io_uring: rearrange defer list checks
[linux-block.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>
2b188cc1 75
c826bd7a
DD
76#define CREATE_TRACE_POINTS
77#include <trace/events/io_uring.h>
78
2b188cc1
JA
79#include <uapi/linux/io_uring.h>
80
561fb04a 81#include "io-wq.h"
2b188cc1 82
de23077e 83#include "io_uring.h"
329061d3 84#include "opdef.h"
e418bbc9 85#include "refs.h"
c9f06aa7 86#include "tctx.h"
17437f31 87#include "sqpoll.h"
a4ad4f74 88#include "fdinfo.h"
3b77495a 89#include "kbuf.h"
73572984 90#include "rsrc.h"
38513c46 91#include "cancel.h"
43e0bbbd 92#include "net.h"
eb42cebb 93#include "notif.h"
e27f928e 94
59915143 95#include "timeout.h"
329061d3 96#include "poll.h"
9b797a37 97#include "alloc_cache.h"
5e2a18d9 98
5277deaa 99#define IORING_MAX_ENTRIES 32768
33a107f0 100#define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
65e19f54 101
21b55dbc
SG
102#define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
103 IORING_REGISTER_LAST + IORING_OP_LAST)
2b188cc1 104
68fe256a
PB
105#define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
106 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
107
5562a8d7
PB
108#define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
109 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
68fe256a 110
c854357b 111#define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
9cae36a0
JA
112 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
113 REQ_F_ASYNC_DATA)
b16fed66 114
a538be5b
PB
115#define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
116 IO_REQ_CLEAN_FLAGS)
117
09899b19
PB
118#define IO_TCTX_REFS_CACHE_NR (1U << 10)
119
6dd0be1e 120#define IO_COMPL_BATCH 32
bf019da7 121#define IO_REQ_ALLOC_BATCH 8
258b29a9 122
10988a0a
DY
123enum {
124 IO_CHECK_CQ_OVERFLOW_BIT,
155bc950 125 IO_CHECK_CQ_DROPPED_BIT,
10988a0a
DY
126};
127
21a091b9
DY
128enum {
129 IO_EVENTFD_OP_SIGNAL_BIT,
130 IO_EVENTFD_OP_FREE_BIT,
131};
132
27dc8338
PB
133struct io_defer_entry {
134 struct list_head list;
135 struct io_kiocb *req;
9cf7c104 136 u32 seq;
2b188cc1
JA
137};
138
0756a869
PB
139/* requests with any of those set should undergo io_disarm_next() */
140#define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
da1a08c5 141#define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
0756a869 142
affa87db 143static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
9936c7c2 144 struct task_struct *task,
3dd0c97a 145 bool cancel_all);
1ffc5422 146
c7dae4ba 147static void io_dismantle_req(struct io_kiocb *req);
68fb8979 148static void io_clean_op(struct io_kiocb *req);
cbc2e203 149static void io_queue_sqe(struct io_kiocb *req);
c0e0d6ba 150static void io_move_task_work_from_local(struct io_ring_ctx *ctx);
c450178d 151static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
77e443ab 152static __cold void io_fallback_tw(struct io_uring_task *tctx);
de0617e4 153
2b188cc1
JA
154static struct kmem_cache *req_cachep;
155
2b188cc1
JA
156struct sock *io_uring_get_socket(struct file *file)
157{
158#if defined(CONFIG_UNIX)
cd40cae2 159 if (io_is_uring_fops(file)) {
2b188cc1
JA
160 struct io_ring_ctx *ctx = file->private_data;
161
162 return ctx->ring_sock->sk;
163 }
164#endif
165 return NULL;
166}
167EXPORT_SYMBOL(io_uring_get_socket);
168
c450178d
PB
169static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
170{
931147dd
DY
171 if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
172 ctx->submit_state.cqes_count)
c450178d
PB
173 __io_submit_flush_completions(ctx);
174}
175
faf88dde
PB
176static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
177{
178 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
179}
180
0fc8c2ac
DY
181static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
182{
183 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
184}
185
9cae36a0
JA
186static bool io_match_linked(struct io_kiocb *head)
187{
188 struct io_kiocb *req;
189
190 io_for_each_link(req, head) {
191 if (req->flags & REQ_F_INFLIGHT)
192 return true;
193 }
194 return false;
6af3f48b
PB
195}
196
197/*
198 * As io_match_task() but protected against racing with linked timeouts.
199 * User must not hold timeout_lock.
200 */
329061d3
JA
201bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
202 bool cancel_all)
6af3f48b 203{
9cae36a0
JA
204 bool matched;
205
6af3f48b
PB
206 if (task && head->task != task)
207 return false;
9cae36a0
JA
208 if (cancel_all)
209 return true;
210
211 if (head->flags & REQ_F_LINK_TIMEOUT) {
212 struct io_ring_ctx *ctx = head->ctx;
213
214 /* protect against races with linked timeouts */
215 spin_lock_irq(&ctx->timeout_lock);
216 matched = io_match_linked(head);
217 spin_unlock_irq(&ctx->timeout_lock);
218 } else {
219 matched = io_match_linked(head);
220 }
221 return matched;
6af3f48b
PB
222}
223
a8295b98
HX
224static inline void req_fail_link_node(struct io_kiocb *req, int res)
225{
226 req_set_fail(req);
97b388d7 227 io_req_set_res(req, res, 0);
a8295b98
HX
228}
229
fa05457a
PB
230static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
231{
232 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
a8295b98
HX
233}
234
c072481d 235static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
2b188cc1
JA
236{
237 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
238
0f158b4c 239 complete(&ctx->ref_comp);
2b188cc1
JA
240}
241
c072481d 242static __cold void io_fallback_req_func(struct work_struct *work)
f56165e6
PB
243{
244 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
245 fallback_work.work);
246 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
247 struct io_kiocb *req, *tmp;
f237c30a 248 bool locked = false;
f56165e6
PB
249
250 percpu_ref_get(&ctx->refs);
3218e5d3 251 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
f237c30a 252 req->io_task_work.func(req, &locked);
5636c00d 253
f237c30a 254 if (locked) {
c450178d 255 io_submit_flush_completions(ctx);
f237c30a
PB
256 mutex_unlock(&ctx->uring_lock);
257 }
f56165e6
PB
258 percpu_ref_put(&ctx->refs);
259}
260
e6f89be6
PB
261static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
262{
263 unsigned hash_buckets = 1U << bits;
264 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
265
266 table->hbs = kmalloc(hash_size, GFP_KERNEL);
267 if (!table->hbs)
268 return -ENOMEM;
269
270 table->hash_bits = bits;
271 init_hash_table(table, hash_buckets);
272 return 0;
273}
274
c072481d 275static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
2b188cc1
JA
276{
277 struct io_ring_ctx *ctx;
9cfc7e94 278 int hash_bits;
2b188cc1
JA
279
280 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
281 if (!ctx)
282 return NULL;
283
9cfc7e94
JA
284 xa_init(&ctx->io_bl_xa);
285
78076bb6
JA
286 /*
287 * Use 5 bits less than the max cq entries, that should give us around
4a07723f
PB
288 * 32 entries per hash list if totally full and uniformly spread, but
289 * don't keep too many buckets to not overconsume memory.
78076bb6 290 */
4a07723f
PB
291 hash_bits = ilog2(p->cq_entries) - 5;
292 hash_bits = clamp(hash_bits, 1, 8);
e6f89be6 293 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
78076bb6 294 goto err;
9ca9fb24
PB
295 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
296 goto err;
38513c46 297
6224843d
PB
298 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
299 if (!ctx->dummy_ubuf)
300 goto err;
301 /* set invalid range, so io_import_fixed() fails meeting it */
302 ctx->dummy_ubuf->ubuf = -1UL;
303
21482896 304 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
48904229 305 0, GFP_KERNEL))
206aefde 306 goto err;
2b188cc1
JA
307
308 ctx->flags = p->flags;
90554200 309 init_waitqueue_head(&ctx->sqo_sq_wait);
69fb2131 310 INIT_LIST_HEAD(&ctx->sqd_list);
1d7bb1d5 311 INIT_LIST_HEAD(&ctx->cq_overflow_list);
cc3cec83 312 INIT_LIST_HEAD(&ctx->io_buffers_cache);
9b797a37 313 io_alloc_cache_init(&ctx->apoll_cache);
43e0bbbd 314 io_alloc_cache_init(&ctx->netmsg_cache);
0f158b4c 315 init_completion(&ctx->ref_comp);
61cf9370 316 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
2b188cc1 317 mutex_init(&ctx->uring_lock);
311997b3 318 init_waitqueue_head(&ctx->cq_wait);
2b188cc1 319 spin_lock_init(&ctx->completion_lock);
89850fce 320 spin_lock_init(&ctx->timeout_lock);
5eef4e87 321 INIT_WQ_LIST(&ctx->iopoll_list);
cc3cec83
JA
322 INIT_LIST_HEAD(&ctx->io_buffers_pages);
323 INIT_LIST_HEAD(&ctx->io_buffers_comp);
de0617e4 324 INIT_LIST_HEAD(&ctx->defer_list);
5262f567 325 INIT_LIST_HEAD(&ctx->timeout_list);
ef9dd637 326 INIT_LIST_HEAD(&ctx->ltimeout_list);
d67d2263
BM
327 spin_lock_init(&ctx->rsrc_ref_lock);
328 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
269bbe5f 329 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
d34b1b0b 330 init_task_work(&ctx->rsrc_put_tw, io_rsrc_put_tw);
269bbe5f 331 init_llist_head(&ctx->rsrc_put_llist);
c0e0d6ba 332 init_llist_head(&ctx->work_llist);
13bf43f5 333 INIT_LIST_HEAD(&ctx->tctx_list);
c2b6c6bc
PB
334 ctx->submit_state.free_list.next = NULL;
335 INIT_WQ_LIST(&ctx->locked_free_list);
9011bf9a 336 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
6f33b0bc 337 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
2b188cc1 338 return ctx;
206aefde 339err:
6224843d 340 kfree(ctx->dummy_ubuf);
e6f89be6 341 kfree(ctx->cancel_table.hbs);
9ca9fb24 342 kfree(ctx->cancel_table_locked.hbs);
9cfc7e94
JA
343 kfree(ctx->io_bl);
344 xa_destroy(&ctx->io_bl_xa);
206aefde
JA
345 kfree(ctx);
346 return NULL;
2b188cc1
JA
347}
348
8f6ed49a
PB
349static void io_account_cq_overflow(struct io_ring_ctx *ctx)
350{
351 struct io_rings *r = ctx->rings;
352
353 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
354 ctx->cq_extra--;
355}
356
9cf7c104 357static bool req_need_defer(struct io_kiocb *req, u32 seq)
7adf4eaf 358{
2bc9930e
JA
359 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
360 struct io_ring_ctx *ctx = req->ctx;
a197f664 361
8f6ed49a 362 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
2bc9930e 363 }
de0617e4 364
9d858b21 365 return false;
de0617e4
JA
366}
367
9cae36a0
JA
368static inline void io_req_track_inflight(struct io_kiocb *req)
369{
370 if (!(req->flags & REQ_F_INFLIGHT)) {
371 req->flags |= REQ_F_INFLIGHT;
386e4fb6 372 atomic_inc(&req->task->io_uring->inflight_tracked);
9cae36a0
JA
373 }
374}
375
fd08e530
PB
376static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
377{
906c6caa
PB
378 if (WARN_ON_ONCE(!req->link))
379 return NULL;
380
4d13d1a4
PB
381 req->flags &= ~REQ_F_ARM_LTIMEOUT;
382 req->flags |= REQ_F_LINK_TIMEOUT;
fd08e530
PB
383
384 /* linked timeouts should have two refs once prep'ed */
48dcd38d 385 io_req_set_refcount(req);
4d13d1a4
PB
386 __io_req_set_refcount(req->link, 2);
387 return req->link;
fd08e530
PB
388}
389
390static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
391{
4d13d1a4 392 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
fd08e530
PB
393 return NULL;
394 return __io_prep_linked_timeout(req);
395}
396
cb2d344c
PB
397static noinline void __io_arm_ltimeout(struct io_kiocb *req)
398{
399 io_queue_linked_timeout(__io_prep_linked_timeout(req));
400}
401
402static inline void io_arm_ltimeout(struct io_kiocb *req)
403{
404 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
405 __io_arm_ltimeout(req);
406}
407
1e6fa521
JA
408static void io_prep_async_work(struct io_kiocb *req)
409{
410 const struct io_op_def *def = &io_op_defs[req->opcode];
1e6fa521
JA
411 struct io_ring_ctx *ctx = req->ctx;
412
b8e64b53
PB
413 if (!(req->flags & REQ_F_CREDS)) {
414 req->flags |= REQ_F_CREDS;
c10d1f98 415 req->creds = get_current_cred();
b8e64b53 416 }
003e8dcc 417
e1d675df
PB
418 req->work.list.next = NULL;
419 req->work.flags = 0;
8e29da69 420 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
feaadc4f
PB
421 if (req->flags & REQ_F_FORCE_ASYNC)
422 req->work.flags |= IO_WQ_WORK_CONCURRENT;
423
f6b543fd
JA
424 if (req->file && !io_req_ffs_set(req))
425 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
426
1e6fa521
JA
427 if (req->flags & REQ_F_ISREG) {
428 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
429 io_wq_hash_work(&req->work, file_inode(req->file));
4b982bd0 430 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
1e6fa521
JA
431 if (def->unbound_nonreg_file)
432 req->work.flags |= IO_WQ_WORK_UNBOUND;
433 }
561fb04a 434}
cccf0ee8 435
cbdcb435 436static void io_prep_async_link(struct io_kiocb *req)
561fb04a 437{
cbdcb435 438 struct io_kiocb *cur;
54a91f3b 439
44eff40a
PB
440 if (req->flags & REQ_F_LINK_TIMEOUT) {
441 struct io_ring_ctx *ctx = req->ctx;
442
674ee8e1 443 spin_lock_irq(&ctx->timeout_lock);
44eff40a
PB
444 io_for_each_link(cur, req)
445 io_prep_async_work(cur);
674ee8e1 446 spin_unlock_irq(&ctx->timeout_lock);
44eff40a
PB
447 } else {
448 io_for_each_link(cur, req)
449 io_prep_async_work(cur);
450 }
561fb04a
JA
451}
452
f3b44f92 453void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
561fb04a 454{
cbdcb435 455 struct io_kiocb *link = io_prep_linked_timeout(req);
5aa75ed5 456 struct io_uring_task *tctx = req->task->io_uring;
561fb04a 457
3bfe6106
JA
458 BUG_ON(!tctx);
459 BUG_ON(!tctx->io_wq);
561fb04a 460
cbdcb435
PB
461 /* init ->work of the whole link before punting */
462 io_prep_async_link(req);
991468dc
JA
463
464 /*
465 * Not expected to happen, but if we do have a bug where this _can_
466 * happen, catch it here and ensure the request is marked as
467 * canceled. That will make io-wq go through the usual work cancel
468 * procedure rather than attempt to run this request (or create a new
469 * worker for it).
470 */
471 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
472 req->work.flags |= IO_WQ_WORK_CANCEL;
473
48863ffd 474 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
ebf93667 475 io_wq_enqueue(tctx->io_wq, &req->work);
7271ef3a
JA
476 if (link)
477 io_queue_linked_timeout(link);
cbdcb435
PB
478}
479
c072481d 480static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
de0617e4 481{
441b8a78 482 while (!list_empty(&ctx->defer_list)) {
27dc8338
PB
483 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
484 struct io_defer_entry, list);
de0617e4 485
9cf7c104 486 if (req_need_defer(de->req, de->seq))
04518945 487 break;
27dc8338 488 list_del_init(&de->list);
907d1df3 489 io_req_task_queue(de->req);
27dc8338 490 kfree(de);
441b8a78 491 }
04518945
PB
492}
493
21a091b9
DY
494
495static void io_eventfd_ops(struct rcu_head *rcu)
f2842ab5 496{
d8e9214f 497 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
21a091b9 498 int ops = atomic_xchg(&ev_fd->ops, 0);
305bef98 499
21a091b9 500 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
44648532 501 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
d8e9214f 502
21a091b9
DY
503 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
504 * ordering in a race but if references are 0 we know we have to free
505 * it regardless.
305bef98 506 */
21a091b9
DY
507 if (atomic_dec_and_test(&ev_fd->refs)) {
508 eventfd_ctx_put(ev_fd->cq_ev_fd);
509 kfree(ev_fd);
510 }
d8e9214f
DY
511}
512
77bc59b4 513static void io_eventfd_signal(struct io_ring_ctx *ctx)
f2842ab5 514{
21a091b9 515 struct io_ev_fd *ev_fd = NULL;
77bc59b4 516
77bc59b4
UA
517 rcu_read_lock();
518 /*
519 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
520 * and eventfd_signal
521 */
522 ev_fd = rcu_dereference(ctx->io_ev_fd);
523
524 /*
525 * Check again if ev_fd exists incase an io_eventfd_unregister call
526 * completed between the NULL check of ctx->io_ev_fd at the start of
527 * the function and rcu_read_lock.
528 */
529 if (unlikely(!ev_fd))
530 goto out;
7e55a19c 531 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
77bc59b4 532 goto out;
21a091b9
DY
533 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
534 goto out;
77bc59b4 535
21a091b9 536 if (likely(eventfd_signal_allowed())) {
44648532 537 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
21a091b9
DY
538 } else {
539 atomic_inc(&ev_fd->refs);
540 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
44a84da4 541 call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
21a091b9
DY
542 else
543 atomic_dec(&ev_fd->refs);
544 }
545
77bc59b4
UA
546out:
547 rcu_read_unlock();
f2842ab5
JA
548}
549
21a091b9
DY
550static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
551{
552 bool skip;
553
554 spin_lock(&ctx->completion_lock);
555
556 /*
557 * Eventfd should only get triggered when at least one event has been
558 * posted. Some applications rely on the eventfd notification count
559 * only changing IFF a new CQE has been added to the CQ ring. There's
560 * no depedency on 1:1 relationship between how many times this
561 * function is called (and hence the eventfd count) and number of CQEs
562 * posted to the CQ ring.
