req_set_fail(req);
97b388d7 249 io_req_set_res(req, res, 0);
a8295b98
HX
250}
251
fa05457a
PB
252static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
253{
254 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
a8295b98
HX
255}
256
c072481d 257static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
2b188cc1
JA
258{
259 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
260
0f158b4c 261 complete(&ctx->ref_comp);
2b188cc1
JA
262}
263
c072481d 264static __cold void io_fallback_req_func(struct work_struct *work)
f56165e6
PB
265{
266 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
267 fallback_work.work);
268 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
269 struct io_kiocb *req, *tmp;
a282967c 270 struct io_tw_state ts = { .locked = true, };
f56165e6 271
31f084b7 272 mutex_lock(&ctx->uring_lock);
3218e5d3 273 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
a282967c
PB
274 req->io_task_work.func(req, &ts);
275 if (WARN_ON_ONCE(!ts.locked))
31f084b7
PB
276 return;
277 io_submit_flush_completions(ctx);
278 mutex_unlock(&ctx->uring_lock);
f56165e6
PB
279}
280
e6f89be6
PB
281static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
282{
283 unsigned hash_buckets = 1U << bits;
284 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
285
286 table->hbs = kmalloc(hash_size, GFP_KERNEL);
287 if (!table->hbs)
288 return -ENOMEM;
289
290 table->hash_bits = bits;
291 init_hash_table(table, hash_buckets);
292 return 0;
293}
294
c072481d 295static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
2b188cc1
JA
296{
297 struct io_ring_ctx *ctx;
9cfc7e94 298 int hash_bits;
2b188cc1
JA
299
300 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
301 if (!ctx)
302 return NULL;
303
9cfc7e94
JA
304 xa_init(&ctx->io_bl_xa);
305
78076bb6
JA
306 /*
307 * Use 5 bits less than the max cq entries, that should give us around
4a07723f
PB
308 * 32 entries per hash list if totally full and uniformly spread, but
309 * don't keep too many buckets to not overconsume memory.
78076bb6 310 */
4a07723f
PB
311 hash_bits = ilog2(p->cq_entries) - 5;
312 hash_bits = clamp(hash_bits, 1, 8);
e6f89be6 313 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
78076bb6 314 goto err;
9ca9fb24
PB
315 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
316 goto err;
21482896 317 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
48904229 318 0, GFP_KERNEL))
206aefde 319 goto err;
2b188cc1
JA
320
321 ctx->flags = p->flags;
90554200 322 init_waitqueue_head(&ctx->sqo_sq_wait);
69fb2131 323 INIT_LIST_HEAD(&ctx->sqd_list);
1d7bb1d5 324 INIT_LIST_HEAD(&ctx->cq_overflow_list);
cc3cec83 325 INIT_LIST_HEAD(&ctx->io_buffers_cache);
69bbc6ad
PB
326 io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX,
327 sizeof(struct io_rsrc_node));
328 io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX,
329 sizeof(struct async_poll));
330 io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX,
331 sizeof(struct io_async_msghdr));
0f158b4c 332 init_completion(&ctx->ref_comp);
61cf9370 333 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
2b188cc1 334 mutex_init(&ctx->uring_lock);
311997b3 335 init_waitqueue_head(&ctx->cq_wait);
7b235dd8 336 init_waitqueue_head(&ctx->poll_wq);
4ea15b56 337 init_waitqueue_head(&ctx->rsrc_quiesce_wq);
2b188cc1 338 spin_lock_init(&ctx->completion_lock);
89850fce 339 spin_lock_init(&ctx->timeout_lock);
5eef4e87 340 INIT_WQ_LIST(&ctx->iopoll_list);
cc3cec83
JA
341 INIT_LIST_HEAD(&ctx->io_buffers_pages);
342 INIT_LIST_HEAD(&ctx->io_buffers_comp);
de0617e4 343 INIT_LIST_HEAD(&ctx->defer_list);
5262f567 344 INIT_LIST_HEAD(&ctx->timeout_list);
ef9dd637 345 INIT_LIST_HEAD(&ctx->ltimeout_list);
d67d2263 346 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
c0e0d6ba 347 init_llist_head(&ctx->work_llist);
13bf43f5 348 INIT_LIST_HEAD(&ctx->tctx_list);
c2b6c6bc
PB
349 ctx->submit_state.free_list.next = NULL;
350 INIT_WQ_LIST(&ctx->locked_free_list);
9011bf9a 351 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
6f33b0bc 352 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
2b188cc1 353 return ctx;
206aefde 354err:
e6f89be6 355 kfree(ctx->cancel_table.hbs);
9ca9fb24 356 kfree(ctx->cancel_table_locked.hbs);
9cfc7e94
JA
357 kfree(ctx->io_bl);
358 xa_destroy(&ctx->io_bl_xa);
206aefde
JA
359 kfree(ctx);
360 return NULL;
2b188cc1
JA
361}
362
8f6ed49a
PB
363static void io_account_cq_overflow(struct io_ring_ctx *ctx)
364{
365 struct io_rings *r = ctx->rings;
366
367 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
368 ctx->cq_extra--;
369}
370
9cf7c104 371static bool req_need_defer(struct io_kiocb *req, u32 seq)
7adf4eaf 372{
2bc9930e
JA
373 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
374 struct io_ring_ctx *ctx = req->ctx;
a197f664 375
8f6ed49a 376 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
2bc9930e 377 }
de0617e4 378
9d858b21 379 return false;
de0617e4
JA
380}
381
5a754dea
PB
382static void io_clean_op(struct io_kiocb *req)
383{
384 if (req->flags & REQ_F_BUFFER_SELECTED) {
385 spin_lock(&req->ctx->completion_lock);
386 io_put_kbuf_comp(req);
387 spin_unlock(&req->ctx->completion_lock);
388 }
389
390 if (req->flags & REQ_F_NEED_CLEANUP) {
391 const struct io_cold_def *def = &io_cold_defs[req->opcode];
392
393 if (def->cleanup)
394 def->cleanup(req);
395 }
396 if ((req->flags & REQ_F_POLLED) && req->apoll) {
397 kfree(req->apoll->double_poll);
398 kfree(req->apoll);
399 req->apoll = NULL;
400 }
401 if (req->flags & REQ_F_INFLIGHT) {
402 struct io_uring_task *tctx = req->task->io_uring;
403
404 atomic_dec(&tctx->inflight_tracked);
405 }
406 if (req->flags & REQ_F_CREDS)
407 put_cred(req->creds);
408 if (req->flags & REQ_F_ASYNC_DATA) {
409 kfree(req->async_data);
410 req->async_data = NULL;
411 }
412 req->flags &= ~IO_REQ_CLEAN_FLAGS;
413}
414
9cae36a0
JA
415static inline void io_req_track_inflight(struct io_kiocb *req)
416{
417 if (!(req->flags & REQ_F_INFLIGHT)) {
418 req->flags |= REQ_F_INFLIGHT;
386e4fb6 419 atomic_inc(&req->task->io_uring->inflight_tracked);
9cae36a0
JA
420 }
421}
422
fd08e530
PB
423static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
424{
906c6caa
PB
425 if (WARN_ON_ONCE(!req->link))
426 return NULL;
427
4d13d1a4
PB
428 req->flags &= ~REQ_F_ARM_LTIMEOUT;
429 req->flags |= REQ_F_LINK_TIMEOUT;
fd08e530
PB
430
431 /* linked timeouts should have two refs once prep'ed */
48dcd38d 432 io_req_set_refcount(req);
4d13d1a4
PB
433 __io_req_set_refcount(req->link, 2);
434 return req->link;
fd08e530
PB
435}
436
437static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
438{
4d13d1a4 439 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
fd08e530
PB
440 return NULL;
441 return __io_prep_linked_timeout(req);
442}
443
cb2d344c
PB
444static noinline void __io_arm_ltimeout(struct io_kiocb *req)
445{
446 io_queue_linked_timeout(__io_prep_linked_timeout(req));
447}
448
449static inline void io_arm_ltimeout(struct io_kiocb *req)
450{
451 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
452 __io_arm_ltimeout(req);
453}
454
1e6fa521
JA
455static void io_prep_async_work(struct io_kiocb *req)
456{
a7dd2782 457 const struct io_issue_def *def = &io_issue_defs[req->opcode];
1e6fa521
JA
458 struct io_ring_ctx *ctx = req->ctx;
459
b8e64b53
PB
460 if (!(req->flags & REQ_F_CREDS)) {
461 req->flags |= REQ_F_CREDS;
c10d1f98 462 req->creds = get_current_cred();
b8e64b53 463 }
003e8dcc 464
e1d675df
PB
465 req->work.list.next = NULL;
466 req->work.flags = 0;
8e29da69 467 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
feaadc4f
PB
468 if (req->flags & REQ_F_FORCE_ASYNC)
469 req->work.flags |= IO_WQ_WORK_CONCURRENT;
470
3beed235 471 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
8487f083 472 req->flags |= io_file_get_flags(req->file);
f6b543fd 473
8b1df11f 474 if (req->file && (req->flags & REQ_F_ISREG)) {
d4755e15
JA
475 bool should_hash = def->hash_reg_file;
476
477 /* don't serialize this request if the fs doesn't need it */
478 if (should_hash && (req->file->f_flags & O_DIRECT) &&
479 (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE))
480 should_hash = false;
481 if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL))
1e6fa521 482 io_wq_hash_work(&req->work, file_inode(req->file));
4b982bd0 483 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1e6fa521
JA
484 if (def->unbound_nonreg_file)
485 req->work.flags |= IO_WQ_WORK_UNBOUND;
486 }
561fb04a 487}
cccf0ee8 488
cbdcb435 489static void io_prep_async_link(struct io_kiocb *req)
561fb04a 490{
cbdcb435 491 struct io_kiocb *cur;
54a91f3b 492
44eff40a
PB
493 if (req->flags & REQ_F_LINK_TIMEOUT) {
494 struct io_ring_ctx *ctx = req->ctx;
495
674ee8e1 496 spin_lock_irq(&ctx->timeout_lock);
44eff40a
PB
497 io_for_each_link(cur, req)
498 io_prep_async_work(cur);
674ee8e1 499 spin_unlock_irq(&ctx->timeout_lock);
44eff40a
PB
500 } else {
501 io_for_each_link(cur, req)
502 io_prep_async_work(cur);
503 }
561fb04a
JA
504}
505
a282967c 506void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use)
561fb04a 507{
cbdcb435 508 struct io_kiocb *link = io_prep_linked_timeout(req);
5aa75ed5 509 struct io_uring_task *tctx = req->task->io_uring;
561fb04a 510
3bfe6106
JA
511 BUG_ON(!tctx);
512 BUG_ON(!tctx->io_wq);
561fb04a 513
cbdcb435
PB
514 /* init ->work of the whole link before punting */
515 io_prep_async_link(req);
991468dc
JA
516
517 /*
518 * Not expected to happen, but if we do have a bug where this _can_
519 * happen, catch it here and ensure the request is marked as
520 * canceled. That will make io-wq go through the usual work cancel
521 * procedure rather than attempt to run this request (or create a new
522 * worker for it).
523 */
524 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
525 req->work.flags |= IO_WQ_WORK_CANCEL;
526
48863ffd 527 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
ebf93667 528 io_wq_enqueue(tctx->io_wq, &req->work);
7271ef3a
JA
529 if (link)
530 io_queue_linked_timeout(link);
cbdcb435
PB
531}
532
c072481d 533static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
de0617e4 534{
441b8a78 535 while (!list_empty(&ctx->defer_list)) {
27dc8338
PB
536 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
537 struct io_defer_entry, list);
de0617e4 538
9cf7c104 539 if (req_need_defer(de->req, de->seq))
04518945 540 break;
27dc8338 541 list_del_init(&de->list);
907d1df3 542 io_req_task_queue(de->req);
27dc8338 543 kfree(de);
441b8a78 544 }
04518945
PB
545}
546
21a091b9
DY
547
548static void io_eventfd_ops(struct rcu_head *rcu)
f2842ab5 549{
d8e9214f 550 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
21a091b9 551 int ops = atomic_xchg(&ev_fd->ops, 0);
305bef98 552
21a091b9 553 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
44648532 554 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
d8e9214f 555
21a091b9
DY
556 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
557 * ordering in a race but if references are 0 we know we have to free
558 * it regardless.
305bef98 559 */
21a091b9
DY
560 if (atomic_dec_and_test(&ev_fd->refs)) {
561 eventfd_ctx_put(ev_fd->cq_ev_fd);
562 kfree(ev_fd);
563 }
d8e9214f
DY
564}
565
77bc59b4 566static void io_eventfd_signal(struct io_ring_ctx *ctx)
f2842ab5 567{
21a091b9 568 struct io_ev_fd *ev_fd = NULL;
77bc59b4 569
77bc59b4
UA
570 rcu_read_lock();
571 /*
572 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
573 * and eventfd_signal
574 */
575 ev_fd = rcu_dereference(ctx->io_ev_fd);
576
577 /*
578 * Check again if ev_fd exists incase an io_eventfd_unregister call
579 * completed between the NULL check of ctx->io_ev_fd at the start of
580 * the function and rcu_read_lock.
581 */
582 if (unlikely(!ev_fd))
583 goto out;
7e55a19c 584 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
77bc59b4 585 goto out;
21a091b9
DY
586 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
587 goto out;
77bc59b4 588
21a091b9 589 if (likely(eventfd_signal_allowed())) {
44648532 590 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
21a091b9
DY
591 } else {
592 atomic_inc(&ev_fd->refs);
593 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
44a84da4 594 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
21a091b9
DY
595 else
596 atomic_dec(&ev_fd->refs);
597 }
598
77bc59b4
UA
599out:
600 rcu_read_unlock();
f2842ab5
JA
601}
602
21a091b9
DY
603static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
604{
605 bool skip;
606
607 spin_lock(&ctx->completion_lock);
608
609 /*
610 * Eventfd should only get triggered when at least one event has been
611 * posted. Some applications rely on the eventfd notification count
612 * only changing IFF a new CQE has been added to the CQ ring. There's
613 * no depedency on 1:1 relationship between how many times this
614 * function is called (and hence the eventfd count) and number of CQEs
615 * posted to the CQ ring.
616 */
617 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
618 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
619 spin_unlock(&ctx->completion_lock);
620 if (skip)
621 return;
622
623 io_eventfd_signal(ctx);
624}
625
a830ffd2
PB
626void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
627{
bca39f39
PB
628 if (ctx->poll_activated)
629 io_poll_wq_wake(ctx);
e5f30f6f
PB
630 if (ctx->off_timeout_used)
631 io_flush_timeouts(ctx);
632 if (ctx->drain_active) {
a830ffd2 633 spin_lock(&ctx->completion_lock);
e5f30f6f 634 io_queue_deferred(ctx);
a830ffd2
PB
635 spin_unlock(&ctx->completion_lock);
636 }
637 if (ctx->has_evfd)
21a091b9 638 io_eventfd_flush_signal(ctx);
a830ffd2
PB
639}
640
f66f7342 641static inline void __io_cq_lock(struct io_ring_ctx *ctx)
f66f7342 642{
ec26c225 643 if (!ctx->lockless_cq)
f66f7342
PB
644 spin_lock(&ctx->completion_lock);
645}
646
6971253f
PB
647static inline void io_cq_lock(struct io_ring_ctx *ctx)
648 __acquires(ctx->completion_lock)
649{
650 spin_lock(&ctx->completion_lock);
651}
652
f66f7342 653static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
3181e22f
PB
654{
655 io_commit_cqring(ctx);
54927baf 656 if (!ctx->task_complete) {
ec26c225
PB
657 if (!ctx->lockless_cq)
658 spin_unlock(&ctx->completion_lock);
659 /* IOPOLL rings only need to wake up if it's also SQPOLL */
660 if (!ctx->syscall_iopoll)
661 io_cqring_wake(ctx);
3181e22f 662 }
54927baf 663 io_commit_cqring_flush(ctx);
3181e22f
PB
664}
665
0fdb9a19 666static void io_cq_unlock_post(struct io_ring_ctx *ctx)
5d772916 667 __releases(ctx->completion_lock)
25399321 668{
f66f7342
PB
669 io_commit_cqring(ctx);
670 spin_unlock(&ctx->completion_lock);
f66f7342 671 io_cqring_wake(ctx);
54927baf 672 io_commit_cqring_flush(ctx);
25399321
PB
673}
674
c4a2ed72 675/* Returns true if there are no backlogged entries after the flush */
a85381d8
PB
676static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
677{
678 struct io_overflow_cqe *ocqe;
679 LIST_HEAD(list);
680
f432b76b 681 spin_lock(&ctx->completion_lock);
a85381d8
PB
682 list_splice_init(&ctx->cq_overflow_list, &list);
683 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
f432b76b 684 spin_unlock(&ctx->completion_lock);
a85381d8
PB
685
686 while (!list_empty(&list)) {
687 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
688 list_del(&ocqe->list);
689 kfree(ocqe);
690 }
691}
692
1b346e4a 693static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1d7bb1d5 694{
e45a3e05 695 size_t cqe_size = sizeof(struct io_uring_cqe);
1d7bb1d5 696
a85381d8 697 if (__io_cqring_events(ctx) == ctx->cq_entries)
1b346e4a 698 return;
1d7bb1d5 699
e45a3e05
SR
700 if (ctx->flags & IORING_SETUP_CQE32)
701 cqe_size <<= 1;
702
25399321 703 io_cq_lock(ctx);
6c2450ae 704 while (!list_empty(&ctx->cq_overflow_list)) {
59fbc409 705 struct io_uring_cqe *cqe;
6c2450ae 706 struct io_overflow_cqe *ocqe;
e6c8aa9a 707
59fbc409 708 if (!io_get_cqe_overflow(ctx, &cqe, true))
1d7bb1d5 709 break;
6c2450ae
PB
710 ocqe = list_first_entry(&ctx->cq_overflow_list,
711 struct io_overflow_cqe, list);
a85381d8 712 memcpy(cqe, &ocqe->cqe, cqe_size);
6c2450ae
PB
713 list_del(&ocqe->list);
714 kfree(ocqe);
1d7bb1d5
JA
715 }
716
1b346e4a 717 if (list_empty(&ctx->cq_overflow_list)) {
10988a0a 718 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
3a4b89a2 719 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
09e88404 720 }
25399321 721 io_cq_unlock_post(ctx);
1d7bb1d5
JA
722}
723
52ea806a
JA
724static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
725{
726 /* iopoll syncs against uring_lock, not completion_lock */
727 if (ctx->flags & IORING_SETUP_IOPOLL)
728 mutex_lock(&ctx->uring_lock);
729 __io_cqring_overflow_flush(ctx);
730 if (ctx->flags & IORING_SETUP_IOPOLL)
731 mutex_unlock(&ctx->uring_lock);
732}
733
1b346e4a 734static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
6c503150 735{
52ea806a
JA
736 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
737 io_cqring_do_overflow_flush(ctx);
6c503150
PB
738}
739
5afa4650 740/* can be called by any task */
2fdd6fb5 741static void io_put_task_remote(struct task_struct *task)
6a290a14
PB
742{
743 struct io_uring_task *tctx = task->io_uring;
744
2fdd6fb5 745 percpu_counter_sub(&tctx->inflight, 1);
8d664282 746 if (unlikely(atomic_read(&tctx->in_cancel)))
9d170164 747 wake_up(&tctx->wait);
2fdd6fb5 748 put_task_struct(task);
9d170164
PB
749}
750
5afa4650 751/* used by a task to put its own references */
2fdd6fb5 752static void io_put_task_local(struct task_struct *task)
5afa4650 753{
2fdd6fb5 754 task->io_uring->cached_refs++;
5afa4650
PB
755}
756
89800a2d 757/* must to be called somewhat shortly after putting a request */
2fdd6fb5 758static inline void io_put_task(struct task_struct *task)
89800a2d
PB
759{
760 if (likely(task == current))
2fdd6fb5 761 io_put_task_local(task);
89800a2d 762 else
2fdd6fb5 763 io_put_task_remote(task);
89800a2d
PB
764}
765
63809137 766void io_task_refs_refill(struct io_uring_task *tctx)
9a10867a
PB
767{
768 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
769
770 percpu_counter_add(&tctx->inflight, refill);
771 refcount_add(refill, &current->usage);
772 tctx->cached_refs += refill;
773}
774
3cc7fdb9
PB
775static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
776{
777 struct io_uring_task *tctx = task->io_uring;
778 unsigned int refs = tctx->cached_refs;
779
780 if (refs) {
781 tctx->cached_refs = 0;
782 percpu_counter_sub(&tctx->inflight, refs);
783 put_task_struct_many(task, refs);
784 }
785}
786
68494a65
PB
787static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
788 s32 res, u32 cflags, u64 extra1, u64 extra2)
2b188cc1 789{
cce4b8b0 790 struct io_overflow_cqe *ocqe;
e45a3e05
SR
791 size_t ocq_size = sizeof(struct io_overflow_cqe);
792 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
2b188cc1 793
f26cc959
PB
794 lockdep_assert_held(&ctx->completion_lock);
795
e45a3e05
SR
796 if (is_cqe32)
797 ocq_size += sizeof(struct io_uring_cqe);
2b188cc1 798
e45a3e05 799 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
08dcd028 800 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
cce4b8b0
PB
801 if (!ocqe) {
802 /*
803 * If we're in ring overflow flush mode, or in task cancel mode,
804 * or cannot allocate an overflow entry, then we need to drop it
805 * on the floor.
