blob: ebfdb2212ec2520829c8d1c209970b5869e1b026 [file] [log] [blame]
Jens Axboe2b188cc2019-01-07 10:46:33 -07001// 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
Stefan Bühler1e84b972019-04-24 23:54:16 +02007 * 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
Pavel Begunkovd068b502021-05-16 22:58:11 +010014 * through a control-dependency in io_get_cqe (smp_store_release to
Stefan Bühler1e84b972019-04-24 23:54:16 +020015 * 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.
Jens Axboe2b188cc2019-01-07 10:46:33 -070029 *
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
Christoph Hellwigc992fe22019-01-11 09:43:02 -070040 * Copyright (c) 2018-2019 Christoph Hellwig
Jens Axboe2b188cc2019-01-07 10:46:33 -070041 */
42#include <linux/kernel.h>
43#include <linux/init.h>
44#include <linux/errno.h>
45#include <linux/syscalls.h>
Jens Axboe52de1fe2020-02-27 10:15:42 -070046#include <net/compat.h>
Jens Axboe2b188cc2019-01-07 10:46:33 -070047#include <linux/refcount.h>
48#include <linux/uio.h>
Pavel Begunkov6b47ee62020-01-18 20:22:41 +030049#include <linux/bits.h>
Jens Axboe2b188cc2019-01-07 10:46:33 -070050
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>
Jens Axboe2b188cc2019-01-07 10:46:33 -070057#include <linux/percpu.h>
58#include <linux/slab.h>
Jens Axboeedafcce2019-01-09 09:16:05 -070059#include <linux/bvec.h>
Jens Axboe2b188cc2019-01-07 10:46:33 -070060#include <linux/net.h>
61#include <net/sock.h>
62#include <net/af_unix.h>
Jens Axboe6b063142019-01-10 22:13:58 -070063#include <net/scm.h>
Jens Axboe2b188cc2019-01-07 10:46:33 -070064#include <linux/anon_inodes.h>
65#include <linux/sched/mm.h>
66#include <linux/uaccess.h>
67#include <linux/nospec.h>
Jens Axboeaa4c3962019-11-29 10:14:00 -070068#include <linux/highmem.h>
Jens Axboe15b71ab2019-12-11 11:20:36 -070069#include <linux/fsnotify.h>
Jens Axboe4840e412019-12-25 22:03:45 -070070#include <linux/fadvise.h>
Jens Axboeb41e9852020-02-17 09:52:41 -070071#include <linux/task_work.h>
Jens Axboe0f212202020-09-13 13:09:39 -060072#include <linux/io_uring.h>
Paul Moore5bd21822021-02-16 19:46:48 -050073#include <linux/audit.h>
Paul Moorecdc14042021-02-01 19:56:49 -050074#include <linux/security.h>
Jens Axboe2b188cc2019-01-07 10:46:33 -070075
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +020076#define CREATE_TRACE_POINTS
77#include <trace/events/io_uring.h>
78
Jens Axboe2b188cc2019-01-07 10:46:33 -070079#include <uapi/linux/io_uring.h>
80
Jens Axboe561fb042019-10-24 07:25:42 -060081#include "io-wq.h"
Jens Axboe2b188cc2019-01-07 10:46:33 -070082
Jens Axboede230772022-05-24 12:45:38 -060083#include "io_uring.h"
Jens Axboe329061d2022-05-25 20:31:09 -060084#include "opdef.h"
Jens Axboee418bbc2022-05-25 08:56:52 -060085#include "refs.h"
Jens Axboec9f06aa2022-05-25 11:01:04 -060086#include "tctx.h"
Jens Axboe17437f32022-05-25 09:13:39 -060087#include "sqpoll.h"
Jens Axboea4ad4f72022-05-25 10:40:19 -060088#include "fdinfo.h"
Jens Axboe3b774952022-06-13 07:07:23 -060089#include "kbuf.h"
Jens Axboe73572982022-06-13 07:12:45 -060090#include "rsrc.h"
Hao Xu38513c42022-06-16 10:22:02 +010091#include "cancel.h"
Jens Axboe43e0bbb2022-07-07 14:30:09 -060092#include "net.h"
Pavel Begunkoveb42ceb2022-07-12 21:52:38 +010093#include "notif.h"
Jens Axboee27f9282022-05-24 10:56:14 -060094
Jens Axboe599151432022-05-25 08:57:27 -060095#include "timeout.h"
Jens Axboe329061d2022-05-25 20:31:09 -060096#include "poll.h"
Jens Axboe9b797a32022-07-07 14:16:20 -060097#include "alloc_cache.h"
Jens Axboe5e2a18d2022-05-24 11:46:43 -060098
Daniel Xu5277dea2019-09-14 14:23:45 -070099#define IORING_MAX_ENTRIES 32768
Jens Axboe33a107f2019-10-04 12:10:03 -0600100#define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
Jens Axboe65e19f52019-10-26 07:20:21 -0600101
Stefano Garzarella21b55db2020-08-27 16:58:30 +0200102#define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
103 IORING_REGISTER_LAST + IORING_OP_LAST)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700104
Pavel Begunkov68fe2562021-09-15 12:03:38 +0100105#define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
106 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
107
Pavel Begunkov5562a8d2021-11-10 15:49:34 +0000108#define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
109 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
Pavel Begunkov68fe2562021-09-15 12:03:38 +0100110
Pavel Begunkovc8543572021-06-17 18:14:04 +0100111#define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
Jens Axboe9cae36a2022-06-01 23:57:02 -0600112 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
113 REQ_F_ASYNC_DATA)
Pavel Begunkovb16fed662021-02-18 18:29:40 +0000114
Pavel Begunkova538be52022-03-21 22:02:22 +0000115#define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
116 IO_REQ_CLEAN_FLAGS)
117
Pavel Begunkov09899b12021-06-14 02:36:22 +0100118#define IO_TCTX_REFS_CACHE_NR (1U << 10)
119
Pavel Begunkov6dd0be12021-02-10 00:03:13 +0000120#define IO_COMPL_BATCH 32
Pavel Begunkovbf019da2021-02-10 00:03:17 +0000121#define IO_REQ_ALLOC_BATCH 8
Pavel Begunkov258b29a2021-02-10 00:03:10 +0000122
Pavel Begunkov992da012021-06-10 16:37:37 +0100123enum {
Dylan Yudaken10988a02022-04-21 02:13:43 -0700124 IO_CHECK_CQ_OVERFLOW_BIT,
Dylan Yudaken155bc952022-04-21 02:13:44 -0700125 IO_CHECK_CQ_DROPPED_BIT,
Dylan Yudaken10988a02022-04-21 02:13:43 -0700126};
127
Pavel Begunkov27dc8332020-07-13 23:37:14 +0300128struct io_defer_entry {
129 struct list_head list;
130 struct io_kiocb *req;
Pavel Begunkov9cf7c102020-07-13 23:37:15 +0300131 u32 seq;
Pavel Begunkov27dc8332020-07-13 23:37:14 +0300132};
133
Pavel Begunkov0756a862021-08-15 10:40:25 +0100134/* requests with any of those set should undergo io_disarm_next() */
135#define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
Pavel Begunkovda1a08c2022-04-15 22:08:29 +0100136#define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
Pavel Begunkov0756a862021-08-15 10:40:25 +0100137
Pavel Begunkovaffa87d2022-06-20 01:25:52 +0100138static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
Pavel Begunkov9936c7c2021-02-04 13:51:56 +0000139 struct task_struct *task,
Pavel Begunkov3dd0c972021-05-16 22:58:04 +0100140 bool cancel_all);
Pavel Begunkov1ffc5422020-12-30 21:34:15 +0000141
Jens Axboec7dae4b2021-02-09 19:53:37 -0700142static void io_dismantle_req(struct io_kiocb *req);
Pavel Begunkov68fb8972021-03-19 17:22:41 +0000143static void io_clean_op(struct io_kiocb *req);
Pavel Begunkovcbc2e202022-04-15 22:08:26 +0100144static void io_queue_sqe(struct io_kiocb *req);
Jens Axboede0617e2019-04-06 21:51:27 -0600145
Pavel Begunkovc4501782021-09-08 16:40:52 +0100146static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
Jens Axboe9a56a232019-01-09 09:06:50 -0700147
Jens Axboe2b188cc2019-01-07 10:46:33 -0700148static struct kmem_cache *req_cachep;
149
Jens Axboe2b188cc2019-01-07 10:46:33 -0700150struct sock *io_uring_get_socket(struct file *file)
151{
152#if defined(CONFIG_UNIX)
Jens Axboecd40cae2022-05-24 21:54:43 -0600153 if (io_is_uring_fops(file)) {
Jens Axboe2b188cc2019-01-07 10:46:33 -0700154 struct io_ring_ctx *ctx = file->private_data;
155
156 return ctx->ring_sock->sk;
157 }
158#endif
159 return NULL;
160}
161EXPORT_SYMBOL(io_uring_get_socket);
162
Pavel Begunkovc4501782021-09-08 16:40:52 +0100163static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
164{
Pavel Begunkov6f33b0b2021-09-24 21:59:44 +0100165 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
Pavel Begunkovc4501782021-09-08 16:40:52 +0100166 __io_submit_flush_completions(ctx);
167}
168
Pavel Begunkovfaf88dd2022-06-17 09:48:01 +0100169static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
170{
171 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
172}
173
Jens Axboe9cae36a2022-06-01 23:57:02 -0600174static bool io_match_linked(struct io_kiocb *head)
175{
176 struct io_kiocb *req;
177
178 io_for_each_link(req, head) {
179 if (req->flags & REQ_F_INFLIGHT)
180 return true;
181 }
182 return false;
Pavel Begunkov6af3f482021-11-26 14:38:15 +0000183}
184
185/*
186 * As io_match_task() but protected against racing with linked timeouts.
187 * User must not hold timeout_lock.
188 */
Jens Axboe329061d2022-05-25 20:31:09 -0600189bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
190 bool cancel_all)
Pavel Begunkov6af3f482021-11-26 14:38:15 +0000191{
Jens Axboe9cae36a2022-06-01 23:57:02 -0600192 bool matched;
193
Pavel Begunkov6af3f482021-11-26 14:38:15 +0000194 if (task && head->task != task)
195 return false;
Jens Axboe9cae36a2022-06-01 23:57:02 -0600196 if (cancel_all)
197 return true;
198
199 if (head->flags & REQ_F_LINK_TIMEOUT) {
200 struct io_ring_ctx *ctx = head->ctx;
201
202 /* protect against races with linked timeouts */
203 spin_lock_irq(&ctx->timeout_lock);
204 matched = io_match_linked(head);
205 spin_unlock_irq(&ctx->timeout_lock);
206 } else {
207 matched = io_match_linked(head);
208 }
209 return matched;
Pavel Begunkov6af3f482021-11-26 14:38:15 +0000210}
211
Hao Xua8295b92021-08-27 17:46:09 +0800212static inline void req_fail_link_node(struct io_kiocb *req, int res)
213{
214 req_set_fail(req);
Jens Axboe97b388d2022-05-24 15:21:00 -0600215 io_req_set_res(req, res, 0);
Hao Xua8295b92021-08-27 17:46:09 +0800216}
217
Pavel Begunkovfa054572022-04-12 15:09:48 +0100218static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
219{
220 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
Jens Axboec40f6372020-06-25 15:39:59 -0600221}
Jens Axboe4a38aed22020-05-14 17:21:15 -0600222
Pavel Begunkovc0724812021-10-04 20:02:54 +0100223static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700224{
225 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
226
Jens Axboe0f158b42020-05-14 17:18:39 -0600227 complete(&ctx->ref_comp);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700228}
229
Pavel Begunkovc0724812021-10-04 20:02:54 +0100230static __cold void io_fallback_req_func(struct work_struct *work)
Pavel Begunkovf56165e2021-08-09 20:18:07 +0100231{
232 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
233 fallback_work.work);
234 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
235 struct io_kiocb *req, *tmp;
Pavel Begunkovf237c302021-08-18 12:42:46 +0100236 bool locked = false;
Pavel Begunkovf56165e2021-08-09 20:18:07 +0100237
238 percpu_ref_get(&ctx->refs);
Pavel Begunkov3218e5d2022-06-25 11:52:59 +0100239 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
Pavel Begunkovf237c302021-08-18 12:42:46 +0100240 req->io_task_work.func(req, &locked);
Pavel Begunkov5636c002021-08-18 12:42:45 +0100241
Pavel Begunkovf237c302021-08-18 12:42:46 +0100242 if (locked) {
Pavel Begunkovc4501782021-09-08 16:40:52 +0100243 io_submit_flush_completions(ctx);
Pavel Begunkovf237c302021-08-18 12:42:46 +0100244 mutex_unlock(&ctx->uring_lock);
245 }
Pavel Begunkovf56165e2021-08-09 20:18:07 +0100246 percpu_ref_put(&ctx->refs);
247}
248
Pavel Begunkove6f89be2022-06-16 10:22:10 +0100249static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
250{
251 unsigned hash_buckets = 1U << bits;
252 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
253
254 table->hbs = kmalloc(hash_size, GFP_KERNEL);
255 if (!table->hbs)
256 return -ENOMEM;
257
258 table->hash_bits = bits;
259 init_hash_table(table, hash_buckets);
260 return 0;
261}
262
Pavel Begunkovc0724812021-10-04 20:02:54 +0100263static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700264{
265 struct io_ring_ctx *ctx;
Jens Axboe9cfc7e92022-05-01 10:52:44 -0600266 int hash_bits;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700267
268 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
269 if (!ctx)
270 return NULL;
271
Jens Axboe9cfc7e92022-05-01 10:52:44 -0600272 xa_init(&ctx->io_bl_xa);
273
Jens Axboe78076bb2019-12-04 19:56:40 -0700274 /*
275 * Use 5 bits less than the max cq entries, that should give us around
Pavel Begunkov4a077232022-06-16 10:22:05 +0100276 * 32 entries per hash list if totally full and uniformly spread, but
277 * don't keep too many buckets to not overconsume memory.
Jens Axboe78076bb2019-12-04 19:56:40 -0700278 */
Pavel Begunkov4a077232022-06-16 10:22:05 +0100279 hash_bits = ilog2(p->cq_entries) - 5;
280 hash_bits = clamp(hash_bits, 1, 8);
Pavel Begunkove6f89be2022-06-16 10:22:10 +0100281 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
Jens Axboe78076bb2019-12-04 19:56:40 -0700282 goto err;
Pavel Begunkov9ca9fb22022-06-16 10:22:12 +0100283 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
284 goto err;
Hao Xu38513c42022-06-16 10:22:02 +0100285
Pavel Begunkov62248432021-04-28 13:11:29 +0100286 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
287 if (!ctx->dummy_ubuf)
288 goto err;
289 /* set invalid range, so io_import_fixed() fails meeting it */
290 ctx->dummy_ubuf->ubuf = -1UL;
291
Roman Gushchin21482892019-05-07 10:01:48 -0700292 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
Michal Koutný48904222022-07-15 19:45:01 +0200293 0, GFP_KERNEL))
Jens Axboe206aefd2019-11-07 18:27:42 -0700294 goto err;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700295
296 ctx->flags = p->flags;
Jens Axboe90554202020-09-03 12:12:41 -0600297 init_waitqueue_head(&ctx->sqo_sq_wait);
Jens Axboe69fb2132020-09-14 11:16:23 -0600298 INIT_LIST_HEAD(&ctx->sqd_list);
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700299 INIT_LIST_HEAD(&ctx->cq_overflow_list);
Jens Axboecc3cec82022-03-08 17:46:52 -0700300 INIT_LIST_HEAD(&ctx->io_buffers_cache);
Jens Axboe9b797a32022-07-07 14:16:20 -0600301 io_alloc_cache_init(&ctx->apoll_cache);
Jens Axboe43e0bbb2022-07-07 14:30:09 -0600302 io_alloc_cache_init(&ctx->netmsg_cache);
Jens Axboe0f158b42020-05-14 17:18:39 -0600303 init_completion(&ctx->ref_comp);
Matthew Wilcox (Oracle)61cf9372021-03-08 14:16:16 +0000304 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700305 mutex_init(&ctx->uring_lock);
Pavel Begunkov311997b2021-06-14 23:37:28 +0100306 init_waitqueue_head(&ctx->cq_wait);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700307 spin_lock_init(&ctx->completion_lock);
Jens Axboe89850fc2021-08-10 15:11:51 -0600308 spin_lock_init(&ctx->timeout_lock);
Pavel Begunkov5eef4e82021-09-24 21:59:49 +0100309 INIT_WQ_LIST(&ctx->iopoll_list);
Jens Axboecc3cec82022-03-08 17:46:52 -0700310 INIT_LIST_HEAD(&ctx->io_buffers_pages);
311 INIT_LIST_HEAD(&ctx->io_buffers_comp);
Jens Axboede0617e2019-04-06 21:51:27 -0600312 INIT_LIST_HEAD(&ctx->defer_list);
Jens Axboe5262f562019-09-17 12:26:57 -0600313 INIT_LIST_HEAD(&ctx->timeout_list);
Pavel Begunkovef9dd632021-08-28 19:54:38 -0600314 INIT_LIST_HEAD(&ctx->ltimeout_list);
Bijan Mottahedehd67d2262021-01-15 17:37:46 +0000315 spin_lock_init(&ctx->rsrc_ref_lock);
316 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
Bijan Mottahedeh269bbe52021-01-15 17:37:44 +0000317 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
318 init_llist_head(&ctx->rsrc_put_llist);
Pavel Begunkov13bf43f2021-03-06 11:02:12 +0000319 INIT_LIST_HEAD(&ctx->tctx_list);
Pavel Begunkovc2b6c6b2021-09-24 21:59:47 +0100320 ctx->submit_state.free_list.next = NULL;
321 INIT_WQ_LIST(&ctx->locked_free_list);
Pavel Begunkov9011bf92021-06-30 21:54:03 +0100322 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
Pavel Begunkov6f33b0b2021-09-24 21:59:44 +0100323 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700324 return ctx;
Jens Axboe206aefd2019-11-07 18:27:42 -0700325err:
Pavel Begunkov62248432021-04-28 13:11:29 +0100326 kfree(ctx->dummy_ubuf);
Pavel Begunkove6f89be2022-06-16 10:22:10 +0100327 kfree(ctx->cancel_table.hbs);
Pavel Begunkov9ca9fb22022-06-16 10:22:12 +0100328 kfree(ctx->cancel_table_locked.hbs);
Jens Axboe9cfc7e92022-05-01 10:52:44 -0600329 kfree(ctx->io_bl);
330 xa_destroy(&ctx->io_bl_xa);
Jens Axboe206aefd2019-11-07 18:27:42 -0700331 kfree(ctx);
332 return NULL;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700333}
334
Pavel Begunkov8f6ed492021-05-16 22:58:10 +0100335static void io_account_cq_overflow(struct io_ring_ctx *ctx)
336{
337 struct io_rings *r = ctx->rings;
338
339 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
340 ctx->cq_extra--;
341}
342
Pavel Begunkov9cf7c102020-07-13 23:37:15 +0300343static bool req_need_defer(struct io_kiocb *req, u32 seq)
Jens Axboede0617e2019-04-06 21:51:27 -0600344{
Jens Axboe2bc99302020-07-09 09:43:27 -0600345 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
346 struct io_ring_ctx *ctx = req->ctx;
Jackie Liua197f662019-11-08 08:09:12 -0700347
Pavel Begunkov8f6ed492021-05-16 22:58:10 +0100348 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
Jens Axboe2bc99302020-07-09 09:43:27 -0600349 }
Jens Axboe7adf4ea2019-10-10 21:42:58 -0600350
Bob Liu9d858b22019-11-13 18:06:25 +0800351 return false;
Jens Axboe7adf4ea2019-10-10 21:42:58 -0600352}
353
Jens Axboe9cae36a2022-06-01 23:57:02 -0600354static inline void io_req_track_inflight(struct io_kiocb *req)
355{
356 if (!(req->flags & REQ_F_INFLIGHT)) {
357 req->flags |= REQ_F_INFLIGHT;
Jens Axboe386e4fb2022-06-23 11:06:43 -0600358 atomic_inc(&req->task->io_uring->inflight_tracked);
Jens Axboe9cae36a2022-06-01 23:57:02 -0600359 }
360}
361
Pavel Begunkovfd08e532021-08-11 19:28:31 +0100362static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
363{
Pavel Begunkov906c6ca2021-08-15 10:40:26 +0100364 if (WARN_ON_ONCE(!req->link))
365 return NULL;
366
Pavel Begunkov4d13d1a2021-08-15 10:40:24 +0100367 req->flags &= ~REQ_F_ARM_LTIMEOUT;
368 req->flags |= REQ_F_LINK_TIMEOUT;
Pavel Begunkovfd08e532021-08-11 19:28:31 +0100369
370 /* linked timeouts should have two refs once prep'ed */
Pavel Begunkov48dcd382021-08-15 10:40:18 +0100371 io_req_set_refcount(req);
Pavel Begunkov4d13d1a2021-08-15 10:40:24 +0100372 __io_req_set_refcount(req->link, 2);
373 return req->link;
Pavel Begunkovfd08e532021-08-11 19:28:31 +0100374}
375
376static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
377{
Pavel Begunkov4d13d1a2021-08-15 10:40:24 +0100378 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
Pavel Begunkovfd08e532021-08-11 19:28:31 +0100379 return NULL;
380 return __io_prep_linked_timeout(req);
381}
382
Pavel Begunkovcb2d3442022-04-15 22:08:25 +0100383static noinline void __io_arm_ltimeout(struct io_kiocb *req)
384{
385 io_queue_linked_timeout(__io_prep_linked_timeout(req));
386}
387
388static inline void io_arm_ltimeout(struct io_kiocb *req)
389{
390 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
391 __io_arm_ltimeout(req);
392}
393
Pavel Begunkovcbdcb432020-06-29 19:18:43 +0300394static void io_prep_async_work(struct io_kiocb *req)
Jens Axboe561fb042019-10-24 07:25:42 -0600395{
Jens Axboed3656342019-12-18 09:50:26 -0700396 const struct io_op_def *def = &io_op_defs[req->opcode];
Pavel Begunkov23329512020-10-10 18:34:06 +0100397 struct io_ring_ctx *ctx = req->ctx;
Jens Axboe54a91f32019-09-10 09:15:04 -0600398
Pavel Begunkovb8e64b52021-06-17 18:14:02 +0100399 if (!(req->flags & REQ_F_CREDS)) {
400 req->flags |= REQ_F_CREDS;
Pavel Begunkovc10d1f92021-06-17 18:14:01 +0100401 req->creds = get_current_cred();
Pavel Begunkovb8e64b52021-06-17 18:14:02 +0100402 }
Jens Axboe003e8dc2021-03-06 09:22:27 -0700403
Pavel Begunkove1d675d2021-03-22 01:58:29 +0000404 req->work.list.next = NULL;
405 req->work.flags = 0;
Jens Axboe8e29da62022-04-18 10:44:00 -0600406 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
Pavel Begunkovfeaadc42020-10-22 16:47:16 +0100407 if (req->flags & REQ_F_FORCE_ASYNC)
408 req->work.flags |= IO_WQ_WORK_CONCURRENT;
409
Jens Axboef6b543f2022-07-21 09:06:47 -0600410 if (req->file && !io_req_ffs_set(req))
411 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
412
Jens Axboed3656342019-12-18 09:50:26 -0700413 if (req->flags & REQ_F_ISREG) {
Pavel Begunkov23329512020-10-10 18:34:06 +0100414 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
Pavel Begunkov8766dd52020-03-14 00:31:04 +0300415 io_wq_hash_work(&req->work, file_inode(req->file));
Jens Axboe4b982bd2021-04-01 08:38:34 -0600416 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
Jens Axboed3656342019-12-18 09:50:26 -0700417 if (def->unbound_nonreg_file)
Jens Axboe3529d8c2019-12-19 18:24:38 -0700418 req->work.flags |= IO_WQ_WORK_UNBOUND;
Jens Axboe54a91f32019-09-10 09:15:04 -0600419 }
Jens Axboe561fb042019-10-24 07:25:42 -0600420}
421
Pavel Begunkovcbdcb432020-06-29 19:18:43 +0300422static void io_prep_async_link(struct io_kiocb *req)
423{
424 struct io_kiocb *cur;
425
Pavel Begunkov44eff402021-07-26 14:14:31 +0100426 if (req->flags & REQ_F_LINK_TIMEOUT) {
427 struct io_ring_ctx *ctx = req->ctx;
428
Pavel Begunkov674ee8e2021-11-23 01:45:35 +0000429 spin_lock_irq(&ctx->timeout_lock);
Pavel Begunkov44eff402021-07-26 14:14:31 +0100430 io_for_each_link(cur, req)
431 io_prep_async_work(cur);
Pavel Begunkov674ee8e2021-11-23 01:45:35 +0000432 spin_unlock_irq(&ctx->timeout_lock);
Pavel Begunkov44eff402021-07-26 14:14:31 +0100433 } else {
434 io_for_each_link(cur, req)
435 io_prep_async_work(cur);
436 }
Pavel Begunkovcbdcb432020-06-29 19:18:43 +0300437}
438
Jens Axboef3b44f92022-06-13 07:27:03 -0600439void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
Jens Axboe561fb042019-10-24 07:25:42 -0600440{
Pavel Begunkovcbdcb432020-06-29 19:18:43 +0300441 struct io_kiocb *link = io_prep_linked_timeout(req);
Jens Axboe5aa75ed2021-02-16 12:56:50 -0700442 struct io_uring_task *tctx = req->task->io_uring;
Jens Axboe561fb042019-10-24 07:25:42 -0600443
Jens Axboe3bfe6102021-02-16 14:15:30 -0700444 BUG_ON(!tctx);
445 BUG_ON(!tctx->io_wq);
Jens Axboe561fb042019-10-24 07:25:42 -0600446
Pavel Begunkovcbdcb432020-06-29 19:18:43 +0300447 /* init ->work of the whole link before punting */
448 io_prep_async_link(req);
Jens Axboe991468d2021-07-23 11:53:54 -0600449
450 /*
451 * Not expected to happen, but if we do have a bug where this _can_
452 * happen, catch it here and ensure the request is marked as
453 * canceled. That will make io-wq go through the usual work cancel
454 * procedure rather than attempt to run this request (or create a new
455 * worker for it).
456 */
457 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
458 req->work.flags |= IO_WQ_WORK_CANCEL;
459
Pavel Begunkov48863ff2022-06-16 13:57:20 +0100460 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
Pavel Begunkovebf93662021-03-01 18:20:47 +0000461 io_wq_enqueue(tctx->io_wq, &req->work);
Jens Axboe7271ef32020-08-10 09:55:22 -0600462 if (link)
463 io_queue_linked_timeout(link);
Pavel Begunkovcbdcb432020-06-29 19:18:43 +0300464}
465
Pavel Begunkovc0724812021-10-04 20:02:54 +0100466static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
Pavel Begunkov04518942020-05-26 20:34:05 +0300467{
Pavel Begunkov441b8a72021-06-14 23:37:31 +0100468 while (!list_empty(&ctx->defer_list)) {
Pavel Begunkov27dc8332020-07-13 23:37:14 +0300469 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
470 struct io_defer_entry, list);
Pavel Begunkov04518942020-05-26 20:34:05 +0300471
Pavel Begunkov9cf7c102020-07-13 23:37:15 +0300472 if (req_need_defer(de->req, de->seq))
Pavel Begunkov04518942020-05-26 20:34:05 +0300473 break;
Pavel Begunkov27dc8332020-07-13 23:37:14 +0300474 list_del_init(&de->list);
Pavel Begunkov907d1df2021-01-26 23:35:10 +0000475 io_req_task_queue(de->req);
Pavel Begunkov27dc8332020-07-13 23:37:14 +0300476 kfree(de);
Pavel Begunkov441b8a72021-06-14 23:37:31 +0100477 }
Pavel Begunkov04518942020-05-26 20:34:05 +0300478}
479
Usama Arif77bc59b2022-02-04 14:51:14 +0000480static void io_eventfd_signal(struct io_ring_ctx *ctx)
Jens Axboef2842ab2020-01-08 11:04:00 -0700481{
Usama Arif77bc59b2022-02-04 14:51:14 +0000482 struct io_ev_fd *ev_fd;
Pavel Begunkov305bef92022-06-20 01:25:55 +0100483 bool skip;
484
485 spin_lock(&ctx->completion_lock);
486 /*
487 * Eventfd should only get triggered when at least one event has been
488 * posted. Some applications rely on the eventfd notification count only
489 * changing IFF a new CQE has been added to the CQ ring. There's no
490 * depedency on 1:1 relationship between how many times this function is
491 * called (and hence the eventfd count) and number of CQEs posted to the
492 * CQ ring.