563 */
564 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
565 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
566 spin_unlock(&ctx->completion_lock);
567 if (skip)
568 return;
569
570 io_eventfd_signal(ctx);
571}
572
a830ffd2
PB
573void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
574{
e5f30f6f
PB
575 if (ctx->off_timeout_used)
576 io_flush_timeouts(ctx);
577 if (ctx->drain_active) {
a830ffd2 578 spin_lock(&ctx->completion_lock);
e5f30f6f 579 io_queue_deferred(ctx);
a830ffd2
PB
580 spin_unlock(&ctx->completion_lock);
581 }
582 if (ctx->has_evfd)
21a091b9 583 io_eventfd_flush_signal(ctx);
a830ffd2
PB
584}
585
f66f7342
PB
586static inline void __io_cq_lock(struct io_ring_ctx *ctx)
587 __acquires(ctx->completion_lock)
588{
589 if (!ctx->task_complete)
590 spin_lock(&ctx->completion_lock);
591}
592
593static inline void __io_cq_unlock(struct io_ring_ctx *ctx)
594{
595 if (!ctx->task_complete)
596 spin_unlock(&ctx->completion_lock);
597}
598
6971253f
PB
599static inline void io_cq_lock(struct io_ring_ctx *ctx)
600 __acquires(ctx->completion_lock)
601{
602 spin_lock(&ctx->completion_lock);
603}
604
605static inline void io_cq_unlock(struct io_ring_ctx *ctx)
606 __releases(ctx->completion_lock)
607{
608 spin_unlock(&ctx->completion_lock);
609}
610
5d772916 611/* keep it inlined for io_submit_flush_completions() */
f66f7342 612static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
25399321
PB
613 __releases(ctx->completion_lock)
614{
615 io_commit_cqring(ctx);
f66f7342 616 __io_cq_unlock(ctx);
6c16fe3c
JA
617 io_commit_cqring_flush(ctx);
618 io_cqring_wake(ctx);
25399321
PB
619}
620
621void io_cq_unlock_post(struct io_ring_ctx *ctx)
5d772916 622 __releases(ctx->completion_lock)
25399321 623{
f66f7342
PB
624 io_commit_cqring(ctx);
625 spin_unlock(&ctx->completion_lock);
626 io_commit_cqring_flush(ctx);
627 io_cqring_wake(ctx);
25399321
PB
628}
629
c4a2ed72 630/* Returns true if there are no backlogged entries after the flush */
a85381d8
PB
631static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
632{
633 struct io_overflow_cqe *ocqe;
634 LIST_HEAD(list);
635
636 io_cq_lock(ctx);
637 list_splice_init(&ctx->cq_overflow_list, &list);
638 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
639 io_cq_unlock(ctx);
640
641 while (!list_empty(&list)) {
642 ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
643 list_del(&ocqe->list);
644 kfree(ocqe);
645 }
646}
647
648/* Returns true if there are no backlogged entries after the flush */
1b346e4a 649static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
1d7bb1d5 650{
e45a3e05 651 size_t cqe_size = sizeof(struct io_uring_cqe);
1d7bb1d5 652
a85381d8 653 if (__io_cqring_events(ctx) == ctx->cq_entries)
1b346e4a 654 return;
1d7bb1d5 655
e45a3e05
SR
656 if (ctx->flags & IORING_SETUP_CQE32)
657 cqe_size <<= 1;
658
25399321 659 io_cq_lock(ctx);
6c2450ae 660 while (!list_empty(&ctx->cq_overflow_list)) {
aa1df3a3 661 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
6c2450ae 662 struct io_overflow_cqe *ocqe;
e6c8aa9a 663
a85381d8 664 if (!cqe)
1d7bb1d5 665 break;
6c2450ae
PB
666 ocqe = list_first_entry(&ctx->cq_overflow_list,
667 struct io_overflow_cqe, list);
a85381d8 668 memcpy(cqe, &ocqe->cqe, cqe_size);
6c2450ae
PB
669 list_del(&ocqe->list);
670 kfree(ocqe);
1d7bb1d5
JA
671 }
672
1b346e4a 673 if (list_empty(&ctx->cq_overflow_list)) {
10988a0a 674 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
3a4b89a2 675 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
09e88404 676 }
25399321 677 io_cq_unlock_post(ctx);
1d7bb1d5
JA
678}
679
52ea806a
JA
680static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
681{
682 /* iopoll syncs against uring_lock, not completion_lock */
683 if (ctx->flags & IORING_SETUP_IOPOLL)
684 mutex_lock(&ctx->uring_lock);
685 __io_cqring_overflow_flush(ctx);
686 if (ctx->flags & IORING_SETUP_IOPOLL)
687 mutex_unlock(&ctx->uring_lock);
688}
689
1b346e4a 690static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
6c503150 691{
52ea806a
JA
692 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
693 io_cqring_do_overflow_flush(ctx);
6c503150
PB
694}
695
e70cb608 696void __io_put_task(struct task_struct *task, int nr)
6a290a14
PB
697{
698 struct io_uring_task *tctx = task->io_uring;
699
9d170164
PB
700 percpu_counter_sub(&tctx->inflight, nr);
701 if (unlikely(atomic_read(&tctx->in_idle)))
702 wake_up(&tctx->wait);
703 put_task_struct_many(task, nr);
704}
705
63809137 706void io_task_refs_refill(struct io_uring_task *tctx)
9a10867a
PB
707{
708 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
709
710 percpu_counter_add(&tctx->inflight, refill);
711 refcount_add(refill, &current->usage);
712 tctx->cached_refs += refill;
713}
714
3cc7fdb9
PB
715static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
716{
717 struct io_uring_task *tctx = task->io_uring;
718 unsigned int refs = tctx->cached_refs;
719
720 if (refs) {
721 tctx->cached_refs = 0;
722 percpu_counter_sub(&tctx->inflight, refs);
723 put_task_struct_many(task, refs);
724 }
725}
726
68494a65
PB
727static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
728 s32 res, u32 cflags, u64 extra1, u64 extra2)
2b188cc1 729{
cce4b8b0 730 struct io_overflow_cqe *ocqe;
e45a3e05
SR
731 size_t ocq_size = sizeof(struct io_overflow_cqe);
732 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
2b188cc1 733
f26cc959
PB
734 lockdep_assert_held(&ctx->completion_lock);
735
e45a3e05
SR
736 if (is_cqe32)
737 ocq_size += sizeof(struct io_uring_cqe);
2b188cc1 738
e45a3e05 739 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
08dcd028 740 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
cce4b8b0
PB
741 if (!ocqe) {
742 /*
743 * If we're in ring overflow flush mode, or in task cancel mode,
744 * or cannot allocate an overflow entry, then we need to drop it
745 * on the floor.
746 */
8f6ed49a 747 io_account_cq_overflow(ctx);
155bc950 748 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
cce4b8b0 749 return false;
2b188cc1 750 }
cce4b8b0 751 if (list_empty(&ctx->cq_overflow_list)) {
10988a0a 752 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
3a4b89a2 753 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
20c0b380 754
cce4b8b0 755 }
d4d19c19 756 ocqe->cqe.user_data = user_data;
cce4b8b0
PB
757 ocqe->cqe.res = res;
758 ocqe->cqe.flags = cflags;
e45a3e05
SR
759 if (is_cqe32) {
760 ocqe->cqe.big_cqe[0] = extra1;
761 ocqe->cqe.big_cqe[1] = extra2;
762 }
cce4b8b0
PB
763 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
764 return true;
2b188cc1
JA
765}
766
68494a65
PB
767bool io_req_cqe_overflow(struct io_kiocb *req)
768{
769 if (!(req->flags & REQ_F_CQE32_INIT)) {
770 req->extra1 = 0;
771 req->extra2 = 0;
772 }
773 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
774 req->cqe.res, req->cqe.flags,
775 req->extra1, req->extra2);
776}
777
faf88dde
PB
778/*
779 * writes to the cq entry need to come after reading head; the
780 * control dependency is enough as we're using WRITE_ONCE to
781 * fill the cq entry
782 */
aa1df3a3 783struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
faf88dde
PB
784{
785 struct io_rings *rings = ctx->rings;
786 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
faf88dde
PB
787 unsigned int free, queued, len;
788
aa1df3a3
PB
789 /*
790 * Posting into the CQ when there are pending overflowed CQEs may break
791 * ordering guarantees, which will affect links, F_MORE users and more.
792 * Force overflow the completion.
793 */
794 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
795 return NULL;
faf88dde
PB
796
797 /* userspace may cheat modifying the tail, be safe and do min */
798 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
799 free = ctx->cq_entries - queued;
800 /* we need a contiguous range, limit based on the current array offset */
801 len = min(free, ctx->cq_entries - off);
802 if (!len)
803 return NULL;
804
b3659a65
PB
805 if (ctx->flags & IORING_SETUP_CQE32) {
806 off <<= 1;
807 len <<= 1;
808 }
809
faf88dde
PB
810 ctx->cqe_cached = &rings->cqes[off];
811 ctx->cqe_sentinel = ctx->cqe_cached + len;
b3659a65
PB
812
813 ctx->cached_cq_tail++;
faf88dde 814 ctx->cqe_cached++;
b3659a65
PB
815 if (ctx->flags & IORING_SETUP_CQE32)
816 ctx->cqe_cached++;
817 return &rings->cqes[off];
faf88dde
PB
818}
819
f66f7342
PB
820static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
821 u32 cflags)
bcda7baa 822{
cd94903d
PB
823 struct io_uring_cqe *cqe;
824
913a571a 825 ctx->cq_extra++;
cd94903d
PB
826
827 /*
828 * If we can't get a cq entry, userspace overflowed the
829 * submission (by quite a lot). Increment the overflow count in
830 * the ring.
831 */
832 cqe = io_get_cqe(ctx);
833 if (likely(cqe)) {
e0486f3f
DY
834 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
835
cd94903d
PB
836 WRITE_ONCE(cqe->user_data, user_data);
837 WRITE_ONCE(cqe->res, res);
838 WRITE_ONCE(cqe->flags, cflags);
c5595975
PB
839
840 if (ctx->flags & IORING_SETUP_CQE32) {
841 WRITE_ONCE(cqe->big_cqe[0], 0);
842 WRITE_ONCE(cqe->big_cqe[1], 0);
843 }
cd94903d
PB
844 return true;
845 }
52120f0f 846 return false;
bcda7baa
JA
847}
848
931147dd
DY
849static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
850 __must_hold(&ctx->uring_lock)
851{
852 struct io_submit_state *state = &ctx->submit_state;
853 unsigned int i;
854
855 lockdep_assert_held(&ctx->uring_lock);
856 for (i = 0; i < state->cqes_count; i++) {
857 struct io_uring_cqe *cqe = &state->cqes[i];
858
f66f7342
PB
859 if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
860 if (ctx->task_complete) {
861 spin_lock(&ctx->completion_lock);
862 io_cqring_event_overflow(ctx, cqe->user_data,
863 cqe->res, cqe->flags, 0, 0);
864 spin_unlock(&ctx->completion_lock);
865 } else {
866 io_cqring_event_overflow(ctx, cqe->user_data,
867 cqe->res, cqe->flags, 0, 0);
868 }
869 }
931147dd
DY
870 }
871 state->cqes_count = 0;
872}
873
b529c96a
DY
874static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
875 bool allow_overflow)
d245bca6
PB
876{
877 bool filled;
878
25399321 879 io_cq_lock(ctx);
f66f7342
PB
880 filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
881 if (!filled && allow_overflow)
882 filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
883
25399321 884 io_cq_unlock_post(ctx);
d245bca6
PB
885 return filled;
886}
887
b529c96a
DY
888bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
889{
890 return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
891}
892
9b8c5475
DY
893bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags,
894 bool allow_overflow)
2b188cc1 895{
9b8c5475
DY
896 struct io_uring_cqe *cqe;
897 unsigned int length;
898
899 if (!defer)
b529c96a 900 return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
9b8c5475
DY
901
902 length = ARRAY_SIZE(ctx->submit_state.cqes);
903
904 lockdep_assert_held(&ctx->uring_lock);
905
906 if (ctx->submit_state.cqes_count == length) {
f66f7342 907 __io_cq_lock(ctx);
9b8c5475
DY
908 __io_flush_post_cqes(ctx);
909 /* no need to flush - flush is deferred */
f66f7342 910 __io_cq_unlock_post(ctx);
9b8c5475
DY
911 }
912
913 /* For defered completions this is not as strict as it is otherwise,
914 * however it's main job is to prevent unbounded posted completions,
915 * and in that it works just as well.
916 */
917 if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
918 return false;
919
920 cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
921 cqe->user_data = user_data;
922 cqe->res = res;
923 cqe->flags = cflags;
924 return true;
925}
926
1bec951c 927static void __io_req_complete_post(struct io_kiocb *req)
2b188cc1 928{
fa18fa22
PB
929 struct io_ring_ctx *ctx = req->ctx;
930
931 io_cq_lock(ctx);
932 if (!(req->flags & REQ_F_CQE_SKIP))
a8cf95f9 933 io_fill_cqe_req(ctx, req);
fa18fa22 934
c7dae4ba
JA
935 /*
936 * If we're the last reference to this request, add to our locked
937 * free_list cache.
938 */
de9b4cca 939 if (req_ref_put_and_test(req)) {
da1a08c5 940 if (req->flags & IO_REQ_LINK_FLAGS) {
0756a869 941 if (req->flags & IO_DISARM_MASK)
7a612350
PB
942 io_disarm_next(req);
943 if (req->link) {
944 io_req_task_queue(req->link);
945 req->link = NULL;
946 }
947 }
7ac1edc4 948 io_req_put_rsrc(req);
8197b053
PB
949 /*
950 * Selected buffer deallocation in io_clean_op() assumes that
951 * we don't hold ->completion_lock. Clean them here to avoid
952 * deadlocks.
953 */
954 io_put_kbuf_comp(req);
c7dae4ba
JA
955 io_dismantle_req(req);
956 io_put_task(req->task, 1);
c2b6c6bc 957 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
d0acdee2 958 ctx->locked_free_nr++;
180f829f 959 }
25399321 960 io_cq_unlock_post(ctx);
4e3d9ff9
JA
961}
962
1bec951c 963void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
bcda7baa 964{
e6aeb272
PB
965 if (req->ctx->task_complete && (issue_flags & IO_URING_F_IOWQ)) {
966 req->io_task_work.func = io_req_task_complete;
967 io_req_task_work_add(req);
968 } else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
969 !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
1bec951c
PB
970 __io_req_complete_post(req);
971 } else {
972 struct io_ring_ctx *ctx = req->ctx;
973
974 mutex_lock(&ctx->uring_lock);
975 __io_req_complete_post(req);
976 mutex_unlock(&ctx->uring_lock);
977 }
0ddf92e8
JA
978}
979
973fc83f 980void io_req_defer_failed(struct io_kiocb *req, s32 res)
e276ae34 981 __must_hold(&ctx->uring_lock)
f41db273 982{
a47b255e
PB
983 const struct io_op_def *def = &io_op_defs[req->opcode];
984
e276ae34
PB
985 lockdep_assert_held(&req->ctx->uring_lock);
986
93d2bcd2 987 req_set_fail(req);
97b388d7 988 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
a47b255e
PB
989 if (def->fail)
990 def->fail(req);
973fc83f 991 io_req_complete_defer(req);
f41db273
PB
992}
993
864ea921
PB
994/*
995 * Don't initialise the fields below on every allocation, but do that in
996 * advance and keep them valid across allocations.
997 */
998static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
999{
1000 req->ctx = ctx;
1001 req->link = NULL;
1002 req->async_data = NULL;
1003 /* not necessary, but safer to zero */
cef216fc 1004 req->cqe.res = 0;
864ea921
PB
1005}
1006
dac7a098 1007static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
cd0ca2e0 1008 struct io_submit_state *state)
dac7a098 1009{
79ebeaee 1010 spin_lock(&ctx->completion_lock);
c2b6c6bc 1011 wq_list_splice(&ctx->locked_free_list, &state->free_list);
d0acdee2 1012 ctx->locked_free_nr = 0;
79ebeaee 1013 spin_unlock(&ctx->completion_lock);
dac7a098
PB
1014}
1015
5d5901a3
PB
1016/*
1017 * A request might get retired back into the request caches even before opcode
1018 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1019 * Because of that, io_alloc_req() should be called only under ->uring_lock
1020 * and with extra caution to not get a request that is still worked on.
1021 */
bd1a3783 1022__cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
5d5901a3 1023 __must_hold(&ctx->uring_lock)
2b188cc1 1024{
864ea921 1025 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
3ab665b7 1026 void *reqs[IO_REQ_ALLOC_BATCH];
864ea921 1027 int ret, i;
e5d1bc0a 1028
23a5c43b
PB
1029 /*
1030 * If we have more than a batch's worth of requests in our IRQ side
1031 * locked cache, grab the lock and move them over to our submission
1032 * side cache.
1033 */
a6d97a8a 1034 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
23a5c43b 1035 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
88ab95be 1036 if (!io_req_cache_empty(ctx))
23a5c43b
PB
1037 return true;
1038 }
e5d1bc0a 1039
3ab665b7 1040 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
fd6fab2c 1041
864ea921
PB
1042 /*
1043 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1044 * retry single alloc to be on the safe side.
1045 */
1046 if (unlikely(ret <= 0)) {
3ab665b7
PB
1047 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1048 if (!reqs[0])
a33ae9ce 1049 return false;
864ea921 1050 ret = 1;
2b188cc1 1051 }
864ea921 1052
37f0e767 1053 percpu_ref_get_many(&ctx->refs, ret);
3ab665b7 1054 for (i = 0; i < ret; i++) {
23a5c43b 1055 struct io_kiocb *req = reqs[i];
3ab665b7
PB
1056
1057 io_preinit_req(req, ctx);
fa05457a 1058 io_req_add_to_cache(req, ctx);
3ab665b7 1059 }
a33ae9ce
PB
1060 return true;
1061}
1062
6b639522 1063static inline void io_dismantle_req(struct io_kiocb *req)
2b188cc1 1064{
094bae49 1065 unsigned int flags = req->flags;
929a3af9 1066
867f8fa5 1067 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
3a0a6902 1068 io_clean_op(req);
e1d767f0
PB
1069 if (!(flags & REQ_F_FIXED_FILE))
1070 io_put_file(req->file);
e65ef56d
JA
1071}
1072
59915143 1073__cold void io_free_req(struct io_kiocb *req)
c6ca97b3 1074{
51a4cc11 1075 struct io_ring_ctx *ctx = req->ctx;
c6ca97b3 1076
7ac1edc4 1077 io_req_put_rsrc(req);
216578e5 1078 io_dismantle_req(req);
7c660731 1079 io_put_task(req->task, 1);
c6ca97b3 1080
79ebeaee 1081 spin_lock(&ctx->completion_lock);
c2b6c6bc 1082 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
c34b025f 1083 ctx->locked_free_nr++;
79ebeaee 1084 spin_unlock(&ctx->completion_lock);
e65ef56d
JA
1085}
1086
d81499bf
PB
1087static void __io_req_find_next_prep(struct io_kiocb *req)
1088{
1089 struct io_ring_ctx *ctx = req->ctx;
d81499bf 1090
6971253f 1091 spin_lock(&ctx->completion_lock);
305bef98 1092 io_disarm_next(req);
6971253f 1093 spin_unlock(&ctx->completion_lock);
d81499bf
PB
1094}
1095
1096static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
c69f8dbe 1097{
33cc89a9 1098 struct io_kiocb *nxt;
944e58bf 1099
9e645e11
JA
1100 /*
1101 * If LINK is set, we have dependent requests in this chain. If we
1102 * didn't fail this request, queue the first one up, moving any other
1103 * dependencies to the next request. In case of failure, fail the rest
1104 * of the chain.
1105 */
d81499bf
PB
1106 if (unlikely(req->flags & IO_DISARM_MASK))
1107 __io_req_find_next_prep(req);
33cc89a9
PB
1108 nxt = req->link;
1109 req->link = NULL;
1110 return nxt;
4d7dd462 1111}
9e645e11 1112
f237c30a 1113static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
2c32395d
PB
1114{
1115 if (!ctx)
1116 return;
ef060ea9
JA
1117 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1118 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
f237c30a 1119 if (*locked) {
c450178d 1120 io_submit_flush_completions(ctx);
2c32395d 1121 mutex_unlock(&ctx->uring_lock);
f237c30a 1122 *locked = false;
2c32395d
PB
1123 }
1124 percpu_ref_put(&ctx->refs);
1125}
1126
c6dd763c
DY
1127static unsigned int handle_tw_list(struct llist_node *node,
1128 struct io_ring_ctx **ctx, bool *locked,
1129 struct llist_node *last)
9f8d032a 1130{
c6dd763c
DY
1131 unsigned int count = 0;
1132
3a0c037b 1133 while (node != last) {
f88262e6 1134 struct llist_node *next = node->next;
9f8d032a
HX
1135 struct io_kiocb *req = container_of(node, struct io_kiocb,
1136 io_task_work.node);
1137
34d2bfe7
JA
1138 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1139
9f8d032a
HX
1140 if (req->ctx != *ctx) {
1141 ctx_flush_and_put(*ctx, locked);
1142 *ctx = req->ctx;
1143 /* if not contended, grab and improve batching */
1144 *locked = mutex_trylock(&(*ctx)->uring_lock);
1145 percpu_ref_get(&(*ctx)->refs);
1146 }
1147 req->io_task_work.func(req, locked);
1148 node = next;
c6dd763c 1149 count++;
3a0c037b 1150 }
c6dd763c
DY
1151
1152 return count;
9f8d032a
HX
1153}
1154
923d1592
DY
1155/**
1156 * io_llist_xchg - swap all entries in a lock-less list
1157 * @head: the head of lock-less list to delete all entries
1158 * @new: new entry as the head of the list
1159 *
1160 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1161 * The order of entries returned is from the newest to the oldest added one.