806 */
8f6ed49a 807 io_account_cq_overflow(ctx);
155bc950 808 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
cce4b8b0 809 return false;
2b188cc1 810 }
cce4b8b0 811 if (list_empty(&ctx->cq_overflow_list)) {
10988a0a 812 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
3a4b89a2 813 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
20c0b380 814
cce4b8b0 815 }
d4d19c19 816 ocqe->cqe.user_data = user_data;
cce4b8b0
PB
817 ocqe->cqe.res = res;
818 ocqe->cqe.flags = cflags;
e45a3e05
SR
819 if (is_cqe32) {
820 ocqe->cqe.big_cqe[0] = extra1;
821 ocqe->cqe.big_cqe[1] = extra2;
822 }
cce4b8b0
PB
823 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
824 return true;
2b188cc1
JA
825}
826
056695bf 827void io_req_cqe_overflow(struct io_kiocb *req)
68494a65 828{
056695bf
PB
829 io_cqring_event_overflow(req->ctx, req->cqe.user_data,
830 req->cqe.res, req->cqe.flags,
b24c5d75
PB
831 req->big_cqe.extra1, req->big_cqe.extra2);
832 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
68494a65
PB
833}
834
faf88dde
PB
835/*
836 * writes to the cq entry need to come after reading head; the
837 * control dependency is enough as we're using WRITE_ONCE to
838 * fill the cq entry
839 */
20d6b633 840bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow)
faf88dde
PB
841{
842 struct io_rings *rings = ctx->rings;
843 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
faf88dde
PB
844 unsigned int free, queued, len;
845
aa1df3a3
PB
846 /*
847 * Posting into the CQ when there are pending overflowed CQEs may break
848 * ordering guarantees, which will affect links, F_MORE users and more.
849 * Force overflow the completion.
850 */
851 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
20d6b633 852 return false;
faf88dde
PB
853
854 /* userspace may cheat modifying the tail, be safe and do min */
855 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
856 free = ctx->cq_entries - queued;
857 /* we need a contiguous range, limit based on the current array offset */
858 len = min(free, ctx->cq_entries - off);
859 if (!len)
20d6b633 860 return false;
faf88dde 861
b3659a65
PB
862 if (ctx->flags & IORING_SETUP_CQE32) {
863 off <<= 1;
864 len <<= 1;
865 }
866
faf88dde
PB
867 ctx->cqe_cached = &rings->cqes[off];
868 ctx->cqe_sentinel = ctx->cqe_cached + len;
20d6b633 869 return true;
faf88dde
PB
870}
871
f66f7342
PB
872static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
873 u32 cflags)
bcda7baa 874{
cd94903d
PB
875 struct io_uring_cqe *cqe;
876
913a571a 877 ctx->cq_extra++;
cd94903d
PB
878
879 /*
880 * If we can't get a cq entry, userspace overflowed the
881 * submission (by quite a lot). Increment the overflow count in
882 * the ring.
883 */
59fbc409 884 if (likely(io_get_cqe(ctx, &cqe))) {
e0486f3f
DY
885 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
886
cd94903d
PB
887 WRITE_ONCE(cqe->user_data, user_data);
888 WRITE_ONCE(cqe->res, res);
889 WRITE_ONCE(cqe->flags, cflags);
c5595975
PB
890
891 if (ctx->flags & IORING_SETUP_CQE32) {
892 WRITE_ONCE(cqe->big_cqe[0], 0);
893 WRITE_ONCE(cqe->big_cqe[1], 0);
894 }
cd94903d
PB
895 return true;
896 }
52120f0f 897 return false;
bcda7baa
JA
898}
899
931147dd
DY
900static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
901 __must_hold(&ctx->uring_lock)
902{
903 struct io_submit_state *state = &ctx->submit_state;
904 unsigned int i;
905
906 lockdep_assert_held(&ctx->uring_lock);
907 for (i = 0; i < state->cqes_count; i++) {
0aa7aa5f 908 struct io_uring_cqe *cqe = &ctx->completion_cqes[i];
931147dd 909
f66f7342 910 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
27122c07 911 if (ctx->lockless_cq) {
f66f7342
PB
912 spin_lock(&ctx->completion_lock);
913 io_cqring_event_overflow(ctx, cqe->user_data,
914 cqe->res, cqe->flags, 0, 0);
915 spin_unlock(&ctx->completion_lock);
916 } else {
917 io_cqring_event_overflow(ctx, cqe->user_data,
918 cqe->res, cqe->flags, 0, 0);
919 }
920 }
931147dd
DY
921 }
922 state->cqes_count = 0;
923}
924
b529c96a
DY
925static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
926 bool allow_overflow)
d245bca6
PB
927{
928 bool filled;
929
25399321 930 io_cq_lock(ctx);
f66f7342
PB
931 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
932 if (!filled && allow_overflow)
933 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
934
25399321 935 io_cq_unlock_post(ctx);
d245bca6
PB
936 return filled;
937}
938
b529c96a
DY
939bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
940{
941 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
942}
943
b6b2bb58
PB
944/*
945 * A helper for multishot requests posting additional CQEs.
946 * Should only be used from a task_work including IO_URING_F_MULTISHOT.
947 */
948bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags)
2b188cc1 949{
d86eaed1
JA
950 struct io_ring_ctx *ctx = req->ctx;
951 u64 user_data = req->cqe.user_data;
9b8c5475 952 struct io_uring_cqe *cqe;
9b8c5475
DY
953
954 if (!defer)
b6b2bb58 955 return __io_post_aux_cqe(ctx, user_data, res, cflags, false);
9b8c5475 956
9b8c5475
DY
957 lockdep_assert_held(&ctx->uring_lock);
958
0aa7aa5f 959 if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->completion_cqes)) {
f66f7342 960 __io_cq_lock(ctx);
9b8c5475
DY
961 __io_flush_post_cqes(ctx);
962 /* no need to flush - flush is deferred */
f66f7342 963 __io_cq_unlock_post(ctx);
9b8c5475
DY
964 }
965
966 /* For defered completions this is not as strict as it is otherwise,
967 * however it's main job is to prevent unbounded posted completions,
968 * and in that it works just as well.
969 */
b6b2bb58 970 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
9b8c5475
DY
971 return false;
972
0aa7aa5f 973 cqe = &ctx->completion_cqes[ctx->submit_state.cqes_count++];
9b8c5475
DY
974 cqe->user_data = user_data;
975 cqe->res = res;
976 cqe->flags = cflags;
977 return true;
978}
979
ef8ae64f 980static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
2b188cc1 981{
fa18fa22 982 struct io_ring_ctx *ctx = req->ctx;
2ad4c6d0 983 struct io_rsrc_node *rsrc_node = NULL;
fa18fa22
PB
984
985 io_cq_lock(ctx);
00b0db56
PB
986 if (!(req->flags & REQ_F_CQE_SKIP)) {
987 if (!io_fill_cqe_req(ctx, req))
988 io_req_cqe_overflow(req);
989 }
fa18fa22 990
c7dae4ba
JA
991 /*
992 * If we're the last reference to this request, add to our locked
993 * free_list cache.
994 */
de9b4cca 995 if (req_ref_put_and_test(req)) {
da1a08c5 996 if (req->flags & IO_REQ_LINK_FLAGS) {
0756a869 997 if (req->flags & IO_DISARM_MASK)
7a612350
PB
998 io_disarm_next(req);
999 if (req->link) {
1000 io_req_task_queue(req->link);
1001 req->link = NULL;
1002 }
1003 }
68a2cc1b 1004 io_put_kbuf_comp(req);
3b7a612f
PB
1005 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1006 io_clean_op(req);
17bc2837 1007 io_put_file(req);
3b7a612f 1008
2ad4c6d0 1009 rsrc_node = req->rsrc_node;
8197b053
PB
1010 /*
1011 * Selected buffer deallocation in io_clean_op() assumes that
1012 * we don't hold ->completion_lock. Clean them here to avoid
1013 * deadlocks.
1014 */
2fdd6fb5 1015 io_put_task_remote(req->task);
c2b6c6bc 1016 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
d0acdee2 1017 ctx->locked_free_nr++;
180f829f 1018 }
25399321 1019 io_cq_unlock_post(ctx);
2ad4c6d0 1020
ef8ae64f
PB
1021 if (rsrc_node) {
1022 io_ring_submit_lock(ctx, issue_flags);
1f2c8f61 1023 io_put_rsrc_node(ctx, rsrc_node);
ef8ae64f
PB
1024 io_ring_submit_unlock(ctx, issue_flags);
1025 }
4e3d9ff9
JA
1026}
1027
1bec951c 1028void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
bcda7baa 1029{
860e1c7f 1030 if (req->ctx->task_complete && req->ctx->submitter_task != current) {
e6aeb272
PB
1031 req->io_task_work.func = io_req_task_complete;
1032 io_req_task_work_add(req);
1033 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
1034 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
ef8ae64f 1035 __io_req_complete_post(req, issue_flags);
1bec951c
PB
1036 } else {
1037 struct io_ring_ctx *ctx = req->ctx;
1038
1039 mutex_lock(&ctx->uring_lock);
ef8ae64f 1040 __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED);
1bec951c
PB
1041 mutex_unlock(&ctx->uring_lock);
1042 }
0ddf92e8
JA
1043}
1044
973fc83f 1045void io_req_defer_failed(struct io_kiocb *req, s32 res)
e276ae34 1046 __must_hold(&ctx->uring_lock)
f41db273 1047{
f30bd4d0 1048 const struct io_cold_def *def = &io_cold_defs[req->opcode];
a47b255e 1049
e276ae34
PB
1050 lockdep_assert_held(&req->ctx->uring_lock);
1051
93d2bcd2 1052 req_set_fail(req);
97b388d7 1053 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
a47b255e
PB
1054 if (def->fail)
1055 def->fail(req);
973fc83f 1056 io_req_complete_defer(req);
f41db273
PB
1057}
1058
864ea921
PB
1059/*
1060 * Don't initialise the fields below on every allocation, but do that in
1061 * advance and keep them valid across allocations.
1062 */
1063static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
1064{
1065 req->ctx = ctx;
1066 req->link = NULL;
1067 req->async_data = NULL;
1068 /* not necessary, but safer to zero */
31d3ba92 1069 memset(&req->cqe, 0, sizeof(req->cqe));
b24c5d75 1070 memset(&req->big_cqe, 0, sizeof(req->big_cqe));
864ea921
PB
1071}
1072
dac7a098 1073static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
cd0ca2e0 1074 struct io_submit_state *state)
dac7a098 1075{
79ebeaee 1076 spin_lock(&ctx->completion_lock);
c2b6c6bc 1077 wq_list_splice(&ctx->locked_free_list, &state->free_list);
d0acdee2 1078 ctx->locked_free_nr = 0;
79ebeaee 1079 spin_unlock(&ctx->completion_lock);
dac7a098
PB
1080}
1081
5d5901a3
PB
1082/*
1083 * A request might get retired back into the request caches even before opcode
1084 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1085 * Because of that, io_alloc_req() should be called only under ->uring_lock
1086 * and with extra caution to not get a request that is still worked on.
1087 */
bd1a3783 1088__cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
5d5901a3 1089 __must_hold(&ctx->uring_lock)
2b188cc1 1090{
864ea921 1091 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
3ab665b7 1092 void *reqs[IO_REQ_ALLOC_BATCH];
864ea921 1093 int ret, i;
e5d1bc0a 1094
23a5c43b
PB
1095 /*
1096 * If we have more than a batch's worth of requests in our IRQ side
1097 * locked cache, grab the lock and move them over to our submission
1098 * side cache.
1099 */
a6d97a8a 1100 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
23a5c43b 1101 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
88ab95be 1102 if (!io_req_cache_empty(ctx))
23a5c43b
PB
1103 return true;
1104 }
e5d1bc0a 1105
3ab665b7 1106 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
fd6fab2c 1107
864ea921
PB
1108 /*
1109 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1110 * retry single alloc to be on the safe side.
1111 */
1112 if (unlikely(ret <= 0)) {
3ab665b7
PB
1113 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1114 if (!reqs[0])
a33ae9ce 1115 return false;
864ea921 1116 ret = 1;
2b188cc1 1117 }
864ea921 1118
37f0e767 1119 percpu_ref_get_many(&ctx->refs, ret);
3ab665b7 1120 for (i = 0; i < ret; i++) {
23a5c43b 1121 struct io_kiocb *req = reqs[i];
3ab665b7
PB
1122
1123 io_preinit_req(req, ctx);
fa05457a 1124 io_req_add_to_cache(req, ctx);
3ab665b7 1125 }
a33ae9ce
PB
1126 return true;
1127}
1128
03adabe8
PB
1129__cold void io_free_req(struct io_kiocb *req)
1130{
6ec9afc7
PB
1131 /* refs were already put, restore them for io_req_task_complete() */
1132 req->flags &= ~REQ_F_REFCOUNT;
1133 /* we only want to free it, don't post CQEs */
1134 req->flags |= REQ_F_CQE_SKIP;
1135 req->io_task_work.func = io_req_task_complete;
03adabe8
PB
1136 io_req_task_work_add(req);
1137}
1138
d81499bf
PB
1139static void __io_req_find_next_prep(struct io_kiocb *req)
1140{
1141 struct io_ring_ctx *ctx = req->ctx;
d81499bf 1142
6971253f 1143 spin_lock(&ctx->completion_lock);
305bef98 1144 io_disarm_next(req);
6971253f 1145 spin_unlock(&ctx->completion_lock);
d81499bf
PB
1146}
1147
1148static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
c69f8dbe 1149{
33cc89a9 1150 struct io_kiocb *nxt;
944e58bf 1151
9e645e11
JA
1152 /*
1153 * If LINK is set, we have dependent requests in this chain. If we
1154 * didn't fail this request, queue the first one up, moving any other
1155 * dependencies to the next request. In case of failure, fail the rest
1156 * of the chain.
1157 */
d81499bf
PB
1158 if (unlikely(req->flags & IO_DISARM_MASK))
1159 __io_req_find_next_prep(req);
33cc89a9
PB
1160 nxt = req->link;
1161 req->link = NULL;
1162 return nxt;
4d7dd462 1163}
9e645e11 1164
a282967c 1165static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts)
2c32395d
PB
1166{
1167 if (!ctx)
1168 return;
ef060ea9
JA
1169 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1170 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
a282967c 1171 if (ts->locked) {
c450178d 1172 io_submit_flush_completions(ctx);
2c32395d 1173 mutex_unlock(&ctx->uring_lock);
a282967c 1174 ts->locked = false;
2c32395d
PB
1175 }
1176 percpu_ref_put(&ctx->refs);
1177}
1178
c6dd763c 1179static unsigned int handle_tw_list(struct llist_node *node,
a282967c
PB
1180 struct io_ring_ctx **ctx,
1181 struct io_tw_state *ts,
c6dd763c 1182 struct llist_node *last)
9f8d032a 1183{
c6dd763c
DY
1184 unsigned int count = 0;
1185
cb6bf7f2 1186 while (node && node != last) {
f88262e6 1187 struct llist_node *next = node->next;
9f8d032a
HX
1188 struct io_kiocb *req = container_of(node, struct io_kiocb,
1189 io_task_work.node);
1190
34d2bfe7
JA
1191 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1192
9f8d032a 1193 if (req->ctx != *ctx) {
a282967c 1194 ctx_flush_and_put(*ctx, ts);
9f8d032a
HX
1195 *ctx = req->ctx;
1196 /* if not contended, grab and improve batching */
a282967c 1197 ts->locked = mutex_trylock(&(*ctx)->uring_lock);
9f8d032a 1198 percpu_ref_get(&(*ctx)->refs);
13bfa6f1 1199 }
c92fcfc2
JA
1200 INDIRECT_CALL_2(req->io_task_work.func,
1201 io_poll_task_func, io_req_rw_complete,
1202 req, ts);
9f8d032a 1203 node = next;
c6dd763c 1204 count++;
f5868008 1205 if (unlikely(need_resched())) {
a282967c 1206 ctx_flush_and_put(*ctx, ts);
f5868008
JA
1207 *ctx = NULL;
1208 cond_resched();
1209 }
3a0c037b 1210 }
c6dd763c
DY
1211
1212 return count;
9f8d032a
HX
1213}
1214
923d1592
DY
1215/**
1216 * io_llist_xchg - swap all entries in a lock-less list
1217 * @head: the head of lock-less list to delete all entries
1218 * @new: new entry as the head of the list
1219 *
1220 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1221 * The order of entries returned is from the newest to the oldest added one.
1222 */
1223static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1224 struct llist_node *new)
1225{
1226 return xchg(&head->first, new);
1227}
1228
1229/**
1230 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1231 * @head: the head of lock-less list to delete all entries
1232 * @old: expected old value of the first entry of the list
1233 * @new: new entry as the head of the list
1234 *
1235 * perform a cmpxchg on the first entry of the list.
1236 */
1237
1238static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1239 struct llist_node *old,
1240 struct llist_node *new)
1241{
1242 return cmpxchg(&head->first, old, new);
1243}
1244
dfbe5561 1245static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync)
10e1c0d5
JA
1246{
1247 struct llist_node *node = llist_del_all(&tctx->task_list);
dfbe5561 1248 struct io_ring_ctx *last_ctx = NULL;
10e1c0d5
JA
1249 struct io_kiocb *req;
1250
1251 while (node) {
1252 req = container_of(node, struct io_kiocb, io_task_work.node);
1253 node = node->next;
dfbe5561
JA
1254 if (sync && last_ctx != req->ctx) {
1255 if (last_ctx) {
1256 flush_delayed_work(&last_ctx->fallback_work);
1257 percpu_ref_put(&last_ctx->refs);
1258 }
1259 last_ctx = req->ctx;
1260 percpu_ref_get(&last_ctx->refs);
1261 }
10e1c0d5
JA
1262 if (llist_add(&req->io_task_work.node,
1263 &req->ctx->fallback_llist))
1264 schedule_delayed_work(&req->ctx->fallback_work, 1);
1265 }
dfbe5561
JA
1266
1267 if (last_ctx) {
1268 flush_delayed_work(&last_ctx->fallback_work);
1269 percpu_ref_put(&last_ctx->refs);
1270 }
10e1c0d5
JA
1271}
1272
c9f06aa7 1273void tctx_task_work(struct callback_head *cb)
c40f6379 1274{
a282967c 1275 struct io_tw_state ts = {};
ebd0df2e 1276 struct io_ring_ctx *ctx = NULL;
3f18407d
PB
1277 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1278 task_work);
3a0c037b 1279 struct llist_node fake = {};
77e443ab 1280 struct llist_node *node;
cb6bf7f2
PB
1281 unsigned int loops = 0;
1282 unsigned int count = 0;
3a0c037b 1283
77e443ab 1284 if (unlikely(current->flags & PF_EXITING)) {
dfbe5561 1285 io_fallback_tw(tctx, true);
77e443ab
PB
1286 return;
1287 }
3a0c037b 1288
cb6bf7f2 1289 do {
c6dd763c 1290 loops++;
3a0c037b 1291 node = io_llist_xchg(&tctx->task_list, &fake);
a282967c 1292 count += handle_tw_list(node, &ctx, &ts, &fake);
50470fc5
PB
1293
1294 /* skip expensive cmpxchg if there are items in the list */
1295 if (READ_ONCE(tctx->task_list.first) != &fake)
1296 continue;
a282967c 1297 if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) {
50470fc5
PB
1298 io_submit_flush_completions(ctx);
1299 if (READ_ONCE(tctx->task_list.first) != &fake)
1300 continue;
1301 }
3a0c037b 1302 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
cb6bf7f2 1303 } while (node != &fake);
ebd0df2e 1304
a282967c 1305 ctx_flush_and_put(ctx, &ts);
3cc7fdb9 1306
8d664282
JA
1307 /* relaxed read is enough as only the task itself sets ->in_cancel */
1308 if (unlikely(atomic_read(&tctx->in_cancel)))
3cc7fdb9 1309 io_uring_drop_tctx_refs(current);
c6dd763c
DY
1310
1311 trace_io_uring_task_work_run(tctx, count, loops);
7cbf1722
JA
1312}
1313
91c7884a 1314static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
c0e0d6ba
DY
1315{
1316 struct io_ring_ctx *ctx = req->ctx;
8751d154 1317 unsigned nr_wait, nr_tw, nr_tw_prev;
51509400 1318 struct llist_node *first;
c0e0d6ba 1319
8751d154
PB
1320 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
1321 flags &= ~IOU_F_TWQ_LAZY_WAKE;
ce8e04f6 1322
51509400
PB
1323 first = READ_ONCE(ctx->work_llist.first);
1324 do {
8751d154
PB
1325 nr_tw_prev = 0;
1326 if (first) {
1327 struct io_kiocb *first_req = container_of(first,
1328 struct io_kiocb,
1329 io_task_work.node);
1330 /*
1331 * Might be executed at any moment, rely on
1332 * SLAB_TYPESAFE_BY_RCU to keep it alive.
