Commit | Line | Data |
---|---|---|
2b188cc1 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Shared application/kernel submission and completion ring pairs, for | |
4 | * supporting fast/efficient IO. | |
5 | * | |
6 | * A note on the read/write ordering memory barriers that are matched between | |
1e84b97b SB |
7 | * the application and kernel side. |
8 | * | |
9 | * After the application reads the CQ ring tail, it must use an | |
10 | * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses | |
11 | * before writing the tail (using smp_load_acquire to read the tail will | |
12 | * do). It also needs a smp_mb() before updating CQ head (ordering the | |
13 | * entry load(s) with the head store), pairing with an implicit barrier | |
d068b506 | 14 | * through a control-dependency in io_get_cqe (smp_store_release to |
1e84b97b SB |
15 | * store head will do). Failure to do so could lead to reading invalid |
16 | * CQ entries. | |
17 | * | |
18 | * Likewise, the application must use an appropriate smp_wmb() before | |
19 | * writing the SQ tail (ordering SQ entry stores with the tail store), | |
20 | * which pairs with smp_load_acquire in io_get_sqring (smp_store_release | |
21 | * to store the tail will do). And it needs a barrier ordering the SQ | |
22 | * head load before writing new SQ entries (smp_load_acquire to read | |
23 | * head will do). | |
24 | * | |
25 | * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application | |
26 | * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after* | |
27 | * updating the SQ tail; a full memory barrier smp_mb() is needed | |
28 | * between. | |
2b188cc1 JA |
29 | * |
30 | * Also see the examples in the liburing library: | |
31 | * | |
32 | * git://git.kernel.dk/liburing | |
33 | * | |
34 | * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens | |
35 | * from data shared between the kernel and application. This is done both | |
36 | * for ordering purposes, but also to ensure that once a value is loaded from | |
37 | * data that the application could potentially modify, it remains stable. | |
38 | * | |
39 | * Copyright (C) 2018-2019 Jens Axboe | |
c992fe29 | 40 | * Copyright (c) 2018-2019 Christoph Hellwig |
2b188cc1 JA |
41 | */ |
42 | #include <linux/kernel.h> | |
43 | #include <linux/init.h> | |
44 | #include <linux/errno.h> | |
45 | #include <linux/syscalls.h> | |
52de1fe1 | 46 | #include <net/compat.h> |
2b188cc1 JA |
47 | #include <linux/refcount.h> |
48 | #include <linux/uio.h> | |
6b47ee6e | 49 | #include <linux/bits.h> |
2b188cc1 JA |
50 | |
51 | #include <linux/sched/signal.h> | |
52 | #include <linux/fs.h> | |
53 | #include <linux/file.h> | |
54 | #include <linux/fdtable.h> | |
55 | #include <linux/mm.h> | |
56 | #include <linux/mman.h> | |
2b188cc1 JA |
57 | #include <linux/percpu.h> |
58 | #include <linux/slab.h> | |
edafccee | 59 | #include <linux/bvec.h> |
2b188cc1 JA |
60 | #include <linux/net.h> |
61 | #include <net/sock.h> | |
62 | #include <net/af_unix.h> | |
63 | #include <linux/anon_inodes.h> | |
64 | #include <linux/sched/mm.h> | |
65 | #include <linux/uaccess.h> | |
66 | #include <linux/nospec.h> | |
aa4c3967 | 67 | #include <linux/highmem.h> |
15b71abe | 68 | #include <linux/fsnotify.h> |
4840e418 | 69 | #include <linux/fadvise.h> |
b41e9852 | 70 | #include <linux/task_work.h> |
0f212204 | 71 | #include <linux/io_uring.h> |
b66509b8 | 72 | #include <linux/io_uring/cmd.h> |
5bd2182d | 73 | #include <linux/audit.h> |
cdc1404a | 74 | #include <linux/security.h> |
d808459b | 75 | #include <asm/shmparam.h> |
2b188cc1 | 76 | |
c826bd7a DD |
77 | #define CREATE_TRACE_POINTS |
78 | #include <trace/events/io_uring.h> | |
79 | ||
2b188cc1 JA |
80 | #include <uapi/linux/io_uring.h> |
81 | ||
561fb04a | 82 | #include "io-wq.h" |
2b188cc1 | 83 | |
de23077e | 84 | #include "io_uring.h" |
329061d3 | 85 | #include "opdef.h" |
e418bbc9 | 86 | #include "refs.h" |
c9f06aa7 | 87 | #include "tctx.h" |
c4320315 | 88 | #include "register.h" |
17437f31 | 89 | #include "sqpoll.h" |
a4ad4f74 | 90 | #include "fdinfo.h" |
3b77495a | 91 | #include "kbuf.h" |
73572984 | 92 | #include "rsrc.h" |
38513c46 | 93 | #include "cancel.h" |
43e0bbbd | 94 | #include "net.h" |
eb42cebb | 95 | #include "notif.h" |
f31ecf67 | 96 | #include "waitid.h" |
194bb58c | 97 | #include "futex.h" |
e27f928e | 98 | |
59915143 | 99 | #include "timeout.h" |
329061d3 | 100 | #include "poll.h" |
c92fcfc2 | 101 | #include "rw.h" |
9b797a37 | 102 | #include "alloc_cache.h" |
5e2a18d9 | 103 | |
5277deaa | 104 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 105 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 | 106 | |
68fe256a PB |
107 | #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \ |
108 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) | |
109 | ||
5562a8d7 PB |
110 | #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \ |
111 | IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS) | |
68fe256a | 112 | |
c854357b | 113 | #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ |
9cae36a0 JA |
114 | REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \ |
115 | REQ_F_ASYNC_DATA) | |
b16fed66 | 116 | |
a538be5b PB |
117 | #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\ |
118 | IO_REQ_CLEAN_FLAGS) | |
119 | ||
09899b19 PB |
120 | #define IO_TCTX_REFS_CACHE_NR (1U << 10) |
121 | ||
6dd0be1e | 122 | #define IO_COMPL_BATCH 32 |
bf019da7 | 123 | #define IO_REQ_ALLOC_BATCH 8 |
258b29a9 | 124 | |
10988a0a DY |
125 | enum { |
126 | IO_CHECK_CQ_OVERFLOW_BIT, | |
155bc950 | 127 | IO_CHECK_CQ_DROPPED_BIT, |
10988a0a DY |
128 | }; |
129 | ||
27dc8338 PB |
130 | struct io_defer_entry { |
131 | struct list_head list; | |
132 | struct io_kiocb *req; | |
9cf7c104 | 133 | u32 seq; |
2b188cc1 JA |
134 | }; |
135 | ||
0756a869 PB |
136 | /* requests with any of those set should undergo io_disarm_next() */ |
137 | #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) | |
da1a08c5 | 138 | #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK) |
0756a869 | 139 | |
b4bc35cf PB |
140 | /* |
141 | * No waiters. It's larger than any valid value of the tw counter | |
142 | * so that tests against ->cq_wait_nr would fail and skip wake_up(). | |
143 | */ | |
144 | #define IO_CQ_WAKE_INIT (-1U) | |
145 | /* Forced wake up if there is a waiter regardless of ->cq_wait_nr */ | |
146 | #define IO_CQ_WAKE_FORCE (IO_CQ_WAKE_INIT >> 1) | |
147 | ||
affa87db | 148 | static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
9936c7c2 | 149 | struct task_struct *task, |
3dd0c97a | 150 | bool cancel_all); |
1ffc5422 | 151 | |
cbc2e203 | 152 | static void io_queue_sqe(struct io_kiocb *req); |
de0617e4 | 153 | |
c1755c25 | 154 | struct kmem_cache *req_cachep; |
2b188cc1 | 155 | |
76d3ccec MR |
156 | static int __read_mostly sysctl_io_uring_disabled; |
157 | static int __read_mostly sysctl_io_uring_group = -1; | |
158 | ||
159 | #ifdef CONFIG_SYSCTL | |
160 | static struct ctl_table kernel_io_uring_disabled_table[] = { | |
161 | { | |
162 | .procname = "io_uring_disabled", | |
163 | .data = &sysctl_io_uring_disabled, | |
164 | .maxlen = sizeof(sysctl_io_uring_disabled), | |
165 | .mode = 0644, | |
166 | .proc_handler = proc_dointvec_minmax, | |
167 | .extra1 = SYSCTL_ZERO, | |
168 | .extra2 = SYSCTL_TWO, | |
169 | }, | |
170 | { | |
171 | .procname = "io_uring_group", | |
172 | .data = &sysctl_io_uring_group, | |
173 | .maxlen = sizeof(gid_t), | |
174 | .mode = 0644, | |
175 | .proc_handler = proc_dointvec, | |
176 | }, | |
177 | {}, | |
178 | }; | |
179 | #endif | |
180 | ||
c450178d PB |
181 | static inline void io_submit_flush_completions(struct io_ring_ctx *ctx) |
182 | { | |
931147dd DY |
183 | if (!wq_list_empty(&ctx->submit_state.compl_reqs) || |
184 | ctx->submit_state.cqes_count) | |
c450178d PB |
185 | __io_submit_flush_completions(ctx); |
186 | } | |
187 | ||
faf88dde PB |
188 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
189 | { | |
190 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
191 | } | |
192 | ||
0fc8c2ac DY |
193 | static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx) |
194 | { | |
195 | return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head); | |
196 | } | |
197 | ||
9cae36a0 JA |
198 | static bool io_match_linked(struct io_kiocb *head) |
199 | { | |
200 | struct io_kiocb *req; | |
201 | ||
202 | io_for_each_link(req, head) { | |
203 | if (req->flags & REQ_F_INFLIGHT) | |
204 | return true; | |
205 | } | |
206 | return false; | |
6af3f48b PB |
207 | } |
208 | ||
209 | /* | |
210 | * As io_match_task() but protected against racing with linked timeouts. | |
211 | * User must not hold timeout_lock. | |
212 | */ | |
329061d3 JA |
213 | bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, |
214 | bool cancel_all) | |
6af3f48b | 215 | { |
9cae36a0 JA |
216 | bool matched; |
217 | ||
6af3f48b PB |
218 | if (task && head->task != task) |
219 | return false; | |
9cae36a0 JA |
220 | if (cancel_all) |
221 | return true; | |
222 | ||
223 | if (head->flags & REQ_F_LINK_TIMEOUT) { | |
224 | struct io_ring_ctx *ctx = head->ctx; | |
225 | ||
226 | /* protect against races with linked timeouts */ | |
227 | spin_lock_irq(&ctx->timeout_lock); | |
228 | matched = io_match_linked(head); | |
229 | spin_unlock_irq(&ctx->timeout_lock); | |
230 | } else { | |
231 | matched = io_match_linked(head); | |
232 | } | |
233 | return matched; | |
6af3f48b PB |
234 | } |
235 | ||
a8295b98 HX |
236 | static inline void req_fail_link_node(struct io_kiocb *req, int res) |
237 | { | |
238 | req_set_fail(req); | |
97b388d7 | 239 | io_req_set_res(req, res, 0); |
a8295b98 HX |
240 | } |
241 | ||
fa05457a PB |
242 | static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) |
243 | { | |
244 | wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); | |
a8295b98 HX |
245 | } |
246 | ||
c072481d | 247 | static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) |
2b188cc1 JA |
248 | { |
249 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
250 | ||
0f158b4c | 251 | complete(&ctx->ref_comp); |
2b188cc1 JA |
252 | } |
253 | ||
c072481d | 254 | static __cold void io_fallback_req_func(struct work_struct *work) |
f56165e6 PB |
255 | { |
256 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
257 | fallback_work.work); | |
258 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
259 | struct io_kiocb *req, *tmp; | |
a282967c | 260 | struct io_tw_state ts = { .locked = true, }; |
f56165e6 | 261 | |
f7b32e78 | 262 | percpu_ref_get(&ctx->refs); |
31f084b7 | 263 | mutex_lock(&ctx->uring_lock); |
3218e5d3 | 264 | llist_for_each_entry_safe(req, tmp, node, io_task_work.node) |
a282967c PB |
265 | req->io_task_work.func(req, &ts); |
266 | if (WARN_ON_ONCE(!ts.locked)) | |
31f084b7 PB |
267 | return; |
268 | io_submit_flush_completions(ctx); | |
269 | mutex_unlock(&ctx->uring_lock); | |
f7b32e78 | 270 | percpu_ref_put(&ctx->refs); |
f56165e6 PB |
271 | } |
272 | ||
e6f89be6 PB |
273 | static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) |
274 | { | |
275 | unsigned hash_buckets = 1U << bits; | |
276 | size_t hash_size = hash_buckets * sizeof(table->hbs[0]); | |
277 | ||
278 | table->hbs = kmalloc(hash_size, GFP_KERNEL); | |
279 | if (!table->hbs) | |
280 | return -ENOMEM; | |
281 | ||
282 | table->hash_bits = bits; | |
283 | init_hash_table(table, hash_buckets); | |
284 | return 0; | |
285 | } | |
286 | ||
c072481d | 287 | static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
2b188cc1 JA |
288 | { |
289 | struct io_ring_ctx *ctx; | |
9cfc7e94 | 290 | int hash_bits; |
2b188cc1 JA |
291 | |
292 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
293 | if (!ctx) | |
294 | return NULL; | |
295 | ||
9cfc7e94 JA |
296 | xa_init(&ctx->io_bl_xa); |
297 | ||
78076bb6 JA |
298 | /* |
299 | * Use 5 bits less than the max cq entries, that should give us around | |
4a07723f PB |
300 | * 32 entries per hash list if totally full and uniformly spread, but |
301 | * don't keep too many buckets to not overconsume memory. | |
78076bb6 | 302 | */ |
4a07723f PB |
303 | hash_bits = ilog2(p->cq_entries) - 5; |
304 | hash_bits = clamp(hash_bits, 1, 8); | |
e6f89be6 | 305 | if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) |
78076bb6 | 306 | goto err; |
9ca9fb24 PB |
307 | if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) |
308 | goto err; | |
21482896 | 309 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
48904229 | 310 | 0, GFP_KERNEL)) |
206aefde | 311 | goto err; |
2b188cc1 JA |
312 | |
313 | ctx->flags = p->flags; | |
b4bc35cf | 314 | atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT); |
90554200 | 315 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 316 | INIT_LIST_HEAD(&ctx->sqd_list); |
1d7bb1d5 | 317 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
cc3cec83 | 318 | INIT_LIST_HEAD(&ctx->io_buffers_cache); |
c392cbec | 319 | INIT_HLIST_HEAD(&ctx->io_buf_list); |
69bbc6ad PB |
320 | io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX, |
321 | sizeof(struct io_rsrc_node)); | |
322 | io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX, | |
323 | sizeof(struct async_poll)); | |
324 | io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX, | |
325 | sizeof(struct io_async_msghdr)); | |
194bb58c | 326 | io_futex_cache_init(ctx); |
0f158b4c | 327 | init_completion(&ctx->ref_comp); |
61cf9370 | 328 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 329 | mutex_init(&ctx->uring_lock); |
311997b3 | 330 | init_waitqueue_head(&ctx->cq_wait); |
7b235dd8 | 331 | init_waitqueue_head(&ctx->poll_wq); |
4ea15b56 | 332 | init_waitqueue_head(&ctx->rsrc_quiesce_wq); |
2b188cc1 | 333 | spin_lock_init(&ctx->completion_lock); |
89850fce | 334 | spin_lock_init(&ctx->timeout_lock); |
5eef4e87 | 335 | INIT_WQ_LIST(&ctx->iopoll_list); |
cc3cec83 | 336 | INIT_LIST_HEAD(&ctx->io_buffers_comp); |
de0617e4 | 337 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 338 | INIT_LIST_HEAD(&ctx->timeout_list); |
ef9dd637 | 339 | INIT_LIST_HEAD(&ctx->ltimeout_list); |
d67d2263 | 340 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); |
c0e0d6ba | 341 | init_llist_head(&ctx->work_llist); |
13bf43f5 | 342 | INIT_LIST_HEAD(&ctx->tctx_list); |
c2b6c6bc PB |
343 | ctx->submit_state.free_list.next = NULL; |
344 | INIT_WQ_LIST(&ctx->locked_free_list); | |
f31ecf67 | 345 | INIT_HLIST_HEAD(&ctx->waitid_list); |
194bb58c JA |
346 | #ifdef CONFIG_FUTEX |
347 | INIT_HLIST_HEAD(&ctx->futex_list); | |
348 | #endif | |
9011bf9a | 349 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
6f33b0bc | 350 | INIT_WQ_LIST(&ctx->submit_state.compl_reqs); |
93b8cc60 | 351 | INIT_HLIST_HEAD(&ctx->cancelable_uring_cmd); |
2b188cc1 | 352 | return ctx; |
206aefde | 353 | err: |
e6f89be6 | 354 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 355 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 JA |
356 | kfree(ctx->io_bl); |
357 | xa_destroy(&ctx->io_bl_xa); | |
206aefde JA |
358 | kfree(ctx); |
359 | return NULL; | |
2b188cc1 JA |
360 | } |
361 | ||
8f6ed49a PB |
362 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
363 | { | |
364 | struct io_rings *r = ctx->rings; | |
365 | ||
366 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
367 | ctx->cq_extra--; | |
368 | } | |
369 | ||
9cf7c104 | 370 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 371 | { |
2bc9930e JA |
372 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
373 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 374 | |
8f6ed49a | 375 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 376 | } |
de0617e4 | 377 | |
9d858b21 | 378 | return false; |
de0617e4 JA |
379 | } |
380 | ||
5a754dea PB |
381 | static void io_clean_op(struct io_kiocb *req) |
382 | { | |
383 | if (req->flags & REQ_F_BUFFER_SELECTED) { | |
384 | spin_lock(&req->ctx->completion_lock); | |
385 | io_put_kbuf_comp(req); | |
386 | spin_unlock(&req->ctx->completion_lock); | |
387 | } | |
388 | ||
389 | if (req->flags & REQ_F_NEED_CLEANUP) { | |
390 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; | |
391 | ||
392 | if (def->cleanup) | |
393 | def->cleanup(req); | |
394 | } | |
395 | if ((req->flags & REQ_F_POLLED) && req->apoll) { | |
396 | kfree(req->apoll->double_poll); | |
397 | kfree(req->apoll); | |
398 | req->apoll = NULL; | |
399 | } | |
400 | if (req->flags & REQ_F_INFLIGHT) { | |
401 | struct io_uring_task *tctx = req->task->io_uring; | |
402 | ||
403 | atomic_dec(&tctx->inflight_tracked); | |
404 | } | |
405 | if (req->flags & REQ_F_CREDS) | |
406 | put_cred(req->creds); | |
407 | if (req->flags & REQ_F_ASYNC_DATA) { | |
408 | kfree(req->async_data); | |
409 | req->async_data = NULL; | |
410 | } | |
411 | req->flags &= ~IO_REQ_CLEAN_FLAGS; | |
412 | } | |
413 | ||
9cae36a0 JA |
414 | static inline void io_req_track_inflight(struct io_kiocb *req) |
415 | { | |
416 | if (!(req->flags & REQ_F_INFLIGHT)) { | |
417 | req->flags |= REQ_F_INFLIGHT; | |
386e4fb6 | 418 | atomic_inc(&req->task->io_uring->inflight_tracked); |
9cae36a0 JA |
419 | } |
420 | } | |
421 | ||
fd08e530 PB |
422 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
423 | { | |
906c6caa PB |
424 | if (WARN_ON_ONCE(!req->link)) |
425 | return NULL; | |
426 | ||
4d13d1a4 PB |
427 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
428 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
429 | |
430 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 431 | io_req_set_refcount(req); |
4d13d1a4 PB |
432 | __io_req_set_refcount(req->link, 2); |
433 | return req->link; | |
fd08e530 PB |
434 | } |
435 | ||
436 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
437 | { | |
4d13d1a4 | 438 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
439 | return NULL; |
440 | return __io_prep_linked_timeout(req); | |
441 | } | |
442 | ||
cb2d344c PB |
443 | static noinline void __io_arm_ltimeout(struct io_kiocb *req) |
444 | { | |
445 | io_queue_linked_timeout(__io_prep_linked_timeout(req)); | |
446 | } | |
447 | ||
448 | static inline void io_arm_ltimeout(struct io_kiocb *req) | |
449 | { | |
450 | if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) | |
451 | __io_arm_ltimeout(req); | |
452 | } | |
453 | ||
1e6fa521 JA |
454 | static void io_prep_async_work(struct io_kiocb *req) |
455 | { | |
a7dd2782 | 456 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
1e6fa521 JA |
457 | struct io_ring_ctx *ctx = req->ctx; |
458 | ||
b8e64b53 PB |
459 | if (!(req->flags & REQ_F_CREDS)) { |
460 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 461 | req->creds = get_current_cred(); |
b8e64b53 | 462 | } |
003e8dcc | 463 | |
e1d675df PB |
464 | req->work.list.next = NULL; |
465 | req->work.flags = 0; | |
feaadc4f PB |
466 | if (req->flags & REQ_F_FORCE_ASYNC) |
467 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
468 | ||
3beed235 | 469 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
8487f083 | 470 | req->flags |= io_file_get_flags(req->file); |
f6b543fd | 471 | |
8b1df11f | 472 | if (req->file && (req->flags & REQ_F_ISREG)) { |
d4755e15 JA |
473 | bool should_hash = def->hash_reg_file; |
474 | ||
475 | /* don't serialize this request if the fs doesn't need it */ | |
476 | if (should_hash && (req->file->f_flags & O_DIRECT) && | |
477 | (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE)) | |
478 | should_hash = false; | |
479 | if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL)) | |
1e6fa521 | 480 | io_wq_hash_work(&req->work, file_inode(req->file)); |
4b982bd0 | 481 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
482 | if (def->unbound_nonreg_file) |
483 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
484 | } | |
561fb04a | 485 | } |
cccf0ee8 | 486 | |
cbdcb435 | 487 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 488 | { |
cbdcb435 | 489 | struct io_kiocb *cur; |
54a91f3b | 490 | |
44eff40a PB |
491 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
492 | struct io_ring_ctx *ctx = req->ctx; | |
493 | ||
674ee8e1 | 494 | spin_lock_irq(&ctx->timeout_lock); |
44eff40a PB |
495 | io_for_each_link(cur, req) |
496 | io_prep_async_work(cur); | |
674ee8e1 | 497 | spin_unlock_irq(&ctx->timeout_lock); |
44eff40a PB |
498 | } else { |
499 | io_for_each_link(cur, req) | |
500 | io_prep_async_work(cur); | |
501 | } | |
561fb04a JA |
502 | } |
503 | ||
a282967c | 504 | void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use) |
561fb04a | 505 | { |
cbdcb435 | 506 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 507 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 508 | |
3bfe6106 JA |
509 | BUG_ON(!tctx); |
510 | BUG_ON(!tctx->io_wq); | |
561fb04a | 511 | |
cbdcb435 PB |
512 | /* init ->work of the whole link before punting */ |
513 | io_prep_async_link(req); | |
991468dc JA |
514 | |
515 | /* | |
516 | * Not expected to happen, but if we do have a bug where this _can_ | |