1162 */
1163static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1164 struct llist_node *new)
1165{
1166 return xchg(&head->first, new);
1167}
1168
1169/**
1170 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1171 * @head: the head of lock-less list to delete all entries
1172 * @old: expected old value of the first entry of the list
1173 * @new: new entry as the head of the list
1174 *
1175 * perform a cmpxchg on the first entry of the list.
1176 */
1177
1178static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1179 struct llist_node *old,
1180 struct llist_node *new)
1181{
1182 return cmpxchg(&head->first, old, new);
1183}
1184
c9f06aa7 1185void tctx_task_work(struct callback_head *cb)
c40f6379 1186{
f28c240e 1187 bool uring_locked = false;
ebd0df2e 1188 struct io_ring_ctx *ctx = NULL;
3f18407d
PB
1189 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1190 task_work);
3a0c037b 1191 struct llist_node fake = {};
77e443ab 1192 struct llist_node *node;
c6dd763c 1193 unsigned int loops = 1;
77e443ab 1194 unsigned int count;
3a0c037b 1195
77e443ab
PB
1196 if (unlikely(current->flags & PF_EXITING)) {
1197 io_fallback_tw(tctx);
1198 return;
1199 }
3a0c037b 1200
77e443ab
PB
1201 node = io_llist_xchg(&tctx->task_list, &fake);
1202 count = handle_tw_list(node, &ctx, &uring_locked, NULL);
3a0c037b
DY
1203 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1204 while (node != &fake) {
c6dd763c 1205 loops++;
3a0c037b 1206 node = io_llist_xchg(&tctx->task_list, &fake);
c6dd763c 1207 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
3a0c037b 1208 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
3f18407d 1209 }
ebd0df2e 1210
f28c240e 1211 ctx_flush_and_put(ctx, &uring_locked);
3cc7fdb9
PB
1212
1213 /* relaxed read is enough as only the task itself sets ->in_idle */
1214 if (unlikely(atomic_read(&tctx->in_idle)))
1215 io_uring_drop_tctx_refs(current);
c6dd763c
DY
1216
1217 trace_io_uring_task_work_run(tctx, count, loops);
7cbf1722
JA
1218}
1219
d7593606
PB
1220static __cold void io_fallback_tw(struct io_uring_task *tctx)
1221{
1222 struct llist_node *node = llist_del_all(&tctx->task_list);
1223 struct io_kiocb *req;
1224
1225 while (node) {
1226 req = container_of(node, struct io_kiocb, io_task_work.node);
1227 node = node->next;
1228 if (llist_add(&req->io_task_work.node,
1229 &req->ctx->fallback_llist))
1230 schedule_delayed_work(&req->ctx->fallback_work, 1);
1231 }
1232}
1233
c0e0d6ba
DY
1234static void io_req_local_work_add(struct io_kiocb *req)
1235{
1236 struct io_ring_ctx *ctx = req->ctx;
1237
9ffa13ff
PB
1238 percpu_ref_get(&ctx->refs);
1239
1240 if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) {
1241 percpu_ref_put(&ctx->refs);
c0e0d6ba 1242 return;
9ffa13ff 1243 }
fc86f9d3
PB
1244 /* need it for the following io_cqring_wake() */
1245 smp_mb__after_atomic();
c0e0d6ba
DY
1246
1247 if (unlikely(atomic_read(&req->task->io_uring->in_idle))) {
1248 io_move_task_work_from_local(ctx);
9ffa13ff 1249 percpu_ref_put(&ctx->refs);
c0e0d6ba
DY
1250 return;
1251 }
1252
1253 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1254 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1255
21a091b9
DY
1256 if (ctx->has_evfd)
1257 io_eventfd_signal(ctx);
fc86f9d3 1258 __io_cqring_wake(ctx);
9ffa13ff 1259 percpu_ref_put(&ctx->refs);
c0e0d6ba
DY
1260}
1261
e52d2e58 1262void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
7cbf1722 1263{
c34398a8 1264 struct io_uring_task *tctx = req->task->io_uring;
9f010507 1265 struct io_ring_ctx *ctx = req->ctx;
7cbf1722 1266
c0e0d6ba
DY
1267 if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1268 io_req_local_work_add(req);
1269 return;
1270 }
7cbf1722
JA
1271
1272 /* task_work already pending, we're done */
32d91f05 1273 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
e09ee510 1274 return;
7cbf1722 1275
ef060ea9
JA
1276 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1277 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1278
3fe07bcd 1279 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
e09ee510 1280 return;
2215bed9 1281
d7593606 1282 io_fallback_tw(tctx);
c0e0d6ba
DY
1283}
1284
1285static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1286{
1287 struct llist_node *node;
1288
1289 node = llist_del_all(&ctx->work_llist);
1290 while (node) {
1291 struct io_kiocb *req = container_of(node, struct io_kiocb,
1292 io_task_work.node);
1293
1294 node = node->next;
1295 __io_req_task_work_add(req, false);
1296 }
1297}
1298
b3026767 1299int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked)
c0e0d6ba 1300{
c0e0d6ba
DY
1301 struct llist_node *node;
1302 struct llist_node fake;
1303 struct llist_node *current_final = NULL;
1304 int ret;
f75d5036 1305 unsigned int loops = 1;
c0e0d6ba 1306
6567506b 1307 if (unlikely(ctx->submitter_task != current))
c0e0d6ba 1308 return -EEXIST;
c0e0d6ba 1309
c0e0d6ba
DY
1310 node = io_llist_xchg(&ctx->work_llist, &fake);
1311 ret = 0;
1312again:
1313 while (node != current_final) {
1314 struct llist_node *next = node->next;
1315 struct io_kiocb *req = container_of(node, struct io_kiocb,
1316 io_task_work.node);
1317 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
b3026767 1318 req->io_task_work.func(req, locked);
c0e0d6ba
DY
1319 ret++;
1320 node = next;
1321 }
1322
1323 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1324 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1325
1326 node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL);
1327 if (node != &fake) {
f75d5036 1328 loops++;
c0e0d6ba
DY
1329 current_final = &fake;
1330 node = io_llist_xchg(&ctx->work_llist, &fake);
1331 goto again;
1332 }
1333
b3026767 1334 if (*locked)
c0e0d6ba 1335 io_submit_flush_completions(ctx);
f75d5036 1336 trace_io_uring_local_work_run(ctx, ret, loops);
c0e0d6ba 1337 return ret;
8ac5d85a
JA
1338
1339}
1340
1341int io_run_local_work(struct io_ring_ctx *ctx)
1342{
1343 bool locked;
1344 int ret;
1345
ec7fd256 1346 __set_current_state(TASK_RUNNING);
8ac5d85a 1347 locked = mutex_trylock(&ctx->uring_lock);
b3026767 1348 ret = __io_run_local_work(ctx, &locked);
8ac5d85a
JA
1349 if (locked)
1350 mutex_unlock(&ctx->uring_lock);
1351
1352 return ret;
c0e0d6ba
DY
1353}
1354
f237c30a 1355static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
c40f6379 1356{
971cf9c1 1357 io_tw_lock(req->ctx, locked);
973fc83f 1358 io_req_defer_failed(req, req->cqe.res);
c40f6379
JA
1359}
1360
329061d3 1361void io_req_task_submit(struct io_kiocb *req, bool *locked)
c40f6379 1362{
971cf9c1 1363 io_tw_lock(req->ctx, locked);
316319e8 1364 /* req->task == current here, checking PF_EXITING is safe */
af066f31 1365 if (likely(!(req->task->flags & PF_EXITING)))
cbc2e203 1366 io_queue_sqe(req);
81b6d05c 1367 else
973fc83f 1368 io_req_defer_failed(req, -EFAULT);
c40f6379
JA
1369}
1370
59915143 1371void io_req_task_queue_fail(struct io_kiocb *req, int ret)
c40f6379 1372{
97b388d7 1373 io_req_set_res(req, ret, 0);
5b0a6acc 1374 req->io_task_work.func = io_req_task_cancel;
3fe07bcd 1375 io_req_task_work_add(req);
c40f6379
JA
1376}
1377
f3b44f92 1378void io_req_task_queue(struct io_kiocb *req)
a3df7698 1379{
5b0a6acc 1380 req->io_task_work.func = io_req_task_submit;
3fe07bcd 1381 io_req_task_work_add(req);
a3df7698
PB
1382}
1383
59915143 1384void io_queue_next(struct io_kiocb *req)
c69f8dbe 1385{
9b5f7bd9 1386 struct io_kiocb *nxt = io_req_find_next(req);
944e58bf
PB
1387
1388 if (nxt)
906a8c3f 1389 io_req_task_queue(nxt);
c69f8dbe
JL
1390}
1391
f3b44f92 1392void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
3aa83bfb 1393 __must_hold(&ctx->uring_lock)
5af1d13e 1394{
d4b7a5ef 1395 struct task_struct *task = NULL;
37f0e767 1396 int task_refs = 0;
5af1d13e 1397
3aa83bfb
PB
1398 do {
1399 struct io_kiocb *req = container_of(node, struct io_kiocb,
1400 comp_list);
2d6500d4 1401
a538be5b
PB
1402 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1403 if (req->flags & REQ_F_REFCOUNT) {
1404 node = req->comp_list.next;
1405 if (!req_ref_put_and_test(req))
1406 continue;
1407 }
b605a7fa
PB
1408 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1409 struct async_poll *apoll = req->apoll;
1410
1411 if (apoll->double_poll)
1412 kfree(apoll->double_poll);
9731bc98
JA
1413 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1414 kfree(apoll);
b605a7fa
PB
1415 req->flags &= ~REQ_F_POLLED;
1416 }
da1a08c5 1417 if (req->flags & IO_REQ_LINK_FLAGS)
57859f4d 1418 io_queue_next(req);
a538be5b
PB
1419 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1420 io_clean_op(req);
c1e53a69 1421 }
a538be5b
PB
1422 if (!(req->flags & REQ_F_FIXED_FILE))
1423 io_put_file(req->file);
2d6500d4 1424
ab409402 1425 io_req_put_rsrc_locked(req, ctx);
5af1d13e 1426
d4b7a5ef
PB
1427 if (req->task != task) {
1428 if (task)
1429 io_put_task(task, task_refs);
1430 task = req->task;
1431 task_refs = 0;
1432 }
1433 task_refs++;
c1e53a69 1434 node = req->comp_list.next;
fa05457a 1435 io_req_add_to_cache(req, ctx);
3aa83bfb 1436 } while (node);
d4b7a5ef 1437
d4b7a5ef
PB
1438 if (task)
1439 io_put_task(task, task_refs);
7a743e22
PB
1440}
1441
c450178d 1442static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
a141dd89 1443 __must_hold(&ctx->uring_lock)
905c172f 1444{
6f33b0bc 1445 struct io_wq_work_node *node, *prev;
cd0ca2e0 1446 struct io_submit_state *state = &ctx->submit_state;
905c172f 1447
f66f7342 1448 __io_cq_lock(ctx);
931147dd
DY
1449 /* must come first to preserve CQE ordering in failure cases */
1450 if (state->cqes_count)
1451 __io_flush_post_cqes(ctx);
d9dee430
PB
1452 wq_list_for_each(node, prev, &state->compl_reqs) {
1453 struct io_kiocb *req = container_of(node, struct io_kiocb,
1454 comp_list);
3d4aeb9f 1455
f66f7342
PB
1456 if (!(req->flags & REQ_F_CQE_SKIP) &&
1457 unlikely(!__io_fill_cqe_req(ctx, req))) {
1458 if (ctx->task_complete) {
1459 spin_lock(&ctx->completion_lock);
1460 io_req_cqe_overflow(req);
1461 spin_unlock(&ctx->completion_lock);
1462 } else {
1463 io_req_cqe_overflow(req);
1464 }
1465 }
905c172f 1466 }
f66f7342 1467 __io_cq_unlock_post(ctx);
d9dee430 1468
931147dd
DY
1469 if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1470 io_free_batch_list(ctx, state->compl_reqs.first);
1471 INIT_WQ_LIST(&state->compl_reqs);
1472 }
7a743e22
PB
1473}
1474
ba816ad6
JA
1475/*
1476 * Drop reference to request, return next in chain (if there is one) if this
1477 * was the last reference to this request.
1478 */
0d85035a 1479static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
e65ef56d 1480{
9b5f7bd9
PB
1481 struct io_kiocb *nxt = NULL;
1482
de9b4cca 1483 if (req_ref_put_and_test(req)) {
da1a08c5 1484 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
7819a1f6 1485 nxt = io_req_find_next(req);
f5c6cf2a 1486 io_free_req(req);
2a44f467 1487 }
9b5f7bd9 1488 return nxt;
2b188cc1
JA
1489}
1490
6c503150 1491static unsigned io_cqring_events(struct io_ring_ctx *ctx)
a3a0e43f
JA
1492{
1493 /* See comment at the top of this file */
1494 smp_rmb();
e23de15f 1495 return __io_cqring_events(ctx);
a3a0e43f
JA
1496}
1497
def596e9
JA
1498/*
1499 * We can't just wait for polled events to come to us, we have to actively
1500 * find and complete them.
1501 */
c072481d 1502static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
def596e9
JA
1503{
1504 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1505 return;
1506
1507 mutex_lock(&ctx->uring_lock);
5eef4e87 1508 while (!wq_list_empty(&ctx->iopoll_list)) {
b2edc0a7 1509 /* let it sleep and repeat later if can't complete a request */
5ba3c874 1510 if (io_do_iopoll(ctx, true) == 0)
b2edc0a7 1511 break;
08f5439f
JA
1512 /*
1513 * Ensure we allow local-to-the-cpu processing to take place,
1514 * in this case we need to ensure that we reap all events.
3fcee5a6 1515 * Also let task_work, etc. to progress by releasing the mutex
08f5439f 1516 */
3fcee5a6
PB
1517 if (need_resched()) {
1518 mutex_unlock(&ctx->uring_lock);
1519 cond_resched();
1520 mutex_lock(&ctx->uring_lock);
1521 }
def596e9
JA
1522 }
1523 mutex_unlock(&ctx->uring_lock);
1524}
1525
7668b92a 1526static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
def596e9 1527{
7668b92a 1528 unsigned int nr_events = 0;
e9979b36 1529 int ret = 0;
155bc950 1530 unsigned long check_cq;
500f9fba 1531
76de6749
PB
1532 if (!io_allowed_run_tw(ctx))
1533 return -EEXIST;
1534
3a08576b
PB
1535 check_cq = READ_ONCE(ctx->check_cq);
1536 if (unlikely(check_cq)) {
1537 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
a85381d8 1538 __io_cqring_overflow_flush(ctx);
3a08576b
PB
1539 /*
1540 * Similarly do not spin if we have not informed the user of any
1541 * dropped CQE.
1542 */
1543 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1544 return -EBADR;
1545 }
f39c8a5b
PB
1546 /*
1547 * Don't enter poll loop if we already have events pending.
1548 * If we do, we can potentially be spinning for commands that
1549 * already triggered a CQE (eg in error).
1550 */
f39c8a5b 1551 if (io_cqring_events(ctx))
d487b43c 1552 return 0;
155bc950 1553
def596e9 1554 do {
500f9fba
JA
1555 /*
1556 * If a submit got punted to a workqueue, we can have the
1557 * application entering polling for a command before it gets
1558 * issued. That app will hold the uring_lock for the duration
1559 * of the poll right here, so we need to take a breather every
1560 * now and then to ensure that the issue has a chance to add
1561 * the poll to the issued list. Otherwise we can spin here
1562 * forever, while the workqueue is stuck trying to acquire the
1563 * very same mutex.
1564 */
dac6a0ea
JA
1565 if (wq_list_empty(&ctx->iopoll_list) ||
1566 io_task_work_pending(ctx)) {
8f487ef2
PB
1567 u32 tail = ctx->cached_cq_tail;
1568
8de11cdc 1569 (void) io_run_local_work_locked(ctx);
def596e9 1570
dac6a0ea
JA
1571 if (task_work_pending(current) ||
1572 wq_list_empty(&ctx->iopoll_list)) {
dac6a0ea 1573 mutex_unlock(&ctx->uring_lock);
9d54bd6a 1574 io_run_task_work();
dac6a0ea 1575 mutex_lock(&ctx->uring_lock);
dac6a0ea 1576 }
8f487ef2
PB
1577 /* some requests don't go through iopoll_list */
1578 if (tail != ctx->cached_cq_tail ||
5eef4e87 1579 wq_list_empty(&ctx->iopoll_list))
e9979b36 1580 break;
500f9fba 1581 }
5ba3c874
PB
1582 ret = io_do_iopoll(ctx, !min);
1583 if (ret < 0)
1584 break;
1585 nr_events += ret;
1586 ret = 0;
1587 } while (nr_events < min && !need_resched());
d487b43c 1588
def596e9
JA
1589 return ret;
1590}
7012c815
PB
1591
1592void io_req_task_complete(struct io_kiocb *req, bool *locked)
8ef12efe 1593{
7012c815 1594 if (*locked)
9da070b1 1595 io_req_complete_defer(req);
7012c815 1596 else
27f35fe9 1597 io_req_complete_post(req, IO_URING_F_UNLOCKED);
8ef12efe
JA
1598}
1599
def596e9
JA
1600/*
1601 * After the iocb has been issued, it's safe to be found on the poll list.
1602 * Adding the kiocb to the list AFTER submission ensures that we don't
f39c8a5b 1603 * find it from a io_do_iopoll() thread before the issuer is done
def596e9
JA
1604 * accessing the kiocb cookie.
1605 */
9882131c 1606static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
def596e9
JA
1607{
1608 struct io_ring_ctx *ctx = req->ctx;
3b44b371 1609 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
cb3d8972
PB
1610
1611 /* workqueue context doesn't hold uring_lock, grab it now */
3b44b371 1612 if (unlikely(needs_lock))
cb3d8972 1613 mutex_lock(&ctx->uring_lock);
def596e9
JA
1614
1615 /*
1616 * Track whether we have multiple files in our lists. This will impact
1617 * how we do polling eventually, not spinning if we're on potentially
1618 * different devices.
1619 */
5eef4e87 1620 if (wq_list_empty(&ctx->iopoll_list)) {
915b3dde
HX
1621 ctx->poll_multi_queue = false;
1622 } else if (!ctx->poll_multi_queue) {
def596e9
JA
1623 struct io_kiocb *list_req;
1624
5eef4e87
PB
1625 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1626 comp_list);
30da1b45 1627 if (list_req->file != req->file)
915b3dde 1628 ctx->poll_multi_queue = true;
def596e9
JA
1629 }
1630
1631 /*
1632 * For fast devices, IO may have already completed. If it has, add
1633 * it to the front so we find it first.
1634 */
65a6543d 1635 if (READ_ONCE(req->iopoll_completed))
5eef4e87 1636 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
def596e9 1637 else
5eef4e87 1638 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
bdcd3eab 1639
3b44b371 1640 if (unlikely(needs_lock)) {
cb3d8972
PB
1641 /*
1642 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1643 * in sq thread task context or in io worker task context. If
1644 * current task context is sq thread, we don't need to check
1645 * whether should wake up sq thread.
1646 */
1647 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1648 wq_has_sleeper(&ctx->sq_data->wait))
1649 wake_up(&ctx->sq_data->wait);
1650
1651 mutex_unlock(&ctx->uring_lock);
1652 }
def596e9
JA
1653}
1654
4503b767
JA
1655static bool io_bdev_nowait(struct block_device *bdev)
1656{
568ec936 1657 return !bdev || bdev_nowait(bdev);
4503b767
JA
1658}
1659
2b188cc1
JA
1660/*
1661 * If we tracked the file through the SCM inflight mechanism, we could support
1662 * any file. For now, just ensure that anything potentially problematic is done
1663 * inline.