1333 */
1334 nr_tw_prev = READ_ONCE(first_req->nr_tw);
1335 }
1336 nr_tw = nr_tw_prev + 1;
1337 /* Large enough to fail the nr_wait comparison below */
1338 if (!(flags & IOU_F_TWQ_LAZY_WAKE))
1339 nr_tw = -1U;
1340
1341 req->nr_tw = nr_tw;
51509400
PB
1342 req->io_task_work.node.next = first;
1343 } while (!try_cmpxchg(&ctx->work_llist.first, &first,
1344 &req->io_task_work.node));
1345
8751d154 1346 if (!first) {
8751d154
PB
1347 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1348 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1349 if (ctx->has_evfd)
1350 io_eventfd_signal(ctx);
c0e0d6ba
DY
1351 }
1352
8751d154
PB
1353 nr_wait = atomic_read(&ctx->cq_wait_nr);
1354 /* no one is waiting */
1355 if (!nr_wait)
1356 return;
1357 /* either not enough or the previous add has already woken it up */
1358 if (nr_wait > nr_tw || nr_tw_prev >= nr_wait)
1359 return;
1360 /* pairs with set_current_state() in io_cqring_wait() */
1361 smp_mb__after_atomic();
1362 wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
c0e0d6ba
DY
1363}
1364
91c7884a 1365static void io_req_normal_work_add(struct io_kiocb *req)
7cbf1722 1366{
c34398a8 1367 struct io_uring_task *tctx = req->task->io_uring;
9f010507 1368 struct io_ring_ctx *ctx = req->ctx;
7cbf1722 1369
7cbf1722 1370 /* task_work already pending, we're done */
32d91f05 1371 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
e09ee510 1372 return;
7cbf1722 1373
ef060ea9
JA
1374 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1375 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1376
3fe07bcd 1377 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
e09ee510 1378 return;
2215bed9 1379
dfbe5561 1380 io_fallback_tw(tctx, false);
c0e0d6ba
DY
1381}
1382
91c7884a
PB
1383void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
1384{
1385 if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1386 rcu_read_lock();
1387 io_req_local_work_add(req, flags);
1388 rcu_read_unlock();
1389 } else {
1390 io_req_normal_work_add(req);
1391 }
1392}
1393
c0e0d6ba
DY
1394static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1395{
1396 struct llist_node *node;
1397
1398 node = llist_del_all(&ctx->work_llist);
1399 while (node) {
1400 struct io_kiocb *req = container_of(node, struct io_kiocb,
1401 io_task_work.node);
1402
1403 node = node->next;
91c7884a 1404 io_req_normal_work_add(req);
c0e0d6ba
DY
1405 }
1406}
1407
a282967c 1408static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts)
c0e0d6ba 1409{
c0e0d6ba 1410 struct llist_node *node;
c3f4d39e 1411 unsigned int loops = 0;
140102ae 1412 int ret = 0;
c0e0d6ba 1413
140102ae 1414 if (WARN_ON_ONCE(ctx->submitter_task != current))
c0e0d6ba 1415 return -EEXIST;
c3f4d39e
PB
1416 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1417 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
c0e0d6ba 1418again:
3af0356c
JA
1419 /*
1420 * llists are in reverse order, flip it back the right way before
1421 * running the pending items.
1422 */
1423 node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL));
c3f4d39e 1424 while (node) {
c0e0d6ba
DY
1425 struct llist_node *next = node->next;
1426 struct io_kiocb *req = container_of(node, struct io_kiocb,
1427 io_task_work.node);
1428 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
c92fcfc2
JA
1429 INDIRECT_CALL_2(req->io_task_work.func,
1430 io_poll_task_func, io_req_rw_complete,
1431 req, ts);
c0e0d6ba
DY
1432 ret++;
1433 node = next;
1434 }
c3f4d39e 1435 loops++;
c0e0d6ba 1436
c3f4d39e 1437 if (!llist_empty(&ctx->work_llist))
c0e0d6ba 1438 goto again;
a282967c 1439 if (ts->locked) {
c0e0d6ba 1440 io_submit_flush_completions(ctx);
b0b7a7d2
PB
1441 if (!llist_empty(&ctx->work_llist))
1442 goto again;
1443 }
f75d5036 1444 trace_io_uring_local_work_run(ctx, ret, loops);
c0e0d6ba 1445 return ret;
8ac5d85a
JA
1446}
1447
360173ab
PB
1448static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
1449{
a282967c 1450 struct io_tw_state ts = { .locked = true, };
360173ab
PB
1451 int ret;
1452
1453 if (llist_empty(&ctx->work_llist))
1454 return 0;
1455
a282967c 1456 ret = __io_run_local_work(ctx, &ts);
360173ab 1457 /* shouldn't happen! */
a282967c 1458 if (WARN_ON_ONCE(!ts.locked))
360173ab
PB
1459 mutex_lock(&ctx->uring_lock);
1460 return ret;
1461}
1462
3e565555 1463static int io_run_local_work(struct io_ring_ctx *ctx)
8ac5d85a 1464{
a282967c 1465 struct io_tw_state ts = {};
8ac5d85a
JA
1466 int ret;
1467
a282967c
PB
1468 ts.locked = mutex_trylock(&ctx->uring_lock);
1469 ret = __io_run_local_work(ctx, &ts);
1470 if (ts.locked)
8ac5d85a
JA
1471 mutex_unlock(&ctx->uring_lock);
1472
1473 return ret;
c0e0d6ba
DY
1474}
1475
a282967c 1476static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts)
c40f6379 1477{
a282967c 1478 io_tw_lock(req->ctx, ts);
973fc83f 1479 io_req_defer_failed(req, req->cqe.res);
c40f6379
JA
1480}
1481
a282967c 1482void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts)
c40f6379 1483{
a282967c 1484 io_tw_lock(req->ctx, ts);
316319e8 1485 /* req->task == current here, checking PF_EXITING is safe */
6bb30855 1486 if (unlikely(req->task->flags & PF_EXITING))
973fc83f 1487 io_req_defer_failed(req, -EFAULT);
6bb30855 1488 else if (req->flags & REQ_F_FORCE_ASYNC)
a282967c 1489 io_queue_iowq(req, ts);
6bb30855
DY
1490 else
1491 io_queue_sqe(req);
c40f6379
JA
1492}
1493
59915143 1494void io_req_task_queue_fail(struct io_kiocb *req, int ret)
c40f6379 1495{
97b388d7 1496 io_req_set_res(req, ret, 0);
5b0a6acc 1497 req->io_task_work.func = io_req_task_cancel;
3fe07bcd 1498 io_req_task_work_add(req);
c40f6379
JA
1499}
1500
f3b44f92 1501void io_req_task_queue(struct io_kiocb *req)
a3df7698 1502{
5b0a6acc 1503 req->io_task_work.func = io_req_task_submit;
3fe07bcd 1504 io_req_task_work_add(req);
a3df7698
PB
1505}
1506
59915143 1507void io_queue_next(struct io_kiocb *req)
c69f8dbe 1508{
9b5f7bd9 1509 struct io_kiocb *nxt = io_req_find_next(req);
944e58bf
PB
1510
1511 if (nxt)
906a8c3f 1512 io_req_task_queue(nxt);
c69f8dbe
JL
1513}
1514
ec26c225
PB
1515static void io_free_batch_list(struct io_ring_ctx *ctx,
1516 struct io_wq_work_node *node)
3aa83bfb 1517 __must_hold(&ctx->uring_lock)
5af1d13e 1518{
3aa83bfb
PB
1519 do {
1520 struct io_kiocb *req = container_of(node, struct io_kiocb,
1521 comp_list);
2d6500d4 1522
a538be5b
PB
1523 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1524 if (req->flags & REQ_F_REFCOUNT) {
1525 node = req->comp_list.next;
1526 if (!req_ref_put_and_test(req))
1527 continue;
1528 }
b605a7fa
PB
1529 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1530 struct async_poll *apoll = req->apoll;
1531
1532 if (apoll->double_poll)
1533 kfree(apoll->double_poll);
9731bc98
JA
1534 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1535 kfree(apoll);
b605a7fa
PB
1536 req->flags &= ~REQ_F_POLLED;
1537 }
da1a08c5 1538 if (req->flags & IO_REQ_LINK_FLAGS)
57859f4d 1539 io_queue_next(req);
a538be5b
PB
1540 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1541 io_clean_op(req);
c1e53a69 1542 }
17bc2837 1543 io_put_file(req);
2d6500d4 1544
ab409402 1545 io_req_put_rsrc_locked(req, ctx);
5af1d13e 1546
2fdd6fb5 1547 io_put_task(req->task);
c1e53a69 1548 node = req->comp_list.next;
fa05457a 1549 io_req_add_to_cache(req, ctx);
3aa83bfb 1550 } while (node);
7a743e22
PB
1551}
1552
ec26c225 1553void __io_submit_flush_completions(struct io_ring_ctx *ctx)
a141dd89 1554 __must_hold(&ctx->uring_lock)
905c172f 1555{
cd0ca2e0 1556 struct io_submit_state *state = &ctx->submit_state;
fa780334 1557 struct io_wq_work_node *node;
905c172f 1558
f66f7342 1559 __io_cq_lock(ctx);
931147dd
DY
1560 /* must come first to preserve CQE ordering in failure cases */
1561 if (state->cqes_count)
1562 __io_flush_post_cqes(ctx);
fa780334 1563 __wq_list_for_each(node, &state->compl_reqs) {
d9dee430
PB
1564 struct io_kiocb *req = container_of(node, struct io_kiocb,
1565 comp_list);
3d4aeb9f 1566
f66f7342 1567 if (!(req->flags & REQ_F_CQE_SKIP) &&
00b0db56 1568 unlikely(!io_fill_cqe_req(ctx, req))) {
27122c07 1569 if (ctx->lockless_cq) {
f66f7342
PB
1570 spin_lock(&ctx->completion_lock);
1571 io_req_cqe_overflow(req);
1572 spin_unlock(&ctx->completion_lock);
1573 } else {
1574 io_req_cqe_overflow(req);
1575 }
1576 }
905c172f 1577 }
c98c81a4 1578 __io_cq_unlock_post(ctx);
d9dee430 1579
931147dd
DY
1580 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1581 io_free_batch_list(ctx, state->compl_reqs.first);
1582 INIT_WQ_LIST(&state->compl_reqs);
1583 }
7a743e22
PB
1584}
1585
6c503150 1586static unsigned io_cqring_events(struct io_ring_ctx *ctx)
a3a0e43f
JA
1587{
1588 /* See comment at the top of this file */
1589 smp_rmb();
e23de15f 1590 return __io_cqring_events(ctx);
a3a0e43f
JA
1591}
1592
def596e9
JA
1593/*
1594 * We can't just wait for polled events to come to us, we have to actively
1595 * find and complete them.
1596 */
c072481d 1597static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
def596e9
JA
1598{
1599 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1600 return;
1601
1602 mutex_lock(&ctx->uring_lock);
5eef4e87 1603 while (!wq_list_empty(&ctx->iopoll_list)) {
b2edc0a7 1604 /* let it sleep and repeat later if can't complete a request */
5ba3c874 1605 if (io_do_iopoll(ctx, true) == 0)
b2edc0a7 1606 break;
08f5439f
JA
1607 /*
1608 * Ensure we allow local-to-the-cpu processing to take place,
1609 * in this case we need to ensure that we reap all events.
3fcee5a6 1610 * Also let task_work, etc. to progress by releasing the mutex
08f5439f 1611 */
3fcee5a6
PB
1612 if (need_resched()) {
1613 mutex_unlock(&ctx->uring_lock);
1614 cond_resched();
1615 mutex_lock(&ctx->uring_lock);
1616 }
def596e9
JA
1617 }
1618 mutex_unlock(&ctx->uring_lock);
1619}
1620
7668b92a 1621static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
def596e9 1622{
7668b92a 1623 unsigned int nr_events = 0;
155bc950 1624 unsigned long check_cq;
500f9fba 1625
76de6749
PB
1626 if (!io_allowed_run_tw(ctx))
1627 return -EEXIST;
1628
3a08576b
PB
1629 check_cq = READ_ONCE(ctx->check_cq);
1630 if (unlikely(check_cq)) {
1631 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
a85381d8 1632 __io_cqring_overflow_flush(ctx);
3a08576b
PB
1633 /*
1634 * Similarly do not spin if we have not informed the user of any
1635 * dropped CQE.
1636 */
1637 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1638 return -EBADR;
1639 }
f39c8a5b
PB
1640 /*
1641 * Don't enter poll loop if we already have events pending.
1642 * If we do, we can potentially be spinning for commands that
1643 * already triggered a CQE (eg in error).
1644 */
f39c8a5b 1645 if (io_cqring_events(ctx))
d487b43c 1646 return 0;
155bc950 1647
def596e9 1648 do {
9e4bef2b
JA
1649 int ret = 0;
1650
500f9fba
JA
1651 /*
1652 * If a submit got punted to a workqueue, we can have the
1653 * application entering polling for a command before it gets
1654 * issued. That app will hold the uring_lock for the duration
1655 * of the poll right here, so we need to take a breather every
1656 * now and then to ensure that the issue has a chance to add
1657 * the poll to the issued list. Otherwise we can spin here
1658 * forever, while the workqueue is stuck trying to acquire the
1659 * very same mutex.
1660 */
dac6a0ea
JA
1661 if (wq_list_empty(&ctx->iopoll_list) ||
1662 io_task_work_pending(ctx)) {
8f487ef2
PB
1663 u32 tail = ctx->cached_cq_tail;
1664
8de11cdc 1665 (void) io_run_local_work_locked(ctx);
def596e9 1666
dac6a0ea
JA
1667 if (task_work_pending(current) ||
1668 wq_list_empty(&ctx->iopoll_list)) {
dac6a0ea 1669 mutex_unlock(&ctx->uring_lock);
9d54bd6a 1670 io_run_task_work();
dac6a0ea 1671 mutex_lock(&ctx->uring_lock);
dac6a0ea 1672 }
8f487ef2
PB
1673 /* some requests don't go through iopoll_list */
1674 if (tail != ctx->cached_cq_tail ||
5eef4e87 1675 wq_list_empty(&ctx->iopoll_list))
e9979b36 1676 break;
500f9fba 1677 }
5ba3c874 1678 ret = io_do_iopoll(ctx, !min);
9e4bef2b
JA
1679 if (unlikely(ret < 0))
1680 return ret;
dc314886
PB
1681
1682 if (task_sigpending(current))
1683 return -EINTR;
9e4bef2b 1684 if (need_resched())
5ba3c874 1685 break;
d487b43c 1686
5ba3c874 1687 nr_events += ret;
9e4bef2b 1688 } while (nr_events < min);
d487b43c 1689
9e4bef2b 1690 return 0;
def596e9 1691}
7012c815 1692
a282967c 1693void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts)
8ef12efe 1694{
a282967c 1695 if (ts->locked)
9da070b1 1696 io_req_complete_defer(req);
7012c815 1697 else
27f35fe9 1698 io_req_complete_post(req, IO_URING_F_UNLOCKED);
8ef12efe
JA
1699}
1700
def596e9
JA
1701/*
1702 * After the iocb has been issued, it's safe to be found on the poll list.
1703 * Adding the kiocb to the list AFTER submission ensures that we don't
f39c8a5b 1704 * find it from a io_do_iopoll() thread before the issuer is done
def596e9
JA
1705 * accessing the kiocb cookie.
1706 */
9882131c 1707static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
def596e9
JA
1708{
1709 struct io_ring_ctx *ctx = req->ctx;
3b44b371 1710 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
cb3d8972
PB
1711
1712 /* workqueue context doesn't hold uring_lock, grab it now */
3b44b371 1713 if (unlikely(needs_lock))
cb3d8972 1714 mutex_lock(&ctx->uring_lock);
def596e9
JA
1715
1716 /*
1717 * Track whether we have multiple files in our lists. This will impact
1718 * how we do polling eventually, not spinning if we're on potentially
1719 * different devices.
1720 */
5eef4e87 1721 if (wq_list_empty(&ctx->iopoll_list)) {
915b3dde
HX
1722 ctx->poll_multi_queue = false;
1723 } else if (!ctx->poll_multi_queue) {
def596e9
JA
1724 struct io_kiocb *list_req;
1725
5eef4e87
PB
1726 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1727 comp_list);
30da1b45 1728 if (list_req->file != req->file)
915b3dde 1729 ctx->poll_multi_queue = true;
def596e9
JA
1730 }
1731
1732 /*
1733 * For fast devices, IO may have already completed. If it has, add
1734 * it to the front so we find it first.
1735 */
65a6543d 1736 if (READ_ONCE(req->iopoll_completed))
5eef4e87 1737 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
def596e9 1738 else
5eef4e87 1739 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
bdcd3eab 1740
3b44b371 1741 if (unlikely(needs_lock)) {
cb3d8972
PB
1742 /*
1743 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1744 * in sq thread task context or in io worker task context. If
1745 * current task context is sq thread, we don't need to check
1746 * whether should wake up sq thread.
1747 */
1748 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1749 wq_has_sleeper(&ctx->sq_data->wait))
1750 wake_up(&ctx->sq_data->wait);
1751
1752 mutex_unlock(&ctx->uring_lock);
1753 }
def596e9
JA
1754}
1755
a4ad4f74 1756unsigned int io_file_get_flags(struct file *file)
88459b50 1757{
88459b50 1758 unsigned int res = 0;
af197f50 1759
53cfd5ce 1760 if (S_ISREG(file_inode(file)->i_mode))
8487f083 1761 res |= REQ_F_ISREG;
b9a6c945 1762 if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT))
8487f083 1763 res |= REQ_F_SUPPORT_NOWAIT;
88459b50 1764 return res;
2b188cc1
JA
1765}
1766
99f15d8d 1767bool io_alloc_async_data(struct io_kiocb *req)
3d9932a8 1768{
f30bd4d0
BL
1769 WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size);
1770 req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL);
d886e185
PB
1771 if (req->async_data) {
1772 req->flags |= REQ_F_ASYNC_DATA;
1773 return false;
1774 }
1775 return true;
3d9932a8
XW
1776}
1777
f3b44f92 1778int io_req_prep_async(struct io_kiocb *req)
f67676d1 1779{
f30bd4d0 1780 const struct io_cold_def *cdef = &io_cold_defs[req->opcode];
a7dd2782 1781 const struct io_issue_def *def = &io_issue_defs[req->opcode];
0702e536
JA
1782
1783 /* assign early for deferred execution for non-fixed file */
54aa7f23 1784 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file)
0702e536 1785 req->file = io_file_get_normal(req, req->cqe.fd);
f30bd4d0 1786 if (!cdef->prep_async)
0702e536
JA
1787 return 0;
1788 if (WARN_ON_ONCE(req_has_async_data(req)))
1789 return -EFAULT;
f30bd4d0 1790 if (!def->manual_alloc) {
59169439
PB
1791 if (io_alloc_async_data(req))
1792 return -EAGAIN;
1793 }
f30bd4d0 1794 return cdef->prep_async(req);
bfe76559
PB
1795}
1796
9cf7c104
PB
1797static u32 io_get_sequence(struct io_kiocb *req)
1798{
a3dbdf54 1799 u32 seq = req->ctx->cached_sq_head;
963c6abb 1800 struct io_kiocb *cur;
9cf7c104 1801
a3dbdf54 1802 /* need original cached_sq_head, but it was increased for each req */
963c6abb 1803 io_for_each_link(cur, req)
a3dbdf54
PB
1804 seq--;
1805 return seq;
9cf7c104
PB
1806}
1807
c072481d 1808static __cold void io_drain_req(struct io_kiocb *req)
e276ae34 1809 __must_hold(&ctx->uring_lock)
de0617e4 1810{
a197f664 1811 struct io_ring_ctx *ctx = req->ctx;
27dc8338 1812 struct io_defer_entry *de;
f67676d1 1813 int ret;
e0eb71dc 1814 u32 seq = io_get_sequence(req);
3c19966d 1815
9d858b21 1816 /* Still need defer if there is pending req in defer list. */
e302f104 1817 spin_lock(&ctx->completion_lock);
5e371265 1818 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
e302f104 1819 spin_unlock(&ctx->completion_lock);
e0eb71dc 1820queue:
10c66904 1821 ctx->drain_active = false;
e0eb71dc
PB
1822 io_req_task_queue(req);
1823 return;
10c66904 1824 }
e302f104 1825 spin_unlock(&ctx->completion_lock);
9cf7c104 1826
cbdcb435 1827 io_prep_async_link(req);
27dc8338 1828 de = kmalloc(sizeof(*de), GFP_KERNEL);
76cc33d7 1829 if (!de) {
1b48773f 1830 ret = -ENOMEM;
ef5c600a
DY
1831 io_req_defer_failed(req, ret);
1832 return;
76cc33d7 1833 }
2d28390a 1834
79ebeaee 1835 spin_lock(&ctx->completion_lock);
9cf7c104 1836 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
79ebeaee 1837 spin_unlock(&ctx->completion_lock);
27dc8338 1838 kfree(de);
e0eb71dc 1839 goto queue;
de0617e4
JA
1840 }
1841
48863ffd 1842 trace_io_uring_defer(req);
27dc8338 1843 de->req = req;
9cf7c104 1844 de->seq = seq;
27dc8338 1845 list_add_tail(&de->list, &ctx->defer_list);
79ebeaee 1846 spin_unlock(&ctx->completion_lock);
de0617e4
JA
1847}
1848
f4992544
JA
1849static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def,
1850 unsigned int issue_flags)
6bf9c47a 1851{
f4992544 1852 if (req->file || !def->needs_file)
6bf9c47a
JA
1853 return true;
1854
1855 if (req->flags & REQ_F_FIXED_FILE)
cef216fc 1856 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
6bf9c47a 1857 else
cef216fc 1858 req->file = io_file_get_normal(req, req->cqe.fd);
6bf9c47a 1859
772f5e00 1860 return !!req->file;
6bf9c47a
JA
1861}
1862
889fca73 1863static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
2b188cc1 1864{
a7dd2782 1865 const struct io_issue_def *def = &io_issue_defs[req->opcode];
5730b27e 1866 const struct cred *creds = NULL;
d625c6ee 1867 int ret;
2b188cc1 1868
f4992544 1869 if (unlikely(!io_assign_file(req, def, issue_flags)))
70152140
JA
1870 return -EBADF;
1871
6878b40e 1872 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
c10d1f98 1873 creds = override_creds(req->creds);
5730b27e 1874
fcde59fe 1875 if (!def->audit_skip)
5bd2182d
PM
1876 audit_uring_entry(req->opcode);
1877
0702e536 1878 ret = def->issue(req, issue_flags);
2b188cc1 1879
fcde59fe 1880 if (!def->audit_skip)
5bd2182d
PM
1881 audit_uring_exit(!ret, ret);
1882
5730b27e
JA
1883 if (creds)
1884 revert_creds(creds);
97b388d7 1885
75d7b3ae
PB
1886 if (ret == IOU_OK) {
1887 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
9da070b1 1888 io_req_complete_defer(req);
75d7b3ae 1889 else
1bec951c 1890 io_req_complete_post(req, issue_flags);
75d7b3ae 1891 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
def596e9 1892 return ret;
97b388d7 1893
b532576e 1894 /* If the op doesn't have a file, we're not polling for it */
ef0ec1ad 1895 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
9882131c 1896 io_iopoll_req_issued(req, issue_flags);
def596e9
JA
1897
1898 return 0;
2b188cc1
JA
1899}
1900
a282967c 1901int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts)
329061d3 1902{
a282967c 1903 io_tw_lock(req->ctx, ts);
9a692451
DY
1904 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1905 IO_URING_F_COMPLETE_DEFER);
329061d3
JA
1906}
1907
c9f06aa7 1908struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
ebc11b6c
PB
1909{
1910 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
247f97a5 1911 struct io_kiocb *nxt = NULL;
ebc11b6c 1912
247f97a5
PB
1913 if (req_ref_put_and_test(req)) {
1914 if (req->flags & IO_REQ_LINK_FLAGS)
1915 nxt = io_req_find_next(req);
1916 io_free_req(req);
1917 }
1918 return nxt ? &nxt->work : NULL;
ebc11b6c
PB
1919}
1920
c9f06aa7 1921void io_wq_submit_work(struct io_wq_work *work)
2b188cc1
JA
1922{
1923 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
a7dd2782 1924 const struct io_issue_def *def = &io_issue_defs[req->opcode];
e6aeb272 1925 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
d01905db 1926 bool needs_poll = false;
6bf9c47a 1927 int ret = 0, err = -ECANCELED;
2b188cc1 1928
23a6c9ac 1929 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
48dcd38d
PB
1930 if (!(req->flags & REQ_F_REFCOUNT))
1931 __io_req_set_refcount(req, 2);
1932 else
1933 req_ref_get(req);
5d5901a3 1934
cb2d344c 1935 io_arm_ltimeout(req);
6bf9c47a 1936
dadebc35 1937 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
d01905db 1938 if (work->flags & IO_WQ_WORK_CANCEL) {
0f8da75b 1939fail:
6bf9c47a 1940 io_req_task_queue_fail(req, err);
d01905db
PB
1941 return;
1942 }
f4992544 1943 if (!io_assign_file(req, def, issue_flags)) {
0f8da75b
PB
1944 err = -EBADF;
1945 work->flags |= IO_WQ_WORK_CANCEL;
1946 goto fail;
1947 }
31b51510 1948
d01905db 1949 if (req->flags & REQ_F_FORCE_ASYNC) {
afb7f56f
PB
1950 bool opcode_poll = def->pollin || def->pollout;
1951
1952 if (opcode_poll && file_can_poll(req->file)) {
1953 needs_poll = true;
d01905db 1954 issue_flags |= IO_URING_F_NONBLOCK;
afb7f56f 1955 }
561fb04a 1956 }
31b51510 1957
d01905db
PB
1958 do {
1959 ret = io_issue_sqe(req, issue_flags);
1960 if (ret != -EAGAIN)
1961 break;
a9be2022
JA
1962
1963 /*
1964 * If REQ_F_NOWAIT is set, then don't wait or retry with
1965 * poll. -EAGAIN is final for that case.
