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> | |
2b188cc1 JA |
62 | #include <linux/anon_inodes.h> |
63 | #include <linux/sched/mm.h> | |
64 | #include <linux/uaccess.h> | |
65 | #include <linux/nospec.h> | |
aa4c3967 | 66 | #include <linux/highmem.h> |
15b71abe | 67 | #include <linux/fsnotify.h> |
4840e418 | 68 | #include <linux/fadvise.h> |
b41e9852 | 69 | #include <linux/task_work.h> |
0f212204 | 70 | #include <linux/io_uring.h> |
b66509b8 | 71 | #include <linux/io_uring/cmd.h> |
5bd2182d | 72 | #include <linux/audit.h> |
cdc1404a | 73 | #include <linux/security.h> |
d808459b | 74 | #include <asm/shmparam.h> |
2b188cc1 | 75 | |
c826bd7a DD |
76 | #define CREATE_TRACE_POINTS |
77 | #include <trace/events/io_uring.h> | |
78 | ||
2b188cc1 JA |
79 | #include <uapi/linux/io_uring.h> |
80 | ||
561fb04a | 81 | #include "io-wq.h" |
2b188cc1 | 82 | |
de23077e | 83 | #include "io_uring.h" |
329061d3 | 84 | #include "opdef.h" |
e418bbc9 | 85 | #include "refs.h" |
c9f06aa7 | 86 | #include "tctx.h" |
c4320315 | 87 | #include "register.h" |
17437f31 | 88 | #include "sqpoll.h" |
a4ad4f74 | 89 | #include "fdinfo.h" |
3b77495a | 90 | #include "kbuf.h" |
73572984 | 91 | #include "rsrc.h" |
38513c46 | 92 | #include "cancel.h" |
43e0bbbd | 93 | #include "net.h" |
eb42cebb | 94 | #include "notif.h" |
f31ecf67 | 95 | #include "waitid.h" |
194bb58c | 96 | #include "futex.h" |
8d0c12a8 | 97 | #include "napi.h" |
da12d9ab | 98 | #include "uring_cmd.h" |
e27f928e | 99 | |
59915143 | 100 | #include "timeout.h" |
329061d3 | 101 | #include "poll.h" |
c92fcfc2 | 102 | #include "rw.h" |
9b797a37 | 103 | #include "alloc_cache.h" |
5e2a18d9 | 104 | |
5277deaa | 105 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 106 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 | 107 | |
68fe256a PB |
108 | #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \ |
109 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) | |
110 | ||
5562a8d7 PB |
111 | #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \ |
112 | IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS) | |
68fe256a | 113 | |
c854357b | 114 | #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ |
9cae36a0 JA |
115 | REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \ |
116 | REQ_F_ASYNC_DATA) | |
b16fed66 | 117 | |
a538be5b PB |
118 | #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\ |
119 | IO_REQ_CLEAN_FLAGS) | |
120 | ||
09899b19 PB |
121 | #define IO_TCTX_REFS_CACHE_NR (1U << 10) |
122 | ||
6dd0be1e | 123 | #define IO_COMPL_BATCH 32 |
bf019da7 | 124 | #define IO_REQ_ALLOC_BATCH 8 |
258b29a9 | 125 | |
27dc8338 PB |
126 | struct io_defer_entry { |
127 | struct list_head list; | |
128 | struct io_kiocb *req; | |
9cf7c104 | 129 | u32 seq; |
2b188cc1 JA |
130 | }; |
131 | ||
0756a869 PB |
132 | /* requests with any of those set should undergo io_disarm_next() */ |
133 | #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) | |
da1a08c5 | 134 | #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK) |
0756a869 | 135 | |
b4bc35cf PB |
136 | /* |
137 | * No waiters. It's larger than any valid value of the tw counter | |
138 | * so that tests against ->cq_wait_nr would fail and skip wake_up(). | |
139 | */ | |
140 | #define IO_CQ_WAKE_INIT (-1U) | |
141 | /* Forced wake up if there is a waiter regardless of ->cq_wait_nr */ | |
142 | #define IO_CQ_WAKE_FORCE (IO_CQ_WAKE_INIT >> 1) | |
143 | ||
affa87db | 144 | static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
9936c7c2 | 145 | struct task_struct *task, |
3dd0c97a | 146 | bool cancel_all); |
1ffc5422 | 147 | |
cbc2e203 | 148 | static void io_queue_sqe(struct io_kiocb *req); |
de0617e4 | 149 | |
c1755c25 | 150 | struct kmem_cache *req_cachep; |
73eaa2b5 | 151 | static struct workqueue_struct *iou_wq __ro_after_init; |
2b188cc1 | 152 | |
76d3ccec MR |
153 | static int __read_mostly sysctl_io_uring_disabled; |
154 | static int __read_mostly sysctl_io_uring_group = -1; | |
155 | ||
156 | #ifdef CONFIG_SYSCTL | |
157 | static struct ctl_table kernel_io_uring_disabled_table[] = { | |
158 | { | |
159 | .procname = "io_uring_disabled", | |
160 | .data = &sysctl_io_uring_disabled, | |
161 | .maxlen = sizeof(sysctl_io_uring_disabled), | |
162 | .mode = 0644, | |
163 | .proc_handler = proc_dointvec_minmax, | |
164 | .extra1 = SYSCTL_ZERO, | |
165 | .extra2 = SYSCTL_TWO, | |
166 | }, | |
167 | { | |
168 | .procname = "io_uring_group", | |
169 | .data = &sysctl_io_uring_group, | |
170 | .maxlen = sizeof(gid_t), | |
171 | .mode = 0644, | |
172 | .proc_handler = proc_dointvec, | |
173 | }, | |
174 | {}, | |
175 | }; | |
176 | #endif | |
177 | ||
faf88dde PB |
178 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
179 | { | |
180 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
181 | } | |
182 | ||
0fc8c2ac DY |
183 | static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx) |
184 | { | |
185 | return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head); | |
186 | } | |
187 | ||
9cae36a0 JA |
188 | static bool io_match_linked(struct io_kiocb *head) |
189 | { | |
190 | struct io_kiocb *req; | |
191 | ||
192 | io_for_each_link(req, head) { | |
193 | if (req->flags & REQ_F_INFLIGHT) | |
194 | return true; | |
195 | } | |
196 | return false; | |
6af3f48b PB |
197 | } |
198 | ||
199 | /* | |
200 | * As io_match_task() but protected against racing with linked timeouts. | |
201 | * User must not hold timeout_lock. | |
202 | */ | |
329061d3 JA |
203 | bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, |
204 | bool cancel_all) | |
6af3f48b | 205 | { |
9cae36a0 JA |
206 | bool matched; |
207 | ||
6af3f48b PB |
208 | if (task && head->task != task) |
209 | return false; | |
9cae36a0 JA |
210 | if (cancel_all) |
211 | return true; | |
212 | ||
213 | if (head->flags & REQ_F_LINK_TIMEOUT) { | |
214 | struct io_ring_ctx *ctx = head->ctx; | |
215 | ||
216 | /* protect against races with linked timeouts */ | |
217 | spin_lock_irq(&ctx->timeout_lock); | |
218 | matched = io_match_linked(head); | |
219 | spin_unlock_irq(&ctx->timeout_lock); | |
220 | } else { | |
221 | matched = io_match_linked(head); | |
222 | } | |
223 | return matched; | |
6af3f48b PB |
224 | } |
225 | ||
a8295b98 HX |
226 | static inline void req_fail_link_node(struct io_kiocb *req, int res) |
227 | { | |
228 | req_set_fail(req); | |
97b388d7 | 229 | io_req_set_res(req, res, 0); |
a8295b98 HX |
230 | } |
231 | ||
fa05457a PB |
232 | static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) |
233 | { | |
234 | wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); | |
a8295b98 HX |
235 | } |
236 | ||
c072481d | 237 | static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) |
2b188cc1 JA |
238 | { |
239 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
240 | ||
0f158b4c | 241 | complete(&ctx->ref_comp); |
2b188cc1 JA |
242 | } |
243 | ||
c072481d | 244 | static __cold void io_fallback_req_func(struct work_struct *work) |
f56165e6 PB |
245 | { |
246 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
247 | fallback_work.work); | |
248 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
249 | struct io_kiocb *req, *tmp; | |
a282967c | 250 | struct io_tw_state ts = { .locked = true, }; |
f56165e6 | 251 | |
f7b32e78 | 252 | percpu_ref_get(&ctx->refs); |
31f084b7 | 253 | mutex_lock(&ctx->uring_lock); |
3218e5d3 | 254 | llist_for_each_entry_safe(req, tmp, node, io_task_work.node) |
a282967c PB |
255 | req->io_task_work.func(req, &ts); |
256 | if (WARN_ON_ONCE(!ts.locked)) | |
31f084b7 PB |
257 | return; |
258 | io_submit_flush_completions(ctx); | |
259 | mutex_unlock(&ctx->uring_lock); | |
f7b32e78 | 260 | percpu_ref_put(&ctx->refs); |
f56165e6 PB |
261 | } |
262 | ||
e6f89be6 PB |
263 | static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) |
264 | { | |
265 | unsigned hash_buckets = 1U << bits; | |
266 | size_t hash_size = hash_buckets * sizeof(table->hbs[0]); | |
267 | ||
268 | table->hbs = kmalloc(hash_size, GFP_KERNEL); | |
269 | if (!table->hbs) | |
270 | return -ENOMEM; | |
271 | ||
272 | table->hash_bits = bits; | |
273 | init_hash_table(table, hash_buckets); | |
274 | return 0; | |
275 | } | |
276 | ||
c072481d | 277 | static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
2b188cc1 JA |
278 | { |
279 | struct io_ring_ctx *ctx; | |
9cfc7e94 | 280 | int hash_bits; |
2b188cc1 JA |
281 | |
282 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
283 | if (!ctx) | |
284 | return NULL; | |
285 | ||
9cfc7e94 JA |
286 | xa_init(&ctx->io_bl_xa); |
287 | ||
78076bb6 JA |
288 | /* |
289 | * Use 5 bits less than the max cq entries, that should give us around | |
4a07723f PB |
290 | * 32 entries per hash list if totally full and uniformly spread, but |
291 | * don't keep too many buckets to not overconsume memory. | |
78076bb6 | 292 | */ |
4a07723f PB |
293 | hash_bits = ilog2(p->cq_entries) - 5; |
294 | hash_bits = clamp(hash_bits, 1, 8); | |
e6f89be6 | 295 | if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) |
78076bb6 | 296 | goto err; |
9ca9fb24 PB |
297 | if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) |
298 | goto err; | |
21482896 | 299 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
48904229 | 300 | 0, GFP_KERNEL)) |
206aefde | 301 | goto err; |
2b188cc1 JA |
302 | |
303 | ctx->flags = p->flags; | |
b4bc35cf | 304 | atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT); |
90554200 | 305 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 306 | INIT_LIST_HEAD(&ctx->sqd_list); |
1d7bb1d5 | 307 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
cc3cec83 | 308 | INIT_LIST_HEAD(&ctx->io_buffers_cache); |
c392cbec | 309 | INIT_HLIST_HEAD(&ctx->io_buf_list); |
69bbc6ad PB |
310 | io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX, |
311 | sizeof(struct io_rsrc_node)); | |
312 | io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX, | |
313 | sizeof(struct async_poll)); | |
314 | io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX, | |
315 | sizeof(struct io_async_msghdr)); | |
194bb58c | 316 | io_futex_cache_init(ctx); |
0f158b4c | 317 | init_completion(&ctx->ref_comp); |
61cf9370 | 318 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 319 | mutex_init(&ctx->uring_lock); |
311997b3 | 320 | init_waitqueue_head(&ctx->cq_wait); |
7b235dd8 | 321 | init_waitqueue_head(&ctx->poll_wq); |
4ea15b56 | 322 | init_waitqueue_head(&ctx->rsrc_quiesce_wq); |
2b188cc1 | 323 | spin_lock_init(&ctx->completion_lock); |
89850fce | 324 | spin_lock_init(&ctx->timeout_lock); |
5eef4e87 | 325 | INIT_WQ_LIST(&ctx->iopoll_list); |
cc3cec83 | 326 | INIT_LIST_HEAD(&ctx->io_buffers_comp); |
de0617e4 | 327 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 328 | INIT_LIST_HEAD(&ctx->timeout_list); |
ef9dd637 | 329 | INIT_LIST_HEAD(&ctx->ltimeout_list); |
d67d2263 | 330 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); |
c0e0d6ba | 331 | init_llist_head(&ctx->work_llist); |
13bf43f5 | 332 | INIT_LIST_HEAD(&ctx->tctx_list); |
c2b6c6bc PB |
333 | ctx->submit_state.free_list.next = NULL; |
334 | INIT_WQ_LIST(&ctx->locked_free_list); | |
f31ecf67 | 335 | INIT_HLIST_HEAD(&ctx->waitid_list); |
194bb58c JA |
336 | #ifdef CONFIG_FUTEX |
337 | INIT_HLIST_HEAD(&ctx->futex_list); | |
338 | #endif | |
9011bf9a | 339 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
6f33b0bc | 340 | INIT_WQ_LIST(&ctx->submit_state.compl_reqs); |
93b8cc60 | 341 | INIT_HLIST_HEAD(&ctx->cancelable_uring_cmd); |
8d0c12a8 SR |
342 | io_napi_init(ctx); |
343 | ||
2b188cc1 | 344 | return ctx; |
206aefde | 345 | err: |
e6f89be6 | 346 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 347 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 | 348 | xa_destroy(&ctx->io_bl_xa); |
206aefde JA |
349 | kfree(ctx); |
350 | return NULL; | |
2b188cc1 JA |
351 | } |
352 | ||
8f6ed49a PB |
353 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
354 | { | |
355 | struct io_rings *r = ctx->rings; | |
356 | ||
357 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
358 | ctx->cq_extra--; | |
359 | } | |
360 | ||
9cf7c104 | 361 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 362 | { |
2bc9930e JA |
363 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
364 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 365 | |
8f6ed49a | 366 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 367 | } |
de0617e4 | 368 | |
9d858b21 | 369 | return false; |
de0617e4 JA |
370 | } |
371 | ||
5a754dea PB |
372 | static void io_clean_op(struct io_kiocb *req) |
373 | { | |
374 | if (req->flags & REQ_F_BUFFER_SELECTED) { | |
375 | spin_lock(&req->ctx->completion_lock); | |
376 | io_put_kbuf_comp(req); | |
377 | spin_unlock(&req->ctx->completion_lock); | |
378 | } | |
379 | ||
380 | if (req->flags & REQ_F_NEED_CLEANUP) { | |
381 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; | |
382 | ||
383 | if (def->cleanup) | |
384 | def->cleanup(req); | |
385 | } | |
386 | if ((req->flags & REQ_F_POLLED) && req->apoll) { | |
387 | kfree(req->apoll->double_poll); | |
388 | kfree(req->apoll); | |
389 | req->apoll = NULL; | |
390 | } | |
391 | if (req->flags & REQ_F_INFLIGHT) { | |
392 | struct io_uring_task *tctx = req->task->io_uring; | |
393 | ||
394 | atomic_dec(&tctx->inflight_tracked); | |
395 | } | |
396 | if (req->flags & REQ_F_CREDS) | |
397 | put_cred(req->creds); | |
398 | if (req->flags & REQ_F_ASYNC_DATA) { | |
399 | kfree(req->async_data); | |
400 | req->async_data = NULL; | |
401 | } | |
402 | req->flags &= ~IO_REQ_CLEAN_FLAGS; | |
403 | } | |
404 | ||
9cae36a0 JA |
405 | static inline void io_req_track_inflight(struct io_kiocb *req) |
406 | { | |
407 | if (!(req->flags & REQ_F_INFLIGHT)) { | |
408 | req->flags |= REQ_F_INFLIGHT; | |
386e4fb6 | 409 | atomic_inc(&req->task->io_uring->inflight_tracked); |
9cae36a0 JA |
410 | } |
411 | } | |
412 | ||
fd08e530 PB |
413 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
414 | { | |
906c6caa PB |
415 | if (WARN_ON_ONCE(!req->link)) |
416 | return NULL; | |
417 | ||
4d13d1a4 PB |
418 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
419 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
420 | |
421 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 422 | io_req_set_refcount(req); |
4d13d1a4 PB |
423 | __io_req_set_refcount(req->link, 2); |
424 | return req->link; | |
fd08e530 PB |
425 | } |
426 | ||
427 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
428 | { | |
4d13d1a4 | 429 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
430 | return NULL; |
431 | return __io_prep_linked_timeout(req); | |
432 | } | |
433 | ||
cb2d344c PB |
434 | static noinline void __io_arm_ltimeout(struct io_kiocb *req) |
435 | { | |
436 | io_queue_linked_timeout(__io_prep_linked_timeout(req)); | |
437 | } | |
438 | ||
439 | static inline void io_arm_ltimeout(struct io_kiocb *req) | |
440 | { | |
441 | if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) | |
442 | __io_arm_ltimeout(req); | |
443 | } | |
444 | ||
1e6fa521 JA |
445 | static void io_prep_async_work(struct io_kiocb *req) |
446 | { | |
a7dd2782 | 447 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
1e6fa521 JA |
448 | struct io_ring_ctx *ctx = req->ctx; |
449 | ||
b8e64b53 PB |
450 | if (!(req->flags & REQ_F_CREDS)) { |
451 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 452 | req->creds = get_current_cred(); |
b8e64b53 | 453 | } |
003e8dcc | 454 | |
e1d675df PB |
455 | req->work.list.next = NULL; |
456 | req->work.flags = 0; | |
feaadc4f PB |
457 | if (req->flags & REQ_F_FORCE_ASYNC) |
458 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
459 | ||
3beed235 | 460 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
8487f083 | 461 | req->flags |= io_file_get_flags(req->file); |
f6b543fd | 462 | |
8b1df11f | 463 | if (req->file && (req->flags & REQ_F_ISREG)) { |
d4755e15 JA |
464 | bool should_hash = def->hash_reg_file; |
465 | ||
466 | /* don't serialize this request if the fs doesn't need it */ | |
467 | if (should_hash && (req->file->f_flags & O_DIRECT) && | |
468 | (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE)) | |
469 | should_hash = false; | |
470 | if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL)) | |
1e6fa521 | 471 | io_wq_hash_work(&req->work, file_inode(req->file)); |
4b982bd0 | 472 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
473 | if (def->unbound_nonreg_file) |
474 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
475 | } | |
561fb04a | 476 | } |
cccf0ee8 | 477 | |
cbdcb435 | 478 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 479 | { |
cbdcb435 | 480 | struct io_kiocb *cur; |
54a91f3b | 481 | |
44eff40a PB |
482 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
483 | struct io_ring_ctx *ctx = req->ctx; | |
484 | ||
674ee8e1 | 485 | spin_lock_irq(&ctx->timeout_lock); |
44eff40a PB |
486 | io_for_each_link(cur, req) |
487 | io_prep_async_work(cur); | |
674ee8e1 | 488 | spin_unlock_irq(&ctx->timeout_lock); |
44eff40a PB |
489 | } else { |
490 | io_for_each_link(cur, req) | |
491 | io_prep_async_work(cur); | |
492 | } | |
561fb04a JA |
493 | } |
494 | ||
a282967c | 495 | void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use) |
561fb04a | 496 | { |
cbdcb435 | 497 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 498 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 499 | |
3bfe6106 JA |
500 | BUG_ON(!tctx); |
501 | BUG_ON(!tctx->io_wq); | |
561fb04a | 502 | |
cbdcb435 PB |
503 | /* init ->work of the whole link before punting */ |
504 | io_prep_async_link(req); | |
991468dc JA |
505 | |
506 | /* | |
507 | * Not expected to happen, but if we do have a bug where this _can_ | |
508 | * happen, catch it here and ensure the request is marked as | |
509 | * canceled. That will make io-wq go through the usual work cancel | |
510 | * procedure rather than attempt to run this request (or create a new | |
511 | * worker for it). | |
512 | */ | |
513 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
514 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
515 | ||
48863ffd | 516 | trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); |
ebf93667 | 517 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
518 | if (link) |
519 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
520 | } |
521 | ||
c072481d | 522 | static __cold void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 523 | { |
441b8a78 | 524 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
525 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
526 | struct io_defer_entry, list); | |
de0617e4 | 527 | |
9cf7c104 | 528 | if (req_need_defer(de->req, de->seq)) |
04518945 | 529 | break; |
27dc8338 | 530 | list_del_init(&de->list); |
907d1df3 | 531 | io_req_task_queue(de->req); |
27dc8338 | 532 | kfree(de); |
441b8a78 | 533 | } |
04518945 PB |
534 | } |
535 | ||
c4320315 | 536 | void io_eventfd_ops(struct rcu_head *rcu) |
f2842ab5 | 537 | { |
d8e9214f | 538 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
21a091b9 | 539 | int ops = atomic_xchg(&ev_fd->ops, 0); |
305bef98 | 540 | |
21a091b9 | 541 | if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) |
120ae585 | 542 | eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE); |
d8e9214f | 543 | |
21a091b9 DY |
544 | /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback |
545 | * ordering in a race but if references are 0 we know we have to free | |
546 | * it regardless. | |
305bef98 | 547 | */ |
21a091b9 DY |
548 | if (atomic_dec_and_test(&ev_fd->refs)) { |
549 | eventfd_ctx_put(ev_fd->cq_ev_fd); | |
550 | kfree(ev_fd); | |
551 | } | |
d8e9214f DY |
552 | } |
553 | ||
77bc59b4 | 554 | static void io_eventfd_signal(struct io_ring_ctx *ctx) |
f2842ab5 | 555 | { |
21a091b9 | 556 | struct io_ev_fd *ev_fd = NULL; |
77bc59b4 | 557 | |
77bc59b4 UA |
558 | rcu_read_lock(); |
559 | /* | |
560 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
561 | * and eventfd_signal | |
562 | */ | |
563 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
564 | ||
565 | /* | |
566 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
567 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
568 | * the function and rcu_read_lock. | |
569 | */ | |
570 | if (unlikely(!ev_fd)) | |
571 | goto out; | |
7e55a19c | 572 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 | 573 | goto out; |
21a091b9 DY |
574 | if (ev_fd->eventfd_async && !io_wq_current_is_worker()) |
575 | goto out; | |
77bc59b4 | 576 | |
21a091b9 | 577 | if (likely(eventfd_signal_allowed())) { |
120ae585 | 578 | eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE); |