493 */
494 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
495 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
496 spin_unlock(&ctx->completion_lock);
497 if (skip)
498 return;
Usama Arif77bc59b2022-02-04 14:51:14 +0000499
Usama Arif77bc59b2022-02-04 14:51:14 +0000500 rcu_read_lock();
501 /*
502 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
503 * and eventfd_signal
504 */
505 ev_fd = rcu_dereference(ctx->io_ev_fd);
506
507 /*
508 * Check again if ev_fd exists incase an io_eventfd_unregister call
509 * completed between the NULL check of ctx->io_ev_fd at the start of
510 * the function and rcu_read_lock.
511 */
512 if (unlikely(!ev_fd))
513 goto out;
Stefano Garzarella7e55a192020-05-15 18:38:05 +0200514 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
Usama Arif77bc59b2022-02-04 14:51:14 +0000515 goto out;
516
Usama Arifc75312d2022-02-04 14:51:15 +0000517 if (!ev_fd->eventfd_async || io_wq_current_is_worker())
Usama Arif77bc59b2022-02-04 14:51:14 +0000518 eventfd_signal(ev_fd->cq_ev_fd, 1);
Usama Arif77bc59b2022-02-04 14:51:14 +0000519out:
520 rcu_read_unlock();
Jens Axboef2842ab2020-01-08 11:04:00 -0700521}
522
Pavel Begunkova830ffd2022-06-19 12:26:06 +0100523void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
524{
525 if (ctx->off_timeout_used || ctx->drain_active) {
526 spin_lock(&ctx->completion_lock);
527 if (ctx->off_timeout_used)
528 io_flush_timeouts(ctx);
529 if (ctx->drain_active)
530 io_queue_deferred(ctx);
531 spin_unlock(&ctx->completion_lock);
532 }
533 if (ctx->has_evfd)
534 io_eventfd_signal(ctx);
535}
536
Pavel Begunkov253993212022-06-20 01:25:56 +0100537static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx)
Jens Axboe8c838782019-03-12 15:48:16 -0600538{
Pavel Begunkov46929b02022-06-20 01:25:57 +0100539 io_commit_cqring_flush(ctx);
Pavel Begunkov9aa8dfd2022-03-17 02:03:42 +0000540 io_cqring_wake(ctx);
Jens Axboe8c838782019-03-12 15:48:16 -0600541}
542
Pavel Begunkov253993212022-06-20 01:25:56 +0100543static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
544 __releases(ctx->completion_lock)
545{
546 io_commit_cqring(ctx);
547 spin_unlock(&ctx->completion_lock);
548 io_cqring_ev_posted(ctx);
549}
550
551void io_cq_unlock_post(struct io_ring_ctx *ctx)
552{
553 __io_cq_unlock_post(ctx);
554}
555
Jens Axboec4a2ed72019-11-21 21:01:26 -0700556/* Returns true if there are no backlogged entries after the flush */
Pavel Begunkov6c2450a2021-02-23 12:40:22 +0000557static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700558{
Pavel Begunkov305bef92022-06-20 01:25:55 +0100559 bool all_flushed;
Stefan Roesche45a3e02022-04-26 11:21:30 -0700560 size_t cqe_size = sizeof(struct io_uring_cqe);
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700561
Pavel Begunkova566c552021-05-16 22:58:08 +0100562 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
Pavel Begunkove23de152020-12-17 00:24:37 +0000563 return false;
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700564
Stefan Roesche45a3e02022-04-26 11:21:30 -0700565 if (ctx->flags & IORING_SETUP_CQE32)
566 cqe_size <<= 1;
567
Pavel Begunkov253993212022-06-20 01:25:56 +0100568 io_cq_lock(ctx);
Pavel Begunkov6c2450a2021-02-23 12:40:22 +0000569 while (!list_empty(&ctx->cq_overflow_list)) {
Pavel Begunkovd068b502021-05-16 22:58:11 +0100570 struct io_uring_cqe *cqe = io_get_cqe(ctx);
Pavel Begunkov6c2450a2021-02-23 12:40:22 +0000571 struct io_overflow_cqe *ocqe;
Jens Axboee6c8aa92020-09-28 13:10:13 -0600572
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700573 if (!cqe && !force)
574 break;
Pavel Begunkov6c2450a2021-02-23 12:40:22 +0000575 ocqe = list_first_entry(&ctx->cq_overflow_list,
576 struct io_overflow_cqe, list);
577 if (cqe)
Stefan Roesche45a3e02022-04-26 11:21:30 -0700578 memcpy(cqe, &ocqe->cqe, cqe_size);
Pavel Begunkov6c2450a2021-02-23 12:40:22 +0000579 else
Pavel Begunkov8f6ed492021-05-16 22:58:10 +0100580 io_account_cq_overflow(ctx);
581
Pavel Begunkov6c2450a2021-02-23 12:40:22 +0000582 list_del(&ocqe->list);
583 kfree(ocqe);
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700584 }
585
Pavel Begunkov09e88402020-12-17 00:24:38 +0000586 all_flushed = list_empty(&ctx->cq_overflow_list);
587 if (all_flushed) {
Dylan Yudaken10988a02022-04-21 02:13:43 -0700588 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
Jens Axboe3a4b89a2022-04-25 19:49:00 -0600589 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
Pavel Begunkov09e88402020-12-17 00:24:38 +0000590 }
Pavel Begunkov46930142020-07-30 18:43:49 +0300591
Pavel Begunkov253993212022-06-20 01:25:56 +0100592 io_cq_unlock_post(ctx);
Pavel Begunkov09e88402020-12-17 00:24:38 +0000593 return all_flushed;
Jens Axboe1d7bb1d2019-11-06 11:31:17 -0700594}
595
Pavel Begunkov90f67362021-08-09 20:18:12 +0100596static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
Pavel Begunkov6c503152021-01-04 20:36:36 +0000597{
Jens Axboeca0a2652021-03-04 17:15:48 -0700598 bool ret = true;
599
Dylan Yudaken10988a02022-04-21 02:13:43 -0700600 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
Pavel Begunkov6c503152021-01-04 20:36:36 +0000601 /* iopoll syncs against uring_lock, not completion_lock */
602 if (ctx->flags & IORING_SETUP_IOPOLL)
603 mutex_lock(&ctx->uring_lock);
Pavel Begunkov90f67362021-08-09 20:18:12 +0100604 ret = __io_cqring_overflow_flush(ctx, false);
Pavel Begunkov6c503152021-01-04 20:36:36 +0000605 if (ctx->flags & IORING_SETUP_IOPOLL)
606 mutex_unlock(&ctx->uring_lock);
607 }
Jens Axboeca0a2652021-03-04 17:15:48 -0700608
609 return ret;
Pavel Begunkov6c503152021-01-04 20:36:36 +0000610}
611
Pavel Begunkove70cb602022-07-12 21:52:37 +0100612void __io_put_task(struct task_struct *task, int nr)
Pavel Begunkov6a290a12021-08-09 13:04:13 +0100613{
614 struct io_uring_task *tctx = task->io_uring;
615
Pavel Begunkov9d170162022-03-25 11:52:15 +0000616 percpu_counter_sub(&tctx->inflight, nr);
617 if (unlikely(atomic_read(&tctx->in_idle)))
618 wake_up(&tctx->wait);
619 put_task_struct_many(task, nr);
620}
621
Pavel Begunkov63809132022-07-12 21:52:47 +0100622void io_task_refs_refill(struct io_uring_task *tctx)
Pavel Begunkov9a108672021-08-27 11:55:01 +0100623{
624 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
625
626 percpu_counter_add(&tctx->inflight, refill);
627 refcount_add(refill, &current->usage);
628 tctx->cached_refs += refill;
629}
630
Pavel Begunkov3cc7fdb2022-01-09 00:53:22 +0000631static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
632{
633 struct io_uring_task *tctx = task->io_uring;
634 unsigned int refs = tctx->cached_refs;
635
636 if (refs) {
637 tctx->cached_refs = 0;
638 percpu_counter_sub(&tctx->inflight, refs);
639 put_task_struct_many(task, refs);
640 }
641}
642
Pavel Begunkov68494a62022-06-17 09:48:02 +0100643static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
644 s32 res, u32 cflags, u64 extra1, u64 extra2)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700645{
Pavel Begunkovcce4b8b2021-04-13 02:58:44 +0100646 struct io_overflow_cqe *ocqe;
Stefan Roesche45a3e02022-04-26 11:21:30 -0700647 size_t ocq_size = sizeof(struct io_overflow_cqe);
648 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700649
Stefan Roesche45a3e02022-04-26 11:21:30 -0700650 if (is_cqe32)
651 ocq_size += sizeof(struct io_uring_cqe);
652
653 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
Dylan Yudaken08dcd022022-04-21 02:13:41 -0700654 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
Pavel Begunkovcce4b8b2021-04-13 02:58:44 +0100655 if (!ocqe) {
656 /*
657 * If we're in ring overflow flush mode, or in task cancel mode,
658 * or cannot allocate an overflow entry, then we need to drop it
659 * on the floor.
660 */
Pavel Begunkov8f6ed492021-05-16 22:58:10 +0100661 io_account_cq_overflow(ctx);
Dylan Yudaken155bc952022-04-21 02:13:44 -0700662 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
Pavel Begunkovcce4b8b2021-04-13 02:58:44 +0100663 return false;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700664 }
Pavel Begunkovcce4b8b2021-04-13 02:58:44 +0100665 if (list_empty(&ctx->cq_overflow_list)) {
Dylan Yudaken10988a02022-04-21 02:13:43 -0700666 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
Jens Axboe3a4b89a2022-04-25 19:49:00 -0600667 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
Nadav Amit20c0b382021-08-07 17:13:42 -0700668
Pavel Begunkovcce4b8b2021-04-13 02:58:44 +0100669 }
Pavel Begunkovd4d19c12021-04-25 14:32:17 +0100670 ocqe->cqe.user_data = user_data;
Pavel Begunkovcce4b8b2021-04-13 02:58:44 +0100671 ocqe->cqe.res = res;
672 ocqe->cqe.flags = cflags;
Stefan Roesche45a3e02022-04-26 11:21:30 -0700673 if (is_cqe32) {
674 ocqe->cqe.big_cqe[0] = extra1;
675 ocqe->cqe.big_cqe[1] = extra2;
676 }
Pavel Begunkovcce4b8b2021-04-13 02:58:44 +0100677 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
678 return true;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700679}
680
Pavel Begunkov68494a62022-06-17 09:48:02 +0100681bool io_req_cqe_overflow(struct io_kiocb *req)
682{
683 if (!(req->flags & REQ_F_CQE32_INIT)) {
684 req->extra1 = 0;
685 req->extra2 = 0;
686 }
687 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
688 req->cqe.res, req->cqe.flags,
689 req->extra1, req->extra2);
690}
691
Pavel Begunkovfaf88dd2022-06-17 09:48:01 +0100692/*
693 * writes to the cq entry need to come after reading head; the
694 * control dependency is enough as we're using WRITE_ONCE to
695 * fill the cq entry
696 */
697struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx)
698{
699 struct io_rings *rings = ctx->rings;
700 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
Pavel Begunkovfaf88dd2022-06-17 09:48:01 +0100701 unsigned int free, queued, len;
702
Pavel Begunkovfaf88dd2022-06-17 09:48:01 +0100703
704 /* userspace may cheat modifying the tail, be safe and do min */
705 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
706 free = ctx->cq_entries - queued;
707 /* we need a contiguous range, limit based on the current array offset */
708 len = min(free, ctx->cq_entries - off);
709 if (!len)
710 return NULL;
711
Pavel Begunkovb3659a62022-06-17 09:48:05 +0100712 if (ctx->flags & IORING_SETUP_CQE32) {
713 off <<= 1;
714 len <<= 1;
715 }
716
Pavel Begunkovfaf88dd2022-06-17 09:48:01 +0100717 ctx->cqe_cached = &rings->cqes[off];
718 ctx->cqe_sentinel = ctx->cqe_cached + len;
Pavel Begunkovb3659a62022-06-17 09:48:05 +0100719
720 ctx->cached_cq_tail++;
Pavel Begunkovfaf88dd2022-06-17 09:48:01 +0100721 ctx->cqe_cached++;
Pavel Begunkovb3659a62022-06-17 09:48:05 +0100722 if (ctx->flags & IORING_SETUP_CQE32)
723 ctx->cqe_cached++;
724 return &rings->cqes[off];
Pavel Begunkovfaf88dd2022-06-17 09:48:01 +0100725}
726
Pavel Begunkoveb42ceb2022-07-12 21:52:38 +0100727bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
728 bool allow_overflow)
Pavel Begunkov913a5712021-11-10 15:49:31 +0000729{
Pavel Begunkovcd949032022-06-15 11:23:06 +0100730 struct io_uring_cqe *cqe;
731
Pavel Begunkov913a5712021-11-10 15:49:31 +0000732 ctx->cq_extra++;
Pavel Begunkovcd949032022-06-15 11:23:06 +0100733
734 /*
735 * If we can't get a cq entry, userspace overflowed the
736 * submission (by quite a lot). Increment the overflow count in
737 * the ring.
738 */
739 cqe = io_get_cqe(ctx);
740 if (likely(cqe)) {
Dylan Yudakene0486f32022-06-30 02:12:31 -0700741 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
742
Pavel Begunkovcd949032022-06-15 11:23:06 +0100743 WRITE_ONCE(cqe->user_data, user_data);
744 WRITE_ONCE(cqe->res, res);
745 WRITE_ONCE(cqe->flags, cflags);
Pavel Begunkovc5595972022-06-15 11:23:07 +0100746
747 if (ctx->flags & IORING_SETUP_CQE32) {
748 WRITE_ONCE(cqe->big_cqe[0], 0);
749 WRITE_ONCE(cqe->big_cqe[1], 0);
750 }
Pavel Begunkovcd949032022-06-15 11:23:06 +0100751 return true;
752 }
Dylan Yudaken52120f02022-06-30 02:12:26 -0700753
754 if (allow_overflow)
755 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
756
757 return false;
Jens Axboebcda7ba2020-02-23 16:42:51 -0700758}
759
Pavel Begunkovd245bca2022-06-17 09:48:00 +0100760bool io_post_aux_cqe(struct io_ring_ctx *ctx,
Dylan Yudaken52120f02022-06-30 02:12:26 -0700761 u64 user_data, s32 res, u32 cflags,
762 bool allow_overflow)
Pavel Begunkovd245bca2022-06-17 09:48:00 +0100763{
764 bool filled;
765
Pavel Begunkov253993212022-06-20 01:25:56 +0100766 io_cq_lock(ctx);
Dylan Yudaken52120f02022-06-30 02:12:26 -0700767 filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow);
Pavel Begunkov253993212022-06-20 01:25:56 +0100768 io_cq_unlock_post(ctx);
Pavel Begunkovd245bca2022-06-17 09:48:00 +0100769 return filled;
770}
771
Stefan Roescheffcf8b2022-04-26 11:21:27 -0700772static void __io_req_complete_put(struct io_kiocb *req)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700773{
Jens Axboec7dae4b2021-02-09 19:53:37 -0700774 /*
775 * If we're the last reference to this request, add to our locked
776 * free_list cache.
777 */
Jens Axboede9b4cc2021-02-24 13:28:27 -0700778 if (req_ref_put_and_test(req)) {
Stefan Roescheffcf8b2022-04-26 11:21:27 -0700779 struct io_ring_ctx *ctx = req->ctx;
780
Pavel Begunkovda1a08c2022-04-15 22:08:29 +0100781 if (req->flags & IO_REQ_LINK_FLAGS) {
Pavel Begunkov0756a862021-08-15 10:40:25 +0100782 if (req->flags & IO_DISARM_MASK)
Pavel Begunkov7a612352021-03-09 00:37:59 +0000783 io_disarm_next(req);
784 if (req->link) {
785 io_req_task_queue(req->link);
786 req->link = NULL;
787 }
788 }
Pavel Begunkov7ac1edc2022-04-18 20:51:15 +0100789 io_req_put_rsrc(req);
Pavel Begunkov8197b052022-03-25 13:00:43 +0000790 /*
791 * Selected buffer deallocation in io_clean_op() assumes that
792 * we don't hold ->completion_lock. Clean them here to avoid
793 * deadlocks.
794 */
795 io_put_kbuf_comp(req);
Jens Axboec7dae4b2021-02-09 19:53:37 -0700796 io_dismantle_req(req);
797 io_put_task(req->task, 1);
Pavel Begunkovc2b6c6b2021-09-24 21:59:47 +0100798 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
Pavel Begunkovd0acdee2021-05-16 22:58:12 +0100799 ctx->locked_free_nr++;
Pavel Begunkov180f8292021-03-14 20:57:09 +0000800 }
Hao Xua37fae82021-12-07 17:39:50 +0800801}
802
Jens Axboe599151432022-05-25 08:57:27 -0600803void __io_req_complete_post(struct io_kiocb *req)
Stefan Roescheffcf8b2022-04-26 11:21:27 -0700804{
Jens Axboe97b388d2022-05-24 15:21:00 -0600805 if (!(req->flags & REQ_F_CQE_SKIP))
Pavel Begunkov91ef75a2022-06-15 11:23:02 +0100806 __io_fill_cqe_req(req->ctx, req);
Stefan Roescheffcf8b2022-04-26 11:21:27 -0700807 __io_req_complete_put(req);
808}
809
Jens Axboe599151432022-05-25 08:57:27 -0600810void io_req_complete_post(struct io_kiocb *req)
Hao Xua37fae82021-12-07 17:39:50 +0800811{
812 struct io_ring_ctx *ctx = req->ctx;
813
Pavel Begunkov253993212022-06-20 01:25:56 +0100814 io_cq_lock(ctx);
Jens Axboe97b388d2022-05-24 15:21:00 -0600815 __io_req_complete_post(req);
Pavel Begunkov253993212022-06-20 01:25:56 +0100816 io_cq_unlock_post(ctx);
Jens Axboe2b188cc2019-01-07 10:46:33 -0700817}
818
Jens Axboe99f15d82022-05-25 05:59:19 -0600819inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
Pavel Begunkova38d68d2021-01-19 13:32:45 +0000820{
Pavel Begunkov75d7b3a2022-06-16 10:21:58 +0100821 io_req_complete_post(req);
Jens Axboebcda7ba2020-02-23 16:42:51 -0700822}
823
Jens Axboe329061d2022-05-25 20:31:09 -0600824void io_req_complete_failed(struct io_kiocb *req, s32 res)
Pavel Begunkovf41db2732021-02-28 22:35:12 +0000825{
Pavel Begunkov93d2bcd2021-05-16 22:58:05 +0100826 req_set_fail(req);
Jens Axboe97b388d2022-05-24 15:21:00 -0600827 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
828 io_req_complete_post(req);
Pavel Begunkovf41db2732021-02-28 22:35:12 +0000829}
830
Pavel Begunkov864ea922021-08-09 13:04:08 +0100831/*
832 * Don't initialise the fields below on every allocation, but do that in
833 * advance and keep them valid across allocations.
834 */
835static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
836{
837 req->ctx = ctx;
838 req->link = NULL;
839 req->async_data = NULL;
840 /* not necessary, but safer to zero */
Pavel Begunkovcef216f2022-04-12 15:09:43 +0100841 req->cqe.res = 0;
Pavel Begunkov864ea922021-08-09 13:04:08 +0100842}
843
Pavel Begunkovdac7a092021-03-19 17:22:39 +0000844static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
Pavel Begunkovcd0ca2e2021-08-09 20:18:11 +0100845 struct io_submit_state *state)
Pavel Begunkovdac7a092021-03-19 17:22:39 +0000846{
Jens Axboe79ebeae2021-08-10 15:18:27 -0600847 spin_lock(&ctx->completion_lock);
Pavel Begunkovc2b6c6b2021-09-24 21:59:47 +0100848 wq_list_splice(&ctx->locked_free_list, &state->free_list);
Pavel Begunkovd0acdee2021-05-16 22:58:12 +0100849 ctx->locked_free_nr = 0;
Jens Axboe79ebeae2021-08-10 15:18:27 -0600850 spin_unlock(&ctx->completion_lock);
Pavel Begunkovdac7a092021-03-19 17:22:39 +0000851}
852
Pavel Begunkov5d5901a2021-08-11 19:28:29 +0100853/*
854 * A request might get retired back into the request caches even before opcode
855 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
856 * Because of that, io_alloc_req() should be called only under ->uring_lock
857 * and with extra caution to not get a request that is still worked on.
858 */
Pavel Begunkovbd1a3782022-07-27 10:30:40 +0100859__cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
Pavel Begunkov5d5901a2021-08-11 19:28:29 +0100860 __must_hold(&ctx->uring_lock)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700861{
Pavel Begunkov864ea922021-08-09 13:04:08 +0100862 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
Pavel Begunkov3ab665b2021-09-24 21:59:45 +0100863 void *reqs[IO_REQ_ALLOC_BATCH];
Pavel Begunkov864ea922021-08-09 13:04:08 +0100864 int ret, i;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700865
Pavel Begunkov23a5c432022-04-12 15:09:46 +0100866 /*
867 * If we have more than a batch's worth of requests in our IRQ side
868 * locked cache, grab the lock and move them over to our submission
869 * side cache.
870 */
Pavel Begunkova6d97a82022-04-15 22:08:33 +0100871 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
Pavel Begunkov23a5c432022-04-12 15:09:46 +0100872 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
Pavel Begunkov88ab95b2022-04-12 15:09:47 +0100873 if (!io_req_cache_empty(ctx))
Pavel Begunkov23a5c432022-04-12 15:09:46 +0100874 return true;
875 }
Jens Axboe2b188cc2019-01-07 10:46:33 -0700876
Pavel Begunkov3ab665b2021-09-24 21:59:45 +0100877 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
Pavel Begunkove5d1bc02021-02-10 00:03:23 +0000878
Pavel Begunkov864ea922021-08-09 13:04:08 +0100879 /*
880 * Bulk alloc is all-or-nothing. If we fail to get a batch,
881 * retry single alloc to be on the safe side.
882 */
883 if (unlikely(ret <= 0)) {
Pavel Begunkov3ab665b2021-09-24 21:59:45 +0100884 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
885 if (!reqs[0])
Pavel Begunkova33ae9c2021-10-04 20:02:49 +0100886 return false;
Pavel Begunkov864ea922021-08-09 13:04:08 +0100887 ret = 1;
Jens Axboe2b188cc2019-01-07 10:46:33 -0700888 }
Pavel Begunkov864ea922021-08-09 13:04:08 +0100889
Pavel Begunkov37f0e762021-10-04 20:02:53 +0100890 percpu_ref_get_many(&ctx->refs, ret);
Pavel Begunkov3ab665b2021-09-24 21:59:45 +0100891 for (i = 0; i < ret; i++) {
Pavel Begunkov23a5c432022-04-12 15:09:46 +0100892 struct io_kiocb *req = reqs[i];
Pavel Begunkov3ab665b2021-09-24 21:59:45 +0100893
894 io_preinit_req(req, ctx);
Pavel Begunkovfa054572022-04-12 15:09:48 +0100895 io_req_add_to_cache(req, ctx);
Pavel Begunkov3ab665b2021-09-24 21:59:45 +0100896 }
Pavel Begunkova33ae9c2021-10-04 20:02:49 +0100897 return true;
898}
899
Pavel Begunkov6b639522021-09-08 16:40:50 +0100900static inline void io_dismantle_req(struct io_kiocb *req)
Jens Axboe2b188cc2019-01-07 10:46:33 -0700901{
Pavel Begunkov094bae42021-03-19 17:22:42 +0000902 unsigned int flags = req->flags;
Pavel Begunkov929a3af2020-02-19 00:19:09 +0300903
Pavel Begunkov867f8fa2021-10-04 20:02:58 +0100904 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
Pavel Begunkov3a0a6902021-04-20 12:03:31 +0100905 io_clean_op(req);
Pavel Begunkove1d767f2021-03-19 17:22:43 +0000906 if (!(flags & REQ_F_FIXED_FILE))
907 io_put_file(req->file);
Pavel Begunkove6543a82020-06-28 12:52:30 +0300908}
Pavel Begunkov2b85edf2019-12-28 14:13:03 +0300909
Jens Axboe599151432022-05-25 08:57:27 -0600910__cold void io_free_req(struct io_kiocb *req)
Pavel Begunkove6543a82020-06-28 12:52:30 +0300911{
Jens Axboe51a4cc12020-08-10 10:55:56 -0600912 struct io_ring_ctx *ctx = req->ctx;
Pavel Begunkovecfc5172020-06-29 13:13:03 +0300913
Pavel Begunkov7ac1edc2022-04-18 20:51:15 +0100914 io_req_put_rsrc(req);
Pavel Begunkov216578e2020-10-13 09:44:00 +0100915 io_dismantle_req(req);
Pavel Begunkov7c660732021-01-25 11:42:21 +0000916 io_put_task(req->task, 1);
Pavel Begunkove6543a82020-06-28 12:52:30 +0300917
Jens Axboe79ebeae2021-08-10 15:18:27 -0600918 spin_lock(&ctx->completion_lock);
Pavel Begunkovc2b6c6b2021-09-24 21:59:47 +0100919 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
Pavel Begunkovc34b0252021-08-09 20:18:08 +0100920 ctx->locked_free_nr++;
Jens Axboe79ebeae2021-08-10 15:18:27 -0600921 spin_unlock(&ctx->completion_lock);
Jens Axboee65ef562019-03-12 10:16:44 -0600922}
923
Pavel Begunkovd81499b2021-09-08 16:40:51 +0100924static void __io_req_find_next_prep(struct io_kiocb *req)
925{
926 struct io_ring_ctx *ctx = req->ctx;
Pavel Begunkovd81499b2021-09-08 16:40:51 +0100927
Pavel Begunkov253993212022-06-20 01:25:56 +0100928 io_cq_lock(ctx);
Pavel Begunkov305bef92022-06-20 01:25:55 +0100929 io_disarm_next(req);
Pavel Begunkov253993212022-06-20 01:25:56 +0100930 io_cq_unlock_post(ctx);
Pavel Begunkovd81499b2021-09-08 16:40:51 +0100931}
932
933static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
Jens Axboe9e645e112019-05-10 16:07:28 -0600934{
Pavel Begunkov33cc89a2021-03-09 00:37:58 +0000935 struct io_kiocb *nxt;
Jens Axboe2665abf2019-11-05 12:40:47 -0700936
Jens Axboe9e645e112019-05-10 16:07:28 -0600937 /*
938 * If LINK is set, we have dependent requests in this chain. If we
939 * didn't fail this request, queue the first one up, moving any other
940 * dependencies to the next request. In case of failure, fail the rest
941 * of the chain.
942 */
Pavel Begunkovd81499b2021-09-08 16:40:51 +0100943 if (unlikely(req->flags & IO_DISARM_MASK))
944 __io_req_find_next_prep(req);
Pavel Begunkov33cc89a2021-03-09 00:37:58 +0000945 nxt = req->link;
946 req->link = NULL;
947 return nxt;
Jens Axboe4d7dd462019-11-20 13:03:52 -0700948}
Jens Axboe2665abf2019-11-05 12:40:47 -0700949
Pavel Begunkovf237c302021-08-18 12:42:46 +0100950static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
Pavel Begunkov2c323952021-02-28 22:04:53 +0000951{
952 if (!ctx)
953 return;
Jens Axboeef060ea2022-04-25 19:49:04 -0600954 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
955 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
Pavel Begunkovf237c302021-08-18 12:42:46 +0100956 if (*locked) {
Pavel Begunkovc4501782021-09-08 16:40:52 +0100957 io_submit_flush_completions(ctx);
Pavel Begunkov2c323952021-02-28 22:04:53 +0000958 mutex_unlock(&ctx->uring_lock);
Pavel Begunkovf237c302021-08-18 12:42:46 +0100959 *locked = false;
Pavel Begunkov2c323952021-02-28 22:04:53 +0000960 }
961 percpu_ref_put(&ctx->refs);
962}
963
Dylan Yudakenc6dd7632022-06-22 06:40:28 -0700964static unsigned int handle_tw_list(struct llist_node *node,
965 struct io_ring_ctx **ctx, bool *locked,
966 struct llist_node *last)
Hao Xu9f8d0322021-12-07 17:39:49 +0800967{
Dylan Yudakenc6dd7632022-06-22 06:40:28 -0700968 unsigned int count = 0;
969
Dylan Yudaken3a0c0372022-06-22 06:40:25 -0700970 while (node != last) {
Dylan Yudakenf88262e2022-06-22 06:40:23 -0700971 struct llist_node *next = node->next;
Hao Xu9f8d0322021-12-07 17:39:49 +0800972 struct io_kiocb *req = container_of(node, struct io_kiocb,
973 io_task_work.node);
974
Jens Axboe34d2bfe2022-03-24 10:17:44 -0600975 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
976
Hao Xu9f8d0322021-12-07 17:39:49 +0800977 if (req->ctx != *ctx) {
978 ctx_flush_and_put(*ctx, locked);
979 *ctx = req->ctx;
980 /* if not contended, grab and improve batching */
981 *locked = mutex_trylock(&(*ctx)->uring_lock);
982 percpu_ref_get(&(*ctx)->refs);
983 }
984 req->io_task_work.func(req, locked);
985 node = next;
Dylan Yudakenc6dd7632022-06-22 06:40:28 -0700986 count++;
Dylan Yudaken3a0c0372022-06-22 06:40:25 -0700987 }
Dylan Yudakenc6dd7632022-06-22 06:40:28 -0700988
989 return count;
Hao Xu9f8d0322021-12-07 17:39:49 +0800990}
991
Dylan Yudaken923d1592022-06-22 06:40:24 -0700992/**
993 * io_llist_xchg - swap all entries in a lock-less list
994 * @head: the head of lock-less list to delete all entries
995 * @new: new entry as the head of the list
996 *
997 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
998 * The order of entries returned is from the newest to the oldest added one.