1664 */
88459b50 1665static bool __io_file_supports_nowait(struct file *file, umode_t mode)
2b188cc1 1666{
4503b767 1667 if (S_ISBLK(mode)) {
4e7b5671
CH
1668 if (IS_ENABLED(CONFIG_BLOCK) &&
1669 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
4503b767
JA
1670 return true;
1671 return false;
1672 }
976517f1 1673 if (S_ISSOCK(mode))
2b188cc1 1674 return true;
4503b767 1675 if (S_ISREG(mode)) {
4e7b5671
CH
1676 if (IS_ENABLED(CONFIG_BLOCK) &&
1677 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
e5550a14 1678 !io_is_uring_fops(file))
4503b767
JA
1679 return true;
1680 return false;
1681 }
2b188cc1 1682
c5b85625
JA
1683 /* any ->read/write should understand O_NONBLOCK */
1684 if (file->f_flags & O_NONBLOCK)
1685 return true;
35645ac3 1686 return file->f_mode & FMODE_NOWAIT;
2b188cc1 1687}
c5b85625 1688
88459b50
PB
1689/*
1690 * If we tracked the file through the SCM inflight mechanism, we could support
1691 * any file. For now, just ensure that anything potentially problematic is done
1692 * inline.
1693 */
a4ad4f74 1694unsigned int io_file_get_flags(struct file *file)
88459b50
PB
1695{
1696 umode_t mode = file_inode(file)->i_mode;
1697 unsigned int res = 0;
af197f50 1698
88459b50
PB
1699 if (S_ISREG(mode))
1700 res |= FFS_ISREG;
1701 if (__io_file_supports_nowait(file, mode))
1702 res |= FFS_NOWAIT;
1703 return res;
2b188cc1
JA
1704}
1705
99f15d8d 1706bool io_alloc_async_data(struct io_kiocb *req)
3d9932a8 1707{
e8c2bc1f
JA
1708 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1709 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
d886e185
PB
1710 if (req->async_data) {
1711 req->flags |= REQ_F_ASYNC_DATA;
1712 return false;
1713 }
1714 return true;
3d9932a8
XW
1715}
1716
f3b44f92 1717int io_req_prep_async(struct io_kiocb *req)
f67676d1 1718{
0702e536
JA
1719 const struct io_op_def *def = &io_op_defs[req->opcode];
1720
1721 /* assign early for deferred execution for non-fixed file */
1722 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1723 req->file = io_file_get_normal(req, req->cqe.fd);
dc919caf 1724 if (!def->prep_async)
0702e536
JA
1725 return 0;
1726 if (WARN_ON_ONCE(req_has_async_data(req)))
1727 return -EFAULT;
59169439
PB
1728 if (!io_op_defs[req->opcode].manual_alloc) {
1729 if (io_alloc_async_data(req))
1730 return -EAGAIN;
1731 }
dc919caf 1732 return def->prep_async(req);
bfe76559
PB
1733}
1734
9cf7c104
PB
1735static u32 io_get_sequence(struct io_kiocb *req)
1736{
a3dbdf54 1737 u32 seq = req->ctx->cached_sq_head;
963c6abb 1738 struct io_kiocb *cur;
9cf7c104 1739
a3dbdf54 1740 /* need original cached_sq_head, but it was increased for each req */
963c6abb 1741 io_for_each_link(cur, req)
a3dbdf54
PB
1742 seq--;
1743 return seq;
9cf7c104
PB
1744}
1745
c072481d 1746static __cold void io_drain_req(struct io_kiocb *req)
e276ae34 1747 __must_hold(&ctx->uring_lock)
de0617e4 1748{
a197f664 1749 struct io_ring_ctx *ctx = req->ctx;
27dc8338 1750 struct io_defer_entry *de;
f67676d1 1751 int ret;
e0eb71dc 1752 u32 seq = io_get_sequence(req);
3c19966d 1753
9d858b21 1754 /* Still need defer if there is pending req in defer list. */
e302f104 1755 spin_lock(&ctx->completion_lock);
5e371265 1756 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
e302f104 1757 spin_unlock(&ctx->completion_lock);
e0eb71dc 1758queue:
10c66904 1759 ctx->drain_active = false;
e0eb71dc
PB
1760 io_req_task_queue(req);
1761 return;
10c66904 1762 }
e302f104 1763 spin_unlock(&ctx->completion_lock);
9cf7c104 1764
cbdcb435 1765 io_prep_async_link(req);
27dc8338 1766 de = kmalloc(sizeof(*de), GFP_KERNEL);
76cc33d7 1767 if (!de) {
1b48773f 1768 ret = -ENOMEM;
ef5c600a
DY
1769 io_req_defer_failed(req, ret);
1770 return;
76cc33d7 1771 }
2d28390a 1772
79ebeaee 1773 spin_lock(&ctx->completion_lock);
9cf7c104 1774 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
79ebeaee 1775 spin_unlock(&ctx->completion_lock);
27dc8338 1776 kfree(de);
e0eb71dc 1777 goto queue;
de0617e4
JA
1778 }
1779
48863ffd 1780 trace_io_uring_defer(req);
27dc8338 1781 de->req = req;
9cf7c104 1782 de->seq = seq;
27dc8338 1783 list_add_tail(&de->list, &ctx->defer_list);
79ebeaee 1784 spin_unlock(&ctx->completion_lock);
de0617e4
JA
1785}
1786
68fb8979 1787static void io_clean_op(struct io_kiocb *req)
99bc4c38 1788{
8197b053
PB
1789 if (req->flags & REQ_F_BUFFER_SELECTED) {
1790 spin_lock(&req->ctx->completion_lock);
cc3cec83 1791 io_put_kbuf_comp(req);
8197b053
PB
1792 spin_unlock(&req->ctx->completion_lock);
1793 }
99bc4c38 1794
0e1b6fe3 1795 if (req->flags & REQ_F_NEED_CLEANUP) {
4d4c9cff 1796 const struct io_op_def *def = &io_op_defs[req->opcode];
bb040a21 1797
4d4c9cff
JA
1798 if (def->cleanup)
1799 def->cleanup(req);
99bc4c38 1800 }
75652a30
JA
1801 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1802 kfree(req->apoll->double_poll);
1803 kfree(req->apoll);
1804 req->apoll = NULL;
1805 }
9cae36a0
JA
1806 if (req->flags & REQ_F_INFLIGHT) {
1807 struct io_uring_task *tctx = req->task->io_uring;
1808
1809 atomic_dec(&tctx->inflight_tracked);
1810 }
c854357b 1811 if (req->flags & REQ_F_CREDS)
b8e64b53 1812 put_cred(req->creds);
d886e185
PB
1813 if (req->flags & REQ_F_ASYNC_DATA) {
1814 kfree(req->async_data);
1815 req->async_data = NULL;
1816 }
c854357b 1817 req->flags &= ~IO_REQ_CLEAN_FLAGS;
99bc4c38
PB
1818}
1819
6bf9c47a
JA
1820static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1821{
1822 if (req->file || !io_op_defs[req->opcode].needs_file)
1823 return true;
1824
1825 if (req->flags & REQ_F_FIXED_FILE)
cef216fc 1826 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
6bf9c47a 1827 else
cef216fc 1828 req->file = io_file_get_normal(req, req->cqe.fd);
6bf9c47a 1829
772f5e00 1830 return !!req->file;
6bf9c47a
JA
1831}
1832
889fca73 1833static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
2b188cc1 1834{
fcde59fe 1835 const struct io_op_def *def = &io_op_defs[req->opcode];
5730b27e 1836 const struct cred *creds = NULL;
d625c6ee 1837 int ret;
2b188cc1 1838
70152140
JA
1839 if (unlikely(!io_assign_file(req, issue_flags)))
1840 return -EBADF;
1841
6878b40e 1842 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
c10d1f98 1843 creds = override_creds(req->creds);
5730b27e 1844
fcde59fe 1845 if (!def->audit_skip)
5bd2182d
PM
1846 audit_uring_entry(req->opcode);
1847
0702e536 1848 ret = def->issue(req, issue_flags);
2b188cc1 1849
fcde59fe 1850 if (!def->audit_skip)
5bd2182d
PM
1851 audit_uring_exit(!ret, ret);
1852
5730b27e
JA
1853 if (creds)
1854 revert_creds(creds);
97b388d7 1855
75d7b3ae
PB
1856 if (ret == IOU_OK) {
1857 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
9da070b1 1858 io_req_complete_defer(req);
75d7b3ae 1859 else
1bec951c 1860 io_req_complete_post(req, issue_flags);
75d7b3ae 1861 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
def596e9 1862 return ret;
97b388d7 1863
b532576e 1864 /* If the op doesn't have a file, we're not polling for it */
ef0ec1ad 1865 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
9882131c 1866 io_iopoll_req_issued(req, issue_flags);
def596e9
JA
1867
1868 return 0;
2b188cc1
JA
1869}
1870
329061d3
JA
1871int io_poll_issue(struct io_kiocb *req, bool *locked)
1872{
1873 io_tw_lock(req->ctx, locked);
9a692451
DY
1874 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1875 IO_URING_F_COMPLETE_DEFER);
329061d3
JA
1876}
1877
c9f06aa7 1878struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
ebc11b6c
PB
1879{
1880 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1881
1882 req = io_put_req_find_next(req);
1883 return req ? &req->work : NULL;
1884}
1885
c9f06aa7 1886void io_wq_submit_work(struct io_wq_work *work)
2b188cc1
JA
1887{
1888 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6bf9c47a 1889 const struct io_op_def *def = &io_op_defs[req->opcode];
e6aeb272 1890 unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
d01905db 1891 bool needs_poll = false;
6bf9c47a 1892 int ret = 0, err = -ECANCELED;
2b188cc1 1893
23a6c9ac 1894 /* one will be dropped by ->io_wq_free_work() after returning to io-wq */
48dcd38d
PB
1895 if (!(req->flags & REQ_F_REFCOUNT))
1896 __io_req_set_refcount(req, 2);
1897 else
1898 req_ref_get(req);
5d5901a3 1899
cb2d344c 1900 io_arm_ltimeout(req);
6bf9c47a 1901
dadebc35 1902 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
d01905db 1903 if (work->flags & IO_WQ_WORK_CANCEL) {
0f8da75b 1904fail:
6bf9c47a 1905 io_req_task_queue_fail(req, err);
d01905db
PB
1906 return;
1907 }
0f8da75b
PB
1908 if (!io_assign_file(req, issue_flags)) {
1909 err = -EBADF;
1910 work->flags |= IO_WQ_WORK_CANCEL;
1911 goto fail;
1912 }
31b51510 1913
d01905db 1914 if (req->flags & REQ_F_FORCE_ASYNC) {
afb7f56f
PB
1915 bool opcode_poll = def->pollin || def->pollout;
1916
1917 if (opcode_poll && file_can_poll(req->file)) {
1918 needs_poll = true;
d01905db 1919 issue_flags |= IO_URING_F_NONBLOCK;
afb7f56f 1920 }
561fb04a 1921 }
31b51510 1922
d01905db
PB
1923 do {
1924 ret = io_issue_sqe(req, issue_flags);
1925 if (ret != -EAGAIN)
1926 break;
1927 /*
1928 * We can get EAGAIN for iopolled IO even though we're
1929 * forcing a sync submission from here, since we can't
1930 * wait for request slots on the block side.
1931 */
1932 if (!needs_poll) {
e0deb6a0
PB
1933 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1934 break;
d01905db
PB
1935 cond_resched();
1936 continue;
90fa0288
HX
1937 }
1938
4d9237e3 1939 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
d01905db
PB
1940 return;
1941 /* aborted or ready, in either case retry blocking */
1942 needs_poll = false;
1943 issue_flags &= ~IO_URING_F_NONBLOCK;
1944 } while (1);
31b51510 1945
a3df7698 1946 /* avoid locking problems by failing it from a clean context */
97b388d7 1947 if (ret < 0)
a3df7698 1948 io_req_task_queue_fail(req, ret);
2b188cc1
JA
1949}
1950
531113bb
JA
1951inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1952 unsigned int issue_flags)
09bb8394 1953{
5106dd6e
JA
1954 struct io_ring_ctx *ctx = req->ctx;
1955 struct file *file = NULL;
ac177053 1956 unsigned long file_ptr;
09bb8394 1957
93f052cb 1958 io_ring_submit_lock(ctx, issue_flags);
5106dd6e 1959
ac177053 1960 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
5106dd6e 1961 goto out;
ac177053
PB
1962 fd = array_index_nospec(fd, ctx->nr_user_files);
1963 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1964 file = (struct file *) (file_ptr & FFS_MASK);
1965 file_ptr &= ~FFS_MASK;
1966 /* mask in overlapping REQ_F and FFS bits */
35645ac3 1967 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
5106dd6e
JA
1968 io_req_set_rsrc_node(req, ctx, 0);
1969out:
93f052cb 1970 io_ring_submit_unlock(ctx, issue_flags);
ac177053
PB
1971 return file;
1972}
d44f554e 1973
531113bb 1974struct file *io_file_get_normal(struct io_kiocb *req, int fd)
ac177053 1975{
62906e89 1976 struct file *file = fget(fd);
ac177053 1977
48863ffd 1978 trace_io_uring_file_get(req, fd);
09bb8394 1979
ac177053 1980 /* we don't allow fixed io_uring files */
e5550a14 1981 if (file && io_is_uring_fops(file))
9cae36a0 1982 io_req_track_inflight(req);
8371adf5 1983 return file;
09bb8394
JA
1984}
1985
7bfa9bad 1986static void io_queue_async(struct io_kiocb *req, int ret)
d475a9a6
PB
1987 __must_hold(&req->ctx->uring_lock)
1988{
7bfa9bad
PB
1989 struct io_kiocb *linked_timeout;
1990
1991 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
973fc83f 1992 io_req_defer_failed(req, ret);
7bfa9bad
PB
1993 return;
1994 }
1995
1996 linked_timeout = io_prep_linked_timeout(req);
d475a9a6 1997
4d9237e3 1998 switch (io_arm_poll_handler(req, 0)) {
d475a9a6 1999 case IO_APOLL_READY:
336d28a8 2000 io_kbuf_recycle(req, 0);
d475a9a6
PB
2001 io_req_task_queue(req);
2002 break;
2003 case IO_APOLL_ABORTED:
6436c770 2004 io_kbuf_recycle(req, 0);
77955efb 2005 io_queue_iowq(req, NULL);
d475a9a6 2006 break;
b1c62645 2007 case IO_APOLL_OK:
b1c62645 2008 break;
d475a9a6
PB
2009 }
2010
2011 if (linked_timeout)
2012 io_queue_linked_timeout(linked_timeout);
2013}
2014
cbc2e203 2015static inline void io_queue_sqe(struct io_kiocb *req)
282cdc86 2016 __must_hold(&req->ctx->uring_lock)
2b188cc1 2017{
e0c5c576 2018 int ret;
2b188cc1 2019
c5eef2b9 2020 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
193155c8 2021
491381ce
JA
2022 /*
2023 * We async punt it if the file wasn't marked NOWAIT, or if the file
2024 * doesn't support non-blocking read/write attempts
2025 */
7bfa9bad 2026 if (likely(!ret))
cb2d344c 2027 io_arm_ltimeout(req);
7bfa9bad
PB
2028 else
2029 io_queue_async(req, ret);
2b188cc1
JA
2030}
2031
4652fe3f 2032static void io_queue_sqe_fallback(struct io_kiocb *req)
282cdc86 2033 __must_hold(&req->ctx->uring_lock)
4fe2c963 2034{
17b147f6
PB
2035 if (unlikely(req->flags & REQ_F_FAIL)) {
2036 /*
2037 * We don't submit, fail them all, for that replace hardlinks
2038 * with normal links. Extra REQ_F_LINK is tolerated.
2039 */
2040 req->flags &= ~REQ_F_HARDLINK;
2041 req->flags |= REQ_F_LINK;
973fc83f 2042 io_req_defer_failed(req, req->cqe.res);
76cc33d7
PB
2043 } else {
2044 int ret = io_req_prep_async(req);
2045
ef5c600a 2046 if (unlikely(ret)) {
973fc83f 2047 io_req_defer_failed(req, ret);
ef5c600a
DY
2048 return;
2049 }
2050
2051 if (unlikely(req->ctx->drain_active))
2052 io_drain_req(req);
76cc33d7 2053 else
77955efb 2054 io_queue_iowq(req, NULL);
ce35a47a 2055 }
4fe2c963
JL
2056}
2057
b16fed66
PB
2058/*
2059 * Check SQE restrictions (opcode and flags).
2060 *
2061 * Returns 'true' if SQE is allowed, 'false' otherwise.
2062 */
2063static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2064 struct io_kiocb *req,
2065 unsigned int sqe_flags)
4fe2c963 2066{
b16fed66
PB
2067 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2068 return false;
2069
2070 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2071 ctx->restrictions.sqe_flags_required)
2072 return false;
2073
2074 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2075 ctx->restrictions.sqe_flags_required))
2076 return false;
2077
2078 return true;
4fe2c963
JL
2079}
2080
22b2ca31
PB
2081static void io_init_req_drain(struct io_kiocb *req)
2082{
2083 struct io_ring_ctx *ctx = req->ctx;
2084 struct io_kiocb *head = ctx->submit_state.link.head;
2085
2086 ctx->drain_active = true;
2087 if (head) {
2088 /*
2089 * If we need to drain a request in the middle of a link, drain
2090 * the head request and the next request/link after the current
2091 * link. Considering sequential execution of links,
b6c7db32 2092 * REQ_F_IO_DRAIN will be maintained for every request of our
22b2ca31
PB
2093 * link.
2094 */
b6c7db32 2095 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
22b2ca31
PB
2096 ctx->drain_next = true;
2097 }
2098}
2099
b16fed66
PB
2100static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2101 const struct io_uring_sqe *sqe)
282cdc86 2102 __must_hold(&ctx->uring_lock)
b16fed66 2103{
fcde59fe 2104 const struct io_op_def *def;
b16fed66 2105 unsigned int sqe_flags;
fc0ae024 2106 int personality;
4a04d1d1 2107 u8 opcode;
b16fed66 2108
864ea921 2109 /* req is partially pre-initialised, see io_preinit_req() */
4a04d1d1 2110 req->opcode = opcode = READ_ONCE(sqe->opcode);
b16fed66
PB
2111 /* same numerical values with corresponding REQ_F_*, safe to copy */
2112 req->flags = sqe_flags = READ_ONCE(sqe->flags);
cef216fc 2113 req->cqe.user_data = READ_ONCE(sqe->user_data);
b16fed66 2114 req->file = NULL;
c1bdf8ed 2115 req->rsrc_node = NULL;
b16fed66 2116 req->task = current;
b16fed66 2117
4a04d1d1
PB
2118 if (unlikely(opcode >= IORING_OP_LAST)) {
2119 req->opcode = 0;
b16fed66 2120 return -EINVAL;
4a04d1d1 2121 }
fcde59fe 2122 def = &io_op_defs[opcode];
68fe256a
PB
2123 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2124 /* enforce forwards compatibility on users */
2125 if (sqe_flags & ~SQE_VALID_FLAGS)
2126 return -EINVAL;
4e906702 2127 if (sqe_flags & IOSQE_BUFFER_SELECT) {
fcde59fe 2128 if (!def->buffer_select)
4e906702
JA
2129 return -EOPNOTSUPP;
2130 req->buf_index = READ_ONCE(sqe->buf_group);
2131 }
5562a8d7
PB
2132 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2133 ctx->drain_disabled = true;
2134 if (sqe_flags & IOSQE_IO_DRAIN) {
2135 if (ctx->drain_disabled)
2136 return -EOPNOTSUPP;
22b2ca31 2137 io_init_req_drain(req);
5562a8d7 2138 }
2a56a9bd
PB
2139 }
2140 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2141 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2142 return -EACCES;
2143 /* knock it to the slow queue path, will be drained there */
2144 if (ctx->drain_active)
2145 req->flags |= REQ_F_FORCE_ASYNC;
2146 /* if there is no link, we're at "next" request and need to drain */
2147 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2148 ctx->drain_next = false;
2149 ctx->drain_active = true;
b6c7db32 2150 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2a56a9bd 2151 }
68fe256a 2152 }
b16fed66 2153
fcde59fe 2154 if (!def->ioprio && sqe->ioprio)
73911426 2155 return -EINVAL;
fcde59fe 2156 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
73911426
JA
2157 return -EINVAL;
2158
fcde59fe 2159 if (def->needs_file) {
6d63416d
PB
2160 struct io_submit_state *state = &ctx->submit_state;
2161
cef216fc 2162 req->cqe.fd = READ_ONCE(sqe->fd);
6bf9c47a 2163
6d63416d
PB
2164 /*
2165 * Plug now if we have more than 2 IO left after this, and the
2166 * target is potentially a read/write to block based storage.