21a091b9 DY |
579 | } else { |
580 | atomic_inc(&ev_fd->refs); | |
581 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) | |
44a84da4 | 582 | call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops); |
21a091b9 DY |
583 | else |
584 | atomic_dec(&ev_fd->refs); | |
585 | } | |
586 | ||
77bc59b4 UA |
587 | out: |
588 | rcu_read_unlock(); | |
f2842ab5 JA |
589 | } |
590 | ||
21a091b9 DY |
591 | static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) |
592 | { | |
593 | bool skip; | |
594 | ||
595 | spin_lock(&ctx->completion_lock); | |
596 | ||
597 | /* | |
598 | * Eventfd should only get triggered when at least one event has been | |
599 | * posted. Some applications rely on the eventfd notification count | |
600 | * only changing IFF a new CQE has been added to the CQ ring. There's | |
601 | * no depedency on 1:1 relationship between how many times this | |
602 | * function is called (and hence the eventfd count) and number of CQEs | |
603 | * posted to the CQ ring. | |
604 | */ | |
605 | skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
606 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
607 | spin_unlock(&ctx->completion_lock); | |
608 | if (skip) | |
609 | return; | |
610 | ||
611 | io_eventfd_signal(ctx); | |
612 | } | |
613 | ||
a830ffd2 PB |
614 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
615 | { | |
bca39f39 PB |
616 | if (ctx->poll_activated) |
617 | io_poll_wq_wake(ctx); | |
e5f30f6f PB |
618 | if (ctx->off_timeout_used) |
619 | io_flush_timeouts(ctx); | |
620 | if (ctx->drain_active) { | |
a830ffd2 | 621 | spin_lock(&ctx->completion_lock); |
e5f30f6f | 622 | io_queue_deferred(ctx); |
a830ffd2 PB |
623 | spin_unlock(&ctx->completion_lock); |
624 | } | |
625 | if (ctx->has_evfd) | |
21a091b9 | 626 | io_eventfd_flush_signal(ctx); |
a830ffd2 PB |
627 | } |
628 | ||
f66f7342 | 629 | static inline void __io_cq_lock(struct io_ring_ctx *ctx) |
f66f7342 | 630 | { |
ec26c225 | 631 | if (!ctx->lockless_cq) |
f66f7342 PB |
632 | spin_lock(&ctx->completion_lock); |
633 | } | |
634 | ||
6971253f PB |
635 | static inline void io_cq_lock(struct io_ring_ctx *ctx) |
636 | __acquires(ctx->completion_lock) | |
637 | { | |
638 | spin_lock(&ctx->completion_lock); | |
639 | } | |
640 | ||
f66f7342 | 641 | static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) |
3181e22f PB |
642 | { |
643 | io_commit_cqring(ctx); | |
54927baf | 644 | if (!ctx->task_complete) { |
ec26c225 PB |
645 | if (!ctx->lockless_cq) |
646 | spin_unlock(&ctx->completion_lock); | |
647 | /* IOPOLL rings only need to wake up if it's also SQPOLL */ | |
648 | if (!ctx->syscall_iopoll) | |
649 | io_cqring_wake(ctx); | |
3181e22f | 650 | } |
54927baf | 651 | io_commit_cqring_flush(ctx); |
3181e22f PB |
652 | } |
653 | ||
0fdb9a19 | 654 | static void io_cq_unlock_post(struct io_ring_ctx *ctx) |
5d772916 | 655 | __releases(ctx->completion_lock) |
25399321 | 656 | { |
f66f7342 PB |
657 | io_commit_cqring(ctx); |
658 | spin_unlock(&ctx->completion_lock); | |
f66f7342 | 659 | io_cqring_wake(ctx); |
54927baf | 660 | io_commit_cqring_flush(ctx); |
25399321 PB |
661 | } |
662 | ||
a85381d8 PB |
663 | static void io_cqring_overflow_kill(struct io_ring_ctx *ctx) |
664 | { | |
665 | struct io_overflow_cqe *ocqe; | |
666 | LIST_HEAD(list); | |
667 | ||
f432b76b | 668 | spin_lock(&ctx->completion_lock); |
a85381d8 PB |
669 | list_splice_init(&ctx->cq_overflow_list, &list); |
670 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); | |
f432b76b | 671 | spin_unlock(&ctx->completion_lock); |
a85381d8 PB |
672 | |
673 | while (!list_empty(&list)) { | |
674 | ocqe = list_first_entry(&list, struct io_overflow_cqe, list); | |
675 | list_del(&ocqe->list); | |
676 | kfree(ocqe); | |
677 | } | |
678 | } | |
679 | ||
1b346e4a | 680 | static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
1d7bb1d5 | 681 | { |
e45a3e05 | 682 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 683 | |
a85381d8 | 684 | if (__io_cqring_events(ctx) == ctx->cq_entries) |
1b346e4a | 685 | return; |
1d7bb1d5 | 686 | |
e45a3e05 SR |
687 | if (ctx->flags & IORING_SETUP_CQE32) |
688 | cqe_size <<= 1; | |
689 | ||
25399321 | 690 | io_cq_lock(ctx); |
6c2450ae | 691 | while (!list_empty(&ctx->cq_overflow_list)) { |
59fbc409 | 692 | struct io_uring_cqe *cqe; |
6c2450ae | 693 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 694 | |
59fbc409 | 695 | if (!io_get_cqe_overflow(ctx, &cqe, true)) |
1d7bb1d5 | 696 | break; |
6c2450ae PB |
697 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
698 | struct io_overflow_cqe, list); | |
a85381d8 | 699 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae PB |
700 | list_del(&ocqe->list); |
701 | kfree(ocqe); | |
1d7bb1d5 JA |
702 | } |
703 | ||
1b346e4a | 704 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 705 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 706 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 707 | } |
25399321 | 708 | io_cq_unlock_post(ctx); |
1d7bb1d5 JA |
709 | } |
710 | ||
52ea806a JA |
711 | static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx) |
712 | { | |
713 | /* iopoll syncs against uring_lock, not completion_lock */ | |
714 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
715 | mutex_lock(&ctx->uring_lock); | |
716 | __io_cqring_overflow_flush(ctx); | |
717 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
718 | mutex_unlock(&ctx->uring_lock); | |
719 | } | |
720 | ||
1b346e4a | 721 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 722 | { |
52ea806a JA |
723 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) |
724 | io_cqring_do_overflow_flush(ctx); | |
6c503150 PB |
725 | } |
726 | ||
5afa4650 | 727 | /* can be called by any task */ |
2fdd6fb5 | 728 | static void io_put_task_remote(struct task_struct *task) |
6a290a14 PB |
729 | { |
730 | struct io_uring_task *tctx = task->io_uring; | |
731 | ||
2fdd6fb5 | 732 | percpu_counter_sub(&tctx->inflight, 1); |
8d664282 | 733 | if (unlikely(atomic_read(&tctx->in_cancel))) |
9d170164 | 734 | wake_up(&tctx->wait); |
2fdd6fb5 | 735 | put_task_struct(task); |
9d170164 PB |
736 | } |
737 | ||
5afa4650 | 738 | /* used by a task to put its own references */ |
2fdd6fb5 | 739 | static void io_put_task_local(struct task_struct *task) |
5afa4650 | 740 | { |
2fdd6fb5 | 741 | task->io_uring->cached_refs++; |
5afa4650 PB |
742 | } |
743 | ||
89800a2d | 744 | /* must to be called somewhat shortly after putting a request */ |
2fdd6fb5 | 745 | static inline void io_put_task(struct task_struct *task) |
89800a2d PB |
746 | { |
747 | if (likely(task == current)) | |
2fdd6fb5 | 748 | io_put_task_local(task); |
89800a2d | 749 | else |
2fdd6fb5 | 750 | io_put_task_remote(task); |
89800a2d PB |
751 | } |
752 | ||
63809137 | 753 | void io_task_refs_refill(struct io_uring_task *tctx) |
9a10867a PB |
754 | { |
755 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
756 | ||
757 | percpu_counter_add(&tctx->inflight, refill); | |
758 | refcount_add(refill, ¤t->usage); | |
759 | tctx->cached_refs += refill; | |
760 | } | |
761 | ||
3cc7fdb9 PB |
762 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
763 | { | |
764 | struct io_uring_task *tctx = task->io_uring; | |
765 | unsigned int refs = tctx->cached_refs; | |
766 | ||
767 | if (refs) { | |
768 | tctx->cached_refs = 0; | |
769 | percpu_counter_sub(&tctx->inflight, refs); | |
770 | put_task_struct_many(task, refs); | |
771 | } | |
772 | } | |
773 | ||
68494a65 PB |
774 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
775 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 776 | { |
cce4b8b0 | 777 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
778 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
779 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 780 | |
f26cc959 PB |
781 | lockdep_assert_held(&ctx->completion_lock); |
782 | ||
e45a3e05 SR |
783 | if (is_cqe32) |
784 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 785 | |
e45a3e05 | 786 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 787 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
788 | if (!ocqe) { |
789 | /* | |
790 | * If we're in ring overflow flush mode, or in task cancel mode, | |
791 | * or cannot allocate an overflow entry, then we need to drop it | |
792 | * on the floor. | |
793 | */ | |
8f6ed49a | 794 | io_account_cq_overflow(ctx); |
155bc950 | 795 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 796 | return false; |
2b188cc1 | 797 | } |
cce4b8b0 | 798 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 799 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 800 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 801 | |
cce4b8b0 | 802 | } |
d4d19c19 | 803 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
804 | ocqe->cqe.res = res; |
805 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
806 | if (is_cqe32) { |
807 | ocqe->cqe.big_cqe[0] = extra1; | |
808 | ocqe->cqe.big_cqe[1] = extra2; | |
809 | } | |
cce4b8b0 PB |
810 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
811 | return true; | |
2b188cc1 JA |
812 | } |
813 | ||
056695bf | 814 | void io_req_cqe_overflow(struct io_kiocb *req) |
68494a65 | 815 | { |
056695bf PB |
816 | io_cqring_event_overflow(req->ctx, req->cqe.user_data, |
817 | req->cqe.res, req->cqe.flags, | |
b24c5d75 PB |
818 | req->big_cqe.extra1, req->big_cqe.extra2); |
819 | memset(&req->big_cqe, 0, sizeof(req->big_cqe)); | |
68494a65 PB |
820 | } |
821 | ||
faf88dde PB |
822 | /* |
823 | * writes to the cq entry need to come after reading head; the | |
824 | * control dependency is enough as we're using WRITE_ONCE to | |
825 | * fill the cq entry | |
826 | */ | |
20d6b633 | 827 | bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow) |
faf88dde PB |
828 | { |
829 | struct io_rings *rings = ctx->rings; | |
830 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
831 | unsigned int free, queued, len; |
832 | ||
aa1df3a3 PB |
833 | /* |
834 | * Posting into the CQ when there are pending overflowed CQEs may break | |
835 | * ordering guarantees, which will affect links, F_MORE users and more. | |
836 | * Force overflow the completion. | |
837 | */ | |
838 | if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) | |
20d6b633 | 839 | return false; |
faf88dde PB |
840 | |
841 | /* userspace may cheat modifying the tail, be safe and do min */ | |
842 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
843 | free = ctx->cq_entries - queued; | |
844 | /* we need a contiguous range, limit based on the current array offset */ | |
845 | len = min(free, ctx->cq_entries - off); | |
846 | if (!len) | |
20d6b633 | 847 | return false; |
faf88dde | 848 | |
b3659a65 PB |
849 | if (ctx->flags & IORING_SETUP_CQE32) { |
850 | off <<= 1; | |
851 | len <<= 1; | |
852 | } | |
853 | ||
faf88dde PB |
854 | ctx->cqe_cached = &rings->cqes[off]; |
855 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
20d6b633 | 856 | return true; |
faf88dde PB |
857 | } |
858 | ||
f66f7342 PB |
859 | static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, |
860 | u32 cflags) | |
bcda7baa | 861 | { |
cd94903d PB |
862 | struct io_uring_cqe *cqe; |
863 | ||
913a571a | 864 | ctx->cq_extra++; |
cd94903d PB |
865 | |
866 | /* | |
867 | * If we can't get a cq entry, userspace overflowed the | |
868 | * submission (by quite a lot). Increment the overflow count in | |
869 | * the ring. | |
870 | */ | |
59fbc409 | 871 | if (likely(io_get_cqe(ctx, &cqe))) { |
e0486f3f DY |
872 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
873 | ||
cd94903d PB |
874 | WRITE_ONCE(cqe->user_data, user_data); |
875 | WRITE_ONCE(cqe->res, res); | |
876 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
877 | |
878 | if (ctx->flags & IORING_SETUP_CQE32) { | |
879 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
880 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
881 | } | |
cd94903d PB |
882 | return true; |
883 | } | |
52120f0f | 884 | return false; |
bcda7baa JA |
885 | } |
886 | ||
931147dd DY |
887 | static void __io_flush_post_cqes(struct io_ring_ctx *ctx) |
888 | __must_hold(&ctx->uring_lock) | |
889 | { | |
890 | struct io_submit_state *state = &ctx->submit_state; | |
891 | unsigned int i; | |
892 | ||
893 | lockdep_assert_held(&ctx->uring_lock); | |
894 | for (i = 0; i < state->cqes_count; i++) { | |
0aa7aa5f | 895 | struct io_uring_cqe *cqe = &ctx->completion_cqes[i]; |
931147dd | 896 | |
f66f7342 | 897 | if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) { |
27122c07 | 898 | if (ctx->lockless_cq) { |
f66f7342 PB |
899 | spin_lock(&ctx->completion_lock); |
900 | io_cqring_event_overflow(ctx, cqe->user_data, | |
901 | cqe->res, cqe->flags, 0, 0); | |
902 | spin_unlock(&ctx->completion_lock); | |
903 | } else { | |
904 | io_cqring_event_overflow(ctx, cqe->user_data, | |
905 | cqe->res, cqe->flags, 0, 0); | |
906 | } | |
907 | } | |
931147dd DY |
908 | } |
909 | state->cqes_count = 0; | |
910 | } | |
911 | ||
b529c96a DY |
912 | static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, |
913 | bool allow_overflow) | |
d245bca6 PB |
914 | { |
915 | bool filled; | |
916 | ||
25399321 | 917 | io_cq_lock(ctx); |
f66f7342 PB |
918 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags); |
919 | if (!filled && allow_overflow) | |
920 | filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); | |
921 | ||
25399321 | 922 | io_cq_unlock_post(ctx); |
d245bca6 PB |
923 | return filled; |
924 | } | |
925 | ||
b529c96a DY |
926 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags) |
927 | { | |
928 | return __io_post_aux_cqe(ctx, user_data, res, cflags, true); | |
929 | } | |
930 | ||
b6b2bb58 PB |
931 | /* |
932 | * A helper for multishot requests posting additional CQEs. | |
933 | * Should only be used from a task_work including IO_URING_F_MULTISHOT. | |
934 | */ | |
935 | bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags) | |
2b188cc1 | 936 | { |
d86eaed1 JA |
937 | struct io_ring_ctx *ctx = req->ctx; |
938 | u64 user_data = req->cqe.user_data; | |
9b8c5475 | 939 | struct io_uring_cqe *cqe; |
9b8c5475 | 940 | |
e0e4ab52 PB |
941 | lockdep_assert(!io_wq_current_is_worker()); |
942 | ||
9b8c5475 | 943 | if (!defer) |
b6b2bb58 | 944 | return __io_post_aux_cqe(ctx, user_data, res, cflags, false); |
9b8c5475 | 945 | |
9b8c5475 DY |
946 | lockdep_assert_held(&ctx->uring_lock); |
947 | ||
0aa7aa5f | 948 | if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->completion_cqes)) { |
f66f7342 | 949 | __io_cq_lock(ctx); |
9b8c5475 DY |
950 | __io_flush_post_cqes(ctx); |
951 | /* no need to flush - flush is deferred */ | |
f66f7342 | 952 | __io_cq_unlock_post(ctx); |
9b8c5475 DY |
953 | } |
954 | ||
955 | /* For defered completions this is not as strict as it is otherwise, | |
956 | * however it's main job is to prevent unbounded posted completions, | |
957 | * and in that it works just as well. | |
958 | */ | |
b6b2bb58 | 959 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) |
9b8c5475 DY |
960 | return false; |
961 | ||
0aa7aa5f | 962 | cqe = &ctx->completion_cqes[ctx->submit_state.cqes_count++]; |
9b8c5475 DY |
963 | cqe->user_data = user_data; |
964 | cqe->res = res; | |
965 | cqe->flags = cflags; | |
966 | return true; | |
967 | } | |
968 | ||
ef8ae64f | 969 | static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
2b188cc1 | 970 | { |
fa18fa22 | 971 | struct io_ring_ctx *ctx = req->ctx; |
2ad4c6d0 | 972 | struct io_rsrc_node *rsrc_node = NULL; |
fa18fa22 PB |
973 | |
974 | io_cq_lock(ctx); | |
00b0db56 PB |
975 | if (!(req->flags & REQ_F_CQE_SKIP)) { |
976 | if (!io_fill_cqe_req(ctx, req)) | |
977 | io_req_cqe_overflow(req); | |
978 | } | |
fa18fa22 | 979 | |
c7dae4ba JA |
980 | /* |
981 | * If we're the last reference to this request, add to our locked | |
982 | * free_list cache. | |
983 | */ | |
de9b4cca | 984 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 985 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 986 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
987 | io_disarm_next(req); |
988 | if (req->link) { | |
989 | io_req_task_queue(req->link); | |
990 | req->link = NULL; | |
991 | } | |
992 | } | |
68a2cc1b | 993 | io_put_kbuf_comp(req); |
3b7a612f PB |
994 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
995 | io_clean_op(req); | |
17bc2837 | 996 | io_put_file(req); |
3b7a612f | 997 | |
2ad4c6d0 | 998 | rsrc_node = req->rsrc_node; |
8197b053 PB |
999 | /* |
1000 | * Selected buffer deallocation in io_clean_op() assumes that | |
1001 | * we don't hold ->completion_lock. Clean them here to avoid | |
1002 | * deadlocks. | |
1003 | */ | |
2fdd6fb5 | 1004 | io_put_task_remote(req->task); |
c2b6c6bc | 1005 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 1006 | ctx->locked_free_nr++; |
180f829f | 1007 | } |
25399321 | 1008 | io_cq_unlock_post(ctx); |
2ad4c6d0 | 1009 | |
ef8ae64f PB |
1010 | if (rsrc_node) { |
1011 | io_ring_submit_lock(ctx, issue_flags); | |
1f2c8f61 | 1012 | io_put_rsrc_node(ctx, rsrc_node); |
ef8ae64f PB |
1013 | io_ring_submit_unlock(ctx, issue_flags); |
1014 | } | |
4e3d9ff9 JA |
1015 | } |
1016 | ||
1bec951c | 1017 | void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
bcda7baa | 1018 | { |
4caa74fd JA |
1019 | struct io_ring_ctx *ctx = req->ctx; |
1020 | ||
1021 | if (ctx->task_complete && ctx->submitter_task != current) { | |
e6aeb272 PB |
1022 | req->io_task_work.func = io_req_task_complete; |
1023 | io_req_task_work_add(req); | |
1024 | } else if (!(issue_flags & IO_URING_F_UNLOCKED) || | |
4caa74fd | 1025 | !(ctx->flags & IORING_SETUP_IOPOLL)) { |
ef8ae64f | 1026 | __io_req_complete_post(req, issue_flags); |
1bec951c | 1027 | } else { |
1bec951c | 1028 | mutex_lock(&ctx->uring_lock); |
ef8ae64f | 1029 | __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED); |
1bec951c PB |
1030 | mutex_unlock(&ctx->uring_lock); |
1031 | } | |
0ddf92e8 JA |
1032 | } |
1033 | ||
973fc83f | 1034 | void io_req_defer_failed(struct io_kiocb *req, s32 res) |
e276ae34 | 1035 | __must_hold(&ctx->uring_lock) |
f41db273 | 1036 | { |
f30bd4d0 | 1037 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; |
a47b255e | 1038 | |
e276ae34 PB |
1039 | lockdep_assert_held(&req->ctx->uring_lock); |
1040 | ||
93d2bcd2 | 1041 | req_set_fail(req); |
97b388d7 | 1042 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
a47b255e PB |
1043 | if (def->fail) |
1044 | def->fail(req); | |
973fc83f | 1045 | io_req_complete_defer(req); |
f41db273 PB |
1046 | } |
1047 | ||
864ea921 PB |
1048 | /* |
1049 | * Don't initialise the fields below on every allocation, but do that in | |
1050 | * advance and keep them valid across allocations. | |
1051 | */ | |
1052 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
1053 | { | |
1054 | req->ctx = ctx; | |
1055 | req->link = NULL; | |
1056 | req->async_data = NULL; | |
1057 | /* not necessary, but safer to zero */ | |
31d3ba92 | 1058 | memset(&req->cqe, 0, sizeof(req->cqe)); |
b24c5d75 | 1059 | memset(&req->big_cqe, 0, sizeof(req->big_cqe)); |
864ea921 PB |
1060 | } |
1061 | ||
dac7a098 | 1062 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 1063 | struct io_submit_state *state) |
dac7a098 | 1064 | { |
79ebeaee | 1065 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1066 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 1067 | ctx->locked_free_nr = 0; |
79ebeaee | 1068 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
1069 | } |
1070 | ||
5d5901a3 PB |
1071 | /* |
1072 | * A request might get retired back into the request caches even before opcode | |
1073 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
1074 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
1075 | * and with extra caution to not get a request that is still worked on. | |
1076 | */ | |
bd1a3783 | 1077 | __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 1078 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 1079 | { |
864ea921 | 1080 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 1081 | void *reqs[IO_REQ_ALLOC_BATCH]; |
864ea921 | 1082 | int ret, i; |
e5d1bc0a | 1083 | |
23a5c43b PB |
1084 | /* |
1085 | * If we have more than a batch's worth of requests in our IRQ side | |
1086 | * locked cache, grab the lock and move them over to our submission | |
1087 | * side cache. | |
1088 | */ | |
a6d97a8a | 1089 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 1090 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 1091 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
1092 | return true; |
1093 | } | |