999 */
1000static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1001 struct llist_node *new)
1002{
1003 return xchg(&head->first, new);
1004}
1005
1006/**
1007 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1008 * @head: the head of lock-less list to delete all entries
1009 * @old: expected old value of the first entry of the list
1010 * @new: new entry as the head of the list
1011 *
1012 * perform a cmpxchg on the first entry of the list.
1013 */
1014
1015static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1016 struct llist_node *old,
1017 struct llist_node *new)
1018{
1019 return cmpxchg(&head->first, old, new);
1020}
1021
Jens Axboec9f06aa2022-05-25 11:01:04 -06001022void tctx_task_work(struct callback_head *cb)
Jens Axboe7cbf1722021-02-10 00:03:20 +00001023{
Hao Xuf28c240e2021-12-08 13:21:25 +08001024 bool uring_locked = false;
Pavel Begunkovebd0df22021-06-17 18:14:07 +01001025 struct io_ring_ctx *ctx = NULL;
Pavel Begunkov3f184072021-06-17 18:14:06 +01001026 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1027 task_work);
Dylan Yudaken3a0c0372022-06-22 06:40:25 -07001028 struct llist_node fake = {};
1029 struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake);
Dylan Yudakenc6dd7632022-06-22 06:40:28 -07001030 unsigned int loops = 1;
1031 unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL);
Jens Axboe7cbf1722021-02-10 00:03:20 +00001032
Dylan Yudaken3a0c0372022-06-22 06:40:25 -07001033 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1034 while (node != &fake) {
Dylan Yudakenc6dd7632022-06-22 06:40:28 -07001035 loops++;
Dylan Yudaken3a0c0372022-06-22 06:40:25 -07001036 node = io_llist_xchg(&tctx->task_list, &fake);
Dylan Yudakenc6dd7632022-06-22 06:40:28 -07001037 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
Dylan Yudaken3a0c0372022-06-22 06:40:25 -07001038 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
Pavel Begunkov3f184072021-06-17 18:14:06 +01001039 }
Pavel Begunkovebd0df22021-06-17 18:14:07 +01001040
Hao Xuf28c240e2021-12-08 13:21:25 +08001041 ctx_flush_and_put(ctx, &uring_locked);
Pavel Begunkov3cc7fdb2022-01-09 00:53:22 +00001042
1043 /* relaxed read is enough as only the task itself sets ->in_idle */
1044 if (unlikely(atomic_read(&tctx->in_idle)))
1045 io_uring_drop_tctx_refs(current);
Dylan Yudakenc6dd7632022-06-22 06:40:28 -07001046
1047 trace_io_uring_task_work_run(tctx, count, loops);
Jens Axboe7cbf1722021-02-10 00:03:20 +00001048}
1049
Dylan Yudakenc34398a2022-06-22 06:40:22 -07001050void io_req_task_work_add(struct io_kiocb *req)
Jens Axboe7cbf1722021-02-10 00:03:20 +00001051{
Dylan Yudakenc34398a2022-06-22 06:40:22 -07001052 struct io_uring_task *tctx = req->task->io_uring;
Jens Axboe9f010502022-04-25 19:49:02 -06001053 struct io_ring_ctx *ctx = req->ctx;
Dylan Yudakenf88262e2022-06-22 06:40:23 -07001054 struct llist_node *node;
Pavel Begunkov6294f362021-08-10 17:53:55 +01001055 bool running;
Jens Axboe7cbf1722021-02-10 00:03:20 +00001056
Dylan Yudakenf88262e2022-06-22 06:40:23 -07001057 running = !llist_add(&req->io_task_work.node, &tctx->task_list);
Jens Axboe7cbf1722021-02-10 00:03:20 +00001058
1059 /* task_work already pending, we're done */
Pavel Begunkov6294f362021-08-10 17:53:55 +01001060 if (running)
Pavel Begunkove09ee512021-07-01 13:26:05 +01001061 return;
Jens Axboe7cbf1722021-02-10 00:03:20 +00001062
Jens Axboeef060ea2022-04-25 19:49:04 -06001063 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1064 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1065
Jens Axboe3fe07bc2022-05-21 09:17:05 -06001066 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
Pavel Begunkove09ee512021-07-01 13:26:05 +01001067 return;
Pavel Begunkov2215bed2021-08-09 13:04:06 +01001068
Dylan Yudakenf88262e2022-06-22 06:40:23 -07001069 node = llist_del_all(&tctx->task_list);
Jens Axboe7cbf1722021-02-10 00:03:20 +00001070
Pavel Begunkove09ee512021-07-01 13:26:05 +01001071 while (node) {
1072 req = container_of(node, struct io_kiocb, io_task_work.node);
1073 node = node->next;
Pavel Begunkov3218e5d2022-06-25 11:52:59 +01001074 if (llist_add(&req->io_task_work.node,
Pavel Begunkove09ee512021-07-01 13:26:05 +01001075 &req->ctx->fallback_llist))
1076 schedule_delayed_work(&req->ctx->fallback_work, 1);
1077 }
Pavel Begunkoveab30c42021-01-19 13:32:42 +00001078}
1079
Pavel Begunkov4e118cd2022-04-15 22:08:23 +01001080static void io_req_tw_post(struct io_kiocb *req, bool *locked)
1081{
Jens Axboe97b388d2022-05-24 15:21:00 -06001082 io_req_complete_post(req);
Pavel Begunkov4e118cd2022-04-15 22:08:23 +01001083}
1084
Jens Axboe599151432022-05-25 08:57:27 -06001085void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags)
Pavel Begunkov4e118cd2022-04-15 22:08:23 +01001086{
Jens Axboe97b388d2022-05-24 15:21:00 -06001087 io_req_set_res(req, res, cflags);
Pavel Begunkov4e118cd2022-04-15 22:08:23 +01001088 req->io_task_work.func = io_req_tw_post;
Jens Axboe3fe07bc2022-05-21 09:17:05 -06001089 io_req_task_work_add(req);
Pavel Begunkov4e118cd2022-04-15 22:08:23 +01001090}
1091
Pavel Begunkovf237c302021-08-18 12:42:46 +01001092static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
Jens Axboec40f6372020-06-25 15:39:59 -06001093{
Pavel Begunkovb18a1a42021-08-25 20:51:39 +01001094 /* not needed for normal modes, but SQPOLL depends on it */
Pavel Begunkov971cf9c2022-04-15 22:08:22 +01001095 io_tw_lock(req->ctx, locked);
Pavel Begunkovcef216f2022-04-12 15:09:43 +01001096 io_req_complete_failed(req, req->cqe.res);
Jens Axboec40f6372020-06-25 15:39:59 -06001097}
1098
Jens Axboe329061d2022-05-25 20:31:09 -06001099void io_req_task_submit(struct io_kiocb *req, bool *locked)
Jens Axboec40f6372020-06-25 15:39:59 -06001100{
Pavel Begunkov971cf9c2022-04-15 22:08:22 +01001101 io_tw_lock(req->ctx, locked);
Jens Axboe316319e2021-08-19 09:41:42 -06001102 /* req->task == current here, checking PF_EXITING is safe */
Pavel Begunkovaf066f32021-08-09 13:04:19 +01001103 if (likely(!(req->task->flags & PF_EXITING)))
Pavel Begunkovcbc2e202022-04-15 22:08:26 +01001104 io_queue_sqe(req);
Pavel Begunkov81b6d052021-01-04 20:36:35 +00001105 else
Pavel Begunkov25935532021-03-19 17:22:40 +00001106 io_req_complete_failed(req, -EFAULT);
Jens Axboe9e645e112019-05-10 16:07:28 -06001107}
1108
Jens Axboe599151432022-05-25 08:57:27 -06001109void io_req_task_queue_fail(struct io_kiocb *req, int ret)
Pavel Begunkova3df76982021-02-18 22:32:52 +00001110{
Jens Axboe97b388d2022-05-24 15:21:00 -06001111 io_req_set_res(req, ret, 0);
Pavel Begunkov5b0a6ac2021-06-30 21:54:04 +01001112 req->io_task_work.func = io_req_task_cancel;
Jens Axboe3fe07bc2022-05-21 09:17:05 -06001113 io_req_task_work_add(req);
Pavel Begunkova3df76982021-02-18 22:32:52 +00001114}
1115
Jens Axboef3b44f92022-06-13 07:27:03 -06001116void io_req_task_queue(struct io_kiocb *req)
Pavel Begunkov2c4b8eb2021-02-28 22:35:10 +00001117{
Pavel Begunkov5b0a6ac2021-06-30 21:54:04 +01001118 req->io_task_work.func = io_req_task_submit;
Jens Axboe3fe07bc2022-05-21 09:17:05 -06001119 io_req_task_work_add(req);
Pavel Begunkov2c4b8eb2021-02-28 22:35:10 +00001120}
1121
Jens Axboe599151432022-05-25 08:57:27 -06001122void io_queue_next(struct io_kiocb *req)
Jackie Liuc69f8db2019-11-09 11:00:08 +08001123{
Pavel Begunkov9b5f7bd92020-06-29 13:13:00 +03001124 struct io_kiocb *nxt = io_req_find_next(req);
Pavel Begunkov944e58b2019-11-21 23:21:01 +03001125
Pavel Begunkov906a8c32020-06-27 14:04:55 +03001126 if (nxt)
1127 io_req_task_queue(nxt);
Jackie Liuc69f8db2019-11-09 11:00:08 +08001128}
1129
Jens Axboef3b44f92022-06-13 07:27:03 -06001130void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
Jens Axboea141dd82021-08-12 12:48:34 -06001131 __must_hold(&ctx->uring_lock)
Pavel Begunkov905c1722021-02-10 00:03:14 +00001132{
Pavel Begunkovd4b7a5e2021-09-24 21:59:53 +01001133 struct task_struct *task = NULL;
Pavel Begunkov37f0e762021-10-04 20:02:53 +01001134 int task_refs = 0;
Pavel Begunkov3aa83bf2021-09-24 21:59:50 +01001135
Pavel Begunkov3aa83bf2021-09-24 21:59:50 +01001136 do {
1137 struct io_kiocb *req = container_of(node, struct io_kiocb,
1138 comp_list);
1139
Pavel Begunkova538be52022-03-21 22:02:22 +00001140 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1141 if (req->flags & REQ_F_REFCOUNT) {
1142 node = req->comp_list.next;
1143 if (!req_ref_put_and_test(req))
1144 continue;
1145 }
Pavel Begunkovb605a7fa2022-03-21 22:02:23 +00001146 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1147 struct async_poll *apoll = req->apoll;
1148
1149 if (apoll->double_poll)
1150 kfree(apoll->double_poll);
Jens Axboe9731bc982022-07-07 14:20:54 -06001151 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1152 kfree(apoll);
Pavel Begunkovb605a7fa2022-03-21 22:02:23 +00001153 req->flags &= ~REQ_F_POLLED;
1154 }
Pavel Begunkovda1a08c2022-04-15 22:08:29 +01001155 if (req->flags & IO_REQ_LINK_FLAGS)
Pavel Begunkov57859f42022-03-21 22:02:24 +00001156 io_queue_next(req);
Pavel Begunkova538be52022-03-21 22:02:22 +00001157 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1158 io_clean_op(req);
Pavel Begunkovc1e53a62021-10-04 20:02:55 +01001159 }
Pavel Begunkova538be52022-03-21 22:02:22 +00001160 if (!(req->flags & REQ_F_FIXED_FILE))
1161 io_put_file(req->file);
Pavel Begunkovd4b7a5e2021-09-24 21:59:53 +01001162
Pavel Begunkovab409402021-10-09 23:14:41 +01001163 io_req_put_rsrc_locked(req, ctx);
Pavel Begunkovd4b7a5e2021-09-24 21:59:53 +01001164
1165 if (req->task != task) {
1166 if (task)
1167 io_put_task(task, task_refs);
1168 task = req->task;
1169 task_refs = 0;
1170 }
1171 task_refs++;
Pavel Begunkovc1e53a62021-10-04 20:02:55 +01001172 node = req->comp_list.next;
Pavel Begunkovfa054572022-04-12 15:09:48 +01001173 io_req_add_to_cache(req, ctx);
Pavel Begunkov3aa83bf2021-09-24 21:59:50 +01001174 } while (node);
Pavel Begunkovd4b7a5e2021-09-24 21:59:53 +01001175
Pavel Begunkovd4b7a5e2021-09-24 21:59:53 +01001176 if (task)
1177 io_put_task(task, task_refs);
Pavel Begunkov3aa83bf2021-09-24 21:59:50 +01001178}
1179
Pavel Begunkovc4501782021-09-08 16:40:52 +01001180static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
Pavel Begunkov905c1722021-02-10 00:03:14 +00001181 __must_hold(&ctx->uring_lock)
1182{
Pavel Begunkov6f33b0b2021-09-24 21:59:44 +01001183 struct io_wq_work_node *node, *prev;
Pavel Begunkovcd0ca2e2021-08-09 20:18:11 +01001184 struct io_submit_state *state = &ctx->submit_state;
Pavel Begunkov905c1722021-02-10 00:03:14 +00001185
Pavel Begunkovd9dee432022-06-19 12:26:08 +01001186 spin_lock(&ctx->completion_lock);
1187 wq_list_for_each(node, prev, &state->compl_reqs) {
1188 struct io_kiocb *req = container_of(node, struct io_kiocb,
1189 comp_list);
Pavel Begunkov5182ed22021-06-26 21:40:48 +01001190
Pavel Begunkovd9dee432022-06-19 12:26:08 +01001191 if (!(req->flags & REQ_F_CQE_SKIP))
1192 __io_fill_cqe_req(ctx, req);
Pavel Begunkov905c1722021-02-10 00:03:14 +00001193 }
Pavel Begunkov253993212022-06-20 01:25:56 +01001194 __io_cq_unlock_post(ctx);
Pavel Begunkovd9dee432022-06-19 12:26:08 +01001195
Pavel Begunkov1cce17a2021-09-24 21:59:54 +01001196 io_free_batch_list(ctx, state->compl_reqs.first);
Pavel Begunkov6f33b0b2021-09-24 21:59:44 +01001197 INIT_WQ_LIST(&state->compl_reqs);
Pavel Begunkov7a743e22020-03-03 21:33:13 +03001198}
1199
Jens Axboeba816ad2019-09-28 11:36:45 -06001200/*
1201 * Drop reference to request, return next in chain (if there is one) if this
1202 * was the last reference to this request.
1203 */
Pavel Begunkov0d850352021-03-19 17:22:37 +00001204static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
Jens Axboee65ef562019-03-12 10:16:44 -06001205{
Pavel Begunkov9b5f7bd92020-06-29 13:13:00 +03001206 struct io_kiocb *nxt = NULL;
1207
Jens Axboede9b4cc2021-02-24 13:28:27 -07001208 if (req_ref_put_and_test(req)) {
Pavel Begunkovda1a08c2022-04-15 22:08:29 +01001209 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
Pavel Begunkov7819a1f2022-03-21 22:02:21 +00001210 nxt = io_req_find_next(req);
Pavel Begunkovf5c6cf22022-04-15 22:08:24 +01001211 io_free_req(req);
Jens Axboe2a44f462020-02-25 13:25:41 -07001212 }
Pavel Begunkov9b5f7bd92020-06-29 13:13:00 +03001213 return nxt;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001214}
1215
Pavel Begunkov6c503152021-01-04 20:36:36 +00001216static unsigned io_cqring_events(struct io_ring_ctx *ctx)
Jens Axboea3a0e432019-08-20 11:03:11 -06001217{
1218 /* See comment at the top of this file */
1219 smp_rmb();
Pavel Begunkove23de152020-12-17 00:24:37 +00001220 return __io_cqring_events(ctx);
Jens Axboea3a0e432019-08-20 11:03:11 -06001221}
1222
Jens Axboedef596e2019-01-09 08:59:42 -07001223/*
Jens Axboedef596e2019-01-09 08:59:42 -07001224 * We can't just wait for polled events to come to us, we have to actively
1225 * find and complete them.
1226 */
Pavel Begunkovc0724812021-10-04 20:02:54 +01001227static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
Jens Axboedef596e2019-01-09 08:59:42 -07001228{
1229 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1230 return;
1231
1232 mutex_lock(&ctx->uring_lock);
Pavel Begunkov5eef4e82021-09-24 21:59:49 +01001233 while (!wq_list_empty(&ctx->iopoll_list)) {
Pavel Begunkovb2edc0a2020-07-07 16:36:22 +03001234 /* let it sleep and repeat later if can't complete a request */
Pavel Begunkov5ba3c872021-09-24 21:59:43 +01001235 if (io_do_iopoll(ctx, true) == 0)
Pavel Begunkovb2edc0a2020-07-07 16:36:22 +03001236 break;
Jens Axboe08f54392019-08-21 22:19:11 -06001237 /*
1238 * Ensure we allow local-to-the-cpu processing to take place,
1239 * in this case we need to ensure that we reap all events.
Pavel Begunkov3fcee5a2020-07-06 17:59:31 +03001240 * Also let task_work, etc. to progress by releasing the mutex
Jens Axboe08f54392019-08-21 22:19:11 -06001241 */
Pavel Begunkov3fcee5a2020-07-06 17:59:31 +03001242 if (need_resched()) {
1243 mutex_unlock(&ctx->uring_lock);
1244 cond_resched();
1245 mutex_lock(&ctx->uring_lock);
1246 }
Jens Axboedef596e2019-01-09 08:59:42 -07001247 }
1248 mutex_unlock(&ctx->uring_lock);
1249}
1250
Pavel Begunkov7668b922020-07-07 16:36:21 +03001251static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
Jens Axboedef596e2019-01-09 08:59:42 -07001252{
Pavel Begunkov7668b922020-07-07 16:36:21 +03001253 unsigned int nr_events = 0;
Pavel Begunkove9979b32021-04-13 02:58:45 +01001254 int ret = 0;
Dylan Yudaken155bc952022-04-21 02:13:44 -07001255 unsigned long check_cq;
Jens Axboedef596e2019-01-09 08:59:42 -07001256
Pavel Begunkov3a085762022-06-15 17:33:55 +01001257 check_cq = READ_ONCE(ctx->check_cq);
1258 if (unlikely(check_cq)) {
1259 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1260 __io_cqring_overflow_flush(ctx, false);
1261 /*
1262 * Similarly do not spin if we have not informed the user of any
1263 * dropped CQE.
1264 */
1265 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1266 return -EBADR;
1267 }
Xiaoguang Wangc7849be2020-02-22 14:46:05 +08001268 /*
Pavel Begunkovf39c8a52021-04-13 02:58:46 +01001269 * Don't enter poll loop if we already have events pending.
1270 * If we do, we can potentially be spinning for commands that
1271 * already triggered a CQE (eg in error).
1272 */
Pavel Begunkovf39c8a52021-04-13 02:58:46 +01001273 if (io_cqring_events(ctx))
Pavel Begunkovd487b432022-03-22 14:07:58 +00001274 return 0;
Dylan Yudaken155bc952022-04-21 02:13:44 -07001275
Jens Axboedef596e2019-01-09 08:59:42 -07001276 do {
Jens Axboe500f9fb2019-08-19 12:15:59 -06001277 /*
1278 * If a submit got punted to a workqueue, we can have the
1279 * application entering polling for a command before it gets
1280 * issued. That app will hold the uring_lock for the duration
1281 * of the poll right here, so we need to take a breather every
1282 * now and then to ensure that the issue has a chance to add
1283 * the poll to the issued list. Otherwise we can spin here
1284 * forever, while the workqueue is stuck trying to acquire the
1285 * very same mutex.
1286 */
Pavel Begunkov5eef4e82021-09-24 21:59:49 +01001287 if (wq_list_empty(&ctx->iopoll_list)) {
Pavel Begunkov8f487ef2021-07-08 13:37:06 +01001288 u32 tail = ctx->cached_cq_tail;
1289
Jens Axboe500f9fb2019-08-19 12:15:59 -06001290 mutex_unlock(&ctx->uring_lock);
Jens Axboe4c6e2772020-07-01 11:29:10 -06001291 io_run_task_work();
Jens Axboe500f9fb2019-08-19 12:15:59 -06001292 mutex_lock(&ctx->uring_lock);
Pavel Begunkove9979b32021-04-13 02:58:45 +01001293
Pavel Begunkov8f487ef2021-07-08 13:37:06 +01001294 /* some requests don't go through iopoll_list */
1295 if (tail != ctx->cached_cq_tail ||
Pavel Begunkov5eef4e82021-09-24 21:59:49 +01001296 wq_list_empty(&ctx->iopoll_list))
Pavel Begunkove9979b32021-04-13 02:58:45 +01001297 break;
Jens Axboe500f9fb2019-08-19 12:15:59 -06001298 }
Pavel Begunkov5ba3c872021-09-24 21:59:43 +01001299 ret = io_do_iopoll(ctx, !min);
1300 if (ret < 0)
1301 break;
1302 nr_events += ret;
1303 ret = 0;
1304 } while (nr_events < min && !need_resched());
Pavel Begunkovd487b432022-03-22 14:07:58 +00001305
Jens Axboedef596e2019-01-09 08:59:42 -07001306 return ret;
1307}
Pavel Begunkov7012c812022-06-16 10:21:59 +01001308
1309void io_req_task_complete(struct io_kiocb *req, bool *locked)
Jens Axboe8ef12ef2021-08-10 15:15:25 -06001310{
Pavel Begunkov7012c812022-06-16 10:21:59 +01001311 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1312 unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1313
1314 req->cqe.flags |= io_put_kbuf(req, issue_flags);
Pavel Begunkov126180b2021-08-18 12:42:47 +01001315 }
Pavel Begunkov7012c812022-06-16 10:21:59 +01001316
1317 if (*locked)
Pavel Begunkov9da070b2022-06-20 01:26:00 +01001318 io_req_complete_defer(req);
Pavel Begunkov7012c812022-06-16 10:21:59 +01001319 else
1320 io_req_complete_post(req);
Jens Axboe8ef12ef2021-08-10 15:15:25 -06001321}
1322
Jens Axboedef596e2019-01-09 08:59:42 -07001323/*
1324 * After the iocb has been issued, it's safe to be found on the poll list.
1325 * Adding the kiocb to the list AFTER submission ensures that we don't
Pavel Begunkovf39c8a52021-04-13 02:58:46 +01001326 * find it from a io_do_iopoll() thread before the issuer is done
Jens Axboedef596e2019-01-09 08:59:42 -07001327 * accessing the kiocb cookie.
1328 */
Pavel Begunkov98821312021-10-15 17:09:12 +01001329static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
Jens Axboedef596e2019-01-09 08:59:42 -07001330{
1331 struct io_ring_ctx *ctx = req->ctx;
Hao Xu3b44b372021-10-18 21:34:31 +08001332 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
Pavel Begunkovcb3d8972021-06-14 02:36:14 +01001333
1334 /* workqueue context doesn't hold uring_lock, grab it now */
Hao Xu3b44b372021-10-18 21:34:31 +08001335 if (unlikely(needs_lock))
Pavel Begunkovcb3d8972021-06-14 02:36:14 +01001336 mutex_lock(&ctx->uring_lock);
Jens Axboedef596e2019-01-09 08:59:42 -07001337
1338 /*
1339 * Track whether we have multiple files in our lists. This will impact
1340 * how we do polling eventually, not spinning if we're on potentially
1341 * different devices.
1342 */
Pavel Begunkov5eef4e82021-09-24 21:59:49 +01001343 if (wq_list_empty(&ctx->iopoll_list)) {
Hao Xu915b3dd2021-06-28 05:37:30 +08001344 ctx->poll_multi_queue = false;
1345 } else if (!ctx->poll_multi_queue) {
Jens Axboedef596e2019-01-09 08:59:42 -07001346 struct io_kiocb *list_req;
1347
Pavel Begunkov5eef4e82021-09-24 21:59:49 +01001348 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1349 comp_list);
Christoph Hellwig30da1b42021-10-12 13:12:14 +02001350 if (list_req->file != req->file)
Hao Xu915b3dd2021-06-28 05:37:30 +08001351 ctx->poll_multi_queue = true;
Jens Axboedef596e2019-01-09 08:59:42 -07001352 }
1353
1354 /*
1355 * For fast devices, IO may have already completed. If it has, add
1356 * it to the front so we find it first.
1357 */
Xiaoguang Wang65a65432020-06-11 23:39:36 +08001358 if (READ_ONCE(req->iopoll_completed))
Pavel Begunkov5eef4e82021-09-24 21:59:49 +01001359 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
Jens Axboedef596e2019-01-09 08:59:42 -07001360 else
Pavel Begunkov5eef4e82021-09-24 21:59:49 +01001361 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
Xiaoguang Wangbdcd3ea2020-02-25 22:12:08 +08001362
Hao Xu3b44b372021-10-18 21:34:31 +08001363 if (unlikely(needs_lock)) {
Pavel Begunkovcb3d8972021-06-14 02:36:14 +01001364 /*
1365 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1366 * in sq thread task context or in io worker task context. If
1367 * current task context is sq thread, we don't need to check
1368 * whether should wake up sq thread.
1369 */
1370 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1371 wq_has_sleeper(&ctx->sq_data->wait))
1372 wake_up(&ctx->sq_data->wait);
1373
1374 mutex_unlock(&ctx->uring_lock);
1375 }
Jens Axboedef596e2019-01-09 08:59:42 -07001376}
1377
Jens Axboe4503b762020-06-01 10:00:27 -06001378static bool io_bdev_nowait(struct block_device *bdev)
1379{
Jeffle Xu9ba0d0c2020-10-19 16:59:42 +08001380 return !bdev || blk_queue_nowait(bdev_get_queue(bdev));
Jens Axboe4503b762020-06-01 10:00:27 -06001381}
1382
Jens Axboe2b188cc2019-01-07 10:46:33 -07001383/*
1384 * If we tracked the file through the SCM inflight mechanism, we could support
1385 * any file. For now, just ensure that anything potentially problematic is done
1386 * inline.
1387 */
Pavel Begunkov88459b52021-10-17 00:07:10 +01001388static bool __io_file_supports_nowait(struct file *file, umode_t mode)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001389{
Jens Axboe4503b762020-06-01 10:00:27 -06001390 if (S_ISBLK(mode)) {
Christoph Hellwig4e7b5672020-11-23 13:38:40 +01001391 if (IS_ENABLED(CONFIG_BLOCK) &&
1392 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
Jens Axboe4503b762020-06-01 10:00:27 -06001393 return true;
1394 return false;
1395 }
Pavel Begunkov976517f2021-06-09 12:07:25 +01001396 if (S_ISSOCK(mode))
Jens Axboe2b188cc2019-01-07 10:46:33 -07001397 return true;
Jens Axboe4503b762020-06-01 10:00:27 -06001398 if (S_ISREG(mode)) {
Christoph Hellwig4e7b5672020-11-23 13:38:40 +01001399 if (IS_ENABLED(CONFIG_BLOCK) &&
1400 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
Jens Axboee5550a12022-05-25 10:28:04 -06001401 !io_is_uring_fops(file))
Jens Axboe4503b762020-06-01 10:00:27 -06001402 return true;
1403 return false;
1404 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07001405
Jens Axboec5b85622020-06-09 19:23:05 -06001406 /* any ->read/write should understand O_NONBLOCK */
1407 if (file->f_flags & O_NONBLOCK)
1408 return true;
Pavel Begunkov35645ac2021-10-17 00:07:09 +01001409 return file->f_mode & FMODE_NOWAIT;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001410}
1411
Pavel Begunkov88459b52021-10-17 00:07:10 +01001412/*
1413 * If we tracked the file through the SCM inflight mechanism, we could support
1414 * any file. For now, just ensure that anything potentially problematic is done
1415 * inline.