1966 */
1967 if (req->flags & REQ_F_NOWAIT)
1968 break;
1969
d01905db
PB
1970 /*
1971 * We can get EAGAIN for iopolled IO even though we're
1972 * forcing a sync submission from here, since we can't
1973 * wait for request slots on the block side.
1974 */
1975 if (!needs_poll) {
e0deb6a0
PB
1976 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1977 break;
45500dc4
PB
1978 if (io_wq_worker_stopped())
1979 break;
d01905db
PB
1980 cond_resched();
1981 continue;
90fa0288
HX
1982 }
1983
4d9237e3 1984 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
d01905db
PB
1985 return;
1986 /* aborted or ready, in either case retry blocking */
1987 needs_poll = false;
1988 issue_flags &= ~IO_URING_F_NONBLOCK;
1989 } while (1);
31b51510 1990
a3df7698 1991 /* avoid locking problems by failing it from a clean context */
97b388d7 1992 if (ret < 0)
a3df7698 1993 io_req_task_queue_fail(req, ret);
2b188cc1
JA
1994}
1995
531113bb
JA
1996inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1997 unsigned int issue_flags)
09bb8394 1998{
5106dd6e 1999 struct io_ring_ctx *ctx = req->ctx;
4bfb0c9a 2000 struct io_fixed_file *slot;
5106dd6e 2001 struct file *file = NULL;
09bb8394 2002
93f052cb 2003 io_ring_submit_lock(ctx, issue_flags);
5106dd6e 2004
ac177053 2005 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
5106dd6e 2006 goto out;
ac177053 2007 fd = array_index_nospec(fd, ctx->nr_user_files);
4bfb0c9a
CH
2008 slot = io_fixed_file_slot(&ctx->file_table, fd);
2009 file = io_slot_file(slot);
2010 req->flags |= io_slot_flags(slot);
5106dd6e
JA
2011 io_req_set_rsrc_node(req, ctx, 0);
2012out:
93f052cb 2013 io_ring_submit_unlock(ctx, issue_flags);
ac177053
PB
2014 return file;
2015}
d44f554e 2016
531113bb 2017struct file *io_file_get_normal(struct io_kiocb *req, int fd)
ac177053 2018{
62906e89 2019 struct file *file = fget(fd);
ac177053 2020
48863ffd 2021 trace_io_uring_file_get(req, fd);
09bb8394 2022
ac177053 2023 /* we don't allow fixed io_uring files */
e5550a14 2024 if (file && io_is_uring_fops(file))
9cae36a0 2025 io_req_track_inflight(req);
8371adf5 2026 return file;
09bb8394
JA
2027}
2028
7bfa9bad 2029static void io_queue_async(struct io_kiocb *req, int ret)
d475a9a6
PB
2030 __must_hold(&req->ctx->uring_lock)
2031{
7bfa9bad
PB
2032 struct io_kiocb *linked_timeout;
2033
2034 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
973fc83f 2035 io_req_defer_failed(req, ret);
7bfa9bad
PB
2036 return;
2037 }
2038
2039 linked_timeout = io_prep_linked_timeout(req);
d475a9a6 2040
4d9237e3 2041 switch (io_arm_poll_handler(req, 0)) {
d475a9a6 2042 case IO_APOLL_READY:
336d28a8 2043 io_kbuf_recycle(req, 0);
d475a9a6
PB
2044 io_req_task_queue(req);
2045 break;
2046 case IO_APOLL_ABORTED:
6436c770 2047 io_kbuf_recycle(req, 0);
77955efb 2048 io_queue_iowq(req, NULL);
d475a9a6 2049 break;
b1c62645 2050 case IO_APOLL_OK:
b1c62645 2051 break;
d475a9a6
PB
2052 }
2053
2054 if (linked_timeout)
2055 io_queue_linked_timeout(linked_timeout);
2056}
2057
cbc2e203 2058static inline void io_queue_sqe(struct io_kiocb *req)
282cdc86 2059 __must_hold(&req->ctx->uring_lock)
2b188cc1 2060{
e0c5c576 2061 int ret;
2b188cc1 2062
c5eef2b9 2063 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
193155c8 2064
491381ce
JA
2065 /*
2066 * We async punt it if the file wasn't marked NOWAIT, or if the file
2067 * doesn't support non-blocking read/write attempts
2068 */
7bfa9bad 2069 if (likely(!ret))
cb2d344c 2070 io_arm_ltimeout(req);
7bfa9bad
PB
2071 else
2072 io_queue_async(req, ret);
2b188cc1
JA
2073}
2074
4652fe3f 2075static void io_queue_sqe_fallback(struct io_kiocb *req)
282cdc86 2076 __must_hold(&req->ctx->uring_lock)
4fe2c963 2077{
17b147f6
PB
2078 if (unlikely(req->flags & REQ_F_FAIL)) {
2079 /*
2080 * We don't submit, fail them all, for that replace hardlinks
2081 * with normal links. Extra REQ_F_LINK is tolerated.
2082 */
2083 req->flags &= ~REQ_F_HARDLINK;
2084 req->flags |= REQ_F_LINK;
973fc83f 2085 io_req_defer_failed(req, req->cqe.res);
76cc33d7
PB
2086 } else {
2087 int ret = io_req_prep_async(req);
2088
ef5c600a 2089 if (unlikely(ret)) {
973fc83f 2090 io_req_defer_failed(req, ret);
ef5c600a
DY
2091 return;
2092 }
2093
2094 if (unlikely(req->ctx->drain_active))
2095 io_drain_req(req);
76cc33d7 2096 else
77955efb 2097 io_queue_iowq(req, NULL);
ce35a47a 2098 }
4fe2c963
JL
2099}
2100
b16fed66
PB
2101/*
2102 * Check SQE restrictions (opcode and flags).
2103 *
2104 * Returns 'true' if SQE is allowed, 'false' otherwise.
2105 */
2106static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2107 struct io_kiocb *req,
2108 unsigned int sqe_flags)
4fe2c963 2109{
b16fed66
PB
2110 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2111 return false;
2112
2113 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2114 ctx->restrictions.sqe_flags_required)
2115 return false;
2116
2117 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2118 ctx->restrictions.sqe_flags_required))
2119 return false;
2120
2121 return true;
4fe2c963
JL
2122}
2123
22b2ca31
PB
2124static void io_init_req_drain(struct io_kiocb *req)
2125{
2126 struct io_ring_ctx *ctx = req->ctx;
2127 struct io_kiocb *head = ctx->submit_state.link.head;
2128
2129 ctx->drain_active = true;
2130 if (head) {
2131 /*
2132 * If we need to drain a request in the middle of a link, drain
2133 * the head request and the next request/link after the current
2134 * link. Considering sequential execution of links,
b6c7db32 2135 * REQ_F_IO_DRAIN will be maintained for every request of our
22b2ca31
PB
2136 * link.
2137 */
b6c7db32 2138 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
22b2ca31
PB
2139 ctx->drain_next = true;
2140 }
2141}
2142
b16fed66
PB
2143static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2144 const struct io_uring_sqe *sqe)
282cdc86 2145 __must_hold(&ctx->uring_lock)
b16fed66 2146{
a7dd2782 2147 const struct io_issue_def *def;
b16fed66 2148 unsigned int sqe_flags;
fc0ae024 2149 int personality;
4a04d1d1 2150 u8 opcode;
b16fed66 2151
864ea921 2152 /* req is partially pre-initialised, see io_preinit_req() */
4a04d1d1 2153 req->opcode = opcode = READ_ONCE(sqe->opcode);
b16fed66
PB
2154 /* same numerical values with corresponding REQ_F_*, safe to copy */
2155 req->flags = sqe_flags = READ_ONCE(sqe->flags);
cef216fc 2156 req->cqe.user_data = READ_ONCE(sqe->user_data);
b16fed66 2157 req->file = NULL;
c1bdf8ed 2158 req->rsrc_node = NULL;
b16fed66 2159 req->task = current;
b16fed66 2160
4a04d1d1
PB
2161 if (unlikely(opcode >= IORING_OP_LAST)) {
2162 req->opcode = 0;
b16fed66 2163 return -EINVAL;
4a04d1d1 2164 }
a7dd2782 2165 def = &io_issue_defs[opcode];
68fe256a
PB
2166 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2167 /* enforce forwards compatibility on users */
2168 if (sqe_flags & ~SQE_VALID_FLAGS)
2169 return -EINVAL;
4e906702 2170 if (sqe_flags & IOSQE_BUFFER_SELECT) {
fcde59fe 2171 if (!def->buffer_select)
4e906702
JA
2172 return -EOPNOTSUPP;
2173 req->buf_index = READ_ONCE(sqe->buf_group);
2174 }
5562a8d7
PB
2175 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2176 ctx->drain_disabled = true;
2177 if (sqe_flags & IOSQE_IO_DRAIN) {
2178 if (ctx->drain_disabled)
2179 return -EOPNOTSUPP;
22b2ca31 2180 io_init_req_drain(req);
5562a8d7 2181 }
2a56a9bd
PB
2182 }
2183 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2184 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2185 return -EACCES;
2186 /* knock it to the slow queue path, will be drained there */
2187 if (ctx->drain_active)
2188 req->flags |= REQ_F_FORCE_ASYNC;
2189 /* if there is no link, we're at "next" request and need to drain */
2190 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2191 ctx->drain_next = false;
2192 ctx->drain_active = true;
b6c7db32 2193 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2a56a9bd 2194 }
68fe256a 2195 }
b16fed66 2196
fcde59fe 2197 if (!def->ioprio && sqe->ioprio)
73911426 2198 return -EINVAL;
fcde59fe 2199 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
73911426
JA
2200 return -EINVAL;
2201
fcde59fe 2202 if (def->needs_file) {
6d63416d
PB
2203 struct io_submit_state *state = &ctx->submit_state;
2204
cef216fc 2205 req->cqe.fd = READ_ONCE(sqe->fd);
6bf9c47a 2206
6d63416d
PB
2207 /*
2208 * Plug now if we have more than 2 IO left after this, and the
2209 * target is potentially a read/write to block based storage.
2210 */
fcde59fe 2211 if (state->need_plug && def->plug) {
6d63416d
PB
2212 state->plug_started = true;
2213 state->need_plug = false;
5ca7a8b3 2214 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
6d63416d 2215 }
b16fed66 2216 }
863e0560 2217
003e8dcc
JA
2218 personality = READ_ONCE(sqe->personality);
2219 if (personality) {
cdab10bf
LT
2220 int ret;
2221
c10d1f98
PB
2222 req->creds = xa_load(&ctx->personalities, personality);
2223 if (!req->creds)
003e8dcc 2224 return -EINVAL;
c10d1f98 2225 get_cred(req->creds);
cdc1404a
PM
2226 ret = security_uring_override_creds(req->creds);
2227 if (ret) {
2228 put_cred(req->creds);
2229 return ret;
2230 }
b8e64b53 2231 req->flags |= REQ_F_CREDS;
003e8dcc 2232 }
b16fed66 2233
0702e536 2234 return def->prep(req, sqe);
b16fed66
PB
2235}
2236
df3becde
PB
2237static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2238 struct io_kiocb *req, int ret)
2239{
2240 struct io_ring_ctx *ctx = req->ctx;
2241 struct io_submit_link *link = &ctx->submit_state.link;
2242 struct io_kiocb *head = link->head;
2243
48863ffd 2244 trace_io_uring_req_failed(sqe, req, ret);
df3becde
PB
2245
2246 /*
2247 * Avoid breaking links in the middle as it renders links with SQPOLL
2248 * unusable. Instead of failing eagerly, continue assembling the link if
2249 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2250 * should find the flag and handle the rest.
2251 */
2252 req_fail_link_node(req, ret);
2253 if (head && !(head->flags & REQ_F_FAIL))
2254 req_fail_link_node(head, -ECANCELED);
2255
2256 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2257 if (head) {
2258 link->last->link = req;
2259 link->head = NULL;
2260 req = head;
2261 }
2262 io_queue_sqe_fallback(req);
2263 return ret;
2264 }
2265
2266 if (head)
2267 link->last->link = req;
2268 else
2269 link->head = req;
2270 link->last = req;
2271 return 0;
2272}
2273
2274static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
a1ab7b35 2275 const struct io_uring_sqe *sqe)
282cdc86 2276 __must_hold(&ctx->uring_lock)
9e645e11 2277{
a1ab7b35 2278 struct io_submit_link *link = &ctx->submit_state.link;
ef4ff581 2279 int ret;
9e645e11 2280
a6b8cadc 2281 ret = io_init_req(ctx, req, sqe);
df3becde
PB
2282 if (unlikely(ret))
2283 return io_submit_fail_init(sqe, req, ret);
441b8a78 2284
2ad57931 2285 trace_io_uring_submit_req(req);
a6b8cadc 2286
9e645e11
JA
2287 /*
2288 * If we already have a head request, queue this one for async
2289 * submittal once the head completes. If we don't have a head but
2290 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2291 * submitted sync once the chain is complete. If none of those
2292 * conditions are true (normal request), then just queue it.
2293 */
924a07e4 2294 if (unlikely(link->head)) {
df3becde
PB
2295 ret = io_req_prep_async(req);
2296 if (unlikely(ret))
2297 return io_submit_fail_init(sqe, req, ret);
2298
48863ffd 2299 trace_io_uring_link(req, link->head);
f2f87370 2300 link->last->link = req;
863e0560 2301 link->last = req;
32fe525b 2302
da1a08c5 2303 if (req->flags & IO_REQ_LINK_FLAGS)
f15a3431 2304 return 0;
df3becde
PB
2305 /* last request of the link, flush it */
2306 req = link->head;
f15a3431 2307 link->head = NULL;
924a07e4
PB
2308 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2309 goto fallback;
2310
2311 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2312 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2313 if (req->flags & IO_REQ_LINK_FLAGS) {
2314 link->head = req;
2315 link->last = req;
2316 } else {
2317fallback:
2318 io_queue_sqe_fallback(req);
2319 }
f15a3431 2320 return 0;
9e645e11 2321 }
2e6e1fde 2322
f15a3431 2323 io_queue_sqe(req);
1d4240cc 2324 return 0;
9e645e11
JA
2325}
2326
9a56a232
JA
2327/*
2328 * Batched submission is done, ensure local IO is flushed out.
2329 */
553deffd 2330static void io_submit_state_end(struct io_ring_ctx *ctx)
9a56a232 2331{
553deffd
PB
2332 struct io_submit_state *state = &ctx->submit_state;
2333
e126391c
PB
2334 if (unlikely(state->link.head))
2335 io_queue_sqe_fallback(state->link.head);
553deffd 2336 /* flush only after queuing links as they can generate completions */
c450178d 2337 io_submit_flush_completions(ctx);
27926b68
JA
2338 if (state->plug_started)
2339 blk_finish_plug(&state->plug);
9a56a232
JA
2340}
2341
2342/*
2343 * Start submission side cache.
2344 */
2345static void io_submit_state_start(struct io_submit_state *state,
ba88ff11 2346 unsigned int max_ios)
9a56a232 2347{
27926b68 2348 state->plug_started = false;
4b628aeb 2349 state->need_plug = max_ios > 2;
5ca7a8b3 2350 state->submit_nr = max_ios;
a1ab7b35
PB
2351 /* set only head, no need to init link_last in advance */
2352 state->link.head = NULL;
9a56a232
JA
2353}
2354
2b188cc1
JA
2355static void io_commit_sqring(struct io_ring_ctx *ctx)
2356{
75b28aff 2357 struct io_rings *rings = ctx->rings;
2b188cc1 2358
caf582c6
PB
2359 /*
2360 * Ensure any loads from the SQEs are done at this point,
2361 * since once we write the new head, the application could
2362 * write new data to them.
2363 */
2364 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2b188cc1
JA
2365}
2366
2b188cc1 2367/*
dd9ae8a0 2368 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2b188cc1
JA
2369 * that is mapped by userspace. This means that care needs to be taken to
2370 * ensure that reads are stable, as we cannot rely on userspace always
2371 * being a good citizen. If members of the sqe are validated and then later
2372 * used, it's important that those reads are done through READ_ONCE() to
2373 * prevent a re-load down the line.
2374 */
b5083dfa 2375static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe)
2b188cc1 2376{
2af89abd
PB
2377 unsigned mask = ctx->sq_entries - 1;
2378 unsigned head = ctx->cached_sq_head++ & mask;
2379
2380 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) {
2381 head = READ_ONCE(ctx->sq_array[head]);
2382 if (unlikely(head >= ctx->sq_entries)) {
2383 /* drop invalid entries */
2384 spin_lock(&ctx->completion_lock);
2385 ctx->cq_extra--;
2386 spin_unlock(&ctx->completion_lock);
2387 WRITE_ONCE(ctx->rings->sq_dropped,
2388 READ_ONCE(ctx->rings->sq_dropped) + 1);
2389 return false;
2390 }
2391 }
2b188cc1
JA
2392
2393 /*
2394 * The cached sq head (or cq tail) serves two purposes:
2395 *
2396 * 1) allows us to batch the cost of updating the user visible
2397 * head updates.
2398 * 2) allows the kernel side to track the head on its own, even
2399 * though the application is the one updating it.