e5d1bc0a | 1094 | |
3ab665b7 | 1095 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 1096 | |
864ea921 PB |
1097 | /* |
1098 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1099 | * retry single alloc to be on the safe side. | |
1100 | */ | |
1101 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
1102 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
1103 | if (!reqs[0]) | |
a33ae9ce | 1104 | return false; |
864ea921 | 1105 | ret = 1; |
2b188cc1 | 1106 | } |
864ea921 | 1107 | |
37f0e767 | 1108 | percpu_ref_get_many(&ctx->refs, ret); |
3ab665b7 | 1109 | for (i = 0; i < ret; i++) { |
23a5c43b | 1110 | struct io_kiocb *req = reqs[i]; |
3ab665b7 PB |
1111 | |
1112 | io_preinit_req(req, ctx); | |
fa05457a | 1113 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 1114 | } |
a33ae9ce PB |
1115 | return true; |
1116 | } | |
1117 | ||
03adabe8 PB |
1118 | __cold void io_free_req(struct io_kiocb *req) |
1119 | { | |
6ec9afc7 PB |
1120 | /* refs were already put, restore them for io_req_task_complete() */ |
1121 | req->flags &= ~REQ_F_REFCOUNT; | |
1122 | /* we only want to free it, don't post CQEs */ | |
1123 | req->flags |= REQ_F_CQE_SKIP; | |
1124 | req->io_task_work.func = io_req_task_complete; | |
03adabe8 PB |
1125 | io_req_task_work_add(req); |
1126 | } | |
1127 | ||
d81499bf PB |
1128 | static void __io_req_find_next_prep(struct io_kiocb *req) |
1129 | { | |
1130 | struct io_ring_ctx *ctx = req->ctx; | |
d81499bf | 1131 | |
6971253f | 1132 | spin_lock(&ctx->completion_lock); |
305bef98 | 1133 | io_disarm_next(req); |
6971253f | 1134 | spin_unlock(&ctx->completion_lock); |
d81499bf PB |
1135 | } |
1136 | ||
1137 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 1138 | { |
33cc89a9 | 1139 | struct io_kiocb *nxt; |
944e58bf | 1140 | |
9e645e11 JA |
1141 | /* |
1142 | * If LINK is set, we have dependent requests in this chain. If we | |
1143 | * didn't fail this request, queue the first one up, moving any other | |
1144 | * dependencies to the next request. In case of failure, fail the rest | |
1145 | * of the chain. | |
1146 | */ | |
d81499bf PB |
1147 | if (unlikely(req->flags & IO_DISARM_MASK)) |
1148 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
1149 | nxt = req->link; |
1150 | req->link = NULL; | |
1151 | return nxt; | |
4d7dd462 | 1152 | } |
9e645e11 | 1153 | |
a282967c | 1154 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts) |
2c32395d PB |
1155 | { |
1156 | if (!ctx) | |
1157 | return; | |
ef060ea9 JA |
1158 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1159 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
a282967c | 1160 | if (ts->locked) { |
c450178d | 1161 | io_submit_flush_completions(ctx); |
2c32395d | 1162 | mutex_unlock(&ctx->uring_lock); |
a282967c | 1163 | ts->locked = false; |
2c32395d PB |
1164 | } |
1165 | percpu_ref_put(&ctx->refs); | |
1166 | } | |
1167 | ||
af5d68f8 JA |
1168 | /* |
1169 | * Run queued task_work, returning the number of entries processed in *count. | |
1170 | * If more entries than max_entries are available, stop processing once this | |
1171 | * is reached and return the rest of the list. | |
1172 | */ | |
1173 | struct llist_node *io_handle_tw_list(struct llist_node *node, | |
1174 | unsigned int *count, | |
1175 | unsigned int max_entries) | |
9f8d032a | 1176 | { |
42c0905f JA |
1177 | struct io_ring_ctx *ctx = NULL; |
1178 | struct io_tw_state ts = { }; | |
c6dd763c | 1179 | |
592b4805 | 1180 | do { |
f88262e6 | 1181 | struct llist_node *next = node->next; |
9f8d032a HX |
1182 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1183 | io_task_work.node); | |
1184 | ||
42c0905f JA |
1185 | if (req->ctx != ctx) { |
1186 | ctx_flush_and_put(ctx, &ts); | |
1187 | ctx = req->ctx; | |
9f8d032a | 1188 | /* if not contended, grab and improve batching */ |
42c0905f JA |
1189 | ts.locked = mutex_trylock(&ctx->uring_lock); |
1190 | percpu_ref_get(&ctx->refs); | |
13bfa6f1 | 1191 | } |
c92fcfc2 JA |
1192 | INDIRECT_CALL_2(req->io_task_work.func, |
1193 | io_poll_task_func, io_req_rw_complete, | |
42c0905f | 1194 | req, &ts); |
9f8d032a | 1195 | node = next; |
2708af1a | 1196 | (*count)++; |
f5868008 | 1197 | if (unlikely(need_resched())) { |
42c0905f JA |
1198 | ctx_flush_and_put(ctx, &ts); |
1199 | ctx = NULL; | |
f5868008 JA |
1200 | cond_resched(); |
1201 | } | |
af5d68f8 | 1202 | } while (node && *count < max_entries); |
c6dd763c | 1203 | |
42c0905f | 1204 | ctx_flush_and_put(ctx, &ts); |
af5d68f8 | 1205 | return node; |
9f8d032a HX |
1206 | } |
1207 | ||
923d1592 DY |
1208 | /** |
1209 | * io_llist_xchg - swap all entries in a lock-less list | |
1210 | * @head: the head of lock-less list to delete all entries | |
1211 | * @new: new entry as the head of the list | |
1212 | * | |
1213 | * If list is empty, return NULL, otherwise, return the pointer to the first entry. | |
1214 | * The order of entries returned is from the newest to the oldest added one. | |
1215 | */ | |
1216 | static inline struct llist_node *io_llist_xchg(struct llist_head *head, | |
1217 | struct llist_node *new) | |
1218 | { | |
1219 | return xchg(&head->first, new); | |
1220 | } | |
1221 | ||
dfbe5561 | 1222 | static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync) |
10e1c0d5 JA |
1223 | { |
1224 | struct llist_node *node = llist_del_all(&tctx->task_list); | |
dfbe5561 | 1225 | struct io_ring_ctx *last_ctx = NULL; |
10e1c0d5 JA |
1226 | struct io_kiocb *req; |
1227 | ||
1228 | while (node) { | |
1229 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1230 | node = node->next; | |
dfbe5561 JA |
1231 | if (sync && last_ctx != req->ctx) { |
1232 | if (last_ctx) { | |
1233 | flush_delayed_work(&last_ctx->fallback_work); | |
1234 | percpu_ref_put(&last_ctx->refs); | |
1235 | } | |
1236 | last_ctx = req->ctx; | |
1237 | percpu_ref_get(&last_ctx->refs); | |
1238 | } | |
10e1c0d5 JA |
1239 | if (llist_add(&req->io_task_work.node, |
1240 | &req->ctx->fallback_llist)) | |
1241 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1242 | } | |
dfbe5561 JA |
1243 | |
1244 | if (last_ctx) { | |
1245 | flush_delayed_work(&last_ctx->fallback_work); | |
1246 | percpu_ref_put(&last_ctx->refs); | |
1247 | } | |
10e1c0d5 JA |
1248 | } |
1249 | ||
af5d68f8 JA |
1250 | struct llist_node *tctx_task_work_run(struct io_uring_task *tctx, |
1251 | unsigned int max_entries, | |
1252 | unsigned int *count) | |
c40f6379 | 1253 | { |
77e443ab | 1254 | struct llist_node *node; |
3a0c037b | 1255 | |
77e443ab | 1256 | if (unlikely(current->flags & PF_EXITING)) { |
dfbe5561 | 1257 | io_fallback_tw(tctx, true); |
af5d68f8 | 1258 | return NULL; |
77e443ab | 1259 | } |
3a0c037b | 1260 | |
592b4805 | 1261 | node = llist_del_all(&tctx->task_list); |
af5d68f8 JA |
1262 | if (node) { |
1263 | node = llist_reverse_order(node); | |
1264 | node = io_handle_tw_list(node, count, max_entries); | |
1265 | } | |
3cc7fdb9 | 1266 | |
8d664282 JA |
1267 | /* relaxed read is enough as only the task itself sets ->in_cancel */ |
1268 | if (unlikely(atomic_read(&tctx->in_cancel))) | |
3cc7fdb9 | 1269 | io_uring_drop_tctx_refs(current); |
c6dd763c | 1270 | |
af5d68f8 JA |
1271 | trace_io_uring_task_work_run(tctx, *count); |
1272 | return node; | |
1273 | } | |
1274 | ||
1275 | void tctx_task_work(struct callback_head *cb) | |
1276 | { | |
1277 | struct io_uring_task *tctx; | |
1278 | struct llist_node *ret; | |
1279 | unsigned int count = 0; | |
1280 | ||
1281 | tctx = container_of(cb, struct io_uring_task, task_work); | |
1282 | ret = tctx_task_work_run(tctx, UINT_MAX, &count); | |
1283 | /* can't happen */ | |
1284 | WARN_ON_ONCE(ret); | |
7cbf1722 JA |
1285 | } |
1286 | ||
91c7884a | 1287 | static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags) |
c0e0d6ba DY |
1288 | { |
1289 | struct io_ring_ctx *ctx = req->ctx; | |
8751d154 | 1290 | unsigned nr_wait, nr_tw, nr_tw_prev; |
e8c40771 | 1291 | struct llist_node *head; |
c0e0d6ba | 1292 | |
b4bc35cf PB |
1293 | /* See comment above IO_CQ_WAKE_INIT */ |
1294 | BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES); | |
c0e0d6ba | 1295 | |
b4bc35cf PB |
1296 | /* |
1297 | * We don't know how many reuqests is there in the link and whether | |
1298 | * they can even be queued lazily, fall back to non-lazy. | |
1299 | */ | |
8751d154 PB |
1300 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) |
1301 | flags &= ~IOU_F_TWQ_LAZY_WAKE; | |
ce8e04f6 | 1302 | |
e8c40771 | 1303 | head = READ_ONCE(ctx->work_llist.first); |
51509400 | 1304 | do { |
8751d154 | 1305 | nr_tw_prev = 0; |
e8c40771 PB |
1306 | if (head) { |
1307 | struct io_kiocb *first_req = container_of(head, | |
8751d154 PB |
1308 | struct io_kiocb, |
1309 | io_task_work.node); | |
1310 | /* | |
1311 | * Might be executed at any moment, rely on | |
1312 | * SLAB_TYPESAFE_BY_RCU to keep it alive. | |
1313 | */ | |
1314 | nr_tw_prev = READ_ONCE(first_req->nr_tw); | |
1315 | } | |
b4bc35cf PB |
1316 | |
1317 | /* | |
1318 | * Theoretically, it can overflow, but that's fine as one of | |
1319 | * previous adds should've tried to wake the task. | |
1320 | */ | |
8751d154 | 1321 | nr_tw = nr_tw_prev + 1; |
8751d154 | 1322 | if (!(flags & IOU_F_TWQ_LAZY_WAKE)) |
b4bc35cf | 1323 | nr_tw = IO_CQ_WAKE_FORCE; |
8751d154 PB |
1324 | |
1325 | req->nr_tw = nr_tw; | |
e8c40771 PB |
1326 | req->io_task_work.node.next = head; |
1327 | } while (!try_cmpxchg(&ctx->work_llist.first, &head, | |
51509400 PB |
1328 | &req->io_task_work.node)); |
1329 | ||
d381099f PB |
1330 | /* |
1331 | * cmpxchg implies a full barrier, which pairs with the barrier | |
1332 | * in set_current_state() on the io_cqring_wait() side. It's used | |
1333 | * to ensure that either we see updated ->cq_wait_nr, or waiters | |
1334 | * going to sleep will observe the work added to the list, which | |
1335 | * is similar to the wait/wawke task state sync. | |
1336 | */ | |
1337 | ||
e8c40771 | 1338 | if (!head) { |
8751d154 PB |
1339 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1340 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1341 | if (ctx->has_evfd) | |
1342 | io_eventfd_signal(ctx); | |
c0e0d6ba DY |
1343 | } |
1344 | ||
8751d154 | 1345 | nr_wait = atomic_read(&ctx->cq_wait_nr); |
b4bc35cf PB |
1346 | /* not enough or no one is waiting */ |
1347 | if (nr_tw < nr_wait) | |
8751d154 | 1348 | return; |
b4bc35cf PB |
1349 | /* the previous add has already woken it up */ |
1350 | if (nr_tw_prev >= nr_wait) | |
8751d154 | 1351 | return; |
8751d154 | 1352 | wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE); |
c0e0d6ba DY |
1353 | } |
1354 | ||
91c7884a | 1355 | static void io_req_normal_work_add(struct io_kiocb *req) |
7cbf1722 | 1356 | { |
c34398a8 | 1357 | struct io_uring_task *tctx = req->task->io_uring; |
9f010507 | 1358 | struct io_ring_ctx *ctx = req->ctx; |
7cbf1722 | 1359 | |
7cbf1722 | 1360 | /* task_work already pending, we're done */ |
32d91f05 | 1361 | if (!llist_add(&req->io_task_work.node, &tctx->task_list)) |
e09ee510 | 1362 | return; |
7cbf1722 | 1363 | |
ef060ea9 JA |
1364 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1365 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1366 | ||
af5d68f8 | 1367 | /* SQPOLL doesn't need the task_work added, it'll run it itself */ |
78f9b61b JA |
1368 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
1369 | struct io_sq_data *sqd = ctx->sq_data; | |
1370 | ||
1371 | if (wq_has_sleeper(&sqd->wait)) | |
1372 | wake_up(&sqd->wait); | |
af5d68f8 | 1373 | return; |
78f9b61b | 1374 | } |
af5d68f8 | 1375 | |
3fe07bcd | 1376 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1377 | return; |
2215bed9 | 1378 | |
dfbe5561 | 1379 | io_fallback_tw(tctx, false); |
c0e0d6ba DY |
1380 | } |
1381 | ||
91c7884a PB |
1382 | void __io_req_task_work_add(struct io_kiocb *req, unsigned flags) |
1383 | { | |
1384 | if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) { | |
1385 | rcu_read_lock(); | |
1386 | io_req_local_work_add(req, flags); | |
1387 | rcu_read_unlock(); | |
1388 | } else { | |
1389 | io_req_normal_work_add(req); | |
1390 | } | |
1391 | } | |
1392 | ||
c0e0d6ba DY |
1393 | static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) |
1394 | { | |
1395 | struct llist_node *node; | |
1396 | ||
1397 | node = llist_del_all(&ctx->work_llist); | |
1398 | while (node) { | |
1399 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1400 | io_task_work.node); | |
1401 | ||
1402 | node = node->next; | |
91c7884a | 1403 | io_req_normal_work_add(req); |
c0e0d6ba DY |
1404 | } |
1405 | } | |
1406 | ||
9fe3eaea JA |
1407 | static bool io_run_local_work_continue(struct io_ring_ctx *ctx, int events, |
1408 | int min_events) | |
1409 | { | |
1410 | if (llist_empty(&ctx->work_llist)) | |
1411 | return false; | |
1412 | if (events < min_events) | |
1413 | return true; | |
1414 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) | |
1415 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1416 | return false; | |
1417 | } | |
1418 | ||
1419 | static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts, | |
1420 | int min_events) | |
c0e0d6ba | 1421 | { |
c0e0d6ba | 1422 | struct llist_node *node; |
c3f4d39e | 1423 | unsigned int loops = 0; |
140102ae | 1424 | int ret = 0; |
c0e0d6ba | 1425 | |
140102ae | 1426 | if (WARN_ON_ONCE(ctx->submitter_task != current)) |
c0e0d6ba | 1427 | return -EEXIST; |
c3f4d39e PB |
1428 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1429 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
c0e0d6ba | 1430 | again: |
3af0356c JA |
1431 | /* |
1432 | * llists are in reverse order, flip it back the right way before | |
1433 | * running the pending items. | |
1434 | */ | |
1435 | node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL)); | |
c3f4d39e | 1436 | while (node) { |
c0e0d6ba DY |
1437 | struct llist_node *next = node->next; |
1438 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1439 | io_task_work.node); | |
c92fcfc2 JA |
1440 | INDIRECT_CALL_2(req->io_task_work.func, |
1441 | io_poll_task_func, io_req_rw_complete, | |
1442 | req, ts); | |
c0e0d6ba DY |
1443 | ret++; |
1444 | node = next; | |
1445 | } | |
c3f4d39e | 1446 | loops++; |
c0e0d6ba | 1447 | |
9fe3eaea | 1448 | if (io_run_local_work_continue(ctx, ret, min_events)) |
c0e0d6ba | 1449 | goto again; |
a282967c | 1450 | if (ts->locked) { |
c0e0d6ba | 1451 | io_submit_flush_completions(ctx); |
9fe3eaea | 1452 | if (io_run_local_work_continue(ctx, ret, min_events)) |
b0b7a7d2 PB |
1453 | goto again; |
1454 | } | |
9fe3eaea | 1455 | |
f75d5036 | 1456 | trace_io_uring_local_work_run(ctx, ret, loops); |
c0e0d6ba | 1457 | return ret; |
8ac5d85a JA |
1458 | } |
1459 | ||
9fe3eaea JA |
1460 | static inline int io_run_local_work_locked(struct io_ring_ctx *ctx, |
1461 | int min_events) | |
360173ab | 1462 | { |
a282967c | 1463 | struct io_tw_state ts = { .locked = true, }; |
360173ab PB |
1464 | int ret; |
1465 | ||
1466 | if (llist_empty(&ctx->work_llist)) | |
1467 | return 0; | |
1468 | ||
9fe3eaea | 1469 | ret = __io_run_local_work(ctx, &ts, min_events); |
360173ab | 1470 | /* shouldn't happen! */ |
a282967c | 1471 | if (WARN_ON_ONCE(!ts.locked)) |
360173ab PB |
1472 | mutex_lock(&ctx->uring_lock); |
1473 | return ret; | |
1474 | } | |
1475 | ||
9fe3eaea | 1476 | static int io_run_local_work(struct io_ring_ctx *ctx, int min_events) |
8ac5d85a | 1477 | { |
a282967c | 1478 | struct io_tw_state ts = {}; |
8ac5d85a JA |
1479 | int ret; |
1480 | ||
a282967c | 1481 | ts.locked = mutex_trylock(&ctx->uring_lock); |
9fe3eaea | 1482 | ret = __io_run_local_work(ctx, &ts, min_events); |
a282967c | 1483 | if (ts.locked) |
8ac5d85a JA |
1484 | mutex_unlock(&ctx->uring_lock); |
1485 | ||
1486 | return ret; | |
c0e0d6ba DY |
1487 | } |
1488 | ||
a282967c | 1489 | static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts) |
c40f6379 | 1490 | { |
a282967c | 1491 | io_tw_lock(req->ctx, ts); |
973fc83f | 1492 | io_req_defer_failed(req, req->cqe.res); |
c40f6379 JA |
1493 | } |
1494 | ||
a282967c | 1495 | void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts) |
c40f6379 | 1496 | { |
a282967c | 1497 | io_tw_lock(req->ctx, ts); |
316319e8 | 1498 | /* req->task == current here, checking PF_EXITING is safe */ |
6bb30855 | 1499 | if (unlikely(req->task->flags & PF_EXITING)) |
973fc83f | 1500 | io_req_defer_failed(req, -EFAULT); |
6bb30855 | 1501 | else if (req->flags & REQ_F_FORCE_ASYNC) |
a282967c | 1502 | io_queue_iowq(req, ts); |
6bb30855 DY |
1503 | else |
1504 | io_queue_sqe(req); | |
c40f6379 JA |
1505 | } |
1506 | ||
59915143 | 1507 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1508 | { |
97b388d7 | 1509 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1510 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1511 | io_req_task_work_add(req); |
c40f6379 JA |
1512 | } |
1513 | ||
f3b44f92 | 1514 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1515 | { |
5b0a6acc | 1516 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1517 | io_req_task_work_add(req); |
a3df7698 PB |
1518 | } |
1519 | ||
59915143 | 1520 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1521 | { |
9b5f7bd9 | 1522 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1523 | |
1524 | if (nxt) | |
906a8c3f | 1525 | io_req_task_queue(nxt); |
c69f8dbe JL |
1526 | } |
1527 | ||
ec26c225 PB |
1528 | static void io_free_batch_list(struct io_ring_ctx *ctx, |
1529 | struct io_wq_work_node *node) | |
3aa83bfb | 1530 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1531 | { |
3aa83bfb PB |
1532 | do { |
1533 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1534 | comp_list); | |
2d6500d4 | 1535 | |
a538be5b PB |
1536 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1537 | if (req->flags & REQ_F_REFCOUNT) { | |
1538 | node = req->comp_list.next; | |
1539 | if (!req_ref_put_and_test(req)) | |
1540 | continue; | |
1541 | } | |
b605a7fa PB |
1542 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1543 | struct async_poll *apoll = req->apoll; | |
1544 | ||
1545 | if (apoll->double_poll) | |
1546 | kfree(apoll->double_poll); | |
9731bc98 JA |
1547 | if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) |
1548 | kfree(apoll); | |
b605a7fa PB |
1549 | req->flags &= ~REQ_F_POLLED; |
1550 | } | |
da1a08c5 | 1551 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1552 | io_queue_next(req); |
a538be5b PB |
1553 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1554 | io_clean_op(req); | |
c1e53a69 | 1555 | } |
17bc2837 | 1556 | io_put_file(req); |
2d6500d4 | 1557 | |
ab409402 | 1558 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1559 | |
2fdd6fb5 | 1560 | io_put_task(req->task); |
c1e53a69 | 1561 | node = req->comp_list.next; |
fa05457a | 1562 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1563 | } while (node); |
7a743e22 PB |
1564 | } |
1565 | ||
ec26c225 | 1566 | void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1567 | __must_hold(&ctx->uring_lock) |
905c172f | 1568 | { |
cd0ca2e0 | 1569 | struct io_submit_state *state = &ctx->submit_state; |
fa780334 | 1570 | struct io_wq_work_node *node; |
905c172f | 1571 | |
f66f7342 | 1572 | __io_cq_lock(ctx); |
931147dd DY |
1573 | /* must come first to preserve CQE ordering in failure cases */ |
1574 | if (state->cqes_count) | |
1575 | __io_flush_post_cqes(ctx); | |
fa780334 | 1576 | __wq_list_for_each(node, &state->compl_reqs) { |
d9dee430 PB |
1577 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1578 | comp_list); | |
3d4aeb9f | 1579 | |
f66f7342 | 1580 | if (!(req->flags & REQ_F_CQE_SKIP) && |
00b0db56 | 1581 | unlikely(!io_fill_cqe_req(ctx, req))) { |
27122c07 | 1582 | if (ctx->lockless_cq) { |
f66f7342 PB |
1583 | spin_lock(&ctx->completion_lock); |
1584 | io_req_cqe_overflow(req); | |
1585 | spin_unlock(&ctx->completion_lock); | |
1586 | } else { | |
1587 | io_req_cqe_overflow(req); | |
1588 | } | |
1589 | } | |
905c172f | 1590 | } |
c98c81a4 | 1591 | __io_cq_unlock_post(ctx); |
d9dee430 | 1592 | |
931147dd DY |
1593 | if (!wq_list_empty(&ctx->submit_state.compl_reqs)) { |
1594 | io_free_batch_list(ctx, state->compl_reqs.first); | |
1595 | INIT_WQ_LIST(&state->compl_reqs); | |
1596 | } | |
7a743e22 PB |
1597 | } |
1598 | ||
6c503150 | 1599 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1600 | { |
1601 | /* See comment at the top of this file */ | |
1602 | smp_rmb(); | |
e23de15f | 1603 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1604 | } |
1605 | ||
def596e9 JA |
1606 | /* |
1607 | * We can't just wait for polled events to come to us, we have to actively | |
1608 | * find and complete them. | |