1416 */
Jens Axboea4ad4f72022-05-25 10:40:19 -06001417unsigned int io_file_get_flags(struct file *file)
Jens Axboe7b29f922021-03-12 08:30:14 -07001418{
Pavel Begunkov88459b52021-10-17 00:07:10 +01001419 umode_t mode = file_inode(file)->i_mode;
1420 unsigned int res = 0;
Jens Axboe7b29f922021-03-12 08:30:14 -07001421
Pavel Begunkov88459b52021-10-17 00:07:10 +01001422 if (S_ISREG(mode))
1423 res |= FFS_ISREG;
1424 if (__io_file_supports_nowait(file, mode))
1425 res |= FFS_NOWAIT;
Jens Axboe5e456902022-04-20 16:15:27 -06001426 if (io_file_need_scm(file))
1427 res |= FFS_SCM;
Pavel Begunkov88459b52021-10-17 00:07:10 +01001428 return res;
Jens Axboe7b29f922021-03-12 08:30:14 -07001429}
1430
Jens Axboe99f15d82022-05-25 05:59:19 -06001431bool io_alloc_async_data(struct io_kiocb *req)
Xiaoguang Wang3d9932a2020-03-27 15:36:52 +08001432{
Jens Axboee8c2bc12020-08-15 18:44:09 -07001433 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1434 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
Pavel Begunkovd886e182021-10-04 20:02:56 +01001435 if (req->async_data) {
1436 req->flags |= REQ_F_ASYNC_DATA;
1437 return false;
1438 }
1439 return true;
Xiaoguang Wang3d9932a2020-03-27 15:36:52 +08001440}
1441
Jens Axboef3b44f92022-06-13 07:27:03 -06001442int io_req_prep_async(struct io_kiocb *req)
Jens Axboef67676d2019-12-02 11:03:47 -07001443{
Jens Axboea196c78b2022-05-01 21:19:50 -06001444 const struct io_op_def *def = &io_op_defs[req->opcode];
1445
1446 /* assign early for deferred execution for non-fixed file */
1447 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
Linus Torvalds3a166bd2022-05-23 12:22:49 -07001448 req->file = io_file_get_normal(req, req->cqe.fd);
Jens Axboedc919ca2022-05-23 17:30:37 -06001449 if (!def->prep_async)
Pavel Begunkovb7e298d2021-02-28 22:35:19 +00001450 return 0;
Pavel Begunkovd886e182021-10-04 20:02:56 +01001451 if (WARN_ON_ONCE(req_has_async_data(req)))
Pavel Begunkovb7e298d2021-02-28 22:35:19 +00001452 return -EFAULT;
1453 if (io_alloc_async_data(req))
1454 return -EAGAIN;
1455
Jens Axboedc919ca2022-05-23 17:30:37 -06001456 return def->prep_async(req);
Jens Axboef67676d2019-12-02 11:03:47 -07001457}
1458
Pavel Begunkov9cf7c102020-07-13 23:37:15 +03001459static u32 io_get_sequence(struct io_kiocb *req)
1460{
Pavel Begunkova3dbdf52021-06-17 18:14:05 +01001461 u32 seq = req->ctx->cached_sq_head;
Pavel Begunkov963c6ab2022-03-25 11:52:16 +00001462 struct io_kiocb *cur;
Pavel Begunkov9cf7c102020-07-13 23:37:15 +03001463
Pavel Begunkova3dbdf52021-06-17 18:14:05 +01001464 /* need original cached_sq_head, but it was increased for each req */
Pavel Begunkov963c6ab2022-03-25 11:52:16 +00001465 io_for_each_link(cur, req)
Pavel Begunkova3dbdf52021-06-17 18:14:05 +01001466 seq--;
1467 return seq;
Pavel Begunkov9cf7c102020-07-13 23:37:15 +03001468}
1469
Pavel Begunkovc0724812021-10-04 20:02:54 +01001470static __cold void io_drain_req(struct io_kiocb *req)
Jens Axboede0617e2019-04-06 21:51:27 -06001471{
Jackie Liua197f662019-11-08 08:09:12 -07001472 struct io_ring_ctx *ctx = req->ctx;
Pavel Begunkov27dc8332020-07-13 23:37:14 +03001473 struct io_defer_entry *de;
Jens Axboef67676d2019-12-02 11:03:47 -07001474 int ret;
Pavel Begunkove0eb71d2021-10-01 18:07:01 +01001475 u32 seq = io_get_sequence(req);
Pavel Begunkov3c199662021-06-15 16:47:57 +01001476
Bob Liu9d858b22019-11-13 18:06:25 +08001477 /* Still need defer if there is pending req in defer list. */
Hao Xue302f102021-11-25 17:21:02 +08001478 spin_lock(&ctx->completion_lock);
Pavel Begunkov5e371262021-09-24 22:00:04 +01001479 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
Hao Xue302f102021-11-25 17:21:02 +08001480 spin_unlock(&ctx->completion_lock);
Pavel Begunkove0eb71d2021-10-01 18:07:01 +01001481queue:
Pavel Begunkov10c66902021-06-15 16:47:56 +01001482 ctx->drain_active = false;
Pavel Begunkove0eb71d2021-10-01 18:07:01 +01001483 io_req_task_queue(req);
1484 return;
Pavel Begunkov10c66902021-06-15 16:47:56 +01001485 }
Hao Xue302f102021-11-25 17:21:02 +08001486 spin_unlock(&ctx->completion_lock);
Pavel Begunkov9cf7c102020-07-13 23:37:15 +03001487
Pavel Begunkovb7e298d2021-02-28 22:35:19 +00001488 ret = io_req_prep_async(req);
Pavel Begunkove0eb71d2021-10-01 18:07:01 +01001489 if (ret) {
1490fail:
1491 io_req_complete_failed(req, ret);
1492 return;
1493 }
Pavel Begunkovcbdcb432020-06-29 19:18:43 +03001494 io_prep_async_link(req);
Pavel Begunkov27dc8332020-07-13 23:37:14 +03001495 de = kmalloc(sizeof(*de), GFP_KERNEL);
Pavel Begunkov76cc33d2021-06-14 23:37:30 +01001496 if (!de) {
Pavel Begunkov1b487732021-07-11 22:41:13 +01001497 ret = -ENOMEM;
Pavel Begunkove0eb71d2021-10-01 18:07:01 +01001498 goto fail;
Pavel Begunkov76cc33d2021-06-14 23:37:30 +01001499 }
Jens Axboe2d283902019-12-04 11:08:05 -07001500
Jens Axboe79ebeae2021-08-10 15:18:27 -06001501 spin_lock(&ctx->completion_lock);
Pavel Begunkov9cf7c102020-07-13 23:37:15 +03001502 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
Jens Axboe79ebeae2021-08-10 15:18:27 -06001503 spin_unlock(&ctx->completion_lock);
Pavel Begunkov27dc8332020-07-13 23:37:14 +03001504 kfree(de);
Pavel Begunkove0eb71d2021-10-01 18:07:01 +01001505 goto queue;
Jens Axboede0617e2019-04-06 21:51:27 -06001506 }
1507
Pavel Begunkov48863ff2022-06-16 13:57:20 +01001508 trace_io_uring_defer(req);
Pavel Begunkov27dc8332020-07-13 23:37:14 +03001509 de->req = req;
Pavel Begunkov9cf7c102020-07-13 23:37:15 +03001510 de->seq = seq;
Pavel Begunkov27dc8332020-07-13 23:37:14 +03001511 list_add_tail(&de->list, &ctx->defer_list);
Jens Axboe79ebeae2021-08-10 15:18:27 -06001512 spin_unlock(&ctx->completion_lock);
Jens Axboede0617e2019-04-06 21:51:27 -06001513}
1514
Pavel Begunkov68fb8972021-03-19 17:22:41 +00001515static void io_clean_op(struct io_kiocb *req)
Pavel Begunkov99bc4c32020-02-07 22:04:45 +03001516{
Pavel Begunkov8197b052022-03-25 13:00:43 +00001517 if (req->flags & REQ_F_BUFFER_SELECTED) {
1518 spin_lock(&req->ctx->completion_lock);
Jens Axboecc3cec82022-03-08 17:46:52 -07001519 io_put_kbuf_comp(req);
Pavel Begunkov8197b052022-03-25 13:00:43 +00001520 spin_unlock(&req->ctx->completion_lock);
1521 }
Pavel Begunkov99bc4c32020-02-07 22:04:45 +03001522
Pavel Begunkov0e1b6fe32020-07-16 23:28:02 +03001523 if (req->flags & REQ_F_NEED_CLEANUP) {
Jens Axboe4d4c9cf2022-05-24 10:26:28 -06001524 const struct io_op_def *def = &io_op_defs[req->opcode];
Pavel Begunkov1dacb4d2021-06-17 18:14:03 +01001525
Jens Axboe4d4c9cf2022-05-24 10:26:28 -06001526 if (def->cleanup)
1527 def->cleanup(req);
Pavel Begunkov0e1b6fe32020-07-16 23:28:02 +03001528 }
Jens Axboe75652a302021-04-15 09:52:40 -06001529 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1530 kfree(req->apoll->double_poll);
1531 kfree(req->apoll);
1532 req->apoll = NULL;
1533 }
Jens Axboe9cae36a2022-06-01 23:57:02 -06001534 if (req->flags & REQ_F_INFLIGHT) {
1535 struct io_uring_task *tctx = req->task->io_uring;
1536
1537 atomic_dec(&tctx->inflight_tracked);
1538 }
Pavel Begunkovc8543572021-06-17 18:14:04 +01001539 if (req->flags & REQ_F_CREDS)
Pavel Begunkovb8e64b52021-06-17 18:14:02 +01001540 put_cred(req->creds);
Pavel Begunkovd886e182021-10-04 20:02:56 +01001541 if (req->flags & REQ_F_ASYNC_DATA) {
1542 kfree(req->async_data);
1543 req->async_data = NULL;
1544 }
Pavel Begunkovc8543572021-06-17 18:14:04 +01001545 req->flags &= ~IO_REQ_CLEAN_FLAGS;
Pavel Begunkov99bc4c32020-02-07 22:04:45 +03001546}
1547
Jens Axboe6bf9c472022-03-29 10:10:08 -06001548static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1549{
1550 if (req->file || !io_op_defs[req->opcode].needs_file)
1551 return true;
1552
1553 if (req->flags & REQ_F_FIXED_FILE)
Pavel Begunkovcef216f2022-04-12 15:09:43 +01001554 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
Jens Axboe6bf9c472022-03-29 10:10:08 -06001555 else
Pavel Begunkovcef216f2022-04-12 15:09:43 +01001556 req->file = io_file_get_normal(req, req->cqe.fd);
Jens Axboe6bf9c472022-03-29 10:10:08 -06001557
Pavel Begunkov772f5e02022-04-18 20:51:12 +01001558 return !!req->file;
Jens Axboe6bf9c472022-03-29 10:10:08 -06001559}
1560
Pavel Begunkov889fca72021-02-10 00:03:09 +00001561static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001562{
Jens Axboefcde59f2022-05-23 16:53:15 -06001563 const struct io_op_def *def = &io_op_defs[req->opcode];
Jens Axboe5730b272021-02-27 15:57:30 -07001564 const struct cred *creds = NULL;
Jens Axboed625c6e2019-12-17 19:53:05 -07001565 int ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001566
Jens Axboe70152142022-04-14 20:23:40 -06001567 if (unlikely(!io_assign_file(req, issue_flags)))
1568 return -EBADF;
1569
Pavel Begunkov6878b402021-09-24 21:59:41 +01001570 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
Pavel Begunkovc10d1f92021-06-17 18:14:01 +01001571 creds = override_creds(req->creds);
Jens Axboe5730b272021-02-27 15:57:30 -07001572
Jens Axboefcde59f2022-05-23 16:53:15 -06001573 if (!def->audit_skip)
Paul Moore5bd21822021-02-16 19:46:48 -05001574 audit_uring_entry(req->opcode);
1575
Jens Axboe0702e532022-05-23 16:56:21 -06001576 ret = def->issue(req, issue_flags);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001577
Jens Axboefcde59f2022-05-23 16:53:15 -06001578 if (!def->audit_skip)
Paul Moore5bd21822021-02-16 19:46:48 -05001579 audit_uring_exit(!ret, ret);
1580
Jens Axboe5730b272021-02-27 15:57:30 -07001581 if (creds)
1582 revert_creds(creds);
Jens Axboe97b388d2022-05-24 15:21:00 -06001583
Pavel Begunkov75d7b3a2022-06-16 10:21:58 +01001584 if (ret == IOU_OK) {
1585 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
Pavel Begunkov9da070b2022-06-20 01:26:00 +01001586 io_req_complete_defer(req);
Pavel Begunkov75d7b3a2022-06-16 10:21:58 +01001587 else
1588 io_req_complete_post(req);
1589 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
Jens Axboedef596e2019-01-09 08:59:42 -07001590 return ret;
Jens Axboe97b388d2022-05-24 15:21:00 -06001591
Jens Axboeb5325762020-05-19 21:20:27 -06001592 /* If the op doesn't have a file, we're not polling for it */
Pavel Begunkov99830282021-10-15 17:09:11 +01001593 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
Pavel Begunkov98821312021-10-15 17:09:12 +01001594 io_iopoll_req_issued(req, issue_flags);
Jens Axboedef596e2019-01-09 08:59:42 -07001595
1596 return 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001597}
1598
Jens Axboe329061d2022-05-25 20:31:09 -06001599int io_poll_issue(struct io_kiocb *req, bool *locked)
1600{
1601 io_tw_lock(req->ctx, locked);
1602 if (unlikely(req->task->flags & PF_EXITING))
1603 return -EFAULT;
Pavel Begunkovaeaa72c2022-06-15 17:33:54 +01001604 return io_issue_sqe(req, IO_URING_F_NONBLOCK);
Jens Axboe329061d2022-05-25 20:31:09 -06001605}
1606
Jens Axboec9f06aa2022-05-25 11:01:04 -06001607struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
Pavel Begunkovebc11b62021-08-09 13:04:05 +01001608{
1609 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1610
1611 req = io_put_req_find_next(req);
1612 return req ? &req->work : NULL;
1613}
1614
Jens Axboec9f06aa2022-05-25 11:01:04 -06001615void io_wq_submit_work(struct io_wq_work *work)
Pavel Begunkovd4c81f32020-06-08 21:08:19 +03001616{
Jens Axboe2b188cc2019-01-07 10:46:33 -07001617 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
Jens Axboe6bf9c472022-03-29 10:10:08 -06001618 const struct io_op_def *def = &io_op_defs[req->opcode];
Pavel Begunkovd01905d2021-10-23 12:13:57 +01001619 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1620 bool needs_poll = false;
Jens Axboe6bf9c472022-03-29 10:10:08 -06001621 int ret = 0, err = -ECANCELED;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001622
Pavel Begunkov48dcd382021-08-15 10:40:18 +01001623 /* one will be dropped by ->io_free_work() after returning to io-wq */
1624 if (!(req->flags & REQ_F_REFCOUNT))
1625 __io_req_set_refcount(req, 2);
1626 else
1627 req_ref_get(req);
Pavel Begunkov5d5901a2021-08-11 19:28:29 +01001628
Pavel Begunkovcb2d3442022-04-15 22:08:25 +01001629 io_arm_ltimeout(req);
Jens Axboe6bf9c472022-03-29 10:10:08 -06001630
Pavel Begunkovdadebc32021-08-23 13:30:44 +01001631 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
Pavel Begunkovd01905d2021-10-23 12:13:57 +01001632 if (work->flags & IO_WQ_WORK_CANCEL) {
Pavel Begunkov0f8da752022-04-12 15:24:43 +01001633fail:
Jens Axboe6bf9c472022-03-29 10:10:08 -06001634 io_req_task_queue_fail(req, err);
Pavel Begunkovd01905d2021-10-23 12:13:57 +01001635 return;
Jens Axboe561fb042019-10-24 07:25:42 -06001636 }
Pavel Begunkov0f8da752022-04-12 15:24:43 +01001637 if (!io_assign_file(req, issue_flags)) {
1638 err = -EBADF;
1639 work->flags |= IO_WQ_WORK_CANCEL;
1640 goto fail;
1641 }
Jens Axboe31b51512019-01-18 22:56:34 -07001642
Pavel Begunkovd01905d2021-10-23 12:13:57 +01001643 if (req->flags & REQ_F_FORCE_ASYNC) {
Pavel Begunkovafb7f56f2021-10-23 12:13:59 +01001644 bool opcode_poll = def->pollin || def->pollout;
1645
1646 if (opcode_poll && file_can_poll(req->file)) {
1647 needs_poll = true;
Pavel Begunkovd01905d2021-10-23 12:13:57 +01001648 issue_flags |= IO_URING_F_NONBLOCK;
Pavel Begunkovafb7f56f2021-10-23 12:13:59 +01001649 }
Pavel Begunkovd01905d2021-10-23 12:13:57 +01001650 }
Hao Xu90fa02882021-10-18 21:34:45 +08001651
Pavel Begunkovd01905d2021-10-23 12:13:57 +01001652 do {
1653 ret = io_issue_sqe(req, issue_flags);
1654 if (ret != -EAGAIN)
1655 break;
1656 /*
1657 * We can get EAGAIN for iopolled IO even though we're
1658 * forcing a sync submission from here, since we can't
1659 * wait for request slots on the block side.
1660 */
1661 if (!needs_poll) {
Pavel Begunkove0deb6a2022-05-13 11:24:56 +01001662 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1663 break;
Pavel Begunkovd01905d2021-10-23 12:13:57 +01001664 cond_resched();
1665 continue;
Hao Xu90fa02882021-10-18 21:34:45 +08001666 }
1667
Jens Axboe4d9237e2022-03-15 10:54:08 -06001668 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
Pavel Begunkovd01905d2021-10-23 12:13:57 +01001669 return;
1670 /* aborted or ready, in either case retry blocking */
1671 needs_poll = false;
1672 issue_flags &= ~IO_URING_F_NONBLOCK;
1673 } while (1);
Jens Axboe31b51512019-01-18 22:56:34 -07001674
Pavel Begunkova3df76982021-02-18 22:32:52 +00001675 /* avoid locking problems by failing it from a clean context */
Jens Axboe97b388d2022-05-24 15:21:00 -06001676 if (ret < 0)
Pavel Begunkova3df76982021-02-18 22:32:52 +00001677 io_req_task_queue_fail(req, ret);
Jens Axboe31b51512019-01-18 22:56:34 -07001678}
Jens Axboe2b188cc2019-01-07 10:46:33 -07001679
Jens Axboe531113b2022-05-24 21:19:47 -06001680inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1681 unsigned int issue_flags)
Pavel Begunkov8da11c12020-02-24 11:32:44 +03001682{
Jens Axboe5106dd62022-04-04 17:18:43 -06001683 struct io_ring_ctx *ctx = req->ctx;
1684 struct file *file = NULL;
Pavel Begunkovac177052021-08-09 13:04:02 +01001685 unsigned long file_ptr;
Pavel Begunkov8da11c12020-02-24 11:32:44 +03001686
Pavel Begunkov93f052c2022-04-18 20:51:11 +01001687 io_ring_submit_lock(ctx, issue_flags);
Jens Axboe5106dd62022-04-04 17:18:43 -06001688
Pavel Begunkovac177052021-08-09 13:04:02 +01001689 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
Jens Axboe5106dd62022-04-04 17:18:43 -06001690 goto out;
Pavel Begunkovac177052021-08-09 13:04:02 +01001691 fd = array_index_nospec(fd, ctx->nr_user_files);
1692 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1693 file = (struct file *) (file_ptr & FFS_MASK);
1694 file_ptr &= ~FFS_MASK;
1695 /* mask in overlapping REQ_F and FFS bits */
Pavel Begunkov35645ac2021-10-17 00:07:09 +01001696 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
Jens Axboe5106dd62022-04-04 17:18:43 -06001697 io_req_set_rsrc_node(req, ctx, 0);
Jens Axboed78bd8a2022-05-07 09:56:13 -06001698 WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap));
Jens Axboe5106dd62022-04-04 17:18:43 -06001699out:
Pavel Begunkov93f052c2022-04-18 20:51:11 +01001700 io_ring_submit_unlock(ctx, issue_flags);
Pavel Begunkov8371adf2020-10-10 18:34:08 +01001701 return file;
Pavel Begunkov8da11c12020-02-24 11:32:44 +03001702}
1703
Jens Axboe531113b2022-05-24 21:19:47 -06001704struct file *io_file_get_normal(struct io_kiocb *req, int fd)
Pavel Begunkovac177052021-08-09 13:04:02 +01001705{
Pavel Begunkov62906e82021-08-10 14:52:47 +01001706 struct file *file = fget(fd);
Pavel Begunkovac177052021-08-09 13:04:02 +01001707
Pavel Begunkov48863ff2022-06-16 13:57:20 +01001708 trace_io_uring_file_get(req, fd);
Pavel Begunkovac177052021-08-09 13:04:02 +01001709
1710 /* we don't allow fixed io_uring files */
Jens Axboee5550a12022-05-25 10:28:04 -06001711 if (file && io_is_uring_fops(file))
Jens Axboe9cae36a2022-06-01 23:57:02 -06001712 io_req_track_inflight(req);
Pavel Begunkovac177052021-08-09 13:04:02 +01001713 return file;
1714}
1715
Pavel Begunkov7bfa9ba2022-04-15 22:08:28 +01001716static void io_queue_async(struct io_kiocb *req, int ret)
Pavel Begunkovd475a9a2021-09-24 21:59:59 +01001717 __must_hold(&req->ctx->uring_lock)
1718{
Pavel Begunkov7bfa9ba2022-04-15 22:08:28 +01001719 struct io_kiocb *linked_timeout;
1720
1721 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1722 io_req_complete_failed(req, ret);
1723 return;
1724 }
1725
1726 linked_timeout = io_prep_linked_timeout(req);
Pavel Begunkovd475a9a2021-09-24 21:59:59 +01001727
Jens Axboe4d9237e2022-03-15 10:54:08 -06001728 switch (io_arm_poll_handler(req, 0)) {
Pavel Begunkovd475a9a2021-09-24 21:59:59 +01001729 case IO_APOLL_READY:
Pavel Begunkovd475a9a2021-09-24 21:59:59 +01001730 io_req_task_queue(req);
1731 break;
1732 case IO_APOLL_ABORTED:
1733 /*
1734 * Queued up for async execution, worker will release
1735 * submit reference when the iocb is actually submitted.
1736 */
Jens Axboe6436c772022-06-17 06:24:26 -06001737 io_kbuf_recycle(req, 0);
Pavel Begunkov77955ef2022-04-15 22:08:27 +01001738 io_queue_iowq(req, NULL);
Pavel Begunkovd475a9a2021-09-24 21:59:59 +01001739 break;
Jens Axboeb1c62642022-03-09 11:27:52 -07001740 case IO_APOLL_OK:
Jens Axboeb1c62642022-03-09 11:27:52 -07001741 break;
Pavel Begunkovd475a9a2021-09-24 21:59:59 +01001742 }
1743
1744 if (linked_timeout)
1745 io_queue_linked_timeout(linked_timeout);
1746}
1747
Pavel Begunkovcbc2e202022-04-15 22:08:26 +01001748static inline void io_queue_sqe(struct io_kiocb *req)
Pavel Begunkov282cdc82021-08-09 13:04:10 +01001749 __must_hold(&req->ctx->uring_lock)
Jens Axboe2b188cc2019-01-07 10:46:33 -07001750{
Jens Axboee0c5c572019-03-12 10:18:47 -06001751 int ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07001752
Pavel Begunkovc5eef2b942021-02-10 00:03:22 +00001753 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
Jens Axboe491381ce2019-10-17 09:20:46 -06001754
1755 /*
1756 * We async punt it if the file wasn't marked NOWAIT, or if the file
1757 * doesn't support non-blocking read/write attempts
1758 */
Pavel Begunkov7bfa9ba2022-04-15 22:08:28 +01001759 if (likely(!ret))
Pavel Begunkovcb2d3442022-04-15 22:08:25 +01001760 io_arm_ltimeout(req);
Pavel Begunkov7bfa9ba2022-04-15 22:08:28 +01001761 else
1762 io_queue_async(req, ret);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001763}
1764
Pavel Begunkov4652fe32021-09-24 21:59:58 +01001765static void io_queue_sqe_fallback(struct io_kiocb *req)
Pavel Begunkov282cdc82021-08-09 13:04:10 +01001766 __must_hold(&req->ctx->uring_lock)
Jackie Liu4fe2c962019-09-09 20:50:40 +08001767{
Pavel Begunkov17b147f2022-04-15 22:08:32 +01001768 if (unlikely(req->flags & REQ_F_FAIL)) {
1769 /*
1770 * We don't submit, fail them all, for that replace hardlinks
1771 * with normal links. Extra REQ_F_LINK is tolerated.
1772 */
1773 req->flags &= ~REQ_F_HARDLINK;
1774 req->flags |= REQ_F_LINK;
1775 io_req_complete_failed(req, req->cqe.res);
Pavel Begunkove0eb71d2021-10-01 18:07:01 +01001776 } else if (unlikely(req->ctx->drain_active)) {
1777 io_drain_req(req);
Pavel Begunkov76cc33d2021-06-14 23:37:30 +01001778 } else {
1779 int ret = io_req_prep_async(req);
1780
1781 if (unlikely(ret))
1782 io_req_complete_failed(req, ret);
1783 else
Pavel Begunkov77955ef2022-04-15 22:08:27 +01001784 io_queue_iowq(req, NULL);
Jens Axboece35a472019-12-17 08:04:44 -07001785 }
Jackie Liu4fe2c962019-09-09 20:50:40 +08001786}
1787
Stefano Garzarella21b55db2020-08-27 16:58:30 +02001788/*
1789 * Check SQE restrictions (opcode and flags).
1790 *
1791 * Returns 'true' if SQE is allowed, 'false' otherwise.
1792 */
1793static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1794 struct io_kiocb *req,
1795 unsigned int sqe_flags)
1796{
Stefano Garzarella21b55db2020-08-27 16:58:30 +02001797 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1798 return false;
1799
1800 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1801 ctx->restrictions.sqe_flags_required)
1802 return false;
1803
1804 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1805 ctx->restrictions.sqe_flags_required))
1806 return false;
1807
1808 return true;
1809}
1810
Pavel Begunkov22b2ca32021-10-01 18:07:00 +01001811static void io_init_req_drain(struct io_kiocb *req)
1812{
1813 struct io_ring_ctx *ctx = req->ctx;
1814 struct io_kiocb *head = ctx->submit_state.link.head;
1815
1816 ctx->drain_active = true;
1817 if (head) {
1818 /*
1819 * If we need to drain a request in the middle of a link, drain
1820 * the head request and the next request/link after the current
1821 * link. Considering sequential execution of links,
Hao Xub6c7db32021-11-25 17:21:03 +08001822 * REQ_F_IO_DRAIN will be maintained for every request of our
Pavel Begunkov22b2ca32021-10-01 18:07:00 +01001823 * link.