517 | * happen, catch it here and ensure the request is marked as | |
518 | * canceled. That will make io-wq go through the usual work cancel | |
519 | * procedure rather than attempt to run this request (or create a new | |
520 | * worker for it). | |
521 | */ | |
522 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
523 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
524 | ||
48863ffd | 525 | trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); |
ebf93667 | 526 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
527 | if (link) |
528 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
529 | } |
530 | ||
c072481d | 531 | static __cold void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 532 | { |
441b8a78 | 533 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
534 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
535 | struct io_defer_entry, list); | |
de0617e4 | 536 | |
9cf7c104 | 537 | if (req_need_defer(de->req, de->seq)) |
04518945 | 538 | break; |
27dc8338 | 539 | list_del_init(&de->list); |
907d1df3 | 540 | io_req_task_queue(de->req); |
27dc8338 | 541 | kfree(de); |
441b8a78 | 542 | } |
04518945 PB |
543 | } |
544 | ||
c4320315 | 545 | void io_eventfd_ops(struct rcu_head *rcu) |
f2842ab5 | 546 | { |
d8e9214f | 547 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
21a091b9 | 548 | int ops = atomic_xchg(&ev_fd->ops, 0); |
305bef98 | 549 | |
21a091b9 | 550 | if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) |
120ae585 | 551 | eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE); |
d8e9214f | 552 | |
21a091b9 DY |
553 | /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback |
554 | * ordering in a race but if references are 0 we know we have to free | |
555 | * it regardless. | |
305bef98 | 556 | */ |
21a091b9 DY |
557 | if (atomic_dec_and_test(&ev_fd->refs)) { |
558 | eventfd_ctx_put(ev_fd->cq_ev_fd); | |
559 | kfree(ev_fd); | |
560 | } | |
d8e9214f DY |
561 | } |
562 | ||
77bc59b4 | 563 | static void io_eventfd_signal(struct io_ring_ctx *ctx) |
f2842ab5 | 564 | { |
21a091b9 | 565 | struct io_ev_fd *ev_fd = NULL; |
77bc59b4 | 566 | |
77bc59b4 UA |
567 | rcu_read_lock(); |
568 | /* | |
569 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
570 | * and eventfd_signal | |
571 | */ | |
572 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
573 | ||
574 | /* | |
575 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
576 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
577 | * the function and rcu_read_lock. | |
578 | */ | |
579 | if (unlikely(!ev_fd)) | |
580 | goto out; | |
7e55a19c | 581 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 | 582 | goto out; |
21a091b9 DY |
583 | if (ev_fd->eventfd_async && !io_wq_current_is_worker()) |
584 | goto out; | |
77bc59b4 | 585 | |
21a091b9 | 586 | if (likely(eventfd_signal_allowed())) { |
120ae585 | 587 | eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE); |
21a091b9 DY |
588 | } else { |
589 | atomic_inc(&ev_fd->refs); | |
590 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) | |
44a84da4 | 591 | call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops); |
21a091b9 DY |
592 | else |
593 | atomic_dec(&ev_fd->refs); | |
594 | } | |
595 | ||
77bc59b4 UA |
596 | out: |
597 | rcu_read_unlock(); | |
f2842ab5 JA |
598 | } |
599 | ||
21a091b9 DY |
600 | static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) |
601 | { | |
602 | bool skip; | |
603 | ||
604 | spin_lock(&ctx->completion_lock); | |
605 | ||
606 | /* | |
607 | * Eventfd should only get triggered when at least one event has been | |
608 | * posted. Some applications rely on the eventfd notification count | |
609 | * only changing IFF a new CQE has been added to the CQ ring. There's | |
610 | * no depedency on 1:1 relationship between how many times this | |
611 | * function is called (and hence the eventfd count) and number of CQEs | |
612 | * posted to the CQ ring. | |
613 | */ | |
614 | skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
615 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
616 | spin_unlock(&ctx->completion_lock); | |
617 | if (skip) | |
618 | return; | |
619 | ||
620 | io_eventfd_signal(ctx); | |
621 | } | |
622 | ||
a830ffd2 PB |
623 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
624 | { | |
bca39f39 PB |
625 | if (ctx->poll_activated) |
626 | io_poll_wq_wake(ctx); | |
e5f30f6f PB |
627 | if (ctx->off_timeout_used) |
628 | io_flush_timeouts(ctx); | |
629 | if (ctx->drain_active) { | |
a830ffd2 | 630 | spin_lock(&ctx->completion_lock); |
e5f30f6f | 631 | io_queue_deferred(ctx); |
a830ffd2 PB |
632 | spin_unlock(&ctx->completion_lock); |
633 | } | |
634 | if (ctx->has_evfd) | |
21a091b9 | 635 | io_eventfd_flush_signal(ctx); |
a830ffd2 PB |
636 | } |
637 | ||
f66f7342 | 638 | static inline void __io_cq_lock(struct io_ring_ctx *ctx) |
f66f7342 | 639 | { |
ec26c225 | 640 | if (!ctx->lockless_cq) |
f66f7342 PB |
641 | spin_lock(&ctx->completion_lock); |
642 | } | |
643 | ||
6971253f PB |
644 | static inline void io_cq_lock(struct io_ring_ctx *ctx) |
645 | __acquires(ctx->completion_lock) | |
646 | { | |
647 | spin_lock(&ctx->completion_lock); | |
648 | } | |
649 | ||
f66f7342 | 650 | static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) |
3181e22f PB |
651 | { |
652 | io_commit_cqring(ctx); | |
54927baf | 653 | if (!ctx->task_complete) { |
ec26c225 PB |
654 | if (!ctx->lockless_cq) |
655 | spin_unlock(&ctx->completion_lock); | |
656 | /* IOPOLL rings only need to wake up if it's also SQPOLL */ | |
657 | if (!ctx->syscall_iopoll) | |
658 | io_cqring_wake(ctx); | |
3181e22f | 659 | } |
54927baf | 660 | io_commit_cqring_flush(ctx); |
3181e22f PB |
661 | } |
662 | ||
0fdb9a19 | 663 | static void io_cq_unlock_post(struct io_ring_ctx *ctx) |
5d772916 | 664 | __releases(ctx->completion_lock) |
25399321 | 665 | { |
f66f7342 PB |
666 | io_commit_cqring(ctx); |
667 | spin_unlock(&ctx->completion_lock); | |
f66f7342 | 668 | io_cqring_wake(ctx); |
54927baf | 669 | io_commit_cqring_flush(ctx); |
25399321 PB |
670 | } |
671 | ||
c4a2ed72 | 672 | /* Returns true if there are no backlogged entries after the flush */ |
a85381d8 PB |
673 | static void io_cqring_overflow_kill(struct io_ring_ctx *ctx) |
674 | { | |
675 | struct io_overflow_cqe *ocqe; | |
676 | LIST_HEAD(list); | |
677 | ||
f432b76b | 678 | spin_lock(&ctx->completion_lock); |
a85381d8 PB |
679 | list_splice_init(&ctx->cq_overflow_list, &list); |
680 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); | |
f432b76b | 681 | spin_unlock(&ctx->completion_lock); |
a85381d8 PB |
682 | |
683 | while (!list_empty(&list)) { | |
684 | ocqe = list_first_entry(&list, struct io_overflow_cqe, list); | |
685 | list_del(&ocqe->list); | |
686 | kfree(ocqe); | |
687 | } | |
688 | } | |
689 | ||
1b346e4a | 690 | static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
1d7bb1d5 | 691 | { |
e45a3e05 | 692 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 693 | |
a85381d8 | 694 | if (__io_cqring_events(ctx) == ctx->cq_entries) |
1b346e4a | 695 | return; |
1d7bb1d5 | 696 | |
e45a3e05 SR |
697 | if (ctx->flags & IORING_SETUP_CQE32) |
698 | cqe_size <<= 1; | |
699 | ||
25399321 | 700 | io_cq_lock(ctx); |
6c2450ae | 701 | while (!list_empty(&ctx->cq_overflow_list)) { |
59fbc409 | 702 | struct io_uring_cqe *cqe; |
6c2450ae | 703 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 704 | |
59fbc409 | 705 | if (!io_get_cqe_overflow(ctx, &cqe, true)) |
1d7bb1d5 | 706 | break; |
6c2450ae PB |
707 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
708 | struct io_overflow_cqe, list); | |
a85381d8 | 709 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae PB |
710 | list_del(&ocqe->list); |
711 | kfree(ocqe); | |
1d7bb1d5 JA |
712 | } |
713 | ||
1b346e4a | 714 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 715 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 716 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 717 | } |
25399321 | 718 | io_cq_unlock_post(ctx); |
1d7bb1d5 JA |
719 | } |
720 | ||
52ea806a JA |
721 | static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx) |
722 | { | |
723 | /* iopoll syncs against uring_lock, not completion_lock */ | |
724 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
725 | mutex_lock(&ctx->uring_lock); | |
726 | __io_cqring_overflow_flush(ctx); | |
727 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
728 | mutex_unlock(&ctx->uring_lock); | |
729 | } | |
730 | ||
1b346e4a | 731 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 732 | { |
52ea806a JA |
733 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) |
734 | io_cqring_do_overflow_flush(ctx); | |
6c503150 PB |
735 | } |
736 | ||
5afa4650 | 737 | /* can be called by any task */ |
2fdd6fb5 | 738 | static void io_put_task_remote(struct task_struct *task) |
6a290a14 PB |
739 | { |
740 | struct io_uring_task *tctx = task->io_uring; | |
741 | ||
2fdd6fb5 | 742 | percpu_counter_sub(&tctx->inflight, 1); |
8d664282 | 743 | if (unlikely(atomic_read(&tctx->in_cancel))) |
9d170164 | 744 | wake_up(&tctx->wait); |
2fdd6fb5 | 745 | put_task_struct(task); |
9d170164 PB |
746 | } |
747 | ||
5afa4650 | 748 | /* used by a task to put its own references */ |
2fdd6fb5 | 749 | static void io_put_task_local(struct task_struct *task) |
5afa4650 | 750 | { |
2fdd6fb5 | 751 | task->io_uring->cached_refs++; |
5afa4650 PB |
752 | } |
753 | ||
89800a2d | 754 | /* must to be called somewhat shortly after putting a request */ |
2fdd6fb5 | 755 | static inline void io_put_task(struct task_struct *task) |
89800a2d PB |
756 | { |
757 | if (likely(task == current)) | |
2fdd6fb5 | 758 | io_put_task_local(task); |
89800a2d | 759 | else |
2fdd6fb5 | 760 | io_put_task_remote(task); |
89800a2d PB |
761 | } |
762 | ||
63809137 | 763 | void io_task_refs_refill(struct io_uring_task *tctx) |
9a10867a PB |
764 | { |
765 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
766 | ||
767 | percpu_counter_add(&tctx->inflight, refill); | |
768 | refcount_add(refill, ¤t->usage); | |
769 | tctx->cached_refs += refill; | |
770 | } | |
771 | ||
3cc7fdb9 PB |
772 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
773 | { | |
774 | struct io_uring_task *tctx = task->io_uring; | |
775 | unsigned int refs = tctx->cached_refs; | |
776 | ||
777 | if (refs) { | |
778 | tctx->cached_refs = 0; | |
779 | percpu_counter_sub(&tctx->inflight, refs); | |
780 | put_task_struct_many(task, refs); | |
781 | } | |
782 | } | |
783 | ||
68494a65 PB |
784 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
785 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 786 | { |
cce4b8b0 | 787 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
788 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
789 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 790 | |
f26cc959 PB |
791 | lockdep_assert_held(&ctx->completion_lock); |
792 | ||
e45a3e05 SR |
793 | if (is_cqe32) |
794 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 795 | |
e45a3e05 | 796 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 797 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
798 | if (!ocqe) { |
799 | /* | |
800 | * If we're in ring overflow flush mode, or in task cancel mode, | |
801 | * or cannot allocate an overflow entry, then we need to drop it | |
802 | * on the floor. | |
803 | */ | |
8f6ed49a | 804 | io_account_cq_overflow(ctx); |
155bc950 | 805 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 806 | return false; |
2b188cc1 | 807 | } |
cce4b8b0 | 808 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 809 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 810 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 811 | |
cce4b8b0 | 812 | } |
d4d19c19 | 813 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
814 | ocqe->cqe.res = res; |
815 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
816 | if (is_cqe32) { |
817 | ocqe->cqe.big_cqe[0] = extra1; | |
818 | ocqe->cqe.big_cqe[1] = extra2; | |
819 | } | |
cce4b8b0 PB |
820 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
821 | return true; | |
2b188cc1 JA |
822 | } |
823 | ||
056695bf | 824 | void io_req_cqe_overflow(struct io_kiocb *req) |
68494a65 | 825 | { |
056695bf PB |
826 | io_cqring_event_overflow(req->ctx, req->cqe.user_data, |
827 | req->cqe.res, req->cqe.flags, | |
b24c5d75 PB |
828 | req->big_cqe.extra1, req->big_cqe.extra2); |
829 | memset(&req->big_cqe, 0, sizeof(req->big_cqe)); | |
68494a65 PB |
830 | } |
831 | ||
faf88dde PB |
832 | /* |
833 | * writes to the cq entry need to come after reading head; the | |
834 | * control dependency is enough as we're using WRITE_ONCE to | |
835 | * fill the cq entry | |
836 | */ | |
20d6b633 | 837 | bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow) |
faf88dde PB |
838 | { |
839 | struct io_rings *rings = ctx->rings; | |
840 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
841 | unsigned int free, queued, len; |
842 | ||
aa1df3a3 PB |
843 | /* |
844 | * Posting into the CQ when there are pending overflowed CQEs may break | |
845 | * ordering guarantees, which will affect links, F_MORE users and more. | |
846 | * Force overflow the completion. | |
847 | */ | |
848 | if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) | |
20d6b633 | 849 | return false; |
faf88dde PB |
850 | |
851 | /* userspace may cheat modifying the tail, be safe and do min */ | |
852 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
853 | free = ctx->cq_entries - queued; | |
854 | /* we need a contiguous range, limit based on the current array offset */ | |
855 | len = min(free, ctx->cq_entries - off); | |
856 | if (!len) | |
20d6b633 | 857 | return false; |
faf88dde | 858 | |
b3659a65 PB |
859 | if (ctx->flags & IORING_SETUP_CQE32) { |
860 | off <<= 1; | |
861 | len <<= 1; | |
862 | } | |
863 | ||
faf88dde PB |
864 | ctx->cqe_cached = &rings->cqes[off]; |
865 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
20d6b633 | 866 | return true; |
faf88dde PB |
867 | } |
868 | ||
f66f7342 PB |
869 | static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, |
870 | u32 cflags) | |
bcda7baa | 871 | { |
cd94903d PB |
872 | struct io_uring_cqe *cqe; |
873 | ||
913a571a | 874 | ctx->cq_extra++; |
cd94903d PB |
875 | |
876 | /* | |
877 | * If we can't get a cq entry, userspace overflowed the | |
878 | * submission (by quite a lot). Increment the overflow count in | |
879 | * the ring. | |
880 | */ | |
59fbc409 | 881 | if (likely(io_get_cqe(ctx, &cqe))) { |
e0486f3f DY |
882 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
883 | ||
cd94903d PB |
884 | WRITE_ONCE(cqe->user_data, user_data); |
885 | WRITE_ONCE(cqe->res, res); | |
886 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
887 | |
888 | if (ctx->flags & IORING_SETUP_CQE32) { | |
889 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
890 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
891 | } | |
cd94903d PB |
892 | return true; |
893 | } | |
52120f0f | 894 | return false; |
bcda7baa JA |
895 | } |
896 | ||
931147dd DY |
897 | static void __io_flush_post_cqes(struct io_ring_ctx *ctx) |
898 | __must_hold(&ctx->uring_lock) | |
899 | { | |
900 | struct io_submit_state *state = &ctx->submit_state; | |
901 | unsigned int i; | |
902 | ||
903 | lockdep_assert_held(&ctx->uring_lock); | |
904 | for (i = 0; i < state->cqes_count; i++) { | |
0aa7aa5f | 905 | struct io_uring_cqe *cqe = &ctx->completion_cqes[i]; |
931147dd | 906 | |
f66f7342 | 907 | if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) { |
27122c07 | 908 | if (ctx->lockless_cq) { |
f66f7342 PB |
909 | spin_lock(&ctx->completion_lock); |
910 | io_cqring_event_overflow(ctx, cqe->user_data, | |
911 | cqe->res, cqe->flags, 0, 0); | |
912 | spin_unlock(&ctx->completion_lock); | |
913 | } else { | |
914 | io_cqring_event_overflow(ctx, cqe->user_data, | |
915 | cqe->res, cqe->flags, 0, 0); | |
916 | } | |
917 | } | |
931147dd DY |
918 | } |
919 | state->cqes_count = 0; | |
920 | } | |
921 | ||
b529c96a DY |
922 | static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, |
923 | bool allow_overflow) | |
d245bca6 PB |
924 | { |
925 | bool filled; | |
926 | ||
25399321 | 927 | io_cq_lock(ctx); |
f66f7342 PB |
928 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags); |
929 | if (!filled && allow_overflow) | |
930 | filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); | |
931 | ||
25399321 | 932 | io_cq_unlock_post(ctx); |
d245bca6 PB |
933 | return filled; |
934 | } | |
935 | ||
b529c96a DY |
936 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags) |
937 | { | |
938 | return __io_post_aux_cqe(ctx, user_data, res, cflags, true); | |
939 | } | |
940 | ||
b6b2bb58 PB |
941 | /* |
942 | * A helper for multishot requests posting additional CQEs. | |
943 | * Should only be used from a task_work including IO_URING_F_MULTISHOT. | |
944 | */ | |
945 | bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags) | |
2b188cc1 | 946 | { |
d86eaed1 JA |
947 | struct io_ring_ctx *ctx = req->ctx; |
948 | u64 user_data = req->cqe.user_data; | |
9b8c5475 | 949 | struct io_uring_cqe *cqe; |
9b8c5475 DY |
950 | |
951 | if (!defer) | |
b6b2bb58 | 952 | return __io_post_aux_cqe(ctx, user_data, res, cflags, false); |
9b8c5475 | 953 | |
9b8c5475 DY |
954 | lockdep_assert_held(&ctx->uring_lock); |
955 | ||
0aa7aa5f | 956 | if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->completion_cqes)) { |
f66f7342 | 957 | __io_cq_lock(ctx); |
9b8c5475 DY |
958 | __io_flush_post_cqes(ctx); |
959 | /* no need to flush - flush is deferred */ | |
f66f7342 | 960 | __io_cq_unlock_post(ctx); |
9b8c5475 DY |
961 | } |
962 | ||
963 | /* For defered completions this is not as strict as it is otherwise, | |
964 | * however it's main job is to prevent unbounded posted completions, | |
965 | * and in that it works just as well. | |
966 | */ | |
b6b2bb58 | 967 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) |
9b8c5475 DY |
968 | return false; |
969 | ||
0aa7aa5f | 970 | cqe = &ctx->completion_cqes[ctx->submit_state.cqes_count++]; |
9b8c5475 DY |
971 | cqe->user_data = user_data; |
972 | cqe->res = res; | |
973 | cqe->flags = cflags; | |
974 | return true; | |
975 | } | |
976 | ||
ef8ae64f | 977 | static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
2b188cc1 | 978 | { |
fa18fa22 | 979 | struct io_ring_ctx *ctx = req->ctx; |
2ad4c6d0 | 980 | struct io_rsrc_node *rsrc_node = NULL; |
fa18fa22 PB |
981 | |
982 | io_cq_lock(ctx); | |
00b0db56 PB |
983 | if (!(req->flags & REQ_F_CQE_SKIP)) { |
984 | if (!io_fill_cqe_req(ctx, req)) | |
985 | io_req_cqe_overflow(req); | |
986 | } | |
fa18fa22 | 987 | |
c7dae4ba JA |
988 | /* |
989 | * If we're the last reference to this request, add to our locked | |
990 | * free_list cache. | |
991 | */ | |
de9b4cca | 992 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 993 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 994 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
995 | io_disarm_next(req); |
996 | if (req->link) { | |
997 | io_req_task_queue(req->link); | |
998 | req->link = NULL; | |
999 | } | |
1000 | } | |
68a2cc1b | 1001 | io_put_kbuf_comp(req); |
3b7a612f PB |
1002 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1003 | io_clean_op(req); | |
17bc2837 | 1004 | io_put_file(req); |
3b7a612f | 1005 | |
2ad4c6d0 | 1006 | rsrc_node = req->rsrc_node; |
8197b053 PB |
1007 | /* |
1008 | * Selected buffer deallocation in io_clean_op() assumes that | |
1009 | * we don't hold ->completion_lock. Clean them here to avoid | |
1010 | * deadlocks. | |
1011 | */ | |
2fdd6fb5 | 1012 | io_put_task_remote(req->task); |
c2b6c6bc | 1013 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 1014 | ctx->locked_free_nr++; |
180f829f | 1015 | } |
25399321 | 1016 | io_cq_unlock_post(ctx); |
2ad4c6d0 | 1017 | |
ef8ae64f PB |
1018 | if (rsrc_node) { |
1019 | io_ring_submit_lock(ctx, issue_flags); | |
1f2c8f61 | 1020 | io_put_rsrc_node(ctx, rsrc_node); |
ef8ae64f PB |
1021 | io_ring_submit_unlock(ctx, issue_flags); |
1022 | } | |
4e3d9ff9 JA |
1023 | } |
1024 | ||
1bec951c | 1025 | void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
bcda7baa | 1026 | { |
4caa74fd JA |
1027 | struct io_ring_ctx *ctx = req->ctx; |
1028 | ||
1029 | if (ctx->task_complete && ctx->submitter_task != current) { | |
e6aeb272 PB |
1030 | req->io_task_work.func = io_req_task_complete; |
1031 | io_req_task_work_add(req); | |
1032 | } else if (!(issue_flags & IO_URING_F_UNLOCKED) || | |
4caa74fd | 1033 | !(ctx->flags & IORING_SETUP_IOPOLL)) { |
ef8ae64f | 1034 | __io_req_complete_post(req, issue_flags); |
1bec951c | 1035 | } else { |
1bec951c | 1036 | mutex_lock(&ctx->uring_lock); |
ef8ae64f | 1037 | __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED); |
1bec951c PB |
1038 | mutex_unlock(&ctx->uring_lock); |
1039 | } | |
0ddf92e8 JA |
1040 | } |