1609 | */ | |
c072481d | 1610 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1611 | { |
1612 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1613 | return; | |
1614 | ||
1615 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1616 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1617 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1618 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1619 | break; |
08f5439f JA |
1620 | /* |
1621 | * Ensure we allow local-to-the-cpu processing to take place, | |
1622 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1623 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1624 | */ |
3fcee5a6 PB |
1625 | if (need_resched()) { |
1626 | mutex_unlock(&ctx->uring_lock); | |
1627 | cond_resched(); | |
1628 | mutex_lock(&ctx->uring_lock); | |
1629 | } | |
def596e9 JA |
1630 | } |
1631 | mutex_unlock(&ctx->uring_lock); | |
1632 | } | |
1633 | ||
7668b92a | 1634 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1635 | { |
7668b92a | 1636 | unsigned int nr_events = 0; |
155bc950 | 1637 | unsigned long check_cq; |
500f9fba | 1638 | |
76de6749 PB |
1639 | if (!io_allowed_run_tw(ctx)) |
1640 | return -EEXIST; | |
1641 | ||
3a08576b PB |
1642 | check_cq = READ_ONCE(ctx->check_cq); |
1643 | if (unlikely(check_cq)) { | |
1644 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
a85381d8 | 1645 | __io_cqring_overflow_flush(ctx); |
3a08576b PB |
1646 | /* |
1647 | * Similarly do not spin if we have not informed the user of any | |
1648 | * dropped CQE. | |
1649 | */ | |
1650 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1651 | return -EBADR; | |
1652 | } | |
f39c8a5b PB |
1653 | /* |
1654 | * Don't enter poll loop if we already have events pending. | |
1655 | * If we do, we can potentially be spinning for commands that | |
1656 | * already triggered a CQE (eg in error). | |
1657 | */ | |
f39c8a5b | 1658 | if (io_cqring_events(ctx)) |
d487b43c | 1659 | return 0; |
155bc950 | 1660 | |
def596e9 | 1661 | do { |
9e4bef2b JA |
1662 | int ret = 0; |
1663 | ||
500f9fba JA |
1664 | /* |
1665 | * If a submit got punted to a workqueue, we can have the | |
1666 | * application entering polling for a command before it gets | |
1667 | * issued. That app will hold the uring_lock for the duration | |
1668 | * of the poll right here, so we need to take a breather every | |
1669 | * now and then to ensure that the issue has a chance to add | |
1670 | * the poll to the issued list. Otherwise we can spin here | |
1671 | * forever, while the workqueue is stuck trying to acquire the | |
1672 | * very same mutex. | |
1673 | */ | |
dac6a0ea JA |
1674 | if (wq_list_empty(&ctx->iopoll_list) || |
1675 | io_task_work_pending(ctx)) { | |
8f487ef2 PB |
1676 | u32 tail = ctx->cached_cq_tail; |
1677 | ||
9fe3eaea | 1678 | (void) io_run_local_work_locked(ctx, min); |
def596e9 | 1679 | |
dac6a0ea JA |
1680 | if (task_work_pending(current) || |
1681 | wq_list_empty(&ctx->iopoll_list)) { | |
dac6a0ea | 1682 | mutex_unlock(&ctx->uring_lock); |
9d54bd6a | 1683 | io_run_task_work(); |
dac6a0ea | 1684 | mutex_lock(&ctx->uring_lock); |
dac6a0ea | 1685 | } |
8f487ef2 PB |
1686 | /* some requests don't go through iopoll_list */ |
1687 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1688 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1689 | break; |
500f9fba | 1690 | } |
5ba3c874 | 1691 | ret = io_do_iopoll(ctx, !min); |
9e4bef2b JA |
1692 | if (unlikely(ret < 0)) |
1693 | return ret; | |
dc314886 PB |
1694 | |
1695 | if (task_sigpending(current)) | |
1696 | return -EINTR; | |
9e4bef2b | 1697 | if (need_resched()) |
5ba3c874 | 1698 | break; |
d487b43c | 1699 | |
5ba3c874 | 1700 | nr_events += ret; |
9e4bef2b | 1701 | } while (nr_events < min); |
d487b43c | 1702 | |
9e4bef2b | 1703 | return 0; |
def596e9 | 1704 | } |
7012c815 | 1705 | |
a282967c | 1706 | void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts) |
8ef12efe | 1707 | { |
a282967c | 1708 | if (ts->locked) |
9da070b1 | 1709 | io_req_complete_defer(req); |
7012c815 | 1710 | else |
27f35fe9 | 1711 | io_req_complete_post(req, IO_URING_F_UNLOCKED); |
8ef12efe JA |
1712 | } |
1713 | ||
def596e9 JA |
1714 | /* |
1715 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1716 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1717 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1718 | * accessing the kiocb cookie. |
1719 | */ | |
9882131c | 1720 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1721 | { |
1722 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1723 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1724 | |
1725 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1726 | if (unlikely(needs_lock)) |
cb3d8972 | 1727 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1728 | |
1729 | /* | |
1730 | * Track whether we have multiple files in our lists. This will impact | |
1731 | * how we do polling eventually, not spinning if we're on potentially | |
1732 | * different devices. | |
1733 | */ | |
5eef4e87 | 1734 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1735 | ctx->poll_multi_queue = false; |
1736 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1737 | struct io_kiocb *list_req; |
1738 | ||
5eef4e87 PB |
1739 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1740 | comp_list); | |
30da1b45 | 1741 | if (list_req->file != req->file) |
915b3dde | 1742 | ctx->poll_multi_queue = true; |
def596e9 JA |
1743 | } |
1744 | ||
1745 | /* | |
1746 | * For fast devices, IO may have already completed. If it has, add | |
1747 | * it to the front so we find it first. | |
1748 | */ | |
65a6543d | 1749 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1750 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1751 | else |
5eef4e87 | 1752 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1753 | |
3b44b371 | 1754 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1755 | /* |
1756 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1757 | * in sq thread task context or in io worker task context. If | |
1758 | * current task context is sq thread, we don't need to check | |
1759 | * whether should wake up sq thread. | |
1760 | */ | |
1761 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1762 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1763 | wake_up(&ctx->sq_data->wait); | |
1764 | ||
1765 | mutex_unlock(&ctx->uring_lock); | |
1766 | } | |
def596e9 JA |
1767 | } |
1768 | ||
4bcb982c | 1769 | io_req_flags_t io_file_get_flags(struct file *file) |
88459b50 | 1770 | { |
4bcb982c | 1771 | io_req_flags_t res = 0; |
af197f50 | 1772 | |
53cfd5ce | 1773 | if (S_ISREG(file_inode(file)->i_mode)) |
8487f083 | 1774 | res |= REQ_F_ISREG; |
b9a6c945 | 1775 | if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT)) |
8487f083 | 1776 | res |= REQ_F_SUPPORT_NOWAIT; |
88459b50 | 1777 | return res; |
2b188cc1 JA |
1778 | } |
1779 | ||
99f15d8d | 1780 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1781 | { |
f30bd4d0 BL |
1782 | WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size); |
1783 | req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL); | |
d886e185 PB |
1784 | if (req->async_data) { |
1785 | req->flags |= REQ_F_ASYNC_DATA; | |
1786 | return false; | |
1787 | } | |
1788 | return true; | |
3d9932a8 XW |
1789 | } |
1790 | ||
f3b44f92 | 1791 | int io_req_prep_async(struct io_kiocb *req) |
f67676d1 | 1792 | { |
f30bd4d0 | 1793 | const struct io_cold_def *cdef = &io_cold_defs[req->opcode]; |
a7dd2782 | 1794 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
0702e536 JA |
1795 | |
1796 | /* assign early for deferred execution for non-fixed file */ | |
54aa7f23 | 1797 | if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file) |
0702e536 | 1798 | req->file = io_file_get_normal(req, req->cqe.fd); |
f30bd4d0 | 1799 | if (!cdef->prep_async) |
0702e536 JA |
1800 | return 0; |
1801 | if (WARN_ON_ONCE(req_has_async_data(req))) | |
1802 | return -EFAULT; | |
f30bd4d0 | 1803 | if (!def->manual_alloc) { |
59169439 PB |
1804 | if (io_alloc_async_data(req)) |
1805 | return -EAGAIN; | |
1806 | } | |
f30bd4d0 | 1807 | return cdef->prep_async(req); |
bfe76559 PB |
1808 | } |
1809 | ||
9cf7c104 PB |
1810 | static u32 io_get_sequence(struct io_kiocb *req) |
1811 | { | |
a3dbdf54 | 1812 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1813 | struct io_kiocb *cur; |
9cf7c104 | 1814 | |
a3dbdf54 | 1815 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1816 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1817 | seq--; |
1818 | return seq; | |
9cf7c104 PB |
1819 | } |
1820 | ||
c072481d | 1821 | static __cold void io_drain_req(struct io_kiocb *req) |
e276ae34 | 1822 | __must_hold(&ctx->uring_lock) |
de0617e4 | 1823 | { |
a197f664 | 1824 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1825 | struct io_defer_entry *de; |
f67676d1 | 1826 | int ret; |
e0eb71dc | 1827 | u32 seq = io_get_sequence(req); |
3c19966d | 1828 | |
9d858b21 | 1829 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1830 | spin_lock(&ctx->completion_lock); |
5e371265 | 1831 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1832 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1833 | queue: |
10c66904 | 1834 | ctx->drain_active = false; |
e0eb71dc PB |
1835 | io_req_task_queue(req); |
1836 | return; | |
10c66904 | 1837 | } |
e302f104 | 1838 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1839 | |
cbdcb435 | 1840 | io_prep_async_link(req); |
27dc8338 | 1841 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1842 | if (!de) { |
1b48773f | 1843 | ret = -ENOMEM; |
ef5c600a DY |
1844 | io_req_defer_failed(req, ret); |
1845 | return; | |
76cc33d7 | 1846 | } |
2d28390a | 1847 | |
79ebeaee | 1848 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1849 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1850 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1851 | kfree(de); |
e0eb71dc | 1852 | goto queue; |
de0617e4 JA |
1853 | } |
1854 | ||
48863ffd | 1855 | trace_io_uring_defer(req); |
27dc8338 | 1856 | de->req = req; |
9cf7c104 | 1857 | de->seq = seq; |
27dc8338 | 1858 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1859 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1860 | } |
1861 | ||
f4992544 JA |
1862 | static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def, |
1863 | unsigned int issue_flags) | |
6bf9c47a | 1864 | { |
f4992544 | 1865 | if (req->file || !def->needs_file) |
6bf9c47a JA |
1866 | return true; |
1867 | ||
1868 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1869 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1870 | else |
cef216fc | 1871 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1872 | |
772f5e00 | 1873 | return !!req->file; |
6bf9c47a JA |
1874 | } |
1875 | ||
889fca73 | 1876 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1877 | { |
a7dd2782 | 1878 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
5730b27e | 1879 | const struct cred *creds = NULL; |
d625c6ee | 1880 | int ret; |
2b188cc1 | 1881 | |
f4992544 | 1882 | if (unlikely(!io_assign_file(req, def, issue_flags))) |
70152140 JA |
1883 | return -EBADF; |
1884 | ||
6878b40e | 1885 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1886 | creds = override_creds(req->creds); |
5730b27e | 1887 | |
fcde59fe | 1888 | if (!def->audit_skip) |
5bd2182d PM |
1889 | audit_uring_entry(req->opcode); |
1890 | ||
0702e536 | 1891 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1892 | |
fcde59fe | 1893 | if (!def->audit_skip) |
5bd2182d PM |
1894 | audit_uring_exit(!ret, ret); |
1895 | ||
5730b27e JA |
1896 | if (creds) |
1897 | revert_creds(creds); | |
97b388d7 | 1898 | |
75d7b3ae PB |
1899 | if (ret == IOU_OK) { |
1900 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
9da070b1 | 1901 | io_req_complete_defer(req); |
75d7b3ae | 1902 | else |
1bec951c | 1903 | io_req_complete_post(req, issue_flags); |
97b388d7 | 1904 | |
9b43ef3d PB |
1905 | return 0; |
1906 | } | |
def596e9 | 1907 | |
e0b23d99 PB |
1908 | if (ret == IOU_ISSUE_SKIP_COMPLETE) { |
1909 | ret = 0; | |
1910 | io_arm_ltimeout(req); | |
def596e9 | 1911 | |
e0b23d99 PB |
1912 | /* If the op doesn't have a file, we're not polling for it */ |
1913 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue) | |
1914 | io_iopoll_req_issued(req, issue_flags); | |
1915 | } | |
1916 | return ret; | |
2b188cc1 JA |
1917 | } |
1918 | ||
a282967c | 1919 | int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts) |
329061d3 | 1920 | { |
a282967c | 1921 | io_tw_lock(req->ctx, ts); |
9a692451 DY |
1922 | return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT| |
1923 | IO_URING_F_COMPLETE_DEFER); | |
329061d3 JA |
1924 | } |
1925 | ||
c9f06aa7 | 1926 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1927 | { |
1928 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
247f97a5 | 1929 | struct io_kiocb *nxt = NULL; |
ebc11b6c | 1930 | |
247f97a5 PB |
1931 | if (req_ref_put_and_test(req)) { |
1932 | if (req->flags & IO_REQ_LINK_FLAGS) | |
1933 | nxt = io_req_find_next(req); | |
1934 | io_free_req(req); | |
1935 | } | |
1936 | return nxt ? &nxt->work : NULL; | |
ebc11b6c PB |
1937 | } |
1938 | ||
c9f06aa7 | 1939 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1940 | { |
1941 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
a7dd2782 | 1942 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
e6aeb272 | 1943 | unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ; |
d01905db | 1944 | bool needs_poll = false; |
6bf9c47a | 1945 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1946 | |
23a6c9ac | 1947 | /* one will be dropped by ->io_wq_free_work() after returning to io-wq */ |
48dcd38d PB |
1948 | if (!(req->flags & REQ_F_REFCOUNT)) |
1949 | __io_req_set_refcount(req, 2); | |
1950 | else | |
1951 | req_ref_get(req); | |
5d5901a3 | 1952 | |
cb2d344c | 1953 | io_arm_ltimeout(req); |
6bf9c47a | 1954 | |
dadebc35 | 1955 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1956 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1957 | fail: |
6bf9c47a | 1958 | io_req_task_queue_fail(req, err); |
d01905db PB |
1959 | return; |
1960 | } | |
f4992544 | 1961 | if (!io_assign_file(req, def, issue_flags)) { |
0f8da75b PB |
1962 | err = -EBADF; |
1963 | work->flags |= IO_WQ_WORK_CANCEL; | |
1964 | goto fail; | |
1965 | } | |
31b51510 | 1966 | |
e0e4ab52 PB |
1967 | /* |
1968 | * If DEFER_TASKRUN is set, it's only allowed to post CQEs from the | |
1969 | * submitter task context. Final request completions are handed to the | |
1970 | * right context, however this is not the case of auxiliary CQEs, | |
1971 | * which is the main mean of operation for multishot requests. | |
1972 | * Don't allow any multishot execution from io-wq. It's more restrictive | |
1973 | * than necessary and also cleaner. | |
1974 | */ | |
1975 | if (req->flags & REQ_F_APOLL_MULTISHOT) { | |
1976 | err = -EBADFD; | |
1977 | if (!io_file_can_poll(req)) | |
1978 | goto fail; | |
bee1d5be JA |
1979 | if (req->file->f_flags & O_NONBLOCK || |
1980 | req->file->f_mode & FMODE_NOWAIT) { | |
1981 | err = -ECANCELED; | |
1982 | if (io_arm_poll_handler(req, issue_flags) != IO_APOLL_OK) | |
1983 | goto fail; | |
1984 | return; | |
1985 | } else { | |
1986 | req->flags &= ~REQ_F_APOLL_MULTISHOT; | |
1987 | } | |
e0e4ab52 PB |
1988 | } |
1989 | ||
d01905db | 1990 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
1991 | bool opcode_poll = def->pollin || def->pollout; |
1992 | ||
95041b93 | 1993 | if (opcode_poll && io_file_can_poll(req)) { |
afb7f56f | 1994 | needs_poll = true; |
d01905db | 1995 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 1996 | } |
561fb04a | 1997 | } |
31b51510 | 1998 | |
d01905db PB |
1999 | do { |
2000 | ret = io_issue_sqe(req, issue_flags); | |
2001 | if (ret != -EAGAIN) | |
2002 | break; | |
a9be2022 JA |
2003 | |
2004 | /* | |
2005 | * If REQ_F_NOWAIT is set, then don't wait or retry with | |
2006 | * poll. -EAGAIN is final for that case. | |
2007 | */ | |
2008 | if (req->flags & REQ_F_NOWAIT) | |
2009 | break; | |
2010 | ||
d01905db PB |
2011 | /* |
2012 | * We can get EAGAIN for iopolled IO even though we're | |
2013 | * forcing a sync submission from here, since we can't | |
2014 | * wait for request slots on the block side. | |
2015 | */ | |
2016 | if (!needs_poll) { | |
e0deb6a0 PB |
2017 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
2018 | break; | |
45500dc4 PB |
2019 | if (io_wq_worker_stopped()) |
2020 | break; | |
d01905db PB |
2021 | cond_resched(); |
2022 | continue; | |
90fa0288 HX |
2023 | } |
2024 | ||
4d9237e3 | 2025 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
2026 | return; |
2027 | /* aborted or ready, in either case retry blocking */ | |
2028 | needs_poll = false; | |
2029 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
2030 | } while (1); | |
31b51510 | 2031 | |
a3df7698 | 2032 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 2033 | if (ret < 0) |
a3df7698 | 2034 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
2035 | } |
2036 | ||
531113bb JA |
2037 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
2038 | unsigned int issue_flags) | |
09bb8394 | 2039 | { |
5106dd6e | 2040 | struct io_ring_ctx *ctx = req->ctx; |
4bfb0c9a | 2041 | struct io_fixed_file *slot; |
5106dd6e | 2042 | struct file *file = NULL; |
09bb8394 | 2043 | |
93f052cb | 2044 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 2045 | |
ac177053 | 2046 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 2047 | goto out; |
ac177053 | 2048 | fd = array_index_nospec(fd, ctx->nr_user_files); |
4bfb0c9a | 2049 | slot = io_fixed_file_slot(&ctx->file_table, fd); |
3f302388 JA |
2050 | if (!req->rsrc_node) |
2051 | __io_req_set_rsrc_node(req, ctx); | |
4bfb0c9a | 2052 | req->flags |= io_slot_flags(slot); |
3f302388 | 2053 | file = io_slot_file(slot); |
5106dd6e | 2054 | out: |
93f052cb | 2055 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
2056 | return file; |
2057 | } | |
d44f554e | 2058 | |
531113bb | 2059 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 2060 | { |
62906e89 | 2061 | struct file *file = fget(fd); |
ac177053 | 2062 | |
48863ffd | 2063 | trace_io_uring_file_get(req, fd); |
09bb8394 | 2064 | |
ac177053 | 2065 | /* we don't allow fixed io_uring files */ |
e5550a14 | 2066 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 2067 | io_req_track_inflight(req); |
8371adf5 | 2068 | return file; |
09bb8394 JA |
2069 | } |
2070 | ||
7bfa9bad | 2071 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
2072 | __must_hold(&req->ctx->uring_lock) |
2073 | { | |
7bfa9bad PB |
2074 | struct io_kiocb *linked_timeout; |
2075 | ||
2076 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
973fc83f | 2077 | io_req_defer_failed(req, ret); |
7bfa9bad PB |
2078 | return; |
2079 | } | |
2080 | ||
2081 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 2082 | |
4d9237e3 | 2083 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 2084 | case IO_APOLL_READY: |
336d28a8 | 2085 | io_kbuf_recycle(req, 0); |
d475a9a6 PB |
2086 | io_req_task_queue(req); |
2087 | break; | |
2088 | case IO_APOLL_ABORTED: | |
6436c770 | 2089 | io_kbuf_recycle(req, 0); |
77955efb | 2090 | io_queue_iowq(req, NULL); |
d475a9a6 | 2091 | break; |
b1c62645 | 2092 | case IO_APOLL_OK: |
b1c62645 | 2093 | break; |
d475a9a6 PB |
2094 | } |
2095 | ||
2096 | if (linked_timeout) | |
2097 | io_queue_linked_timeout(linked_timeout); | |
2098 | } | |
2099 | ||
cbc2e203 | 2100 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 2101 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 2102 | { |
e0c5c576 | 2103 | int ret; |
2b188cc1 | 2104 | |
c5eef2b9 | 2105 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 2106 | |
491381ce JA |
2107 | /* |
2108 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2109 | * doesn't support non-blocking read/write attempts | |
2110 | */ | |
e0b23d99 | 2111 | if (unlikely(ret)) |
7bfa9bad | 2112 | io_queue_async(req, ret); |
2b188cc1 JA |
2113 | } |
2114 | ||
4652fe3f | 2115 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 2116 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 2117 | { |
17b147f6 PB |
2118 | if (unlikely(req->flags & REQ_F_FAIL)) { |
2119 | /* | |
2120 | * We don't submit, fail them all, for that replace hardlinks | |
2121 | * with normal links. Extra REQ_F_LINK is tolerated. | |
2122 | */ | |
2123 | req->flags &= ~REQ_F_HARDLINK; | |
2124 | req->flags |= REQ_F_LINK; | |
973fc83f | 2125 | io_req_defer_failed(req, req->cqe.res); |
76cc33d7 PB |