2167 */
fcde59fe 2168 if (state->need_plug && def->plug) {
6d63416d
PB
2169 state->plug_started = true;
2170 state->need_plug = false;
5ca7a8b3 2171 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
6d63416d 2172 }
b16fed66 2173 }
863e0560 2174
003e8dcc
JA
2175 personality = READ_ONCE(sqe->personality);
2176 if (personality) {
cdab10bf
LT
2177 int ret;
2178
c10d1f98
PB
2179 req->creds = xa_load(&ctx->personalities, personality);
2180 if (!req->creds)
003e8dcc 2181 return -EINVAL;
c10d1f98 2182 get_cred(req->creds);
cdc1404a
PM
2183 ret = security_uring_override_creds(req->creds);
2184 if (ret) {
2185 put_cred(req->creds);
2186 return ret;
2187 }
b8e64b53 2188 req->flags |= REQ_F_CREDS;
003e8dcc 2189 }
b16fed66 2190
0702e536 2191 return def->prep(req, sqe);
b16fed66
PB
2192}
2193
df3becde
PB
2194static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2195 struct io_kiocb *req, int ret)
2196{
2197 struct io_ring_ctx *ctx = req->ctx;
2198 struct io_submit_link *link = &ctx->submit_state.link;
2199 struct io_kiocb *head = link->head;
2200
48863ffd 2201 trace_io_uring_req_failed(sqe, req, ret);
df3becde
PB
2202
2203 /*
2204 * Avoid breaking links in the middle as it renders links with SQPOLL
2205 * unusable. Instead of failing eagerly, continue assembling the link if
2206 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2207 * should find the flag and handle the rest.
2208 */
2209 req_fail_link_node(req, ret);
2210 if (head && !(head->flags & REQ_F_FAIL))
2211 req_fail_link_node(head, -ECANCELED);
2212
2213 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2214 if (head) {
2215 link->last->link = req;
2216 link->head = NULL;
2217 req = head;
2218 }
2219 io_queue_sqe_fallback(req);
2220 return ret;
2221 }
2222
2223 if (head)
2224 link->last->link = req;
2225 else
2226 link->head = req;
2227 link->last = req;
2228 return 0;
2229}
2230
2231static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
a1ab7b35 2232 const struct io_uring_sqe *sqe)
282cdc86 2233 __must_hold(&ctx->uring_lock)
9e645e11 2234{
a1ab7b35 2235 struct io_submit_link *link = &ctx->submit_state.link;
ef4ff581 2236 int ret;
9e645e11 2237
a6b8cadc 2238 ret = io_init_req(ctx, req, sqe);
df3becde
PB
2239 if (unlikely(ret))
2240 return io_submit_fail_init(sqe, req, ret);
441b8a78 2241
be7053b7 2242 /* don't need @sqe from now on */
48863ffd 2243 trace_io_uring_submit_sqe(req, true);
a6b8cadc 2244
9e645e11
JA
2245 /*
2246 * If we already have a head request, queue this one for async
2247 * submittal once the head completes. If we don't have a head but
2248 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2249 * submitted sync once the chain is complete. If none of those
2250 * conditions are true (normal request), then just queue it.
2251 */
924a07e4 2252 if (unlikely(link->head)) {
df3becde
PB
2253 ret = io_req_prep_async(req);
2254 if (unlikely(ret))
2255 return io_submit_fail_init(sqe, req, ret);
2256
48863ffd 2257 trace_io_uring_link(req, link->head);
f2f87370 2258 link->last->link = req;
863e0560 2259 link->last = req;
32fe525b 2260
da1a08c5 2261 if (req->flags & IO_REQ_LINK_FLAGS)
f15a3431 2262 return 0;
df3becde
PB
2263 /* last request of the link, flush it */
2264 req = link->head;
f15a3431 2265 link->head = NULL;
924a07e4
PB
2266 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2267 goto fallback;
2268
2269 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2270 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2271 if (req->flags & IO_REQ_LINK_FLAGS) {
2272 link->head = req;
2273 link->last = req;
2274 } else {
2275fallback:
2276 io_queue_sqe_fallback(req);
2277 }
f15a3431 2278 return 0;
9e645e11 2279 }
2e6e1fde 2280
f15a3431 2281 io_queue_sqe(req);
1d4240cc 2282 return 0;
9e645e11
JA
2283}
2284
9a56a232
JA
2285/*
2286 * Batched submission is done, ensure local IO is flushed out.
2287 */
553deffd 2288static void io_submit_state_end(struct io_ring_ctx *ctx)
9a56a232 2289{
553deffd
PB
2290 struct io_submit_state *state = &ctx->submit_state;
2291
e126391c
PB
2292 if (unlikely(state->link.head))
2293 io_queue_sqe_fallback(state->link.head);
553deffd 2294 /* flush only after queuing links as they can generate completions */
c450178d 2295 io_submit_flush_completions(ctx);
27926b68
JA
2296 if (state->plug_started)
2297 blk_finish_plug(&state->plug);
9a56a232
JA
2298}
2299
2300/*
2301 * Start submission side cache.
2302 */
2303static void io_submit_state_start(struct io_submit_state *state,
ba88ff11 2304 unsigned int max_ios)
9a56a232 2305{
27926b68 2306 state->plug_started = false;
4b628aeb 2307 state->need_plug = max_ios > 2;
5ca7a8b3 2308 state->submit_nr = max_ios;
a1ab7b35
PB
2309 /* set only head, no need to init link_last in advance */
2310 state->link.head = NULL;
9a56a232
JA
2311}
2312
2b188cc1
JA
2313static void io_commit_sqring(struct io_ring_ctx *ctx)
2314{
75b28aff 2315 struct io_rings *rings = ctx->rings;
2b188cc1 2316
caf582c6
PB
2317 /*
2318 * Ensure any loads from the SQEs are done at this point,
2319 * since once we write the new head, the application could
2320 * write new data to them.
2321 */
2322 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2b188cc1
JA
2323}
2324
2b188cc1 2325/*
dd9ae8a0 2326 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2b188cc1
JA
2327 * that is mapped by userspace. This means that care needs to be taken to
2328 * ensure that reads are stable, as we cannot rely on userspace always
2329 * being a good citizen. If members of the sqe are validated and then later
2330 * used, it's important that those reads are done through READ_ONCE() to
2331 * prevent a re-load down the line.
2332 */
709b302f 2333static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2b188cc1 2334{
ea5ab3b5 2335 unsigned head, mask = ctx->sq_entries - 1;
17d3aeb3 2336 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2b188cc1
JA
2337
2338 /*
2339 * The cached sq head (or cq tail) serves two purposes:
2340 *
2341 * 1) allows us to batch the cost of updating the user visible
2342 * head updates.
2343 * 2) allows the kernel side to track the head on its own, even
2344 * though the application is the one updating it.
2345 */
17d3aeb3 2346 head = READ_ONCE(ctx->sq_array[sq_idx]);
ebdeb7c0
JA
2347 if (likely(head < ctx->sq_entries)) {
2348 /* double index for 128-byte SQEs, twice as long */
2349 if (ctx->flags & IORING_SETUP_SQE128)
2350 head <<= 1;
709b302f 2351 return &ctx->sq_sqes[head];
ebdeb7c0 2352 }
2b188cc1
JA
2353
2354 /* drop invalid entries */
15641e42
PB
2355 ctx->cq_extra--;
2356 WRITE_ONCE(ctx->rings->sq_dropped,
2357 READ_ONCE(ctx->rings->sq_dropped) + 1);
709b302f
PB
2358 return NULL;
2359}
2360
17437f31 2361int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
282cdc86 2362 __must_hold(&ctx->uring_lock)
6c271ce2 2363{
69629809 2364 unsigned int entries = io_sqring_entries(ctx);
8e6971a8
PB
2365 unsigned int left;
2366 int ret;
6c271ce2 2367
51d48dab 2368 if (unlikely(!entries))
69629809 2369 return 0;
ee7d46d9 2370 /* make sure SQ entry isn't read before tail */
8e6971a8
PB
2371 ret = left = min3(nr, ctx->sq_entries, entries);
2372 io_get_task_refs(left);
2373 io_submit_state_start(&ctx->submit_state, left);
6c271ce2 2374
69629809 2375 do {
3529d8c2 2376 const struct io_uring_sqe *sqe;
196be95c 2377 struct io_kiocb *req;
fb5ccc98 2378
8e6971a8 2379 if (unlikely(!io_alloc_req_refill(ctx)))
fb5ccc98 2380 break;
a33ae9ce 2381 req = io_alloc_req(ctx);
4fccfcbb
PB
2382 sqe = io_get_sqe(ctx);
2383 if (unlikely(!sqe)) {
fa05457a 2384 io_req_add_to_cache(req, ctx);
4fccfcbb
PB
2385 break;
2386 }
6c271ce2 2387
1cd15904
PB
2388 /*
2389 * Continue submitting even for sqe failure if the
2390 * ring was setup with IORING_SETUP_SUBMIT_ALL
2391 */
2392 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2393 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2394 left--;
2395 break;
bcbb7bf6 2396 }
1cd15904 2397 } while (--left);
9466f437 2398
8e6971a8
PB
2399 if (unlikely(left)) {
2400 ret -= left;
2401 /* try again if it submitted nothing and can't allocate a req */
2402 if (!ret && io_req_cache_empty(ctx))
2403 ret = -EAGAIN;
2404 current->io_uring->cached_refs += left;
9466f437 2405 }
6c271ce2 2406
553deffd 2407 io_submit_state_end(ctx);
ae9428ca
PB
2408 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2409 io_commit_sqring(ctx);
8e6971a8 2410 return ret;
6c271ce2
JA
2411}
2412
bda52162
JA
2413struct io_wait_queue {
2414 struct wait_queue_entry wq;
2415 struct io_ring_ctx *ctx;
5fd46178 2416 unsigned cq_tail;
bda52162
JA
2417 unsigned nr_timeouts;
2418};
2419
b4c98d59
DY
2420static inline bool io_has_work(struct io_ring_ctx *ctx)
2421{
c0e0d6ba
DY
2422 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2423 ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
2424 !llist_empty(&ctx->work_llist));
b4c98d59
DY
2425}
2426
6c503150 2427static inline bool io_should_wake(struct io_wait_queue *iowq)
bda52162
JA
2428{
2429 struct io_ring_ctx *ctx = iowq->ctx;
0fc8c2ac 2430 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
bda52162
JA
2431
2432 /*
d195a66e 2433 * Wake up if we have enough events, or if a timeout occurred since we
bda52162
JA
2434 * started waiting. For timeouts, we always want to return to userspace,
2435 * regardless of event count.
2436 */
5fd46178 2437 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
bda52162
JA
2438}
2439
2440static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2441 int wake_flags, void *key)
2442{
2443 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2444 wq);
b4c98d59 2445 struct io_ring_ctx *ctx = iowq->ctx;
bda52162 2446
6c503150
PB
2447 /*
2448 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2449 * the task, and the next invocation will do it.
2450 */
b4c98d59 2451 if (io_should_wake(iowq) || io_has_work(ctx))
6c503150
PB
2452 return autoremove_wake_function(curr, mode, wake_flags, key);
2453 return -1;
bda52162
JA
2454}
2455
c0e0d6ba 2456int io_run_task_work_sig(struct io_ring_ctx *ctx)
af9c1a44 2457{
c0e0d6ba 2458 if (io_run_task_work_ctx(ctx) > 0)
af9c1a44 2459 return 1;
c5020bc8
OL
2460 if (task_sigpending(current))
2461 return -EINTR;
2462 return 0;
af9c1a44
JA
2463}
2464
eeb60b9a
PB
2465/* when returns >0, the caller should retry */
2466static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2467 struct io_wait_queue *iowq,
12521a5d 2468 ktime_t *timeout)
eeb60b9a
PB
2469{
2470 int ret;
155bc950 2471 unsigned long check_cq;
eeb60b9a
PB
2472
2473 /* make sure we run task_work before checking for signals */
c0e0d6ba 2474 ret = io_run_task_work_sig(ctx);
eeb60b9a
PB
2475 if (ret || io_should_wake(iowq))
2476 return ret;
3a08576b 2477
155bc950 2478 check_cq = READ_ONCE(ctx->check_cq);
3a08576b
PB
2479 if (unlikely(check_cq)) {
2480 /* let the caller flush overflows, retry */
2481 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2482 return 1;
2483 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2484 return -EBADR;
2485 }
12521a5d 2486 if (!schedule_hrtimeout(timeout, HRTIMER_MODE_ABS))
22833966 2487 return -ETIME;
35d90f95
JA
2488
2489 /*
2490 * Run task_work after scheduling. If we got woken because of
2491 * task_work being processed, run it now rather than let the caller
2492 * do another wait loop.
2493 */
2494 ret = io_run_task_work_sig(ctx);
2495 return ret < 0 ? ret : 1;
eeb60b9a
PB
2496}
2497
2b188cc1
JA
2498/*
2499 * Wait until events become available, if we don't already have some. The
2500 * application must reap them itself, as they reside on the shared cq ring.
2501 */
2502static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
c73ebb68
HX
2503 const sigset_t __user *sig, size_t sigsz,
2504 struct __kernel_timespec __user *uts)
2b188cc1 2505{
90291099 2506 struct io_wait_queue iowq;
75b28aff 2507 struct io_rings *rings = ctx->rings;
22833966 2508 ktime_t timeout = KTIME_MAX;
c1d5a224 2509 int ret;
2b188cc1 2510
76de6749
PB
2511 if (!io_allowed_run_tw(ctx))
2512 return -EEXIST;
2513
b41e9852 2514 do {
c0e0d6ba
DY
2515 /* always run at least 1 task work to process local work */
2516 ret = io_run_task_work_ctx(ctx);
2517 if (ret < 0)
2518 return ret;
90f67366 2519 io_cqring_overflow_flush(ctx);
aa1df3a3 2520
0fc8c2ac
DY
2521 /* if user messes with these they will just get an early return */
2522 if (__io_cqring_events_user(ctx) >= min_events)
b41e9852 2523 return 0;
c0e0d6ba 2524 } while (ret > 0);
2b188cc1
JA
2525
2526 if (sig) {
9e75ad5d
AB
2527#ifdef CONFIG_COMPAT
2528 if (in_compat_syscall())
2529 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
b772434b 2530 sigsz);
9e75ad5d
AB
2531 else
2532#endif
b772434b 2533 ret = set_user_sigmask(sig, sigsz);
9e75ad5d 2534
2b188cc1
JA
2535 if (ret)
2536 return ret;
2537 }
2538
950e79dd
OL
2539 if (uts) {
2540 struct timespec64 ts;
2541
2542 if (get_timespec64(&ts, uts))
2543 return -EFAULT;
2544 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2545 }
2546
90291099
PB
2547 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2548 iowq.wq.private = current;
2549 INIT_LIST_HEAD(&iowq.wq.entry);
2550 iowq.ctx = ctx;
bda52162 2551 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
5fd46178 2552 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
90291099 2553
c826bd7a 2554 trace_io_uring_cqring_wait(ctx, min_events);
bda52162 2555 do {
52ea806a
JA
2556 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2557 finish_wait(&ctx->cq_wait, &iowq.wq);
2558 io_cqring_do_overflow_flush(ctx);
2559 }
311997b3 2560 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
bda52162 2561 TASK_INTERRUPTIBLE);
12521a5d 2562 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
35d90f95
JA
2563 if (__io_cqring_events_user(ctx) >= min_events)
2564 break;
ca0a2651 2565 cond_resched();
eeb60b9a 2566 } while (ret > 0);
bda52162 2567
b4f20bb4 2568 finish_wait(&ctx->cq_wait, &iowq.wq);
b7db41c9 2569 restore_saved_sigmask_unless(ret == -EINTR);
2b188cc1 2570
75b28aff 2571 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2b188cc1
JA
2572}
2573
73572984 2574static void io_mem_free(void *ptr)
b9bd2bea 2575{
73572984 2576 struct page *page;
b36a2050 2577
73572984
JA
2578 if (!ptr)
2579 return;
b9bd2bea 2580
73572984
JA
2581 page = virt_to_head_page(ptr);
2582 if (put_page_testzero(page))
2583 free_compound_page(page);
b9bd2bea
PB
2584}
2585
73572984 2586static void *io_mem_alloc(size_t size)
b9bd2bea 2587{
73572984 2588 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
b9bd2bea 2589
73572984 2590 return (void *) __get_free_pages(gfp, get_order(size));
b9bd2bea
PB
2591}
2592
73572984
JA
2593static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2594 unsigned int cq_entries, size_t *sq_offset)
6b06314c 2595{
73572984
JA
2596 struct io_rings *rings;
2597 size_t off, sq_array_size;
6b06314c 2598
73572984
JA
2599 off = struct_size(rings, cqes, cq_entries);
2600 if (off == SIZE_MAX)
2601 return SIZE_MAX;
2602 if (ctx->flags & IORING_SETUP_CQE32) {
2603 if (check_shl_overflow(off, 1, &off))
2604 return SIZE_MAX;
2605 }
ab409402 2606
73572984
JA
2607#ifdef CONFIG_SMP
2608 off = ALIGN(off, SMP_CACHE_BYTES);
2609 if (off == 0)
2610 return SIZE_MAX;
2611#endif
82fbcfa9 2612
73572984
JA
2613 if (sq_offset)
2614 *sq_offset = off;
82fbcfa9 2615
73572984
JA
2616 sq_array_size = array_size(sizeof(u32), sq_entries);
2617 if (sq_array_size == SIZE_MAX)
2618 return SIZE_MAX;
6b06314c 2619
73572984
JA
2620 if (check_add_overflow(off, sq_array_size, &off))
2621 return SIZE_MAX;
8bad28d8 2622
73572984 2623 return off;
8bad28d8
HX
2624}
2625
73572984
JA
2626static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2627 unsigned int eventfd_async)
8bad28d8 2628{
73572984
JA
2629 struct io_ev_fd *ev_fd;
2630 __s32 __user *fds = arg;
2631 int fd;
f2303b1f 2632
73572984
JA
2633 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2634 lockdep_is_held(&ctx->uring_lock));
2635 if (ev_fd)
2636 return -EBUSY;
8bad28d8 2637
73572984
JA
2638 if (copy_from_user(&fd, fds, sizeof(*fds)))
2639 return -EFAULT;
8dd03afe 2640
73572984
JA
2641 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2642 if (!ev_fd)
2643 return -ENOMEM;
05f3fb3c 2644
73572984
JA
2645 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2646 if (IS_ERR(ev_fd->cq_ev_fd)) {
2647 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2648 kfree(ev_fd);
2649 return ret;
2650 }
305bef98
PB
2651
2652 spin_lock(&ctx->completion_lock);
2653 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2654 spin_unlock(&ctx->completion_lock);
2655
73572984
JA
2656 ev_fd->eventfd_async = eventfd_async;
2657 ctx->has_evfd = true;
2658 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
21a091b9
DY
2659 atomic_set(&ev_fd->refs, 1);
2660 atomic_set(&ev_fd->ops, 0);
73572984 2661 return 0;
d7954b2b
BM
2662}
2663
73572984 2664static int io_eventfd_unregister(struct io_ring_ctx *ctx)
1ad555c6 2665{
73572984
JA
2666 struct io_ev_fd *ev_fd;
2667
2668 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2669 lockdep_is_held(&ctx->uring_lock));
2670 if (ev_fd) {
2671 ctx->has_evfd = false;
2672 rcu_assign_pointer(ctx->io_ev_fd, NULL);
21a091b9
DY
2673 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2674 call_rcu(&ev_fd->rcu, io_eventfd_ops);
73572984
JA
2675 return 0;
2676 }
2d091d62 2677
73572984 2678 return -ENXIO;
44b31f2f
PB
2679}
2680
73572984 2681static void io_req_caches_free(struct io_ring_ctx *ctx)
2b188cc1 2682{
37f0e767 2683 int nr = 0;
bf019da7 2684
9a4fdbd8 2685 mutex_lock(&ctx->uring_lock);
34f0bc42 2686 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
9a4fdbd8 2687
88ab95be 2688 while (!io_req_cache_empty(ctx)) {
34f0bc42 2689 struct io_kiocb *req = io_alloc_req(ctx);
9a4fdbd8 2690
c2b6c6bc 2691 kmem_cache_free(req_cachep, req);
37f0e767 2692 nr++;
c2b6c6bc 2693 }
37f0e767
PB
2694 if (nr)
2695 percpu_ref_put_many(&ctx->refs, nr);
9a4fdbd8
JA
2696 mutex_unlock(&ctx->uring_lock);
2697}
2698
c072481d 2699static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2b188cc1 2700{
37d1e2e3 2701 io_sq_thread_finish(ctx);
ab409402 2702 io_rsrc_refs_drop(ctx);
43597aac
PB
2703 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2704 io_wait_rsrc_data(ctx->buf_data);
2705 io_wait_rsrc_data(ctx->file_data);
2706
8bad28d8 2707 mutex_lock(&ctx->uring_lock);
43597aac 2708 if (ctx->buf_data)
bd54b6fe 2709 __io_sqe_buffers_unregister(ctx);
43597aac 2710 if (ctx->file_data)
08480400 2711 __io_sqe_files_unregister(ctx);
a85381d8 2712 io_cqring_overflow_kill(ctx);
9b402849 2713 io_eventfd_unregister(ctx);
9b797a37 2714 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
43e0bbbd 2715 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
77bc59b4 2716 mutex_unlock(&ctx->uring_lock);
5a2e745d 2717 io_destroy_buffers(ctx);
07db298a
PB
2718 if (ctx->sq_creds)
2719 put_cred(ctx->sq_creds);
97bbdc06
PB
2720 if (ctx->submitter_task)
2721 put_task_struct(ctx->submitter_task);
def596e9 2722
a7f0ed5a
PB
2723 /* there are no registered resources left, nobody uses it */
2724 if (ctx->rsrc_node)
2725 io_rsrc_node_destroy(ctx->rsrc_node);
8dd03afe 2726 if (ctx->rsrc_backup_node)
b895c9a6 2727 io_rsrc_node_destroy(ctx->rsrc_backup_node);
a7f0ed5a 2728 flush_delayed_work(&ctx->rsrc_put_work);
756ab7c0 2729 flush_delayed_work(&ctx->fallback_work);
a7f0ed5a
PB
2730
2731 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2732 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
def596e9 2733
2b188cc1 2734#if defined(CONFIG_UNIX)
355e8d26
EB
2735 if (ctx->ring_sock) {
2736 ctx->ring_sock->file = NULL; /* so that iput() is called */
2b188cc1 2737 sock_release(ctx->ring_sock);
355e8d26 2738 }
2b188cc1 2739#endif
ef9dd637 2740 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2b188cc1 2741
42b6419d
PB
2742 if (ctx->mm_account) {
2743 mmdrop(ctx->mm_account);
2744 ctx->mm_account = NULL;
2745 }
75b28aff 2746 io_mem_free(ctx->rings);
2b188cc1 2747 io_mem_free(ctx->sq_sqes);
2b188cc1
JA
2748
2749 percpu_ref_exit(&ctx->refs);
2b188cc1 2750 free_uid(ctx->user);
4010fec4 2751 io_req_caches_free(ctx);
e941894e
JA
2752 if (ctx->hash_map)
2753 io_wq_put_hash(ctx->hash_map);
e6f89be6 2754 kfree(ctx->cancel_table.hbs);
9ca9fb24 2755 kfree(ctx->cancel_table_locked.hbs);
6224843d 2756 kfree(ctx->dummy_ubuf);
9cfc7e94
JA
2757 kfree(ctx->io_bl);
2758 xa_destroy(&ctx->io_bl_xa);
2b188cc1
JA
2759 kfree(ctx);
2760}
2761
2762static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2763{
2764 struct io_ring_ctx *ctx = file->private_data;
2765 __poll_t mask = 0;
2766
d60aa65b 2767 poll_wait(file, &ctx->cq_wait, wait);
4f7067c3
SB
2768 /*
2769 * synchronizes with barrier from wq_has_sleeper call in
2770 * io_commit_cqring
2771 */
2b188cc1 2772 smp_rmb();
90554200 2773 if (!io_sqring_full(ctx))
2b188cc1 2774 mask |= EPOLLOUT | EPOLLWRNORM;
ed670c3f
HX
2775
2776 /*
2777 * Don't flush cqring overflow list here, just do a simple check.