1041 | ||
973fc83f | 1042 | void io_req_defer_failed(struct io_kiocb *req, s32 res) |
e276ae34 | 1043 | __must_hold(&ctx->uring_lock) |
f41db273 | 1044 | { |
f30bd4d0 | 1045 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; |
a47b255e | 1046 | |
e276ae34 PB |
1047 | lockdep_assert_held(&req->ctx->uring_lock); |
1048 | ||
93d2bcd2 | 1049 | req_set_fail(req); |
97b388d7 | 1050 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
a47b255e PB |
1051 | if (def->fail) |
1052 | def->fail(req); | |
973fc83f | 1053 | io_req_complete_defer(req); |
f41db273 PB |
1054 | } |
1055 | ||
864ea921 PB |
1056 | /* |
1057 | * Don't initialise the fields below on every allocation, but do that in | |
1058 | * advance and keep them valid across allocations. | |
1059 | */ | |
1060 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
1061 | { | |
1062 | req->ctx = ctx; | |
1063 | req->link = NULL; | |
1064 | req->async_data = NULL; | |
1065 | /* not necessary, but safer to zero */ | |
31d3ba92 | 1066 | memset(&req->cqe, 0, sizeof(req->cqe)); |
b24c5d75 | 1067 | memset(&req->big_cqe, 0, sizeof(req->big_cqe)); |
864ea921 PB |
1068 | } |
1069 | ||
dac7a098 | 1070 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 1071 | struct io_submit_state *state) |
dac7a098 | 1072 | { |
79ebeaee | 1073 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1074 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 1075 | ctx->locked_free_nr = 0; |
79ebeaee | 1076 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
1077 | } |
1078 | ||
5d5901a3 PB |
1079 | /* |
1080 | * A request might get retired back into the request caches even before opcode | |
1081 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
1082 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
1083 | * and with extra caution to not get a request that is still worked on. | |
1084 | */ | |
bd1a3783 | 1085 | __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 1086 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 1087 | { |
864ea921 | 1088 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 1089 | void *reqs[IO_REQ_ALLOC_BATCH]; |
864ea921 | 1090 | int ret, i; |
e5d1bc0a | 1091 | |
23a5c43b PB |
1092 | /* |
1093 | * If we have more than a batch's worth of requests in our IRQ side | |
1094 | * locked cache, grab the lock and move them over to our submission | |
1095 | * side cache. | |
1096 | */ | |
a6d97a8a | 1097 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 1098 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 1099 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
1100 | return true; |
1101 | } | |
e5d1bc0a | 1102 | |
3ab665b7 | 1103 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 1104 | |
864ea921 PB |
1105 | /* |
1106 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1107 | * retry single alloc to be on the safe side. | |
1108 | */ | |
1109 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
1110 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
1111 | if (!reqs[0]) | |
a33ae9ce | 1112 | return false; |
864ea921 | 1113 | ret = 1; |
2b188cc1 | 1114 | } |
864ea921 | 1115 | |
37f0e767 | 1116 | percpu_ref_get_many(&ctx->refs, ret); |
3ab665b7 | 1117 | for (i = 0; i < ret; i++) { |
23a5c43b | 1118 | struct io_kiocb *req = reqs[i]; |
3ab665b7 PB |
1119 | |
1120 | io_preinit_req(req, ctx); | |
fa05457a | 1121 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 1122 | } |
a33ae9ce PB |
1123 | return true; |
1124 | } | |
1125 | ||
03adabe8 PB |
1126 | __cold void io_free_req(struct io_kiocb *req) |
1127 | { | |
6ec9afc7 PB |
1128 | /* refs were already put, restore them for io_req_task_complete() */ |
1129 | req->flags &= ~REQ_F_REFCOUNT; | |
1130 | /* we only want to free it, don't post CQEs */ | |
1131 | req->flags |= REQ_F_CQE_SKIP; | |
1132 | req->io_task_work.func = io_req_task_complete; | |
03adabe8 PB |
1133 | io_req_task_work_add(req); |
1134 | } | |
1135 | ||
d81499bf PB |
1136 | static void __io_req_find_next_prep(struct io_kiocb *req) |
1137 | { | |
1138 | struct io_ring_ctx *ctx = req->ctx; | |
d81499bf | 1139 | |
6971253f | 1140 | spin_lock(&ctx->completion_lock); |
305bef98 | 1141 | io_disarm_next(req); |
6971253f | 1142 | spin_unlock(&ctx->completion_lock); |
d81499bf PB |
1143 | } |
1144 | ||
1145 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 1146 | { |
33cc89a9 | 1147 | struct io_kiocb *nxt; |
944e58bf | 1148 | |
9e645e11 JA |
1149 | /* |
1150 | * If LINK is set, we have dependent requests in this chain. If we | |
1151 | * didn't fail this request, queue the first one up, moving any other | |
1152 | * dependencies to the next request. In case of failure, fail the rest | |
1153 | * of the chain. | |
1154 | */ | |
d81499bf PB |
1155 | if (unlikely(req->flags & IO_DISARM_MASK)) |
1156 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
1157 | nxt = req->link; |
1158 | req->link = NULL; | |
1159 | return nxt; | |
4d7dd462 | 1160 | } |
9e645e11 | 1161 | |
a282967c | 1162 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts) |
2c32395d PB |
1163 | { |
1164 | if (!ctx) | |
1165 | return; | |
ef060ea9 JA |
1166 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1167 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
a282967c | 1168 | if (ts->locked) { |
c450178d | 1169 | io_submit_flush_completions(ctx); |
2c32395d | 1170 | mutex_unlock(&ctx->uring_lock); |
a282967c | 1171 | ts->locked = false; |
2c32395d PB |
1172 | } |
1173 | percpu_ref_put(&ctx->refs); | |
1174 | } | |
1175 | ||
42c0905f | 1176 | static unsigned int handle_tw_list(struct llist_node *node) |
9f8d032a | 1177 | { |
42c0905f JA |
1178 | struct io_ring_ctx *ctx = NULL; |
1179 | struct io_tw_state ts = { }; | |
c6dd763c DY |
1180 | unsigned int count = 0; |
1181 | ||
592b4805 | 1182 | do { |
f88262e6 | 1183 | struct llist_node *next = node->next; |
9f8d032a HX |
1184 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1185 | io_task_work.node); | |
1186 | ||
42c0905f JA |
1187 | if (req->ctx != ctx) { |
1188 | ctx_flush_and_put(ctx, &ts); | |
1189 | ctx = req->ctx; | |
9f8d032a | 1190 | /* if not contended, grab and improve batching */ |
42c0905f JA |
1191 | ts.locked = mutex_trylock(&ctx->uring_lock); |
1192 | percpu_ref_get(&ctx->refs); | |
13bfa6f1 | 1193 | } |
c92fcfc2 JA |
1194 | INDIRECT_CALL_2(req->io_task_work.func, |
1195 | io_poll_task_func, io_req_rw_complete, | |
42c0905f | 1196 | req, &ts); |
9f8d032a | 1197 | node = next; |
c6dd763c | 1198 | count++; |
f5868008 | 1199 | if (unlikely(need_resched())) { |
42c0905f JA |
1200 | ctx_flush_and_put(ctx, &ts); |
1201 | ctx = NULL; | |
f5868008 JA |
1202 | cond_resched(); |
1203 | } | |
592b4805 | 1204 | } while (node); |
c6dd763c | 1205 | |
42c0905f | 1206 | ctx_flush_and_put(ctx, &ts); |
c6dd763c | 1207 | return count; |
9f8d032a HX |
1208 | } |
1209 | ||
923d1592 DY |
1210 | /** |
1211 | * io_llist_xchg - swap all entries in a lock-less list | |
1212 | * @head: the head of lock-less list to delete all entries | |
1213 | * @new: new entry as the head of the list | |
1214 | * | |
1215 | * If list is empty, return NULL, otherwise, return the pointer to the first entry. | |
1216 | * The order of entries returned is from the newest to the oldest added one. | |
1217 | */ | |
1218 | static inline struct llist_node *io_llist_xchg(struct llist_head *head, | |
1219 | struct llist_node *new) | |
1220 | { | |
1221 | return xchg(&head->first, new); | |
1222 | } | |
1223 | ||
dfbe5561 | 1224 | static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync) |
10e1c0d5 JA |
1225 | { |
1226 | struct llist_node *node = llist_del_all(&tctx->task_list); | |
dfbe5561 | 1227 | struct io_ring_ctx *last_ctx = NULL; |
10e1c0d5 JA |
1228 | struct io_kiocb *req; |
1229 | ||
1230 | while (node) { | |
1231 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1232 | node = node->next; | |
dfbe5561 JA |
1233 | if (sync && last_ctx != req->ctx) { |
1234 | if (last_ctx) { | |
1235 | flush_delayed_work(&last_ctx->fallback_work); | |
1236 | percpu_ref_put(&last_ctx->refs); | |
1237 | } | |
1238 | last_ctx = req->ctx; | |
1239 | percpu_ref_get(&last_ctx->refs); | |
1240 | } | |
10e1c0d5 JA |
1241 | if (llist_add(&req->io_task_work.node, |
1242 | &req->ctx->fallback_llist)) | |
1243 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1244 | } | |
dfbe5561 JA |
1245 | |
1246 | if (last_ctx) { | |
1247 | flush_delayed_work(&last_ctx->fallback_work); | |
1248 | percpu_ref_put(&last_ctx->refs); | |
1249 | } | |
10e1c0d5 JA |
1250 | } |
1251 | ||
c9f06aa7 | 1252 | void tctx_task_work(struct callback_head *cb) |
c40f6379 | 1253 | { |
3f18407d PB |
1254 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
1255 | task_work); | |
77e443ab | 1256 | struct llist_node *node; |
cb6bf7f2 | 1257 | unsigned int count = 0; |
3a0c037b | 1258 | |
77e443ab | 1259 | if (unlikely(current->flags & PF_EXITING)) { |
dfbe5561 | 1260 | io_fallback_tw(tctx, true); |
77e443ab PB |
1261 | return; |
1262 | } | |
3a0c037b | 1263 | |
592b4805 JA |
1264 | node = llist_del_all(&tctx->task_list); |
1265 | if (node) | |
42c0905f | 1266 | count = handle_tw_list(llist_reverse_order(node)); |
3cc7fdb9 | 1267 | |
8d664282 JA |
1268 | /* relaxed read is enough as only the task itself sets ->in_cancel */ |
1269 | if (unlikely(atomic_read(&tctx->in_cancel))) | |
3cc7fdb9 | 1270 | io_uring_drop_tctx_refs(current); |
c6dd763c | 1271 | |
4c98b891 | 1272 | trace_io_uring_task_work_run(tctx, count); |
7cbf1722 JA |
1273 | } |
1274 | ||
91c7884a | 1275 | static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags) |
c0e0d6ba DY |
1276 | { |
1277 | struct io_ring_ctx *ctx = req->ctx; | |
8751d154 | 1278 | unsigned nr_wait, nr_tw, nr_tw_prev; |
e8c40771 | 1279 | struct llist_node *head; |
c0e0d6ba | 1280 | |
b4bc35cf PB |
1281 | /* See comment above IO_CQ_WAKE_INIT */ |
1282 | BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES); | |
c0e0d6ba | 1283 | |
b4bc35cf PB |
1284 | /* |
1285 | * We don't know how many reuqests is there in the link and whether | |
1286 | * they can even be queued lazily, fall back to non-lazy. | |
1287 | */ | |
8751d154 PB |
1288 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) |
1289 | flags &= ~IOU_F_TWQ_LAZY_WAKE; | |
ce8e04f6 | 1290 | |
e8c40771 | 1291 | head = READ_ONCE(ctx->work_llist.first); |
51509400 | 1292 | do { |
8751d154 | 1293 | nr_tw_prev = 0; |
e8c40771 PB |
1294 | if (head) { |
1295 | struct io_kiocb *first_req = container_of(head, | |
8751d154 PB |
1296 | struct io_kiocb, |
1297 | io_task_work.node); | |
1298 | /* | |
1299 | * Might be executed at any moment, rely on | |
1300 | * SLAB_TYPESAFE_BY_RCU to keep it alive. | |
1301 | */ | |
1302 | nr_tw_prev = READ_ONCE(first_req->nr_tw); | |
1303 | } | |
b4bc35cf PB |
1304 | |
1305 | /* | |
1306 | * Theoretically, it can overflow, but that's fine as one of | |
1307 | * previous adds should've tried to wake the task. | |
1308 | */ | |
8751d154 | 1309 | nr_tw = nr_tw_prev + 1; |
8751d154 | 1310 | if (!(flags & IOU_F_TWQ_LAZY_WAKE)) |
b4bc35cf | 1311 | nr_tw = IO_CQ_WAKE_FORCE; |
8751d154 PB |
1312 | |
1313 | req->nr_tw = nr_tw; | |
e8c40771 PB |
1314 | req->io_task_work.node.next = head; |
1315 | } while (!try_cmpxchg(&ctx->work_llist.first, &head, | |
51509400 PB |
1316 | &req->io_task_work.node)); |
1317 | ||
d381099f PB |
1318 | /* |
1319 | * cmpxchg implies a full barrier, which pairs with the barrier | |
1320 | * in set_current_state() on the io_cqring_wait() side. It's used | |
1321 | * to ensure that either we see updated ->cq_wait_nr, or waiters | |
1322 | * going to sleep will observe the work added to the list, which | |
1323 | * is similar to the wait/wawke task state sync. | |
1324 | */ | |
1325 | ||
e8c40771 | 1326 | if (!head) { |
8751d154 PB |
1327 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1328 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1329 | if (ctx->has_evfd) | |
1330 | io_eventfd_signal(ctx); | |
c0e0d6ba DY |
1331 | } |
1332 | ||
8751d154 | 1333 | nr_wait = atomic_read(&ctx->cq_wait_nr); |
b4bc35cf PB |
1334 | /* not enough or no one is waiting */ |
1335 | if (nr_tw < nr_wait) | |
8751d154 | 1336 | return; |
b4bc35cf PB |
1337 | /* the previous add has already woken it up */ |
1338 | if (nr_tw_prev >= nr_wait) | |
8751d154 | 1339 | return; |
8751d154 | 1340 | wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE); |
c0e0d6ba DY |
1341 | } |
1342 | ||
91c7884a | 1343 | static void io_req_normal_work_add(struct io_kiocb *req) |
7cbf1722 | 1344 | { |
c34398a8 | 1345 | struct io_uring_task *tctx = req->task->io_uring; |
9f010507 | 1346 | struct io_ring_ctx *ctx = req->ctx; |
7cbf1722 | 1347 | |
7cbf1722 | 1348 | /* task_work already pending, we're done */ |
32d91f05 | 1349 | if (!llist_add(&req->io_task_work.node, &tctx->task_list)) |
e09ee510 | 1350 | return; |
7cbf1722 | 1351 | |
ef060ea9 JA |
1352 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1353 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1354 | ||
3fe07bcd | 1355 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1356 | return; |
2215bed9 | 1357 | |
dfbe5561 | 1358 | io_fallback_tw(tctx, false); |
c0e0d6ba DY |
1359 | } |
1360 | ||
91c7884a PB |
1361 | void __io_req_task_work_add(struct io_kiocb *req, unsigned flags) |
1362 | { | |
1363 | if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) { | |
1364 | rcu_read_lock(); | |
1365 | io_req_local_work_add(req, flags); | |
1366 | rcu_read_unlock(); | |
1367 | } else { | |
1368 | io_req_normal_work_add(req); | |
1369 | } | |
1370 | } | |
1371 | ||
c0e0d6ba DY |
1372 | static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) |
1373 | { | |
1374 | struct llist_node *node; | |
1375 | ||
1376 | node = llist_del_all(&ctx->work_llist); | |
1377 | while (node) { | |
1378 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1379 | io_task_work.node); | |
1380 | ||
1381 | node = node->next; | |
91c7884a | 1382 | io_req_normal_work_add(req); |
c0e0d6ba DY |
1383 | } |
1384 | } | |
1385 | ||
9fe3eaea JA |
1386 | static bool io_run_local_work_continue(struct io_ring_ctx *ctx, int events, |
1387 | int min_events) | |
1388 | { | |
1389 | if (llist_empty(&ctx->work_llist)) | |
1390 | return false; | |
1391 | if (events < min_events) | |
1392 | return true; | |
1393 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) | |
1394 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1395 | return false; | |
1396 | } | |
1397 | ||
1398 | static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts, | |
1399 | int min_events) | |
c0e0d6ba | 1400 | { |
c0e0d6ba | 1401 | struct llist_node *node; |
c3f4d39e | 1402 | unsigned int loops = 0; |
140102ae | 1403 | int ret = 0; |
c0e0d6ba | 1404 | |
140102ae | 1405 | if (WARN_ON_ONCE(ctx->submitter_task != current)) |
c0e0d6ba | 1406 | return -EEXIST; |
c3f4d39e PB |
1407 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1408 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
c0e0d6ba | 1409 | again: |
3af0356c JA |
1410 | /* |
1411 | * llists are in reverse order, flip it back the right way before | |
1412 | * running the pending items. | |
1413 | */ | |
1414 | node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL)); | |
c3f4d39e | 1415 | while (node) { |
c0e0d6ba DY |
1416 | struct llist_node *next = node->next; |
1417 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1418 | io_task_work.node); | |
c92fcfc2 JA |
1419 | INDIRECT_CALL_2(req->io_task_work.func, |
1420 | io_poll_task_func, io_req_rw_complete, | |
1421 | req, ts); | |
c0e0d6ba DY |
1422 | ret++; |
1423 | node = next; | |
1424 | } | |
c3f4d39e | 1425 | loops++; |
c0e0d6ba | 1426 | |
9fe3eaea | 1427 | if (io_run_local_work_continue(ctx, ret, min_events)) |
c0e0d6ba | 1428 | goto again; |
a282967c | 1429 | if (ts->locked) { |
c0e0d6ba | 1430 | io_submit_flush_completions(ctx); |
9fe3eaea | 1431 | if (io_run_local_work_continue(ctx, ret, min_events)) |
b0b7a7d2 PB |
1432 | goto again; |
1433 | } | |
9fe3eaea | 1434 | |
f75d5036 | 1435 | trace_io_uring_local_work_run(ctx, ret, loops); |
c0e0d6ba | 1436 | return ret; |
8ac5d85a JA |
1437 | } |
1438 | ||
9fe3eaea JA |
1439 | static inline int io_run_local_work_locked(struct io_ring_ctx *ctx, |
1440 | int min_events) | |
360173ab | 1441 | { |
a282967c | 1442 | struct io_tw_state ts = { .locked = true, }; |
360173ab PB |
1443 | int ret; |
1444 | ||
1445 | if (llist_empty(&ctx->work_llist)) | |
1446 | return 0; | |
1447 | ||
9fe3eaea | 1448 | ret = __io_run_local_work(ctx, &ts, min_events); |
360173ab | 1449 | /* shouldn't happen! */ |
a282967c | 1450 | if (WARN_ON_ONCE(!ts.locked)) |
360173ab PB |
1451 | mutex_lock(&ctx->uring_lock); |
1452 | return ret; | |
1453 | } | |
1454 | ||
9fe3eaea | 1455 | static int io_run_local_work(struct io_ring_ctx *ctx, int min_events) |
8ac5d85a | 1456 | { |
a282967c | 1457 | struct io_tw_state ts = {}; |
8ac5d85a JA |
1458 | int ret; |
1459 | ||
a282967c | 1460 | ts.locked = mutex_trylock(&ctx->uring_lock); |
9fe3eaea | 1461 | ret = __io_run_local_work(ctx, &ts, min_events); |
a282967c | 1462 | if (ts.locked) |
8ac5d85a JA |
1463 | mutex_unlock(&ctx->uring_lock); |
1464 | ||
1465 | return ret; | |
c0e0d6ba DY |
1466 | } |
1467 | ||
a282967c | 1468 | static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts) |
c40f6379 | 1469 | { |
a282967c | 1470 | io_tw_lock(req->ctx, ts); |
973fc83f | 1471 | io_req_defer_failed(req, req->cqe.res); |
c40f6379 JA |
1472 | } |
1473 | ||
a282967c | 1474 | void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts) |
c40f6379 | 1475 | { |
a282967c | 1476 | io_tw_lock(req->ctx, ts); |
316319e8 | 1477 | /* req->task == current here, checking PF_EXITING is safe */ |
6bb30855 | 1478 | if (unlikely(req->task->flags & PF_EXITING)) |
973fc83f | 1479 | io_req_defer_failed(req, -EFAULT); |
6bb30855 | 1480 | else if (req->flags & REQ_F_FORCE_ASYNC) |
a282967c | 1481 | io_queue_iowq(req, ts); |
6bb30855 DY |
1482 | else |
1483 | io_queue_sqe(req); | |
c40f6379 JA |
1484 | } |
1485 | ||
59915143 | 1486 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1487 | { |
97b388d7 | 1488 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1489 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1490 | io_req_task_work_add(req); |
c40f6379 JA |
1491 | } |
1492 | ||
f3b44f92 | 1493 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1494 | { |
5b0a6acc | 1495 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1496 | io_req_task_work_add(req); |
a3df7698 PB |
1497 | } |
1498 | ||
59915143 | 1499 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1500 | { |
9b5f7bd9 | 1501 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1502 | |
1503 | if (nxt) | |
906a8c3f | 1504 | io_req_task_queue(nxt); |
c69f8dbe JL |
1505 | } |
1506 | ||
ec26c225 PB |
1507 | static void io_free_batch_list(struct io_ring_ctx *ctx, |
1508 | struct io_wq_work_node *node) | |
3aa83bfb | 1509 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1510 | { |
3aa83bfb PB |
1511 | do { |
1512 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1513 | comp_list); | |
2d6500d4 | 1514 | |
a538be5b PB |
1515 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1516 | if (req->flags & REQ_F_REFCOUNT) { | |
1517 | node = req->comp_list.next; | |
1518 | if (!req_ref_put_and_test(req)) | |
1519 | continue; | |
1520 | } | |
b605a7fa PB |
1521 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1522 | struct async_poll *apoll = req->apoll; | |
1523 | ||
1524 | if (apoll->double_poll) | |
1525 | kfree(apoll->double_poll); | |
9731bc98 JA |
1526 | if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) |
1527 | kfree(apoll); | |
b605a7fa PB |
1528 | req->flags &= ~REQ_F_POLLED; |
1529 | } | |
da1a08c5 | 1530 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1531 | io_queue_next(req); |
a538be5b PB |
1532 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1533 | io_clean_op(req); | |
c1e53a69 | 1534 | } |
17bc2837 | 1535 | io_put_file(req); |
2d6500d4 | 1536 | |
ab409402 | 1537 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1538 | |
2fdd6fb5 | 1539 | io_put_task(req->task); |
c1e53a69 | 1540 | node = req->comp_list.next; |
fa05457a | 1541 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1542 | } while (node); |
7a743e22 PB |
1543 | } |
1544 | ||
ec26c225 | 1545 | void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1546 | __must_hold(&ctx->uring_lock) |
905c172f | 1547 | { |
cd0ca2e0 | 1548 | struct io_submit_state *state = &ctx->submit_state; |
fa780334 | 1549 | struct io_wq_work_node *node; |
905c172f | 1550 | |
f66f7342 | 1551 | __io_cq_lock(ctx); |
931147dd DY |
1552 | /* must come first to preserve CQE ordering in failure cases */ |
1553 | if (state->cqes_count) | |
1554 | __io_flush_post_cqes(ctx); | |
fa780334 | 1555 | __wq_list_for_each(node, &state->compl_reqs) { |
d9dee430 PB |