2126 | } else { |
2127 | int ret = io_req_prep_async(req); | |
2128 | ||
ef5c600a | 2129 | if (unlikely(ret)) { |
973fc83f | 2130 | io_req_defer_failed(req, ret); |
ef5c600a DY |
2131 | return; |
2132 | } | |
2133 | ||
2134 | if (unlikely(req->ctx->drain_active)) | |
2135 | io_drain_req(req); | |
76cc33d7 | 2136 | else |
77955efb | 2137 | io_queue_iowq(req, NULL); |
ce35a47a | 2138 | } |
4fe2c963 JL |
2139 | } |
2140 | ||
b16fed66 PB |
2141 | /* |
2142 | * Check SQE restrictions (opcode and flags). | |
2143 | * | |
2144 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
2145 | */ | |
2146 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
2147 | struct io_kiocb *req, | |
2148 | unsigned int sqe_flags) | |
4fe2c963 | 2149 | { |
b16fed66 PB |
2150 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
2151 | return false; | |
2152 | ||
2153 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
2154 | ctx->restrictions.sqe_flags_required) | |
2155 | return false; | |
2156 | ||
2157 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
2158 | ctx->restrictions.sqe_flags_required)) | |
2159 | return false; | |
2160 | ||
2161 | return true; | |
4fe2c963 JL |
2162 | } |
2163 | ||
22b2ca31 PB |
2164 | static void io_init_req_drain(struct io_kiocb *req) |
2165 | { | |
2166 | struct io_ring_ctx *ctx = req->ctx; | |
2167 | struct io_kiocb *head = ctx->submit_state.link.head; | |
2168 | ||
2169 | ctx->drain_active = true; | |
2170 | if (head) { | |
2171 | /* | |
2172 | * If we need to drain a request in the middle of a link, drain | |
2173 | * the head request and the next request/link after the current | |
2174 | * link. Considering sequential execution of links, | |
b6c7db32 | 2175 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
2176 | * link. |
2177 | */ | |
b6c7db32 | 2178 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
2179 | ctx->drain_next = true; |
2180 | } | |
2181 | } | |
2182 | ||
e21e1c45 JA |
2183 | static __cold int io_init_fail_req(struct io_kiocb *req, int err) |
2184 | { | |
2185 | /* ensure per-opcode data is cleared if we fail before prep */ | |
2186 | memset(&req->cmd.data, 0, sizeof(req->cmd.data)); | |
2187 | return err; | |
2188 | } | |
2189 | ||
b16fed66 PB |
2190 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2191 | const struct io_uring_sqe *sqe) | |
282cdc86 | 2192 | __must_hold(&ctx->uring_lock) |
b16fed66 | 2193 | { |
a7dd2782 | 2194 | const struct io_issue_def *def; |
b16fed66 | 2195 | unsigned int sqe_flags; |
fc0ae024 | 2196 | int personality; |
4a04d1d1 | 2197 | u8 opcode; |
b16fed66 | 2198 | |
864ea921 | 2199 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 2200 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 | 2201 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
4bcb982c JA |
2202 | sqe_flags = READ_ONCE(sqe->flags); |
2203 | req->flags = (io_req_flags_t) sqe_flags; | |
cef216fc | 2204 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 2205 | req->file = NULL; |
c1bdf8ed | 2206 | req->rsrc_node = NULL; |
b16fed66 | 2207 | req->task = current; |
b16fed66 | 2208 | |
4a04d1d1 PB |
2209 | if (unlikely(opcode >= IORING_OP_LAST)) { |
2210 | req->opcode = 0; | |
e21e1c45 | 2211 | return io_init_fail_req(req, -EINVAL); |
4a04d1d1 | 2212 | } |
a7dd2782 | 2213 | def = &io_issue_defs[opcode]; |
68fe256a PB |
2214 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
2215 | /* enforce forwards compatibility on users */ | |
2216 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
e21e1c45 | 2217 | return io_init_fail_req(req, -EINVAL); |
4e906702 | 2218 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 2219 | if (!def->buffer_select) |
e21e1c45 | 2220 | return io_init_fail_req(req, -EOPNOTSUPP); |
4e906702 JA |
2221 | req->buf_index = READ_ONCE(sqe->buf_group); |
2222 | } | |
5562a8d7 PB |
2223 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
2224 | ctx->drain_disabled = true; | |
2225 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
2226 | if (ctx->drain_disabled) | |
e21e1c45 | 2227 | return io_init_fail_req(req, -EOPNOTSUPP); |
22b2ca31 | 2228 | io_init_req_drain(req); |
5562a8d7 | 2229 | } |
2a56a9bd PB |
2230 | } |
2231 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
2232 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
e21e1c45 | 2233 | return io_init_fail_req(req, -EACCES); |
2a56a9bd PB |
2234 | /* knock it to the slow queue path, will be drained there */ |
2235 | if (ctx->drain_active) | |
2236 | req->flags |= REQ_F_FORCE_ASYNC; | |
2237 | /* if there is no link, we're at "next" request and need to drain */ | |
2238 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
2239 | ctx->drain_next = false; | |
2240 | ctx->drain_active = true; | |
b6c7db32 | 2241 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 2242 | } |
68fe256a | 2243 | } |
b16fed66 | 2244 | |
fcde59fe | 2245 | if (!def->ioprio && sqe->ioprio) |
e21e1c45 | 2246 | return io_init_fail_req(req, -EINVAL); |
fcde59fe | 2247 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
e21e1c45 | 2248 | return io_init_fail_req(req, -EINVAL); |
73911426 | 2249 | |
fcde59fe | 2250 | if (def->needs_file) { |
6d63416d PB |
2251 | struct io_submit_state *state = &ctx->submit_state; |
2252 | ||
cef216fc | 2253 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 2254 | |
6d63416d PB |
2255 | /* |
2256 | * Plug now if we have more than 2 IO left after this, and the | |
2257 | * target is potentially a read/write to block based storage. | |
2258 | */ | |
fcde59fe | 2259 | if (state->need_plug && def->plug) { |
6d63416d PB |
2260 | state->plug_started = true; |
2261 | state->need_plug = false; | |
5ca7a8b3 | 2262 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 2263 | } |
b16fed66 | 2264 | } |
863e0560 | 2265 | |
003e8dcc JA |
2266 | personality = READ_ONCE(sqe->personality); |
2267 | if (personality) { | |
cdab10bf LT |
2268 | int ret; |
2269 | ||
c10d1f98 PB |
2270 | req->creds = xa_load(&ctx->personalities, personality); |
2271 | if (!req->creds) | |
e21e1c45 | 2272 | return io_init_fail_req(req, -EINVAL); |
c10d1f98 | 2273 | get_cred(req->creds); |
cdc1404a PM |
2274 | ret = security_uring_override_creds(req->creds); |
2275 | if (ret) { | |
2276 | put_cred(req->creds); | |
e21e1c45 | 2277 | return io_init_fail_req(req, ret); |
cdc1404a | 2278 | } |
b8e64b53 | 2279 | req->flags |= REQ_F_CREDS; |
003e8dcc | 2280 | } |
b16fed66 | 2281 | |
0702e536 | 2282 | return def->prep(req, sqe); |
b16fed66 PB |
2283 | } |
2284 | ||
df3becde PB |
2285 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
2286 | struct io_kiocb *req, int ret) | |
2287 | { | |
2288 | struct io_ring_ctx *ctx = req->ctx; | |
2289 | struct io_submit_link *link = &ctx->submit_state.link; | |
2290 | struct io_kiocb *head = link->head; | |
2291 | ||
48863ffd | 2292 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
2293 | |
2294 | /* | |
2295 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
2296 | * unusable. Instead of failing eagerly, continue assembling the link if | |
2297 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
2298 | * should find the flag and handle the rest. | |
2299 | */ | |
2300 | req_fail_link_node(req, ret); | |
2301 | if (head && !(head->flags & REQ_F_FAIL)) | |
2302 | req_fail_link_node(head, -ECANCELED); | |
2303 | ||
2304 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
2305 | if (head) { | |
2306 | link->last->link = req; | |
2307 | link->head = NULL; | |
2308 | req = head; | |
2309 | } | |
2310 | io_queue_sqe_fallback(req); | |
2311 | return ret; | |
2312 | } | |
2313 | ||
2314 | if (head) | |
2315 | link->last->link = req; | |
2316 | else | |
2317 | link->head = req; | |
2318 | link->last = req; | |
2319 | return 0; | |
2320 | } | |
2321 | ||
2322 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 2323 | const struct io_uring_sqe *sqe) |
282cdc86 | 2324 | __must_hold(&ctx->uring_lock) |
9e645e11 | 2325 | { |
a1ab7b35 | 2326 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 2327 | int ret; |
9e645e11 | 2328 | |
a6b8cadc | 2329 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
2330 | if (unlikely(ret)) |
2331 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 2332 | |
2ad57931 | 2333 | trace_io_uring_submit_req(req); |
a6b8cadc | 2334 | |
9e645e11 JA |
2335 | /* |
2336 | * If we already have a head request, queue this one for async | |
2337 | * submittal once the head completes. If we don't have a head but | |
2338 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2339 | * submitted sync once the chain is complete. If none of those | |
2340 | * conditions are true (normal request), then just queue it. | |
2341 | */ | |
924a07e4 | 2342 | if (unlikely(link->head)) { |
df3becde PB |
2343 | ret = io_req_prep_async(req); |
2344 | if (unlikely(ret)) | |
2345 | return io_submit_fail_init(sqe, req, ret); | |
2346 | ||
48863ffd | 2347 | trace_io_uring_link(req, link->head); |
f2f87370 | 2348 | link->last->link = req; |
863e0560 | 2349 | link->last = req; |
32fe525b | 2350 | |
da1a08c5 | 2351 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 2352 | return 0; |
df3becde PB |
2353 | /* last request of the link, flush it */ |
2354 | req = link->head; | |
f15a3431 | 2355 | link->head = NULL; |
924a07e4 PB |
2356 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
2357 | goto fallback; | |
2358 | ||
2359 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2360 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2361 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2362 | link->head = req; | |
2363 | link->last = req; | |
2364 | } else { | |
2365 | fallback: | |
2366 | io_queue_sqe_fallback(req); | |
2367 | } | |
f15a3431 | 2368 | return 0; |
9e645e11 | 2369 | } |
2e6e1fde | 2370 | |
f15a3431 | 2371 | io_queue_sqe(req); |
1d4240cc | 2372 | return 0; |
9e645e11 JA |
2373 | } |
2374 | ||
9a56a232 JA |
2375 | /* |
2376 | * Batched submission is done, ensure local IO is flushed out. | |
2377 | */ | |
553deffd | 2378 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2379 | { |
553deffd PB |
2380 | struct io_submit_state *state = &ctx->submit_state; |
2381 | ||
e126391c PB |
2382 | if (unlikely(state->link.head)) |
2383 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2384 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2385 | io_submit_flush_completions(ctx); |
27926b68 JA |
2386 | if (state->plug_started) |
2387 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2388 | } |
2389 | ||
2390 | /* | |
2391 | * Start submission side cache. | |
2392 | */ | |
2393 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2394 | unsigned int max_ios) |
9a56a232 | 2395 | { |
27926b68 | 2396 | state->plug_started = false; |
4b628aeb | 2397 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2398 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2399 | /* set only head, no need to init link_last in advance */ |
2400 | state->link.head = NULL; | |
9a56a232 JA |
2401 | } |
2402 | ||
2b188cc1 JA |
2403 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2404 | { | |
75b28aff | 2405 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2406 | |
caf582c6 PB |
2407 | /* |
2408 | * Ensure any loads from the SQEs are done at this point, | |
2409 | * since once we write the new head, the application could | |
2410 | * write new data to them. | |
2411 | */ | |
2412 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2413 | } |
2414 | ||
2b188cc1 | 2415 | /* |
dd9ae8a0 | 2416 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2417 | * that is mapped by userspace. This means that care needs to be taken to |
2418 | * ensure that reads are stable, as we cannot rely on userspace always | |
2419 | * being a good citizen. If members of the sqe are validated and then later | |
2420 | * used, it's important that those reads are done through READ_ONCE() to | |
2421 | * prevent a re-load down the line. | |
2422 | */ | |
b5083dfa | 2423 | static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe) |
2b188cc1 | 2424 | { |
2af89abd PB |
2425 | unsigned mask = ctx->sq_entries - 1; |
2426 | unsigned head = ctx->cached_sq_head++ & mask; | |
2427 | ||
2428 | if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) { | |
2429 | head = READ_ONCE(ctx->sq_array[head]); | |
2430 | if (unlikely(head >= ctx->sq_entries)) { | |
2431 | /* drop invalid entries */ | |
2432 | spin_lock(&ctx->completion_lock); | |
2433 | ctx->cq_extra--; | |
2434 | spin_unlock(&ctx->completion_lock); | |
2435 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2436 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
2437 | return false; | |
2438 | } | |
2439 | } | |
2b188cc1 JA |
2440 | |
2441 | /* | |
2442 | * The cached sq head (or cq tail) serves two purposes: | |
2443 | * | |
2444 | * 1) allows us to batch the cost of updating the user visible | |
2445 | * head updates. | |
2446 | * 2) allows the kernel side to track the head on its own, even | |
2447 | * though the application is the one updating it. | |
2448 | */ | |
2b188cc1 | 2449 | |
2af89abd PB |
2450 | /* double index for 128-byte SQEs, twice as long */ |
2451 | if (ctx->flags & IORING_SETUP_SQE128) | |
2452 | head <<= 1; | |
2453 | *sqe = &ctx->sq_sqes[head]; | |
2454 | return true; | |
709b302f PB |
2455 | } |
2456 | ||
17437f31 | 2457 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2458 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2459 | { |
69629809 | 2460 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2461 | unsigned int left; |
2462 | int ret; | |
6c271ce2 | 2463 | |
51d48dab | 2464 | if (unlikely(!entries)) |
69629809 | 2465 | return 0; |
ee7d46d9 | 2466 | /* make sure SQ entry isn't read before tail */ |
e3ef728f | 2467 | ret = left = min(nr, entries); |
8e6971a8 PB |
2468 | io_get_task_refs(left); |
2469 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2470 | |
69629809 | 2471 | do { |
3529d8c2 | 2472 | const struct io_uring_sqe *sqe; |
196be95c | 2473 | struct io_kiocb *req; |
fb5ccc98 | 2474 | |
c8576f3e | 2475 | if (unlikely(!io_alloc_req(ctx, &req))) |
fb5ccc98 | 2476 | break; |
b5083dfa | 2477 | if (unlikely(!io_get_sqe(ctx, &sqe))) { |
fa05457a | 2478 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2479 | break; |
2480 | } | |
6c271ce2 | 2481 | |
1cd15904 PB |
2482 | /* |
2483 | * Continue submitting even for sqe failure if the | |
2484 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2485 | */ | |
2486 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2487 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2488 | left--; | |
2489 | break; | |
bcbb7bf6 | 2490 | } |
1cd15904 | 2491 | } while (--left); |
9466f437 | 2492 | |
8e6971a8 PB |
2493 | if (unlikely(left)) { |
2494 | ret -= left; | |
2495 | /* try again if it submitted nothing and can't allocate a req */ | |
2496 | if (!ret && io_req_cache_empty(ctx)) | |
2497 | ret = -EAGAIN; | |
2498 | current->io_uring->cached_refs += left; | |
9466f437 | 2499 | } |
6c271ce2 | 2500 | |
553deffd | 2501 | io_submit_state_end(ctx); |
ae9428ca PB |
2502 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2503 | io_commit_sqring(ctx); | |
8e6971a8 | 2504 | return ret; |
6c271ce2 JA |
2505 | } |
2506 | ||
bda52162 JA |
2507 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, |
2508 | int wake_flags, void *key) | |
2509 | { | |
bd550173 | 2510 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq); |
bda52162 | 2511 | |
6c503150 PB |
2512 | /* |
2513 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2514 | * the task, and the next invocation will do it. | |
2515 | */ | |
bd550173 | 2516 | if (io_should_wake(iowq) || io_has_work(iowq->ctx)) |
6c503150 PB |
2517 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2518 | return -1; | |
bda52162 JA |
2519 | } |
2520 | ||
c0e0d6ba | 2521 | int io_run_task_work_sig(struct io_ring_ctx *ctx) |
af9c1a44 | 2522 | { |
1414d629 | 2523 | if (!llist_empty(&ctx->work_llist)) { |
2f413956 | 2524 | __set_current_state(TASK_RUNNING); |
9fe3eaea | 2525 | if (io_run_local_work(ctx, INT_MAX) > 0) |
d246c759 | 2526 | return 0; |
1414d629 PB |
2527 | } |
2528 | if (io_run_task_work() > 0) | |
d246c759 | 2529 | return 0; |
c5020bc8 OL |
2530 | if (task_sigpending(current)) |
2531 | return -EINTR; | |
2532 | return 0; | |
af9c1a44 JA |
2533 | } |
2534 | ||
7b72d661 JA |
2535 | static bool current_pending_io(void) |
2536 | { | |
2537 | struct io_uring_task *tctx = current->io_uring; | |
2538 | ||
2539 | if (!tctx) | |
2540 | return false; | |
2541 | return percpu_counter_read_positive(&tctx->inflight); | |
2542 | } | |
2543 | ||
eeb60b9a PB |
2544 | /* when returns >0, the caller should retry */ |
2545 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
d33a39e5 | 2546 | struct io_wait_queue *iowq) |
eeb60b9a | 2547 | { |
6f0974ec | 2548 | int ret; |
8a796565 | 2549 | |
3fcf19d5 PB |
2550 | if (unlikely(READ_ONCE(ctx->check_cq))) |
2551 | return 1; | |
846072f1 PB |
2552 | if (unlikely(!llist_empty(&ctx->work_llist))) |
2553 | return 1; | |
2554 | if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) | |
2555 | return 1; | |
2556 | if (unlikely(task_sigpending(current))) | |
2557 | return -EINTR; | |
2558 | if (unlikely(io_should_wake(iowq))) | |
2559 | return 0; | |
8a796565 AF |
2560 | |
2561 | /* | |
7b72d661 JA |
2562 | * Mark us as being in io_wait if we have pending requests, so cpufreq |
2563 | * can take into account that the task is waiting for IO - turns out | |
2564 | * to be important for low QD IO. | |
8a796565 | 2565 | */ |
7b72d661 JA |
2566 | if (current_pending_io()) |
2567 | current->in_iowait = 1; | |
8a796565 | 2568 | ret = 0; |
d33a39e5 | 2569 | if (iowq->timeout == KTIME_MAX) |
46ae7eef | 2570 | schedule(); |
d33a39e5 | 2571 | else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS)) |
8a796565 | 2572 | ret = -ETIME; |
6f0974ec | 2573 | current->in_iowait = 0; |
8a796565 | 2574 | return ret; |
eeb60b9a PB |
2575 | } |
2576 | ||
2b188cc1 JA |
2577 | /* |
2578 | * Wait until events become available, if we don't already have some. The | |
2579 | * application must reap them itself, as they reside on the shared cq ring. | |
2580 | */ | |
2581 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2582 | const sigset_t __user *sig, size_t sigsz, |
2583 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2584 | { |
90291099 | 2585 | struct io_wait_queue iowq; |
75b28aff | 2586 | struct io_rings *rings = ctx->rings; |
c1d5a224 | 2587 | int ret; |
2b188cc1 | 2588 | |
76de6749 PB |
2589 | if (!io_allowed_run_tw(ctx)) |
2590 | return -EEXIST; | |
140102ae | 2591 | if (!llist_empty(&ctx->work_llist)) |
9fe3eaea | 2592 | io_run_local_work(ctx, min_events); |
f36ba6cf PB |
2593 | io_run_task_work(); |
2594 | io_cqring_overflow_flush(ctx); | |
2595 | /* if user messes with these they will just get an early return */ | |
2596 | if (__io_cqring_events_user(ctx) >= min_events) | |
2597 | return 0; | |
2b188cc1 | 2598 | |
90291099 PB |
2599 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
2600 | iowq.wq.private = current; | |
2601 | INIT_LIST_HEAD(&iowq.wq.entry); | |
2602 | iowq.ctx = ctx; | |
bda52162 | 2603 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 2604 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
d33a39e5 PB |
2605 | iowq.timeout = KTIME_MAX; |
2606 | ||
2607 | if (uts) { | |
2608 | struct timespec64 ts; | |
2609 | ||
2610 | if (get_timespec64(&ts, uts)) | |
2611 | return -EFAULT; | |
8d0c12a8 | 2612 | |
d33a39e5 | 2613 | iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); |
8d0c12a8 | 2614 | io_napi_adjust_timeout(ctx, &iowq, &ts); |
d33a39e5 | 2615 | } |
90291099 | 2616 | |
978e5c19 AI |
2617 | if (sig) { |
2618 | #ifdef CONFIG_COMPAT | |
2619 | if (in_compat_syscall()) | |
2620 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
2621 | sigsz); | |
2622 | else | |
2623 | #endif | |
2624 | ret = set_user_sigmask(sig, sigsz); | |
2625 | ||
2626 | if (ret) | |
2627 | return ret; | |
2628 | } | |
2629 | ||
8d0c12a8 SR |
2630 | io_napi_busy_loop(ctx, &iowq); |
2631 | ||
c826bd7a | 2632 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2633 | do { |
9fe3eaea | 2634 | int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail); |
3fcf19d5 PB |
2635 | unsigned long check_cq; |
2636 | ||
130bd686 | 2637 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
8751d154 | 2638 | atomic_set(&ctx->cq_wait_nr, nr_wait); |
130bd686 PB |
2639 | set_current_state(TASK_INTERRUPTIBLE); |
2640 | } else { | |
2641 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, | |