2400 */
2b188cc1 2401
2af89abd
PB
2402 /* double index for 128-byte SQEs, twice as long */
2403 if (ctx->flags & IORING_SETUP_SQE128)
2404 head <<= 1;
2405 *sqe = &ctx->sq_sqes[head];
2406 return true;
709b302f
PB
2407}
2408
17437f31 2409int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
282cdc86 2410 __must_hold(&ctx->uring_lock)
6c271ce2 2411{
69629809 2412 unsigned int entries = io_sqring_entries(ctx);
8e6971a8
PB
2413 unsigned int left;
2414 int ret;
6c271ce2 2415
51d48dab 2416 if (unlikely(!entries))
69629809 2417 return 0;
ee7d46d9 2418 /* make sure SQ entry isn't read before tail */
e3ef728f 2419 ret = left = min(nr, entries);
8e6971a8
PB
2420 io_get_task_refs(left);
2421 io_submit_state_start(&ctx->submit_state, left);
6c271ce2 2422
69629809 2423 do {
3529d8c2 2424 const struct io_uring_sqe *sqe;
196be95c 2425 struct io_kiocb *req;
fb5ccc98 2426
c8576f3e 2427 if (unlikely(!io_alloc_req(ctx, &req)))
fb5ccc98 2428 break;
b5083dfa 2429 if (unlikely(!io_get_sqe(ctx, &sqe))) {
fa05457a 2430 io_req_add_to_cache(req, ctx);
4fccfcbb
PB
2431 break;
2432 }
6c271ce2 2433
1cd15904
PB
2434 /*
2435 * Continue submitting even for sqe failure if the
2436 * ring was setup with IORING_SETUP_SUBMIT_ALL
2437 */
2438 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2439 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2440 left--;
2441 break;
bcbb7bf6 2442 }
1cd15904 2443 } while (--left);
9466f437 2444
8e6971a8
PB
2445 if (unlikely(left)) {
2446 ret -= left;
2447 /* try again if it submitted nothing and can't allocate a req */
2448 if (!ret && io_req_cache_empty(ctx))
2449 ret = -EAGAIN;
2450 current->io_uring->cached_refs += left;
9466f437 2451 }
6c271ce2 2452
553deffd 2453 io_submit_state_end(ctx);
ae9428ca
PB
2454 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2455 io_commit_sqring(ctx);
8e6971a8 2456 return ret;
6c271ce2
JA
2457}
2458
bda52162
JA
2459struct io_wait_queue {
2460 struct wait_queue_entry wq;
2461 struct io_ring_ctx *ctx;
5fd46178 2462 unsigned cq_tail;
bda52162 2463 unsigned nr_timeouts;
d33a39e5 2464 ktime_t timeout;
bda52162
JA
2465};
2466
b4c98d59
DY
2467static inline bool io_has_work(struct io_ring_ctx *ctx)
2468{
c0e0d6ba 2469 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
490c00eb 2470 !llist_empty(&ctx->work_llist);
b4c98d59
DY
2471}
2472
6c503150 2473static inline bool io_should_wake(struct io_wait_queue *iowq)
bda52162
JA
2474{
2475 struct io_ring_ctx *ctx = iowq->ctx;
0fc8c2ac 2476 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
bda52162
JA
2477
2478 /*
d195a66e 2479 * Wake up if we have enough events, or if a timeout occurred since we
bda52162
JA
2480 * started waiting. For timeouts, we always want to return to userspace,
2481 * regardless of event count.
2482 */
5fd46178 2483 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
bda52162
JA
2484}
2485
2486static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2487 int wake_flags, void *key)
2488{
bd550173 2489 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq);
bda52162 2490
6c503150
PB
2491 /*
2492 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2493 * the task, and the next invocation will do it.
2494 */
bd550173 2495 if (io_should_wake(iowq) || io_has_work(iowq->ctx))
6c503150
PB
2496 return autoremove_wake_function(curr, mode, wake_flags, key);
2497 return -1;
bda52162
JA
2498}
2499
c0e0d6ba 2500int io_run_task_work_sig(struct io_ring_ctx *ctx)
af9c1a44 2501{
1414d629 2502 if (!llist_empty(&ctx->work_llist)) {
2f413956 2503 __set_current_state(TASK_RUNNING);
1414d629 2504 if (io_run_local_work(ctx) > 0)
d246c759 2505 return 0;
1414d629
PB
2506 }
2507 if (io_run_task_work() > 0)
d246c759 2508 return 0;
c5020bc8
OL
2509 if (task_sigpending(current))
2510 return -EINTR;
2511 return 0;
af9c1a44
JA
2512}
2513
7b72d661
JA
2514static bool current_pending_io(void)
2515{
2516 struct io_uring_task *tctx = current->io_uring;
2517
2518 if (!tctx)
2519 return false;
2520 return percpu_counter_read_positive(&tctx->inflight);
2521}
2522
eeb60b9a
PB
2523/* when returns >0, the caller should retry */
2524static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
d33a39e5 2525 struct io_wait_queue *iowq)
eeb60b9a 2526{
7b72d661 2527 int io_wait, ret;
8a796565 2528
3fcf19d5
PB
2529 if (unlikely(READ_ONCE(ctx->check_cq)))
2530 return 1;
846072f1
PB
2531 if (unlikely(!llist_empty(&ctx->work_llist)))
2532 return 1;
2533 if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL)))
2534 return 1;
2535 if (unlikely(task_sigpending(current)))
2536 return -EINTR;
2537 if (unlikely(io_should_wake(iowq)))
2538 return 0;
8a796565
AF
2539
2540 /*
7b72d661
JA
2541 * Mark us as being in io_wait if we have pending requests, so cpufreq
2542 * can take into account that the task is waiting for IO - turns out
2543 * to be important for low QD IO.
8a796565 2544 */
7b72d661
JA
2545 io_wait = current->in_iowait;
2546 if (current_pending_io())
2547 current->in_iowait = 1;
8a796565 2548 ret = 0;
d33a39e5 2549 if (iowq->timeout == KTIME_MAX)
46ae7eef 2550 schedule();
d33a39e5 2551 else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS))
8a796565 2552 ret = -ETIME;
7b72d661 2553 current->in_iowait = io_wait;
8a796565 2554 return ret;
eeb60b9a
PB
2555}
2556
2b188cc1
JA
2557/*
2558 * Wait until events become available, if we don't already have some. The
2559 * application must reap them itself, as they reside on the shared cq ring.
2560 */
2561static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
c73ebb68
HX
2562 const sigset_t __user *sig, size_t sigsz,
2563 struct __kernel_timespec __user *uts)
2b188cc1 2564{
90291099 2565 struct io_wait_queue iowq;
75b28aff 2566 struct io_rings *rings = ctx->rings;
c1d5a224 2567 int ret;
2b188cc1 2568
76de6749
PB
2569 if (!io_allowed_run_tw(ctx))
2570 return -EEXIST;
140102ae
PB
2571 if (!llist_empty(&ctx->work_llist))
2572 io_run_local_work(ctx);
f36ba6cf
PB
2573 io_run_task_work();
2574 io_cqring_overflow_flush(ctx);
2575 /* if user messes with these they will just get an early return */
2576 if (__io_cqring_events_user(ctx) >= min_events)
2577 return 0;
2b188cc1
JA
2578
2579 if (sig) {
9e75ad5d
AB
2580#ifdef CONFIG_COMPAT
2581 if (in_compat_syscall())
2582 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
b772434b 2583 sigsz);
9e75ad5d
AB
2584 else
2585#endif
b772434b 2586 ret = set_user_sigmask(sig, sigsz);
9e75ad5d 2587
2b188cc1
JA
2588 if (ret)
2589 return ret;
2590 }
2591
90291099
PB
2592 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2593 iowq.wq.private = current;
2594 INIT_LIST_HEAD(&iowq.wq.entry);
2595 iowq.ctx = ctx;
bda52162 2596 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
5fd46178 2597 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
d33a39e5
PB
2598 iowq.timeout = KTIME_MAX;
2599
2600 if (uts) {
2601 struct timespec64 ts;
2602
2603 if (get_timespec64(&ts, uts))
2604 return -EFAULT;
2605 iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2606 }
90291099 2607
c826bd7a 2608 trace_io_uring_cqring_wait(ctx, min_events);
bda52162 2609 do {
3fcf19d5
PB
2610 unsigned long check_cq;
2611
130bd686 2612 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
8751d154
PB
2613 int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
2614
2615 atomic_set(&ctx->cq_wait_nr, nr_wait);
130bd686
PB
2616 set_current_state(TASK_INTERRUPTIBLE);
2617 } else {
2618 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2619 TASK_INTERRUPTIBLE);
2620 }
2621
d33a39e5 2622 ret = io_cqring_wait_schedule(ctx, &iowq);
130bd686 2623 __set_current_state(TASK_RUNNING);
8751d154 2624 atomic_set(&ctx->cq_wait_nr, 0);
d80c0f00 2625
846072f1
PB
2626 if (ret < 0)
2627 break;
2628 /*
2629 * Run task_work after scheduling and before io_should_wake().
2630 * If we got woken because of task_work being processed, run it
2631 * now rather than let the caller do another wait loop.
2632 */
2633 io_run_task_work();
2634 if (!llist_empty(&ctx->work_llist))
2635 io_run_local_work(ctx);
3fcf19d5
PB
2636
2637 check_cq = READ_ONCE(ctx->check_cq);
2638 if (unlikely(check_cq)) {
2639 /* let the caller flush overflows, retry */
326a9e48 2640 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
3fcf19d5 2641 io_cqring_do_overflow_flush(ctx);
3fcf19d5
PB
2642 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) {
2643 ret = -EBADR;
2644 break;
2645 }
2646 }
2647
846072f1
PB
2648 if (io_should_wake(&iowq)) {
2649 ret = 0;
35d90f95 2650 break;
846072f1 2651 }
ca0a2651 2652 cond_resched();
846072f1 2653 } while (1);
bda52162 2654
130bd686
PB
2655 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
2656 finish_wait(&ctx->cq_wait, &iowq.wq);
b7db41c9 2657 restore_saved_sigmask_unless(ret == -EINTR);
2b188cc1 2658
75b28aff 2659 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2b188cc1
JA
2660}
2661
73572984 2662static void io_mem_free(void *ptr)
b9bd2bea 2663{
73572984
JA
2664 if (!ptr)
2665 return;
b9bd2bea 2666
99a9e0b8 2667 folio_put(virt_to_folio(ptr));
b9bd2bea
PB
2668}
2669
03d89a2d
JA
2670static void io_pages_free(struct page ***pages, int npages)
2671{
2672 struct page **page_array;
2673 int i;
2674
2675 if (!pages)
2676 return;
2677 page_array = *pages;
2678 for (i = 0; i < npages; i++)
2679 unpin_user_page(page_array[i]);
2680 kvfree(page_array);
2681 *pages = NULL;
2682}
2683
2684static void *__io_uaddr_map(struct page ***pages, unsigned short *npages,
2685 unsigned long uaddr, size_t size)
2686{
2687 struct page **page_array;
2688 unsigned int nr_pages;
223ef474 2689 int ret, i;
03d89a2d
JA
2690
2691 *npages = 0;
2692
2693 if (uaddr & (PAGE_SIZE - 1) || !size)
2694 return ERR_PTR(-EINVAL);
2695
2696 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2697 if (nr_pages > USHRT_MAX)
2698 return ERR_PTR(-EINVAL);
2699 page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL);
2700 if (!page_array)
2701 return ERR_PTR(-ENOMEM);
2702
2703 ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM,
2704 page_array);
2705 if (ret != nr_pages) {
2706err:
2707 io_pages_free(&page_array, ret > 0 ? ret : 0);
2708 return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT);
2709 }
2710 /*
2711 * Should be a single page. If the ring is small enough that we can
2712 * use a normal page, that is fine. If we need multiple pages, then
2713 * userspace should use a huge page. That's the only way to guarantee
2714 * that we get contigious memory, outside of just being lucky or
2715 * (currently) having low memory fragmentation.
2716 */
2717 if (page_array[0] != page_array[ret - 1])
2718 goto err;
223ef474
JA
2719
2720 /*
2721 * Can't support mapping user allocated ring memory on 32-bit archs
2722 * where it could potentially reside in highmem. Just fail those with
2723 * -EINVAL, just like we did on kernels that didn't support this
2724 * feature.
2725 */
2726 for (i = 0; i < nr_pages; i++) {
2727 if (PageHighMem(page_array[i])) {
2728 ret = -EINVAL;
2729 goto err;
2730 }
2731 }
2732
03d89a2d
JA
2733 *pages = page_array;
2734 *npages = nr_pages;
2735 return page_to_virt(page_array[0]);
2736}
2737
2738static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2739 size_t size)
2740{
2741 return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr,
2742 size);
2743}
2744
2745static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr,
2746 size_t size)
2747{
2748 return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr,
2749 size);
2750}
2751
9c189eee
JA
2752static void io_rings_free(struct io_ring_ctx *ctx)
2753{
03d89a2d
JA
2754 if (!(ctx->flags & IORING_SETUP_NO_MMAP)) {
2755 io_mem_free(ctx->rings);
2756 io_mem_free(ctx->sq_sqes);
2757 ctx->rings = NULL;
2758 ctx->sq_sqes = NULL;
2759 } else {
2760 io_pages_free(&ctx->ring_pages, ctx->n_ring_pages);
2761 io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages);
2762 }
9c189eee
JA
2763}
2764
73572984 2765static void *io_mem_alloc(size_t size)
b9bd2bea 2766{
73572984 2767 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
e27cef86 2768 void *ret;
b9bd2bea 2769
e27cef86
JA
2770 ret = (void *) __get_free_pages(gfp, get_order(size));
2771 if (ret)
2772 return ret;
2773 return ERR_PTR(-ENOMEM);
b9bd2bea
PB
2774}
2775
73572984
JA
2776static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2777 unsigned int cq_entries, size_t *sq_offset)
6b06314c 2778{
73572984
JA
2779 struct io_rings *rings;
2780 size_t off, sq_array_size;
6b06314c 2781
73572984
JA
2782 off = struct_size(rings, cqes, cq_entries);
2783 if (off == SIZE_MAX)
2784 return SIZE_MAX;
2785 if (ctx->flags & IORING_SETUP_CQE32) {
2786 if (check_shl_overflow(off, 1, &off))
2787 return SIZE_MAX;
2788 }
ab409402 2789
73572984
JA
2790#ifdef CONFIG_SMP
2791 off = ALIGN(off, SMP_CACHE_BYTES);
2792 if (off == 0)
2793 return SIZE_MAX;
2794#endif
82fbcfa9 2795
2af89abd
PB
2796 if (ctx->flags & IORING_SETUP_NO_SQARRAY) {
2797 if (sq_offset)
2798 *sq_offset = SIZE_MAX;
2799 return off;
2800 }
2801
73572984
JA
2802 if (sq_offset)
2803 *sq_offset = off;
82fbcfa9 2804
73572984
JA
2805 sq_array_size = array_size(sizeof(u32), sq_entries);
2806 if (sq_array_size == SIZE_MAX)
2807 return SIZE_MAX;
6b06314c 2808
73572984
JA
2809 if (check_add_overflow(off, sq_array_size, &off))
2810 return SIZE_MAX;
8bad28d8 2811
73572984 2812 return off;
8bad28d8
HX
2813}
2814
73572984
JA
2815static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2816 unsigned int eventfd_async)
8bad28d8 2817{
73572984
JA
2818 struct io_ev_fd *ev_fd;
2819 __s32 __user *fds = arg;
2820 int fd;
f2303b1f 2821
73572984
JA
2822 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2823 lockdep_is_held(&ctx->uring_lock));
2824 if (ev_fd)
2825 return -EBUSY;
8bad28d8 2826
73572984
JA
2827 if (copy_from_user(&fd, fds, sizeof(*fds)))
2828 return -EFAULT;
8dd03afe 2829
73572984
JA
2830 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2831 if (!ev_fd)
2832 return -ENOMEM;
05f3fb3c 2833
73572984
JA
2834 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2835 if (IS_ERR(ev_fd->cq_ev_fd)) {
2836 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2837 kfree(ev_fd);
2838 return ret;
2839 }
305bef98
PB
2840
2841 spin_lock(&ctx->completion_lock);
2842 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2843 spin_unlock(&ctx->completion_lock);
2844
73572984
JA
2845 ev_fd->eventfd_async = eventfd_async;
2846 ctx->has_evfd = true;
2847 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
21a091b9
DY
2848 atomic_set(&ev_fd->refs, 1);
2849 atomic_set(&ev_fd->ops, 0);
73572984 2850 return 0;
d7954b2b
BM
2851}
2852
73572984 2853static int io_eventfd_unregister(struct io_ring_ctx *ctx)
1ad555c6 2854{
73572984
JA
2855 struct io_ev_fd *ev_fd;
2856
2857 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2858 lockdep_is_held(&ctx->uring_lock));
2859 if (ev_fd) {
2860 ctx->has_evfd = false;
2861 rcu_assign_pointer(ctx->io_ev_fd, NULL);
21a091b9
DY
2862 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2863 call_rcu(&ev_fd->rcu, io_eventfd_ops);
73572984
JA
2864 return 0;
2865 }
2d091d62 2866
73572984 2867 return -ENXIO;
44b31f2f
PB
2868}
2869
73572984 2870static void io_req_caches_free(struct io_ring_ctx *ctx)
2b188cc1 2871{
c8576f3e 2872 struct io_kiocb *req;
37f0e767 2873 int nr = 0;
bf019da7 2874
9a4fdbd8 2875 mutex_lock(&ctx->uring_lock);
34f0bc42 2876 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
9a4fdbd8 2877
88ab95be 2878 while (!io_req_cache_empty(ctx)) {
c8576f3e 2879 req = io_extract_req(ctx);
c2b6c6bc 2880 kmem_cache_free(req_cachep, req);
37f0e767 2881 nr++;
c2b6c6bc 2882 }
37f0e767
PB
2883 if (nr)
2884 percpu_ref_put_many(&ctx->refs, nr);
9a4fdbd8
JA
2885 mutex_unlock(&ctx->uring_lock);
2886}
2887
9eae8655
PB
2888static void io_rsrc_node_cache_free(struct io_cache_entry *entry)
2889{
2890 kfree(container_of(entry, struct io_rsrc_node, cache));
2891}
2892
c072481d 2893static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2b188cc1 2894{
37d1e2e3 2895 io_sq_thread_finish(ctx);
43597aac 2896 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
0b222eeb
PB
2897 if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)))
2898 return;
43597aac 2899
8bad28d8 2900 mutex_lock(&ctx->uring_lock);
43597aac 2901 if (ctx->buf_data)
bd54b6fe 2902 __io_sqe_buffers_unregister(ctx);
43597aac 2903 if (ctx->file_data)
08480400 2904 __io_sqe_files_unregister(ctx);
a85381d8 2905 io_cqring_overflow_kill(ctx);
9b402849 2906 io_eventfd_unregister(ctx);
9b797a37 2907 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
43e0bbbd 2908 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
5a2e745d 2909 io_destroy_buffers(ctx);
b4a72c05 2910 mutex_unlock(&ctx->uring_lock);
07db298a
PB
2911 if (ctx->sq_creds)
2912 put_cred(ctx->sq_creds);
97bbdc06
PB
2913 if (ctx->submitter_task)
2914 put_task_struct(ctx->submitter_task);
def596e9 2915
a7f0ed5a
PB
2916 /* there are no registered resources left, nobody uses it */
2917 if (ctx->rsrc_node)
9eae8655 2918 io_rsrc_node_destroy(ctx, ctx->rsrc_node);
a7f0ed5a
PB
2919
2920 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
def596e9 2921
2b188cc1 2922#if defined(CONFIG_UNIX)
355e8d26
EB
2923 if (ctx->ring_sock) {
2924 ctx->ring_sock->file = NULL; /* so that iput() is called */
2b188cc1 2925 sock_release(ctx->ring_sock);
355e8d26 2926 }
2b188cc1 2927#endif
ef9dd637 2928 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2b188cc1 2929
9eae8655 2930 io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free);
42b6419d
PB
2931 if (ctx->mm_account) {
2932 mmdrop(ctx->mm_account);
2933 ctx->mm_account = NULL;
2934 }
9c189eee 2935 io_rings_free(ctx);
2b188cc1
JA
2936
2937 percpu_ref_exit(&ctx->refs);
2b188cc1 2938 free_uid(ctx->user);
4010fec4 2939 io_req_caches_free(ctx);
e941894e
JA
2940 if (ctx->hash_map)
2941 io_wq_put_hash(ctx->hash_map);
e6f89be6 2942 kfree(ctx->cancel_table.hbs);
9ca9fb24 2943 kfree(ctx->cancel_table_locked.hbs);
9cfc7e94
JA
2944 kfree(ctx->io_bl);
2945 xa_destroy(&ctx->io_bl_xa);
2b188cc1
JA
2946 kfree(ctx);
2947}
2948
bca39f39
PB
2949static __cold void io_activate_pollwq_cb(struct callback_head *cb)
2950{
2951 struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx,
2952 poll_wq_task_work);
2953
2954 mutex_lock(&ctx->uring_lock);
2955 ctx->poll_activated = true;
2956 mutex_unlock(&ctx->uring_lock);
2957
2958 /*
2959 * Wake ups for some events between start of polling and activation
2960 * might've been lost due to loose synchronisation.