1556 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1557 | comp_list); | |
3d4aeb9f | 1558 | |
f66f7342 | 1559 | if (!(req->flags & REQ_F_CQE_SKIP) && |
00b0db56 | 1560 | unlikely(!io_fill_cqe_req(ctx, req))) { |
27122c07 | 1561 | if (ctx->lockless_cq) { |
f66f7342 PB |
1562 | spin_lock(&ctx->completion_lock); |
1563 | io_req_cqe_overflow(req); | |
1564 | spin_unlock(&ctx->completion_lock); | |
1565 | } else { | |
1566 | io_req_cqe_overflow(req); | |
1567 | } | |
1568 | } | |
905c172f | 1569 | } |
c98c81a4 | 1570 | __io_cq_unlock_post(ctx); |
d9dee430 | 1571 | |
931147dd DY |
1572 | if (!wq_list_empty(&ctx->submit_state.compl_reqs)) { |
1573 | io_free_batch_list(ctx, state->compl_reqs.first); | |
1574 | INIT_WQ_LIST(&state->compl_reqs); | |
1575 | } | |
7a743e22 PB |
1576 | } |
1577 | ||
6c503150 | 1578 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1579 | { |
1580 | /* See comment at the top of this file */ | |
1581 | smp_rmb(); | |
e23de15f | 1582 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1583 | } |
1584 | ||
def596e9 JA |
1585 | /* |
1586 | * We can't just wait for polled events to come to us, we have to actively | |
1587 | * find and complete them. | |
1588 | */ | |
c072481d | 1589 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1590 | { |
1591 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1592 | return; | |
1593 | ||
1594 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1595 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1596 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1597 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1598 | break; |
08f5439f JA |
1599 | /* |
1600 | * Ensure we allow local-to-the-cpu processing to take place, | |
1601 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1602 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1603 | */ |
3fcee5a6 PB |
1604 | if (need_resched()) { |
1605 | mutex_unlock(&ctx->uring_lock); | |
1606 | cond_resched(); | |
1607 | mutex_lock(&ctx->uring_lock); | |
1608 | } | |
def596e9 JA |
1609 | } |
1610 | mutex_unlock(&ctx->uring_lock); | |
1611 | } | |
1612 | ||
7668b92a | 1613 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1614 | { |
7668b92a | 1615 | unsigned int nr_events = 0; |
155bc950 | 1616 | unsigned long check_cq; |
500f9fba | 1617 | |
76de6749 PB |
1618 | if (!io_allowed_run_tw(ctx)) |
1619 | return -EEXIST; | |
1620 | ||
3a08576b PB |
1621 | check_cq = READ_ONCE(ctx->check_cq); |
1622 | if (unlikely(check_cq)) { | |
1623 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
a85381d8 | 1624 | __io_cqring_overflow_flush(ctx); |
3a08576b PB |
1625 | /* |
1626 | * Similarly do not spin if we have not informed the user of any | |
1627 | * dropped CQE. | |
1628 | */ | |
1629 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1630 | return -EBADR; | |
1631 | } | |
f39c8a5b PB |
1632 | /* |
1633 | * Don't enter poll loop if we already have events pending. | |
1634 | * If we do, we can potentially be spinning for commands that | |
1635 | * already triggered a CQE (eg in error). | |
1636 | */ | |
f39c8a5b | 1637 | if (io_cqring_events(ctx)) |
d487b43c | 1638 | return 0; |
155bc950 | 1639 | |
def596e9 | 1640 | do { |
9e4bef2b JA |
1641 | int ret = 0; |
1642 | ||
500f9fba JA |
1643 | /* |
1644 | * If a submit got punted to a workqueue, we can have the | |
1645 | * application entering polling for a command before it gets | |
1646 | * issued. That app will hold the uring_lock for the duration | |
1647 | * of the poll right here, so we need to take a breather every | |
1648 | * now and then to ensure that the issue has a chance to add | |
1649 | * the poll to the issued list. Otherwise we can spin here | |
1650 | * forever, while the workqueue is stuck trying to acquire the | |
1651 | * very same mutex. | |
1652 | */ | |
dac6a0ea JA |
1653 | if (wq_list_empty(&ctx->iopoll_list) || |
1654 | io_task_work_pending(ctx)) { | |
8f487ef2 PB |
1655 | u32 tail = ctx->cached_cq_tail; |
1656 | ||
9fe3eaea | 1657 | (void) io_run_local_work_locked(ctx, min); |
def596e9 | 1658 | |
dac6a0ea JA |
1659 | if (task_work_pending(current) || |
1660 | wq_list_empty(&ctx->iopoll_list)) { | |
dac6a0ea | 1661 | mutex_unlock(&ctx->uring_lock); |
9d54bd6a | 1662 | io_run_task_work(); |
dac6a0ea | 1663 | mutex_lock(&ctx->uring_lock); |
dac6a0ea | 1664 | } |
8f487ef2 PB |
1665 | /* some requests don't go through iopoll_list */ |
1666 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1667 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1668 | break; |
500f9fba | 1669 | } |
5ba3c874 | 1670 | ret = io_do_iopoll(ctx, !min); |
9e4bef2b JA |
1671 | if (unlikely(ret < 0)) |
1672 | return ret; | |
dc314886 PB |
1673 | |
1674 | if (task_sigpending(current)) | |
1675 | return -EINTR; | |
9e4bef2b | 1676 | if (need_resched()) |
5ba3c874 | 1677 | break; |
d487b43c | 1678 | |
5ba3c874 | 1679 | nr_events += ret; |
9e4bef2b | 1680 | } while (nr_events < min); |
d487b43c | 1681 | |
9e4bef2b | 1682 | return 0; |
def596e9 | 1683 | } |
7012c815 | 1684 | |
a282967c | 1685 | void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts) |
8ef12efe | 1686 | { |
a282967c | 1687 | if (ts->locked) |
9da070b1 | 1688 | io_req_complete_defer(req); |
7012c815 | 1689 | else |
27f35fe9 | 1690 | io_req_complete_post(req, IO_URING_F_UNLOCKED); |
8ef12efe JA |
1691 | } |
1692 | ||
def596e9 JA |
1693 | /* |
1694 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1695 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1696 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1697 | * accessing the kiocb cookie. |
1698 | */ | |
9882131c | 1699 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1700 | { |
1701 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1702 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1703 | |
1704 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1705 | if (unlikely(needs_lock)) |
cb3d8972 | 1706 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1707 | |
1708 | /* | |
1709 | * Track whether we have multiple files in our lists. This will impact | |
1710 | * how we do polling eventually, not spinning if we're on potentially | |
1711 | * different devices. | |
1712 | */ | |
5eef4e87 | 1713 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1714 | ctx->poll_multi_queue = false; |
1715 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1716 | struct io_kiocb *list_req; |
1717 | ||
5eef4e87 PB |
1718 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1719 | comp_list); | |
30da1b45 | 1720 | if (list_req->file != req->file) |
915b3dde | 1721 | ctx->poll_multi_queue = true; |
def596e9 JA |
1722 | } |
1723 | ||
1724 | /* | |
1725 | * For fast devices, IO may have already completed. If it has, add | |
1726 | * it to the front so we find it first. | |
1727 | */ | |
65a6543d | 1728 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1729 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1730 | else |
5eef4e87 | 1731 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1732 | |
3b44b371 | 1733 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1734 | /* |
1735 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1736 | * in sq thread task context or in io worker task context. If | |
1737 | * current task context is sq thread, we don't need to check | |
1738 | * whether should wake up sq thread. | |
1739 | */ | |
1740 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1741 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1742 | wake_up(&ctx->sq_data->wait); | |
1743 | ||
1744 | mutex_unlock(&ctx->uring_lock); | |
1745 | } | |
def596e9 JA |
1746 | } |
1747 | ||
4bcb982c | 1748 | io_req_flags_t io_file_get_flags(struct file *file) |
88459b50 | 1749 | { |
4bcb982c | 1750 | io_req_flags_t res = 0; |
af197f50 | 1751 | |
53cfd5ce | 1752 | if (S_ISREG(file_inode(file)->i_mode)) |
8487f083 | 1753 | res |= REQ_F_ISREG; |
b9a6c945 | 1754 | if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT)) |
8487f083 | 1755 | res |= REQ_F_SUPPORT_NOWAIT; |
88459b50 | 1756 | return res; |
2b188cc1 JA |
1757 | } |
1758 | ||
99f15d8d | 1759 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1760 | { |
f30bd4d0 BL |
1761 | WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size); |
1762 | req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL); | |
d886e185 PB |
1763 | if (req->async_data) { |
1764 | req->flags |= REQ_F_ASYNC_DATA; | |
1765 | return false; | |
1766 | } | |
1767 | return true; | |
3d9932a8 XW |
1768 | } |
1769 | ||
f3b44f92 | 1770 | int io_req_prep_async(struct io_kiocb *req) |
f67676d1 | 1771 | { |
f30bd4d0 | 1772 | const struct io_cold_def *cdef = &io_cold_defs[req->opcode]; |
a7dd2782 | 1773 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
0702e536 JA |
1774 | |
1775 | /* assign early for deferred execution for non-fixed file */ | |
54aa7f23 | 1776 | if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file) |
0702e536 | 1777 | req->file = io_file_get_normal(req, req->cqe.fd); |
f30bd4d0 | 1778 | if (!cdef->prep_async) |
0702e536 JA |
1779 | return 0; |
1780 | if (WARN_ON_ONCE(req_has_async_data(req))) | |
1781 | return -EFAULT; | |
f30bd4d0 | 1782 | if (!def->manual_alloc) { |
59169439 PB |
1783 | if (io_alloc_async_data(req)) |
1784 | return -EAGAIN; | |
1785 | } | |
f30bd4d0 | 1786 | return cdef->prep_async(req); |
bfe76559 PB |
1787 | } |
1788 | ||
9cf7c104 PB |
1789 | static u32 io_get_sequence(struct io_kiocb *req) |
1790 | { | |
a3dbdf54 | 1791 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1792 | struct io_kiocb *cur; |
9cf7c104 | 1793 | |
a3dbdf54 | 1794 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1795 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1796 | seq--; |
1797 | return seq; | |
9cf7c104 PB |
1798 | } |
1799 | ||
c072481d | 1800 | static __cold void io_drain_req(struct io_kiocb *req) |
e276ae34 | 1801 | __must_hold(&ctx->uring_lock) |
de0617e4 | 1802 | { |
a197f664 | 1803 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1804 | struct io_defer_entry *de; |
f67676d1 | 1805 | int ret; |
e0eb71dc | 1806 | u32 seq = io_get_sequence(req); |
3c19966d | 1807 | |
9d858b21 | 1808 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1809 | spin_lock(&ctx->completion_lock); |
5e371265 | 1810 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1811 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1812 | queue: |
10c66904 | 1813 | ctx->drain_active = false; |
e0eb71dc PB |
1814 | io_req_task_queue(req); |
1815 | return; | |
10c66904 | 1816 | } |
e302f104 | 1817 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1818 | |
cbdcb435 | 1819 | io_prep_async_link(req); |
27dc8338 | 1820 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1821 | if (!de) { |
1b48773f | 1822 | ret = -ENOMEM; |
ef5c600a DY |
1823 | io_req_defer_failed(req, ret); |
1824 | return; | |
76cc33d7 | 1825 | } |
2d28390a | 1826 | |
79ebeaee | 1827 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1828 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1829 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1830 | kfree(de); |
e0eb71dc | 1831 | goto queue; |
de0617e4 JA |
1832 | } |
1833 | ||
48863ffd | 1834 | trace_io_uring_defer(req); |
27dc8338 | 1835 | de->req = req; |
9cf7c104 | 1836 | de->seq = seq; |
27dc8338 | 1837 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1838 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1839 | } |
1840 | ||
f4992544 JA |
1841 | static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def, |
1842 | unsigned int issue_flags) | |
6bf9c47a | 1843 | { |
f4992544 | 1844 | if (req->file || !def->needs_file) |
6bf9c47a JA |
1845 | return true; |
1846 | ||
1847 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1848 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1849 | else |
cef216fc | 1850 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1851 | |
772f5e00 | 1852 | return !!req->file; |
6bf9c47a JA |
1853 | } |
1854 | ||
889fca73 | 1855 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1856 | { |
a7dd2782 | 1857 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
5730b27e | 1858 | const struct cred *creds = NULL; |
d625c6ee | 1859 | int ret; |
2b188cc1 | 1860 | |
f4992544 | 1861 | if (unlikely(!io_assign_file(req, def, issue_flags))) |
70152140 JA |
1862 | return -EBADF; |
1863 | ||
6878b40e | 1864 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1865 | creds = override_creds(req->creds); |
5730b27e | 1866 | |
fcde59fe | 1867 | if (!def->audit_skip) |
5bd2182d PM |
1868 | audit_uring_entry(req->opcode); |
1869 | ||
0702e536 | 1870 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1871 | |
fcde59fe | 1872 | if (!def->audit_skip) |
5bd2182d PM |
1873 | audit_uring_exit(!ret, ret); |
1874 | ||
5730b27e JA |
1875 | if (creds) |
1876 | revert_creds(creds); | |
97b388d7 | 1877 | |
75d7b3ae PB |
1878 | if (ret == IOU_OK) { |
1879 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
9da070b1 | 1880 | io_req_complete_defer(req); |
75d7b3ae | 1881 | else |
1bec951c | 1882 | io_req_complete_post(req, issue_flags); |
97b388d7 | 1883 | |
9b43ef3d PB |
1884 | return 0; |
1885 | } | |
def596e9 | 1886 | |
e0b23d99 PB |
1887 | if (ret == IOU_ISSUE_SKIP_COMPLETE) { |
1888 | ret = 0; | |
1889 | io_arm_ltimeout(req); | |
def596e9 | 1890 | |
e0b23d99 PB |
1891 | /* If the op doesn't have a file, we're not polling for it */ |
1892 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue) | |
1893 | io_iopoll_req_issued(req, issue_flags); | |
1894 | } | |
1895 | return ret; | |
2b188cc1 JA |
1896 | } |
1897 | ||
a282967c | 1898 | int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts) |
329061d3 | 1899 | { |
a282967c | 1900 | io_tw_lock(req->ctx, ts); |
9a692451 DY |
1901 | return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT| |
1902 | IO_URING_F_COMPLETE_DEFER); | |
329061d3 JA |
1903 | } |
1904 | ||
c9f06aa7 | 1905 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1906 | { |
1907 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
247f97a5 | 1908 | struct io_kiocb *nxt = NULL; |
ebc11b6c | 1909 | |
247f97a5 PB |
1910 | if (req_ref_put_and_test(req)) { |
1911 | if (req->flags & IO_REQ_LINK_FLAGS) | |
1912 | nxt = io_req_find_next(req); | |
1913 | io_free_req(req); | |
1914 | } | |
1915 | return nxt ? &nxt->work : NULL; | |
ebc11b6c PB |
1916 | } |
1917 | ||
c9f06aa7 | 1918 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1919 | { |
1920 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
a7dd2782 | 1921 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
e6aeb272 | 1922 | unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ; |
d01905db | 1923 | bool needs_poll = false; |
6bf9c47a | 1924 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1925 | |
23a6c9ac | 1926 | /* one will be dropped by ->io_wq_free_work() after returning to io-wq */ |
48dcd38d PB |
1927 | if (!(req->flags & REQ_F_REFCOUNT)) |
1928 | __io_req_set_refcount(req, 2); | |
1929 | else | |
1930 | req_ref_get(req); | |
5d5901a3 | 1931 | |
cb2d344c | 1932 | io_arm_ltimeout(req); |
6bf9c47a | 1933 | |
dadebc35 | 1934 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1935 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1936 | fail: |
6bf9c47a | 1937 | io_req_task_queue_fail(req, err); |
d01905db PB |
1938 | return; |
1939 | } | |
f4992544 | 1940 | if (!io_assign_file(req, def, issue_flags)) { |
0f8da75b PB |
1941 | err = -EBADF; |
1942 | work->flags |= IO_WQ_WORK_CANCEL; | |
1943 | goto fail; | |
1944 | } | |
31b51510 | 1945 | |
d01905db | 1946 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
1947 | bool opcode_poll = def->pollin || def->pollout; |
1948 | ||
95041b93 | 1949 | if (opcode_poll && io_file_can_poll(req)) { |
afb7f56f | 1950 | needs_poll = true; |
d01905db | 1951 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 1952 | } |
561fb04a | 1953 | } |
31b51510 | 1954 | |
d01905db PB |
1955 | do { |
1956 | ret = io_issue_sqe(req, issue_flags); | |
1957 | if (ret != -EAGAIN) | |
1958 | break; | |
a9be2022 JA |
1959 | |
1960 | /* | |
1961 | * If REQ_F_NOWAIT is set, then don't wait or retry with | |
1962 | * poll. -EAGAIN is final for that case. | |
1963 | */ | |
1964 | if (req->flags & REQ_F_NOWAIT) | |
1965 | break; | |
1966 | ||
d01905db PB |
1967 | /* |
1968 | * We can get EAGAIN for iopolled IO even though we're | |
1969 | * forcing a sync submission from here, since we can't | |
1970 | * wait for request slots on the block side. | |
1971 | */ | |
1972 | if (!needs_poll) { | |
e0deb6a0 PB |
1973 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
1974 | break; | |
45500dc4 PB |
1975 | if (io_wq_worker_stopped()) |
1976 | break; | |
d01905db PB |
1977 | cond_resched(); |
1978 | continue; | |
90fa0288 HX |
1979 | } |
1980 | ||
4d9237e3 | 1981 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
1982 | return; |
1983 | /* aborted or ready, in either case retry blocking */ | |
1984 | needs_poll = false; | |
1985 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
1986 | } while (1); | |
31b51510 | 1987 | |
a3df7698 | 1988 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 1989 | if (ret < 0) |
a3df7698 | 1990 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
1991 | } |
1992 | ||
531113bb JA |
1993 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
1994 | unsigned int issue_flags) | |
09bb8394 | 1995 | { |
5106dd6e | 1996 | struct io_ring_ctx *ctx = req->ctx; |
4bfb0c9a | 1997 | struct io_fixed_file *slot; |
5106dd6e | 1998 | struct file *file = NULL; |
09bb8394 | 1999 | |
93f052cb | 2000 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 2001 | |
ac177053 | 2002 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 2003 | goto out; |
ac177053 | 2004 | fd = array_index_nospec(fd, ctx->nr_user_files); |
4bfb0c9a | 2005 | slot = io_fixed_file_slot(&ctx->file_table, fd); |
3f302388 JA |
2006 | if (!req->rsrc_node) |
2007 | __io_req_set_rsrc_node(req, ctx); | |
4bfb0c9a | 2008 | req->flags |= io_slot_flags(slot); |
3f302388 | 2009 | file = io_slot_file(slot); |
5106dd6e | 2010 | out: |
93f052cb | 2011 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
2012 | return file; |
2013 | } | |
d44f554e | 2014 | |
531113bb | 2015 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 2016 | { |
62906e89 | 2017 | struct file *file = fget(fd); |
ac177053 | 2018 | |
48863ffd | 2019 | trace_io_uring_file_get(req, fd); |
09bb8394 | 2020 | |
ac177053 | 2021 | /* we don't allow fixed io_uring files */ |
e5550a14 | 2022 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 2023 | io_req_track_inflight(req); |
8371adf5 | 2024 | return file; |
09bb8394 JA |
2025 | } |
2026 | ||
7bfa9bad | 2027 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
2028 | __must_hold(&req->ctx->uring_lock) |
2029 | { | |
7bfa9bad PB |
2030 | struct io_kiocb *linked_timeout; |
2031 | ||
2032 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
973fc83f | 2033 | io_req_defer_failed(req, ret); |
7bfa9bad PB |
2034 | return; |
2035 | } | |
2036 | ||
2037 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 2038 | |
4d9237e3 | 2039 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 2040 | case IO_APOLL_READY: |
336d28a8 | 2041 | io_kbuf_recycle(req, 0); |
d475a9a6 PB |
2042 | io_req_task_queue(req); |
2043 | break; | |
2044 | case IO_APOLL_ABORTED: | |
6436c770 | 2045 | io_kbuf_recycle(req, 0); |
77955efb | 2046 | io_queue_iowq(req, NULL); |
d475a9a6 | 2047 | break; |
b1c62645 | 2048 | case IO_APOLL_OK: |
b1c62645 | 2049 | break; |
d475a9a6 PB |
2050 | } |
2051 | ||
2052 | if (linked_timeout) | |
2053 | io_queue_linked_timeout(linked_timeout); | |
2054 | } | |
2055 | ||
cbc2e203 | 2056 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 2057 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 2058 | { |
e0c5c576 | 2059 | int ret; |
2b188cc1 | 2060 | |
c5eef2b9 | 2061 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 2062 | |
491381ce JA |
2063 | /* |
2064 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2065 | * doesn't support non-blocking read/write attempts | |
2066 | */ | |
e0b23d99 | 2067 | if (unlikely(ret)) |
7bfa9bad | 2068 | io_queue_async(req, ret); |
2b188cc1 JA |
2069 | } |
2070 | ||
4652fe3f | 2071 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 2072 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 2073 | { |
17b147f6 PB |
2074 | if (unlikely(req->flags & REQ_F_FAIL)) { |
2075 | /* | |
2076 | * We don't submit, fail them all, for that replace hardlinks | |
2077 | * with normal links. Extra REQ_F_LINK is tolerated. | |
2078 | */ | |
2079 | req->flags &= ~REQ_F_HARDLINK; | |
2080 | req->flags |= REQ_F_LINK; | |
973fc83f | 2081 | io_req_defer_failed(req, req->cqe.res); |
76cc33d7 PB |
2082 | } else { |
2083 | int ret = io_req_prep_async(req); | |
2084 | ||
ef5c600a | 2085 | if (unlikely(ret)) { |