1824 */
Hao Xub6c7db32021-11-25 17:21:03 +08001825 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
Pavel Begunkov22b2ca32021-10-01 18:07:00 +01001826 ctx->drain_next = true;
1827 }
1828}
1829
Pavel Begunkovef4ff582020-04-12 02:05:05 +03001830static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
Pavel Begunkov258b29a2021-02-10 00:03:10 +00001831 const struct io_uring_sqe *sqe)
Pavel Begunkov282cdc82021-08-09 13:04:10 +01001832 __must_hold(&ctx->uring_lock)
Pavel Begunkov0553b8b2020-04-08 08:58:45 +03001833{
Jens Axboefcde59f2022-05-23 16:53:15 -06001834 const struct io_op_def *def;
Pavel Begunkovef4ff582020-04-12 02:05:05 +03001835 unsigned int sqe_flags;
Pavel Begunkovfc0ae022021-10-01 18:07:02 +01001836 int personality;
Pavel Begunkov4a04d1d2021-10-06 16:06:49 +01001837 u8 opcode;
Pavel Begunkovef4ff582020-04-12 02:05:05 +03001838
Pavel Begunkov864ea922021-08-09 13:04:08 +01001839 /* req is partially pre-initialised, see io_preinit_req() */
Pavel Begunkov4a04d1d2021-10-06 16:06:49 +01001840 req->opcode = opcode = READ_ONCE(sqe->opcode);
Pavel Begunkov5be9ad12021-02-12 18:41:17 +00001841 /* same numerical values with corresponding REQ_F_*, safe to copy */
1842 req->flags = sqe_flags = READ_ONCE(sqe->flags);
Pavel Begunkovcef216f2022-04-12 15:09:43 +01001843 req->cqe.user_data = READ_ONCE(sqe->user_data);
Pavel Begunkov0553b8b2020-04-08 08:58:45 +03001844 req->file = NULL;
Pavel Begunkovc1bdf8e2022-04-18 20:51:13 +01001845 req->rsrc_node = NULL;
Pavel Begunkov4dd28242020-06-15 10:33:13 +03001846 req->task = current;
Pavel Begunkovef4ff582020-04-12 02:05:05 +03001847
Pavel Begunkov4a04d1d2021-10-06 16:06:49 +01001848 if (unlikely(opcode >= IORING_OP_LAST)) {
1849 req->opcode = 0;
Pavel Begunkov5be9ad12021-02-12 18:41:17 +00001850 return -EINVAL;
Pavel Begunkov4a04d1d2021-10-06 16:06:49 +01001851 }
Jens Axboefcde59f2022-05-23 16:53:15 -06001852 def = &io_op_defs[opcode];
Pavel Begunkov68fe2562021-09-15 12:03:38 +01001853 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
1854 /* enforce forwards compatibility on users */
1855 if (sqe_flags & ~SQE_VALID_FLAGS)
1856 return -EINVAL;
Jens Axboe4e906702022-04-28 19:09:43 -06001857 if (sqe_flags & IOSQE_BUFFER_SELECT) {
Jens Axboefcde59f2022-05-23 16:53:15 -06001858 if (!def->buffer_select)
Jens Axboe4e906702022-04-28 19:09:43 -06001859 return -EOPNOTSUPP;
1860 req->buf_index = READ_ONCE(sqe->buf_group);
1861 }
Pavel Begunkov5562a8d2021-11-10 15:49:34 +00001862 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
1863 ctx->drain_disabled = true;
1864 if (sqe_flags & IOSQE_IO_DRAIN) {
1865 if (ctx->drain_disabled)
1866 return -EOPNOTSUPP;
Pavel Begunkov22b2ca32021-10-01 18:07:00 +01001867 io_init_req_drain(req);
Pavel Begunkov5562a8d2021-11-10 15:49:34 +00001868 }
Pavel Begunkov68fe2562021-09-15 12:03:38 +01001869 }
Pavel Begunkov2a56a9b2021-09-24 21:59:57 +01001870 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
1871 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
1872 return -EACCES;
1873 /* knock it to the slow queue path, will be drained there */
1874 if (ctx->drain_active)
1875 req->flags |= REQ_F_FORCE_ASYNC;
1876 /* if there is no link, we're at "next" request and need to drain */
1877 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
1878 ctx->drain_next = false;
1879 ctx->drain_active = true;
Hao Xub6c7db32021-11-25 17:21:03 +08001880 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
Pavel Begunkov2a56a9b2021-09-24 21:59:57 +01001881 }
1882 }
Stefano Garzarella21b55db2020-08-27 16:58:30 +02001883
Jens Axboefcde59f2022-05-23 16:53:15 -06001884 if (!def->ioprio && sqe->ioprio)
Jens Axboe73911422022-04-26 11:34:56 -06001885 return -EINVAL;
Jens Axboefcde59f2022-05-23 16:53:15 -06001886 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
Jens Axboe73911422022-04-26 11:34:56 -06001887 return -EINVAL;
1888
Jens Axboefcde59f2022-05-23 16:53:15 -06001889 if (def->needs_file) {
Pavel Begunkov6d634162021-10-06 16:06:46 +01001890 struct io_submit_state *state = &ctx->submit_state;
1891
Pavel Begunkovcef216f2022-04-12 15:09:43 +01001892 req->cqe.fd = READ_ONCE(sqe->fd);
Jens Axboe6bf9c472022-03-29 10:10:08 -06001893
Pavel Begunkov6d634162021-10-06 16:06:46 +01001894 /*
1895 * Plug now if we have more than 2 IO left after this, and the
1896 * target is potentially a read/write to block based storage.
1897 */
Jens Axboefcde59f2022-05-23 16:53:15 -06001898 if (state->need_plug && def->plug) {
Pavel Begunkov6d634162021-10-06 16:06:46 +01001899 state->plug_started = true;
1900 state->need_plug = false;
Jens Axboe5ca7a8b2021-10-06 11:01:42 -06001901 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
Pavel Begunkov6d634162021-10-06 16:06:46 +01001902 }
Pavel Begunkovbd5bbda2020-11-20 15:50:51 +00001903 }
Pavel Begunkovc11368a52020-05-17 14:13:42 +03001904
Pavel Begunkovef4ff582020-04-12 02:05:05 +03001905 personality = READ_ONCE(sqe->personality);
Jens Axboe63ff8222020-05-07 14:56:15 -06001906 if (personality) {
Linus Torvaldscdab10b2021-11-01 21:06:18 -07001907 int ret;
1908
Jens Axboe63ff8222020-05-07 14:56:15 -06001909 req->creds = xa_load(&ctx->personalities, personality);
1910 if (!req->creds)
Jens Axboe27926b62020-10-28 09:33:23 -06001911 return -EINVAL;
Jens Axboe63ff8222020-05-07 14:56:15 -06001912 get_cred(req->creds);
Paul Moorecdc14042021-02-01 19:56:49 -05001913 ret = security_uring_override_creds(req->creds);
1914 if (ret) {
1915 put_cred(req->creds);
1916 return ret;
1917 }
Jens Axboe63ff8222020-05-07 14:56:15 -06001918 req->flags |= REQ_F_CREDS;
1919 }
Pavel Begunkovbd5bbda2020-11-20 15:50:51 +00001920
Jens Axboe0702e532022-05-23 16:56:21 -06001921 return def->prep(req, sqe);
Pavel Begunkov0553b8b2020-04-08 08:58:45 +03001922}
1923
Pavel Begunkovdf3becd2022-04-15 22:08:30 +01001924static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
1925 struct io_kiocb *req, int ret)
1926{
1927 struct io_ring_ctx *ctx = req->ctx;
1928 struct io_submit_link *link = &ctx->submit_state.link;
1929 struct io_kiocb *head = link->head;
1930
Pavel Begunkov48863ff2022-06-16 13:57:20 +01001931 trace_io_uring_req_failed(sqe, req, ret);
Pavel Begunkovdf3becd2022-04-15 22:08:30 +01001932
1933 /*
1934 * Avoid breaking links in the middle as it renders links with SQPOLL
1935 * unusable. Instead of failing eagerly, continue assembling the link if
1936 * applicable and mark the head with REQ_F_FAIL. The link flushing code
1937 * should find the flag and handle the rest.
1938 */
1939 req_fail_link_node(req, ret);
1940 if (head && !(head->flags & REQ_F_FAIL))
1941 req_fail_link_node(head, -ECANCELED);
1942
1943 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
1944 if (head) {
1945 link->last->link = req;
1946 link->head = NULL;
1947 req = head;
1948 }
1949 io_queue_sqe_fallback(req);
1950 return ret;
1951 }
1952
1953 if (head)
1954 link->last->link = req;
1955 else
1956 link->head = req;
1957 link->last = req;
1958 return 0;
1959}
1960
1961static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
Pavel Begunkova1ab7b32021-02-18 18:29:42 +00001962 const struct io_uring_sqe *sqe)
Pavel Begunkov282cdc82021-08-09 13:04:10 +01001963 __must_hold(&ctx->uring_lock)
Jens Axboe6c271ce2019-01-10 11:22:30 -07001964{
Pavel Begunkova1ab7b32021-02-18 18:29:42 +00001965 struct io_submit_link *link = &ctx->submit_state.link;
Jens Axboe6c271ce2019-01-10 11:22:30 -07001966 int ret;
1967
Pavel Begunkova6b8cadc2021-02-18 18:29:41 +00001968 ret = io_init_req(ctx, req, sqe);
Pavel Begunkovdf3becd2022-04-15 22:08:30 +01001969 if (unlikely(ret))
1970 return io_submit_fail_init(sqe, req, ret);
Pavel Begunkov441b8a72021-06-14 23:37:31 +01001971
Pavel Begunkovbe7053b2021-02-18 18:29:45 +00001972 /* don't need @sqe from now on */
Pavel Begunkov48863ff2022-06-16 13:57:20 +01001973 trace_io_uring_submit_sqe(req, true);
Pavel Begunkova6b8cadc2021-02-18 18:29:41 +00001974
Jens Axboe6c271ce2019-01-10 11:22:30 -07001975 /*
1976 * If we already have a head request, queue this one for async
1977 * submittal once the head completes. If we don't have a head but
1978 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
1979 * submitted sync once the chain is complete. If none of those
1980 * conditions are true (normal request), then just queue it.
1981 */
Pavel Begunkov924a07e2022-04-15 22:08:31 +01001982 if (unlikely(link->head)) {
Pavel Begunkovdf3becd2022-04-15 22:08:30 +01001983 ret = io_req_prep_async(req);
1984 if (unlikely(ret))
1985 return io_submit_fail_init(sqe, req, ret);
Jens Axboe6c271ce2019-01-10 11:22:30 -07001986
Pavel Begunkov48863ff2022-06-16 13:57:20 +01001987 trace_io_uring_link(req, link->head);
Jens Axboe2b188cc2019-01-07 10:46:33 -07001988 link->last->link = req;
1989 link->last = req;
1990
Pavel Begunkovda1a08c2022-04-15 22:08:29 +01001991 if (req->flags & IO_REQ_LINK_FLAGS)
Pavel Begunkovf15a3432021-09-24 21:59:56 +01001992 return 0;
Pavel Begunkovdf3becd2022-04-15 22:08:30 +01001993 /* last request of the link, flush it */
1994 req = link->head;
Pavel Begunkovf15a3432021-09-24 21:59:56 +01001995 link->head = NULL;
Pavel Begunkov924a07e2022-04-15 22:08:31 +01001996 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
1997 goto fallback;
1998
1999 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2000 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2001 if (req->flags & IO_REQ_LINK_FLAGS) {
2002 link->head = req;
2003 link->last = req;
2004 } else {
2005fallback:
2006 io_queue_sqe_fallback(req);
2007 }
Pavel Begunkovf15a3432021-09-24 21:59:56 +01002008 return 0;
Jackie Liu4fe2c962019-09-09 20:50:40 +08002009 }
2010
Pavel Begunkovf15a3432021-09-24 21:59:56 +01002011 io_queue_sqe(req);
Jackie Liu4fe2c962019-09-09 20:50:40 +08002012 return 0;
2013}
2014
2015/*
2016 * Batched submission is done, ensure local IO is flushed out.
Pavel Begunkov1b4a51b2019-11-21 11:54:28 +03002017 */
Pavel Begunkov553deff2021-09-24 21:59:55 +01002018static void io_submit_state_end(struct io_ring_ctx *ctx)
Jackie Liua197f662019-11-08 08:09:12 -07002019{
Pavel Begunkov553deff2021-09-24 21:59:55 +01002020 struct io_submit_state *state = &ctx->submit_state;
2021
Pavel Begunkove1263912022-04-12 15:09:45 +01002022 if (unlikely(state->link.head))
2023 io_queue_sqe_fallback(state->link.head);
Pavel Begunkov553deff2021-09-24 21:59:55 +01002024 /* flush only after queuing links as they can generate completions */
Pavel Begunkovc4501782021-09-08 16:40:52 +01002025 io_submit_flush_completions(ctx);
Jens Axboe9e645e112019-05-10 16:07:28 -06002026 if (state->plug_started)
Pavel Begunkov32fe5252019-12-17 22:26:58 +03002027 blk_finish_plug(&state->plug);
Jens Axboe9e645e112019-05-10 16:07:28 -06002028}
Pavel Begunkov32fe5252019-12-17 22:26:58 +03002029
Jens Axboe9e645e112019-05-10 16:07:28 -06002030/*
Pavel Begunkov196be952019-11-07 01:41:06 +03002031 * Start submission side cache.
Jens Axboe9e645e112019-05-10 16:07:28 -06002032 */
2033static void io_submit_state_start(struct io_submit_state *state,
Jens Axboebcda7ba2020-02-23 16:42:51 -07002034 unsigned int max_ios)
2035{
2036 state->plug_started = false;
Pavel Begunkov4b628ae2021-09-08 16:40:49 +01002037 state->need_plug = max_ios > 2;
Jens Axboe5ca7a8b2021-10-06 11:01:42 -06002038 state->submit_nr = max_ios;
Pavel Begunkova1ab7b32021-02-18 18:29:42 +00002039 /* set only head, no need to init link_last in advance */
2040 state->link.head = NULL;
Jens Axboe193155c2020-02-22 23:22:19 -07002041}
Jens Axboe75c6a032020-01-28 10:15:23 -07002042
2043static void io_commit_sqring(struct io_ring_ctx *ctx)
2044{
Jens Axboe193155c2020-02-22 23:22:19 -07002045 struct io_rings *rings = ctx->rings;
Jens Axboe75c6a032020-01-28 10:15:23 -07002046
2047 /*
Pavel Begunkov6b47ee62020-01-18 20:22:41 +03002048 * Ensure any loads from the SQEs are done at this point,
Pavel Begunkov8da11c12020-02-24 11:32:44 +03002049 * since once we write the new head, the application could
Jens Axboebcda7ba2020-02-23 16:42:51 -07002050 * write new data to them.
2051 */
Jens Axboe9e645e112019-05-10 16:07:28 -06002052 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
Jens Axboe3529d8c2019-12-19 18:24:38 -07002053}
Jens Axboe9e645e112019-05-10 16:07:28 -06002054
2055/*
Fam Zhengdd9ae8a2021-06-04 17:42:56 +01002056 * Fetch an sqe, if one is available. Note this returns a pointer to memory
Jens Axboe78e19bb2019-11-06 15:21:34 -07002057 * that is mapped by userspace. This means that care needs to be taken to
Pavel Begunkov2e6e1fd2019-12-05 16:15:45 +03002058 * ensure that reads are stable, as we cannot rely on userspace always
Jens Axboe9e645e112019-05-10 16:07:28 -06002059 * being a good citizen. If members of the sqe are validated and then later
2060 * used, it's important that those reads are done through READ_ONCE() to
Jens Axboe9e645e112019-05-10 16:07:28 -06002061 * prevent a re-load down the line.
2062 */
2063static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2064{
Pavel Begunkovea5ab3b2021-05-16 22:58:09 +01002065 unsigned head, mask = ctx->sq_entries - 1;
Pavel Begunkov17d3aeb2021-06-14 23:37:23 +01002066 unsigned sq_idx = ctx->cached_sq_head++ & mask;
Jens Axboe9e645e112019-05-10 16:07:28 -06002067
2068 /*
Pavel Begunkov9d763772019-12-17 02:22:07 +03002069 * The cached sq head (or cq tail) serves two purposes:
Jens Axboe9e645e112019-05-10 16:07:28 -06002070 *
Pavel Begunkov8cdf2192020-01-25 00:40:24 +03002071 * 1) allows us to batch the cost of updating the user visible
2072 * head updates.
2073 * 2) allows the kernel side to track the head on its own, even
2074 * though the application is the one updating it.
2075 */
Pavel Begunkov17d3aeb2021-06-14 23:37:23 +01002076 head = READ_ONCE(ctx->sq_array[sq_idx]);
Jens Axboeebdeb7c02022-03-31 19:27:52 -06002077 if (likely(head < ctx->sq_entries)) {
2078 /* double index for 128-byte SQEs, twice as long */
2079 if (ctx->flags & IORING_SETUP_SQE128)
2080 head <<= 1;
Pavel Begunkov711be032020-01-17 03:57:59 +03002081 return &ctx->sq_sqes[head];
Jens Axboeebdeb7c02022-03-31 19:27:52 -06002082 }
Pavel Begunkov711be032020-01-17 03:57:59 +03002083
2084 /* drop invalid entries */
Pavel Begunkov15641e42021-06-14 23:37:24 +01002085 ctx->cq_extra--;
2086 WRITE_ONCE(ctx->rings->sq_dropped,
2087 READ_ONCE(ctx->rings->sq_dropped) + 1);
Jens Axboe9e645e112019-05-10 16:07:28 -06002088 return NULL;
2089}
2090
Jens Axboe17437f32022-05-25 09:13:39 -06002091int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
Pavel Begunkov282cdc82021-08-09 13:04:10 +01002092 __must_hold(&ctx->uring_lock)
Jens Axboe6c271ce2019-01-10 11:22:30 -07002093{
Pavel Begunkov69629802021-09-24 22:00:01 +01002094 unsigned int entries = io_sqring_entries(ctx);
Pavel Begunkov8e6971a2022-04-12 15:09:49 +01002095 unsigned int left;
2096 int ret;
Jens Axboe6c271ce2019-01-10 11:22:30 -07002097
Pavel Begunkov51d48da2021-10-04 20:02:47 +01002098 if (unlikely(!entries))
Pavel Begunkov69629802021-09-24 22:00:01 +01002099 return 0;
Pavel Begunkovee7d46d92019-12-30 21:24:45 +03002100 /* make sure SQ entry isn't read before tail */
Pavel Begunkov8e6971a2022-04-12 15:09:49 +01002101 ret = left = min3(nr, ctx->sq_entries, entries);
2102 io_get_task_refs(left);
2103 io_submit_state_start(&ctx->submit_state, left);
Jens Axboe6c271ce2019-01-10 11:22:30 -07002104
Pavel Begunkov69629802021-09-24 22:00:01 +01002105 do {
Jens Axboe3529d8c2019-12-19 18:24:38 -07002106 const struct io_uring_sqe *sqe;
Pavel Begunkov196be952019-11-07 01:41:06 +03002107 struct io_kiocb *req;
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03002108
Pavel Begunkov8e6971a2022-04-12 15:09:49 +01002109 if (unlikely(!io_alloc_req_refill(ctx)))
Pavel Begunkovfb5ccc92019-10-25 12:31:30 +03002110 break;
Pavel Begunkova33ae9c2021-10-04 20:02:49 +01002111 req = io_alloc_req(ctx);
Pavel Begunkov4fccfcb2021-02-12 11:55:17 +00002112 sqe = io_get_sqe(ctx);
2113 if (unlikely(!sqe)) {
Pavel Begunkovfa054572022-04-12 15:09:48 +01002114 io_req_add_to_cache(req, ctx);
Pavel Begunkov4fccfcb2021-02-12 11:55:17 +00002115 break;
2116 }
Pavel Begunkov1cd15902022-04-12 15:09:50 +01002117
2118 /*
2119 * Continue submitting even for sqe failure if the
2120 * ring was setup with IORING_SETUP_SUBMIT_ALL
2121 */
2122 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2123 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2124 left--;
2125 break;
Jens Axboebcbb7bf2022-03-10 12:59:35 -07002126 }
Pavel Begunkov1cd15902022-04-12 15:09:50 +01002127 } while (--left);
Jens Axboe6c271ce2019-01-10 11:22:30 -07002128
Pavel Begunkov8e6971a2022-04-12 15:09:49 +01002129 if (unlikely(left)) {
2130 ret -= left;
2131 /* try again if it submitted nothing and can't allocate a req */
2132 if (!ret && io_req_cache_empty(ctx))
2133 ret = -EAGAIN;
2134 current->io_uring->cached_refs += left;
Pavel Begunkov9466f432020-01-25 22:34:01 +03002135 }
Jens Axboe6c271ce2019-01-10 11:22:30 -07002136
Pavel Begunkov553deff2021-09-24 21:59:55 +01002137 io_submit_state_end(ctx);
Pavel Begunkovae9428c2019-11-06 00:22:14 +03002138 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2139 io_commit_sqring(ctx);
Pavel Begunkov8e6971a2022-04-12 15:09:49 +01002140 return ret;
Jens Axboe6c271ce2019-01-10 11:22:30 -07002141}
2142
Jens Axboebda52162019-09-24 13:47:15 -06002143struct io_wait_queue {
2144 struct wait_queue_entry wq;
2145 struct io_ring_ctx *ctx;
Jens Axboe5fd46172021-08-06 14:04:31 -06002146 unsigned cq_tail;
Jens Axboebda52162019-09-24 13:47:15 -06002147 unsigned nr_timeouts;
2148};
2149
Pavel Begunkov6c503152021-01-04 20:36:36 +00002150static inline bool io_should_wake(struct io_wait_queue *iowq)
Jens Axboebda52162019-09-24 13:47:15 -06002151{
2152 struct io_ring_ctx *ctx = iowq->ctx;
Jens Axboe5fd46172021-08-06 14:04:31 -06002153 int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
Jens Axboebda52162019-09-24 13:47:15 -06002154
2155 /*
Brian Gianforcarod195a662019-12-13 03:09:50 -08002156 * Wake up if we have enough events, or if a timeout occurred since we
Jens Axboebda52162019-09-24 13:47:15 -06002157 * started waiting. For timeouts, we always want to return to userspace,
2158 * regardless of event count.
2159 */
Jens Axboe5fd46172021-08-06 14:04:31 -06002160 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
Jens Axboebda52162019-09-24 13:47:15 -06002161}
2162
2163static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2164 int wake_flags, void *key)
2165{
2166 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2167 wq);
2168
Pavel Begunkov6c503152021-01-04 20:36:36 +00002169 /*
2170 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2171 * the task, and the next invocation will do it.
2172 */
Dylan Yudaken10988a02022-04-21 02:13:43 -07002173 if (io_should_wake(iowq) ||
2174 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq))
Pavel Begunkov6c503152021-01-04 20:36:36 +00002175 return autoremove_wake_function(curr, mode, wake_flags, key);
2176 return -1;
Jens Axboebda52162019-09-24 13:47:15 -06002177}
2178
Jens Axboe73572982022-06-13 07:12:45 -06002179int io_run_task_work_sig(void)
Jens Axboeaf9c1a42020-09-24 13:32:18 -06002180{
2181 if (io_run_task_work())
2182 return 1;
Olivier Langloisc5020bc2022-02-16 14:53:42 -05002183 if (task_sigpending(current))
2184 return -EINTR;
2185 return 0;
Jens Axboeaf9c1a42020-09-24 13:32:18 -06002186}
2187
Pavel Begunkoveeb60b92021-02-04 13:51:58 +00002188/* when returns >0, the caller should retry */
2189static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2190 struct io_wait_queue *iowq,
Jens Axboe22833962022-02-21 05:49:30 -07002191 ktime_t timeout)
Pavel Begunkoveeb60b92021-02-04 13:51:58 +00002192{
2193 int ret;
Dylan Yudaken155bc952022-04-21 02:13:44 -07002194 unsigned long check_cq;
Pavel Begunkoveeb60b92021-02-04 13:51:58 +00002195
2196 /* make sure we run task_work before checking for signals */
2197 ret = io_run_task_work_sig();
2198 if (ret || io_should_wake(iowq))
2199 return ret;
Pavel Begunkov3a085762022-06-15 17:33:55 +01002200
Dylan Yudaken155bc952022-04-21 02:13:44 -07002201 check_cq = READ_ONCE(ctx->check_cq);
Pavel Begunkov3a085762022-06-15 17:33:55 +01002202 if (unlikely(check_cq)) {
2203 /* let the caller flush overflows, retry */
2204 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2205 return 1;
2206 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2207 return -EBADR;
2208 }
Jens Axboe22833962022-02-21 05:49:30 -07002209 if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2210 return -ETIME;
2211 return 1;
Pavel Begunkoveeb60b92021-02-04 13:51:58 +00002212}
2213
Jens Axboe2b188cc2019-01-07 10:46:33 -07002214/*
2215 * Wait until events become available, if we don't already have some. The
2216 * application must reap them itself, as they reside on the shared cq ring.