2961 */
2962 wake_up_all(&ctx->poll_wq);
2963 percpu_ref_put(&ctx->refs);
2964}
2965
2966static __cold void io_activate_pollwq(struct io_ring_ctx *ctx)
2967{
2968 spin_lock(&ctx->completion_lock);
2969 /* already activated or in progress */
2970 if (ctx->poll_activated || ctx->poll_wq_task_work.func)
2971 goto out;
2972 if (WARN_ON_ONCE(!ctx->task_complete))
2973 goto out;
2974 if (!ctx->submitter_task)
2975 goto out;
2976 /*
2977 * with ->submitter_task only the submitter task completes requests, we
2978 * only need to sync with it, which is done by injecting a tw
2979 */
2980 init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb);
2981 percpu_ref_get(&ctx->refs);
2982 if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL))
2983 percpu_ref_put(&ctx->refs);
2984out:
2985 spin_unlock(&ctx->completion_lock);
2986}
2987
2b188cc1
JA
2988static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2989{
2990 struct io_ring_ctx *ctx = file->private_data;
2991 __poll_t mask = 0;
2992
bca39f39
PB
2993 if (unlikely(!ctx->poll_activated))
2994 io_activate_pollwq(ctx);
2995
7b235dd8 2996 poll_wait(file, &ctx->poll_wq, wait);
4f7067c3
SB
2997 /*
2998 * synchronizes with barrier from wq_has_sleeper call in
2999 * io_commit_cqring
3000 */
2b188cc1 3001 smp_rmb();
90554200 3002 if (!io_sqring_full(ctx))
2b188cc1 3003 mask |= EPOLLOUT | EPOLLWRNORM;
ed670c3f
HX
3004
3005 /*
3006 * Don't flush cqring overflow list here, just do a simple check.
3007 * Otherwise there could possible be ABBA deadlock:
3008 * CPU0 CPU1
3009 * ---- ----
3010 * lock(&ctx->uring_lock);
3011 * lock(&ep->mtx);
3012 * lock(&ctx->uring_lock);
3013 * lock(&ep->mtx);
3014 *
3015 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
10d8bc35 3016 * pushes them to do the flush.
ed670c3f 3017 */
b4c98d59 3018
c10bb646 3019 if (__io_cqring_events_user(ctx) || io_has_work(ctx))
2b188cc1
JA
3020 mask |= EPOLLIN | EPOLLRDNORM;
3021
3022 return mask;
3023}
3024
0bead8cd 3025static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
071698e1 3026{
4379bf8b 3027 const struct cred *creds;
071698e1 3028
61cf9370 3029 creds = xa_erase(&ctx->personalities, id);
4379bf8b
JA
3030 if (creds) {
3031 put_cred(creds);
0bead8cd 3032 return 0;
1e6fa521 3033 }
0bead8cd
YD
3034
3035 return -EINVAL;
3036}
3037
d56d938b
PB
3038struct io_tctx_exit {
3039 struct callback_head task_work;
3040 struct completion completion;
baf186c4 3041 struct io_ring_ctx *ctx;
d56d938b
PB
3042};
3043
c072481d 3044static __cold void io_tctx_exit_cb(struct callback_head *cb)
d56d938b
PB
3045{
3046 struct io_uring_task *tctx = current->io_uring;
3047 struct io_tctx_exit *work;
3048
3049 work = container_of(cb, struct io_tctx_exit, task_work);
3050 /*
8d664282 3051 * When @in_cancel, we're in cancellation and it's racy to remove the
d56d938b 3052 * node. It'll be removed by the end of cancellation, just ignore it.
998b30c3
HM
3053 * tctx can be NULL if the queueing of this task_work raced with
3054 * work cancelation off the exec path.
d56d938b 3055 */
8d664282 3056 if (tctx && !atomic_read(&tctx->in_cancel))
eef51daa 3057 io_uring_del_tctx_node((unsigned long)work->ctx);
d56d938b
PB
3058 complete(&work->completion);
3059}
3060
c072481d 3061static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
28090c13
PB
3062{
3063 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3064
3065 return req->ctx == data;
3066}
3067
c072481d 3068static __cold void io_ring_exit_work(struct work_struct *work)
85faa7b8 3069{
d56d938b 3070 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
b5bb3a24 3071 unsigned long timeout = jiffies + HZ * 60 * 5;
58d3be2c 3072 unsigned long interval = HZ / 20;
d56d938b
PB
3073 struct io_tctx_exit exit;
3074 struct io_tctx_node *node;
3075 int ret;
85faa7b8 3076
56952e91
JA
3077 /*
3078 * If we're doing polled IO and end up having requests being
3079 * submitted async (out-of-line), then completions can come in while
3080 * we're waiting for refs to drop. We need to reap these manually,
3081 * as nobody else will be looking for them.
3082 */
b2edc0a7 3083 do {
a85381d8
PB
3084 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
3085 mutex_lock(&ctx->uring_lock);
3086 io_cqring_overflow_kill(ctx);
3087 mutex_unlock(&ctx->uring_lock);
3088 }
3089
c0e0d6ba
DY
3090 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3091 io_move_task_work_from_local(ctx);
3092
affa87db
PB
3093 while (io_uring_try_cancel_requests(ctx, NULL, true))
3094 cond_resched();
3095
28090c13
PB
3096 if (ctx->sq_data) {
3097 struct io_sq_data *sqd = ctx->sq_data;
3098 struct task_struct *tsk;
3099
3100 io_sq_thread_park(sqd);
3101 tsk = sqd->thread;
3102 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
3103 io_wq_cancel_cb(tsk->io_uring->io_wq,
3104 io_cancel_ctx_cb, ctx, true);
3105 io_sq_thread_unpark(sqd);
3106 }
b5bb3a24 3107
37f0e767
PB
3108 io_req_caches_free(ctx);
3109
58d3be2c
PB
3110 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
3111 /* there is little hope left, don't run it too often */
3112 interval = HZ * 60;
3113 }
4826c594
JA
3114 /*
3115 * This is really an uninterruptible wait, as it has to be
3116 * complete. But it's also run from a kworker, which doesn't
3117 * take signals, so it's fine to make it interruptible. This
3118 * avoids scenarios where we knowingly can wait much longer
3119 * on completions, for example if someone does a SIGSTOP on
3120 * a task that needs to finish task_work to make this loop
3121 * complete. That's a synthetic situation that should not
3122 * cause a stuck task backtrace, and hence a potential panic
3123 * on stuck tasks if that is enabled.
3124 */
3125 } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval));
d56d938b 3126
7f00651a
PB
3127 init_completion(&exit.completion);
3128 init_task_work(&exit.task_work, io_tctx_exit_cb);
3129 exit.ctx = ctx;
89b5066e
PB
3130 /*
3131 * Some may use context even when all refs and requests have been put,
3132 * and they are free to do so while still holding uring_lock or
5b0a6acc 3133 * completion_lock, see io_req_task_submit(). Apart from other work,
89b5066e
PB
3134 * this lock/unlock section also waits them to finish.
3135 */
d56d938b
PB
3136 mutex_lock(&ctx->uring_lock);
3137 while (!list_empty(&ctx->tctx_list)) {
b5bb3a24
PB
3138 WARN_ON_ONCE(time_after(jiffies, timeout));
3139
d56d938b
PB
3140 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
3141 ctx_node);
7f00651a
PB
3142 /* don't spin on a single task if cancellation failed */
3143 list_rotate_left(&ctx->tctx_list);
d56d938b
PB
3144 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
3145 if (WARN_ON_ONCE(ret))
3146 continue;
d56d938b
PB
3147
3148 mutex_unlock(&ctx->uring_lock);
4826c594
JA
3149 /*
3150 * See comment above for
3151 * wait_for_completion_interruptible_timeout() on why this
3152 * wait is marked as interruptible.
3153 */
3154 wait_for_completion_interruptible(&exit.completion);
d56d938b
PB
3155 mutex_lock(&ctx->uring_lock);
3156 }
3157 mutex_unlock(&ctx->uring_lock);
79ebeaee
JA
3158 spin_lock(&ctx->completion_lock);
3159 spin_unlock(&ctx->completion_lock);
d56d938b 3160
d73a572d
PB
3161 /* pairs with RCU read section in io_req_local_work_add() */
3162 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3163 synchronize_rcu();
3164
85faa7b8
JA
3165 io_ring_ctx_free(ctx);
3166}
3167
c072481d 3168static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2b188cc1 3169{
61cf9370
MWO
3170 unsigned long index;
3171 struct creds *creds;
3172
2b188cc1
JA
3173 mutex_lock(&ctx->uring_lock);
3174 percpu_ref_kill(&ctx->refs);
61cf9370
MWO
3175 xa_for_each(&ctx->personalities, index, creds)
3176 io_unregister_personality(ctx, index);
9ca9fb24
PB
3177 if (ctx->rings)
3178 io_poll_remove_all(ctx, NULL, true);
2b188cc1
JA
3179 mutex_unlock(&ctx->uring_lock);
3180
02bac94b
PB
3181 /*
3182 * If we failed setting up the ctx, we might not have any rings
3183 * and therefore did not submit any requests
3184 */
3185 if (ctx->rings)
60053be8 3186 io_kill_timeouts(ctx, NULL, true);
309fc03a 3187
dfbe5561
JA
3188 flush_delayed_work(&ctx->fallback_work);
3189
85faa7b8 3190 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
fc666777
JA
3191 /*
3192 * Use system_unbound_wq to avoid spawning tons of event kworkers
3193 * if we're exiting a ton of rings at the same time. It just adds
3194 * noise and overhead, there's no discernable change in runtime
3195 * over using system_wq.
3196 */
3197 queue_work(system_unbound_wq, &ctx->exit_work);
2b188cc1
JA
3198}
3199
3200static int io_uring_release(struct inode *inode, struct file *file)
3201{
3202 struct io_ring_ctx *ctx = file->private_data;
3203
3204 file->private_data = NULL;
3205 io_ring_ctx_wait_and_kill(ctx);
3206 return 0;
3207}
3208
f6edbabb
PB
3209struct io_task_cancel {
3210 struct task_struct *task;
3dd0c97a 3211 bool all;
f6edbabb 3212};
f254ac04 3213
f6edbabb 3214static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
b711d4ea 3215{
9a472ef7 3216 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
f6edbabb 3217 struct io_task_cancel *cancel = data;
9a472ef7 3218
6af3f48b 3219 return io_match_task_safe(req, cancel->task, cancel->all);
b711d4ea
JA
3220}
3221
c072481d
PB
3222static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
3223 struct task_struct *task,
3224 bool cancel_all)
b7ddce3c 3225{
e1915f76 3226 struct io_defer_entry *de;
b7ddce3c
PB
3227 LIST_HEAD(list);
3228
79ebeaee 3229 spin_lock(&ctx->completion_lock);
b7ddce3c 3230 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
6af3f48b 3231 if (io_match_task_safe(de->req, task, cancel_all)) {
b7ddce3c
PB
3232 list_cut_position(&list, &ctx->defer_list, &de->list);
3233 break;
3234 }
3235 }
79ebeaee 3236 spin_unlock(&ctx->completion_lock);
e1915f76
PB
3237 if (list_empty(&list))
3238 return false;
b7ddce3c
PB
3239
3240 while (!list_empty(&list)) {
3241 de = list_first_entry(&list, struct io_defer_entry, list);
3242 list_del_init(&de->list);
e276ae34 3243 io_req_task_queue_fail(de->req, -ECANCELED);
b7ddce3c
PB
3244 kfree(de);
3245 }
e1915f76 3246 return true;
b7ddce3c
PB
3247}
3248
c072481d 3249static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
1b00764f
PB
3250{
3251 struct io_tctx_node *node;
3252 enum io_wq_cancel cret;
3253 bool ret = false;
3254
3255 mutex_lock(&ctx->uring_lock);
3256 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3257 struct io_uring_task *tctx = node->task->io_uring;
3258
3259 /*
3260 * io_wq will stay alive while we hold uring_lock, because it's
3261 * killed after ctx nodes, which requires to take the lock.
3262 */
3263 if (!tctx || !tctx->io_wq)
3264 continue;
3265 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3266 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3267 }
3268 mutex_unlock(&ctx->uring_lock);
3269
3270 return ret;
3271}
3272
affa87db 3273static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
c072481d
PB
3274 struct task_struct *task,
3275 bool cancel_all)
9936c7c2 3276{
3dd0c97a 3277 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
1b00764f 3278 struct io_uring_task *tctx = task ? task->io_uring : NULL;
affa87db
PB
3279 enum io_wq_cancel cret;
3280 bool ret = false;
9936c7c2 3281
360cd42c
PB
3282 /* set it so io_req_local_work_add() would wake us up */
3283 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
3284 atomic_set(&ctx->cq_wait_nr, 1);
3285 smp_mb();
3286 }
3287
60053be8
PB
3288 /* failed during ring init, it couldn't have issued any requests */
3289 if (!ctx->rings)
affa87db 3290 return false;
60053be8 3291
affa87db
PB
3292 if (!task) {
3293 ret |= io_uring_try_cancel_iowq(ctx);
3294 } else if (tctx && tctx->io_wq) {
3295 /*
3296 * Cancels requests of all rings, not only @ctx, but
3297 * it's fine as the task is in exit/exec.
3298 */
3299 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3300 &cancel, true);
3301 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3302 }
9936c7c2 3303
affa87db
PB
3304 /* SQPOLL thread does its own polling */
3305 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3306 (ctx->sq_data && ctx->sq_data->thread == current)) {
3307 while (!wq_list_empty(&ctx->iopoll_list)) {
3308 io_iopoll_try_reap_events(ctx);
3309 ret = true;
fcc926bb 3310 cond_resched();
9936c7c2 3311 }
9936c7c2 3312 }
affa87db 3313
140102ae
PB
3314 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3315 io_allowed_defer_tw_run(ctx))
c0e0d6ba 3316 ret |= io_run_local_work(ctx) > 0;
affa87db
PB
3317 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3318 mutex_lock(&ctx->uring_lock);
3319 ret |= io_poll_remove_all(ctx, task, cancel_all);
3320 mutex_unlock(&ctx->uring_lock);
3321 ret |= io_kill_timeouts(ctx, task, cancel_all);
3322 if (task)
c0e0d6ba 3323 ret |= io_run_task_work() > 0;
affa87db 3324 return ret;
9936c7c2
PB
3325}
3326
3f48cf18 3327static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
521d6a73 3328{
3f48cf18 3329 if (tracked)
9cae36a0 3330 return atomic_read(&tctx->inflight_tracked);
521d6a73
PB
3331 return percpu_counter_sum(&tctx->inflight);
3332}
3333
78cc687b
PB
3334/*
3335 * Find any io_uring ctx that this task has registered or done IO on, and cancel
78a78060 3336 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
78cc687b 3337 */
17437f31 3338__cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
0e9ddb39 3339{
521d6a73 3340 struct io_uring_task *tctx = current->io_uring;
734551df 3341 struct io_ring_ctx *ctx;
360cd42c
PB
3342 struct io_tctx_node *node;
3343 unsigned long index;
0e9ddb39
PB
3344 s64 inflight;
3345 DEFINE_WAIT(wait);
fdaf083c 3346
78cc687b
PB
3347 WARN_ON_ONCE(sqd && sqd->thread != current);
3348
6d042ffb
PO
3349 if (!current->io_uring)
3350 return;
17a91051
PB
3351 if (tctx->io_wq)
3352 io_wq_exit_start(tctx->io_wq);
3353
8d664282 3354 atomic_inc(&tctx->in_cancel);
0e9ddb39 3355 do {
affa87db
PB
3356 bool loop = false;
3357
e9dbe221 3358 io_uring_drop_tctx_refs(current);
0e9ddb39 3359 /* read completions before cancelations */
78cc687b 3360 inflight = tctx_inflight(tctx, !cancel_all);
0e9ddb39
PB
3361 if (!inflight)
3362 break;
fdaf083c 3363
78cc687b 3364 if (!sqd) {
78cc687b
PB
3365 xa_for_each(&tctx->xa, index, node) {
3366 /* sqpoll task will cancel all its requests */
3367 if (node->ctx->sq_data)
3368 continue;
affa87db
PB
3369 loop |= io_uring_try_cancel_requests(node->ctx,
3370 current, cancel_all);
78cc687b
PB
3371 }
3372 } else {
3373 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
affa87db
PB
3374 loop |= io_uring_try_cancel_requests(ctx,
3375 current,
3376 cancel_all);
3377 }
3378
3379 if (loop) {
3380 cond_resched();
3381 continue;
78cc687b 3382 }
17a91051 3383
78a78060
JA
3384 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3385 io_run_task_work();
e9dbe221 3386 io_uring_drop_tctx_refs(current);
360cd42c
PB
3387 xa_for_each(&tctx->xa, index, node) {
3388 if (!llist_empty(&node->ctx->work_llist)) {
3389 WARN_ON_ONCE(node->ctx->submitter_task &&
3390 node->ctx->submitter_task != current);
3391 goto end_wait;
3392 }
3393 }
0f212204 3394 /*
a1bb3cd5
PB
3395 * If we've seen completions, retry without waiting. This
3396 * avoids a race where a completion comes in before we did
3397 * prepare_to_wait().
0f212204 3398 */
3dd0c97a 3399 if (inflight == tctx_inflight(tctx, !cancel_all))
a1bb3cd5 3400 schedule();
360cd42c 3401end_wait:
f57555ed 3402 finish_wait(&tctx->wait, &wait);
d8a6df10 3403 } while (1);
de7f1d9e 3404
8452d4a6 3405 io_uring_clean_tctx(tctx);
3dd0c97a 3406 if (cancel_all) {
3cc7fdb9
PB
3407 /*
3408 * We shouldn't run task_works after cancel, so just leave
8d664282 3409 * ->in_cancel set for normal exit.
3cc7fdb9 3410 */
8d664282 3411 atomic_dec(&tctx->in_cancel);
3f48cf18
PB
3412 /* for exec all current's requests should be gone, kill tctx */
3413 __io_uring_free(current);
3414 }
44e728b8
PB
3415}
3416
f552a27a 3417void __io_uring_cancel(bool cancel_all)
78cc687b 3418{
f552a27a 3419 io_uring_cancel_generic(cancel_all, NULL);
78cc687b
PB
3420}
3421
6c5c240e
RP
3422static void *io_uring_validate_mmap_request(struct file *file,
3423 loff_t pgoff, size_t sz)
2b188cc1 3424{
2b188cc1 3425 struct io_ring_ctx *ctx = file->private_data;
6c5c240e 3426 loff_t offset = pgoff << PAGE_SHIFT;
2b188cc1
JA
3427 struct page *page;
3428 void *ptr;
3429
03d89a2d
JA
3430 /* Don't allow mmap if the ring was setup without it */
3431 if (ctx->flags & IORING_SETUP_NO_MMAP)
3432 return ERR_PTR(-EINVAL);
3433
c56e022c 3434 switch (offset & IORING_OFF_MMAP_MASK) {
2b188cc1 3435 case IORING_OFF_SQ_RING:
75b28aff
HV
3436 case IORING_OFF_CQ_RING:
3437 ptr = ctx->rings;
2b188cc1
JA
3438 break;
3439 case IORING_OFF_SQES:
3440 ptr = ctx->sq_sqes;
3441 break;
c56e022c
JA
3442 case IORING_OFF_PBUF_RING: {
3443 unsigned int bgid;
3444
3445 bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT;
3446 mutex_lock(&ctx->uring_lock);
3447 ptr = io_pbuf_get_address(ctx, bgid);
3448 mutex_unlock(&ctx->uring_lock);
3449 if (!ptr)
3450 return ERR_PTR(-EINVAL);
3451 break;
3452 }
2b188cc1 3453 default:
6c5c240e 3454 return ERR_PTR(-EINVAL);
2b188cc1
JA
3455 }
3456
3457 page = virt_to_head_page(ptr);
a50b854e 3458 if (sz > page_size(page))
6c5c240e
RP
3459 return ERR_PTR(-EINVAL);
3460
3461 return ptr;
3462}
3463
3464#ifdef CONFIG_MMU
3465
c072481d 3466static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6c5c240e
RP
3467{
3468 size_t sz = vma->vm_end - vma->vm_start;
3469 unsigned long pfn;
3470 void *ptr;
3471
3472 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3473 if (IS_ERR(ptr))
3474 return PTR_ERR(ptr);
2b188cc1
JA
3475
3476 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3477 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3478}
3479
d808459b
HD
3480static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
3481 unsigned long addr, unsigned long len,
3482 unsigned long pgoff, unsigned long flags)
3483{
d808459b
HD
3484 void *ptr;
3485
3486 /*
3487 * Do not allow to map to user-provided address to avoid breaking the
3488 * aliasing rules. Userspace is not able to guess the offset address of
3489 * kernel kmalloc()ed memory area.
3490 */
3491 if (addr)
3492 return -EINVAL;
3493
3494 ptr = io_uring_validate_mmap_request(filp, pgoff, len);
3495 if (IS_ERR(ptr))
3496 return -ENOMEM;
3497
32832a40
HD
3498 /*
3499 * Some architectures have strong cache aliasing requirements.
3500 * For such architectures we need a coherent mapping which aliases
3501 * kernel memory *and* userspace memory. To achieve that:
3502 * - use a NULL file pointer to reference physical memory, and
3503 * - use the kernel virtual address of the shared io_uring context
3504 * (instead of the userspace-provided address, which has to be 0UL
3505 * anyway).