2642 | TASK_INTERRUPTIBLE); | |
2643 | } | |
2644 | ||
d33a39e5 | 2645 | ret = io_cqring_wait_schedule(ctx, &iowq); |
130bd686 | 2646 | __set_current_state(TASK_RUNNING); |
b4bc35cf | 2647 | atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT); |
d80c0f00 | 2648 | |
846072f1 PB |
2649 | /* |
2650 | * Run task_work after scheduling and before io_should_wake(). | |
2651 | * If we got woken because of task_work being processed, run it | |
2652 | * now rather than let the caller do another wait loop. | |
2653 | */ | |
2654 | io_run_task_work(); | |
2655 | if (!llist_empty(&ctx->work_llist)) | |
9fe3eaea | 2656 | io_run_local_work(ctx, nr_wait); |
3fcf19d5 | 2657 | |
6ff1407e JA |
2658 | /* |
2659 | * Non-local task_work will be run on exit to userspace, but | |
2660 | * if we're using DEFER_TASKRUN, then we could have waited | |
2661 | * with a timeout for a number of requests. If the timeout | |
2662 | * hits, we could have some requests ready to process. Ensure | |
2663 | * this break is _after_ we have run task_work, to avoid | |
2664 | * deferring running potentially pending requests until the | |
2665 | * next time we wait for events. | |
2666 | */ | |
2667 | if (ret < 0) | |
2668 | break; | |
2669 | ||
3fcf19d5 PB |
2670 | check_cq = READ_ONCE(ctx->check_cq); |
2671 | if (unlikely(check_cq)) { | |
2672 | /* let the caller flush overflows, retry */ | |
326a9e48 | 2673 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) |
3fcf19d5 | 2674 | io_cqring_do_overflow_flush(ctx); |
3fcf19d5 PB |
2675 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) { |
2676 | ret = -EBADR; | |
2677 | break; | |
2678 | } | |
2679 | } | |
2680 | ||
846072f1 PB |
2681 | if (io_should_wake(&iowq)) { |
2682 | ret = 0; | |
35d90f95 | 2683 | break; |
846072f1 | 2684 | } |
ca0a2651 | 2685 | cond_resched(); |
846072f1 | 2686 | } while (1); |
bda52162 | 2687 | |
130bd686 PB |
2688 | if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) |
2689 | finish_wait(&ctx->cq_wait, &iowq.wq); | |
b7db41c9 | 2690 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2691 | |
75b28aff | 2692 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2693 | } |
2694 | ||
edecf168 | 2695 | void io_mem_free(void *ptr) |
b9bd2bea | 2696 | { |
73572984 JA |
2697 | if (!ptr) |
2698 | return; | |
b9bd2bea | 2699 | |
99a9e0b8 | 2700 | folio_put(virt_to_folio(ptr)); |
b9bd2bea PB |
2701 | } |
2702 | ||
03d89a2d JA |
2703 | static void io_pages_free(struct page ***pages, int npages) |
2704 | { | |
2c5c0ba1 | 2705 | struct page **page_array = *pages; |
03d89a2d JA |
2706 | int i; |
2707 | ||
8b51a395 JA |
2708 | if (!page_array) |
2709 | return; | |
2710 | ||
03d89a2d JA |
2711 | for (i = 0; i < npages; i++) |
2712 | unpin_user_page(page_array[i]); | |
2713 | kvfree(page_array); | |
2714 | *pages = NULL; | |
2715 | } | |
2716 | ||
2717 | static void *__io_uaddr_map(struct page ***pages, unsigned short *npages, | |
2718 | unsigned long uaddr, size_t size) | |
2719 | { | |
2720 | struct page **page_array; | |
2721 | unsigned int nr_pages; | |
820d070f | 2722 | void *page_addr; |
67d1189d | 2723 | int ret, i, pinned; |
03d89a2d JA |
2724 | |
2725 | *npages = 0; | |
2726 | ||
2727 | if (uaddr & (PAGE_SIZE - 1) || !size) | |
2728 | return ERR_PTR(-EINVAL); | |
2729 | ||
2730 | nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2731 | if (nr_pages > USHRT_MAX) | |
2732 | return ERR_PTR(-EINVAL); | |
2733 | page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); | |
2734 | if (!page_array) | |
2735 | return ERR_PTR(-ENOMEM); | |
2736 | ||
67d1189d GKB |
2737 | |
2738 | pinned = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM, | |
2739 | page_array); | |
2740 | if (pinned != nr_pages) { | |
2741 | ret = (pinned < 0) ? pinned : -EFAULT; | |
2742 | goto free_pages; | |
03d89a2d | 2743 | } |
223ef474 | 2744 | |
820d070f | 2745 | page_addr = page_address(page_array[0]); |
223ef474 | 2746 | for (i = 0; i < nr_pages; i++) { |
820d070f JA |
2747 | ret = -EINVAL; |
2748 | ||
2749 | /* | |
2750 | * Can't support mapping user allocated ring memory on 32-bit | |
2751 | * archs where it could potentially reside in highmem. Just | |
2752 | * fail those with -EINVAL, just like we did on kernels that | |
2753 | * didn't support this feature. | |
2754 | */ | |
2755 | if (PageHighMem(page_array[i])) | |
67d1189d | 2756 | goto free_pages; |
820d070f JA |
2757 | |
2758 | /* | |
2759 | * No support for discontig pages for now, should either be a | |
2760 | * single normal page, or a huge page. Later on we can add | |
2761 | * support for remapping discontig pages, for now we will | |
2762 | * just fail them with EINVAL. | |
2763 | */ | |
2764 | if (page_address(page_array[i]) != page_addr) | |
67d1189d | 2765 | goto free_pages; |
820d070f | 2766 | page_addr += PAGE_SIZE; |
223ef474 JA |
2767 | } |
2768 | ||
03d89a2d JA |
2769 | *pages = page_array; |
2770 | *npages = nr_pages; | |
2771 | return page_to_virt(page_array[0]); | |
67d1189d GKB |
2772 | |
2773 | free_pages: | |
2774 | io_pages_free(&page_array, pinned > 0 ? pinned : 0); | |
2775 | return ERR_PTR(ret); | |
03d89a2d JA |
2776 | } |
2777 | ||
2778 | static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr, | |
2779 | size_t size) | |
2780 | { | |
2781 | return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr, | |
2782 | size); | |
2783 | } | |
2784 | ||
2785 | static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr, | |
2786 | size_t size) | |
2787 | { | |
2788 | return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr, | |
2789 | size); | |
2790 | } | |
2791 | ||
9c189eee JA |
2792 | static void io_rings_free(struct io_ring_ctx *ctx) |
2793 | { | |
03d89a2d JA |
2794 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) { |
2795 | io_mem_free(ctx->rings); | |
2796 | io_mem_free(ctx->sq_sqes); | |
03d89a2d JA |
2797 | } else { |
2798 | io_pages_free(&ctx->ring_pages, ctx->n_ring_pages); | |
8b51a395 | 2799 | ctx->n_ring_pages = 0; |
03d89a2d | 2800 | io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages); |
8b51a395 | 2801 | ctx->n_sqe_pages = 0; |
03d89a2d | 2802 | } |
cef59d1e PB |
2803 | |
2804 | ctx->rings = NULL; | |
2805 | ctx->sq_sqes = NULL; | |
9c189eee JA |
2806 | } |
2807 | ||
edecf168 | 2808 | void *io_mem_alloc(size_t size) |
b9bd2bea | 2809 | { |
73572984 | 2810 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; |
e27cef86 | 2811 | void *ret; |
b9bd2bea | 2812 | |
e27cef86 JA |
2813 | ret = (void *) __get_free_pages(gfp, get_order(size)); |
2814 | if (ret) | |
2815 | return ret; | |
2816 | return ERR_PTR(-ENOMEM); | |
b9bd2bea PB |
2817 | } |
2818 | ||
73572984 JA |
2819 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2820 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2821 | { |
73572984 JA |
2822 | struct io_rings *rings; |
2823 | size_t off, sq_array_size; | |
6b06314c | 2824 | |
73572984 JA |
2825 | off = struct_size(rings, cqes, cq_entries); |
2826 | if (off == SIZE_MAX) | |
2827 | return SIZE_MAX; | |
2828 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2829 | if (check_shl_overflow(off, 1, &off)) | |
2830 | return SIZE_MAX; | |
2831 | } | |
ab409402 | 2832 | |
73572984 JA |
2833 | #ifdef CONFIG_SMP |
2834 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2835 | if (off == 0) | |
2836 | return SIZE_MAX; | |
2837 | #endif | |
82fbcfa9 | 2838 | |
2af89abd PB |
2839 | if (ctx->flags & IORING_SETUP_NO_SQARRAY) { |
2840 | if (sq_offset) | |
2841 | *sq_offset = SIZE_MAX; | |
2842 | return off; | |
2843 | } | |
2844 | ||
73572984 JA |
2845 | if (sq_offset) |
2846 | *sq_offset = off; | |
82fbcfa9 | 2847 | |
73572984 JA |
2848 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2849 | if (sq_array_size == SIZE_MAX) | |
2850 | return SIZE_MAX; | |
6b06314c | 2851 | |
73572984 JA |
2852 | if (check_add_overflow(off, sq_array_size, &off)) |
2853 | return SIZE_MAX; | |
8bad28d8 | 2854 | |
73572984 | 2855 | return off; |
8bad28d8 HX |
2856 | } |
2857 | ||
73572984 | 2858 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2859 | { |
c8576f3e | 2860 | struct io_kiocb *req; |
37f0e767 | 2861 | int nr = 0; |
bf019da7 | 2862 | |
9a4fdbd8 | 2863 | mutex_lock(&ctx->uring_lock); |
34f0bc42 | 2864 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
9a4fdbd8 | 2865 | |
88ab95be | 2866 | while (!io_req_cache_empty(ctx)) { |
c8576f3e | 2867 | req = io_extract_req(ctx); |
c2b6c6bc | 2868 | kmem_cache_free(req_cachep, req); |
37f0e767 | 2869 | nr++; |
c2b6c6bc | 2870 | } |
37f0e767 PB |
2871 | if (nr) |
2872 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2873 | mutex_unlock(&ctx->uring_lock); |
2874 | } | |
2875 | ||
9eae8655 PB |
2876 | static void io_rsrc_node_cache_free(struct io_cache_entry *entry) |
2877 | { | |
2878 | kfree(container_of(entry, struct io_rsrc_node, cache)); | |
2879 | } | |
2880 | ||
c072481d | 2881 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2882 | { |
37d1e2e3 | 2883 | io_sq_thread_finish(ctx); |
43597aac | 2884 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
0b222eeb PB |
2885 | if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list))) |
2886 | return; | |
43597aac | 2887 | |
8bad28d8 | 2888 | mutex_lock(&ctx->uring_lock); |
43597aac | 2889 | if (ctx->buf_data) |
bd54b6fe | 2890 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2891 | if (ctx->file_data) |
08480400 | 2892 | __io_sqe_files_unregister(ctx); |
a85381d8 | 2893 | io_cqring_overflow_kill(ctx); |
9b402849 | 2894 | io_eventfd_unregister(ctx); |
9b797a37 | 2895 | io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); |
43e0bbbd | 2896 | io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); |
194bb58c | 2897 | io_futex_cache_free(ctx); |
5a2e745d | 2898 | io_destroy_buffers(ctx); |
b4a72c05 | 2899 | mutex_unlock(&ctx->uring_lock); |
07db298a PB |
2900 | if (ctx->sq_creds) |
2901 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2902 | if (ctx->submitter_task) |
2903 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2904 | |
a7f0ed5a PB |
2905 | /* there are no registered resources left, nobody uses it */ |
2906 | if (ctx->rsrc_node) | |
9eae8655 | 2907 | io_rsrc_node_destroy(ctx, ctx->rsrc_node); |
a7f0ed5a PB |
2908 | |
2909 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
ef9dd637 | 2910 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
2b188cc1 | 2911 | |
9eae8655 | 2912 | io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free); |
42b6419d PB |
2913 | if (ctx->mm_account) { |
2914 | mmdrop(ctx->mm_account); | |
2915 | ctx->mm_account = NULL; | |
2916 | } | |
9c189eee | 2917 | io_rings_free(ctx); |
c392cbec | 2918 | io_kbuf_mmap_list_free(ctx); |
2b188cc1 JA |
2919 | |
2920 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2921 | free_uid(ctx->user); |
4010fec4 | 2922 | io_req_caches_free(ctx); |
e941894e JA |
2923 | if (ctx->hash_map) |
2924 | io_wq_put_hash(ctx->hash_map); | |
8d0c12a8 | 2925 | io_napi_free(ctx); |
e6f89be6 | 2926 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2927 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 | 2928 | xa_destroy(&ctx->io_bl_xa); |
2b188cc1 JA |
2929 | kfree(ctx); |
2930 | } | |
2931 | ||
bca39f39 PB |
2932 | static __cold void io_activate_pollwq_cb(struct callback_head *cb) |
2933 | { | |
2934 | struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx, | |
2935 | poll_wq_task_work); | |
2936 | ||
2937 | mutex_lock(&ctx->uring_lock); | |
2938 | ctx->poll_activated = true; | |
2939 | mutex_unlock(&ctx->uring_lock); | |
2940 | ||
2941 | /* | |
2942 | * Wake ups for some events between start of polling and activation | |
2943 | * might've been lost due to loose synchronisation. | |
2944 | */ | |
2945 | wake_up_all(&ctx->poll_wq); | |
2946 | percpu_ref_put(&ctx->refs); | |
2947 | } | |
2948 | ||
c4320315 | 2949 | __cold void io_activate_pollwq(struct io_ring_ctx *ctx) |
bca39f39 PB |
2950 | { |
2951 | spin_lock(&ctx->completion_lock); | |
2952 | /* already activated or in progress */ | |
2953 | if (ctx->poll_activated || ctx->poll_wq_task_work.func) | |
2954 | goto out; | |
2955 | if (WARN_ON_ONCE(!ctx->task_complete)) | |
2956 | goto out; | |
2957 | if (!ctx->submitter_task) | |
2958 | goto out; | |
2959 | /* | |
2960 | * with ->submitter_task only the submitter task completes requests, we | |
2961 | * only need to sync with it, which is done by injecting a tw | |
2962 | */ | |
2963 | init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb); | |
2964 | percpu_ref_get(&ctx->refs); | |
2965 | if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL)) | |
2966 | percpu_ref_put(&ctx->refs); | |
2967 | out: | |
2968 | spin_unlock(&ctx->completion_lock); | |
2969 | } | |
2970 | ||
2b188cc1 JA |
2971 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) |
2972 | { | |
2973 | struct io_ring_ctx *ctx = file->private_data; | |
2974 | __poll_t mask = 0; | |
2975 | ||
bca39f39 PB |
2976 | if (unlikely(!ctx->poll_activated)) |
2977 | io_activate_pollwq(ctx); | |
2978 | ||
7b235dd8 | 2979 | poll_wait(file, &ctx->poll_wq, wait); |
4f7067c3 SB |
2980 | /* |
2981 | * synchronizes with barrier from wq_has_sleeper call in | |
2982 | * io_commit_cqring | |
2983 | */ | |
2b188cc1 | 2984 | smp_rmb(); |
90554200 | 2985 | if (!io_sqring_full(ctx)) |
2b188cc1 | 2986 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
2987 | |
2988 | /* | |
2989 | * Don't flush cqring overflow list here, just do a simple check. | |
2990 | * Otherwise there could possible be ABBA deadlock: | |
2991 | * CPU0 CPU1 | |
2992 | * ---- ---- | |
2993 | * lock(&ctx->uring_lock); | |
2994 | * lock(&ep->mtx); | |
2995 | * lock(&ctx->uring_lock); | |
2996 | * lock(&ep->mtx); | |
2997 | * | |
2998 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
10d8bc35 | 2999 | * pushes them to do the flush. |
ed670c3f | 3000 | */ |
b4c98d59 | 3001 | |
c10bb646 | 3002 | if (__io_cqring_events_user(ctx) || io_has_work(ctx)) |
2b188cc1 JA |
3003 | mask |= EPOLLIN | EPOLLRDNORM; |
3004 | ||
3005 | return mask; | |
3006 | } | |
3007 | ||
d56d938b PB |
3008 | struct io_tctx_exit { |
3009 | struct callback_head task_work; | |
3010 | struct completion completion; | |
baf186c4 | 3011 | struct io_ring_ctx *ctx; |
d56d938b PB |
3012 | }; |
3013 | ||
c072481d | 3014 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
3015 | { |
3016 | struct io_uring_task *tctx = current->io_uring; | |
3017 | struct io_tctx_exit *work; | |
3018 | ||
3019 | work = container_of(cb, struct io_tctx_exit, task_work); | |
3020 | /* | |
8d664282 | 3021 | * When @in_cancel, we're in cancellation and it's racy to remove the |
d56d938b | 3022 | * node. It'll be removed by the end of cancellation, just ignore it. |
998b30c3 HM |
3023 | * tctx can be NULL if the queueing of this task_work raced with |
3024 | * work cancelation off the exec path. | |
d56d938b | 3025 | */ |
8d664282 | 3026 | if (tctx && !atomic_read(&tctx->in_cancel)) |
eef51daa | 3027 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
3028 | complete(&work->completion); |
3029 | } | |
3030 | ||
c072481d | 3031 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
3032 | { |
3033 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
3034 | ||
3035 | return req->ctx == data; | |
3036 | } | |
3037 | ||
c072481d | 3038 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 3039 | { |
d56d938b | 3040 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 3041 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 3042 | unsigned long interval = HZ / 20; |
d56d938b PB |
3043 | struct io_tctx_exit exit; |
3044 | struct io_tctx_node *node; | |
3045 | int ret; | |
85faa7b8 | 3046 | |
56952e91 JA |
3047 | /* |
3048 | * If we're doing polled IO and end up having requests being | |
3049 | * submitted async (out-of-line), then completions can come in while | |
3050 | * we're waiting for refs to drop. We need to reap these manually, | |
3051 | * as nobody else will be looking for them. | |
3052 | */ | |
b2edc0a7 | 3053 | do { |
a85381d8 PB |
3054 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
3055 | mutex_lock(&ctx->uring_lock); | |
3056 | io_cqring_overflow_kill(ctx); | |
3057 | mutex_unlock(&ctx->uring_lock); | |
3058 | } | |
3059 | ||
c0e0d6ba DY |
3060 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) |
3061 | io_move_task_work_from_local(ctx); | |
3062 | ||
affa87db PB |
3063 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
3064 | cond_resched(); | |
3065 | ||
28090c13 PB |
3066 | if (ctx->sq_data) { |
3067 | struct io_sq_data *sqd = ctx->sq_data; | |
3068 | struct task_struct *tsk; | |
3069 | ||
3070 | io_sq_thread_park(sqd); | |
3071 | tsk = sqd->thread; | |
3072 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
3073 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
3074 | io_cancel_ctx_cb, ctx, true); | |
3075 | io_sq_thread_unpark(sqd); | |
3076 | } | |
b5bb3a24 | 3077 | |
37f0e767 PB |
3078 | io_req_caches_free(ctx); |
3079 | ||
58d3be2c PB |
3080 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
3081 | /* there is little hope left, don't run it too often */ | |
3082 | interval = HZ * 60; | |
3083 | } | |
4826c594 JA |
3084 | /* |
3085 | * This is really an uninterruptible wait, as it has to be | |
3086 | * complete. But it's also run from a kworker, which doesn't | |
3087 | * take signals, so it's fine to make it interruptible. This | |
3088 | * avoids scenarios where we knowingly can wait much longer | |
3089 | * on completions, for example if someone does a SIGSTOP on | |
3090 | * a task that needs to finish task_work to make this loop | |
3091 | * complete. That's a synthetic situation that should not | |
3092 | * cause a stuck task backtrace, and hence a potential panic | |
3093 | * on stuck tasks if that is enabled. | |
3094 | */ | |
3095 | } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 3096 | |
7f00651a PB |
3097 | init_completion(&exit.completion); |
3098 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
3099 | exit.ctx = ctx; | |
f7b32e78 | 3100 | |
d56d938b PB |
3101 | mutex_lock(&ctx->uring_lock); |
3102 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
3103 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
3104 | ||
d56d938b PB |
3105 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
3106 | ctx_node); | |
7f00651a PB |
3107 | /* don't spin on a single task if cancellation failed */ |
3108 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
3109 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
3110 | if (WARN_ON_ONCE(ret)) | |
3111 | continue; | |
d56d938b PB |
3112 | |
3113 | mutex_unlock(&ctx->uring_lock); | |
4826c594 JA |
3114 | /* |
3115 | * See comment above for | |
3116 | * wait_for_completion_interruptible_timeout() on why this | |
3117 | * wait is marked as interruptible. | |
3118 | */ | |
3119 | wait_for_completion_interruptible(&exit.completion); | |
d56d938b PB |
3120 | mutex_lock(&ctx->uring_lock); |
3121 | } | |
3122 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
3123 | spin_lock(&ctx->completion_lock); |
3124 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 3125 | |
d73a572d PB |
3126 | /* pairs with RCU read section in io_req_local_work_add() */ |
3127 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3128 | synchronize_rcu(); | |
3129 | ||
85faa7b8 JA |
3130 | io_ring_ctx_free(ctx); |
3131 | } | |
3132 | ||
c072481d | 3133 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 3134 | { |
61cf9370 MWO |
3135 | unsigned long index; |
3136 | struct creds *creds; | |
3137 | ||
2b188cc1 JA |
3138 | mutex_lock(&ctx->uring_lock); |
3139 | percpu_ref_kill(&ctx->refs); | |
61cf9370 MWO |
3140 | xa_for_each(&ctx->personalities, index, creds) |
3141 | io_unregister_personality(ctx, index); | |
9ca9fb24 PB |
3142 | if (ctx->rings) |
3143 | io_poll_remove_all(ctx, NULL, true); | |
2b188cc1 JA |
3144 | mutex_unlock(&ctx->uring_lock); |
3145 | ||
02bac94b PB |
3146 | /* |
3147 | * If we failed setting up the ctx, we might not have any rings | |
3148 | * and therefore did not submit any requests | |
3149 | */ | |
3150 | if (ctx->rings) | |
60053be8 | 3151 | io_kill_timeouts(ctx, NULL, true); |
309fc03a | 3152 | |
dfbe5561 JA |
3153 | flush_delayed_work(&ctx->fallback_work); |
3154 | ||
85faa7b8 | 3155 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
3156 | /* |
3157 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
3158 | * if we're exiting a ton of rings at the same time. It just adds | |
3159 | * noise and overhead, there's no discernable change in runtime | |
3160 | * over using system_wq. | |