2778 * Otherwise there could possible be ABBA deadlock:
2779 * CPU0 CPU1
2780 * ---- ----
2781 * lock(&ctx->uring_lock);
2782 * lock(&ep->mtx);
2783 * lock(&ctx->uring_lock);
2784 * lock(&ep->mtx);
2785 *
2786 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
10d8bc35 2787 * pushes them to do the flush.
ed670c3f 2788 */
b4c98d59
DY
2789
2790 if (io_cqring_events(ctx) || io_has_work(ctx))
2b188cc1
JA
2791 mask |= EPOLLIN | EPOLLRDNORM;
2792
2793 return mask;
2794}
2795
0bead8cd 2796static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
071698e1 2797{
4379bf8b 2798 const struct cred *creds;
071698e1 2799
61cf9370 2800 creds = xa_erase(&ctx->personalities, id);
4379bf8b
JA
2801 if (creds) {
2802 put_cred(creds);
0bead8cd 2803 return 0;
1e6fa521 2804 }
0bead8cd
YD
2805
2806 return -EINVAL;
2807}
2808
d56d938b
PB
2809struct io_tctx_exit {
2810 struct callback_head task_work;
2811 struct completion completion;
baf186c4 2812 struct io_ring_ctx *ctx;
d56d938b
PB
2813};
2814
c072481d 2815static __cold void io_tctx_exit_cb(struct callback_head *cb)
d56d938b
PB
2816{
2817 struct io_uring_task *tctx = current->io_uring;
2818 struct io_tctx_exit *work;
2819
2820 work = container_of(cb, struct io_tctx_exit, task_work);
2821 /*
2822 * When @in_idle, we're in cancellation and it's racy to remove the
2823 * node. It'll be removed by the end of cancellation, just ignore it.
998b30c3
HM
2824 * tctx can be NULL if the queueing of this task_work raced with
2825 * work cancelation off the exec path.
d56d938b 2826 */
998b30c3 2827 if (tctx && !atomic_read(&tctx->in_idle))
eef51daa 2828 io_uring_del_tctx_node((unsigned long)work->ctx);
d56d938b
PB
2829 complete(&work->completion);
2830}
2831
c072481d 2832static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
28090c13
PB
2833{
2834 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2835
2836 return req->ctx == data;
2837}
2838
c072481d 2839static __cold void io_ring_exit_work(struct work_struct *work)
85faa7b8 2840{
d56d938b 2841 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
b5bb3a24 2842 unsigned long timeout = jiffies + HZ * 60 * 5;
58d3be2c 2843 unsigned long interval = HZ / 20;
d56d938b
PB
2844 struct io_tctx_exit exit;
2845 struct io_tctx_node *node;
2846 int ret;
85faa7b8 2847
56952e91
JA
2848 /*
2849 * If we're doing polled IO and end up having requests being
2850 * submitted async (out-of-line), then completions can come in while
2851 * we're waiting for refs to drop. We need to reap these manually,
2852 * as nobody else will be looking for them.
2853 */
b2edc0a7 2854 do {
a85381d8
PB
2855 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2856 mutex_lock(&ctx->uring_lock);
2857 io_cqring_overflow_kill(ctx);
2858 mutex_unlock(&ctx->uring_lock);
2859 }
2860
c0e0d6ba
DY
2861 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2862 io_move_task_work_from_local(ctx);
2863
affa87db
PB
2864 while (io_uring_try_cancel_requests(ctx, NULL, true))
2865 cond_resched();
2866
28090c13
PB
2867 if (ctx->sq_data) {
2868 struct io_sq_data *sqd = ctx->sq_data;
2869 struct task_struct *tsk;
2870
2871 io_sq_thread_park(sqd);
2872 tsk = sqd->thread;
2873 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2874 io_wq_cancel_cb(tsk->io_uring->io_wq,
2875 io_cancel_ctx_cb, ctx, true);
2876 io_sq_thread_unpark(sqd);
2877 }
b5bb3a24 2878
37f0e767
PB
2879 io_req_caches_free(ctx);
2880
58d3be2c
PB
2881 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2882 /* there is little hope left, don't run it too often */
2883 interval = HZ * 60;
2884 }
2885 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
d56d938b 2886
7f00651a
PB
2887 init_completion(&exit.completion);
2888 init_task_work(&exit.task_work, io_tctx_exit_cb);
2889 exit.ctx = ctx;
89b5066e
PB
2890 /*
2891 * Some may use context even when all refs and requests have been put,
2892 * and they are free to do so while still holding uring_lock or
5b0a6acc 2893 * completion_lock, see io_req_task_submit(). Apart from other work,
89b5066e
PB
2894 * this lock/unlock section also waits them to finish.
2895 */
d56d938b
PB
2896 mutex_lock(&ctx->uring_lock);
2897 while (!list_empty(&ctx->tctx_list)) {
b5bb3a24
PB
2898 WARN_ON_ONCE(time_after(jiffies, timeout));
2899
d56d938b
PB
2900 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2901 ctx_node);
7f00651a
PB
2902 /* don't spin on a single task if cancellation failed */
2903 list_rotate_left(&ctx->tctx_list);
d56d938b
PB
2904 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2905 if (WARN_ON_ONCE(ret))
2906 continue;
d56d938b
PB
2907
2908 mutex_unlock(&ctx->uring_lock);
2909 wait_for_completion(&exit.completion);
d56d938b
PB
2910 mutex_lock(&ctx->uring_lock);
2911 }
2912 mutex_unlock(&ctx->uring_lock);
79ebeaee
JA
2913 spin_lock(&ctx->completion_lock);
2914 spin_unlock(&ctx->completion_lock);
d56d938b 2915
85faa7b8
JA
2916 io_ring_ctx_free(ctx);
2917}
2918
c072481d 2919static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2b188cc1 2920{
61cf9370
MWO
2921 unsigned long index;
2922 struct creds *creds;
2923
2b188cc1
JA
2924 mutex_lock(&ctx->uring_lock);
2925 percpu_ref_kill(&ctx->refs);
61cf9370
MWO
2926 xa_for_each(&ctx->personalities, index, creds)
2927 io_unregister_personality(ctx, index);
9ca9fb24
PB
2928 if (ctx->rings)
2929 io_poll_remove_all(ctx, NULL, true);
2b188cc1
JA
2930 mutex_unlock(&ctx->uring_lock);
2931
02bac94b
PB
2932 /*
2933 * If we failed setting up the ctx, we might not have any rings
2934 * and therefore did not submit any requests
2935 */
2936 if (ctx->rings)
60053be8 2937 io_kill_timeouts(ctx, NULL, true);
309fc03a 2938
85faa7b8 2939 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
fc666777
JA
2940 /*
2941 * Use system_unbound_wq to avoid spawning tons of event kworkers
2942 * if we're exiting a ton of rings at the same time. It just adds
2943 * noise and overhead, there's no discernable change in runtime
2944 * over using system_wq.
2945 */
2946 queue_work(system_unbound_wq, &ctx->exit_work);
2b188cc1
JA
2947}
2948
2949static int io_uring_release(struct inode *inode, struct file *file)
2950{
2951 struct io_ring_ctx *ctx = file->private_data;
2952
2953 file->private_data = NULL;
2954 io_ring_ctx_wait_and_kill(ctx);
2955 return 0;
2956}
2957
f6edbabb
PB
2958struct io_task_cancel {
2959 struct task_struct *task;
3dd0c97a 2960 bool all;
f6edbabb 2961};
f254ac04 2962
f6edbabb 2963static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
b711d4ea 2964{
9a472ef7 2965 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
f6edbabb 2966 struct io_task_cancel *cancel = data;
9a472ef7 2967
6af3f48b 2968 return io_match_task_safe(req, cancel->task, cancel->all);
b711d4ea
JA
2969}
2970
c072481d
PB
2971static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2972 struct task_struct *task,
2973 bool cancel_all)
b7ddce3c 2974{
e1915f76 2975 struct io_defer_entry *de;
b7ddce3c
PB
2976 LIST_HEAD(list);
2977
79ebeaee 2978 spin_lock(&ctx->completion_lock);
b7ddce3c 2979 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
6af3f48b 2980 if (io_match_task_safe(de->req, task, cancel_all)) {
b7ddce3c
PB
2981 list_cut_position(&list, &ctx->defer_list, &de->list);
2982 break;
2983 }
2984 }
79ebeaee 2985 spin_unlock(&ctx->completion_lock);
e1915f76
PB
2986 if (list_empty(&list))
2987 return false;
b7ddce3c
PB
2988
2989 while (!list_empty(&list)) {
2990 de = list_first_entry(&list, struct io_defer_entry, list);
2991 list_del_init(&de->list);
e276ae34 2992 io_req_task_queue_fail(de->req, -ECANCELED);
b7ddce3c
PB
2993 kfree(de);
2994 }
e1915f76 2995 return true;
b7ddce3c
PB
2996}
2997
c072481d 2998static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
1b00764f
PB
2999{
3000 struct io_tctx_node *node;
3001 enum io_wq_cancel cret;
3002 bool ret = false;
3003
3004 mutex_lock(&ctx->uring_lock);
3005 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3006 struct io_uring_task *tctx = node->task->io_uring;
3007
3008 /*
3009 * io_wq will stay alive while we hold uring_lock, because it's
3010 * killed after ctx nodes, which requires to take the lock.
3011 */
3012 if (!tctx || !tctx->io_wq)
3013 continue;
3014 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3015 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3016 }
3017 mutex_unlock(&ctx->uring_lock);
3018
3019 return ret;
3020}
3021
affa87db 3022static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
c072481d
PB
3023 struct task_struct *task,
3024 bool cancel_all)
9936c7c2 3025{
3dd0c97a 3026 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
1b00764f 3027 struct io_uring_task *tctx = task ? task->io_uring : NULL;
affa87db
PB
3028 enum io_wq_cancel cret;
3029 bool ret = false;
9936c7c2 3030
60053be8
PB
3031 /* failed during ring init, it couldn't have issued any requests */
3032 if (!ctx->rings)
affa87db 3033 return false;
60053be8 3034
affa87db
PB
3035 if (!task) {
3036 ret |= io_uring_try_cancel_iowq(ctx);
3037 } else if (tctx && tctx->io_wq) {
3038 /*
3039 * Cancels requests of all rings, not only @ctx, but
3040 * it's fine as the task is in exit/exec.
3041 */
3042 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3043 &cancel, true);
3044 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3045 }
9936c7c2 3046
affa87db
PB
3047 /* SQPOLL thread does its own polling */
3048 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3049 (ctx->sq_data && ctx->sq_data->thread == current)) {
3050 while (!wq_list_empty(&ctx->iopoll_list)) {
3051 io_iopoll_try_reap_events(ctx);
3052 ret = true;
9936c7c2 3053 }
9936c7c2 3054 }
affa87db 3055
c0e0d6ba
DY
3056 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3057 ret |= io_run_local_work(ctx) > 0;
affa87db
PB
3058 ret |= io_cancel_defer_files(ctx, task, cancel_all);
3059 mutex_lock(&ctx->uring_lock);
3060 ret |= io_poll_remove_all(ctx, task, cancel_all);
3061 mutex_unlock(&ctx->uring_lock);
3062 ret |= io_kill_timeouts(ctx, task, cancel_all);
3063 if (task)
c0e0d6ba 3064 ret |= io_run_task_work() > 0;
affa87db 3065 return ret;
9936c7c2
PB
3066}
3067
3f48cf18 3068static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
521d6a73 3069{
3f48cf18 3070 if (tracked)
9cae36a0 3071 return atomic_read(&tctx->inflight_tracked);
521d6a73
PB
3072 return percpu_counter_sum(&tctx->inflight);
3073}
3074
78cc687b
PB
3075/*
3076 * Find any io_uring ctx that this task has registered or done IO on, and cancel
78a78060 3077 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
78cc687b 3078 */
17437f31 3079__cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
0e9ddb39 3080{
521d6a73 3081 struct io_uring_task *tctx = current->io_uring;
734551df 3082 struct io_ring_ctx *ctx;
0e9ddb39
PB
3083 s64 inflight;
3084 DEFINE_WAIT(wait);
fdaf083c 3085
78cc687b
PB
3086 WARN_ON_ONCE(sqd && sqd->thread != current);
3087
6d042ffb
PO
3088 if (!current->io_uring)
3089 return;
17a91051
PB
3090 if (tctx->io_wq)
3091 io_wq_exit_start(tctx->io_wq);
3092
0e9ddb39
PB
3093 atomic_inc(&tctx->in_idle);
3094 do {
affa87db
PB
3095 bool loop = false;
3096
e9dbe221 3097 io_uring_drop_tctx_refs(current);
0e9ddb39 3098 /* read completions before cancelations */
78cc687b 3099 inflight = tctx_inflight(tctx, !cancel_all);
0e9ddb39
PB
3100 if (!inflight)
3101 break;
fdaf083c 3102
78cc687b
PB
3103 if (!sqd) {
3104 struct io_tctx_node *node;
3105 unsigned long index;
0f212204 3106
78cc687b
PB
3107 xa_for_each(&tctx->xa, index, node) {
3108 /* sqpoll task will cancel all its requests */
3109 if (node->ctx->sq_data)
3110 continue;
affa87db
PB
3111 loop |= io_uring_try_cancel_requests(node->ctx,
3112 current, cancel_all);
78cc687b
PB
3113 }
3114 } else {
3115 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
affa87db
PB
3116 loop |= io_uring_try_cancel_requests(ctx,
3117 current,
3118 cancel_all);
3119 }
3120
3121 if (loop) {
3122 cond_resched();
3123 continue;
78cc687b 3124 }
17a91051 3125
78a78060
JA
3126 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3127 io_run_task_work();
e9dbe221 3128 io_uring_drop_tctx_refs(current);
78a78060 3129
0f212204 3130 /*
a1bb3cd5
PB
3131 * If we've seen completions, retry without waiting. This
3132 * avoids a race where a completion comes in before we did
3133 * prepare_to_wait().
0f212204 3134 */
3dd0c97a 3135 if (inflight == tctx_inflight(tctx, !cancel_all))
a1bb3cd5 3136 schedule();
f57555ed 3137 finish_wait(&tctx->wait, &wait);
d8a6df10 3138 } while (1);
de7f1d9e 3139
8452d4a6 3140 io_uring_clean_tctx(tctx);
3dd0c97a 3141 if (cancel_all) {
3cc7fdb9
PB
3142 /*
3143 * We shouldn't run task_works after cancel, so just leave
3144 * ->in_idle set for normal exit.