56675f8b
HD
3506 * - use the same pgoff which the get_unmapped_area() uses to
3507 * calculate the page colouring.
32832a40
HD
3508 * For architectures without such aliasing requirements, the
3509 * architecture will return any suitable mapping because addr is 0.
3510 */
3511 filp = NULL;
3512 flags |= MAP_SHARED;
3513 pgoff = 0; /* has been translated to ptr above */
d808459b 3514#ifdef SHM_COLOUR
32832a40 3515 addr = (uintptr_t) ptr;
56675f8b 3516 pgoff = addr >> PAGE_SHIFT;
d808459b 3517#else
32832a40 3518 addr = 0UL;
d808459b 3519#endif
32832a40 3520 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
d808459b
HD
3521}
3522
6c5c240e
RP
3523#else /* !CONFIG_MMU */
3524
3525static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3526{
fc4f4be9 3527 return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL;
6c5c240e
RP
3528}
3529
3530static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3531{
3532 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3533}
3534
3535static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3536 unsigned long addr, unsigned long len,
3537 unsigned long pgoff, unsigned long flags)
3538{
3539 void *ptr;
3540
3541 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3542 if (IS_ERR(ptr))
3543 return PTR_ERR(ptr);
3544
3545 return (unsigned long) ptr;
3546}
3547
3548#endif /* !CONFIG_MMU */
3549
f81440d3
PB
3550static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3551{
3552 if (flags & IORING_ENTER_EXT_ARG) {
3553 struct io_uring_getevents_arg arg;
3554
3555 if (argsz != sizeof(arg))
3556 return -EINVAL;
3557 if (copy_from_user(&arg, argp, sizeof(arg)))
3558 return -EFAULT;
3559 }
3560 return 0;
3561}
3562
c73ebb68
HX
3563static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3564 struct __kernel_timespec __user **ts,
3565 const sigset_t __user **sig)
3566{
3567 struct io_uring_getevents_arg arg;
3568
3569 /*
3570 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3571 * is just a pointer to the sigset_t.
3572 */
3573 if (!(flags & IORING_ENTER_EXT_ARG)) {
3574 *sig = (const sigset_t __user *) argp;
3575 *ts = NULL;
3576 return 0;
3577 }
3578
3579 /*
3580 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3581 * timespec and sigset_t pointers if good.
3582 */
3583 if (*argsz != sizeof(arg))
3584 return -EINVAL;
3585 if (copy_from_user(&arg, argp, sizeof(arg)))
3586 return -EFAULT;
d2347b96
DY
3587 if (arg.pad)
3588 return -EINVAL;
c73ebb68
HX
3589 *sig = u64_to_user_ptr(arg.sigmask);
3590 *argsz = arg.sigmask_sz;
3591 *ts = u64_to_user_ptr(arg.ts);
3592 return 0;
3593}
3594
2b188cc1 3595SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
c73ebb68
HX
3596 u32, min_complete, u32, flags, const void __user *, argp,
3597 size_t, argsz)
2b188cc1
JA
3598{
3599 struct io_ring_ctx *ctx;
2b188cc1 3600 struct fd f;
33f993da 3601 long ret;
2b188cc1 3602
33f993da 3603 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
e7a6c00d
JA
3604 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3605 IORING_ENTER_REGISTERED_RING)))
2b188cc1
JA
3606 return -EINVAL;
3607
e7a6c00d
JA
3608 /*
3609 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3610 * need only dereference our task private array to find it.
3611 */
3612 if (flags & IORING_ENTER_REGISTERED_RING) {
3613 struct io_uring_task *tctx = current->io_uring;
3614
3273c440 3615 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
e7a6c00d
JA
3616 return -EINVAL;
3617 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3618 f.file = tctx->registered_rings[fd];
4329490a 3619 f.flags = 0;
3273c440
PB
3620 if (unlikely(!f.file))
3621 return -EBADF;
e7a6c00d
JA
3622 } else {
3623 f = fdget(fd);
3273c440
PB
3624 if (unlikely(!f.file))
3625 return -EBADF;
3626 ret = -EOPNOTSUPP;
3627 if (unlikely(!io_is_uring_fops(f.file)))
fbb8bb02 3628 goto out;
e7a6c00d 3629 }
2b188cc1 3630
2b188cc1 3631 ctx = f.file->private_data;
7e84e1c7 3632 ret = -EBADFD;
33f993da 3633 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
7e84e1c7
SG
3634 goto out;
3635
6c271ce2
JA
3636 /*
3637 * For SQ polling, the thread will do all submissions and completions.
3638 * Just return the requested submit count, and wake the thread if
3639 * we were asked to.
3640 */
b2a9eada 3641 ret = 0;
6c271ce2 3642 if (ctx->flags & IORING_SETUP_SQPOLL) {
90f67366 3643 io_cqring_overflow_flush(ctx);
89448c47 3644
21f96522
JA
3645 if (unlikely(ctx->sq_data->thread == NULL)) {
3646 ret = -EOWNERDEAD;
04147488 3647 goto out;
21f96522 3648 }
6c271ce2 3649 if (flags & IORING_ENTER_SQ_WAKEUP)
534ca6d6 3650 wake_up(&ctx->sq_data->wait);
88b80534
QF
3651 if (flags & IORING_ENTER_SQ_WAIT)
3652 io_sqpoll_wait_sq(ctx);
3653
3e813c90 3654 ret = to_submit;
b2a9eada 3655 } else if (to_submit) {
eef51daa 3656 ret = io_uring_add_tctx_node(ctx);
0f212204
JA
3657 if (unlikely(ret))
3658 goto out;
7c504e65 3659
2b188cc1 3660 mutex_lock(&ctx->uring_lock);
3e813c90
DY
3661 ret = io_submit_sqes(ctx, to_submit);
3662 if (ret != to_submit) {
d487b43c 3663 mutex_unlock(&ctx->uring_lock);
7c504e65 3664 goto out;
d487b43c 3665 }
44f87745
PB
3666 if (flags & IORING_ENTER_GETEVENTS) {
3667 if (ctx->syscall_iopoll)
3668 goto iopoll_locked;
3669 /*
3670 * Ignore errors, we'll soon call io_cqring_wait() and
3671 * it should handle ownership problems if any.
3672 */
3673 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3674 (void)io_run_local_work_locked(ctx);
3675 }
d487b43c 3676 mutex_unlock(&ctx->uring_lock);
2b188cc1 3677 }
c0e0d6ba 3678
2b188cc1 3679 if (flags & IORING_ENTER_GETEVENTS) {
3e813c90 3680 int ret2;
c0e0d6ba 3681
773697b6 3682 if (ctx->syscall_iopoll) {
d487b43c
PB
3683 /*
3684 * We disallow the app entering submit/complete with
3685 * polling, but we still need to lock the ring to
3686 * prevent racing with polled issue that got punted to
3687 * a workqueue.
3688 */
3689 mutex_lock(&ctx->uring_lock);
3690iopoll_locked:
3e813c90
DY
3691 ret2 = io_validate_ext_arg(flags, argp, argsz);
3692 if (likely(!ret2)) {
3693 min_complete = min(min_complete,
3694 ctx->cq_entries);
3695 ret2 = io_iopoll_check(ctx, min_complete);
d487b43c
PB
3696 }
3697 mutex_unlock(&ctx->uring_lock);
def596e9 3698 } else {
f81440d3
PB
3699 const sigset_t __user *sig;
3700 struct __kernel_timespec __user *ts;
3701
3e813c90
DY
3702 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3703 if (likely(!ret2)) {
3704 min_complete = min(min_complete,
3705 ctx->cq_entries);
3706 ret2 = io_cqring_wait(ctx, min_complete, sig,
3707 argsz, ts);
3708 }
def596e9 3709 }
c73ebb68 3710
155bc950 3711 if (!ret) {
3e813c90 3712 ret = ret2;
2b188cc1 3713
155bc950
DY
3714 /*
3715 * EBADR indicates that one or more CQE were dropped.
3716 * Once the user has been informed we can clear the bit
3717 * as they are obviously ok with those drops.
3718 */
3719 if (unlikely(ret2 == -EBADR))
3720 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3721 &ctx->check_cq);
def596e9 3722 }
2b188cc1 3723 }
7c504e65 3724out:
4329490a 3725 fdput(f);
3e813c90 3726 return ret;
2b188cc1
JA
3727}
3728
3729static const struct file_operations io_uring_fops = {
3730 .release = io_uring_release,
3731 .mmap = io_uring_mmap,
6c5c240e
RP
3732#ifndef CONFIG_MMU
3733 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3734 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
d808459b
HD
3735#else
3736 .get_unmapped_area = io_uring_mmu_get_unmapped_area,
6c5c240e 3737#endif
2b188cc1 3738 .poll = io_uring_poll,
bebdb65e 3739#ifdef CONFIG_PROC_FS
87ce955b 3740 .show_fdinfo = io_uring_show_fdinfo,
bebdb65e 3741#endif
2b188cc1
JA
3742};
3743
92ac8bea
JA
3744bool io_is_uring_fops(struct file *file)
3745{
3746 return file->f_op == &io_uring_fops;
3747}
3748
c072481d
PB
3749static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3750 struct io_uring_params *p)
2b188cc1 3751{
75b28aff
HV
3752 struct io_rings *rings;
3753 size_t size, sq_array_offset;
e27cef86 3754 void *ptr;
2b188cc1 3755
bd740481
JA
3756 /* make sure these are sane, as we already accounted them */
3757 ctx->sq_entries = p->sq_entries;
3758 ctx->cq_entries = p->cq_entries;
3759
baf9cb64 3760 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
75b28aff
HV
3761 if (size == SIZE_MAX)
3762 return -EOVERFLOW;
3763
03d89a2d
JA
3764 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3765 rings = io_mem_alloc(size);
3766 else
3767 rings = io_rings_map(ctx, p->cq_off.user_addr, size);
3768
e27cef86
JA
3769 if (IS_ERR(rings))
3770 return PTR_ERR(rings);
2b188cc1 3771
75b28aff 3772 ctx->rings = rings;
2af89abd
PB
3773 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3774 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
75b28aff
HV
3775 rings->sq_ring_mask = p->sq_entries - 1;
3776 rings->cq_ring_mask = p->cq_entries - 1;
3777 rings->sq_ring_entries = p->sq_entries;
3778 rings->cq_ring_entries = p->cq_entries;
2b188cc1 3779
ebdeb7c0
JA
3780 if (p->flags & IORING_SETUP_SQE128)
3781 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3782 else
3783 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
eb065d30 3784 if (size == SIZE_MAX) {
9c189eee 3785 io_rings_free(ctx);
2b188cc1 3786 return -EOVERFLOW;
eb065d30 3787 }
2b188cc1 3788
03d89a2d
JA
3789 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3790 ptr = io_mem_alloc(size);
3791 else
3792 ptr = io_sqes_map(ctx, p->sq_off.user_addr, size);
3793
e27cef86 3794 if (IS_ERR(ptr)) {
9c189eee 3795 io_rings_free(ctx);
e27cef86 3796 return PTR_ERR(ptr);
eb065d30 3797 }
2b188cc1 3798
e27cef86 3799 ctx->sq_sqes = ptr;
2b188cc1
JA
3800 return 0;
3801}
3802
6e76ac59 3803static int io_uring_install_fd(struct file *file)
9faadcc8 3804{
6e76ac59 3805 int fd;
9faadcc8
PB
3806
3807 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3808 if (fd < 0)
3809 return fd;
9faadcc8
PB
3810 fd_install(fd, file);
3811 return fd;
3812}
3813
2b188cc1
JA
3814/*
3815 * Allocate an anonymous fd, this is what constitutes the application
3816 * visible backing of an io_uring instance. The application mmaps this
3817 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3818 * we have to tie this fd to a socket for file garbage collection purposes.
3819 */
9faadcc8 3820static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
2b188cc1
JA
3821{
3822 struct file *file;
9faadcc8 3823#if defined(CONFIG_UNIX)
2b188cc1
JA
3824 int ret;
3825
2b188cc1
JA
3826 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3827 &ctx->ring_sock);
3828 if (ret)
9faadcc8 3829 return ERR_PTR(ret);
2b188cc1
JA
3830#endif
3831
91a9ab7c
PM
3832 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3833 O_RDWR | O_CLOEXEC, NULL);
2b188cc1 3834#if defined(CONFIG_UNIX)
9faadcc8
PB
3835 if (IS_ERR(file)) {
3836 sock_release(ctx->ring_sock);
3837 ctx->ring_sock = NULL;
3838 } else {
3839 ctx->ring_sock->file = file;
0f212204 3840 }
2b188cc1 3841#endif
9faadcc8 3842 return file;
2b188cc1
JA
3843}
3844
c072481d
PB
3845static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3846 struct io_uring_params __user *params)
2b188cc1 3847{
2b188cc1 3848 struct io_ring_ctx *ctx;
6e76ac59 3849 struct io_uring_task *tctx;
9faadcc8 3850 struct file *file;
2b188cc1
JA
3851 int ret;
3852
8110c1a6 3853 if (!entries)
2b188cc1 3854 return -EINVAL;
8110c1a6
JA
3855 if (entries > IORING_MAX_ENTRIES) {
3856 if (!(p->flags & IORING_SETUP_CLAMP))
3857 return -EINVAL;
3858 entries = IORING_MAX_ENTRIES;
3859 }
2b188cc1 3860
6e76ac59
JT
3861 if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
3862 && !(p->flags & IORING_SETUP_NO_MMAP))
3863 return -EINVAL;
3864
2b188cc1
JA
3865 /*
3866 * Use twice as many entries for the CQ ring. It's possible for the
3867 * application to drive a higher depth than the size of the SQ ring,
3868 * since the sqes are only used at submission time. This allows for
33a107f0
JA
3869 * some flexibility in overcommitting a bit. If the application has
3870 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3871 * of CQ ring entries manually.
2b188cc1
JA
3872 */
3873 p->sq_entries = roundup_pow_of_two(entries);
33a107f0
JA
3874 if (p->flags & IORING_SETUP_CQSIZE) {
3875 /*
3876 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3877 * to a power-of-two, if it isn't already. We do NOT impose
3878 * any cq vs sq ring sizing.
3879 */
eb2667b3 3880 if (!p->cq_entries)
33a107f0 3881 return -EINVAL;
8110c1a6
JA
3882 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3883 if (!(p->flags & IORING_SETUP_CLAMP))
3884 return -EINVAL;
3885 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3886 }
eb2667b3
JQ
3887 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3888 if (p->cq_entries < p->sq_entries)
3889 return -EINVAL;
33a107f0
JA
3890 } else {
3891 p->cq_entries = 2 * p->sq_entries;
3892 }
2b188cc1 3893
2b188cc1 3894 ctx = io_ring_ctx_alloc(p);
62e398be 3895 if (!ctx)
2b188cc1 3896 return -ENOMEM;
773697b6 3897
e6aeb272
PB
3898 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3899 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3900 !(ctx->flags & IORING_SETUP_SQPOLL))
3901 ctx->task_complete = true;
3902
ec26c225
PB
3903 if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL))
3904 ctx->lockless_cq = true;
3905
bca39f39
PB
3906 /*
3907 * lazy poll_wq activation relies on ->task_complete for synchronisation
3908 * purposes, see io_activate_pollwq()
3909 */
3910 if (!ctx->task_complete)
3911 ctx->poll_activated = true;
3912
773697b6
PB
3913 /*
3914 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3915 * space applications don't need to do io completion events
3916 * polling again, they can rely on io_sq_thread to do polling
3917 * work, which can reduce cpu usage and uring_lock contention.
3918 */
3919 if (ctx->flags & IORING_SETUP_IOPOLL &&
3920 !(ctx->flags & IORING_SETUP_SQPOLL))
3921 ctx->syscall_iopoll = 1;
3922
2b188cc1 3923 ctx->compat = in_compat_syscall();
6adc2272 3924 if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK))
62e398be 3925 ctx->user = get_uid(current_user());
2aede0e4 3926
9f010507 3927 /*
e1169f06
JA
3928 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3929 * COOP_TASKRUN is set, then IPIs are never needed by the app.
9f010507 3930 */
e1169f06
JA
3931 ret = -EINVAL;
3932 if (ctx->flags & IORING_SETUP_SQPOLL) {
3933 /* IPI related flags don't make sense with SQPOLL */
ef060ea9 3934 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
c0e0d6ba
DY
3935 IORING_SETUP_TASKRUN_FLAG |
3936 IORING_SETUP_DEFER_TASKRUN))
e1169f06 3937 goto err;
9f010507 3938 ctx->notify_method = TWA_SIGNAL_NO_IPI;
e1169f06
JA
3939 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3940 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3941 } else {
c0e0d6ba
DY
3942 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3943 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
ef060ea9 3944 goto err;
9f010507 3945 ctx->notify_method = TWA_SIGNAL;
e1169f06 3946 }
9f010507 3947
c0e0d6ba
DY
3948 /*
3949 * For DEFER_TASKRUN we require the completion task to be the same as the
3950 * submission task. This implies that there is only one submitter, so enforce
3951 * that.
3952 */
3953 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3954 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3955 goto err;
3956 }
3957
2aede0e4
JA
3958 /*
3959 * This is just grabbed for accounting purposes. When a process exits,
3960 * the mm is exited and dropped before the files, hence we need to hang
3961 * on to this mm purely for the purposes of being able to unaccount
3962 * memory (locked/pinned vm). It's not used for anything else.
3963 */
6b7898eb 3964 mmgrab(current->mm);
2aede0e4 3965 ctx->mm_account = current->mm;
6b7898eb 3966
2b188cc1
JA
3967 ret = io_allocate_scq_urings(ctx, p);
3968 if (ret)
3969 goto err;
3970
7e84e1c7 3971 ret = io_sq_offload_create(ctx, p);
2b188cc1
JA
3972 if (ret)
3973 goto err;
2933ae6e
PB
3974
3975 ret = io_rsrc_init(ctx);
47b228ce
PB
3976 if (ret)
3977 goto err;
2b188cc1 3978
75b28aff
HV
3979 p->sq_off.head = offsetof(struct io_rings, sq.head);
3980 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3981 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3982 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3983 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3984 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
2af89abd
PB
3985 if (!(ctx->flags & IORING_SETUP_NO_SQARRAY))
3986 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
9b1b58ca 3987 p->sq_off.resv1 = 0;
03d89a2d
JA
3988 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
3989 p->sq_off.user_addr = 0;
2b188cc1 3990
75b28aff
HV
3991 p->cq_off.head = offsetof(struct io_rings, cq.head);
3992 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3993 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3994 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3995 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3996 p->cq_off.cqes = offsetof(struct io_rings, cqes);
0d9b5b3a 3997 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
9b1b58ca 3998 p->cq_off.resv1 = 0;
03d89a2d
JA
3999 if (!(ctx->flags & IORING_SETUP_NO_MMAP))
4000 p->cq_off.user_addr = 0;
ac90f249 4001
7f13657d
XW
4002 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
4003 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
5769a351 4004 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
c73ebb68 4005 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9690557e 4006 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
c4212f3e 4007 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
7d3fd88d 4008 IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING;
7f13657d
XW
4009
4010 if (copy_to_user(params, p, sizeof(*p))) {
4011 ret = -EFAULT;
4012 goto err;
4013 }
d1719f70 4014
7cae596b
DY
4015 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
4016 && !(ctx->flags & IORING_SETUP_R_DISABLED))
8579538c 4017 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
7cae596b 4018
9faadcc8
PB
4019 file = io_uring_get_file(ctx);
4020 if (IS_ERR(file)) {
4021 ret = PTR_ERR(file);
4022 goto err;
4023 }
4024
6e76ac59
JT
4025 ret = __io_uring_add_tctx_node(ctx);
4026 if (ret)
4027 goto err_fput;
4028 tctx = current->io_uring;
4029
044c1ab3
JA
4030 /*
4031 * Install ring fd as the very last thing, so we don't risk someone
4032 * having closed it before we finish setup
4033 */
6e76ac59
JT
4034 if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY)
4035 ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX);
4036 else
4037 ret = io_uring_install_fd(file);
4038 if (ret < 0)
4039 goto err_fput;
044c1ab3 4040
c826bd7a 4041 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
2b188cc1
JA
4042 return ret;
4043err:
4044 io_ring_ctx_wait_and_kill(ctx);
4045 return ret;
6e76ac59
JT
4046err_fput:
4047 fput(file);
4048 return ret;
2b188cc1
JA
4049}
4050
4051/*
4052 * Sets up an aio uring context, and returns the fd. Applications asks for a
4053 * ring size, we return the actual sq/cq ring sizes (among other things) in the
4054 * params structure passed in.