2217 */
2218static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
Hao Xuc73ebb62020-11-03 10:54:37 +08002219 const sigset_t __user *sig, size_t sigsz,
2220 struct __kernel_timespec __user *uts)
Jens Axboe2b188cc2019-01-07 10:46:33 -07002221{
Pavel Begunkov902910992021-08-09 09:07:32 -06002222 struct io_wait_queue iowq;
Hristo Venev75b28af2019-08-26 17:23:46 +00002223 struct io_rings *rings = ctx->rings;
Jens Axboe22833962022-02-21 05:49:30 -07002224 ktime_t timeout = KTIME_MAX;
Pavel Begunkovc1d5a222021-02-04 13:51:57 +00002225 int ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002226
Jens Axboeb41e9852020-02-17 09:52:41 -07002227 do {
Pavel Begunkov90f67362021-08-09 20:18:12 +01002228 io_cqring_overflow_flush(ctx);
Pavel Begunkov6c503152021-01-04 20:36:36 +00002229 if (io_cqring_events(ctx) >= min_events)
Jens Axboeb41e9852020-02-17 09:52:41 -07002230 return 0;
Jens Axboe4c6e2772020-07-01 11:29:10 -06002231 if (!io_run_task_work())
Jens Axboeb41e9852020-02-17 09:52:41 -07002232 break;
Jens Axboeb41e9852020-02-17 09:52:41 -07002233 } while (1);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002234
2235 if (sig) {
Arnd Bergmann9e75ad52019-03-25 15:34:53 +01002236#ifdef CONFIG_COMPAT
2237 if (in_compat_syscall())
2238 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
Oleg Nesterovb7724342019-07-16 16:29:53 -07002239 sigsz);
Arnd Bergmann9e75ad52019-03-25 15:34:53 +01002240 else
2241#endif
Oleg Nesterovb7724342019-07-16 16:29:53 -07002242 ret = set_user_sigmask(sig, sigsz);
Arnd Bergmann9e75ad52019-03-25 15:34:53 +01002243
Jens Axboe2b188cc2019-01-07 10:46:33 -07002244 if (ret)
2245 return ret;
2246 }
2247
Olivier Langlois950e79d2022-03-08 17:17:21 -05002248 if (uts) {
2249 struct timespec64 ts;
2250
2251 if (get_timespec64(&ts, uts))
2252 return -EFAULT;
2253 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2254 }
2255
Pavel Begunkov902910992021-08-09 09:07:32 -06002256 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2257 iowq.wq.private = current;
2258 INIT_LIST_HEAD(&iowq.wq.entry);
2259 iowq.ctx = ctx;
Jens Axboebda52162019-09-24 13:47:15 -06002260 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
Jens Axboe5fd46172021-08-06 14:04:31 -06002261 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
Pavel Begunkov902910992021-08-09 09:07:32 -06002262
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02002263 trace_io_uring_cqring_wait(ctx, min_events);
Jens Axboebda52162019-09-24 13:47:15 -06002264 do {
Jens Axboeca0a2652021-03-04 17:15:48 -07002265 /* if we can't even flush overflow, don't wait for more */
Pavel Begunkov90f67362021-08-09 20:18:12 +01002266 if (!io_cqring_overflow_flush(ctx)) {
Jens Axboeca0a2652021-03-04 17:15:48 -07002267 ret = -EBUSY;
2268 break;
2269 }
Pavel Begunkov311997b2021-06-14 23:37:28 +01002270 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
Jens Axboebda52162019-09-24 13:47:15 -06002271 TASK_INTERRUPTIBLE);
Jens Axboe22833962022-02-21 05:49:30 -07002272 ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
Jens Axboeca0a2652021-03-04 17:15:48 -07002273 cond_resched();
Pavel Begunkoveeb60b92021-02-04 13:51:58 +00002274 } while (ret > 0);
Jens Axboebda52162019-09-24 13:47:15 -06002275
Jens Axboeb4f20bb2022-03-25 16:39:57 -06002276 finish_wait(&ctx->cq_wait, &iowq.wq);
Jens Axboeb7db41c2020-07-04 08:55:50 -06002277 restore_saved_sigmask_unless(ret == -EINTR);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002278
Hristo Venev75b28af2019-08-26 17:23:46 +00002279 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002280}
2281
Jens Axboe2b188cc2019-01-07 10:46:33 -07002282static void io_mem_free(void *ptr)
2283{
Mark Rutland52e04ef2019-04-30 17:30:21 +01002284 struct page *page;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002285
Mark Rutland52e04ef2019-04-30 17:30:21 +01002286 if (!ptr)
2287 return;
2288
2289 page = virt_to_head_page(ptr);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002290 if (put_page_testzero(page))
2291 free_compound_page(page);
2292}
2293
2294static void *io_mem_alloc(size_t size)
2295{
Shakeel Butt0a3f1e02022-01-24 21:17:36 -08002296 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002297
Shakeel Butt0a3f1e02022-01-24 21:17:36 -08002298 return (void *) __get_free_pages(gfp, get_order(size));
Jens Axboe2b188cc2019-01-07 10:46:33 -07002299}
2300
Stefan Roeschbaf9cb62022-04-26 11:21:25 -07002301static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2302 unsigned int cq_entries, size_t *sq_offset)
Hristo Venev75b28af2019-08-26 17:23:46 +00002303{
2304 struct io_rings *rings;
2305 size_t off, sq_array_size;
2306
2307 off = struct_size(rings, cqes, cq_entries);
2308 if (off == SIZE_MAX)
2309 return SIZE_MAX;
Stefan Roeschbaf9cb62022-04-26 11:21:25 -07002310 if (ctx->flags & IORING_SETUP_CQE32) {
2311 if (check_shl_overflow(off, 1, &off))
2312 return SIZE_MAX;
2313 }
Hristo Venev75b28af2019-08-26 17:23:46 +00002314
2315#ifdef CONFIG_SMP
2316 off = ALIGN(off, SMP_CACHE_BYTES);
2317 if (off == 0)
2318 return SIZE_MAX;
2319#endif
2320
Dmitry Vyukovb36200f2020-07-11 11:31:11 +02002321 if (sq_offset)
2322 *sq_offset = off;
2323
Hristo Venev75b28af2019-08-26 17:23:46 +00002324 sq_array_size = array_size(sizeof(u32), sq_entries);
2325 if (sq_array_size == SIZE_MAX)
2326 return SIZE_MAX;
2327
2328 if (check_add_overflow(off, sq_array_size, &off))
2329 return SIZE_MAX;
2330
Hristo Venev75b28af2019-08-26 17:23:46 +00002331 return off;
2332}
2333
Usama Arifc75312d2022-02-04 14:51:15 +00002334static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2335 unsigned int eventfd_async)
Jens Axboe9b402842019-04-11 11:45:41 -06002336{
Usama Arif77bc59b2022-02-04 14:51:14 +00002337 struct io_ev_fd *ev_fd;
Jens Axboe9b402842019-04-11 11:45:41 -06002338 __s32 __user *fds = arg;
Nathan Chancellorf0a4e62b2022-02-07 09:24:11 -07002339 int fd;
Jens Axboe9b402842019-04-11 11:45:41 -06002340
Usama Arif77bc59b2022-02-04 14:51:14 +00002341 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2342 lockdep_is_held(&ctx->uring_lock));
2343 if (ev_fd)
Jens Axboe9b402842019-04-11 11:45:41 -06002344 return -EBUSY;
2345
2346 if (copy_from_user(&fd, fds, sizeof(*fds)))
2347 return -EFAULT;
2348
Usama Arif77bc59b2022-02-04 14:51:14 +00002349 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2350 if (!ev_fd)
2351 return -ENOMEM;
Pavel Begunkovfe7e3252021-06-24 15:09:57 +01002352
Usama Arif77bc59b2022-02-04 14:51:14 +00002353 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2354 if (IS_ERR(ev_fd->cq_ev_fd)) {
Nathan Chancellorf0a4e62b2022-02-07 09:24:11 -07002355 int ret = PTR_ERR(ev_fd->cq_ev_fd);
Usama Arif77bc59b2022-02-04 14:51:14 +00002356 kfree(ev_fd);
Jens Axboe9b402842019-04-11 11:45:41 -06002357 return ret;
2358 }
Pavel Begunkov305bef92022-06-20 01:25:55 +01002359
2360 spin_lock(&ctx->completion_lock);
2361 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2362 spin_unlock(&ctx->completion_lock);
2363
Usama Arifc75312d2022-02-04 14:51:15 +00002364 ev_fd->eventfd_async = eventfd_async;
Pavel Begunkov9aa8dfd2022-03-17 02:03:42 +00002365 ctx->has_evfd = true;
Usama Arif77bc59b2022-02-04 14:51:14 +00002366 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
Nathan Chancellorf0a4e62b2022-02-07 09:24:11 -07002367 return 0;
Usama Arif77bc59b2022-02-04 14:51:14 +00002368}
2369
2370static void io_eventfd_put(struct rcu_head *rcu)
2371{
2372 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
2373
2374 eventfd_ctx_put(ev_fd->cq_ev_fd);
2375 kfree(ev_fd);
Jens Axboe9b402842019-04-11 11:45:41 -06002376}
2377
2378static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2379{
Usama Arif77bc59b2022-02-04 14:51:14 +00002380 struct io_ev_fd *ev_fd;
2381
2382 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2383 lockdep_is_held(&ctx->uring_lock));
2384 if (ev_fd) {
Pavel Begunkov9aa8dfd2022-03-17 02:03:42 +00002385 ctx->has_evfd = false;
Usama Arif77bc59b2022-02-04 14:51:14 +00002386 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2387 call_rcu(&ev_fd->rcu, io_eventfd_put);
Jens Axboe9b402842019-04-11 11:45:41 -06002388 return 0;
2389 }
2390
2391 return -ENXIO;
2392}
2393
Jens Axboe4010fec2021-02-27 15:04:18 -07002394static void io_req_caches_free(struct io_ring_ctx *ctx)
Jens Axboe2b188cc2019-01-07 10:46:33 -07002395{
Pavel Begunkovcd0ca2e2021-08-09 20:18:11 +01002396 struct io_submit_state *state = &ctx->submit_state;
Pavel Begunkov37f0e762021-10-04 20:02:53 +01002397 int nr = 0;
Pavel Begunkovbf019da2021-02-10 00:03:17 +00002398
Jens Axboe9a4fdbd2021-02-13 09:09:44 -07002399 mutex_lock(&ctx->uring_lock);
Pavel Begunkovcd0ca2e2021-08-09 20:18:11 +01002400 io_flush_cached_locked_reqs(ctx, state);
Pavel Begunkovc2b6c6b2021-09-24 21:59:47 +01002401
Pavel Begunkov88ab95b2022-04-12 15:09:47 +01002402 while (!io_req_cache_empty(ctx)) {
Pavel Begunkovc2b6c6b2021-09-24 21:59:47 +01002403 struct io_wq_work_node *node;
2404 struct io_kiocb *req;
2405
2406 node = wq_stack_extract(&state->free_list);
2407 req = container_of(node, struct io_kiocb, comp_list);
2408 kmem_cache_free(req_cachep, req);
Pavel Begunkov37f0e762021-10-04 20:02:53 +01002409 nr++;
Pavel Begunkovc2b6c6b2021-09-24 21:59:47 +01002410 }
Pavel Begunkov37f0e762021-10-04 20:02:53 +01002411 if (nr)
2412 percpu_ref_put_many(&ctx->refs, nr);
Jens Axboe9a4fdbd2021-02-13 09:09:44 -07002413 mutex_unlock(&ctx->uring_lock);
2414}
2415
Pavel Begunkovc0724812021-10-04 20:02:54 +01002416static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
Jens Axboe2b188cc2019-01-07 10:46:33 -07002417{
Jens Axboe37d1e2e2021-02-17 21:03:43 -07002418 io_sq_thread_finish(ctx);
Jens Axboe2aede0e2020-09-14 10:45:53 -06002419
Jens Axboe37d1e2e2021-02-17 21:03:43 -07002420 if (ctx->mm_account) {
Jens Axboe2aede0e2020-09-14 10:45:53 -06002421 mmdrop(ctx->mm_account);
2422 ctx->mm_account = NULL;
Bijan Mottahedeh30975822020-06-16 16:36:09 -07002423 }
Jens Axboedef596e2019-01-09 08:59:42 -07002424
Pavel Begunkovab409402021-10-09 23:14:41 +01002425 io_rsrc_refs_drop(ctx);
Pavel Begunkov43597aa2021-08-10 02:44:23 +01002426 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2427 io_wait_rsrc_data(ctx->buf_data);
2428 io_wait_rsrc_data(ctx->file_data);
2429
Hao Xu8bad28d2021-02-19 17:19:36 +08002430 mutex_lock(&ctx->uring_lock);
Pavel Begunkov43597aa2021-08-10 02:44:23 +01002431 if (ctx->buf_data)
Bijan Mottahedehbd54b6f2021-04-25 14:32:25 +01002432 __io_sqe_buffers_unregister(ctx);
Pavel Begunkov43597aa2021-08-10 02:44:23 +01002433 if (ctx->file_data)
Pavel Begunkov08480402021-04-13 02:58:38 +01002434 __io_sqe_files_unregister(ctx);
Pavel Begunkovc4ea0602021-04-01 15:43:58 +01002435 if (ctx->rings)
2436 __io_cqring_overflow_flush(ctx, true);
Jens Axboe9b402842019-04-11 11:45:41 -06002437 io_eventfd_unregister(ctx);
Jens Axboe9b797a32022-07-07 14:16:20 -06002438 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
Jens Axboe43e0bbb2022-07-07 14:30:09 -06002439 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
Usama Arif77bc59b2022-02-04 14:51:14 +00002440 mutex_unlock(&ctx->uring_lock);
Jens Axboe5a2e7452020-02-23 16:23:11 -07002441 io_destroy_buffers(ctx);
Pavel Begunkov07db2982021-04-20 12:03:32 +01002442 if (ctx->sq_creds)
2443 put_cred(ctx->sq_creds);
Pavel Begunkov97bbdc02022-06-16 10:22:08 +01002444 if (ctx->submitter_task)
2445 put_task_struct(ctx->submitter_task);
Jens Axboedef596e2019-01-09 08:59:42 -07002446
Pavel Begunkova7f0ed52021-04-01 15:43:46 +01002447 /* there are no registered resources left, nobody uses it */
2448 if (ctx->rsrc_node)
2449 io_rsrc_node_destroy(ctx->rsrc_node);
Pavel Begunkov8dd03af2021-03-19 17:22:36 +00002450 if (ctx->rsrc_backup_node)
Pavel Begunkovb895c9a2021-04-01 15:43:40 +01002451 io_rsrc_node_destroy(ctx->rsrc_backup_node);
Pavel Begunkova7f0ed52021-04-01 15:43:46 +01002452 flush_delayed_work(&ctx->rsrc_put_work);
Pavel Begunkov756ab7c2021-10-06 16:06:47 +01002453 flush_delayed_work(&ctx->fallback_work);
Pavel Begunkova7f0ed52021-04-01 15:43:46 +01002454
2455 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2456 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
Jens Axboe2b188cc2019-01-07 10:46:33 -07002457
2458#if defined(CONFIG_UNIX)
Eric Biggers355e8d262019-06-12 14:58:43 -07002459 if (ctx->ring_sock) {
2460 ctx->ring_sock->file = NULL; /* so that iput() is called */
Jens Axboe2b188cc2019-01-07 10:46:33 -07002461 sock_release(ctx->ring_sock);
Eric Biggers355e8d262019-06-12 14:58:43 -07002462 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07002463#endif
Pavel Begunkovef9dd632021-08-28 19:54:38 -06002464 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
Pavel Begunkoveb42ceb2022-07-12 21:52:38 +01002465 WARN_ON_ONCE(ctx->notif_slots || ctx->nr_notif_slots);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002466
Hristo Venev75b28af2019-08-26 17:23:46 +00002467 io_mem_free(ctx->rings);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002468 io_mem_free(ctx->sq_sqes);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002469
2470 percpu_ref_exit(&ctx->refs);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002471 free_uid(ctx->user);
Jens Axboe4010fec2021-02-27 15:04:18 -07002472 io_req_caches_free(ctx);
Jens Axboee9418942021-02-19 12:33:30 -07002473 if (ctx->hash_map)
2474 io_wq_put_hash(ctx->hash_map);
Pavel Begunkove6f89be2022-06-16 10:22:10 +01002475 kfree(ctx->cancel_table.hbs);
Pavel Begunkov9ca9fb22022-06-16 10:22:12 +01002476 kfree(ctx->cancel_table_locked.hbs);
Pavel Begunkov62248432021-04-28 13:11:29 +01002477 kfree(ctx->dummy_ubuf);
Jens Axboe9cfc7e92022-05-01 10:52:44 -06002478 kfree(ctx->io_bl);
2479 xa_destroy(&ctx->io_bl_xa);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002480 kfree(ctx);
2481}
2482
2483static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2484{
2485 struct io_ring_ctx *ctx = file->private_data;
2486 __poll_t mask = 0;
2487
Pavel Begunkovd60aa652021-10-04 20:02:52 +01002488 poll_wait(file, &ctx->cq_wait, wait);
Stefan Bühler4f7067c2019-04-24 23:54:17 +02002489 /*
2490 * synchronizes with barrier from wq_has_sleeper call in
2491 * io_commit_cqring
2492 */
Jens Axboe2b188cc2019-01-07 10:46:33 -07002493 smp_rmb();
Jens Axboe90554202020-09-03 12:12:41 -06002494 if (!io_sqring_full(ctx))
Jens Axboe2b188cc2019-01-07 10:46:33 -07002495 mask |= EPOLLOUT | EPOLLWRNORM;
Hao Xued670c32021-02-05 16:34:21 +08002496
2497 /*
2498 * Don't flush cqring overflow list here, just do a simple check.
2499 * Otherwise there could possible be ABBA deadlock:
2500 * CPU0 CPU1
2501 * ---- ----
2502 * lock(&ctx->uring_lock);
2503 * lock(&ep->mtx);
2504 * lock(&ctx->uring_lock);
2505 * lock(&ep->mtx);
2506 *
2507 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2508 * pushs them to do the flush.
2509 */
Dylan Yudaken10988a02022-04-21 02:13:43 -07002510 if (io_cqring_events(ctx) ||
2511 test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
Jens Axboe2b188cc2019-01-07 10:46:33 -07002512 mask |= EPOLLIN | EPOLLRDNORM;
2513
2514 return mask;
2515}
2516
Yejune Deng0bead8c2020-12-24 11:02:20 +08002517static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
Jens Axboe071698e2020-01-28 10:04:42 -07002518{
Jens Axboe4379bf82021-02-15 13:40:22 -07002519 const struct cred *creds;
Jens Axboe071698e2020-01-28 10:04:42 -07002520
Matthew Wilcox (Oracle)61cf9372021-03-08 14:16:16 +00002521 creds = xa_erase(&ctx->personalities, id);
Jens Axboe4379bf82021-02-15 13:40:22 -07002522 if (creds) {
2523 put_cred(creds);
Yejune Deng0bead8c2020-12-24 11:02:20 +08002524 return 0;
Jens Axboe1e6fa522020-10-15 08:46:24 -06002525 }
Yejune Deng0bead8c2020-12-24 11:02:20 +08002526
2527 return -EINVAL;
2528}
2529
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002530struct io_tctx_exit {
2531 struct callback_head task_work;
2532 struct completion completion;
Pavel Begunkovbaf186c2021-03-06 11:02:15 +00002533 struct io_ring_ctx *ctx;
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002534};
2535
Pavel Begunkovc0724812021-10-04 20:02:54 +01002536static __cold void io_tctx_exit_cb(struct callback_head *cb)
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002537{
2538 struct io_uring_task *tctx = current->io_uring;
2539 struct io_tctx_exit *work;
2540
2541 work = container_of(cb, struct io_tctx_exit, task_work);
2542 /*
2543 * When @in_idle, we're in cancellation and it's racy to remove the
2544 * node. It'll be removed by the end of cancellation, just ignore it.
2545 */
2546 if (!atomic_read(&tctx->in_idle))
Pavel Begunkoveef51da2021-06-14 02:36:15 +01002547 io_uring_del_tctx_node((unsigned long)work->ctx);
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002548 complete(&work->completion);
2549}
2550
Pavel Begunkovc0724812021-10-04 20:02:54 +01002551static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
Pavel Begunkov28090c12021-04-25 23:34:45 +01002552{
2553 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2554
2555 return req->ctx == data;
2556}
2557
Pavel Begunkovc0724812021-10-04 20:02:54 +01002558static __cold void io_ring_exit_work(struct work_struct *work)
Jens Axboe85faa7b2020-04-09 18:14:00 -06002559{
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002560 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
Pavel Begunkovb5bb3a22021-03-06 11:02:16 +00002561 unsigned long timeout = jiffies + HZ * 60 * 5;
Pavel Begunkov58d3be22021-08-09 13:04:17 +01002562 unsigned long interval = HZ / 20;
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002563 struct io_tctx_exit exit;
2564 struct io_tctx_node *node;
2565 int ret;
Jens Axboe85faa7b2020-04-09 18:14:00 -06002566
Jens Axboe56952e92020-06-17 15:00:04 -06002567 /*
2568 * If we're doing polled IO and end up having requests being
2569 * submitted async (out-of-line), then completions can come in while
2570 * we're waiting for refs to drop. We need to reap these manually,
2571 * as nobody else will be looking for them.
2572 */
Pavel Begunkovb2edc0a2020-07-07 16:36:22 +03002573 do {
Pavel Begunkovaffa87d2022-06-20 01:25:52 +01002574 while (io_uring_try_cancel_requests(ctx, NULL, true))
2575 cond_resched();
2576
Pavel Begunkov28090c12021-04-25 23:34:45 +01002577 if (ctx->sq_data) {
2578 struct io_sq_data *sqd = ctx->sq_data;
2579 struct task_struct *tsk;
2580
2581 io_sq_thread_park(sqd);
2582 tsk = sqd->thread;
2583 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2584 io_wq_cancel_cb(tsk->io_uring->io_wq,
2585 io_cancel_ctx_cb, ctx, true);
2586 io_sq_thread_unpark(sqd);
2587 }
Pavel Begunkovb5bb3a22021-03-06 11:02:16 +00002588
Pavel Begunkov37f0e762021-10-04 20:02:53 +01002589 io_req_caches_free(ctx);
2590
Pavel Begunkov58d3be22021-08-09 13:04:17 +01002591 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2592 /* there is little hope left, don't run it too often */
2593 interval = HZ * 60;
2594 }
2595 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002596
Pavel Begunkov7f006512021-04-14 13:38:34 +01002597 init_completion(&exit.completion);
2598 init_task_work(&exit.task_work, io_tctx_exit_cb);
2599 exit.ctx = ctx;
Pavel Begunkov89b50662021-04-01 15:43:50 +01002600 /*
2601 * Some may use context even when all refs and requests have been put,
2602 * and they are free to do so while still holding uring_lock or
Pavel Begunkov5b0a6ac2021-06-30 21:54:04 +01002603 * completion_lock, see io_req_task_submit(). Apart from other work,
Pavel Begunkov89b50662021-04-01 15:43:50 +01002604 * this lock/unlock section also waits them to finish.
2605 */
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002606 mutex_lock(&ctx->uring_lock);
2607 while (!list_empty(&ctx->tctx_list)) {
Pavel Begunkovb5bb3a22021-03-06 11:02:16 +00002608 WARN_ON_ONCE(time_after(jiffies, timeout));
2609
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002610 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2611 ctx_node);
Pavel Begunkov7f006512021-04-14 13:38:34 +01002612 /* don't spin on a single task if cancellation failed */
2613 list_rotate_left(&ctx->tctx_list);
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002614 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2615 if (WARN_ON_ONCE(ret))
2616 continue;
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002617
2618 mutex_unlock(&ctx->uring_lock);
2619 wait_for_completion(&exit.completion);
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002620 mutex_lock(&ctx->uring_lock);
2621 }
2622 mutex_unlock(&ctx->uring_lock);
Jens Axboe79ebeae2021-08-10 15:18:27 -06002623 spin_lock(&ctx->completion_lock);
2624 spin_unlock(&ctx->completion_lock);
Pavel Begunkovd56d9382021-03-06 11:02:13 +00002625
Jens Axboe85faa7b2020-04-09 18:14:00 -06002626 io_ring_ctx_free(ctx);
2627}
2628
Pavel Begunkovc0724812021-10-04 20:02:54 +01002629static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
Jens Axboe2b188cc2019-01-07 10:46:33 -07002630{
Matthew Wilcox (Oracle)61cf9372021-03-08 14:16:16 +00002631 unsigned long index;
2632 struct creds *creds;
2633
Jens Axboe2b188cc2019-01-07 10:46:33 -07002634 mutex_lock(&ctx->uring_lock);
2635 percpu_ref_kill(&ctx->refs);
Pavel Begunkov634578f2020-12-06 22:22:44 +00002636 if (ctx->rings)
Pavel Begunkov6c2450a2021-02-23 12:40:22 +00002637 __io_cqring_overflow_flush(ctx, true);
Matthew Wilcox (Oracle)61cf9372021-03-08 14:16:16 +00002638 xa_for_each(&ctx->personalities, index, creds)
2639 io_unregister_personality(ctx, index);
Pavel Begunkov9ca9fb22022-06-16 10:22:12 +01002640 if (ctx->rings)
2641 io_poll_remove_all(ctx, NULL, true);
Pavel Begunkoveb42ceb2022-07-12 21:52:38 +01002642 io_notif_unregister(ctx);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002643 mutex_unlock(&ctx->uring_lock);
2644
Pavel Begunkov60053be2022-03-21 22:02:20 +00002645 /* failed during ring init, it couldn't have issued any requests */
2646 if (ctx->rings) {
2647 io_kill_timeouts(ctx, NULL, true);
Pavel Begunkov60053be2022-03-21 22:02:20 +00002648 /* if we failed setting up the ctx, we might not have any rings */
2649 io_iopoll_try_reap_events(ctx);
2650 }
Jens Axboe309fc032020-07-10 09:13:34 -06002651
Jens Axboe85faa7b2020-04-09 18:14:00 -06002652 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
Jens Axboefc666772020-08-19 11:10:51 -06002653 /*
2654 * Use system_unbound_wq to avoid spawning tons of event kworkers
2655 * if we're exiting a ton of rings at the same time. It just adds
2656 * noise and overhead, there's no discernable change in runtime
2657 * over using system_wq.
2658 */
2659 queue_work(system_unbound_wq, &ctx->exit_work);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002660}
2661
2662static int io_uring_release(struct inode *inode, struct file *file)
2663{
2664 struct io_ring_ctx *ctx = file->private_data;
2665
2666 file->private_data = NULL;
2667 io_ring_ctx_wait_and_kill(ctx);
2668 return 0;
2669}
2670
Pavel Begunkovf6edbab2020-11-06 13:00:26 +00002671struct io_task_cancel {
2672 struct task_struct *task;
Pavel Begunkov3dd0c972021-05-16 22:58:04 +01002673 bool all;
Pavel Begunkovf6edbab2020-11-06 13:00:26 +00002674};
Pavel Begunkov67c4d9e2020-06-15 10:24:05 +03002675
Pavel Begunkovf6edbab2020-11-06 13:00:26 +00002676static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
Jens Axboeb711d4e2020-08-16 08:23:05 -07002677{
Pavel Begunkov9a472ef2020-11-05 22:31:37 +00002678 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
Pavel Begunkovf6edbab2020-11-06 13:00:26 +00002679 struct io_task_cancel *cancel = data;
Pavel Begunkov9a472ef2020-11-05 22:31:37 +00002680
Pavel Begunkov6af3f482021-11-26 14:38:15 +00002681 return io_match_task_safe(req, cancel->task, cancel->all);
Jens Axboeb711d4e2020-08-16 08:23:05 -07002682}
2683
Pavel Begunkovc0724812021-10-04 20:02:54 +01002684static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2685 struct task_struct *task,
2686 bool cancel_all)
Pavel Begunkovb7ddce32020-09-06 00:45:14 +03002687{
Pavel Begunkove1915f72021-03-11 23:29:35 +00002688 struct io_defer_entry *de;
Pavel Begunkovb7ddce32020-09-06 00:45:14 +03002689 LIST_HEAD(list);
2690
Jens Axboe79ebeae2021-08-10 15:18:27 -06002691 spin_lock(&ctx->completion_lock);
Pavel Begunkovb7ddce32020-09-06 00:45:14 +03002692 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
Pavel Begunkov6af3f482021-11-26 14:38:15 +00002693 if (io_match_task_safe(de->req, task, cancel_all)) {
Pavel Begunkovb7ddce32020-09-06 00:45:14 +03002694 list_cut_position(&list, &ctx->defer_list, &de->list);
2695 break;
2696 }
2697 }
Jens Axboe79ebeae2021-08-10 15:18:27 -06002698 spin_unlock(&ctx->completion_lock);
Pavel Begunkove1915f72021-03-11 23:29:35 +00002699 if (list_empty(&list))
2700 return false;
Pavel Begunkovb7ddce32020-09-06 00:45:14 +03002701
2702 while (!list_empty(&list)) {
2703 de = list_first_entry(&list, struct io_defer_entry, list);
2704 list_del_init(&de->list);
Pavel Begunkovf41db2732021-02-28 22:35:12 +00002705 io_req_complete_failed(de->req, -ECANCELED);
Pavel Begunkovb7ddce32020-09-06 00:45:14 +03002706 kfree(de);
2707 }
Pavel Begunkove1915f72021-03-11 23:29:35 +00002708 return true;
Pavel Begunkovb7ddce32020-09-06 00:45:14 +03002709}
2710
Pavel Begunkovc0724812021-10-04 20:02:54 +01002711static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
Pavel Begunkov1b007642021-03-06 11:02:17 +00002712{
2713 struct io_tctx_node *node;
2714 enum io_wq_cancel cret;
2715 bool ret = false;
2716
2717 mutex_lock(&ctx->uring_lock);
2718 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2719 struct io_uring_task *tctx = node->task->io_uring;
2720
2721 /*
2722 * io_wq will stay alive while we hold uring_lock, because it's
2723 * killed after ctx nodes, which requires to take the lock.
2724 */
2725 if (!tctx || !tctx->io_wq)
2726 continue;
2727 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2728 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2729 }
2730 mutex_unlock(&ctx->uring_lock);
2731
2732 return ret;
2733}
2734
Pavel Begunkovaffa87d2022-06-20 01:25:52 +01002735static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
Pavel Begunkovc0724812021-10-04 20:02:54 +01002736 struct task_struct *task,
2737 bool cancel_all)
Pavel Begunkov9936c7c2021-02-04 13:51:56 +00002738{
Pavel Begunkov3dd0c972021-05-16 22:58:04 +01002739 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
Pavel Begunkov1b007642021-03-06 11:02:17 +00002740 struct io_uring_task *tctx = task ? task->io_uring : NULL;
Pavel Begunkovaffa87d2022-06-20 01:25:52 +01002741 enum io_wq_cancel cret;
2742 bool ret = false;
Pavel Begunkov9936c7c2021-02-04 13:51:56 +00002743
Pavel Begunkov60053be2022-03-21 22:02:20 +00002744 /* failed during ring init, it couldn't have issued any requests */
2745 if (!ctx->rings)
Pavel Begunkovaffa87d2022-06-20 01:25:52 +01002746 return false;
Pavel Begunkov60053be2022-03-21 22:02:20 +00002747
Pavel Begunkovaffa87d2022-06-20 01:25:52 +01002748 if (!task) {
2749 ret |= io_uring_try_cancel_iowq(ctx);
2750 } else if (tctx && tctx->io_wq) {
2751 /*
2752 * Cancels requests of all rings, not only @ctx, but
2753 * it's fine as the task is in exit/exec.