3145 */
3146 atomic_dec(&tctx->in_idle);
3f48cf18
PB
3147 /* for exec all current's requests should be gone, kill tctx */
3148 __io_uring_free(current);
3149 }
44e728b8
PB
3150}
3151
f552a27a 3152void __io_uring_cancel(bool cancel_all)
78cc687b 3153{
f552a27a 3154 io_uring_cancel_generic(cancel_all, NULL);
78cc687b
PB
3155}
3156
6c5c240e
RP
3157static void *io_uring_validate_mmap_request(struct file *file,
3158 loff_t pgoff, size_t sz)
2b188cc1 3159{
2b188cc1 3160 struct io_ring_ctx *ctx = file->private_data;
6c5c240e 3161 loff_t offset = pgoff << PAGE_SHIFT;
2b188cc1
JA
3162 struct page *page;
3163 void *ptr;
3164
3165 switch (offset) {
3166 case IORING_OFF_SQ_RING:
75b28aff
HV
3167 case IORING_OFF_CQ_RING:
3168 ptr = ctx->rings;
2b188cc1
JA
3169 break;
3170 case IORING_OFF_SQES:
3171 ptr = ctx->sq_sqes;
3172 break;
2b188cc1 3173 default:
6c5c240e 3174 return ERR_PTR(-EINVAL);
2b188cc1
JA
3175 }
3176
3177 page = virt_to_head_page(ptr);
a50b854e 3178 if (sz > page_size(page))
6c5c240e
RP
3179 return ERR_PTR(-EINVAL);
3180
3181 return ptr;
3182}
3183
3184#ifdef CONFIG_MMU
3185
c072481d 3186static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
6c5c240e
RP
3187{
3188 size_t sz = vma->vm_end - vma->vm_start;
3189 unsigned long pfn;
3190 void *ptr;
3191
3192 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3193 if (IS_ERR(ptr))
3194 return PTR_ERR(ptr);
2b188cc1
JA
3195
3196 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3197 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3198}
3199
6c5c240e
RP
3200#else /* !CONFIG_MMU */
3201
3202static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3203{
3204 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
3205}
3206
3207static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3208{
3209 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3210}
3211
3212static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3213 unsigned long addr, unsigned long len,
3214 unsigned long pgoff, unsigned long flags)
3215{
3216 void *ptr;
3217
3218 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3219 if (IS_ERR(ptr))
3220 return PTR_ERR(ptr);
3221
3222 return (unsigned long) ptr;
3223}
3224
3225#endif /* !CONFIG_MMU */
3226
f81440d3
PB
3227static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3228{
3229 if (flags & IORING_ENTER_EXT_ARG) {
3230 struct io_uring_getevents_arg arg;
3231
3232 if (argsz != sizeof(arg))
3233 return -EINVAL;
3234 if (copy_from_user(&arg, argp, sizeof(arg)))
3235 return -EFAULT;
3236 }
3237 return 0;
3238}
3239
c73ebb68
HX
3240static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3241 struct __kernel_timespec __user **ts,
3242 const sigset_t __user **sig)
3243{
3244 struct io_uring_getevents_arg arg;
3245
3246 /*
3247 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3248 * is just a pointer to the sigset_t.
3249 */
3250 if (!(flags & IORING_ENTER_EXT_ARG)) {
3251 *sig = (const sigset_t __user *) argp;
3252 *ts = NULL;
3253 return 0;
3254 }
3255
3256 /*
3257 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3258 * timespec and sigset_t pointers if good.
3259 */
3260 if (*argsz != sizeof(arg))
3261 return -EINVAL;
3262 if (copy_from_user(&arg, argp, sizeof(arg)))
3263 return -EFAULT;
d2347b96
DY
3264 if (arg.pad)
3265 return -EINVAL;
c73ebb68
HX
3266 *sig = u64_to_user_ptr(arg.sigmask);
3267 *argsz = arg.sigmask_sz;
3268 *ts = u64_to_user_ptr(arg.ts);
3269 return 0;
3270}
3271
2b188cc1 3272SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
c73ebb68
HX
3273 u32, min_complete, u32, flags, const void __user *, argp,
3274 size_t, argsz)
2b188cc1
JA
3275{
3276 struct io_ring_ctx *ctx;
2b188cc1 3277 struct fd f;
33f993da 3278 long ret;
2b188cc1 3279
33f993da 3280 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
e7a6c00d
JA
3281 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3282 IORING_ENTER_REGISTERED_RING)))
2b188cc1
JA
3283 return -EINVAL;
3284
e7a6c00d
JA
3285 /*
3286 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3287 * need only dereference our task private array to find it.
3288 */
3289 if (flags & IORING_ENTER_REGISTERED_RING) {
3290 struct io_uring_task *tctx = current->io_uring;
3291
3273c440 3292 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
e7a6c00d
JA
3293 return -EINVAL;
3294 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3295 f.file = tctx->registered_rings[fd];
4329490a 3296 f.flags = 0;
3273c440
PB
3297 if (unlikely(!f.file))
3298 return -EBADF;
e7a6c00d
JA
3299 } else {
3300 f = fdget(fd);
3273c440
PB
3301 if (unlikely(!f.file))
3302 return -EBADF;
3303 ret = -EOPNOTSUPP;
3304 if (unlikely(!io_is_uring_fops(f.file)))
fbb8bb02 3305 goto out;
e7a6c00d 3306 }
2b188cc1 3307
2b188cc1 3308 ctx = f.file->private_data;
7e84e1c7 3309 ret = -EBADFD;
33f993da 3310 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
7e84e1c7
SG
3311 goto out;
3312
6c271ce2
JA
3313 /*
3314 * For SQ polling, the thread will do all submissions and completions.
3315 * Just return the requested submit count, and wake the thread if
3316 * we were asked to.
3317 */
b2a9eada 3318 ret = 0;
6c271ce2 3319 if (ctx->flags & IORING_SETUP_SQPOLL) {
90f67366 3320 io_cqring_overflow_flush(ctx);
89448c47 3321
21f96522
JA
3322 if (unlikely(ctx->sq_data->thread == NULL)) {
3323 ret = -EOWNERDEAD;
04147488 3324 goto out;
21f96522 3325 }
6c271ce2 3326 if (flags & IORING_ENTER_SQ_WAKEUP)
534ca6d6 3327 wake_up(&ctx->sq_data->wait);
d9d05217
PB
3328 if (flags & IORING_ENTER_SQ_WAIT) {
3329 ret = io_sqpoll_wait_sq(ctx);
3330 if (ret)
3331 goto out;
3332 }
3e813c90 3333 ret = to_submit;
b2a9eada 3334 } else if (to_submit) {
eef51daa 3335 ret = io_uring_add_tctx_node(ctx);
0f212204
JA
3336 if (unlikely(ret))
3337 goto out;
7c504e65 3338
2b188cc1 3339 mutex_lock(&ctx->uring_lock);
3e813c90
DY
3340 ret = io_submit_sqes(ctx, to_submit);
3341 if (ret != to_submit) {
d487b43c 3342 mutex_unlock(&ctx->uring_lock);
7c504e65 3343 goto out;
d487b43c 3344 }
44f87745
PB
3345 if (flags & IORING_ENTER_GETEVENTS) {
3346 if (ctx->syscall_iopoll)
3347 goto iopoll_locked;
3348 /*
3349 * Ignore errors, we'll soon call io_cqring_wait() and
3350 * it should handle ownership problems if any.
3351 */
3352 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3353 (void)io_run_local_work_locked(ctx);
3354 }
d487b43c 3355 mutex_unlock(&ctx->uring_lock);
2b188cc1 3356 }
c0e0d6ba 3357
2b188cc1 3358 if (flags & IORING_ENTER_GETEVENTS) {
3e813c90 3359 int ret2;
c0e0d6ba 3360
773697b6 3361 if (ctx->syscall_iopoll) {
d487b43c
PB
3362 /*
3363 * We disallow the app entering submit/complete with
3364 * polling, but we still need to lock the ring to
3365 * prevent racing with polled issue that got punted to
3366 * a workqueue.
3367 */
3368 mutex_lock(&ctx->uring_lock);
3369iopoll_locked:
3e813c90
DY
3370 ret2 = io_validate_ext_arg(flags, argp, argsz);
3371 if (likely(!ret2)) {
3372 min_complete = min(min_complete,
3373 ctx->cq_entries);
3374 ret2 = io_iopoll_check(ctx, min_complete);
d487b43c
PB
3375 }
3376 mutex_unlock(&ctx->uring_lock);
def596e9 3377 } else {
f81440d3
PB
3378 const sigset_t __user *sig;
3379 struct __kernel_timespec __user *ts;
3380
3e813c90
DY
3381 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3382 if (likely(!ret2)) {
3383 min_complete = min(min_complete,
3384 ctx->cq_entries);
3385 ret2 = io_cqring_wait(ctx, min_complete, sig,
3386 argsz, ts);
3387 }
def596e9 3388 }
c73ebb68 3389
155bc950 3390 if (!ret) {
3e813c90 3391 ret = ret2;
2b188cc1 3392
155bc950
DY
3393 /*
3394 * EBADR indicates that one or more CQE were dropped.
3395 * Once the user has been informed we can clear the bit
3396 * as they are obviously ok with those drops.
3397 */
3398 if (unlikely(ret2 == -EBADR))
3399 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3400 &ctx->check_cq);
def596e9 3401 }
2b188cc1 3402 }
7c504e65 3403out:
4329490a 3404 fdput(f);
3e813c90 3405 return ret;
2b188cc1
JA
3406}
3407
3408static const struct file_operations io_uring_fops = {
3409 .release = io_uring_release,
3410 .mmap = io_uring_mmap,
6c5c240e
RP
3411#ifndef CONFIG_MMU
3412 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3413 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3414#endif
2b188cc1 3415 .poll = io_uring_poll,
bebdb65e 3416#ifdef CONFIG_PROC_FS
87ce955b 3417 .show_fdinfo = io_uring_show_fdinfo,
bebdb65e 3418#endif
2b188cc1
JA
3419};
3420
92ac8bea
JA
3421bool io_is_uring_fops(struct file *file)
3422{
3423 return file->f_op == &io_uring_fops;
3424}
3425
c072481d
PB
3426static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3427 struct io_uring_params *p)
2b188cc1 3428{
75b28aff
HV
3429 struct io_rings *rings;
3430 size_t size, sq_array_offset;
2b188cc1 3431
bd740481
JA
3432 /* make sure these are sane, as we already accounted them */
3433 ctx->sq_entries = p->sq_entries;
3434 ctx->cq_entries = p->cq_entries;
3435
baf9cb64 3436 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
75b28aff
HV
3437 if (size == SIZE_MAX)
3438 return -EOVERFLOW;
3439
3440 rings = io_mem_alloc(size);
3441 if (!rings)
2b188cc1
JA
3442 return -ENOMEM;
3443
75b28aff
HV
3444 ctx->rings = rings;
3445 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3446 rings->sq_ring_mask = p->sq_entries - 1;
3447 rings->cq_ring_mask = p->cq_entries - 1;
3448 rings->sq_ring_entries = p->sq_entries;
3449 rings->cq_ring_entries = p->cq_entries;
2b188cc1 3450
ebdeb7c0
JA
3451 if (p->flags & IORING_SETUP_SQE128)
3452 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3453 else
3454 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
eb065d30
JA
3455 if (size == SIZE_MAX) {
3456 io_mem_free(ctx->rings);
3457 ctx->rings = NULL;
2b188cc1 3458 return -EOVERFLOW;
eb065d30 3459 }
2b188cc1
JA
3460
3461 ctx->sq_sqes = io_mem_alloc(size);
eb065d30
JA
3462 if (!ctx->sq_sqes) {
3463 io_mem_free(ctx->rings);
3464 ctx->rings = NULL;
2b188cc1 3465 return -ENOMEM;
eb065d30 3466 }
2b188cc1 3467
2b188cc1
JA
3468 return 0;
3469}
3470
9faadcc8
PB
3471static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3472{
3473 int ret, fd;
3474
3475 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3476 if (fd < 0)
3477 return fd;
3478
97c96e9f 3479 ret = __io_uring_add_tctx_node(ctx);
9faadcc8
PB
3480 if (ret) {
3481 put_unused_fd(fd);
3482 return ret;
3483 }
3484 fd_install(fd, file);
3485 return fd;
3486}
3487
2b188cc1
JA
3488/*
3489 * Allocate an anonymous fd, this is what constitutes the application
3490 * visible backing of an io_uring instance. The application mmaps this
3491 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3492 * we have to tie this fd to a socket for file garbage collection purposes.
3493 */
9faadcc8 3494static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
2b188cc1
JA
3495{
3496 struct file *file;
9faadcc8 3497#if defined(CONFIG_UNIX)
2b188cc1
JA
3498 int ret;
3499
2b188cc1
JA
3500 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3501 &ctx->ring_sock);
3502 if (ret)
9faadcc8 3503 return ERR_PTR(ret);
2b188cc1
JA
3504#endif
3505
91a9ab7c
PM
3506 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3507 O_RDWR | O_CLOEXEC, NULL);
2b188cc1 3508#if defined(CONFIG_UNIX)
9faadcc8
PB
3509 if (IS_ERR(file)) {
3510 sock_release(ctx->ring_sock);
3511 ctx->ring_sock = NULL;
3512 } else {
3513 ctx->ring_sock->file = file;
0f212204 3514 }
2b188cc1 3515#endif
9faadcc8 3516 return file;
2b188cc1
JA
3517}
3518
c072481d
PB
3519static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3520 struct io_uring_params __user *params)
2b188cc1 3521{
2b188cc1 3522 struct io_ring_ctx *ctx;
9faadcc8 3523 struct file *file;
2b188cc1
JA
3524 int ret;
3525
8110c1a6 3526 if (!entries)
2b188cc1 3527 return -EINVAL;
8110c1a6
JA
3528 if (entries > IORING_MAX_ENTRIES) {
3529 if (!(p->flags & IORING_SETUP_CLAMP))
3530 return -EINVAL;
3531 entries = IORING_MAX_ENTRIES;
3532 }
2b188cc1
JA
3533
3534 /*
3535 * Use twice as many entries for the CQ ring. It's possible for the
3536 * application to drive a higher depth than the size of the SQ ring,
3537 * since the sqes are only used at submission time. This allows for
33a107f0
JA
3538 * some flexibility in overcommitting a bit. If the application has
3539 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3540 * of CQ ring entries manually.
2b188cc1
JA
3541 */
3542 p->sq_entries = roundup_pow_of_two(entries);
33a107f0
JA
3543 if (p->flags & IORING_SETUP_CQSIZE) {
3544 /*
3545 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3546 * to a power-of-two, if it isn't already. We do NOT impose
3547 * any cq vs sq ring sizing.
3548 */
eb2667b3 3549 if (!p->cq_entries)
33a107f0 3550 return -EINVAL;
8110c1a6
JA
3551 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3552 if (!(p->flags & IORING_SETUP_CLAMP))
3553 return -EINVAL;
3554 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3555 }
eb2667b3
JQ
3556 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3557 if (p->cq_entries < p->sq_entries)
3558 return -EINVAL;
33a107f0
JA
3559 } else {
3560 p->cq_entries = 2 * p->sq_entries;
3561 }
2b188cc1 3562
2b188cc1 3563 ctx = io_ring_ctx_alloc(p);
62e398be 3564 if (!ctx)
2b188cc1 3565 return -ENOMEM;
773697b6 3566
e6aeb272
PB
3567 if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3568 !(ctx->flags & IORING_SETUP_IOPOLL) &&
3569 !(ctx->flags & IORING_SETUP_SQPOLL))
3570 ctx->task_complete = true;
3571
773697b6
PB
3572 /*
3573 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3574 * space applications don't need to do io completion events
3575 * polling again, they can rely on io_sq_thread to do polling
3576 * work, which can reduce cpu usage and uring_lock contention.
3577 */
3578 if (ctx->flags & IORING_SETUP_IOPOLL &&
3579 !(ctx->flags & IORING_SETUP_SQPOLL))
3580 ctx->syscall_iopoll = 1;
3581
2b188cc1 3582 ctx->compat = in_compat_syscall();
62e398be
JA
3583 if (!capable(CAP_IPC_LOCK))
3584 ctx->user = get_uid(current_user());
2aede0e4 3585
9f010507 3586 /*
e1169f06
JA
3587 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3588 * COOP_TASKRUN is set, then IPIs are never needed by the app.
9f010507 3589 */
e1169f06
JA
3590 ret = -EINVAL;
3591 if (ctx->flags & IORING_SETUP_SQPOLL) {
3592 /* IPI related flags don't make sense with SQPOLL */
ef060ea9 3593 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
c0e0d6ba
DY
3594 IORING_SETUP_TASKRUN_FLAG |
3595 IORING_SETUP_DEFER_TASKRUN))
e1169f06 3596 goto err;
9f010507 3597 ctx->notify_method = TWA_SIGNAL_NO_IPI;
e1169f06
JA
3598 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3599 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3600 } else {
c0e0d6ba
DY
3601 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3602 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
ef060ea9 3603 goto err;
9f010507 3604 ctx->notify_method = TWA_SIGNAL;
e1169f06 3605 }
9f010507 3606
c0e0d6ba
DY
3607 /*
3608 * For DEFER_TASKRUN we require the completion task to be the same as the
3609 * submission task. This implies that there is only one submitter, so enforce
3610 * that.
3611 */
3612 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3613 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3614 goto err;
3615 }
3616
2aede0e4
JA
3617 /*
3618 * This is just grabbed for accounting purposes. When a process exits,
3619 * the mm is exited and dropped before the files, hence we need to hang
3620 * on to this mm purely for the purposes of being able to unaccount
3621 * memory (locked/pinned vm). It's not used for anything else.
3622 */
6b7898eb 3623 mmgrab(current->mm);
2aede0e4 3624 ctx->mm_account = current->mm;
6b7898eb 3625
2b188cc1
JA
3626 ret = io_allocate_scq_urings(ctx, p);
3627 if (ret)
3628 goto err;
3629
7e84e1c7 3630 ret = io_sq_offload_create(ctx, p);
2b188cc1
JA
3631 if (ret)
3632 goto err;
eae071c9 3633 /* always set a rsrc node */
47b228ce
PB
3634 ret = io_rsrc_node_switch_start(ctx);
3635 if (ret)
3636 goto err;
eae071c9 3637 io_rsrc_node_switch(ctx, NULL);
2b188cc1 3638
2b188cc1 3639 memset(&p->sq_off, 0, sizeof(p->sq_off));
75b28aff
HV
3640 p->sq_off.head = offsetof(struct io_rings, sq.head);
3641 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3642 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3643 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3644 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3645 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3646 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
2b188cc1
JA
3647
3648 memset(&p->cq_off, 0, sizeof(p->cq_off));
75b28aff
HV
3649 p->cq_off.head = offsetof(struct io_rings, cq.head);
3650 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3651 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3652 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3653 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3654 p->cq_off.cqes = offsetof(struct io_rings, cqes);
0d9b5b3a 3655 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
ac90f249 3656
7f13657d
XW
3657 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3658 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
5769a351 3659 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
c73ebb68 3660 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
9690557e 3661 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
c4212f3e
JA
3662 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3663 IORING_FEAT_LINKED_FILE;
7f13657d
XW
3664
3665 if (copy_to_user(params, p, sizeof(*p))) {
3666 ret = -EFAULT;
3667 goto err;
3668 }
d1719f70 3669
7cae596b
DY
3670 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3671 && !(ctx->flags & IORING_SETUP_R_DISABLED))
8579538c 3672 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
7cae596b 3673
9faadcc8
PB
3674 file = io_uring_get_file(ctx);
3675 if (IS_ERR(file)) {
3676 ret = PTR_ERR(file);
3677 goto err;
3678 }
3679
044c1ab3
JA
3680 /*
3681 * Install ring fd as the very last thing, so we don't risk someone
3682 * having closed it before we finish setup
3683 */
9faadcc8
PB
3684 ret = io_uring_install_fd(ctx, file);
3685 if (ret < 0) {
3686 /* fput will clean it up */
3687 fput(file);
3688 return ret;
3689 }
044c1ab3 3690
c826bd7a 3691 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
2b188cc1
JA
3692 return ret;
3693err:
3694 io_ring_ctx_wait_and_kill(ctx);
3695 return ret;
3696}
3697
3698/*
3699 * Sets up an aio uring context, and returns the fd. Applications asks for a
3700 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3701 * params structure passed in.