3161 | */ | |
73eaa2b5 | 3162 | queue_work(iou_wq, &ctx->exit_work); |
2b188cc1 JA |
3163 | } |
3164 | ||
3165 | static int io_uring_release(struct inode *inode, struct file *file) | |
3166 | { | |
3167 | struct io_ring_ctx *ctx = file->private_data; | |
3168 | ||
3169 | file->private_data = NULL; | |
3170 | io_ring_ctx_wait_and_kill(ctx); | |
3171 | return 0; | |
3172 | } | |
3173 | ||
f6edbabb PB |
3174 | struct io_task_cancel { |
3175 | struct task_struct *task; | |
3dd0c97a | 3176 | bool all; |
f6edbabb | 3177 | }; |
f254ac04 | 3178 | |
f6edbabb | 3179 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 3180 | { |
9a472ef7 | 3181 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 3182 | struct io_task_cancel *cancel = data; |
9a472ef7 | 3183 | |
6af3f48b | 3184 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
3185 | } |
3186 | ||
c072481d PB |
3187 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
3188 | struct task_struct *task, | |
3189 | bool cancel_all) | |
b7ddce3c | 3190 | { |
e1915f76 | 3191 | struct io_defer_entry *de; |
b7ddce3c PB |
3192 | LIST_HEAD(list); |
3193 | ||
79ebeaee | 3194 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 3195 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 3196 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
3197 | list_cut_position(&list, &ctx->defer_list, &de->list); |
3198 | break; | |
3199 | } | |
3200 | } | |
79ebeaee | 3201 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
3202 | if (list_empty(&list)) |
3203 | return false; | |
b7ddce3c PB |
3204 | |
3205 | while (!list_empty(&list)) { | |
3206 | de = list_first_entry(&list, struct io_defer_entry, list); | |
3207 | list_del_init(&de->list); | |
e276ae34 | 3208 | io_req_task_queue_fail(de->req, -ECANCELED); |
b7ddce3c PB |
3209 | kfree(de); |
3210 | } | |
e1915f76 | 3211 | return true; |
b7ddce3c PB |
3212 | } |
3213 | ||
c072481d | 3214 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
3215 | { |
3216 | struct io_tctx_node *node; | |
3217 | enum io_wq_cancel cret; | |
3218 | bool ret = false; | |
3219 | ||
3220 | mutex_lock(&ctx->uring_lock); | |
3221 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3222 | struct io_uring_task *tctx = node->task->io_uring; | |
3223 | ||
3224 | /* | |
3225 | * io_wq will stay alive while we hold uring_lock, because it's | |
3226 | * killed after ctx nodes, which requires to take the lock. | |
3227 | */ | |
3228 | if (!tctx || !tctx->io_wq) | |
3229 | continue; | |
3230 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
3231 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3232 | } | |
3233 | mutex_unlock(&ctx->uring_lock); | |
3234 | ||
3235 | return ret; | |
3236 | } | |
3237 | ||
affa87db | 3238 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
3239 | struct task_struct *task, |
3240 | bool cancel_all) | |
9936c7c2 | 3241 | { |
3dd0c97a | 3242 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 3243 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
3244 | enum io_wq_cancel cret; |
3245 | bool ret = false; | |
9936c7c2 | 3246 | |
360cd42c PB |
3247 | /* set it so io_req_local_work_add() would wake us up */ |
3248 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { | |
3249 | atomic_set(&ctx->cq_wait_nr, 1); | |
3250 | smp_mb(); | |
3251 | } | |
3252 | ||
60053be8 PB |
3253 | /* failed during ring init, it couldn't have issued any requests */ |
3254 | if (!ctx->rings) | |
affa87db | 3255 | return false; |
60053be8 | 3256 | |
affa87db PB |
3257 | if (!task) { |
3258 | ret |= io_uring_try_cancel_iowq(ctx); | |
3259 | } else if (tctx && tctx->io_wq) { | |
3260 | /* | |
3261 | * Cancels requests of all rings, not only @ctx, but | |
3262 | * it's fine as the task is in exit/exec. | |
3263 | */ | |
3264 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
3265 | &cancel, true); | |
3266 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3267 | } | |
9936c7c2 | 3268 | |
affa87db PB |
3269 | /* SQPOLL thread does its own polling */ |
3270 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
3271 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
3272 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
3273 | io_iopoll_try_reap_events(ctx); | |
3274 | ret = true; | |
fcc926bb | 3275 | cond_resched(); |
9936c7c2 | 3276 | } |
9936c7c2 | 3277 | } |
affa87db | 3278 | |
140102ae PB |
3279 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3280 | io_allowed_defer_tw_run(ctx)) | |
9fe3eaea | 3281 | ret |= io_run_local_work(ctx, INT_MAX) > 0; |
affa87db PB |
3282 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
3283 | mutex_lock(&ctx->uring_lock); | |
3284 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
f31ecf67 | 3285 | ret |= io_waitid_remove_all(ctx, task, cancel_all); |
194bb58c | 3286 | ret |= io_futex_remove_all(ctx, task, cancel_all); |
93b8cc60 | 3287 | ret |= io_uring_try_cancel_uring_cmd(ctx, task, cancel_all); |
affa87db PB |
3288 | mutex_unlock(&ctx->uring_lock); |
3289 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
3290 | if (task) | |
c0e0d6ba | 3291 | ret |= io_run_task_work() > 0; |
affa87db | 3292 | return ret; |
9936c7c2 PB |
3293 | } |
3294 | ||
3f48cf18 | 3295 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 3296 | { |
3f48cf18 | 3297 | if (tracked) |
9cae36a0 | 3298 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
3299 | return percpu_counter_sum(&tctx->inflight); |
3300 | } | |
3301 | ||
78cc687b PB |
3302 | /* |
3303 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 3304 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 3305 | */ |
17437f31 | 3306 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 3307 | { |
521d6a73 | 3308 | struct io_uring_task *tctx = current->io_uring; |
734551df | 3309 | struct io_ring_ctx *ctx; |
360cd42c PB |
3310 | struct io_tctx_node *node; |
3311 | unsigned long index; | |
0e9ddb39 PB |
3312 | s64 inflight; |
3313 | DEFINE_WAIT(wait); | |
fdaf083c | 3314 | |
78cc687b PB |
3315 | WARN_ON_ONCE(sqd && sqd->thread != current); |
3316 | ||
6d042ffb PO |
3317 | if (!current->io_uring) |
3318 | return; | |
17a91051 PB |
3319 | if (tctx->io_wq) |
3320 | io_wq_exit_start(tctx->io_wq); | |
3321 | ||
8d664282 | 3322 | atomic_inc(&tctx->in_cancel); |
0e9ddb39 | 3323 | do { |
affa87db PB |
3324 | bool loop = false; |
3325 | ||
e9dbe221 | 3326 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 3327 | /* read completions before cancelations */ |
78cc687b | 3328 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
3329 | if (!inflight) |
3330 | break; | |
fdaf083c | 3331 | |
78cc687b | 3332 | if (!sqd) { |
78cc687b PB |
3333 | xa_for_each(&tctx->xa, index, node) { |
3334 | /* sqpoll task will cancel all its requests */ | |
3335 | if (node->ctx->sq_data) | |
3336 | continue; | |
affa87db PB |
3337 | loop |= io_uring_try_cancel_requests(node->ctx, |
3338 | current, cancel_all); | |
78cc687b PB |
3339 | } |
3340 | } else { | |
3341 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
3342 | loop |= io_uring_try_cancel_requests(ctx, |
3343 | current, | |
3344 | cancel_all); | |
3345 | } | |
3346 | ||
3347 | if (loop) { | |
3348 | cond_resched(); | |
3349 | continue; | |
78cc687b | 3350 | } |
17a91051 | 3351 | |
78a78060 JA |
3352 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
3353 | io_run_task_work(); | |
e9dbe221 | 3354 | io_uring_drop_tctx_refs(current); |
360cd42c PB |
3355 | xa_for_each(&tctx->xa, index, node) { |
3356 | if (!llist_empty(&node->ctx->work_llist)) { | |
3357 | WARN_ON_ONCE(node->ctx->submitter_task && | |
3358 | node->ctx->submitter_task != current); | |
3359 | goto end_wait; | |
3360 | } | |
3361 | } | |
0f212204 | 3362 | /* |
a1bb3cd5 PB |
3363 | * If we've seen completions, retry without waiting. This |
3364 | * avoids a race where a completion comes in before we did | |
3365 | * prepare_to_wait(). | |
0f212204 | 3366 | */ |
3dd0c97a | 3367 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 3368 | schedule(); |
360cd42c | 3369 | end_wait: |
f57555ed | 3370 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 3371 | } while (1); |
de7f1d9e | 3372 | |
8452d4a6 | 3373 | io_uring_clean_tctx(tctx); |
3dd0c97a | 3374 | if (cancel_all) { |
3cc7fdb9 PB |
3375 | /* |
3376 | * We shouldn't run task_works after cancel, so just leave | |
8d664282 | 3377 | * ->in_cancel set for normal exit. |
3cc7fdb9 | 3378 | */ |
8d664282 | 3379 | atomic_dec(&tctx->in_cancel); |
3f48cf18 PB |
3380 | /* for exec all current's requests should be gone, kill tctx */ |
3381 | __io_uring_free(current); | |
3382 | } | |
44e728b8 PB |
3383 | } |
3384 | ||
f552a27a | 3385 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 3386 | { |
f552a27a | 3387 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
3388 | } |
3389 | ||
6c5c240e RP |
3390 | static void *io_uring_validate_mmap_request(struct file *file, |
3391 | loff_t pgoff, size_t sz) | |
2b188cc1 | 3392 | { |
2b188cc1 | 3393 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 3394 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
3395 | struct page *page; |
3396 | void *ptr; | |
3397 | ||
c56e022c | 3398 | switch (offset & IORING_OFF_MMAP_MASK) { |
2b188cc1 | 3399 | case IORING_OFF_SQ_RING: |
75b28aff | 3400 | case IORING_OFF_CQ_RING: |
6f007b14 JA |
3401 | /* Don't allow mmap if the ring was setup without it */ |
3402 | if (ctx->flags & IORING_SETUP_NO_MMAP) | |
3403 | return ERR_PTR(-EINVAL); | |
75b28aff | 3404 | ptr = ctx->rings; |
2b188cc1 JA |
3405 | break; |
3406 | case IORING_OFF_SQES: | |
6f007b14 JA |
3407 | /* Don't allow mmap if the ring was setup without it */ |
3408 | if (ctx->flags & IORING_SETUP_NO_MMAP) | |
3409 | return ERR_PTR(-EINVAL); | |
2b188cc1 JA |
3410 | ptr = ctx->sq_sqes; |
3411 | break; | |
c56e022c | 3412 | case IORING_OFF_PBUF_RING: { |
561e4f94 | 3413 | struct io_buffer_list *bl; |
c56e022c JA |
3414 | unsigned int bgid; |
3415 | ||
3416 | bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT; | |
561e4f94 JA |
3417 | bl = io_pbuf_get_bl(ctx, bgid); |
3418 | if (IS_ERR(bl)) | |
3419 | return bl; | |
3420 | ptr = bl->buf_ring; | |
3421 | io_put_bl(ctx, bl); | |
c56e022c JA |
3422 | break; |
3423 | } | |
2b188cc1 | 3424 | default: |
6c5c240e | 3425 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
3426 | } |
3427 | ||
3428 | page = virt_to_head_page(ptr); | |
a50b854e | 3429 | if (sz > page_size(page)) |
6c5c240e RP |
3430 | return ERR_PTR(-EINVAL); |
3431 | ||
3432 | return ptr; | |
3433 | } | |
3434 | ||
3435 | #ifdef CONFIG_MMU | |
3436 | ||
c072481d | 3437 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e RP |
3438 | { |
3439 | size_t sz = vma->vm_end - vma->vm_start; | |
3440 | unsigned long pfn; | |
3441 | void *ptr; | |
3442 | ||
3443 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
3444 | if (IS_ERR(ptr)) | |
3445 | return PTR_ERR(ptr); | |
2b188cc1 JA |
3446 | |
3447 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3448 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3449 | } | |
3450 | ||
d808459b HD |
3451 | static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp, |
3452 | unsigned long addr, unsigned long len, | |
3453 | unsigned long pgoff, unsigned long flags) | |
3454 | { | |
d808459b HD |
3455 | void *ptr; |
3456 | ||
3457 | /* | |
3458 | * Do not allow to map to user-provided address to avoid breaking the | |
3459 | * aliasing rules. Userspace is not able to guess the offset address of | |
3460 | * kernel kmalloc()ed memory area. | |
3461 | */ | |
3462 | if (addr) | |
3463 | return -EINVAL; | |
3464 | ||
3465 | ptr = io_uring_validate_mmap_request(filp, pgoff, len); | |
3466 | if (IS_ERR(ptr)) | |
3467 | return -ENOMEM; | |
3468 | ||
32832a40 HD |
3469 | /* |
3470 | * Some architectures have strong cache aliasing requirements. | |
3471 | * For such architectures we need a coherent mapping which aliases | |
3472 | * kernel memory *and* userspace memory. To achieve that: | |
3473 | * - use a NULL file pointer to reference physical memory, and | |
3474 | * - use the kernel virtual address of the shared io_uring context | |
3475 | * (instead of the userspace-provided address, which has to be 0UL | |
3476 | * anyway). | |
56675f8b HD |
3477 | * - use the same pgoff which the get_unmapped_area() uses to |
3478 | * calculate the page colouring. | |
32832a40 HD |
3479 | * For architectures without such aliasing requirements, the |
3480 | * architecture will return any suitable mapping because addr is 0. | |
3481 | */ | |
3482 | filp = NULL; | |
3483 | flags |= MAP_SHARED; | |
3484 | pgoff = 0; /* has been translated to ptr above */ | |
d808459b | 3485 | #ifdef SHM_COLOUR |
32832a40 | 3486 | addr = (uintptr_t) ptr; |
56675f8b | 3487 | pgoff = addr >> PAGE_SHIFT; |
d808459b | 3488 | #else |
32832a40 | 3489 | addr = 0UL; |
d808459b | 3490 | #endif |
32832a40 | 3491 | return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); |
d808459b HD |
3492 | } |
3493 | ||
6c5c240e RP |
3494 | #else /* !CONFIG_MMU */ |
3495 | ||
3496 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3497 | { | |
fc4f4be9 | 3498 | return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL; |
6c5c240e RP |
3499 | } |
3500 | ||
3501 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
3502 | { | |
3503 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
3504 | } | |
3505 | ||
3506 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
3507 | unsigned long addr, unsigned long len, | |
3508 | unsigned long pgoff, unsigned long flags) | |
3509 | { | |
3510 | void *ptr; | |
3511 | ||
3512 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
3513 | if (IS_ERR(ptr)) | |
3514 | return PTR_ERR(ptr); | |
3515 | ||
3516 | return (unsigned long) ptr; | |
3517 | } | |
3518 | ||
3519 | #endif /* !CONFIG_MMU */ | |
3520 | ||
f81440d3 PB |
3521 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
3522 | { | |
3523 | if (flags & IORING_ENTER_EXT_ARG) { | |
3524 | struct io_uring_getevents_arg arg; | |
3525 | ||
3526 | if (argsz != sizeof(arg)) | |
3527 | return -EINVAL; | |
3528 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3529 | return -EFAULT; | |
3530 | } | |
3531 | return 0; | |
3532 | } | |
3533 | ||
c73ebb68 HX |
3534 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
3535 | struct __kernel_timespec __user **ts, | |
3536 | const sigset_t __user **sig) | |
3537 | { | |
3538 | struct io_uring_getevents_arg arg; | |
3539 | ||
3540 | /* | |
3541 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
3542 | * is just a pointer to the sigset_t. | |
3543 | */ | |
3544 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
3545 | *sig = (const sigset_t __user *) argp; | |
3546 | *ts = NULL; | |
3547 | return 0; | |
3548 | } | |
3549 | ||
3550 | /* | |
3551 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
3552 | * timespec and sigset_t pointers if good. | |
3553 | */ | |
3554 | if (*argsz != sizeof(arg)) | |
3555 | return -EINVAL; | |
3556 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3557 | return -EFAULT; | |
d2347b96 DY |
3558 | if (arg.pad) |
3559 | return -EINVAL; | |
c73ebb68 HX |
3560 | *sig = u64_to_user_ptr(arg.sigmask); |
3561 | *argsz = arg.sigmask_sz; | |
3562 | *ts = u64_to_user_ptr(arg.ts); | |
3563 | return 0; | |
3564 | } | |
3565 | ||
2b188cc1 | 3566 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
3567 | u32, min_complete, u32, flags, const void __user *, argp, |
3568 | size_t, argsz) | |
2b188cc1 JA |
3569 | { |
3570 | struct io_ring_ctx *ctx; | |
73363c26 | 3571 | struct file *file; |
33f993da | 3572 | long ret; |
2b188cc1 | 3573 | |
33f993da | 3574 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
3575 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
3576 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
3577 | return -EINVAL; |
3578 | ||
e7a6c00d JA |
3579 | /* |
3580 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3581 | * need only dereference our task private array to find it. | |
3582 | */ | |
3583 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3584 | struct io_uring_task *tctx = current->io_uring; | |
3585 | ||
3273c440 | 3586 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
e7a6c00d JA |
3587 | return -EINVAL; |
3588 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
73363c26 JA |
3589 | file = tctx->registered_rings[fd]; |
3590 | if (unlikely(!file)) | |
3273c440 | 3591 | return -EBADF; |
e7a6c00d | 3592 | } else { |
73363c26 JA |
3593 | file = fget(fd); |
3594 | if (unlikely(!file)) | |
3273c440 PB |
3595 | return -EBADF; |
3596 | ret = -EOPNOTSUPP; | |
73363c26 | 3597 | if (unlikely(!io_is_uring_fops(file))) |
fbb8bb02 | 3598 | goto out; |
e7a6c00d | 3599 | } |
2b188cc1 | 3600 | |
73363c26 | 3601 | ctx = file->private_data; |
7e84e1c7 | 3602 | ret = -EBADFD; |
33f993da | 3603 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3604 | goto out; |
3605 | ||
6c271ce2 JA |
3606 | /* |
3607 | * For SQ polling, the thread will do all submissions and completions. | |
3608 | * Just return the requested submit count, and wake the thread if | |
3609 | * we were asked to. | |
3610 | */ | |
b2a9eada | 3611 | ret = 0; |
6c271ce2 | 3612 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3613 | io_cqring_overflow_flush(ctx); |
89448c47 | 3614 | |
21f96522 JA |
3615 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3616 | ret = -EOWNERDEAD; | |
04147488 | 3617 | goto out; |
21f96522 | 3618 | } |
6c271ce2 | 3619 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3620 | wake_up(&ctx->sq_data->wait); |
88b80534 QF |
3621 | if (flags & IORING_ENTER_SQ_WAIT) |
3622 | io_sqpoll_wait_sq(ctx); | |
3623 | ||
3e813c90 | 3624 | ret = to_submit; |
b2a9eada | 3625 | } else if (to_submit) { |
eef51daa | 3626 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3627 | if (unlikely(ret)) |
3628 | goto out; | |
7c504e65 | 3629 | |
2b188cc1 | 3630 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3631 | ret = io_submit_sqes(ctx, to_submit); |
3632 | if (ret != to_submit) { | |
d487b43c | 3633 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3634 | goto out; |
d487b43c | 3635 | } |
44f87745 PB |
3636 | if (flags & IORING_ENTER_GETEVENTS) { |
3637 | if (ctx->syscall_iopoll) | |
3638 | goto iopoll_locked; | |
3639 | /* | |
3640 | * Ignore errors, we'll soon call io_cqring_wait() and | |
3641 | * it should handle ownership problems if any. | |
3642 | */ | |
3643 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
9fe3eaea | 3644 | (void)io_run_local_work_locked(ctx, min_complete); |
44f87745 | 3645 | } |
d487b43c | 3646 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 | 3647 | } |
c0e0d6ba | 3648 | |
2b188cc1 | 3649 | if (flags & IORING_ENTER_GETEVENTS) { |
3e813c90 | 3650 | int ret2; |
c0e0d6ba | 3651 | |
773697b6 | 3652 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3653 | /* |
3654 | * We disallow the app entering submit/complete with | |
3655 | * polling, but we still need to lock the ring to | |
3656 | * prevent racing with polled issue that got punted to | |
3657 | * a workqueue. | |
3658 | */ | |
3659 | mutex_lock(&ctx->uring_lock); | |
3660 | iopoll_locked: | |
3e813c90 DY |
3661 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3662 | if (likely(!ret2)) { | |
3663 | min_complete = min(min_complete, | |
3664 | ctx->cq_entries); | |
3665 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3666 | } |
3667 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3668 | } else { |
f81440d3 PB |
3669 | const sigset_t __user *sig; |
3670 | struct __kernel_timespec __user *ts; | |
3671 | ||
3e813c90 DY |
3672 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3673 | if (likely(!ret2)) { | |
3674 | min_complete = min(min_complete, | |
3675 | ctx->cq_entries); | |
3676 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3677 | argsz, ts); | |
3678 | } | |
def596e9 | 3679 | } |
c73ebb68 | 3680 | |
155bc950 | 3681 | if (!ret) { |
3e813c90 | 3682 | ret = ret2; |
2b188cc1 | 3683 | |
155bc950 DY |
3684 | /* |
3685 | * EBADR indicates that one or more CQE were dropped. | |
3686 | * Once the user has been informed we can clear the bit | |
3687 | * as they are obviously ok with those drops. | |
3688 | */ | |
3689 | if (unlikely(ret2 == -EBADR)) | |
3690 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3691 | &ctx->check_cq); | |
def596e9 | 3692 | } |
2b188cc1 | 3693 | } |
7c504e65 | 3694 | out: |
73363c26 JA |
3695 | if (!(flags & IORING_ENTER_REGISTERED_RING)) |
3696 | fput(file); | |
3e813c90 | 3697 | return ret; |
2b188cc1 JA |
3698 | } |