4055 */
4056static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
4057{
4058 struct io_uring_params p;
2b188cc1
JA
4059 int i;
4060
4061 if (copy_from_user(&p, params, sizeof(p)))
4062 return -EFAULT;
4063 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
4064 if (p.resv[i])
4065 return -EINVAL;
4066 }
4067
6c271ce2 4068 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8110c1a6 4069 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7e84e1c7 4070 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
e1169f06 4071 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
ebdeb7c0 4072 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
97bbdc06 4073 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
03d89a2d 4074 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN |
2af89abd
PB
4075 IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY |
4076 IORING_SETUP_NO_SQARRAY))
2b188cc1
JA
4077 return -EINVAL;
4078
ef060ea9 4079 return io_uring_create(entries, &p, params);
2b188cc1
JA
4080}
4081
76d3ccec
MR
4082static inline bool io_uring_allowed(void)
4083{
4084 int disabled = READ_ONCE(sysctl_io_uring_disabled);
4085 kgid_t io_uring_group;
4086
4087 if (disabled == 2)
4088 return false;
4089
4090 if (disabled == 0 || capable(CAP_SYS_ADMIN))
4091 return true;
4092
4093 io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group);
4094 if (!gid_valid(io_uring_group))
4095 return false;
4096
4097 return in_group_p(io_uring_group);
4098}
4099
2b188cc1
JA
4100SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4101 struct io_uring_params __user *, params)
4102{
76d3ccec
MR
4103 if (!io_uring_allowed())
4104 return -EPERM;
4105
2b188cc1
JA
4106 return io_uring_setup(entries, params);
4107}
4108
c072481d
PB
4109static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
4110 unsigned nr_args)
66f4af93
JA
4111{
4112 struct io_uring_probe *p;
4113 size_t size;
4114 int i, ret;
4115
4116 size = struct_size(p, ops, nr_args);
4117 if (size == SIZE_MAX)
4118 return -EOVERFLOW;
4119 p = kzalloc(size, GFP_KERNEL);
4120 if (!p)
4121 return -ENOMEM;
4122
4123 ret = -EFAULT;
4124 if (copy_from_user(p, arg, size))
4125 goto out;
4126 ret = -EINVAL;
4127 if (memchr_inv(p, 0, size))
4128 goto out;
4129
4130 p->last_op = IORING_OP_LAST - 1;
4131 if (nr_args > IORING_OP_LAST)
4132 nr_args = IORING_OP_LAST;
4133
4134 for (i = 0; i < nr_args; i++) {
4135 p->ops[i].op = i;
a7dd2782 4136 if (!io_issue_defs[i].not_supported)
66f4af93
JA
4137 p->ops[i].flags = IO_URING_OP_SUPPORTED;
4138 }
4139 p->ops_len = i;
4140
4141 ret = 0;
4142 if (copy_to_user(arg, p, size))
4143 ret = -EFAULT;
4144out:
4145 kfree(p);
4146 return ret;
4147}
4148
071698e1
JA
4149static int io_register_personality(struct io_ring_ctx *ctx)
4150{
4379bf8b 4151 const struct cred *creds;
61cf9370 4152 u32 id;
1e6fa521 4153 int ret;
071698e1 4154
4379bf8b 4155 creds = get_current_cred();
1e6fa521 4156
61cf9370
MWO
4157 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
4158 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
a30f895a
JA
4159 if (ret < 0) {
4160 put_cred(creds);
4161 return ret;
4162 }
4163 return id;
071698e1
JA
4164}
4165
c072481d
PB
4166static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
4167 void __user *arg, unsigned int nr_args)
21b55dbc
SG
4168{
4169 struct io_uring_restriction *res;
4170 size_t size;
4171 int i, ret;
4172
7e84e1c7
SG
4173 /* Restrictions allowed only if rings started disabled */
4174 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4175 return -EBADFD;
4176
21b55dbc 4177 /* We allow only a single restrictions registration */
7e84e1c7 4178 if (ctx->restrictions.registered)
21b55dbc
SG
4179 return -EBUSY;
4180
4181 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
4182 return -EINVAL;
4183
4184 size = array_size(nr_args, sizeof(*res));
4185 if (size == SIZE_MAX)
4186 return -EOVERFLOW;
4187
4188 res = memdup_user(arg, size);
4189 if (IS_ERR(res))
4190 return PTR_ERR(res);
4191
4192 ret = 0;
4193
4194 for (i = 0; i < nr_args; i++) {
4195 switch (res[i].opcode) {
4196 case IORING_RESTRICTION_REGISTER_OP:
4197 if (res[i].register_op >= IORING_REGISTER_LAST) {
4198 ret = -EINVAL;
4199 goto out;
4200 }
4201
4202 __set_bit(res[i].register_op,
4203 ctx->restrictions.register_op);
4204 break;
4205 case IORING_RESTRICTION_SQE_OP:
4206 if (res[i].sqe_op >= IORING_OP_LAST) {
4207 ret = -EINVAL;
4208 goto out;
4209 }
4210
4211 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
4212 break;
4213 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
4214 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
4215 break;
4216 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
4217 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
4218 break;
4219 default:
4220 ret = -EINVAL;
4221 goto out;
4222 }
4223 }
4224
4225out:
4226 /* Reset all restrictions if an error happened */
4227 if (ret != 0)
4228 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
4229 else
7e84e1c7 4230 ctx->restrictions.registered = true;
21b55dbc
SG
4231
4232 kfree(res);
4233 return ret;
4234}
4235
7e84e1c7
SG
4236static int io_register_enable_rings(struct io_ring_ctx *ctx)
4237{
4238 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
4239 return -EBADFD;
4240
bca39f39 4241 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) {
8579538c 4242 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
bca39f39
PB
4243 /*
4244 * Lazy activation attempts would fail if it was polled before
4245 * submitter_task is set.
4246 */
4247 if (wq_has_sleeper(&ctx->poll_wq))
4248 io_activate_pollwq(ctx);
4249 }
7cae596b 4250
7e84e1c7
SG
4251 if (ctx->restrictions.registered)
4252 ctx->restricted = 1;
4253
0298ef96
PB
4254 ctx->flags &= ~IORING_SETUP_R_DISABLED;
4255 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
4256 wake_up(&ctx->sq_data->wait);
7e84e1c7
SG
4257 return 0;
4258}
4259
ebdfefc0
JA
4260static __cold int __io_register_iowq_aff(struct io_ring_ctx *ctx,
4261 cpumask_var_t new_mask)
4262{
4263 int ret;
4264
4265 if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
4266 ret = io_wq_cpu_affinity(current->io_uring, new_mask);
4267 } else {
4268 mutex_unlock(&ctx->uring_lock);
4269 ret = io_sqpoll_wq_cpu_affinity(ctx, new_mask);
4270 mutex_lock(&ctx->uring_lock);
4271 }
4272
4273 return ret;
4274}
4275
c072481d
PB
4276static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
4277 void __user *arg, unsigned len)
fe76421d 4278{
fe76421d
JA
4279 cpumask_var_t new_mask;
4280 int ret;
4281
fe76421d
JA
4282 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4283 return -ENOMEM;
4284
4285 cpumask_clear(new_mask);
4286 if (len > cpumask_size())
4287 len = cpumask_size();
4288
0f5e4b83
ES
4289 if (in_compat_syscall()) {
4290 ret = compat_get_bitmap(cpumask_bits(new_mask),
4291 (const compat_ulong_t __user *)arg,
4292 len * 8 /* CHAR_BIT */);
4293 } else {
4294 ret = copy_from_user(new_mask, arg, len);
4295 }
4296
4297 if (ret) {
fe76421d
JA
4298 free_cpumask_var(new_mask);
4299 return -EFAULT;
4300 }
4301
ebdfefc0 4302 ret = __io_register_iowq_aff(ctx, new_mask);
fe76421d
JA
4303 free_cpumask_var(new_mask);
4304 return ret;
4305}
4306
c072481d 4307static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
fe76421d 4308{
ebdfefc0 4309 return __io_register_iowq_aff(ctx, NULL);
fe76421d
JA
4310}
4311
c072481d
PB
4312static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
4313 void __user *arg)
b22fa62a 4314 __must_hold(&ctx->uring_lock)
2e480058 4315{
b22fa62a 4316 struct io_tctx_node *node;
fa84693b
JA
4317 struct io_uring_task *tctx = NULL;
4318 struct io_sq_data *sqd = NULL;
2e480058
JA
4319 __u32 new_count[2];
4320 int i, ret;
4321
2e480058
JA
4322 if (copy_from_user(new_count, arg, sizeof(new_count)))
4323 return -EFAULT;
4324 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4325 if (new_count[i] > INT_MAX)
4326 return -EINVAL;
4327
fa84693b
JA
4328 if (ctx->flags & IORING_SETUP_SQPOLL) {
4329 sqd = ctx->sq_data;
4330 if (sqd) {
009ad9f0
JA
4331 /*
4332 * Observe the correct sqd->lock -> ctx->uring_lock
4333 * ordering. Fine to drop uring_lock here, we hold
4334 * a ref to the ctx.
4335 */
41d3a6bd 4336 refcount_inc(&sqd->refs);
009ad9f0 4337 mutex_unlock(&ctx->uring_lock);
fa84693b 4338 mutex_lock(&sqd->lock);
009ad9f0 4339 mutex_lock(&ctx->uring_lock);
41d3a6bd
JA
4340 if (sqd->thread)
4341 tctx = sqd->thread->io_uring;
fa84693b
JA
4342 }
4343 } else {
4344 tctx = current->io_uring;
4345 }
4346
e139a1ec 4347 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
fa84693b 4348
bad119b9
PB
4349 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4350 if (new_count[i])
4351 ctx->iowq_limits[i] = new_count[i];
e139a1ec
PB
4352 ctx->iowq_limits_set = true;
4353
e139a1ec
PB
4354 if (tctx && tctx->io_wq) {
4355 ret = io_wq_max_workers(tctx->io_wq, new_count);
4356 if (ret)
4357 goto err;
4358 } else {
4359 memset(new_count, 0, sizeof(new_count));
4360 }
fa84693b 4361
41d3a6bd 4362 if (sqd) {
fa84693b 4363 mutex_unlock(&sqd->lock);
41d3a6bd
JA
4364 io_put_sq_data(sqd);
4365 }
2e480058
JA
4366
4367 if (copy_to_user(arg, new_count, sizeof(new_count)))
4368 return -EFAULT;
4369
b22fa62a
PB
4370 /* that's it for SQPOLL, only the SQPOLL task creates requests */
4371 if (sqd)
4372 return 0;
4373
4374 /* now propagate the restriction to all registered users */
4375 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
4376 struct io_uring_task *tctx = node->task->io_uring;
4377
4378 if (WARN_ON_ONCE(!tctx->io_wq))
4379 continue;
4380
4381 for (i = 0; i < ARRAY_SIZE(new_count); i++)
4382 new_count[i] = ctx->iowq_limits[i];
4383 /* ignore errors, it always returns zero anyway */
4384 (void)io_wq_max_workers(tctx->io_wq, new_count);
4385 }
2e480058 4386 return 0;
fa84693b 4387err:
41d3a6bd 4388 if (sqd) {
fa84693b 4389 mutex_unlock(&sqd->lock);
41d3a6bd
JA
4390 io_put_sq_data(sqd);
4391 }
fa84693b 4392 return ret;
2e480058
JA
4393}
4394
edafccee
JA
4395static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4396 void __user *arg, unsigned nr_args)
b19062a5
JA
4397 __releases(ctx->uring_lock)
4398 __acquires(ctx->uring_lock)
edafccee
JA
4399{
4400 int ret;
4401
35fa71a0 4402 /*
fbb8bb02
PB
4403 * We don't quiesce the refs for register anymore and so it can't be
4404 * dying as we're holding a file ref here.
35fa71a0 4405 */
fbb8bb02 4406 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
35fa71a0
JA
4407 return -ENXIO;
4408
d7cce96c
PB
4409 if (ctx->submitter_task && ctx->submitter_task != current)
4410 return -EEXIST;
4411
75c4021a 4412 if (ctx->restricted) {
75c4021a
PB
4413 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4414 if (!test_bit(opcode, ctx->restrictions.register_op))
4415 return -EACCES;
4416 }
4417
edafccee
JA
4418 switch (opcode) {
4419 case IORING_REGISTER_BUFFERS:
0184f08e
PB
4420 ret = -EFAULT;
4421 if (!arg)
4422 break;
634d00df 4423 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
edafccee
JA
4424 break;
4425 case IORING_UNREGISTER_BUFFERS:
4426 ret = -EINVAL;
4427 if (arg || nr_args)
4428 break;
0a96bbe4 4429 ret = io_sqe_buffers_unregister(ctx);
edafccee 4430 break;
6b06314c 4431 case IORING_REGISTER_FILES:
a8da73a3
JA
4432 ret = -EFAULT;
4433 if (!arg)
4434 break;
792e3582 4435 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
6b06314c
JA
4436 break;
4437 case IORING_UNREGISTER_FILES:
4438 ret = -EINVAL;
4439 if (arg || nr_args)
4440 break;
4441 ret = io_sqe_files_unregister(ctx);
4442 break;
c3a31e60 4443 case IORING_REGISTER_FILES_UPDATE:
c3bdad02 4444 ret = io_register_files_update(ctx, arg, nr_args);
c3a31e60 4445 break;
9b402849
JA
4446 case IORING_REGISTER_EVENTFD:
4447 ret = -EINVAL;
4448 if (nr_args != 1)
4449 break;
c75312dd
UA
4450 ret = io_eventfd_register(ctx, arg, 0);
4451 break;
4452 case IORING_REGISTER_EVENTFD_ASYNC:
4453 ret = -EINVAL;
4454 if (nr_args != 1)
f2842ab5 4455 break;
c75312dd 4456 ret = io_eventfd_register(ctx, arg, 1);
9b402849
JA
4457 break;
4458 case IORING_UNREGISTER_EVENTFD:
4459 ret = -EINVAL;
4460 if (arg || nr_args)
4461 break;
4462 ret = io_eventfd_unregister(ctx);
4463 break;
66f4af93
JA
4464 case IORING_REGISTER_PROBE:
4465 ret = -EINVAL;
4466 if (!arg || nr_args > 256)
4467 break;
4468 ret = io_probe(ctx, arg, nr_args);
4469 break;
071698e1
JA
4470 case IORING_REGISTER_PERSONALITY:
4471 ret = -EINVAL;
4472 if (arg || nr_args)
4473 break;
4474 ret = io_register_personality(ctx);
4475 break;
4476 case IORING_UNREGISTER_PERSONALITY:
4477 ret = -EINVAL;
4478 if (arg)
4479 break;
4480 ret = io_unregister_personality(ctx, nr_args);
4481 break;
7e84e1c7
SG
4482 case IORING_REGISTER_ENABLE_RINGS:
4483 ret = -EINVAL;
4484 if (arg || nr_args)
4485 break;
4486 ret = io_register_enable_rings(ctx);
4487 break;
21b55dbc
SG
4488 case IORING_REGISTER_RESTRICTIONS:
4489 ret = io_register_restrictions(ctx, arg, nr_args);
4490 break;
992da01a
PB
4491 case IORING_REGISTER_FILES2:
4492 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4493 break;
4494 case IORING_REGISTER_FILES_UPDATE2:
4495 ret = io_register_rsrc_update(ctx, arg, nr_args,
4496 IORING_RSRC_FILE);
4497 break;
4498 case IORING_REGISTER_BUFFERS2:
4499 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
792e3582 4500 break;
992da01a
PB
4501 case IORING_REGISTER_BUFFERS_UPDATE:
4502 ret = io_register_rsrc_update(ctx, arg, nr_args,
4503 IORING_RSRC_BUFFER);
c3bdad02 4504 break;
fe76421d
JA
4505 case IORING_REGISTER_IOWQ_AFF:
4506 ret = -EINVAL;
4507 if (!arg || !nr_args)
4508 break;
4509 ret = io_register_iowq_aff(ctx, arg, nr_args);
4510 break;
4511 case IORING_UNREGISTER_IOWQ_AFF:
4512 ret = -EINVAL;
4513 if (arg || nr_args)
4514 break;
4515 ret = io_unregister_iowq_aff(ctx);
4516 break;
2e480058
JA
4517 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4518 ret = -EINVAL;
4519 if (!arg || nr_args != 2)
4520 break;
4521 ret = io_register_iowq_max_workers(ctx, arg);
4522 break;
e7a6c00d
JA
4523 case IORING_REGISTER_RING_FDS:
4524 ret = io_ringfd_register(ctx, arg, nr_args);
4525 break;
4526 case IORING_UNREGISTER_RING_FDS:
4527 ret = io_ringfd_unregister(ctx, arg, nr_args);
4528 break;
c7fb1942
JA
4529 case IORING_REGISTER_PBUF_RING:
4530 ret = -EINVAL;
4531 if (!arg || nr_args != 1)
4532 break;
4533 ret = io_register_pbuf_ring(ctx, arg);
4534 break;
4535 case IORING_UNREGISTER_PBUF_RING:
4536 ret = -EINVAL;
4537 if (!arg || nr_args != 1)
4538 break;
4539 ret = io_unregister_pbuf_ring(ctx, arg);
4540 break;
78a861b9
JA
4541 case IORING_REGISTER_SYNC_CANCEL:
4542 ret = -EINVAL;
4543 if (!arg || nr_args != 1)
4544 break;
4545 ret = io_sync_cancel(ctx, arg);
4546 break;
6e73dffb
PB
4547 case IORING_REGISTER_FILE_ALLOC_RANGE:
4548 ret = -EINVAL;
4549 if (!arg || nr_args)
4550 break;
4551 ret = io_register_file_alloc_range(ctx, arg);
4552 break;
edafccee
JA
4553 default:
4554 ret = -EINVAL;
4555 break;
4556 }
4557
edafccee
JA
4558 return ret;
4559}
4560
4561SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4562 void __user *, arg, unsigned int, nr_args)
4563{
4564 struct io_ring_ctx *ctx;
4565 long ret = -EBADF;
4566 struct fd f;
7d3fd88d
JT
4567 bool use_registered_ring;
4568
4569 use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING);
4570 opcode &= ~IORING_REGISTER_USE_REGISTERED_RING;
edafccee 4571
34319084
JA
4572 if (opcode >= IORING_REGISTER_LAST)
4573 return -EINVAL;
4574
7d3fd88d
JT
4575 if (use_registered_ring) {
4576 /*
4577 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
4578 * need only dereference our task private array to find it.
4579 */
4580 struct io_uring_task *tctx = current->io_uring;
edafccee 4581
7d3fd88d
JT
4582 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
4583 return -EINVAL;
4584 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
4585 f.file = tctx->registered_rings[fd];
4586 f.flags = 0;
4587 if (unlikely(!f.file))
4588 return -EBADF;
4589 } else {
4590 f = fdget(fd);
4591 if (unlikely(!f.file))
4592 return -EBADF;
4593 ret = -EOPNOTSUPP;
4594 if (!io_is_uring_fops(f.file))
4595 goto out_fput;
4596 }
edafccee
JA
4597
4598 ctx = f.file->private_data;
4599
4600 mutex_lock(&ctx->uring_lock);
4601 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4602 mutex_unlock(&ctx->uring_lock);
2757be22 4603 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
edafccee
JA
4604out_fput:
4605 fdput(f);
4606 return ret;
4607}
4608
2b188cc1
JA
4609static int __init io_uring_init(void)
4610{
9c71d39a 4611#define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
d7f62e82 4612 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9c71d39a 4613 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
d7f62e82
SM
4614} while (0)
4615
4616#define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9c71d39a
SM
4617 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4618#define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4619 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
d7f62e82
SM
4620 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4621 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4622 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4623 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4624 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4625 BUILD_BUG_SQE_ELEM(8, __u64, off);
4626 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9c71d39a
SM
4627 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4628 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
d7f62e82 4629 BUILD_BUG_SQE_ELEM(16, __u64, addr);
7d67af2c 4630 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
d7f62e82
SM
4631 BUILD_BUG_SQE_ELEM(24, __u32, len);
4632 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4633 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4634 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4635 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
5769a351
JX
4636 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4637 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
d7f62e82
SM
4638 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4639 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4640 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4641 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4642 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4643 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4644 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4645 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
7d67af2c 4646 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9c71d39a
SM
4647 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4648 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4649 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4650 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4651 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
d7f62e82
SM
4652 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4653 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
16340eab 4654 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
d7f62e82 4655 BUILD_BUG_SQE_ELEM(42, __u16, personality);
7d67af2c 4656 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
b9445598 4657 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
b48c312b
PB
4658 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4659 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
e9621e2b 4660 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
9c71d39a
SM
4661 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4662 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
d7f62e82 4663
b0d658ec
PB
4664 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4665 sizeof(struct io_uring_rsrc_update));
4666 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4667 sizeof(struct io_uring_rsrc_update2));
90499ad0
PB
4668
4669 /* ->buf_index is u16 */
c7fb1942
JA
4670 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4671 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4672 offsetof(struct io_uring_buf_ring, tail));
90499ad0 4673
b0d658ec
PB
4674 /* should fit into one byte */
4675 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
68fe256a
PB
4676 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4677 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
b0d658ec 4678
32c2d33e 4679 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
16340eab 4680
3a4b89a2
JA
4681 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4682
d9b57aa3 4683 io_uring_optable_init();
0702e536 4684
b97f96e2
JA
4685 /*
4686 * Allow user copy in the per-command field, which starts after the
4687 * file in io_kiocb and until the opcode field. The openat2 handling
4688 * requires copying in user memory into the io_kiocb object in that
4689 * range, and HARDENED_USERCOPY will complain if we haven't
4690 * correctly annotated this range.
4691 */
4692 req_cachep = kmem_cache_create_usercopy("io_kiocb",
4693 sizeof(struct io_kiocb), 0,
4694 SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4695 SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU,
4696 offsetof(struct io_kiocb, cmd.data),
4697 sizeof_field(struct io_kiocb, cmd.data), NULL);
4698
76d3ccec
MR
4699#ifdef CONFIG_SYSCTL
4700 register_sysctl_init("kernel", kernel_io_uring_disabled_table);
4701#endif
4702
2b188cc1
JA
4703 return 0;
4704};
4705__initcall(io_uring_init);