3702 */
3703static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3704{
3705 struct io_uring_params p;
2b188cc1
JA
3706 int i;
3707
3708 if (copy_from_user(&p, params, sizeof(p)))
3709 return -EFAULT;
3710 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3711 if (p.resv[i])
3712 return -EINVAL;
3713 }
3714
6c271ce2 3715 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
8110c1a6 3716 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7e84e1c7 3717 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
e1169f06 3718 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
ebdeb7c0 3719 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
97bbdc06 3720 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
c0e0d6ba 3721 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
2b188cc1
JA
3722 return -EINVAL;
3723
ef060ea9 3724 return io_uring_create(entries, &p, params);
2b188cc1
JA
3725}
3726
3727SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3728 struct io_uring_params __user *, params)
3729{
3730 return io_uring_setup(entries, params);
3731}
3732
c072481d
PB
3733static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3734 unsigned nr_args)
66f4af93
JA
3735{
3736 struct io_uring_probe *p;
3737 size_t size;
3738 int i, ret;
3739
3740 size = struct_size(p, ops, nr_args);
3741 if (size == SIZE_MAX)
3742 return -EOVERFLOW;
3743 p = kzalloc(size, GFP_KERNEL);
3744 if (!p)
3745 return -ENOMEM;
3746
3747 ret = -EFAULT;
3748 if (copy_from_user(p, arg, size))
3749 goto out;
3750 ret = -EINVAL;
3751 if (memchr_inv(p, 0, size))
3752 goto out;
3753
3754 p->last_op = IORING_OP_LAST - 1;
3755 if (nr_args > IORING_OP_LAST)
3756 nr_args = IORING_OP_LAST;
3757
3758 for (i = 0; i < nr_args; i++) {
3759 p->ops[i].op = i;
3760 if (!io_op_defs[i].not_supported)
3761 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3762 }
3763 p->ops_len = i;
3764
3765 ret = 0;
3766 if (copy_to_user(arg, p, size))
3767 ret = -EFAULT;
3768out:
3769 kfree(p);
3770 return ret;
3771}
3772
071698e1
JA
3773static int io_register_personality(struct io_ring_ctx *ctx)
3774{
4379bf8b 3775 const struct cred *creds;
61cf9370 3776 u32 id;
1e6fa521 3777 int ret;
071698e1 3778
4379bf8b 3779 creds = get_current_cred();
1e6fa521 3780
61cf9370
MWO
3781 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3782 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
a30f895a
JA
3783 if (ret < 0) {
3784 put_cred(creds);
3785 return ret;
3786 }
3787 return id;
071698e1
JA
3788}
3789
c072481d
PB
3790static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3791 void __user *arg, unsigned int nr_args)
21b55dbc
SG
3792{
3793 struct io_uring_restriction *res;
3794 size_t size;
3795 int i, ret;
3796
7e84e1c7
SG
3797 /* Restrictions allowed only if rings started disabled */
3798 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3799 return -EBADFD;
3800
21b55dbc 3801 /* We allow only a single restrictions registration */
7e84e1c7 3802 if (ctx->restrictions.registered)
21b55dbc
SG
3803 return -EBUSY;
3804
3805 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3806 return -EINVAL;
3807
3808 size = array_size(nr_args, sizeof(*res));
3809 if (size == SIZE_MAX)
3810 return -EOVERFLOW;
3811
3812 res = memdup_user(arg, size);
3813 if (IS_ERR(res))
3814 return PTR_ERR(res);
3815
3816 ret = 0;
3817
3818 for (i = 0; i < nr_args; i++) {
3819 switch (res[i].opcode) {
3820 case IORING_RESTRICTION_REGISTER_OP:
3821 if (res[i].register_op >= IORING_REGISTER_LAST) {
3822 ret = -EINVAL;
3823 goto out;
3824 }
3825
3826 __set_bit(res[i].register_op,
3827 ctx->restrictions.register_op);
3828 break;
3829 case IORING_RESTRICTION_SQE_OP:
3830 if (res[i].sqe_op >= IORING_OP_LAST) {
3831 ret = -EINVAL;
3832 goto out;
3833 }
3834
3835 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3836 break;
3837 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3838 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3839 break;
3840 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3841 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3842 break;
3843 default:
3844 ret = -EINVAL;
3845 goto out;
3846 }
3847 }
3848
3849out:
3850 /* Reset all restrictions if an error happened */
3851 if (ret != 0)
3852 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3853 else
7e84e1c7 3854 ctx->restrictions.registered = true;
21b55dbc
SG
3855
3856 kfree(res);
3857 return ret;
3858}
3859
7e84e1c7
SG
3860static int io_register_enable_rings(struct io_ring_ctx *ctx)
3861{
3862 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3863 return -EBADFD;
3864
7cae596b 3865 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task)
8579538c 3866 WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
7cae596b 3867
7e84e1c7
SG
3868 if (ctx->restrictions.registered)
3869 ctx->restricted = 1;
3870
0298ef96
PB
3871 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3872 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3873 wake_up(&ctx->sq_data->wait);
7e84e1c7
SG
3874 return 0;
3875}
3876
c072481d
PB
3877static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3878 void __user *arg, unsigned len)
fe76421d
JA
3879{
3880 struct io_uring_task *tctx = current->io_uring;
3881 cpumask_var_t new_mask;
3882 int ret;
3883
3884 if (!tctx || !tctx->io_wq)
3885 return -EINVAL;
3886
3887 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3888 return -ENOMEM;
3889
3890 cpumask_clear(new_mask);
3891 if (len > cpumask_size())
3892 len = cpumask_size();
3893
0f5e4b83
ES
3894 if (in_compat_syscall()) {
3895 ret = compat_get_bitmap(cpumask_bits(new_mask),
3896 (const compat_ulong_t __user *)arg,
3897 len * 8 /* CHAR_BIT */);
3898 } else {
3899 ret = copy_from_user(new_mask, arg, len);
3900 }
3901
3902 if (ret) {
fe76421d
JA
3903 free_cpumask_var(new_mask);
3904 return -EFAULT;
3905 }
3906
3907 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3908 free_cpumask_var(new_mask);
3909 return ret;
3910}
3911
c072481d 3912static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
fe76421d
JA
3913{
3914 struct io_uring_task *tctx = current->io_uring;
3915
3916 if (!tctx || !tctx->io_wq)
3917 return -EINVAL;
3918
3919 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3920}
3921
c072481d
PB
3922static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3923 void __user *arg)
b22fa62a 3924 __must_hold(&ctx->uring_lock)
2e480058 3925{
b22fa62a 3926 struct io_tctx_node *node;
fa84693b
JA
3927 struct io_uring_task *tctx = NULL;
3928 struct io_sq_data *sqd = NULL;
2e480058
JA
3929 __u32 new_count[2];
3930 int i, ret;
3931
2e480058
JA
3932 if (copy_from_user(new_count, arg, sizeof(new_count)))
3933 return -EFAULT;
3934 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3935 if (new_count[i] > INT_MAX)
3936 return -EINVAL;
3937
fa84693b
JA
3938 if (ctx->flags & IORING_SETUP_SQPOLL) {
3939 sqd = ctx->sq_data;
3940 if (sqd) {
009ad9f0
JA
3941 /*
3942 * Observe the correct sqd->lock -> ctx->uring_lock
3943 * ordering. Fine to drop uring_lock here, we hold
3944 * a ref to the ctx.
3945 */
41d3a6bd 3946 refcount_inc(&sqd->refs);
009ad9f0 3947 mutex_unlock(&ctx->uring_lock);
fa84693b 3948 mutex_lock(&sqd->lock);
009ad9f0 3949 mutex_lock(&ctx->uring_lock);
41d3a6bd
JA
3950 if (sqd->thread)
3951 tctx = sqd->thread->io_uring;
fa84693b
JA
3952 }
3953 } else {
3954 tctx = current->io_uring;
3955 }
3956
e139a1ec 3957 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
fa84693b 3958
bad119b9
PB
3959 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3960 if (new_count[i])
3961 ctx->iowq_limits[i] = new_count[i];
e139a1ec
PB
3962 ctx->iowq_limits_set = true;
3963
e139a1ec
PB
3964 if (tctx && tctx->io_wq) {
3965 ret = io_wq_max_workers(tctx->io_wq, new_count);
3966 if (ret)
3967 goto err;
3968 } else {
3969 memset(new_count, 0, sizeof(new_count));
3970 }
fa84693b 3971
41d3a6bd 3972 if (sqd) {
fa84693b 3973 mutex_unlock(&sqd->lock);
41d3a6bd
JA
3974 io_put_sq_data(sqd);
3975 }
2e480058
JA
3976
3977 if (copy_to_user(arg, new_count, sizeof(new_count)))
3978 return -EFAULT;
3979
b22fa62a
PB
3980 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3981 if (sqd)
3982 return 0;
3983
3984 /* now propagate the restriction to all registered users */
3985 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3986 struct io_uring_task *tctx = node->task->io_uring;
3987
3988 if (WARN_ON_ONCE(!tctx->io_wq))
3989 continue;
3990
3991 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3992 new_count[i] = ctx->iowq_limits[i];
3993 /* ignore errors, it always returns zero anyway */
3994 (void)io_wq_max_workers(tctx->io_wq, new_count);
3995 }
2e480058 3996 return 0;
fa84693b 3997err:
41d3a6bd 3998 if (sqd) {
fa84693b 3999 mutex_unlock(&sqd->lock);
41d3a6bd
JA
4000 io_put_sq_data(sqd);
4001 }
fa84693b 4002 return ret;
2e480058
JA
4003}
4004
edafccee
JA
4005static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4006 void __user *arg, unsigned nr_args)
b19062a5
JA
4007 __releases(ctx->uring_lock)
4008 __acquires(ctx->uring_lock)
edafccee
JA
4009{
4010 int ret;
4011
35fa71a0 4012 /*
fbb8bb02
PB
4013 * We don't quiesce the refs for register anymore and so it can't be
4014 * dying as we're holding a file ref here.
35fa71a0 4015 */
fbb8bb02 4016 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
35fa71a0
JA
4017 return -ENXIO;
4018
d7cce96c
PB
4019 if (ctx->submitter_task && ctx->submitter_task != current)
4020 return -EEXIST;
4021
75c4021a 4022 if (ctx->restricted) {
75c4021a
PB
4023 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4024 if (!test_bit(opcode, ctx->restrictions.register_op))
4025 return -EACCES;
4026 }
4027
edafccee
JA
4028 switch (opcode) {
4029 case IORING_REGISTER_BUFFERS:
0184f08e
PB
4030 ret = -EFAULT;
4031 if (!arg)
4032 break;
634d00df 4033 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
edafccee
JA
4034 break;
4035 case IORING_UNREGISTER_BUFFERS:
4036 ret = -EINVAL;
4037 if (arg || nr_args)
4038 break;
0a96bbe4 4039 ret = io_sqe_buffers_unregister(ctx);
edafccee 4040 break;
6b06314c 4041 case IORING_REGISTER_FILES:
a8da73a3
JA
4042 ret = -EFAULT;
4043 if (!arg)
4044 break;
792e3582 4045 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
6b06314c
JA
4046 break;
4047 case IORING_UNREGISTER_FILES:
4048 ret = -EINVAL;
4049 if (arg || nr_args)
4050 break;
4051 ret = io_sqe_files_unregister(ctx);
4052 break;
c3a31e60 4053 case IORING_REGISTER_FILES_UPDATE:
c3bdad02 4054 ret = io_register_files_update(ctx, arg, nr_args);
c3a31e60 4055 break;
9b402849
JA
4056 case IORING_REGISTER_EVENTFD:
4057 ret = -EINVAL;
4058 if (nr_args != 1)
4059 break;
c75312dd
UA
4060 ret = io_eventfd_register(ctx, arg, 0);
4061 break;
4062 case IORING_REGISTER_EVENTFD_ASYNC:
4063 ret = -EINVAL;
4064 if (nr_args != 1)
f2842ab5 4065 break;
c75312dd 4066 ret = io_eventfd_register(ctx, arg, 1);
9b402849
JA
4067 break;
4068 case IORING_UNREGISTER_EVENTFD:
4069 ret = -EINVAL;
4070 if (arg || nr_args)
4071 break;
4072 ret = io_eventfd_unregister(ctx);
4073 break;
66f4af93
JA
4074 case IORING_REGISTER_PROBE:
4075 ret = -EINVAL;
4076 if (!arg || nr_args > 256)
4077 break;
4078 ret = io_probe(ctx, arg, nr_args);
4079 break;
071698e1
JA
4080 case IORING_REGISTER_PERSONALITY:
4081 ret = -EINVAL;
4082 if (arg || nr_args)
4083 break;
4084 ret = io_register_personality(ctx);
4085 break;
4086 case IORING_UNREGISTER_PERSONALITY:
4087 ret = -EINVAL;
4088 if (arg)
4089 break;
4090 ret = io_unregister_personality(ctx, nr_args);
4091 break;
7e84e1c7
SG
4092 case IORING_REGISTER_ENABLE_RINGS:
4093 ret = -EINVAL;
4094 if (arg || nr_args)
4095 break;
4096 ret = io_register_enable_rings(ctx);
4097 break;
21b55dbc
SG
4098 case IORING_REGISTER_RESTRICTIONS:
4099 ret = io_register_restrictions(ctx, arg, nr_args);
4100 break;
992da01a
PB
4101 case IORING_REGISTER_FILES2:
4102 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4103 break;
4104 case IORING_REGISTER_FILES_UPDATE2:
4105 ret = io_register_rsrc_update(ctx, arg, nr_args,
4106 IORING_RSRC_FILE);
4107 break;
4108 case IORING_REGISTER_BUFFERS2:
4109 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
792e3582 4110 break;
992da01a
PB
4111 case IORING_REGISTER_BUFFERS_UPDATE:
4112 ret = io_register_rsrc_update(ctx, arg, nr_args,
4113 IORING_RSRC_BUFFER);
c3bdad02 4114 break;
fe76421d
JA
4115 case IORING_REGISTER_IOWQ_AFF:
4116 ret = -EINVAL;
4117 if (!arg || !nr_args)
4118 break;
4119 ret = io_register_iowq_aff(ctx, arg, nr_args);
4120 break;
4121 case IORING_UNREGISTER_IOWQ_AFF:
4122 ret = -EINVAL;
4123 if (arg || nr_args)
4124 break;
4125 ret = io_unregister_iowq_aff(ctx);
4126 break;
2e480058
JA
4127 case IORING_REGISTER_IOWQ_MAX_WORKERS:
4128 ret = -EINVAL;
4129 if (!arg || nr_args != 2)
4130 break;
4131 ret = io_register_iowq_max_workers(ctx, arg);
4132 break;
e7a6c00d
JA
4133 case IORING_REGISTER_RING_FDS:
4134 ret = io_ringfd_register(ctx, arg, nr_args);
4135 break;
4136 case IORING_UNREGISTER_RING_FDS:
4137 ret = io_ringfd_unregister(ctx, arg, nr_args);
4138 break;
c7fb1942
JA
4139 case IORING_REGISTER_PBUF_RING:
4140 ret = -EINVAL;
4141 if (!arg || nr_args != 1)
4142 break;
4143 ret = io_register_pbuf_ring(ctx, arg);
4144 break;
4145 case IORING_UNREGISTER_PBUF_RING:
4146 ret = -EINVAL;
4147 if (!arg || nr_args != 1)
4148 break;
4149 ret = io_unregister_pbuf_ring(ctx, arg);
4150 break;
78a861b9
JA
4151 case IORING_REGISTER_SYNC_CANCEL:
4152 ret = -EINVAL;
4153 if (!arg || nr_args != 1)
4154 break;
4155 ret = io_sync_cancel(ctx, arg);
4156 break;
6e73dffb
PB
4157 case IORING_REGISTER_FILE_ALLOC_RANGE:
4158 ret = -EINVAL;
4159 if (!arg || nr_args)
4160 break;
4161 ret = io_register_file_alloc_range(ctx, arg);
4162 break;
edafccee
JA
4163 default:
4164 ret = -EINVAL;
4165 break;
4166 }
4167
edafccee
JA
4168 return ret;
4169}
4170
4171SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4172 void __user *, arg, unsigned int, nr_args)
4173{
4174 struct io_ring_ctx *ctx;
4175 long ret = -EBADF;
4176 struct fd f;
4177
34319084
JA
4178 if (opcode >= IORING_REGISTER_LAST)
4179 return -EINVAL;
4180
edafccee
JA
4181 f = fdget(fd);
4182 if (!f.file)
4183 return -EBADF;
4184
4185 ret = -EOPNOTSUPP;
e5550a14 4186 if (!io_is_uring_fops(f.file))
edafccee
JA
4187 goto out_fput;
4188
4189 ctx = f.file->private_data;
4190
4191 mutex_lock(&ctx->uring_lock);
4192 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4193 mutex_unlock(&ctx->uring_lock);
2757be22 4194 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
edafccee
JA
4195out_fput:
4196 fdput(f);
4197 return ret;
4198}
4199
2b188cc1
JA
4200static int __init io_uring_init(void)
4201{
9c71d39a 4202#define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
d7f62e82 4203 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
9c71d39a 4204 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
d7f62e82
SM
4205} while (0)
4206
4207#define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
9c71d39a
SM
4208 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4209#define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4210 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
d7f62e82
SM
4211 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4212 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4213 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4214 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4215 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4216 BUILD_BUG_SQE_ELEM(8, __u64, off);
4217 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
9c71d39a
SM
4218 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4219 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
d7f62e82 4220 BUILD_BUG_SQE_ELEM(16, __u64, addr);
7d67af2c 4221 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
d7f62e82
SM
4222 BUILD_BUG_SQE_ELEM(24, __u32, len);
4223 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4224 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4225 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4226 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
5769a351
JX
4227 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4228 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
d7f62e82
SM
4229 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4230 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4231 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4232 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4233 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4234 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4235 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4236 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
7d67af2c 4237 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
9c71d39a
SM
4238 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4239 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4240 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4241 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4242 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
d7f62e82
SM
4243 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4244 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
16340eab 4245 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
d7f62e82 4246 BUILD_BUG_SQE_ELEM(42, __u16, personality);
7d67af2c 4247 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
b9445598 4248 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
b48c312b
PB
4249 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4250 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
e9621e2b 4251 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
9c71d39a
SM
4252 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4253 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
d7f62e82 4254
b0d658ec
PB
4255 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4256 sizeof(struct io_uring_rsrc_update));
4257 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4258 sizeof(struct io_uring_rsrc_update2));
90499ad0
PB
4259
4260 /* ->buf_index is u16 */
c7fb1942
JA
4261 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4262 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4263 offsetof(struct io_uring_buf_ring, tail));
90499ad0 4264
b0d658ec
PB
4265 /* should fit into one byte */
4266 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
68fe256a
PB
4267 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4268 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
b0d658ec 4269
32c2d33e 4270 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
16340eab 4271
3a4b89a2
JA
4272 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4273
d9b57aa3 4274 io_uring_optable_init();
0702e536 4275
91f245d5
JA
4276 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4277 SLAB_ACCOUNT);
2b188cc1
JA
4278 return 0;
4279};
4280__initcall(io_uring_init);