973fc83f | 2086 | io_req_defer_failed(req, ret); |
ef5c600a DY |
2087 | return; |
2088 | } | |
2089 | ||
2090 | if (unlikely(req->ctx->drain_active)) | |
2091 | io_drain_req(req); | |
76cc33d7 | 2092 | else |
77955efb | 2093 | io_queue_iowq(req, NULL); |
ce35a47a | 2094 | } |
4fe2c963 JL |
2095 | } |
2096 | ||
b16fed66 PB |
2097 | /* |
2098 | * Check SQE restrictions (opcode and flags). | |
2099 | * | |
2100 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
2101 | */ | |
2102 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
2103 | struct io_kiocb *req, | |
2104 | unsigned int sqe_flags) | |
4fe2c963 | 2105 | { |
b16fed66 PB |
2106 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
2107 | return false; | |
2108 | ||
2109 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
2110 | ctx->restrictions.sqe_flags_required) | |
2111 | return false; | |
2112 | ||
2113 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
2114 | ctx->restrictions.sqe_flags_required)) | |
2115 | return false; | |
2116 | ||
2117 | return true; | |
4fe2c963 JL |
2118 | } |
2119 | ||
22b2ca31 PB |
2120 | static void io_init_req_drain(struct io_kiocb *req) |
2121 | { | |
2122 | struct io_ring_ctx *ctx = req->ctx; | |
2123 | struct io_kiocb *head = ctx->submit_state.link.head; | |
2124 | ||
2125 | ctx->drain_active = true; | |
2126 | if (head) { | |
2127 | /* | |
2128 | * If we need to drain a request in the middle of a link, drain | |
2129 | * the head request and the next request/link after the current | |
2130 | * link. Considering sequential execution of links, | |
b6c7db32 | 2131 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
2132 | * link. |
2133 | */ | |
b6c7db32 | 2134 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
2135 | ctx->drain_next = true; |
2136 | } | |
2137 | } | |
2138 | ||
b16fed66 PB |
2139 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2140 | const struct io_uring_sqe *sqe) | |
282cdc86 | 2141 | __must_hold(&ctx->uring_lock) |
b16fed66 | 2142 | { |
a7dd2782 | 2143 | const struct io_issue_def *def; |
b16fed66 | 2144 | unsigned int sqe_flags; |
fc0ae024 | 2145 | int personality; |
4a04d1d1 | 2146 | u8 opcode; |
b16fed66 | 2147 | |
864ea921 | 2148 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 2149 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 | 2150 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
4bcb982c JA |
2151 | sqe_flags = READ_ONCE(sqe->flags); |
2152 | req->flags = (io_req_flags_t) sqe_flags; | |
cef216fc | 2153 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 2154 | req->file = NULL; |
c1bdf8ed | 2155 | req->rsrc_node = NULL; |
b16fed66 | 2156 | req->task = current; |
b16fed66 | 2157 | |
4a04d1d1 PB |
2158 | if (unlikely(opcode >= IORING_OP_LAST)) { |
2159 | req->opcode = 0; | |
b16fed66 | 2160 | return -EINVAL; |
4a04d1d1 | 2161 | } |
a7dd2782 | 2162 | def = &io_issue_defs[opcode]; |
68fe256a PB |
2163 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
2164 | /* enforce forwards compatibility on users */ | |
2165 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
2166 | return -EINVAL; | |
4e906702 | 2167 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 2168 | if (!def->buffer_select) |
4e906702 JA |
2169 | return -EOPNOTSUPP; |
2170 | req->buf_index = READ_ONCE(sqe->buf_group); | |
2171 | } | |
5562a8d7 PB |
2172 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
2173 | ctx->drain_disabled = true; | |
2174 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
2175 | if (ctx->drain_disabled) | |
2176 | return -EOPNOTSUPP; | |
22b2ca31 | 2177 | io_init_req_drain(req); |
5562a8d7 | 2178 | } |
2a56a9bd PB |
2179 | } |
2180 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
2181 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
2182 | return -EACCES; | |
2183 | /* knock it to the slow queue path, will be drained there */ | |
2184 | if (ctx->drain_active) | |
2185 | req->flags |= REQ_F_FORCE_ASYNC; | |
2186 | /* if there is no link, we're at "next" request and need to drain */ | |
2187 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
2188 | ctx->drain_next = false; | |
2189 | ctx->drain_active = true; | |
b6c7db32 | 2190 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 2191 | } |
68fe256a | 2192 | } |
b16fed66 | 2193 | |
fcde59fe | 2194 | if (!def->ioprio && sqe->ioprio) |
73911426 | 2195 | return -EINVAL; |
fcde59fe | 2196 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
73911426 JA |
2197 | return -EINVAL; |
2198 | ||
fcde59fe | 2199 | if (def->needs_file) { |
6d63416d PB |
2200 | struct io_submit_state *state = &ctx->submit_state; |
2201 | ||
cef216fc | 2202 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 2203 | |
6d63416d PB |
2204 | /* |
2205 | * Plug now if we have more than 2 IO left after this, and the | |
2206 | * target is potentially a read/write to block based storage. | |
2207 | */ | |
fcde59fe | 2208 | if (state->need_plug && def->plug) { |
6d63416d PB |
2209 | state->plug_started = true; |
2210 | state->need_plug = false; | |
5ca7a8b3 | 2211 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 2212 | } |
b16fed66 | 2213 | } |
863e0560 | 2214 | |
003e8dcc JA |
2215 | personality = READ_ONCE(sqe->personality); |
2216 | if (personality) { | |
cdab10bf LT |
2217 | int ret; |
2218 | ||
c10d1f98 PB |
2219 | req->creds = xa_load(&ctx->personalities, personality); |
2220 | if (!req->creds) | |
003e8dcc | 2221 | return -EINVAL; |
c10d1f98 | 2222 | get_cred(req->creds); |
cdc1404a PM |
2223 | ret = security_uring_override_creds(req->creds); |
2224 | if (ret) { | |
2225 | put_cred(req->creds); | |
2226 | return ret; | |
2227 | } | |
b8e64b53 | 2228 | req->flags |= REQ_F_CREDS; |
003e8dcc | 2229 | } |
b16fed66 | 2230 | |
0702e536 | 2231 | return def->prep(req, sqe); |
b16fed66 PB |
2232 | } |
2233 | ||
df3becde PB |
2234 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
2235 | struct io_kiocb *req, int ret) | |
2236 | { | |
2237 | struct io_ring_ctx *ctx = req->ctx; | |
2238 | struct io_submit_link *link = &ctx->submit_state.link; | |
2239 | struct io_kiocb *head = link->head; | |
2240 | ||
48863ffd | 2241 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
2242 | |
2243 | /* | |
2244 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
2245 | * unusable. Instead of failing eagerly, continue assembling the link if | |
2246 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
2247 | * should find the flag and handle the rest. | |
2248 | */ | |
2249 | req_fail_link_node(req, ret); | |
2250 | if (head && !(head->flags & REQ_F_FAIL)) | |
2251 | req_fail_link_node(head, -ECANCELED); | |
2252 | ||
2253 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
2254 | if (head) { | |
2255 | link->last->link = req; | |
2256 | link->head = NULL; | |
2257 | req = head; | |
2258 | } | |
2259 | io_queue_sqe_fallback(req); | |
2260 | return ret; | |
2261 | } | |
2262 | ||
2263 | if (head) | |
2264 | link->last->link = req; | |
2265 | else | |
2266 | link->head = req; | |
2267 | link->last = req; | |
2268 | return 0; | |
2269 | } | |
2270 | ||
2271 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 2272 | const struct io_uring_sqe *sqe) |
282cdc86 | 2273 | __must_hold(&ctx->uring_lock) |
9e645e11 | 2274 | { |
a1ab7b35 | 2275 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 2276 | int ret; |
9e645e11 | 2277 | |
a6b8cadc | 2278 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
2279 | if (unlikely(ret)) |
2280 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 2281 | |
2ad57931 | 2282 | trace_io_uring_submit_req(req); |
a6b8cadc | 2283 | |
9e645e11 JA |
2284 | /* |
2285 | * If we already have a head request, queue this one for async | |
2286 | * submittal once the head completes. If we don't have a head but | |
2287 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2288 | * submitted sync once the chain is complete. If none of those | |
2289 | * conditions are true (normal request), then just queue it. | |
2290 | */ | |
924a07e4 | 2291 | if (unlikely(link->head)) { |
df3becde PB |
2292 | ret = io_req_prep_async(req); |
2293 | if (unlikely(ret)) | |
2294 | return io_submit_fail_init(sqe, req, ret); | |
2295 | ||
48863ffd | 2296 | trace_io_uring_link(req, link->head); |
f2f87370 | 2297 | link->last->link = req; |
863e0560 | 2298 | link->last = req; |
32fe525b | 2299 | |
da1a08c5 | 2300 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 2301 | return 0; |
df3becde PB |
2302 | /* last request of the link, flush it */ |
2303 | req = link->head; | |
f15a3431 | 2304 | link->head = NULL; |
924a07e4 PB |
2305 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
2306 | goto fallback; | |
2307 | ||
2308 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2309 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2310 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2311 | link->head = req; | |
2312 | link->last = req; | |
2313 | } else { | |
2314 | fallback: | |
2315 | io_queue_sqe_fallback(req); | |
2316 | } | |
f15a3431 | 2317 | return 0; |
9e645e11 | 2318 | } |
2e6e1fde | 2319 | |
f15a3431 | 2320 | io_queue_sqe(req); |
1d4240cc | 2321 | return 0; |
9e645e11 JA |
2322 | } |
2323 | ||
9a56a232 JA |
2324 | /* |
2325 | * Batched submission is done, ensure local IO is flushed out. | |
2326 | */ | |
553deffd | 2327 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2328 | { |
553deffd PB |
2329 | struct io_submit_state *state = &ctx->submit_state; |
2330 | ||
e126391c PB |
2331 | if (unlikely(state->link.head)) |
2332 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2333 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2334 | io_submit_flush_completions(ctx); |
27926b68 JA |
2335 | if (state->plug_started) |
2336 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2337 | } |
2338 | ||
2339 | /* | |
2340 | * Start submission side cache. | |
2341 | */ | |
2342 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2343 | unsigned int max_ios) |
9a56a232 | 2344 | { |
27926b68 | 2345 | state->plug_started = false; |
4b628aeb | 2346 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2347 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2348 | /* set only head, no need to init link_last in advance */ |
2349 | state->link.head = NULL; | |
9a56a232 JA |
2350 | } |
2351 | ||
2b188cc1 JA |
2352 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2353 | { | |
75b28aff | 2354 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2355 | |
caf582c6 PB |
2356 | /* |
2357 | * Ensure any loads from the SQEs are done at this point, | |
2358 | * since once we write the new head, the application could | |
2359 | * write new data to them. | |
2360 | */ | |
2361 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2362 | } |
2363 | ||
2b188cc1 | 2364 | /* |
dd9ae8a0 | 2365 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2366 | * that is mapped by userspace. This means that care needs to be taken to |
2367 | * ensure that reads are stable, as we cannot rely on userspace always | |
2368 | * being a good citizen. If members of the sqe are validated and then later | |
2369 | * used, it's important that those reads are done through READ_ONCE() to | |
2370 | * prevent a re-load down the line. | |
2371 | */ | |
b5083dfa | 2372 | static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe) |
2b188cc1 | 2373 | { |
2af89abd PB |
2374 | unsigned mask = ctx->sq_entries - 1; |
2375 | unsigned head = ctx->cached_sq_head++ & mask; | |
2376 | ||
2377 | if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) { | |
2378 | head = READ_ONCE(ctx->sq_array[head]); | |
2379 | if (unlikely(head >= ctx->sq_entries)) { | |
2380 | /* drop invalid entries */ | |
2381 | spin_lock(&ctx->completion_lock); | |
2382 | ctx->cq_extra--; | |
2383 | spin_unlock(&ctx->completion_lock); | |
2384 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2385 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
2386 | return false; | |
2387 | } | |
2388 | } | |
2b188cc1 JA |
2389 | |
2390 | /* | |
2391 | * The cached sq head (or cq tail) serves two purposes: | |
2392 | * | |
2393 | * 1) allows us to batch the cost of updating the user visible | |
2394 | * head updates. | |
2395 | * 2) allows the kernel side to track the head on its own, even | |
2396 | * though the application is the one updating it. | |
2397 | */ | |
2b188cc1 | 2398 | |
2af89abd PB |
2399 | /* double index for 128-byte SQEs, twice as long */ |
2400 | if (ctx->flags & IORING_SETUP_SQE128) | |
2401 | head <<= 1; | |
2402 | *sqe = &ctx->sq_sqes[head]; | |
2403 | return true; | |
709b302f PB |
2404 | } |
2405 | ||
17437f31 | 2406 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2407 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2408 | { |
69629809 | 2409 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2410 | unsigned int left; |
2411 | int ret; | |
6c271ce2 | 2412 | |
51d48dab | 2413 | if (unlikely(!entries)) |
69629809 | 2414 | return 0; |
ee7d46d9 | 2415 | /* make sure SQ entry isn't read before tail */ |
e3ef728f | 2416 | ret = left = min(nr, entries); |
8e6971a8 PB |
2417 | io_get_task_refs(left); |
2418 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2419 | |
69629809 | 2420 | do { |
3529d8c2 | 2421 | const struct io_uring_sqe *sqe; |
196be95c | 2422 | struct io_kiocb *req; |
fb5ccc98 | 2423 | |
c8576f3e | 2424 | if (unlikely(!io_alloc_req(ctx, &req))) |
fb5ccc98 | 2425 | break; |
b5083dfa | 2426 | if (unlikely(!io_get_sqe(ctx, &sqe))) { |
fa05457a | 2427 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2428 | break; |
2429 | } | |
6c271ce2 | 2430 | |
1cd15904 PB |
2431 | /* |
2432 | * Continue submitting even for sqe failure if the | |
2433 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2434 | */ | |
2435 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2436 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2437 | left--; | |
2438 | break; | |
bcbb7bf6 | 2439 | } |
1cd15904 | 2440 | } while (--left); |
9466f437 | 2441 | |
8e6971a8 PB |
2442 | if (unlikely(left)) { |
2443 | ret -= left; | |
2444 | /* try again if it submitted nothing and can't allocate a req */ | |
2445 | if (!ret && io_req_cache_empty(ctx)) | |
2446 | ret = -EAGAIN; | |
2447 | current->io_uring->cached_refs += left; | |
9466f437 | 2448 | } |
6c271ce2 | 2449 | |
553deffd | 2450 | io_submit_state_end(ctx); |
ae9428ca PB |
2451 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2452 | io_commit_sqring(ctx); | |
8e6971a8 | 2453 | return ret; |
6c271ce2 JA |
2454 | } |
2455 | ||
bda52162 JA |
2456 | struct io_wait_queue { |
2457 | struct wait_queue_entry wq; | |
2458 | struct io_ring_ctx *ctx; | |
5fd46178 | 2459 | unsigned cq_tail; |
bda52162 | 2460 | unsigned nr_timeouts; |
d33a39e5 | 2461 | ktime_t timeout; |
bda52162 JA |
2462 | }; |
2463 | ||
b4c98d59 DY |
2464 | static inline bool io_has_work(struct io_ring_ctx *ctx) |
2465 | { | |
c0e0d6ba | 2466 | return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) || |
490c00eb | 2467 | !llist_empty(&ctx->work_llist); |
b4c98d59 DY |
2468 | } |
2469 | ||
6c503150 | 2470 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
2471 | { |
2472 | struct io_ring_ctx *ctx = iowq->ctx; | |
0fc8c2ac | 2473 | int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail; |
bda52162 JA |
2474 | |
2475 | /* | |
d195a66e | 2476 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
2477 | * started waiting. For timeouts, we always want to return to userspace, |
2478 | * regardless of event count. | |
2479 | */ | |
5fd46178 | 2480 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
2481 | } |
2482 | ||
2483 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2484 | int wake_flags, void *key) | |
2485 | { | |
bd550173 | 2486 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq); |
bda52162 | 2487 | |
6c503150 PB |
2488 | /* |
2489 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2490 | * the task, and the next invocation will do it. | |
2491 | */ | |
bd550173 | 2492 | if (io_should_wake(iowq) || io_has_work(iowq->ctx)) |
6c503150 PB |
2493 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2494 | return -1; | |
bda52162 JA |
2495 | } |
2496 | ||
c0e0d6ba | 2497 | int io_run_task_work_sig(struct io_ring_ctx *ctx) |
af9c1a44 | 2498 | { |
1414d629 | 2499 | if (!llist_empty(&ctx->work_llist)) { |
2f413956 | 2500 | __set_current_state(TASK_RUNNING); |
9fe3eaea | 2501 | if (io_run_local_work(ctx, INT_MAX) > 0) |
d246c759 | 2502 | return 0; |
1414d629 PB |
2503 | } |
2504 | if (io_run_task_work() > 0) | |
d246c759 | 2505 | return 0; |
c5020bc8 OL |
2506 | if (task_sigpending(current)) |
2507 | return -EINTR; | |
2508 | return 0; | |
af9c1a44 JA |
2509 | } |
2510 | ||
7b72d661 JA |
2511 | static bool current_pending_io(void) |
2512 | { | |
2513 | struct io_uring_task *tctx = current->io_uring; | |
2514 | ||
2515 | if (!tctx) | |
2516 | return false; | |
2517 | return percpu_counter_read_positive(&tctx->inflight); | |
2518 | } | |
2519 | ||
eeb60b9a PB |
2520 | /* when returns >0, the caller should retry */ |
2521 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
d33a39e5 | 2522 | struct io_wait_queue *iowq) |
eeb60b9a | 2523 | { |
7b72d661 | 2524 | int io_wait, ret; |
8a796565 | 2525 | |
3fcf19d5 PB |
2526 | if (unlikely(READ_ONCE(ctx->check_cq))) |
2527 | return 1; | |
846072f1 PB |
2528 | if (unlikely(!llist_empty(&ctx->work_llist))) |
2529 | return 1; | |
2530 | if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) | |
2531 | return 1; | |
2532 | if (unlikely(task_sigpending(current))) | |
2533 | return -EINTR; | |
2534 | if (unlikely(io_should_wake(iowq))) | |
2535 | return 0; | |
8a796565 AF |
2536 | |
2537 | /* | |
7b72d661 JA |
2538 | * Mark us as being in io_wait if we have pending requests, so cpufreq |
2539 | * can take into account that the task is waiting for IO - turns out | |
2540 | * to be important for low QD IO. | |
8a796565 | 2541 | */ |
7b72d661 JA |
2542 | io_wait = current->in_iowait; |
2543 | if (current_pending_io()) | |
2544 | current->in_iowait = 1; | |
8a796565 | 2545 | ret = 0; |
d33a39e5 | 2546 | if (iowq->timeout == KTIME_MAX) |
46ae7eef | 2547 | schedule(); |
d33a39e5 | 2548 | else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS)) |
8a796565 | 2549 | ret = -ETIME; |
7b72d661 | 2550 | current->in_iowait = io_wait; |
8a796565 | 2551 | return ret; |
eeb60b9a PB |
2552 | } |
2553 | ||
2b188cc1 JA |
2554 | /* |
2555 | * Wait until events become available, if we don't already have some. The | |
2556 | * application must reap them itself, as they reside on the shared cq ring. | |
2557 | */ | |
2558 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2559 | const sigset_t __user *sig, size_t sigsz, |
2560 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2561 | { |
90291099 | 2562 | struct io_wait_queue iowq; |
75b28aff | 2563 | struct io_rings *rings = ctx->rings; |
c1d5a224 | 2564 | int ret; |
2b188cc1 | 2565 | |
76de6749 PB |
2566 | if (!io_allowed_run_tw(ctx)) |
2567 | return -EEXIST; | |
140102ae | 2568 | if (!llist_empty(&ctx->work_llist)) |
9fe3eaea | 2569 | io_run_local_work(ctx, min_events); |
f36ba6cf PB |
2570 | io_run_task_work(); |
2571 | io_cqring_overflow_flush(ctx); | |
2572 | /* if user messes with these they will just get an early return */ | |
2573 | if (__io_cqring_events_user(ctx) >= min_events) | |
2574 | return 0; | |
2b188cc1 JA |
2575 | |
2576 | if (sig) { | |
9e75ad5d AB |
2577 | #ifdef CONFIG_COMPAT |
2578 | if (in_compat_syscall()) | |
2579 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2580 | sigsz); |
9e75ad5d AB |
2581 | else |
2582 | #endif | |
b772434b | 2583 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2584 | |
2b188cc1 JA |
2585 | if (ret) |
2586 | return ret; | |
2587 | } | |
2588 | ||
90291099 PB |
2589 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
2590 | iowq.wq.private = current; | |
2591 | INIT_LIST_HEAD(&iowq.wq.entry); | |
2592 | iowq.ctx = ctx; | |
bda52162 | 2593 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 2594 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
d33a39e5 PB |
2595 | iowq.timeout = KTIME_MAX; |
2596 | ||
2597 | if (uts) { | |
2598 | struct timespec64 ts; | |
2599 | ||
2600 | if (get_timespec64(&ts, uts)) | |
2601 | return -EFAULT; | |
2602 | iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); | |
2603 | } | |
90291099 | 2604 | |
c826bd7a | 2605 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2606 | do { |
9fe3eaea | 2607 | int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail); |
3fcf19d5 PB |
2608 | unsigned long check_cq; |
2609 | ||
130bd686 | 2610 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
8751d154 | 2611 | atomic_set(&ctx->cq_wait_nr, nr_wait); |
130bd686 PB |
2612 | set_current_state(TASK_INTERRUPTIBLE); |
2613 | } else { | |