2754 */
2755 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2756 &cancel, true);
2757 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
Pavel Begunkov9936c7c2021-02-04 13:51:56 +00002758 }
Pavel Begunkovaffa87d2022-06-20 01:25:52 +01002759
2760 /* SQPOLL thread does its own polling */
2761 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2762 (ctx->sq_data && ctx->sq_data->thread == current)) {
2763 while (!wq_list_empty(&ctx->iopoll_list)) {
2764 io_iopoll_try_reap_events(ctx);
2765 ret = true;
2766 }
2767 }
2768
2769 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2770 mutex_lock(&ctx->uring_lock);
2771 ret |= io_poll_remove_all(ctx, task, cancel_all);
2772 mutex_unlock(&ctx->uring_lock);
2773 ret |= io_kill_timeouts(ctx, task, cancel_all);
2774 if (task)
2775 ret |= io_run_task_work();
2776 return ret;
Pavel Begunkov9936c7c2021-02-04 13:51:56 +00002777}
2778
Pavel Begunkov3f48cf12021-04-11 01:46:27 +01002779static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
Pavel Begunkov521d6a72021-03-11 23:29:38 +00002780{
Pavel Begunkov3f48cf12021-04-11 01:46:27 +01002781 if (tracked)
Jens Axboe9cae36a2022-06-01 23:57:02 -06002782 return atomic_read(&tctx->inflight_tracked);
Pavel Begunkov521d6a72021-03-11 23:29:38 +00002783 return percpu_counter_sum(&tctx->inflight);
2784}
2785
Pavel Begunkov78cc6872021-06-14 02:36:23 +01002786/*
2787 * Find any io_uring ctx that this task has registered or done IO on, and cancel
Jens Axboe78a78062021-12-09 08:54:29 -07002788 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
Pavel Begunkov78cc6872021-06-14 02:36:23 +01002789 */
Jens Axboe17437f32022-05-25 09:13:39 -06002790__cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
Pavel Begunkov0e9ddb32021-02-07 22:34:26 +00002791{
Pavel Begunkov521d6a72021-03-11 23:29:38 +00002792 struct io_uring_task *tctx = current->io_uring;
Pavel Begunkov734551d2021-04-18 14:52:09 +01002793 struct io_ring_ctx *ctx;
Jens Axboefdaf0832020-10-30 09:37:30 -06002794 s64 inflight;
Pavel Begunkov0e9ddb32021-02-07 22:34:26 +00002795 DEFINE_WAIT(wait);
Jens Axboefdaf0832020-10-30 09:37:30 -06002796
Pavel Begunkov78cc6872021-06-14 02:36:23 +01002797 WARN_ON_ONCE(sqd && sqd->thread != current);
2798
Palash Oswal6d042ff2021-04-27 18:21:49 +05302799 if (!current->io_uring)
2800 return;
Pavel Begunkov17a91052021-05-23 15:48:39 +01002801 if (tctx->io_wq)
2802 io_wq_exit_start(tctx->io_wq);
2803
Jens Axboefdaf0832020-10-30 09:37:30 -06002804 atomic_inc(&tctx->in_idle);
Jens Axboed8a6df12020-10-15 16:24:45 -06002805 do {
Pavel Begunkovaffa87d2022-06-20 01:25:52 +01002806 bool loop = false;
2807
Pavel Begunkove9dbe222021-08-09 13:04:20 +01002808 io_uring_drop_tctx_refs(current);
Jens Axboe0f212202020-09-13 13:09:39 -06002809 /* read completions before cancelations */
Pavel Begunkov3dd0c972021-05-16 22:58:04 +01002810 inflight = tctx_inflight(tctx, !cancel_all);
Jens Axboed8a6df12020-10-15 16:24:45 -06002811 if (!inflight)
2812 break;
Jens Axboe0f212202020-09-13 13:09:39 -06002813
Pavel Begunkov78cc6872021-06-14 02:36:23 +01002814 if (!sqd) {
2815 struct io_tctx_node *node;
2816 unsigned long index;
2817
2818 xa_for_each(&tctx->xa, index, node) {
2819 /* sqpoll task will cancel all its requests */
2820 if (node->ctx->sq_data)
2821 continue;
Pavel Begunkovaffa87d2022-06-20 01:25:52 +01002822 loop |= io_uring_try_cancel_requests(node->ctx,
2823 current, cancel_all);
Pavel Begunkov78cc6872021-06-14 02:36:23 +01002824 }
2825 } else {
2826 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
Pavel Begunkovaffa87d2022-06-20 01:25:52 +01002827 loop |= io_uring_try_cancel_requests(ctx,
2828 current,
2829 cancel_all);
2830 }
2831
2832 if (loop) {
2833 cond_resched();
2834 continue;
Pavel Begunkov78cc6872021-06-14 02:36:23 +01002835 }
2836
Jens Axboe78a78062021-12-09 08:54:29 -07002837 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
2838 io_run_task_work();
Pavel Begunkove9dbe222021-08-09 13:04:20 +01002839 io_uring_drop_tctx_refs(current);
Jens Axboe78a78062021-12-09 08:54:29 -07002840
Jens Axboe0f212202020-09-13 13:09:39 -06002841 /*
Pavel Begunkova1bb3cd2021-01-26 15:28:26 +00002842 * If we've seen completions, retry without waiting. This
2843 * avoids a race where a completion comes in before we did
2844 * prepare_to_wait().
Jens Axboe0f212202020-09-13 13:09:39 -06002845 */
Pavel Begunkov3dd0c972021-05-16 22:58:04 +01002846 if (inflight == tctx_inflight(tctx, !cancel_all))
Pavel Begunkova1bb3cd2021-01-26 15:28:26 +00002847 schedule();
Pavel Begunkovf57555e2020-12-20 13:21:44 +00002848 finish_wait(&tctx->wait, &wait);
Jens Axboed8a6df12020-10-15 16:24:45 -06002849 } while (1);
Pavel Begunkovde7f1d92021-01-04 20:43:29 +00002850
Pavel Begunkov8452d4a2021-02-27 11:16:46 +00002851 io_uring_clean_tctx(tctx);
Pavel Begunkov3dd0c972021-05-16 22:58:04 +01002852 if (cancel_all) {
Pavel Begunkov3cc7fdb2022-01-09 00:53:22 +00002853 /*
2854 * We shouldn't run task_works after cancel, so just leave
2855 * ->in_idle set for normal exit.
2856 */
2857 atomic_dec(&tctx->in_idle);
Pavel Begunkov3f48cf12021-04-11 01:46:27 +01002858 /* for exec all current's requests should be gone, kill tctx */
2859 __io_uring_free(current);
2860 }
Pavel Begunkov44e728b2020-06-15 10:24:04 +03002861}
2862
Hao Xuf552a272021-08-12 12:14:35 +08002863void __io_uring_cancel(bool cancel_all)
Pavel Begunkov78cc6872021-06-14 02:36:23 +01002864{
Hao Xuf552a272021-08-12 12:14:35 +08002865 io_uring_cancel_generic(cancel_all, NULL);
Pavel Begunkov78cc6872021-06-14 02:36:23 +01002866}
2867
Roman Penyaev6c5c2402019-11-28 12:53:22 +01002868static void *io_uring_validate_mmap_request(struct file *file,
2869 loff_t pgoff, size_t sz)
Jens Axboe2b188cc2019-01-07 10:46:33 -07002870{
Jens Axboe2b188cc2019-01-07 10:46:33 -07002871 struct io_ring_ctx *ctx = file->private_data;
Roman Penyaev6c5c2402019-11-28 12:53:22 +01002872 loff_t offset = pgoff << PAGE_SHIFT;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002873 struct page *page;
2874 void *ptr;
2875
2876 switch (offset) {
2877 case IORING_OFF_SQ_RING:
Hristo Venev75b28af2019-08-26 17:23:46 +00002878 case IORING_OFF_CQ_RING:
2879 ptr = ctx->rings;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002880 break;
2881 case IORING_OFF_SQES:
2882 ptr = ctx->sq_sqes;
2883 break;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002884 default:
Roman Penyaev6c5c2402019-11-28 12:53:22 +01002885 return ERR_PTR(-EINVAL);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002886 }
2887
2888 page = virt_to_head_page(ptr);
Matthew Wilcox (Oracle)a50b8542019-09-23 15:34:25 -07002889 if (sz > page_size(page))
Roman Penyaev6c5c2402019-11-28 12:53:22 +01002890 return ERR_PTR(-EINVAL);
2891
2892 return ptr;
2893}
2894
2895#ifdef CONFIG_MMU
2896
Pavel Begunkovc0724812021-10-04 20:02:54 +01002897static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
Roman Penyaev6c5c2402019-11-28 12:53:22 +01002898{
2899 size_t sz = vma->vm_end - vma->vm_start;
2900 unsigned long pfn;
2901 void *ptr;
2902
2903 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
2904 if (IS_ERR(ptr))
2905 return PTR_ERR(ptr);
Jens Axboe2b188cc2019-01-07 10:46:33 -07002906
2907 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2908 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2909}
2910
Roman Penyaev6c5c2402019-11-28 12:53:22 +01002911#else /* !CONFIG_MMU */
2912
2913static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2914{
2915 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
2916}
2917
2918static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
2919{
2920 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
2921}
2922
2923static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
2924 unsigned long addr, unsigned long len,
2925 unsigned long pgoff, unsigned long flags)
2926{
2927 void *ptr;
2928
2929 ptr = io_uring_validate_mmap_request(file, pgoff, len);
2930 if (IS_ERR(ptr))
2931 return PTR_ERR(ptr);
2932
2933 return (unsigned long) ptr;
2934}
2935
2936#endif /* !CONFIG_MMU */
2937
Pavel Begunkovf81440d2022-03-22 14:07:56 +00002938static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
2939{
2940 if (flags & IORING_ENTER_EXT_ARG) {
2941 struct io_uring_getevents_arg arg;
2942
2943 if (argsz != sizeof(arg))
2944 return -EINVAL;
2945 if (copy_from_user(&arg, argp, sizeof(arg)))
2946 return -EFAULT;
2947 }
2948 return 0;
2949}
2950
Hao Xuc73ebb62020-11-03 10:54:37 +08002951static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
2952 struct __kernel_timespec __user **ts,
2953 const sigset_t __user **sig)
2954{
2955 struct io_uring_getevents_arg arg;
2956
2957 /*
2958 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
2959 * is just a pointer to the sigset_t.
2960 */
2961 if (!(flags & IORING_ENTER_EXT_ARG)) {
2962 *sig = (const sigset_t __user *) argp;
2963 *ts = NULL;
2964 return 0;
2965 }
2966
2967 /*
2968 * EXT_ARG is set - ensure we agree on the size of it and copy in our
2969 * timespec and sigset_t pointers if good.
2970 */
2971 if (*argsz != sizeof(arg))
2972 return -EINVAL;
2973 if (copy_from_user(&arg, argp, sizeof(arg)))
2974 return -EFAULT;
Dylan Yudakend2347b92022-04-12 09:30:42 -07002975 if (arg.pad)
2976 return -EINVAL;
Hao Xuc73ebb62020-11-03 10:54:37 +08002977 *sig = u64_to_user_ptr(arg.sigmask);
2978 *argsz = arg.sigmask_sz;
2979 *ts = u64_to_user_ptr(arg.ts);
2980 return 0;
2981}
2982
Jens Axboe2b188cc2019-01-07 10:46:33 -07002983SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
Hao Xuc73ebb62020-11-03 10:54:37 +08002984 u32, min_complete, u32, flags, const void __user *, argp,
2985 size_t, argsz)
Jens Axboe2b188cc2019-01-07 10:46:33 -07002986{
2987 struct io_ring_ctx *ctx;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002988 struct fd f;
Pavel Begunkov33f993d2021-03-19 17:22:30 +00002989 long ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07002990
Jens Axboe4c6e2772020-07-01 11:29:10 -06002991 io_run_task_work();
Jens Axboeb41e9852020-02-17 09:52:41 -07002992
Pavel Begunkov33f993d2021-03-19 17:22:30 +00002993 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
Jens Axboee7a6c002022-03-04 08:22:22 -07002994 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
2995 IORING_ENTER_REGISTERED_RING)))
Jens Axboe2b188cc2019-01-07 10:46:33 -07002996 return -EINVAL;
2997
Jens Axboee7a6c002022-03-04 08:22:22 -07002998 /*
2999 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3000 * need only dereference our task private array to find it.
3001 */
3002 if (flags & IORING_ENTER_REGISTERED_RING) {
3003 struct io_uring_task *tctx = current->io_uring;
3004
Pavel Begunkov3273c442022-06-25 11:53:01 +01003005 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
Jens Axboee7a6c002022-03-04 08:22:22 -07003006 return -EINVAL;
3007 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3008 f.file = tctx->registered_rings[fd];
Al Viro43294902022-05-11 20:30:20 -04003009 f.flags = 0;
Pavel Begunkov3273c442022-06-25 11:53:01 +01003010 if (unlikely(!f.file))
3011 return -EBADF;
Jens Axboee7a6c002022-03-04 08:22:22 -07003012 } else {
3013 f = fdget(fd);
Pavel Begunkov3273c442022-06-25 11:53:01 +01003014 if (unlikely(!f.file))
3015 return -EBADF;
3016 ret = -EOPNOTSUPP;
3017 if (unlikely(!io_is_uring_fops(f.file)))
Pavel Begunkovfbb8bb02022-06-25 11:53:02 +01003018 goto out;
Jens Axboee7a6c002022-03-04 08:22:22 -07003019 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07003020
Jens Axboe2b188cc2019-01-07 10:46:33 -07003021 ctx = f.file->private_data;
Stefano Garzarella7e84e1c2020-08-27 16:58:31 +02003022 ret = -EBADFD;
Pavel Begunkov33f993d2021-03-19 17:22:30 +00003023 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
Stefano Garzarella7e84e1c2020-08-27 16:58:31 +02003024 goto out;
3025
Jens Axboe6c271ce2019-01-10 11:22:30 -07003026 /*
3027 * For SQ polling, the thread will do all submissions and completions.
3028 * Just return the requested submit count, and wake the thread if
3029 * we were asked to.
3030 */
Jens Axboeb2a9ead2019-09-12 14:19:16 -06003031 ret = 0;
Jens Axboe6c271ce2019-01-10 11:22:30 -07003032 if (ctx->flags & IORING_SETUP_SQPOLL) {
Pavel Begunkov90f67362021-08-09 20:18:12 +01003033 io_cqring_overflow_flush(ctx);
Pavel Begunkov89448c42020-12-17 00:24:39 +00003034
Jens Axboe21f96522021-08-14 09:04:40 -06003035 if (unlikely(ctx->sq_data->thread == NULL)) {
3036 ret = -EOWNERDEAD;
Stefan Metzmacher04147482021-03-07 11:54:29 +01003037 goto out;
Jens Axboe21f96522021-08-14 09:04:40 -06003038 }
Jens Axboe6c271ce2019-01-10 11:22:30 -07003039 if (flags & IORING_ENTER_SQ_WAKEUP)
Jens Axboe534ca6d2020-09-02 13:52:19 -06003040 wake_up(&ctx->sq_data->wait);
Pavel Begunkovd9d05212021-01-08 20:57:25 +00003041 if (flags & IORING_ENTER_SQ_WAIT) {
3042 ret = io_sqpoll_wait_sq(ctx);
3043 if (ret)
3044 goto out;
3045 }
Dylan Yudaken3e813c92022-04-21 02:13:42 -07003046 ret = to_submit;
Jens Axboeb2a9ead2019-09-12 14:19:16 -06003047 } else if (to_submit) {
Pavel Begunkoveef51da2021-06-14 02:36:15 +01003048 ret = io_uring_add_tctx_node(ctx);
Jens Axboe0f212202020-09-13 13:09:39 -06003049 if (unlikely(ret))
3050 goto out;
Pavel Begunkov7c504e652019-12-18 19:53:45 +03003051
Jens Axboe2b188cc2019-01-07 10:46:33 -07003052 mutex_lock(&ctx->uring_lock);
Dylan Yudaken3e813c92022-04-21 02:13:42 -07003053 ret = io_submit_sqes(ctx, to_submit);
3054 if (ret != to_submit) {
Pavel Begunkovd487b432022-03-22 14:07:58 +00003055 mutex_unlock(&ctx->uring_lock);
Pavel Begunkov7c504e652019-12-18 19:53:45 +03003056 goto out;
Pavel Begunkovd487b432022-03-22 14:07:58 +00003057 }
3058 if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll)
3059 goto iopoll_locked;
3060 mutex_unlock(&ctx->uring_lock);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003061 }
3062 if (flags & IORING_ENTER_GETEVENTS) {
Dylan Yudaken3e813c92022-04-21 02:13:42 -07003063 int ret2;
Pavel Begunkov773697b2022-03-22 14:07:57 +00003064 if (ctx->syscall_iopoll) {
Pavel Begunkovd487b432022-03-22 14:07:58 +00003065 /*
3066 * We disallow the app entering submit/complete with
3067 * polling, but we still need to lock the ring to
3068 * prevent racing with polled issue that got punted to
3069 * a workqueue.
3070 */
3071 mutex_lock(&ctx->uring_lock);
3072iopoll_locked:
Dylan Yudaken3e813c92022-04-21 02:13:42 -07003073 ret2 = io_validate_ext_arg(flags, argp, argsz);
3074 if (likely(!ret2)) {
3075 min_complete = min(min_complete,
3076 ctx->cq_entries);
3077 ret2 = io_iopoll_check(ctx, min_complete);
Pavel Begunkovd487b432022-03-22 14:07:58 +00003078 }
3079 mutex_unlock(&ctx->uring_lock);
Jens Axboedef596e2019-01-09 08:59:42 -07003080 } else {
Pavel Begunkovf81440d2022-03-22 14:07:56 +00003081 const sigset_t __user *sig;
3082 struct __kernel_timespec __user *ts;
3083
Dylan Yudaken3e813c92022-04-21 02:13:42 -07003084 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3085 if (likely(!ret2)) {
3086 min_complete = min(min_complete,
3087 ctx->cq_entries);
3088 ret2 = io_cqring_wait(ctx, min_complete, sig,
3089 argsz, ts);
3090 }
Jens Axboedef596e2019-01-09 08:59:42 -07003091 }
Dylan Yudaken3e813c92022-04-21 02:13:42 -07003092
Dylan Yudaken155bc952022-04-21 02:13:44 -07003093 if (!ret) {
Dylan Yudaken3e813c92022-04-21 02:13:42 -07003094 ret = ret2;
3095
Dylan Yudaken155bc952022-04-21 02:13:44 -07003096 /*
3097 * EBADR indicates that one or more CQE were dropped.
3098 * Once the user has been informed we can clear the bit
3099 * as they are obviously ok with those drops.
3100 */
3101 if (unlikely(ret2 == -EBADR))
3102 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3103 &ctx->check_cq);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003104 }
3105 }
Pavel Begunkov7c504e652019-12-18 19:53:45 +03003106out:
Al Viro43294902022-05-11 20:30:20 -04003107 fdput(f);
Dylan Yudaken3e813c92022-04-21 02:13:42 -07003108 return ret;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003109}
3110
3111static const struct file_operations io_uring_fops = {
3112 .release = io_uring_release,
3113 .mmap = io_uring_mmap,
Roman Penyaev6c5c2402019-11-28 12:53:22 +01003114#ifndef CONFIG_MMU
3115 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3116 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3117#endif
Jens Axboe2b188cc2019-01-07 10:46:33 -07003118 .poll = io_uring_poll,
Tobias Klauserbebdb652020-02-26 18:38:32 +01003119#ifdef CONFIG_PROC_FS
Jens Axboe87ce9552020-01-30 08:25:34 -07003120 .show_fdinfo = io_uring_show_fdinfo,
Tobias Klauserbebdb652020-02-26 18:38:32 +01003121#endif
Jens Axboe2b188cc2019-01-07 10:46:33 -07003122};
3123
Jens Axboe92ac8be2022-05-25 11:48:35 -06003124bool io_is_uring_fops(struct file *file)
3125{
3126 return file->f_op == &io_uring_fops;
3127}
3128
Pavel Begunkovc0724812021-10-04 20:02:54 +01003129static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3130 struct io_uring_params *p)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003131{
Hristo Venev75b28af2019-08-26 17:23:46 +00003132 struct io_rings *rings;
3133 size_t size, sq_array_offset;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003134
Jens Axboebd740482020-08-05 12:58:23 -06003135 /* make sure these are sane, as we already accounted them */
3136 ctx->sq_entries = p->sq_entries;
3137 ctx->cq_entries = p->cq_entries;
3138
Stefan Roeschbaf9cb62022-04-26 11:21:25 -07003139 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
Hristo Venev75b28af2019-08-26 17:23:46 +00003140 if (size == SIZE_MAX)
3141 return -EOVERFLOW;
3142
3143 rings = io_mem_alloc(size);
3144 if (!rings)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003145 return -ENOMEM;
3146
Hristo Venev75b28af2019-08-26 17:23:46 +00003147 ctx->rings = rings;
3148 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3149 rings->sq_ring_mask = p->sq_entries - 1;
3150 rings->cq_ring_mask = p->cq_entries - 1;
3151 rings->sq_ring_entries = p->sq_entries;
3152 rings->cq_ring_entries = p->cq_entries;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003153
Jens Axboeebdeb7c02022-03-31 19:27:52 -06003154 if (p->flags & IORING_SETUP_SQE128)
3155 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3156 else
3157 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
Jens Axboeeb065d32019-11-20 09:26:29 -07003158 if (size == SIZE_MAX) {
3159 io_mem_free(ctx->rings);
3160 ctx->rings = NULL;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003161 return -EOVERFLOW;
Jens Axboeeb065d32019-11-20 09:26:29 -07003162 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07003163
3164 ctx->sq_sqes = io_mem_alloc(size);
Jens Axboeeb065d32019-11-20 09:26:29 -07003165 if (!ctx->sq_sqes) {
3166 io_mem_free(ctx->rings);
3167 ctx->rings = NULL;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003168 return -ENOMEM;
Jens Axboeeb065d32019-11-20 09:26:29 -07003169 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07003170
Jens Axboe2b188cc2019-01-07 10:46:33 -07003171 return 0;
3172}
3173
Pavel Begunkov9faadcc2020-12-21 18:34:05 +00003174static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3175{
3176 int ret, fd;
3177
3178 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3179 if (fd < 0)
3180 return fd;
3181
Pavel Begunkov97bbdc02022-06-16 10:22:08 +01003182 ret = __io_uring_add_tctx_node(ctx, false);
Pavel Begunkov9faadcc2020-12-21 18:34:05 +00003183 if (ret) {
3184 put_unused_fd(fd);
3185 return ret;
3186 }
3187 fd_install(fd, file);
3188 return fd;
3189}
3190
Jens Axboe2b188cc2019-01-07 10:46:33 -07003191/*
3192 * Allocate an anonymous fd, this is what constitutes the application
3193 * visible backing of an io_uring instance. The application mmaps this
3194 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3195 * we have to tie this fd to a socket for file garbage collection purposes.
3196 */
Pavel Begunkov9faadcc2020-12-21 18:34:05 +00003197static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003198{
3199 struct file *file;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003200#if defined(CONFIG_UNIX)
Pavel Begunkov9faadcc2020-12-21 18:34:05 +00003201 int ret;
3202
Jens Axboe2b188cc2019-01-07 10:46:33 -07003203 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3204 &ctx->ring_sock);
3205 if (ret)
Pavel Begunkov9faadcc2020-12-21 18:34:05 +00003206 return ERR_PTR(ret);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003207#endif
3208
Paul Moore91a9ab72021-02-01 19:33:52 -05003209 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3210 O_RDWR | O_CLOEXEC, NULL);
Pavel Begunkov9faadcc2020-12-21 18:34:05 +00003211#if defined(CONFIG_UNIX)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003212 if (IS_ERR(file)) {
Pavel Begunkov9faadcc2020-12-21 18:34:05 +00003213 sock_release(ctx->ring_sock);
3214 ctx->ring_sock = NULL;
3215 } else {
3216 ctx->ring_sock->file = file;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003217 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07003218#endif
Pavel Begunkov9faadcc2020-12-21 18:34:05 +00003219 return file;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003220}
3221
Pavel Begunkovc0724812021-10-04 20:02:54 +01003222static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3223 struct io_uring_params __user *params)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003224{
Jens Axboe2b188cc2019-01-07 10:46:33 -07003225 struct io_ring_ctx *ctx;
Pavel Begunkov9faadcc2020-12-21 18:34:05 +00003226 struct file *file;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003227 int ret;
3228
Jens Axboe8110c1a2019-12-28 15:39:54 -07003229 if (!entries)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003230 return -EINVAL;
Jens Axboe8110c1a2019-12-28 15:39:54 -07003231 if (entries > IORING_MAX_ENTRIES) {
3232 if (!(p->flags & IORING_SETUP_CLAMP))
3233 return -EINVAL;
3234 entries = IORING_MAX_ENTRIES;
3235 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07003236
3237 /*
3238 * Use twice as many entries for the CQ ring. It's possible for the
3239 * application to drive a higher depth than the size of the SQ ring,
3240 * since the sqes are only used at submission time. This allows for
Jens Axboe33a107f2019-10-04 12:10:03 -06003241 * some flexibility in overcommitting a bit. If the application has
3242 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3243 * of CQ ring entries manually.
Jens Axboe2b188cc2019-01-07 10:46:33 -07003244 */
3245 p->sq_entries = roundup_pow_of_two(entries);
Jens Axboe33a107f2019-10-04 12:10:03 -06003246 if (p->flags & IORING_SETUP_CQSIZE) {
3247 /*
3248 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3249 * to a power-of-two, if it isn't already. We do NOT impose
3250 * any cq vs sq ring sizing.
3251 */
Joseph Qieb2667b32020-11-24 15:03:03 +08003252 if (!p->cq_entries)
Jens Axboe33a107f2019-10-04 12:10:03 -06003253 return -EINVAL;
Jens Axboe8110c1a2019-12-28 15:39:54 -07003254 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3255 if (!(p->flags & IORING_SETUP_CLAMP))
3256 return -EINVAL;
3257 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3258 }
Joseph Qieb2667b32020-11-24 15:03:03 +08003259 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3260 if (p->cq_entries < p->sq_entries)
3261 return -EINVAL;
Jens Axboe33a107f2019-10-04 12:10:03 -06003262 } else {
3263 p->cq_entries = 2 * p->sq_entries;
3264 }
Jens Axboe2b188cc2019-01-07 10:46:33 -07003265
Jens Axboe2b188cc2019-01-07 10:46:33 -07003266 ctx = io_ring_ctx_alloc(p);
Jens Axboe62e398b2021-02-21 16:19:37 -07003267 if (!ctx)
Jens Axboe2b188cc2019-01-07 10:46:33 -07003268 return -ENOMEM;
Pavel Begunkov773697b2022-03-22 14:07:57 +00003269
3270 /*
3271 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3272 * space applications don't need to do io completion events
3273 * polling again, they can rely on io_sq_thread to do polling
3274 * work, which can reduce cpu usage and uring_lock contention.
3275 */
3276 if (ctx->flags & IORING_SETUP_IOPOLL &&
3277 !(ctx->flags & IORING_SETUP_SQPOLL))
3278 ctx->syscall_iopoll = 1;
3279
Jens Axboe2b188cc2019-01-07 10:46:33 -07003280 ctx->compat = in_compat_syscall();
Jens Axboe62e398b2021-02-21 16:19:37 -07003281 if (!capable(CAP_IPC_LOCK))
3282 ctx->user = get_uid(current_user());
Jens Axboe2aede0e2020-09-14 10:45:53 -06003283
3284 /*
Jens Axboee1169f02022-04-25 19:49:03 -06003285 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3286 * COOP_TASKRUN is set, then IPIs are never needed by the app.
Jens Axboe9f010502022-04-25 19:49:02 -06003287 */
Jens Axboee1169f02022-04-25 19:49:03 -06003288 ret = -EINVAL;
3289 if (ctx->flags & IORING_SETUP_SQPOLL) {
3290 /* IPI related flags don't make sense with SQPOLL */
Jens Axboeef060ea2022-04-25 19:49:04 -06003291 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3292 IORING_SETUP_TASKRUN_FLAG))
Jens Axboee1169f02022-04-25 19:49:03 -06003293 goto err;
Jens Axboe9f010502022-04-25 19:49:02 -06003294 ctx->notify_method = TWA_SIGNAL_NO_IPI;
Jens Axboee1169f02022-04-25 19:49:03 -06003295 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3296 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3297 } else {
Jens Axboeef060ea2022-04-25 19:49:04 -06003298 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
3299 goto err;
Jens Axboe9f010502022-04-25 19:49:02 -06003300 ctx->notify_method = TWA_SIGNAL;
Jens Axboee1169f02022-04-25 19:49:03 -06003301 }
Jens Axboe9f010502022-04-25 19:49:02 -06003302
3303 /*
Jens Axboe2aede0e2020-09-14 10:45:53 -06003304 * This is just grabbed for accounting purposes. When a process exits,
3305 * the mm is exited and dropped before the files, hence we need to hang
3306 * on to this mm purely for the purposes of being able to unaccount
3307 * memory (locked/pinned vm). It's not used for anything else.
3308 */
Jens Axboe6b7898e2020-08-25 07:58:00 -06003309 mmgrab(current->mm);
Jens Axboe2aede0e2020-09-14 10:45:53 -06003310 ctx->mm_account = current->mm;
Jens Axboe6b7898e2020-08-25 07:58:00 -06003311
Jens Axboe2b188cc2019-01-07 10:46:33 -07003312 ret = io_allocate_scq_urings(ctx, p);
3313 if (ret)
3314 goto err;
3315
Stefano Garzarella7e84e1c2020-08-27 16:58:31 +02003316 ret = io_sq_offload_create(ctx, p);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003317 if (ret)
3318 goto err;
Pavel Begunkoveae071c2021-04-25 14:32:24 +01003319 /* always set a rsrc node */
Pavel Begunkov47b228c2021-04-29 11:46:48 +01003320 ret = io_rsrc_node_switch_start(ctx);
3321 if (ret)
3322 goto err;
Pavel Begunkoveae071c2021-04-25 14:32:24 +01003323 io_rsrc_node_switch(ctx, NULL);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003324
Jens Axboe2b188cc2019-01-07 10:46:33 -07003325 memset(&p->sq_off, 0, sizeof(p->sq_off));
Hristo Venev75b28af2019-08-26 17:23:46 +00003326 p->sq_off.head = offsetof(struct io_rings, sq.head);
3327 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3328 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3329 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3330 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3331 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3332 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003333
3334 memset(&p->cq_off, 0, sizeof(p->cq_off));
Hristo Venev75b28af2019-08-26 17:23:46 +00003335 p->cq_off.head = offsetof(struct io_rings, cq.head);
3336 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3337 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3338 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3339 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3340 p->cq_off.cqes = offsetof(struct io_rings, cqes);
Stefano Garzarella0d9b5b32020-05-15 18:38:04 +02003341 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
Jens Axboeac90f242019-09-06 10:26:21 -06003342
Xiaoguang Wang7f136572020-05-05 16:28:53 +08003343 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3344 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
Jiufei Xue5769a352020-06-17 17:53:55 +08003345 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
Hao Xuc73ebb62020-11-03 10:54:37 +08003346 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
Pavel Begunkov96905572021-06-10 16:37:38 +01003347 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
Jens Axboec4212f32022-04-10 15:13:24 -06003348 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3349 IORING_FEAT_LINKED_FILE;
Xiaoguang Wang7f136572020-05-05 16:28:53 +08003350
3351 if (copy_to_user(params, p, sizeof(*p))) {
3352 ret = -EFAULT;
3353 goto err;
3354 }
Jens Axboed1719f72020-07-30 13:43:53 -06003355
Pavel Begunkov9faadcc2020-12-21 18:34:05 +00003356 file = io_uring_get_file(ctx);
3357 if (IS_ERR(file)) {
3358 ret = PTR_ERR(file);
3359 goto err;
3360 }
3361
Jens Axboed1719f72020-07-30 13:43:53 -06003362 /*
Jens Axboe044c1ab2019-10-28 09:15:33 -06003363 * Install ring fd as the very last thing, so we don't risk someone
3364 * having closed it before we finish setup
3365 */
Pavel Begunkov9faadcc2020-12-21 18:34:05 +00003366 ret = io_uring_install_fd(ctx, file);
3367 if (ret < 0) {
3368 /* fput will clean it up */
3369 fput(file);
3370 return ret;
3371 }
Jens Axboe044c1ab2019-10-28 09:15:33 -06003372
Dmitrii Dolgovc826bd72019-10-15 19:02:01 +02003373 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003374 return ret;
3375err:
3376 io_ring_ctx_wait_and_kill(ctx);
3377 return ret;
3378}
3379
3380/*
3381 * Sets up an aio uring context, and returns the fd. Applications asks for a
3382 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3383 * params structure passed in.