3699 | ||
3700 | static const struct file_operations io_uring_fops = { | |
3701 | .release = io_uring_release, | |
3702 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3703 | #ifndef CONFIG_MMU |
3704 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3705 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
d808459b HD |
3706 | #else |
3707 | .get_unmapped_area = io_uring_mmu_get_unmapped_area, | |
6c5c240e | 3708 | #endif |
2b188cc1 | 3709 | .poll = io_uring_poll, |
bebdb65e | 3710 | #ifdef CONFIG_PROC_FS |
87ce955b | 3711 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3712 | #endif |
2b188cc1 JA |
3713 | }; |
3714 | ||
92ac8bea JA |
3715 | bool io_is_uring_fops(struct file *file) |
3716 | { | |
3717 | return file->f_op == &io_uring_fops; | |
3718 | } | |
3719 | ||
c072481d PB |
3720 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3721 | struct io_uring_params *p) | |
2b188cc1 | 3722 | { |
75b28aff HV |
3723 | struct io_rings *rings; |
3724 | size_t size, sq_array_offset; | |
e27cef86 | 3725 | void *ptr; |
2b188cc1 | 3726 | |
bd740481 JA |
3727 | /* make sure these are sane, as we already accounted them */ |
3728 | ctx->sq_entries = p->sq_entries; | |
3729 | ctx->cq_entries = p->cq_entries; | |
3730 | ||
baf9cb64 | 3731 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3732 | if (size == SIZE_MAX) |
3733 | return -EOVERFLOW; | |
3734 | ||
03d89a2d JA |
3735 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3736 | rings = io_mem_alloc(size); | |
3737 | else | |
3738 | rings = io_rings_map(ctx, p->cq_off.user_addr, size); | |
3739 | ||
e27cef86 JA |
3740 | if (IS_ERR(rings)) |
3741 | return PTR_ERR(rings); | |
2b188cc1 | 3742 | |
75b28aff | 3743 | ctx->rings = rings; |
2af89abd PB |
3744 | if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) |
3745 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
75b28aff HV |
3746 | rings->sq_ring_mask = p->sq_entries - 1; |
3747 | rings->cq_ring_mask = p->cq_entries - 1; | |
3748 | rings->sq_ring_entries = p->sq_entries; | |
3749 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3750 | |
ebdeb7c0 JA |
3751 | if (p->flags & IORING_SETUP_SQE128) |
3752 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3753 | else | |
3754 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 | 3755 | if (size == SIZE_MAX) { |
9c189eee | 3756 | io_rings_free(ctx); |
2b188cc1 | 3757 | return -EOVERFLOW; |
eb065d30 | 3758 | } |
2b188cc1 | 3759 | |
03d89a2d JA |
3760 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3761 | ptr = io_mem_alloc(size); | |
3762 | else | |
3763 | ptr = io_sqes_map(ctx, p->sq_off.user_addr, size); | |
3764 | ||
e27cef86 | 3765 | if (IS_ERR(ptr)) { |
9c189eee | 3766 | io_rings_free(ctx); |
e27cef86 | 3767 | return PTR_ERR(ptr); |
eb065d30 | 3768 | } |
2b188cc1 | 3769 | |
e27cef86 | 3770 | ctx->sq_sqes = ptr; |
2b188cc1 JA |
3771 | return 0; |
3772 | } | |
3773 | ||
6e76ac59 | 3774 | static int io_uring_install_fd(struct file *file) |
9faadcc8 | 3775 | { |
6e76ac59 | 3776 | int fd; |
9faadcc8 PB |
3777 | |
3778 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3779 | if (fd < 0) | |
3780 | return fd; | |
9faadcc8 PB |
3781 | fd_install(fd, file); |
3782 | return fd; | |
3783 | } | |
3784 | ||
2b188cc1 JA |
3785 | /* |
3786 | * Allocate an anonymous fd, this is what constitutes the application | |
3787 | * visible backing of an io_uring instance. The application mmaps this | |
6e5e6d27 | 3788 | * fd to gain access to the SQ/CQ ring details. |
2b188cc1 | 3789 | */ |
9faadcc8 | 3790 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 | 3791 | { |
4f0b9194 | 3792 | /* Create a new inode so that the LSM can block the creation. */ |
09d1c6a8 | 3793 | return anon_inode_create_getfile("[io_uring]", &io_uring_fops, ctx, |
91a9ab7c | 3794 | O_RDWR | O_CLOEXEC, NULL); |
2b188cc1 JA |
3795 | } |
3796 | ||
c072481d PB |
3797 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3798 | struct io_uring_params __user *params) | |
2b188cc1 | 3799 | { |
2b188cc1 | 3800 | struct io_ring_ctx *ctx; |
6e76ac59 | 3801 | struct io_uring_task *tctx; |
9faadcc8 | 3802 | struct file *file; |
2b188cc1 JA |
3803 | int ret; |
3804 | ||
8110c1a6 | 3805 | if (!entries) |
2b188cc1 | 3806 | return -EINVAL; |
8110c1a6 JA |
3807 | if (entries > IORING_MAX_ENTRIES) { |
3808 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3809 | return -EINVAL; | |
3810 | entries = IORING_MAX_ENTRIES; | |
3811 | } | |
2b188cc1 | 3812 | |
6e76ac59 JT |
3813 | if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY) |
3814 | && !(p->flags & IORING_SETUP_NO_MMAP)) | |
3815 | return -EINVAL; | |
3816 | ||
2b188cc1 JA |
3817 | /* |
3818 | * Use twice as many entries for the CQ ring. It's possible for the | |
3819 | * application to drive a higher depth than the size of the SQ ring, | |
3820 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3821 | * some flexibility in overcommitting a bit. If the application has |
3822 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3823 | * of CQ ring entries manually. | |
2b188cc1 JA |
3824 | */ |
3825 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3826 | if (p->flags & IORING_SETUP_CQSIZE) { |
3827 | /* | |
3828 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3829 | * to a power-of-two, if it isn't already. We do NOT impose | |
3830 | * any cq vs sq ring sizing. | |
3831 | */ | |
eb2667b3 | 3832 | if (!p->cq_entries) |
33a107f0 | 3833 | return -EINVAL; |
8110c1a6 JA |
3834 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3835 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3836 | return -EINVAL; | |
3837 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3838 | } | |
eb2667b3 JQ |
3839 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3840 | if (p->cq_entries < p->sq_entries) | |
3841 | return -EINVAL; | |
33a107f0 JA |
3842 | } else { |
3843 | p->cq_entries = 2 * p->sq_entries; | |
3844 | } | |
2b188cc1 | 3845 | |
2b188cc1 | 3846 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3847 | if (!ctx) |
2b188cc1 | 3848 | return -ENOMEM; |
773697b6 | 3849 | |
e6aeb272 PB |
3850 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3851 | !(ctx->flags & IORING_SETUP_IOPOLL) && | |
3852 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3853 | ctx->task_complete = true; | |
3854 | ||
ec26c225 PB |
3855 | if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL)) |
3856 | ctx->lockless_cq = true; | |
3857 | ||
bca39f39 PB |
3858 | /* |
3859 | * lazy poll_wq activation relies on ->task_complete for synchronisation | |
3860 | * purposes, see io_activate_pollwq() | |
3861 | */ | |
3862 | if (!ctx->task_complete) | |
3863 | ctx->poll_activated = true; | |
3864 | ||
773697b6 PB |
3865 | /* |
3866 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3867 | * space applications don't need to do io completion events | |
3868 | * polling again, they can rely on io_sq_thread to do polling | |
3869 | * work, which can reduce cpu usage and uring_lock contention. | |
3870 | */ | |
3871 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3872 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3873 | ctx->syscall_iopoll = 1; | |
3874 | ||
2b188cc1 | 3875 | ctx->compat = in_compat_syscall(); |
6adc2272 | 3876 | if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK)) |
62e398be | 3877 | ctx->user = get_uid(current_user()); |
2aede0e4 | 3878 | |
9f010507 | 3879 | /* |
e1169f06 JA |
3880 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3881 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3882 | */ |
e1169f06 JA |
3883 | ret = -EINVAL; |
3884 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3885 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 | 3886 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
c0e0d6ba DY |
3887 | IORING_SETUP_TASKRUN_FLAG | |
3888 | IORING_SETUP_DEFER_TASKRUN)) | |
e1169f06 | 3889 | goto err; |
9f010507 | 3890 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3891 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3892 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3893 | } else { | |
c0e0d6ba DY |
3894 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && |
3895 | !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) | |
ef060ea9 | 3896 | goto err; |
9f010507 | 3897 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3898 | } |
9f010507 | 3899 | |
c0e0d6ba DY |
3900 | /* |
3901 | * For DEFER_TASKRUN we require the completion task to be the same as the | |
3902 | * submission task. This implies that there is only one submitter, so enforce | |
3903 | * that. | |
3904 | */ | |
3905 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && | |
3906 | !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { | |
3907 | goto err; | |
3908 | } | |
3909 | ||
2aede0e4 JA |
3910 | /* |
3911 | * This is just grabbed for accounting purposes. When a process exits, | |
3912 | * the mm is exited and dropped before the files, hence we need to hang | |
3913 | * on to this mm purely for the purposes of being able to unaccount | |
3914 | * memory (locked/pinned vm). It's not used for anything else. | |
3915 | */ | |
6b7898eb | 3916 | mmgrab(current->mm); |
2aede0e4 | 3917 | ctx->mm_account = current->mm; |
6b7898eb | 3918 | |
2b188cc1 JA |
3919 | ret = io_allocate_scq_urings(ctx, p); |
3920 | if (ret) | |
3921 | goto err; | |
3922 | ||
7e84e1c7 | 3923 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
3924 | if (ret) |
3925 | goto err; | |
2933ae6e PB |
3926 | |
3927 | ret = io_rsrc_init(ctx); | |
47b228ce PB |
3928 | if (ret) |
3929 | goto err; | |
2b188cc1 | 3930 | |
75b28aff HV |
3931 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3932 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3933 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3934 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3935 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3936 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
2af89abd PB |
3937 | if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) |
3938 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
9b1b58ca | 3939 | p->sq_off.resv1 = 0; |
03d89a2d JA |
3940 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3941 | p->sq_off.user_addr = 0; | |
2b188cc1 | 3942 | |
75b28aff HV |
3943 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3944 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3945 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3946 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3947 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3948 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 3949 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
9b1b58ca | 3950 | p->cq_off.resv1 = 0; |
03d89a2d JA |
3951 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3952 | p->cq_off.user_addr = 0; | |
ac90f249 | 3953 | |
7f13657d XW |
3954 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
3955 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 3956 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 3957 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 3958 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e | 3959 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
7d3fd88d | 3960 | IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING; |
7f13657d XW |
3961 | |
3962 | if (copy_to_user(params, p, sizeof(*p))) { | |
3963 | ret = -EFAULT; | |
3964 | goto err; | |
3965 | } | |
d1719f70 | 3966 | |
7cae596b DY |
3967 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER |
3968 | && !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
8579538c | 3969 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
7cae596b | 3970 | |
9faadcc8 PB |
3971 | file = io_uring_get_file(ctx); |
3972 | if (IS_ERR(file)) { | |
3973 | ret = PTR_ERR(file); | |
3974 | goto err; | |
3975 | } | |
3976 | ||
6e76ac59 JT |
3977 | ret = __io_uring_add_tctx_node(ctx); |
3978 | if (ret) | |
3979 | goto err_fput; | |
3980 | tctx = current->io_uring; | |
3981 | ||
044c1ab3 JA |
3982 | /* |
3983 | * Install ring fd as the very last thing, so we don't risk someone | |
3984 | * having closed it before we finish setup | |
3985 | */ | |
6e76ac59 JT |
3986 | if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY) |
3987 | ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX); | |
3988 | else | |
3989 | ret = io_uring_install_fd(file); | |
3990 | if (ret < 0) | |
3991 | goto err_fput; | |
044c1ab3 | 3992 | |
c826bd7a | 3993 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
3994 | return ret; |
3995 | err: | |
3996 | io_ring_ctx_wait_and_kill(ctx); | |
3997 | return ret; | |
6e76ac59 JT |
3998 | err_fput: |
3999 | fput(file); | |
4000 | return ret; | |
2b188cc1 JA |
4001 | } |
4002 | ||
4003 | /* | |
4004 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
4005 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
4006 | * params structure passed in. | |
4007 | */ | |
4008 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
4009 | { | |
4010 | struct io_uring_params p; | |
2b188cc1 JA |
4011 | int i; |
4012 | ||
4013 | if (copy_from_user(&p, params, sizeof(p))) | |
4014 | return -EFAULT; | |
4015 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
4016 | if (p.resv[i]) | |
4017 | return -EINVAL; | |
4018 | } | |
4019 | ||
6c271ce2 | 4020 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 4021 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 4022 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 4023 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 4024 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 | 4025 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
03d89a2d | 4026 | IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN | |
2af89abd PB |
4027 | IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY | |
4028 | IORING_SETUP_NO_SQARRAY)) | |
2b188cc1 JA |
4029 | return -EINVAL; |
4030 | ||
ef060ea9 | 4031 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
4032 | } |
4033 | ||
76d3ccec MR |
4034 | static inline bool io_uring_allowed(void) |
4035 | { | |
4036 | int disabled = READ_ONCE(sysctl_io_uring_disabled); | |
4037 | kgid_t io_uring_group; | |
4038 | ||
4039 | if (disabled == 2) | |
4040 | return false; | |
4041 | ||
4042 | if (disabled == 0 || capable(CAP_SYS_ADMIN)) | |
4043 | return true; | |
4044 | ||
4045 | io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group); | |
4046 | if (!gid_valid(io_uring_group)) | |
4047 | return false; | |
4048 | ||
4049 | return in_group_p(io_uring_group); | |
4050 | } | |
4051 | ||
2b188cc1 JA |
4052 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, |
4053 | struct io_uring_params __user *, params) | |
4054 | { | |
76d3ccec MR |
4055 | if (!io_uring_allowed()) |
4056 | return -EPERM; | |
4057 | ||
2b188cc1 JA |
4058 | return io_uring_setup(entries, params); |
4059 | } | |
4060 | ||
4061 | static int __init io_uring_init(void) | |
4062 | { | |
9c71d39a | 4063 | #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ |
d7f62e82 | 4064 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ |
9c71d39a | 4065 | BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ |
d7f62e82 SM |
4066 | } while (0) |
4067 | ||
4068 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9c71d39a SM |
4069 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) |
4070 | #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ | |
4071 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) | |
d7f62e82 SM |
4072 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); |
4073 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
4074 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
4075 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
4076 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
4077 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
4078 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9c71d39a SM |
4079 | BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); |
4080 | BUILD_BUG_SQE_ELEM(12, __u32, __pad1); | |
d7f62e82 | 4081 | BUILD_BUG_SQE_ELEM(16, __u64, addr); |
7d67af2c | 4082 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
4083 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
4084 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
4085 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
4086 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
4087 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
4088 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
4089 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
4090 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
4091 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
4092 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
4093 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
4094 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
4095 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
4096 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
4097 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 4098 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
9c71d39a SM |
4099 | BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); |
4100 | BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); | |
4101 | BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); | |
4102 | BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); | |
4103 | BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); | |
d7f62e82 SM |
4104 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
4105 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 4106 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 4107 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 4108 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 4109 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
b48c312b PB |
4110 | BUILD_BUG_SQE_ELEM(44, __u16, addr_len); |
4111 | BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); | |
e9621e2b | 4112 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
9c71d39a SM |
4113 | BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); |
4114 | BUILD_BUG_SQE_ELEM(56, __u64, __pad2); | |
d7f62e82 | 4115 | |
b0d658ec PB |
4116 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
4117 | sizeof(struct io_uring_rsrc_update)); | |
4118 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
4119 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
4120 | |
4121 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
4122 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
4123 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
4124 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 4125 | |
b0d658ec PB |
4126 | /* should fit into one byte */ |
4127 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
4128 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
4129 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 4130 | |
4bcb982c | 4131 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof_field(struct io_kiocb, flags)); |
16340eab | 4132 | |
3a4b89a2 JA |
4133 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
4134 | ||
528ce678 ML |
4135 | /* top 8bits are for internal use */ |
4136 | BUILD_BUG_ON((IORING_URING_CMD_MASK & 0xff000000) != 0); | |
4137 | ||
d9b57aa3 | 4138 | io_uring_optable_init(); |
0702e536 | 4139 | |
b97f96e2 JA |
4140 | /* |
4141 | * Allow user copy in the per-command field, which starts after the | |
4142 | * file in io_kiocb and until the opcode field. The openat2 handling | |
4143 | * requires copying in user memory into the io_kiocb object in that | |
4144 | * range, and HARDENED_USERCOPY will complain if we haven't | |
4145 | * correctly annotated this range. | |
4146 | */ | |
4147 | req_cachep = kmem_cache_create_usercopy("io_kiocb", | |
4148 | sizeof(struct io_kiocb), 0, | |
4149 | SLAB_HWCACHE_ALIGN | SLAB_PANIC | | |
4150 | SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU, | |
4151 | offsetof(struct io_kiocb, cmd.data), | |
4152 | sizeof_field(struct io_kiocb, cmd.data), NULL); | |
a6e959bd KC |
4153 | io_buf_cachep = KMEM_CACHE(io_buffer, |
4154 | SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT); | |
b97f96e2 | 4155 | |
73eaa2b5 JA |
4156 | iou_wq = alloc_workqueue("iou_exit", WQ_UNBOUND, 64); |
4157 | ||
76d3ccec MR |
4158 | #ifdef CONFIG_SYSCTL |
4159 | register_sysctl_init("kernel", kernel_io_uring_disabled_table); | |
4160 | #endif | |
4161 | ||
2b188cc1 JA |
4162 | return 0; |
4163 | }; | |
4164 | __initcall(io_uring_init); |