2614 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, | |
2615 | TASK_INTERRUPTIBLE); | |
2616 | } | |
2617 | ||
d33a39e5 | 2618 | ret = io_cqring_wait_schedule(ctx, &iowq); |
130bd686 | 2619 | __set_current_state(TASK_RUNNING); |
b4bc35cf | 2620 | atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT); |
d80c0f00 | 2621 | |
846072f1 PB |
2622 | /* |
2623 | * Run task_work after scheduling and before io_should_wake(). | |
2624 | * If we got woken because of task_work being processed, run it | |
2625 | * now rather than let the caller do another wait loop. | |
2626 | */ | |
2627 | io_run_task_work(); | |
2628 | if (!llist_empty(&ctx->work_llist)) | |
9fe3eaea | 2629 | io_run_local_work(ctx, nr_wait); |
3fcf19d5 | 2630 | |
6ff1407e JA |
2631 | /* |
2632 | * Non-local task_work will be run on exit to userspace, but | |
2633 | * if we're using DEFER_TASKRUN, then we could have waited | |
2634 | * with a timeout for a number of requests. If the timeout | |
2635 | * hits, we could have some requests ready to process. Ensure | |
2636 | * this break is _after_ we have run task_work, to avoid | |
2637 | * deferring running potentially pending requests until the | |
2638 | * next time we wait for events. | |
2639 | */ | |
2640 | if (ret < 0) | |
2641 | break; | |
2642 | ||
3fcf19d5 PB |
2643 | check_cq = READ_ONCE(ctx->check_cq); |
2644 | if (unlikely(check_cq)) { | |
2645 | /* let the caller flush overflows, retry */ | |
326a9e48 | 2646 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) |
3fcf19d5 | 2647 | io_cqring_do_overflow_flush(ctx); |
3fcf19d5 PB |
2648 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) { |
2649 | ret = -EBADR; | |
2650 | break; | |
2651 | } | |
2652 | } | |
2653 | ||
846072f1 PB |
2654 | if (io_should_wake(&iowq)) { |
2655 | ret = 0; | |
35d90f95 | 2656 | break; |
846072f1 | 2657 | } |
ca0a2651 | 2658 | cond_resched(); |
846072f1 | 2659 | } while (1); |
bda52162 | 2660 | |
130bd686 PB |
2661 | if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) |
2662 | finish_wait(&ctx->cq_wait, &iowq.wq); | |
b7db41c9 | 2663 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2664 | |
75b28aff | 2665 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2666 | } |
2667 | ||
edecf168 | 2668 | void io_mem_free(void *ptr) |
b9bd2bea | 2669 | { |
73572984 JA |
2670 | if (!ptr) |
2671 | return; | |
b9bd2bea | 2672 | |
99a9e0b8 | 2673 | folio_put(virt_to_folio(ptr)); |
b9bd2bea PB |
2674 | } |
2675 | ||
03d89a2d JA |
2676 | static void io_pages_free(struct page ***pages, int npages) |
2677 | { | |
2678 | struct page **page_array; | |
2679 | int i; | |
2680 | ||
2681 | if (!pages) | |
2682 | return; | |
8b51a395 | 2683 | |
03d89a2d | 2684 | page_array = *pages; |
8b51a395 JA |
2685 | if (!page_array) |
2686 | return; | |
2687 | ||
03d89a2d JA |
2688 | for (i = 0; i < npages; i++) |
2689 | unpin_user_page(page_array[i]); | |
2690 | kvfree(page_array); | |
2691 | *pages = NULL; | |
2692 | } | |
2693 | ||
2694 | static void *__io_uaddr_map(struct page ***pages, unsigned short *npages, | |
2695 | unsigned long uaddr, size_t size) | |
2696 | { | |
2697 | struct page **page_array; | |
2698 | unsigned int nr_pages; | |
820d070f | 2699 | void *page_addr; |
223ef474 | 2700 | int ret, i; |
03d89a2d JA |
2701 | |
2702 | *npages = 0; | |
2703 | ||
2704 | if (uaddr & (PAGE_SIZE - 1) || !size) | |
2705 | return ERR_PTR(-EINVAL); | |
2706 | ||
2707 | nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2708 | if (nr_pages > USHRT_MAX) | |
2709 | return ERR_PTR(-EINVAL); | |
2710 | page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); | |
2711 | if (!page_array) | |
2712 | return ERR_PTR(-ENOMEM); | |
2713 | ||
2714 | ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM, | |
2715 | page_array); | |
2716 | if (ret != nr_pages) { | |
2717 | err: | |
2718 | io_pages_free(&page_array, ret > 0 ? ret : 0); | |
2719 | return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT); | |
2720 | } | |
223ef474 | 2721 | |
820d070f | 2722 | page_addr = page_address(page_array[0]); |
223ef474 | 2723 | for (i = 0; i < nr_pages; i++) { |
820d070f JA |
2724 | ret = -EINVAL; |
2725 | ||
2726 | /* | |
2727 | * Can't support mapping user allocated ring memory on 32-bit | |
2728 | * archs where it could potentially reside in highmem. Just | |
2729 | * fail those with -EINVAL, just like we did on kernels that | |
2730 | * didn't support this feature. | |
2731 | */ | |
2732 | if (PageHighMem(page_array[i])) | |
223ef474 | 2733 | goto err; |
820d070f JA |
2734 | |
2735 | /* | |
2736 | * No support for discontig pages for now, should either be a | |
2737 | * single normal page, or a huge page. Later on we can add | |
2738 | * support for remapping discontig pages, for now we will | |
2739 | * just fail them with EINVAL. | |
2740 | */ | |
2741 | if (page_address(page_array[i]) != page_addr) | |
2742 | goto err; | |
2743 | page_addr += PAGE_SIZE; | |
223ef474 JA |
2744 | } |
2745 | ||
03d89a2d JA |
2746 | *pages = page_array; |
2747 | *npages = nr_pages; | |
2748 | return page_to_virt(page_array[0]); | |
2749 | } | |
2750 | ||
2751 | static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr, | |
2752 | size_t size) | |
2753 | { | |
2754 | return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr, | |
2755 | size); | |
2756 | } | |
2757 | ||
2758 | static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr, | |
2759 | size_t size) | |
2760 | { | |
2761 | return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr, | |
2762 | size); | |
2763 | } | |
2764 | ||
9c189eee JA |
2765 | static void io_rings_free(struct io_ring_ctx *ctx) |
2766 | { | |
03d89a2d JA |
2767 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) { |
2768 | io_mem_free(ctx->rings); | |
2769 | io_mem_free(ctx->sq_sqes); | |
2770 | ctx->rings = NULL; | |
2771 | ctx->sq_sqes = NULL; | |
2772 | } else { | |
2773 | io_pages_free(&ctx->ring_pages, ctx->n_ring_pages); | |
8b51a395 | 2774 | ctx->n_ring_pages = 0; |
03d89a2d | 2775 | io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages); |
8b51a395 | 2776 | ctx->n_sqe_pages = 0; |
03d89a2d | 2777 | } |
9c189eee JA |
2778 | } |
2779 | ||
edecf168 | 2780 | void *io_mem_alloc(size_t size) |
b9bd2bea | 2781 | { |
73572984 | 2782 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; |
e27cef86 | 2783 | void *ret; |
b9bd2bea | 2784 | |
e27cef86 JA |
2785 | ret = (void *) __get_free_pages(gfp, get_order(size)); |
2786 | if (ret) | |
2787 | return ret; | |
2788 | return ERR_PTR(-ENOMEM); | |
b9bd2bea PB |
2789 | } |
2790 | ||
73572984 JA |
2791 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2792 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2793 | { |
73572984 JA |
2794 | struct io_rings *rings; |
2795 | size_t off, sq_array_size; | |
6b06314c | 2796 | |
73572984 JA |
2797 | off = struct_size(rings, cqes, cq_entries); |
2798 | if (off == SIZE_MAX) | |
2799 | return SIZE_MAX; | |
2800 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2801 | if (check_shl_overflow(off, 1, &off)) | |
2802 | return SIZE_MAX; | |
2803 | } | |
ab409402 | 2804 | |
73572984 JA |
2805 | #ifdef CONFIG_SMP |
2806 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2807 | if (off == 0) | |
2808 | return SIZE_MAX; | |
2809 | #endif | |
82fbcfa9 | 2810 | |
2af89abd PB |
2811 | if (ctx->flags & IORING_SETUP_NO_SQARRAY) { |
2812 | if (sq_offset) | |
2813 | *sq_offset = SIZE_MAX; | |
2814 | return off; | |
2815 | } | |
2816 | ||
73572984 JA |
2817 | if (sq_offset) |
2818 | *sq_offset = off; | |
82fbcfa9 | 2819 | |
73572984 JA |
2820 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2821 | if (sq_array_size == SIZE_MAX) | |
2822 | return SIZE_MAX; | |
6b06314c | 2823 | |
73572984 JA |
2824 | if (check_add_overflow(off, sq_array_size, &off)) |
2825 | return SIZE_MAX; | |
8bad28d8 | 2826 | |
73572984 | 2827 | return off; |
8bad28d8 HX |
2828 | } |
2829 | ||
73572984 | 2830 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2831 | { |
c8576f3e | 2832 | struct io_kiocb *req; |
37f0e767 | 2833 | int nr = 0; |
bf019da7 | 2834 | |
9a4fdbd8 | 2835 | mutex_lock(&ctx->uring_lock); |
34f0bc42 | 2836 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
9a4fdbd8 | 2837 | |
88ab95be | 2838 | while (!io_req_cache_empty(ctx)) { |
c8576f3e | 2839 | req = io_extract_req(ctx); |
c2b6c6bc | 2840 | kmem_cache_free(req_cachep, req); |
37f0e767 | 2841 | nr++; |
c2b6c6bc | 2842 | } |
37f0e767 PB |
2843 | if (nr) |
2844 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2845 | mutex_unlock(&ctx->uring_lock); |
2846 | } | |
2847 | ||
9eae8655 PB |
2848 | static void io_rsrc_node_cache_free(struct io_cache_entry *entry) |
2849 | { | |
2850 | kfree(container_of(entry, struct io_rsrc_node, cache)); | |
2851 | } | |
2852 | ||
c072481d | 2853 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2854 | { |
37d1e2e3 | 2855 | io_sq_thread_finish(ctx); |
43597aac | 2856 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
0b222eeb PB |
2857 | if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list))) |
2858 | return; | |
43597aac | 2859 | |
8bad28d8 | 2860 | mutex_lock(&ctx->uring_lock); |
43597aac | 2861 | if (ctx->buf_data) |
bd54b6fe | 2862 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2863 | if (ctx->file_data) |
08480400 | 2864 | __io_sqe_files_unregister(ctx); |
a85381d8 | 2865 | io_cqring_overflow_kill(ctx); |
9b402849 | 2866 | io_eventfd_unregister(ctx); |
9b797a37 | 2867 | io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); |
43e0bbbd | 2868 | io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); |
194bb58c | 2869 | io_futex_cache_free(ctx); |
5a2e745d | 2870 | io_destroy_buffers(ctx); |
b4a72c05 | 2871 | mutex_unlock(&ctx->uring_lock); |
07db298a PB |
2872 | if (ctx->sq_creds) |
2873 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2874 | if (ctx->submitter_task) |
2875 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2876 | |
a7f0ed5a PB |
2877 | /* there are no registered resources left, nobody uses it */ |
2878 | if (ctx->rsrc_node) | |
9eae8655 | 2879 | io_rsrc_node_destroy(ctx, ctx->rsrc_node); |
a7f0ed5a PB |
2880 | |
2881 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
ef9dd637 | 2882 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
2b188cc1 | 2883 | |
9eae8655 | 2884 | io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free); |
42b6419d PB |
2885 | if (ctx->mm_account) { |
2886 | mmdrop(ctx->mm_account); | |
2887 | ctx->mm_account = NULL; | |
2888 | } | |
9c189eee | 2889 | io_rings_free(ctx); |
c392cbec | 2890 | io_kbuf_mmap_list_free(ctx); |
2b188cc1 JA |
2891 | |
2892 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2893 | free_uid(ctx->user); |
4010fec4 | 2894 | io_req_caches_free(ctx); |
e941894e JA |
2895 | if (ctx->hash_map) |
2896 | io_wq_put_hash(ctx->hash_map); | |
e6f89be6 | 2897 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2898 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 JA |
2899 | kfree(ctx->io_bl); |
2900 | xa_destroy(&ctx->io_bl_xa); | |
2b188cc1 JA |
2901 | kfree(ctx); |
2902 | } | |
2903 | ||
bca39f39 PB |
2904 | static __cold void io_activate_pollwq_cb(struct callback_head *cb) |
2905 | { | |
2906 | struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx, | |
2907 | poll_wq_task_work); | |
2908 | ||
2909 | mutex_lock(&ctx->uring_lock); | |
2910 | ctx->poll_activated = true; | |
2911 | mutex_unlock(&ctx->uring_lock); | |
2912 | ||
2913 | /* | |
2914 | * Wake ups for some events between start of polling and activation | |
2915 | * might've been lost due to loose synchronisation. | |
2916 | */ | |
2917 | wake_up_all(&ctx->poll_wq); | |
2918 | percpu_ref_put(&ctx->refs); | |
2919 | } | |
2920 | ||
c4320315 | 2921 | __cold void io_activate_pollwq(struct io_ring_ctx *ctx) |
bca39f39 PB |
2922 | { |
2923 | spin_lock(&ctx->completion_lock); | |
2924 | /* already activated or in progress */ | |
2925 | if (ctx->poll_activated || ctx->poll_wq_task_work.func) | |
2926 | goto out; | |
2927 | if (WARN_ON_ONCE(!ctx->task_complete)) | |
2928 | goto out; | |
2929 | if (!ctx->submitter_task) | |
2930 | goto out; | |
2931 | /* | |
2932 | * with ->submitter_task only the submitter task completes requests, we | |
2933 | * only need to sync with it, which is done by injecting a tw | |
2934 | */ | |
2935 | init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb); | |
2936 | percpu_ref_get(&ctx->refs); | |
2937 | if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL)) | |
2938 | percpu_ref_put(&ctx->refs); | |
2939 | out: | |
2940 | spin_unlock(&ctx->completion_lock); | |
2941 | } | |
2942 | ||
2b188cc1 JA |
2943 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) |
2944 | { | |
2945 | struct io_ring_ctx *ctx = file->private_data; | |
2946 | __poll_t mask = 0; | |
2947 | ||
bca39f39 PB |
2948 | if (unlikely(!ctx->poll_activated)) |
2949 | io_activate_pollwq(ctx); | |
2950 | ||
7b235dd8 | 2951 | poll_wait(file, &ctx->poll_wq, wait); |
4f7067c3 SB |
2952 | /* |
2953 | * synchronizes with barrier from wq_has_sleeper call in | |
2954 | * io_commit_cqring | |
2955 | */ | |
2b188cc1 | 2956 | smp_rmb(); |
90554200 | 2957 | if (!io_sqring_full(ctx)) |
2b188cc1 | 2958 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
2959 | |
2960 | /* | |
2961 | * Don't flush cqring overflow list here, just do a simple check. | |
2962 | * Otherwise there could possible be ABBA deadlock: | |
2963 | * CPU0 CPU1 | |
2964 | * ---- ---- | |
2965 | * lock(&ctx->uring_lock); | |
2966 | * lock(&ep->mtx); | |
2967 | * lock(&ctx->uring_lock); | |
2968 | * lock(&ep->mtx); | |
2969 | * | |
2970 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
10d8bc35 | 2971 | * pushes them to do the flush. |
ed670c3f | 2972 | */ |
b4c98d59 | 2973 | |
c10bb646 | 2974 | if (__io_cqring_events_user(ctx) || io_has_work(ctx)) |
2b188cc1 JA |
2975 | mask |= EPOLLIN | EPOLLRDNORM; |
2976 | ||
2977 | return mask; | |
2978 | } | |
2979 | ||
d56d938b PB |
2980 | struct io_tctx_exit { |
2981 | struct callback_head task_work; | |
2982 | struct completion completion; | |
baf186c4 | 2983 | struct io_ring_ctx *ctx; |
d56d938b PB |
2984 | }; |
2985 | ||
c072481d | 2986 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
2987 | { |
2988 | struct io_uring_task *tctx = current->io_uring; | |
2989 | struct io_tctx_exit *work; | |
2990 | ||
2991 | work = container_of(cb, struct io_tctx_exit, task_work); | |
2992 | /* | |
8d664282 | 2993 | * When @in_cancel, we're in cancellation and it's racy to remove the |
d56d938b | 2994 | * node. It'll be removed by the end of cancellation, just ignore it. |
998b30c3 HM |
2995 | * tctx can be NULL if the queueing of this task_work raced with |
2996 | * work cancelation off the exec path. | |
d56d938b | 2997 | */ |
8d664282 | 2998 | if (tctx && !atomic_read(&tctx->in_cancel)) |
eef51daa | 2999 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
3000 | complete(&work->completion); |
3001 | } | |
3002 | ||
c072481d | 3003 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
3004 | { |
3005 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
3006 | ||
3007 | return req->ctx == data; | |
3008 | } | |
3009 | ||
c072481d | 3010 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 3011 | { |
d56d938b | 3012 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 3013 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 3014 | unsigned long interval = HZ / 20; |
d56d938b PB |
3015 | struct io_tctx_exit exit; |
3016 | struct io_tctx_node *node; | |
3017 | int ret; | |
85faa7b8 | 3018 | |
56952e91 JA |
3019 | /* |
3020 | * If we're doing polled IO and end up having requests being | |
3021 | * submitted async (out-of-line), then completions can come in while | |
3022 | * we're waiting for refs to drop. We need to reap these manually, | |
3023 | * as nobody else will be looking for them. | |
3024 | */ | |
b2edc0a7 | 3025 | do { |
a85381d8 PB |
3026 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
3027 | mutex_lock(&ctx->uring_lock); | |
3028 | io_cqring_overflow_kill(ctx); | |
3029 | mutex_unlock(&ctx->uring_lock); | |
3030 | } | |
3031 | ||
c0e0d6ba DY |
3032 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) |
3033 | io_move_task_work_from_local(ctx); | |
3034 | ||
affa87db PB |
3035 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
3036 | cond_resched(); | |
3037 | ||
28090c13 PB |
3038 | if (ctx->sq_data) { |
3039 | struct io_sq_data *sqd = ctx->sq_data; | |
3040 | struct task_struct *tsk; | |
3041 | ||
3042 | io_sq_thread_park(sqd); | |
3043 | tsk = sqd->thread; | |
3044 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
3045 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
3046 | io_cancel_ctx_cb, ctx, true); | |
3047 | io_sq_thread_unpark(sqd); | |
3048 | } | |
b5bb3a24 | 3049 | |
37f0e767 PB |
3050 | io_req_caches_free(ctx); |
3051 | ||
58d3be2c PB |
3052 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
3053 | /* there is little hope left, don't run it too often */ | |
3054 | interval = HZ * 60; | |
3055 | } | |
4826c594 JA |
3056 | /* |
3057 | * This is really an uninterruptible wait, as it has to be | |
3058 | * complete. But it's also run from a kworker, which doesn't | |
3059 | * take signals, so it's fine to make it interruptible. This | |
3060 | * avoids scenarios where we knowingly can wait much longer | |
3061 | * on completions, for example if someone does a SIGSTOP on | |
3062 | * a task that needs to finish task_work to make this loop | |
3063 | * complete. That's a synthetic situation that should not | |
3064 | * cause a stuck task backtrace, and hence a potential panic | |
3065 | * on stuck tasks if that is enabled. | |
3066 | */ | |
3067 | } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 3068 | |
7f00651a PB |
3069 | init_completion(&exit.completion); |
3070 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
3071 | exit.ctx = ctx; | |
f7b32e78 | 3072 | |
d56d938b PB |
3073 | mutex_lock(&ctx->uring_lock); |
3074 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
3075 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
3076 | ||
d56d938b PB |
3077 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
3078 | ctx_node); | |
7f00651a PB |
3079 | /* don't spin on a single task if cancellation failed */ |
3080 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
3081 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
3082 | if (WARN_ON_ONCE(ret)) | |
3083 | continue; | |
d56d938b PB |
3084 | |
3085 | mutex_unlock(&ctx->uring_lock); | |
4826c594 JA |
3086 | /* |
3087 | * See comment above for | |
3088 | * wait_for_completion_interruptible_timeout() on why this | |
3089 | * wait is marked as interruptible. | |
3090 | */ | |
3091 | wait_for_completion_interruptible(&exit.completion); | |
d56d938b PB |
3092 | mutex_lock(&ctx->uring_lock); |
3093 | } | |
3094 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
3095 | spin_lock(&ctx->completion_lock); |
3096 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 3097 | |
d73a572d PB |
3098 | /* pairs with RCU read section in io_req_local_work_add() */ |
3099 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3100 | synchronize_rcu(); | |
3101 | ||
85faa7b8 JA |
3102 | io_ring_ctx_free(ctx); |
3103 | } | |
3104 | ||
c072481d | 3105 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 3106 | { |
61cf9370 MWO |
3107 | unsigned long index; |
3108 | struct creds *creds; | |
3109 | ||
2b188cc1 JA |
3110 | mutex_lock(&ctx->uring_lock); |
3111 | percpu_ref_kill(&ctx->refs); | |
61cf9370 MWO |
3112 | xa_for_each(&ctx->personalities, index, creds) |
3113 | io_unregister_personality(ctx, index); | |
9ca9fb24 PB |
3114 | if (ctx->rings) |
3115 | io_poll_remove_all(ctx, NULL, true); | |
2b188cc1 JA |
3116 | mutex_unlock(&ctx->uring_lock); |
3117 | ||
02bac94b PB |
3118 | /* |
3119 | * If we failed setting up the ctx, we might not have any rings | |
3120 | * and therefore did not submit any requests | |
3121 | */ | |
3122 | if (ctx->rings) | |
60053be8 | 3123 | io_kill_timeouts(ctx, NULL, true); |
309fc03a | 3124 | |
dfbe5561 JA |
3125 | flush_delayed_work(&ctx->fallback_work); |
3126 | ||
85faa7b8 | 3127 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
3128 | /* |
3129 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
3130 | * if we're exiting a ton of rings at the same time. It just adds | |
3131 | * noise and overhead, there's no discernable change in runtime | |
3132 | * over using system_wq. | |
3133 | */ | |
3134 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
3135 | } |
3136 | ||
3137 | static int io_uring_release(struct inode *inode, struct file *file) | |
3138 | { | |
3139 | struct io_ring_ctx *ctx = file->private_data; | |
3140 | ||
3141 | file->private_data = NULL; | |
3142 | io_ring_ctx_wait_and_kill(ctx); | |
3143 | return 0; | |
3144 | } | |