3384 */
3385static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3386{
3387 struct io_uring_params p;
Jens Axboe2b188cc2019-01-07 10:46:33 -07003388 int i;
3389
3390 if (copy_from_user(&p, params, sizeof(p)))
3391 return -EFAULT;
3392 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3393 if (p.resv[i])
3394 return -EINVAL;
3395 }
3396
Jens Axboe6c271ce2019-01-10 11:22:30 -07003397 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
Jens Axboe8110c1a2019-12-28 15:39:54 -07003398 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
Stefano Garzarella7e84e1c2020-08-27 16:58:31 +02003399 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
Jens Axboee1169f02022-04-25 19:49:03 -06003400 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
Jens Axboeebdeb7c02022-03-31 19:27:52 -06003401 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
Pavel Begunkov97bbdc02022-06-16 10:22:08 +01003402 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3403 IORING_SETUP_SINGLE_ISSUER))
Jens Axboe2b188cc2019-01-07 10:46:33 -07003404 return -EINVAL;
3405
Jens Axboeef060ea2022-04-25 19:49:04 -06003406 return io_uring_create(entries, &p, params);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003407}
3408
3409SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3410 struct io_uring_params __user *, params)
3411{
3412 return io_uring_setup(entries, params);
3413}
3414
Pavel Begunkovc0724812021-10-04 20:02:54 +01003415static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3416 unsigned nr_args)
Jens Axboe66f4af92020-01-16 15:36:52 -07003417{
3418 struct io_uring_probe *p;
3419 size_t size;
3420 int i, ret;
3421
3422 size = struct_size(p, ops, nr_args);
3423 if (size == SIZE_MAX)
3424 return -EOVERFLOW;
3425 p = kzalloc(size, GFP_KERNEL);
3426 if (!p)
3427 return -ENOMEM;
3428
3429 ret = -EFAULT;
3430 if (copy_from_user(p, arg, size))
3431 goto out;
3432 ret = -EINVAL;
3433 if (memchr_inv(p, 0, size))
3434 goto out;
3435
3436 p->last_op = IORING_OP_LAST - 1;
3437 if (nr_args > IORING_OP_LAST)
3438 nr_args = IORING_OP_LAST;
3439
3440 for (i = 0; i < nr_args; i++) {
3441 p->ops[i].op = i;
3442 if (!io_op_defs[i].not_supported)
3443 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3444 }
3445 p->ops_len = i;
3446
3447 ret = 0;
3448 if (copy_to_user(arg, p, size))
3449 ret = -EFAULT;
3450out:
3451 kfree(p);
3452 return ret;
3453}
3454
Jens Axboe071698e2020-01-28 10:04:42 -07003455static int io_register_personality(struct io_ring_ctx *ctx)
3456{
Jens Axboe4379bf82021-02-15 13:40:22 -07003457 const struct cred *creds;
Matthew Wilcox (Oracle)61cf9372021-03-08 14:16:16 +00003458 u32 id;
Jens Axboe1e6fa522020-10-15 08:46:24 -06003459 int ret;
Jens Axboe071698e2020-01-28 10:04:42 -07003460
Jens Axboe4379bf82021-02-15 13:40:22 -07003461 creds = get_current_cred();
Jens Axboe1e6fa522020-10-15 08:46:24 -06003462
Matthew Wilcox (Oracle)61cf9372021-03-08 14:16:16 +00003463 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3464 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
Jens Axboea30f8952021-08-20 14:53:59 -06003465 if (ret < 0) {
3466 put_cred(creds);
3467 return ret;
3468 }
3469 return id;
Jens Axboe071698e2020-01-28 10:04:42 -07003470}
3471
Pavel Begunkovc0724812021-10-04 20:02:54 +01003472static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3473 void __user *arg, unsigned int nr_args)
Stefano Garzarella21b55db2020-08-27 16:58:30 +02003474{
3475 struct io_uring_restriction *res;
3476 size_t size;
3477 int i, ret;
3478
Stefano Garzarella7e84e1c2020-08-27 16:58:31 +02003479 /* Restrictions allowed only if rings started disabled */
3480 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3481 return -EBADFD;
3482
Stefano Garzarella21b55db2020-08-27 16:58:30 +02003483 /* We allow only a single restrictions registration */
Stefano Garzarella7e84e1c2020-08-27 16:58:31 +02003484 if (ctx->restrictions.registered)
Stefano Garzarella21b55db2020-08-27 16:58:30 +02003485 return -EBUSY;
3486
3487 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3488 return -EINVAL;
3489
3490 size = array_size(nr_args, sizeof(*res));
3491 if (size == SIZE_MAX)
3492 return -EOVERFLOW;
3493
3494 res = memdup_user(arg, size);
3495 if (IS_ERR(res))
3496 return PTR_ERR(res);
3497
3498 ret = 0;
3499
3500 for (i = 0; i < nr_args; i++) {
3501 switch (res[i].opcode) {
3502 case IORING_RESTRICTION_REGISTER_OP:
3503 if (res[i].register_op >= IORING_REGISTER_LAST) {
3504 ret = -EINVAL;
3505 goto out;
3506 }
3507
3508 __set_bit(res[i].register_op,
3509 ctx->restrictions.register_op);
3510 break;
3511 case IORING_RESTRICTION_SQE_OP:
3512 if (res[i].sqe_op >= IORING_OP_LAST) {
3513 ret = -EINVAL;
3514 goto out;
3515 }
3516
3517 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3518 break;
3519 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3520 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3521 break;
3522 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3523 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3524 break;
3525 default:
3526 ret = -EINVAL;
3527 goto out;
3528 }
3529 }
3530
3531out:
3532 /* Reset all restrictions if an error happened */
3533 if (ret != 0)
3534 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3535 else
Stefano Garzarella7e84e1c2020-08-27 16:58:31 +02003536 ctx->restrictions.registered = true;
Stefano Garzarella21b55db2020-08-27 16:58:30 +02003537
3538 kfree(res);
3539 return ret;
3540}
3541
Stefano Garzarella7e84e1c2020-08-27 16:58:31 +02003542static int io_register_enable_rings(struct io_ring_ctx *ctx)
3543{
3544 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3545 return -EBADFD;
3546
3547 if (ctx->restrictions.registered)
3548 ctx->restricted = 1;
3549
Pavel Begunkov0298ef92021-03-08 13:20:57 +00003550 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3551 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3552 wake_up(&ctx->sq_data->wait);
Stefano Garzarella7e84e1c2020-08-27 16:58:31 +02003553 return 0;
3554}
3555
Pavel Begunkovc0724812021-10-04 20:02:54 +01003556static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3557 void __user *arg, unsigned len)
Jens Axboefe764212021-06-17 10:19:54 -06003558{
3559 struct io_uring_task *tctx = current->io_uring;
3560 cpumask_var_t new_mask;
3561 int ret;
3562
3563 if (!tctx || !tctx->io_wq)
3564 return -EINVAL;
3565
3566 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3567 return -ENOMEM;
3568
3569 cpumask_clear(new_mask);
3570 if (len > cpumask_size())
3571 len = cpumask_size();
3572
Eugene Syromiatnikov0f5e4b82022-04-06 13:55:33 +02003573 if (in_compat_syscall()) {
3574 ret = compat_get_bitmap(cpumask_bits(new_mask),
3575 (const compat_ulong_t __user *)arg,
3576 len * 8 /* CHAR_BIT */);
3577 } else {
3578 ret = copy_from_user(new_mask, arg, len);
3579 }
3580
3581 if (ret) {
Jens Axboefe764212021-06-17 10:19:54 -06003582 free_cpumask_var(new_mask);
3583 return -EFAULT;
3584 }
3585
3586 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3587 free_cpumask_var(new_mask);
3588 return ret;
3589}
3590
Pavel Begunkovc0724812021-10-04 20:02:54 +01003591static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
Jens Axboefe764212021-06-17 10:19:54 -06003592{
3593 struct io_uring_task *tctx = current->io_uring;
3594
3595 if (!tctx || !tctx->io_wq)
3596 return -EINVAL;
3597
3598 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3599}
3600
Pavel Begunkovc0724812021-10-04 20:02:54 +01003601static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3602 void __user *arg)
Pavel Begunkovb22fa622021-10-21 13:20:29 +01003603 __must_hold(&ctx->uring_lock)
Jens Axboe2e480052021-08-27 11:33:19 -06003604{
Pavel Begunkovb22fa622021-10-21 13:20:29 +01003605 struct io_tctx_node *node;
Jens Axboefa846932021-09-01 14:15:59 -06003606 struct io_uring_task *tctx = NULL;
3607 struct io_sq_data *sqd = NULL;
Jens Axboe2e480052021-08-27 11:33:19 -06003608 __u32 new_count[2];
3609 int i, ret;
3610
Jens Axboe2e480052021-08-27 11:33:19 -06003611 if (copy_from_user(new_count, arg, sizeof(new_count)))
3612 return -EFAULT;
3613 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3614 if (new_count[i] > INT_MAX)
3615 return -EINVAL;
3616
Jens Axboefa846932021-09-01 14:15:59 -06003617 if (ctx->flags & IORING_SETUP_SQPOLL) {
3618 sqd = ctx->sq_data;
3619 if (sqd) {
Jens Axboe009ad9f2021-09-08 19:07:26 -06003620 /*
3621 * Observe the correct sqd->lock -> ctx->uring_lock
3622 * ordering. Fine to drop uring_lock here, we hold
3623 * a ref to the ctx.
3624 */
Jens Axboe41d3a6b2021-09-13 13:08:51 -06003625 refcount_inc(&sqd->refs);
Jens Axboe009ad9f2021-09-08 19:07:26 -06003626 mutex_unlock(&ctx->uring_lock);
Jens Axboefa846932021-09-01 14:15:59 -06003627 mutex_lock(&sqd->lock);
Jens Axboe009ad9f2021-09-08 19:07:26 -06003628 mutex_lock(&ctx->uring_lock);
Jens Axboe41d3a6b2021-09-13 13:08:51 -06003629 if (sqd->thread)
3630 tctx = sqd->thread->io_uring;
Jens Axboefa846932021-09-01 14:15:59 -06003631 }
3632 } else {
3633 tctx = current->io_uring;
3634 }
3635
Pavel Begunkove139a1e2021-10-19 23:43:46 +01003636 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
Jens Axboefa846932021-09-01 14:15:59 -06003637
Pavel Begunkovbad119b2021-11-08 15:10:03 +00003638 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3639 if (new_count[i])
3640 ctx->iowq_limits[i] = new_count[i];
Pavel Begunkove139a1e2021-10-19 23:43:46 +01003641 ctx->iowq_limits_set = true;
3642
Pavel Begunkove139a1e2021-10-19 23:43:46 +01003643 if (tctx && tctx->io_wq) {
3644 ret = io_wq_max_workers(tctx->io_wq, new_count);
3645 if (ret)
3646 goto err;
3647 } else {
3648 memset(new_count, 0, sizeof(new_count));
3649 }
Jens Axboefa846932021-09-01 14:15:59 -06003650
Jens Axboe41d3a6b2021-09-13 13:08:51 -06003651 if (sqd) {
Jens Axboefa846932021-09-01 14:15:59 -06003652 mutex_unlock(&sqd->lock);
Jens Axboe41d3a6b2021-09-13 13:08:51 -06003653 io_put_sq_data(sqd);
3654 }
Jens Axboe2e480052021-08-27 11:33:19 -06003655
3656 if (copy_to_user(arg, new_count, sizeof(new_count)))
3657 return -EFAULT;
3658
Pavel Begunkovb22fa622021-10-21 13:20:29 +01003659 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3660 if (sqd)
3661 return 0;
3662
3663 /* now propagate the restriction to all registered users */
3664 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3665 struct io_uring_task *tctx = node->task->io_uring;
3666
3667 if (WARN_ON_ONCE(!tctx->io_wq))
3668 continue;
3669
3670 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3671 new_count[i] = ctx->iowq_limits[i];
3672 /* ignore errors, it always returns zero anyway */
3673 (void)io_wq_max_workers(tctx->io_wq, new_count);
3674 }
Jens Axboe2e480052021-08-27 11:33:19 -06003675 return 0;
Jens Axboefa846932021-09-01 14:15:59 -06003676err:
Jens Axboe41d3a6b2021-09-13 13:08:51 -06003677 if (sqd) {
Jens Axboefa846932021-09-01 14:15:59 -06003678 mutex_unlock(&sqd->lock);
Jens Axboe41d3a6b2021-09-13 13:08:51 -06003679 io_put_sq_data(sqd);
3680 }
Jens Axboefa846932021-09-01 14:15:59 -06003681 return ret;
Jens Axboe2e480052021-08-27 11:33:19 -06003682}
3683
Jens Axboeedafcce2019-01-09 09:16:05 -07003684static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3685 void __user *arg, unsigned nr_args)
Jens Axboeb19062a2019-04-15 10:49:38 -06003686 __releases(ctx->uring_lock)
3687 __acquires(ctx->uring_lock)
Jens Axboeedafcce2019-01-09 09:16:05 -07003688{
3689 int ret;
3690
Jens Axboe35fa71a2019-04-22 10:23:23 -06003691 /*
Pavel Begunkovfbb8bb02022-06-25 11:53:02 +01003692 * We don't quiesce the refs for register anymore and so it can't be
3693 * dying as we're holding a file ref here.
Jens Axboe35fa71a2019-04-22 10:23:23 -06003694 */
Pavel Begunkovfbb8bb02022-06-25 11:53:02 +01003695 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
Jens Axboe35fa71a2019-04-22 10:23:23 -06003696 return -ENXIO;
3697
Pavel Begunkov75c40212021-04-15 13:07:40 +01003698 if (ctx->restricted) {
3699 if (opcode >= IORING_REGISTER_LAST)
3700 return -EINVAL;
3701 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3702 if (!test_bit(opcode, ctx->restrictions.register_op))
3703 return -EACCES;
3704 }
3705
Jens Axboeedafcce2019-01-09 09:16:05 -07003706 switch (opcode) {
3707 case IORING_REGISTER_BUFFERS:
Pavel Begunkov0184f082022-05-18 19:13:49 +01003708 ret = -EFAULT;
3709 if (!arg)
3710 break;
Pavel Begunkov634d00d2021-04-25 14:32:26 +01003711 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
Jens Axboeedafcce2019-01-09 09:16:05 -07003712 break;
3713 case IORING_UNREGISTER_BUFFERS:
3714 ret = -EINVAL;
3715 if (arg || nr_args)
3716 break;
Bijan Mottahedeh0a96bbe2021-01-06 12:39:10 -08003717 ret = io_sqe_buffers_unregister(ctx);
Jens Axboeedafcce2019-01-09 09:16:05 -07003718 break;
Jens Axboe6b063142019-01-10 22:13:58 -07003719 case IORING_REGISTER_FILES:
Jens Axboea8da73a2022-05-09 09:29:14 -06003720 ret = -EFAULT;
3721 if (!arg)
3722 break;
Pavel Begunkov792e3582021-04-25 14:32:21 +01003723 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
Jens Axboe6b063142019-01-10 22:13:58 -07003724 break;
3725 case IORING_UNREGISTER_FILES:
3726 ret = -EINVAL;
3727 if (arg || nr_args)
3728 break;
3729 ret = io_sqe_files_unregister(ctx);
3730 break;
Jens Axboec3a31e62019-10-03 13:59:56 -06003731 case IORING_REGISTER_FILES_UPDATE:
Pavel Begunkovc3bdad02021-04-25 14:32:22 +01003732 ret = io_register_files_update(ctx, arg, nr_args);
Jens Axboec3a31e62019-10-03 13:59:56 -06003733 break;
Jens Axboe9b402842019-04-11 11:45:41 -06003734 case IORING_REGISTER_EVENTFD:
Usama Arifc75312d2022-02-04 14:51:15 +00003735 ret = -EINVAL;
3736 if (nr_args != 1)
3737 break;
3738 ret = io_eventfd_register(ctx, arg, 0);
3739 break;
Jens Axboef2842ab2020-01-08 11:04:00 -07003740 case IORING_REGISTER_EVENTFD_ASYNC:
Jens Axboe9b402842019-04-11 11:45:41 -06003741 ret = -EINVAL;
3742 if (nr_args != 1)
3743 break;
Usama Arifc75312d2022-02-04 14:51:15 +00003744 ret = io_eventfd_register(ctx, arg, 1);
Jens Axboe9b402842019-04-11 11:45:41 -06003745 break;
3746 case IORING_UNREGISTER_EVENTFD:
3747 ret = -EINVAL;
3748 if (arg || nr_args)
3749 break;
3750 ret = io_eventfd_unregister(ctx);
3751 break;
Jens Axboe66f4af92020-01-16 15:36:52 -07003752 case IORING_REGISTER_PROBE:
3753 ret = -EINVAL;
3754 if (!arg || nr_args > 256)
3755 break;
3756 ret = io_probe(ctx, arg, nr_args);
3757 break;
Jens Axboe071698e2020-01-28 10:04:42 -07003758 case IORING_REGISTER_PERSONALITY:
3759 ret = -EINVAL;
3760 if (arg || nr_args)
3761 break;
3762 ret = io_register_personality(ctx);
3763 break;
3764 case IORING_UNREGISTER_PERSONALITY:
3765 ret = -EINVAL;
3766 if (arg)
3767 break;
3768 ret = io_unregister_personality(ctx, nr_args);
3769 break;
Stefano Garzarella7e84e1c2020-08-27 16:58:31 +02003770 case IORING_REGISTER_ENABLE_RINGS:
3771 ret = -EINVAL;
3772 if (arg || nr_args)
3773 break;
3774 ret = io_register_enable_rings(ctx);
3775 break;
Stefano Garzarella21b55db2020-08-27 16:58:30 +02003776 case IORING_REGISTER_RESTRICTIONS:
3777 ret = io_register_restrictions(ctx, arg, nr_args);
3778 break;
Pavel Begunkov992da012021-06-10 16:37:37 +01003779 case IORING_REGISTER_FILES2:
3780 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
Pavel Begunkov792e3582021-04-25 14:32:21 +01003781 break;
Pavel Begunkov992da012021-06-10 16:37:37 +01003782 case IORING_REGISTER_FILES_UPDATE2:
3783 ret = io_register_rsrc_update(ctx, arg, nr_args,
3784 IORING_RSRC_FILE);
3785 break;
3786 case IORING_REGISTER_BUFFERS2:
3787 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3788 break;
3789 case IORING_REGISTER_BUFFERS_UPDATE:
3790 ret = io_register_rsrc_update(ctx, arg, nr_args,
3791 IORING_RSRC_BUFFER);
Pavel Begunkovc3bdad02021-04-25 14:32:22 +01003792 break;
Jens Axboefe764212021-06-17 10:19:54 -06003793 case IORING_REGISTER_IOWQ_AFF:
3794 ret = -EINVAL;
3795 if (!arg || !nr_args)
3796 break;
3797 ret = io_register_iowq_aff(ctx, arg, nr_args);
3798 break;
3799 case IORING_UNREGISTER_IOWQ_AFF:
3800 ret = -EINVAL;
3801 if (arg || nr_args)
3802 break;
3803 ret = io_unregister_iowq_aff(ctx);
3804 break;
Jens Axboe2e480052021-08-27 11:33:19 -06003805 case IORING_REGISTER_IOWQ_MAX_WORKERS:
3806 ret = -EINVAL;
3807 if (!arg || nr_args != 2)
3808 break;
3809 ret = io_register_iowq_max_workers(ctx, arg);
3810 break;
Jens Axboee7a6c002022-03-04 08:22:22 -07003811 case IORING_REGISTER_RING_FDS:
3812 ret = io_ringfd_register(ctx, arg, nr_args);
3813 break;
3814 case IORING_UNREGISTER_RING_FDS:
3815 ret = io_ringfd_unregister(ctx, arg, nr_args);
3816 break;
Jens Axboec7fb1942022-04-30 14:38:53 -06003817 case IORING_REGISTER_PBUF_RING:
3818 ret = -EINVAL;
3819 if (!arg || nr_args != 1)
3820 break;
3821 ret = io_register_pbuf_ring(ctx, arg);
3822 break;
3823 case IORING_UNREGISTER_PBUF_RING:
3824 ret = -EINVAL;
3825 if (!arg || nr_args != 1)
3826 break;
3827 ret = io_unregister_pbuf_ring(ctx, arg);
3828 break;
Jens Axboe78a861b2022-06-18 10:00:50 -06003829 case IORING_REGISTER_SYNC_CANCEL:
3830 ret = -EINVAL;
3831 if (!arg || nr_args != 1)
3832 break;
3833 ret = io_sync_cancel(ctx, arg);
3834 break;
Pavel Begunkov6e73dff2022-06-25 11:55:38 +01003835 case IORING_REGISTER_FILE_ALLOC_RANGE:
3836 ret = -EINVAL;
3837 if (!arg || nr_args)
3838 break;
3839 ret = io_register_file_alloc_range(ctx, arg);
3840 break;
Pavel Begunkovbc24d6b2022-07-12 21:52:42 +01003841 case IORING_REGISTER_NOTIFIERS:
3842 ret = io_notif_register(ctx, arg, nr_args);
3843 break;
3844 case IORING_UNREGISTER_NOTIFIERS:
3845 ret = -EINVAL;
3846 if (arg || nr_args)
3847 break;
3848 ret = io_notif_unregister(ctx);
3849 break;
Jens Axboeedafcce2019-01-09 09:16:05 -07003850 default:
3851 ret = -EINVAL;
3852 break;
3853 }
3854
Jens Axboeedafcce2019-01-09 09:16:05 -07003855 return ret;
3856}
3857
3858SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3859 void __user *, arg, unsigned int, nr_args)
3860{
3861 struct io_ring_ctx *ctx;
3862 long ret = -EBADF;
3863 struct fd f;
3864
3865 f = fdget(fd);
3866 if (!f.file)
3867 return -EBADF;
3868
3869 ret = -EOPNOTSUPP;
Jens Axboee5550a12022-05-25 10:28:04 -06003870 if (!io_is_uring_fops(f.file))
Jens Axboeedafcce2019-01-09 09:16:05 -07003871 goto out_fput;
3872
3873 ctx = f.file->private_data;
3874
Pavel Begunkovb6c23dd2021-02-20 15:17:18 +00003875 io_run_task_work();
3876
Jens Axboeedafcce2019-01-09 09:16:05 -07003877 mutex_lock(&ctx->uring_lock);
3878 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3879 mutex_unlock(&ctx->uring_lock);
Usama Arif2757be22022-02-04 14:51:13 +00003880 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
Jens Axboeedafcce2019-01-09 09:16:05 -07003881out_fput:
3882 fdput(f);
3883 return ret;
3884}
3885
Jens Axboe2b188cc2019-01-07 10:46:33 -07003886static int __init io_uring_init(void)
3887{
Stefan Metzmacher9c71d392022-08-11 09:11:16 +02003888#define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
Stefan Metzmacherd7f62e82020-01-29 14:39:41 +01003889 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
Stefan Metzmacher9c71d392022-08-11 09:11:16 +02003890 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
Stefan Metzmacherd7f62e82020-01-29 14:39:41 +01003891} while (0)
3892
3893#define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
Stefan Metzmacher9c71d392022-08-11 09:11:16 +02003894 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
3895#define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
3896 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
Stefan Metzmacherd7f62e82020-01-29 14:39:41 +01003897 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
3898 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
3899 BUILD_BUG_SQE_ELEM(1, __u8, flags);
3900 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
3901 BUILD_BUG_SQE_ELEM(4, __s32, fd);
3902 BUILD_BUG_SQE_ELEM(8, __u64, off);
3903 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
Stefan Metzmacher9c71d392022-08-11 09:11:16 +02003904 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
3905 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
Stefan Metzmacherd7f62e82020-01-29 14:39:41 +01003906 BUILD_BUG_SQE_ELEM(16, __u64, addr);
Pavel Begunkov7d67af22020-02-24 11:32:45 +03003907 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
Stefan Metzmacherd7f62e82020-01-29 14:39:41 +01003908 BUILD_BUG_SQE_ELEM(24, __u32, len);
3909 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
3910 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
3911 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
3912 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
Jiufei Xue5769a352020-06-17 17:53:55 +08003913 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
3914 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
Stefan Metzmacherd7f62e82020-01-29 14:39:41 +01003915 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
3916 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
3917 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
3918 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
3919 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
3920 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
3921 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
3922 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
Pavel Begunkov7d67af22020-02-24 11:32:45 +03003923 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
Stefan Metzmacher9c71d392022-08-11 09:11:16 +02003924 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
3925 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
3926 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
3927 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
3928 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
Stefan Metzmacherd7f62e82020-01-29 14:39:41 +01003929 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
3930 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
Pavel Begunkov16340ea2021-06-24 15:09:58 +01003931 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
Stefan Metzmacherd7f62e82020-01-29 14:39:41 +01003932 BUILD_BUG_SQE_ELEM(42, __u16, personality);
Pavel Begunkov7d67af22020-02-24 11:32:45 +03003933 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
Pavel Begunkovb9445592021-08-25 12:25:45 +01003934 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
Stefan Metzmacher9c71d392022-08-11 09:11:16 +02003935 BUILD_BUG_SQE_ELEM(44, __u16, notification_idx);
3936 BUILD_BUG_SQE_ELEM(46, __u16, addr_len);
Stefan Roesche9621e2b2022-03-23 08:44:19 -07003937 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
Stefan Metzmacher9c71d392022-08-11 09:11:16 +02003938 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
3939 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
Stefan Metzmacherd7f62e82020-01-29 14:39:41 +01003940
Pavel Begunkovb0d658ec2021-04-27 16:13:53 +01003941 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
3942 sizeof(struct io_uring_rsrc_update));
3943 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
3944 sizeof(struct io_uring_rsrc_update2));
Pavel Begunkov90499ad2021-08-25 20:51:40 +01003945
3946 /* ->buf_index is u16 */
Jens Axboec7fb1942022-04-30 14:38:53 -06003947 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
3948 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
3949 offsetof(struct io_uring_buf_ring, tail));
Pavel Begunkov90499ad2021-08-25 20:51:40 +01003950
Pavel Begunkovb0d658ec2021-04-27 16:13:53 +01003951 /* should fit into one byte */
3952 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
Pavel Begunkov68fe2562021-09-15 12:03:38 +01003953 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
3954 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
Pavel Begunkovb0d658ec2021-04-27 16:13:53 +01003955
Hao Xu32c2d332021-09-07 11:22:43 +08003956 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
Pavel Begunkov16340ea2021-06-24 15:09:58 +01003957
Jens Axboe3a4b89a2022-04-25 19:49:00 -06003958 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
3959
Jens Axboed9b57aa2022-06-15 16:27:42 -06003960 io_uring_optable_init();
Jens Axboe0702e532022-05-23 16:56:21 -06003961
Jens Axboe91f245d2021-02-09 13:48:50 -07003962 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
3963 SLAB_ACCOUNT);
Jens Axboe2b188cc2019-01-07 10:46:33 -07003964 return 0;
3965};
3966__initcall(io_uring_init);