3145 | ||
f6edbabb PB |
3146 | struct io_task_cancel { |
3147 | struct task_struct *task; | |
3dd0c97a | 3148 | bool all; |
f6edbabb | 3149 | }; |
f254ac04 | 3150 | |
f6edbabb | 3151 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 3152 | { |
9a472ef7 | 3153 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 3154 | struct io_task_cancel *cancel = data; |
9a472ef7 | 3155 | |
6af3f48b | 3156 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
3157 | } |
3158 | ||
c072481d PB |
3159 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
3160 | struct task_struct *task, | |
3161 | bool cancel_all) | |
b7ddce3c | 3162 | { |
e1915f76 | 3163 | struct io_defer_entry *de; |
b7ddce3c PB |
3164 | LIST_HEAD(list); |
3165 | ||
79ebeaee | 3166 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 3167 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 3168 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
3169 | list_cut_position(&list, &ctx->defer_list, &de->list); |
3170 | break; | |
3171 | } | |
3172 | } | |
79ebeaee | 3173 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
3174 | if (list_empty(&list)) |
3175 | return false; | |
b7ddce3c PB |
3176 | |
3177 | while (!list_empty(&list)) { | |
3178 | de = list_first_entry(&list, struct io_defer_entry, list); | |
3179 | list_del_init(&de->list); | |
e276ae34 | 3180 | io_req_task_queue_fail(de->req, -ECANCELED); |
b7ddce3c PB |
3181 | kfree(de); |
3182 | } | |
e1915f76 | 3183 | return true; |
b7ddce3c PB |
3184 | } |
3185 | ||
c072481d | 3186 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
3187 | { |
3188 | struct io_tctx_node *node; | |
3189 | enum io_wq_cancel cret; | |
3190 | bool ret = false; | |
3191 | ||
3192 | mutex_lock(&ctx->uring_lock); | |
3193 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3194 | struct io_uring_task *tctx = node->task->io_uring; | |
3195 | ||
3196 | /* | |
3197 | * io_wq will stay alive while we hold uring_lock, because it's | |
3198 | * killed after ctx nodes, which requires to take the lock. | |
3199 | */ | |
3200 | if (!tctx || !tctx->io_wq) | |
3201 | continue; | |
3202 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
3203 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3204 | } | |
3205 | mutex_unlock(&ctx->uring_lock); | |
3206 | ||
3207 | return ret; | |
3208 | } | |
3209 | ||
93b8cc60 ML |
3210 | static bool io_uring_try_cancel_uring_cmd(struct io_ring_ctx *ctx, |
3211 | struct task_struct *task, bool cancel_all) | |
3212 | { | |
3213 | struct hlist_node *tmp; | |
3214 | struct io_kiocb *req; | |
3215 | bool ret = false; | |
3216 | ||
3217 | lockdep_assert_held(&ctx->uring_lock); | |
3218 | ||
3219 | hlist_for_each_entry_safe(req, tmp, &ctx->cancelable_uring_cmd, | |
3220 | hash_node) { | |
3221 | struct io_uring_cmd *cmd = io_kiocb_to_cmd(req, | |
3222 | struct io_uring_cmd); | |
3223 | struct file *file = req->file; | |
3224 | ||
3225 | if (!cancel_all && req->task != task) | |
3226 | continue; | |
3227 | ||
3228 | if (cmd->flags & IORING_URING_CMD_CANCELABLE) { | |
3229 | /* ->sqe isn't available if no async data */ | |
3230 | if (!req_has_async_data(req)) | |
3231 | cmd->sqe = NULL; | |
3232 | file->f_op->uring_cmd(cmd, IO_URING_F_CANCEL); | |
3233 | ret = true; | |
3234 | } | |
3235 | } | |
3236 | io_submit_flush_completions(ctx); | |
3237 | ||
3238 | return ret; | |
3239 | } | |
3240 | ||
affa87db | 3241 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
3242 | struct task_struct *task, |
3243 | bool cancel_all) | |
9936c7c2 | 3244 | { |
3dd0c97a | 3245 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 3246 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
3247 | enum io_wq_cancel cret; |
3248 | bool ret = false; | |
9936c7c2 | 3249 | |
360cd42c PB |
3250 | /* set it so io_req_local_work_add() would wake us up */ |
3251 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { | |
3252 | atomic_set(&ctx->cq_wait_nr, 1); | |
3253 | smp_mb(); | |
3254 | } | |
3255 | ||
60053be8 PB |
3256 | /* failed during ring init, it couldn't have issued any requests */ |
3257 | if (!ctx->rings) | |
affa87db | 3258 | return false; |
60053be8 | 3259 | |
affa87db PB |
3260 | if (!task) { |
3261 | ret |= io_uring_try_cancel_iowq(ctx); | |
3262 | } else if (tctx && tctx->io_wq) { | |
3263 | /* | |
3264 | * Cancels requests of all rings, not only @ctx, but | |
3265 | * it's fine as the task is in exit/exec. | |
3266 | */ | |
3267 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
3268 | &cancel, true); | |
3269 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3270 | } | |
9936c7c2 | 3271 | |
affa87db PB |
3272 | /* SQPOLL thread does its own polling */ |
3273 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
3274 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
3275 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
3276 | io_iopoll_try_reap_events(ctx); | |
3277 | ret = true; | |
fcc926bb | 3278 | cond_resched(); |
9936c7c2 | 3279 | } |
9936c7c2 | 3280 | } |
affa87db | 3281 | |
140102ae PB |
3282 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3283 | io_allowed_defer_tw_run(ctx)) | |
9fe3eaea | 3284 | ret |= io_run_local_work(ctx, INT_MAX) > 0; |
affa87db PB |
3285 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
3286 | mutex_lock(&ctx->uring_lock); | |
3287 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
f31ecf67 | 3288 | ret |= io_waitid_remove_all(ctx, task, cancel_all); |
194bb58c | 3289 | ret |= io_futex_remove_all(ctx, task, cancel_all); |
93b8cc60 | 3290 | ret |= io_uring_try_cancel_uring_cmd(ctx, task, cancel_all); |
affa87db PB |
3291 | mutex_unlock(&ctx->uring_lock); |
3292 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
3293 | if (task) | |
c0e0d6ba | 3294 | ret |= io_run_task_work() > 0; |
affa87db | 3295 | return ret; |
9936c7c2 PB |
3296 | } |
3297 | ||
3f48cf18 | 3298 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 3299 | { |
3f48cf18 | 3300 | if (tracked) |
9cae36a0 | 3301 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
3302 | return percpu_counter_sum(&tctx->inflight); |
3303 | } | |
3304 | ||
78cc687b PB |
3305 | /* |
3306 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 3307 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 3308 | */ |
17437f31 | 3309 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 3310 | { |
521d6a73 | 3311 | struct io_uring_task *tctx = current->io_uring; |
734551df | 3312 | struct io_ring_ctx *ctx; |
360cd42c PB |
3313 | struct io_tctx_node *node; |
3314 | unsigned long index; | |
0e9ddb39 PB |
3315 | s64 inflight; |
3316 | DEFINE_WAIT(wait); | |
fdaf083c | 3317 | |
78cc687b PB |
3318 | WARN_ON_ONCE(sqd && sqd->thread != current); |
3319 | ||
6d042ffb PO |
3320 | if (!current->io_uring) |
3321 | return; | |
17a91051 PB |
3322 | if (tctx->io_wq) |
3323 | io_wq_exit_start(tctx->io_wq); | |
3324 | ||
8d664282 | 3325 | atomic_inc(&tctx->in_cancel); |
0e9ddb39 | 3326 | do { |
affa87db PB |
3327 | bool loop = false; |
3328 | ||
e9dbe221 | 3329 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 3330 | /* read completions before cancelations */ |
78cc687b | 3331 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
3332 | if (!inflight) |
3333 | break; | |
fdaf083c | 3334 | |
78cc687b | 3335 | if (!sqd) { |
78cc687b PB |
3336 | xa_for_each(&tctx->xa, index, node) { |
3337 | /* sqpoll task will cancel all its requests */ | |
3338 | if (node->ctx->sq_data) | |
3339 | continue; | |
affa87db PB |
3340 | loop |= io_uring_try_cancel_requests(node->ctx, |
3341 | current, cancel_all); | |
78cc687b PB |
3342 | } |
3343 | } else { | |
3344 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
3345 | loop |= io_uring_try_cancel_requests(ctx, |
3346 | current, | |
3347 | cancel_all); | |
3348 | } | |
3349 | ||
3350 | if (loop) { | |
3351 | cond_resched(); | |
3352 | continue; | |
78cc687b | 3353 | } |
17a91051 | 3354 | |
78a78060 JA |
3355 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
3356 | io_run_task_work(); | |
e9dbe221 | 3357 | io_uring_drop_tctx_refs(current); |
360cd42c PB |
3358 | xa_for_each(&tctx->xa, index, node) { |
3359 | if (!llist_empty(&node->ctx->work_llist)) { | |
3360 | WARN_ON_ONCE(node->ctx->submitter_task && | |
3361 | node->ctx->submitter_task != current); | |
3362 | goto end_wait; | |
3363 | } | |
3364 | } | |
0f212204 | 3365 | /* |
a1bb3cd5 PB |
3366 | * If we've seen completions, retry without waiting. This |
3367 | * avoids a race where a completion comes in before we did | |
3368 | * prepare_to_wait(). | |
0f212204 | 3369 | */ |
3dd0c97a | 3370 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 3371 | schedule(); |
360cd42c | 3372 | end_wait: |
f57555ed | 3373 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 3374 | } while (1); |
de7f1d9e | 3375 | |
8452d4a6 | 3376 | io_uring_clean_tctx(tctx); |
3dd0c97a | 3377 | if (cancel_all) { |
3cc7fdb9 PB |
3378 | /* |
3379 | * We shouldn't run task_works after cancel, so just leave | |
8d664282 | 3380 | * ->in_cancel set for normal exit. |
3cc7fdb9 | 3381 | */ |
8d664282 | 3382 | atomic_dec(&tctx->in_cancel); |
3f48cf18 PB |
3383 | /* for exec all current's requests should be gone, kill tctx */ |
3384 | __io_uring_free(current); | |
3385 | } | |
44e728b8 PB |
3386 | } |
3387 | ||
f552a27a | 3388 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 3389 | { |
f552a27a | 3390 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
3391 | } |
3392 | ||
6c5c240e RP |
3393 | static void *io_uring_validate_mmap_request(struct file *file, |
3394 | loff_t pgoff, size_t sz) | |
2b188cc1 | 3395 | { |
2b188cc1 | 3396 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 3397 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
3398 | struct page *page; |
3399 | void *ptr; | |
3400 | ||
c56e022c | 3401 | switch (offset & IORING_OFF_MMAP_MASK) { |
2b188cc1 | 3402 | case IORING_OFF_SQ_RING: |
75b28aff | 3403 | case IORING_OFF_CQ_RING: |
6f007b14 JA |
3404 | /* Don't allow mmap if the ring was setup without it */ |
3405 | if (ctx->flags & IORING_SETUP_NO_MMAP) | |
3406 | return ERR_PTR(-EINVAL); | |
75b28aff | 3407 | ptr = ctx->rings; |
2b188cc1 JA |
3408 | break; |
3409 | case IORING_OFF_SQES: | |
6f007b14 JA |
3410 | /* Don't allow mmap if the ring was setup without it */ |
3411 | if (ctx->flags & IORING_SETUP_NO_MMAP) | |
3412 | return ERR_PTR(-EINVAL); | |
2b188cc1 JA |
3413 | ptr = ctx->sq_sqes; |
3414 | break; | |
c56e022c JA |
3415 | case IORING_OFF_PBUF_RING: { |
3416 | unsigned int bgid; | |
3417 | ||
3418 | bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT; | |
5cf4f52e | 3419 | rcu_read_lock(); |
c56e022c | 3420 | ptr = io_pbuf_get_address(ctx, bgid); |
5cf4f52e | 3421 | rcu_read_unlock(); |
c56e022c JA |
3422 | if (!ptr) |
3423 | return ERR_PTR(-EINVAL); | |
3424 | break; | |
3425 | } | |
2b188cc1 | 3426 | default: |
6c5c240e | 3427 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
3428 | } |
3429 | ||
3430 | page = virt_to_head_page(ptr); | |
a50b854e | 3431 | if (sz > page_size(page)) |
6c5c240e RP |
3432 | return ERR_PTR(-EINVAL); |
3433 | ||
3434 | return ptr; | |
3435 | } | |
3436 | ||
3437 | #ifdef CONFIG_MMU | |
3438 | ||
c072481d | 3439 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e RP |
3440 | { |
3441 | size_t sz = vma->vm_end - vma->vm_start; | |
3442 | unsigned long pfn; | |
3443 | void *ptr; | |
3444 | ||
3445 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
3446 | if (IS_ERR(ptr)) | |
3447 | return PTR_ERR(ptr); | |
2b188cc1 JA |
3448 | |
3449 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3450 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3451 | } | |
3452 | ||
d808459b HD |
3453 | static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp, |
3454 | unsigned long addr, unsigned long len, | |
3455 | unsigned long pgoff, unsigned long flags) | |
3456 | { | |
d808459b HD |
3457 | void *ptr; |
3458 | ||
3459 | /* | |
3460 | * Do not allow to map to user-provided address to avoid breaking the | |
3461 | * aliasing rules. Userspace is not able to guess the offset address of | |
3462 | * kernel kmalloc()ed memory area. | |
3463 | */ | |
3464 | if (addr) | |
3465 | return -EINVAL; | |
3466 | ||
3467 | ptr = io_uring_validate_mmap_request(filp, pgoff, len); | |
3468 | if (IS_ERR(ptr)) | |
3469 | return -ENOMEM; | |
3470 | ||
32832a40 HD |
3471 | /* |
3472 | * Some architectures have strong cache aliasing requirements. | |
3473 | * For such architectures we need a coherent mapping which aliases | |
3474 | * kernel memory *and* userspace memory. To achieve that: | |
3475 | * - use a NULL file pointer to reference physical memory, and | |
3476 | * - use the kernel virtual address of the shared io_uring context | |
3477 | * (instead of the userspace-provided address, which has to be 0UL | |
3478 | * anyway). | |
56675f8b HD |
3479 | * - use the same pgoff which the get_unmapped_area() uses to |
3480 | * calculate the page colouring. | |
32832a40 HD |
3481 | * For architectures without such aliasing requirements, the |
3482 | * architecture will return any suitable mapping because addr is 0. | |
3483 | */ | |
3484 | filp = NULL; | |
3485 | flags |= MAP_SHARED; | |
3486 | pgoff = 0; /* has been translated to ptr above */ | |
d808459b | 3487 | #ifdef SHM_COLOUR |
32832a40 | 3488 | addr = (uintptr_t) ptr; |
56675f8b | 3489 | pgoff = addr >> PAGE_SHIFT; |
d808459b | 3490 | #else |
32832a40 | 3491 | addr = 0UL; |
d808459b | 3492 | #endif |
32832a40 | 3493 | return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); |
d808459b HD |
3494 | } |
3495 | ||
6c5c240e RP |
3496 | #else /* !CONFIG_MMU */ |
3497 | ||
3498 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3499 | { | |
fc4f4be9 | 3500 | return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL; |
6c5c240e RP |
3501 | } |
3502 | ||
3503 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
3504 | { | |
3505 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
3506 | } | |
3507 | ||
3508 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
3509 | unsigned long addr, unsigned long len, | |
3510 | unsigned long pgoff, unsigned long flags) | |
3511 | { | |
3512 | void *ptr; | |
3513 | ||
3514 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
3515 | if (IS_ERR(ptr)) | |
3516 | return PTR_ERR(ptr); | |
3517 | ||
3518 | return (unsigned long) ptr; | |
3519 | } | |
3520 | ||
3521 | #endif /* !CONFIG_MMU */ | |
3522 | ||
f81440d3 PB |
3523 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
3524 | { | |
3525 | if (flags & IORING_ENTER_EXT_ARG) { | |
3526 | struct io_uring_getevents_arg arg; | |
3527 | ||
3528 | if (argsz != sizeof(arg)) | |
3529 | return -EINVAL; | |
3530 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3531 | return -EFAULT; | |
3532 | } | |
3533 | return 0; | |
3534 | } | |
3535 | ||
c73ebb68 HX |
3536 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
3537 | struct __kernel_timespec __user **ts, | |
3538 | const sigset_t __user **sig) | |
3539 | { | |
3540 | struct io_uring_getevents_arg arg; | |
3541 | ||
3542 | /* | |
3543 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
3544 | * is just a pointer to the sigset_t. | |
3545 | */ | |
3546 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
3547 | *sig = (const sigset_t __user *) argp; | |
3548 | *ts = NULL; | |
3549 | return 0; | |
3550 | } | |
3551 | ||
3552 | /* | |
3553 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
3554 | * timespec and sigset_t pointers if good. | |
3555 | */ | |
3556 | if (*argsz != sizeof(arg)) | |
3557 | return -EINVAL; | |
3558 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3559 | return -EFAULT; | |
d2347b96 DY |
3560 | if (arg.pad) |
3561 | return -EINVAL; | |
c73ebb68 HX |
3562 | *sig = u64_to_user_ptr(arg.sigmask); |
3563 | *argsz = arg.sigmask_sz; | |
3564 | *ts = u64_to_user_ptr(arg.ts); | |
3565 | return 0; | |
3566 | } | |
3567 | ||
2b188cc1 | 3568 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
3569 | u32, min_complete, u32, flags, const void __user *, argp, |
3570 | size_t, argsz) | |
2b188cc1 JA |
3571 | { |
3572 | struct io_ring_ctx *ctx; | |
73363c26 | 3573 | struct file *file; |
33f993da | 3574 | long ret; |
2b188cc1 | 3575 | |
33f993da | 3576 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
3577 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
3578 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
3579 | return -EINVAL; |
3580 | ||
e7a6c00d JA |
3581 | /* |
3582 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3583 | * need only dereference our task private array to find it. | |
3584 | */ | |
3585 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3586 | struct io_uring_task *tctx = current->io_uring; | |
3587 | ||
3273c440 | 3588 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
e7a6c00d JA |
3589 | return -EINVAL; |
3590 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
73363c26 JA |
3591 | file = tctx->registered_rings[fd]; |
3592 | if (unlikely(!file)) | |
3273c440 | 3593 | return -EBADF; |
e7a6c00d | 3594 | } else { |
73363c26 JA |
3595 | file = fget(fd); |
3596 | if (unlikely(!file)) | |
3273c440 PB |
3597 | return -EBADF; |
3598 | ret = -EOPNOTSUPP; | |
73363c26 | 3599 | if (unlikely(!io_is_uring_fops(file))) |
fbb8bb02 | 3600 | goto out; |
e7a6c00d | 3601 | } |
2b188cc1 | 3602 | |
73363c26 | 3603 | ctx = file->private_data; |
7e84e1c7 | 3604 | ret = -EBADFD; |
33f993da | 3605 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3606 | goto out; |
3607 | ||
6c271ce2 JA |
3608 | /* |
3609 | * For SQ polling, the thread will do all submissions and completions. | |
3610 | * Just return the requested submit count, and wake the thread if | |
3611 | * we were asked to. | |
3612 | */ | |
b2a9eada | 3613 | ret = 0; |
6c271ce2 | 3614 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3615 | io_cqring_overflow_flush(ctx); |
89448c47 | 3616 | |
21f96522 JA |
3617 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3618 | ret = -EOWNERDEAD; | |
04147488 | 3619 | goto out; |
21f96522 | 3620 | } |
6c271ce2 | 3621 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3622 | wake_up(&ctx->sq_data->wait); |
88b80534 QF |
3623 | if (flags & IORING_ENTER_SQ_WAIT) |
3624 | io_sqpoll_wait_sq(ctx); | |
3625 | ||
3e813c90 | 3626 | ret = to_submit; |
b2a9eada | 3627 | } else if (to_submit) { |
eef51daa | 3628 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3629 | if (unlikely(ret)) |
3630 | goto out; | |
7c504e65 | 3631 | |
2b188cc1 | 3632 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3633 | ret = io_submit_sqes(ctx, to_submit); |
3634 | if (ret != to_submit) { | |
d487b43c | 3635 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3636 | goto out; |
d487b43c | 3637 | } |
44f87745 PB |
3638 | if (flags & IORING_ENTER_GETEVENTS) { |
3639 | if (ctx->syscall_iopoll) | |
3640 | goto iopoll_locked; | |
3641 | /* | |
3642 | * Ignore errors, we'll soon call io_cqring_wait() and | |
3643 | * it should handle ownership problems if any. | |
3644 | */ | |
3645 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
9fe3eaea | 3646 | (void)io_run_local_work_locked(ctx, min_complete); |
44f87745 | 3647 | } |
d487b43c | 3648 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 | 3649 | } |
c0e0d6ba | 3650 | |
2b188cc1 | 3651 | if (flags & IORING_ENTER_GETEVENTS) { |
3e813c90 | 3652 | int ret2; |
c0e0d6ba | 3653 | |
773697b6 | 3654 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3655 | /* |
3656 | * We disallow the app entering submit/complete with | |
3657 | * polling, but we still need to lock the ring to | |
3658 | * prevent racing with polled issue that got punted to | |
3659 | * a workqueue. | |
3660 | */ | |
3661 | mutex_lock(&ctx->uring_lock); | |
3662 | iopoll_locked: | |
3e813c90 DY |
3663 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3664 | if (likely(!ret2)) { | |
3665 | min_complete = min(min_complete, | |
3666 | ctx->cq_entries); | |
3667 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3668 | } |
3669 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3670 | } else { |
f81440d3 PB |
3671 | const sigset_t __user *sig; |
3672 | struct __kernel_timespec __user *ts; | |
3673 | ||
3e813c90 DY |
3674 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3675 | if (likely(!ret2)) { | |
3676 | min_complete = min(min_complete, | |
3677 | ctx->cq_entries); | |
3678 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3679 | argsz, ts); | |
3680 | } | |
def596e9 | 3681 | } |
c73ebb68 | 3682 | |
155bc950 | 3683 | if (!ret) { |
3e813c90 | 3684 | ret = ret2; |
2b188cc1 | 3685 | |
155bc950 DY |
3686 | /* |
3687 | * EBADR indicates that one or more CQE were dropped. | |
3688 | * Once the user has been informed we can clear the bit | |
3689 | * as they are obviously ok with those drops. | |
3690 | */ | |
3691 | if (unlikely(ret2 == -EBADR)) | |
3692 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3693 | &ctx->check_cq); | |
def596e9 | 3694 | } |
2b188cc1 | 3695 | } |
7c504e65 | 3696 | out: |
73363c26 JA |
3697 | if (!(flags & IORING_ENTER_REGISTERED_RING)) |
3698 | fput(file); | |
3e813c90 | 3699 | return ret; |
2b188cc1 JA |
3700 | } |
3701 | ||
3702 | static const struct file_operations io_uring_fops = { | |
3703 | .release = io_uring_release, | |
3704 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3705 | #ifndef CONFIG_MMU |
3706 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3707 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
d808459b HD |
3708 | #else |
3709 | .get_unmapped_area = io_uring_mmu_get_unmapped_area, | |
6c5c240e | 3710 | #endif |
2b188cc1 | 3711 | .poll = io_uring_poll, |
bebdb65e | 3712 | #ifdef CONFIG_PROC_FS |
87ce955b | 3713 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3714 | #endif |
2b188cc1 JA |
3715 | }; |
3716 | ||
92ac8bea JA |
3717 | bool io_is_uring_fops(struct file *file) |
3718 | { | |
3719 | return file->f_op == &io_uring_fops; | |
3720 | } | |
3721 | ||
c072481d PB |
3722 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3723 | struct io_uring_params *p) | |
2b188cc1 | 3724 | { |
75b28aff HV |
3725 | struct io_rings *rings; |
3726 | size_t size, sq_array_offset; | |
e27cef86 | 3727 | void *ptr; |
2b188cc1 | 3728 | |
bd740481 JA |
3729 | /* make sure these are sane, as we already accounted them */ |
3730 | ctx->sq_entries = p->sq_entries; | |
3731 | ctx->cq_entries = p->cq_entries; | |
3732 | ||
baf9cb64 | 3733 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3734 | if (size == SIZE_MAX) |
3735 | return -EOVERFLOW; | |
3736 | ||
03d89a2d JA |
3737 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3738 | rings = io_mem_alloc(size); | |
3739 | else | |
3740 | rings = io_rings_map(ctx, p->cq_off.user_addr, size); | |
3741 | ||
e27cef86 JA |
3742 | if (IS_ERR(rings)) |
3743 | return PTR_ERR(rings); | |
2b188cc1 | 3744 | |
75b28aff | 3745 | ctx->rings = rings; |
2af89abd PB |
3746 | if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) |
3747 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
75b28aff HV |
3748 | rings->sq_ring_mask = p->sq_entries - 1; |
3749 | rings->cq_ring_mask = p->cq_entries - 1; | |
3750 | rings->sq_ring_entries = p->sq_entries; | |
3751 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3752 | |
ebdeb7c0 JA |
3753 | if (p->flags & IORING_SETUP_SQE128) |
3754 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3755 | else | |
3756 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 | 3757 | if (size == SIZE_MAX) { |
9c189eee | 3758 | io_rings_free(ctx); |
2b188cc1 | 3759 | return -EOVERFLOW; |
eb065d30 | 3760 | } |
2b188cc1 | 3761 | |
03d89a2d JA |
3762 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3763 | ptr = io_mem_alloc(size); | |
3764 | else | |
3765 | ptr = io_sqes_map(ctx, p->sq_off.user_addr, size); | |
3766 | ||
e27cef86 | 3767 | if (IS_ERR(ptr)) { |
9c189eee | 3768 | io_rings_free(ctx); |
e27cef86 | 3769 | return PTR_ERR(ptr); |
eb065d30 | 3770 | } |
2b188cc1 | 3771 | |
e27cef86 | 3772 | ctx->sq_sqes = ptr; |
2b188cc1 JA |
3773 | return 0; |
3774 | } | |
3775 | ||
6e76ac59 | 3776 | static int io_uring_install_fd(struct file *file) |
9faadcc8 | 3777 | { |
6e76ac59 | 3778 | int fd; |
9faadcc8 PB |
3779 | |
3780 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3781 | if (fd < 0) | |
3782 | return fd; | |
9faadcc8 PB |
3783 | fd_install(fd, file); |
3784 | return fd; | |
3785 | } | |
3786 | ||
2b188cc1 JA |
3787 | /* |
3788 | * Allocate an anonymous fd, this is what constitutes the application | |
3789 | * visible backing of an io_uring instance. The application mmaps this | |
6e5e6d27 | 3790 | * fd to gain access to the SQ/CQ ring details. |
2b188cc1 | 3791 | */ |
9faadcc8 | 3792 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 | 3793 | { |
4f0b9194 | 3794 | /* Create a new inode so that the LSM can block the creation. */ |
09d1c6a8 | 3795 | return anon_inode_create_getfile("[io_uring]", &io_uring_fops, ctx, |
91a9ab7c | 3796 | O_RDWR | O_CLOEXEC, NULL); |
2b188cc1 JA |
3797 | } |
3798 | ||
c072481d PB |
3799 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3800 | struct io_uring_params __user *params) | |
2b188cc1 | 3801 | { |
2b188cc1 | 3802 | struct io_ring_ctx *ctx; |
6e76ac59 | 3803 | struct io_uring_task *tctx; |
9faadcc8 | 3804 | struct file *file; |
2b188cc1 JA |
3805 | int ret; |
3806 | ||
8110c1a6 | 3807 | if (!entries) |
2b188cc1 | 3808 | return -EINVAL; |
8110c1a6 JA |
3809 | if (entries > IORING_MAX_ENTRIES) { |
3810 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3811 | return -EINVAL; | |
3812 | entries = IORING_MAX_ENTRIES; | |
3813 | } | |
2b188cc1 | 3814 | |
6e76ac59 JT |
3815 | if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY) |
3816 | && !(p->flags & IORING_SETUP_NO_MMAP)) | |
3817 | return -EINVAL; | |
3818 | ||
2b188cc1 JA |
3819 | /* |
3820 | * Use twice as many entries for the CQ ring. It's possible for the | |
3821 | * application to drive a higher depth than the size of the SQ ring, | |
3822 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3823 | * some flexibility in overcommitting a bit. If the application has |
3824 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3825 | * of CQ ring entries manually. | |
2b188cc1 JA |
3826 | */ |
3827 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3828 | if (p->flags & IORING_SETUP_CQSIZE) { |
3829 | /* | |
3830 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3831 | * to a power-of-two, if it isn't already. We do NOT impose | |
3832 | * any cq vs sq ring sizing. | |
3833 | */ | |
eb2667b3 | 3834 | if (!p->cq_entries) |
33a107f0 | 3835 | return -EINVAL; |
8110c1a6 JA |
3836 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3837 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3838 | return -EINVAL; | |
3839 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3840 | } | |
eb2667b3 JQ |
3841 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3842 | if (p->cq_entries < p->sq_entries) | |
3843 | return -EINVAL; | |
33a107f0 JA |
3844 | } else { |
3845 | p->cq_entries = 2 * p->sq_entries; | |
3846 | } | |
2b188cc1 | 3847 | |
2b188cc1 | 3848 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3849 | if (!ctx) |
2b188cc1 | 3850 | return -ENOMEM; |
773697b6 | 3851 | |
e6aeb272 PB |
3852 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3853 | !(ctx->flags & IORING_SETUP_IOPOLL) && | |
3854 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3855 | ctx->task_complete = true; | |
3856 | ||
ec26c225 PB |
3857 | if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL)) |
3858 | ctx->lockless_cq = true; | |
3859 | ||
bca39f39 PB |
3860 | /* |
3861 | * lazy poll_wq activation relies on ->task_complete for synchronisation | |
3862 | * purposes, see io_activate_pollwq() | |
3863 | */ | |
3864 | if (!ctx->task_complete) | |
3865 | ctx->poll_activated = true; | |
3866 | ||
773697b6 PB |
3867 | /* |
3868 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3869 | * space applications don't need to do io completion events | |
3870 | * polling again, they can rely on io_sq_thread to do polling | |
3871 | * work, which can reduce cpu usage and uring_lock contention. | |
3872 | */ | |
3873 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3874 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3875 | ctx->syscall_iopoll = 1; | |
3876 | ||
2b188cc1 | 3877 | ctx->compat = in_compat_syscall(); |
6adc2272 | 3878 | if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK)) |
62e398be | 3879 | ctx->user = get_uid(current_user()); |
2aede0e4 | 3880 | |
9f010507 | 3881 | /* |
e1169f06 JA |
3882 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3883 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3884 | */ |
e1169f06 JA |
3885 | ret = -EINVAL; |
3886 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3887 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 | 3888 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
c0e0d6ba DY |
3889 | IORING_SETUP_TASKRUN_FLAG | |
3890 | IORING_SETUP_DEFER_TASKRUN)) | |
e1169f06 | 3891 | goto err; |
9f010507 | 3892 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3893 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3894 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3895 | } else { | |
c0e0d6ba DY |
3896 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && |
3897 | !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) | |
ef060ea9 | 3898 | goto err; |
9f010507 | 3899 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3900 | } |
9f010507 | 3901 | |
c0e0d6ba DY |
3902 | /* |
3903 | * For DEFER_TASKRUN we require the completion task to be the same as the | |
3904 | * submission task. This implies that there is only one submitter, so enforce | |
3905 | * that. | |
3906 | */ | |
3907 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && | |
3908 | !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { | |
3909 | goto err; | |
3910 | } | |
3911 | ||
2aede0e4 JA |
3912 | /* |
3913 | * This is just grabbed for accounting purposes. When a process exits, | |
3914 | * the mm is exited and dropped before the files, hence we need to hang | |
3915 | * on to this mm purely for the purposes of being able to unaccount | |
3916 | * memory (locked/pinned vm). It's not used for anything else. | |
3917 | */ | |
6b7898eb | 3918 | mmgrab(current->mm); |
2aede0e4 | 3919 | ctx->mm_account = current->mm; |
6b7898eb | 3920 | |
2b188cc1 JA |
3921 | ret = io_allocate_scq_urings(ctx, p); |
3922 | if (ret) | |
3923 | goto err; | |
3924 | ||
7e84e1c7 | 3925 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
3926 | if (ret) |
3927 | goto err; | |
2933ae6e PB |
3928 | |
3929 | ret = io_rsrc_init(ctx); | |
47b228ce PB |
3930 | if (ret) |
3931 | goto err; | |
2b188cc1 | 3932 | |
75b28aff HV |
3933 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3934 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3935 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3936 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3937 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3938 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
2af89abd PB |
3939 | if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) |
3940 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
9b1b58ca | 3941 | p->sq_off.resv1 = 0; |
03d89a2d JA |
3942 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3943 | p->sq_off.user_addr = 0; | |
2b188cc1 | 3944 | |
75b28aff HV |
3945 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3946 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3947 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3948 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3949 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3950 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 3951 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
9b1b58ca | 3952 | p->cq_off.resv1 = 0; |
03d89a2d JA |
3953 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3954 | p->cq_off.user_addr = 0; | |
ac90f249 | 3955 | |
7f13657d XW |
3956 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
3957 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 3958 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 3959 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 3960 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e | 3961 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
7d3fd88d | 3962 | IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING; |
7f13657d XW |
3963 | |
3964 | if (copy_to_user(params, p, sizeof(*p))) { | |
3965 | ret = -EFAULT; | |
3966 | goto err; | |
3967 | } | |
d1719f70 | 3968 | |
7cae596b DY |
3969 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER |
3970 | && !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
8579538c | 3971 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
7cae596b | 3972 | |
9faadcc8 PB |
3973 | file = io_uring_get_file(ctx); |
3974 | if (IS_ERR(file)) { | |
3975 | ret = PTR_ERR(file); | |
3976 | goto err; | |
3977 | } | |
3978 | ||
6e76ac59 JT |
3979 | ret = __io_uring_add_tctx_node(ctx); |
3980 | if (ret) | |
3981 | goto err_fput; | |
3982 | tctx = current->io_uring; | |
3983 | ||
044c1ab3 JA |
3984 | /* |
3985 | * Install ring fd as the very last thing, so we don't risk someone | |
3986 | * having closed it before we finish setup | |
3987 | */ | |
6e76ac59 JT |
3988 | if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY) |
3989 | ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX); | |
3990 | else | |
3991 | ret = io_uring_install_fd(file); | |
3992 | if (ret < 0) | |
3993 | goto err_fput; | |
044c1ab3 | 3994 | |
c826bd7a | 3995 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
3996 | return ret; |
3997 | err: | |
3998 | io_ring_ctx_wait_and_kill(ctx); | |
3999 | return ret; | |
6e76ac59 JT |
4000 | err_fput: |
4001 | fput(file); | |
4002 | return ret; | |
2b188cc1 JA |
4003 | } |
4004 | ||
4005 | /* | |
4006 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
4007 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
4008 | * params structure passed in. | |
4009 | */ | |
4010 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
4011 | { | |
4012 | struct io_uring_params p; | |
2b188cc1 JA |
4013 | int i; |
4014 | ||
4015 | if (copy_from_user(&p, params, sizeof(p))) | |
4016 | return -EFAULT; | |
4017 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
4018 | if (p.resv[i]) | |
4019 | return -EINVAL; | |
4020 | } | |
4021 | ||
6c271ce2 | 4022 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 4023 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 4024 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 4025 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 4026 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 | 4027 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
03d89a2d | 4028 | IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN | |
2af89abd PB |
4029 | IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY | |
4030 | IORING_SETUP_NO_SQARRAY)) | |
2b188cc1 JA |
4031 | return -EINVAL; |
4032 | ||
ef060ea9 | 4033 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
4034 | } |
4035 | ||
76d3ccec MR |
4036 | static inline bool io_uring_allowed(void) |
4037 | { | |
4038 | int disabled = READ_ONCE(sysctl_io_uring_disabled); | |
4039 | kgid_t io_uring_group; | |
4040 | ||
4041 | if (disabled == 2) | |
4042 | return false; | |
4043 | ||
4044 | if (disabled == 0 || capable(CAP_SYS_ADMIN)) | |
4045 | return true; | |
4046 | ||
4047 | io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group); | |
4048 | if (!gid_valid(io_uring_group)) | |
4049 | return false; | |
4050 | ||
4051 | return in_group_p(io_uring_group); | |
4052 | } | |
4053 | ||
2b188cc1 JA |
4054 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, |
4055 | struct io_uring_params __user *, params) | |
4056 | { | |
76d3ccec MR |
4057 | if (!io_uring_allowed()) |
4058 | return -EPERM; | |
4059 | ||
2b188cc1 JA |
4060 | return io_uring_setup(entries, params); |
4061 | } | |
4062 | ||
4063 | static int __init io_uring_init(void) | |
4064 | { | |
9c71d39a | 4065 | #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ |
d7f62e82 | 4066 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ |
9c71d39a | 4067 | BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ |
d7f62e82 SM |
4068 | } while (0) |
4069 | ||
4070 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9c71d39a SM |
4071 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) |
4072 | #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ | |
4073 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) | |
d7f62e82 SM |
4074 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); |
4075 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
4076 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
4077 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
4078 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
4079 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
4080 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9c71d39a SM |
4081 | BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); |
4082 | BUILD_BUG_SQE_ELEM(12, __u32, __pad1); | |
d7f62e82 | 4083 | BUILD_BUG_SQE_ELEM(16, __u64, addr); |
7d67af2c | 4084 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
4085 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
4086 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
4087 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
4088 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
4089 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
4090 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
4091 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
4092 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
4093 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
4094 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
4095 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
4096 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
4097 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
4098 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
4099 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 4100 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
9c71d39a SM |
4101 | BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); |
4102 | BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); | |
4103 | BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); | |
4104 | BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); | |
4105 | BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); | |
d7f62e82 SM |
4106 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
4107 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 4108 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 4109 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 4110 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 4111 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
b48c312b PB |
4112 | BUILD_BUG_SQE_ELEM(44, __u16, addr_len); |
4113 | BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); | |
e9621e2b | 4114 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
9c71d39a SM |
4115 | BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); |
4116 | BUILD_BUG_SQE_ELEM(56, __u64, __pad2); | |
d7f62e82 | 4117 | |
b0d658ec PB |
4118 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
4119 | sizeof(struct io_uring_rsrc_update)); | |
4120 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
4121 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
4122 | |
4123 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
4124 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
4125 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
4126 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 4127 | |
b0d658ec PB |
4128 | /* should fit into one byte */ |
4129 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
4130 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
4131 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 4132 | |
4bcb982c | 4133 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof_field(struct io_kiocb, flags)); |
16340eab | 4134 | |
3a4b89a2 JA |
4135 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
4136 | ||
528ce678 ML |
4137 | /* top 8bits are for internal use */ |
4138 | BUILD_BUG_ON((IORING_URING_CMD_MASK & 0xff000000) != 0); | |
4139 | ||
d9b57aa3 | 4140 | io_uring_optable_init(); |
0702e536 | 4141 | |
b97f96e2 JA |
4142 | /* |
4143 | * Allow user copy in the per-command field, which starts after the | |
4144 | * file in io_kiocb and until the opcode field. The openat2 handling | |
4145 | * requires copying in user memory into the io_kiocb object in that | |
4146 | * range, and HARDENED_USERCOPY will complain if we haven't | |
4147 | * correctly annotated this range. | |
4148 | */ | |
4149 | req_cachep = kmem_cache_create_usercopy("io_kiocb", | |
4150 | sizeof(struct io_kiocb), 0, | |
4151 | SLAB_HWCACHE_ALIGN | SLAB_PANIC | | |
4152 | SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU, | |
4153 | offsetof(struct io_kiocb, cmd.data), | |
4154 | sizeof_field(struct io_kiocb, cmd.data), NULL); | |
b3a4dbc8 GKB |
4155 | io_buf_cachep = kmem_cache_create("io_buffer", sizeof(struct io_buffer), 0, |
4156 | SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT, | |
4157 | NULL); | |
b97f96e2 | 4158 | |
76d3ccec MR |
4159 | #ifdef CONFIG_SYSCTL |
4160 | register_sysctl_init("kernel", kernel_io_uring_disabled_table); | |
4161 | #endif | |
4162 | ||
2b188cc1 JA |
4163 | return 0; |
4164 | }; | |
4165 | __initcall(io_uring_init); |