Commit | Line | Data |
---|---|---|
2b188cc1 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Shared application/kernel submission and completion ring pairs, for | |
4 | * supporting fast/efficient IO. | |
5 | * | |
6 | * A note on the read/write ordering memory barriers that are matched between | |
1e84b97b SB |
7 | * the application and kernel side. |
8 | * | |
9 | * After the application reads the CQ ring tail, it must use an | |
10 | * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses | |
11 | * before writing the tail (using smp_load_acquire to read the tail will | |
12 | * do). It also needs a smp_mb() before updating CQ head (ordering the | |
13 | * entry load(s) with the head store), pairing with an implicit barrier | |
d068b506 | 14 | * through a control-dependency in io_get_cqe (smp_store_release to |
1e84b97b SB |
15 | * store head will do). Failure to do so could lead to reading invalid |
16 | * CQ entries. | |
17 | * | |
18 | * Likewise, the application must use an appropriate smp_wmb() before | |
19 | * writing the SQ tail (ordering SQ entry stores with the tail store), | |
20 | * which pairs with smp_load_acquire in io_get_sqring (smp_store_release | |
21 | * to store the tail will do). And it needs a barrier ordering the SQ | |
22 | * head load before writing new SQ entries (smp_load_acquire to read | |
23 | * head will do). | |
24 | * | |
25 | * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application | |
26 | * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after* | |
27 | * updating the SQ tail; a full memory barrier smp_mb() is needed | |
28 | * between. | |
2b188cc1 JA |
29 | * |
30 | * Also see the examples in the liburing library: | |
31 | * | |
32 | * git://git.kernel.dk/liburing | |
33 | * | |
34 | * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens | |
35 | * from data shared between the kernel and application. This is done both | |
36 | * for ordering purposes, but also to ensure that once a value is loaded from | |
37 | * data that the application could potentially modify, it remains stable. | |
38 | * | |
39 | * Copyright (C) 2018-2019 Jens Axboe | |
c992fe29 | 40 | * Copyright (c) 2018-2019 Christoph Hellwig |
2b188cc1 JA |
41 | */ |
42 | #include <linux/kernel.h> | |
43 | #include <linux/init.h> | |
44 | #include <linux/errno.h> | |
45 | #include <linux/syscalls.h> | |
52de1fe1 | 46 | #include <net/compat.h> |
2b188cc1 JA |
47 | #include <linux/refcount.h> |
48 | #include <linux/uio.h> | |
6b47ee6e | 49 | #include <linux/bits.h> |
2b188cc1 JA |
50 | |
51 | #include <linux/sched/signal.h> | |
52 | #include <linux/fs.h> | |
53 | #include <linux/file.h> | |
54 | #include <linux/fdtable.h> | |
55 | #include <linux/mm.h> | |
56 | #include <linux/mman.h> | |
2b188cc1 JA |
57 | #include <linux/percpu.h> |
58 | #include <linux/slab.h> | |
edafccee | 59 | #include <linux/bvec.h> |
2b188cc1 JA |
60 | #include <linux/net.h> |
61 | #include <net/sock.h> | |
62 | #include <net/af_unix.h> | |
6b06314c | 63 | #include <net/scm.h> |
2b188cc1 JA |
64 | #include <linux/anon_inodes.h> |
65 | #include <linux/sched/mm.h> | |
66 | #include <linux/uaccess.h> | |
67 | #include <linux/nospec.h> | |
aa4c3967 | 68 | #include <linux/highmem.h> |
15b71abe | 69 | #include <linux/fsnotify.h> |
4840e418 | 70 | #include <linux/fadvise.h> |
b41e9852 | 71 | #include <linux/task_work.h> |
0f212204 | 72 | #include <linux/io_uring.h> |
5bd2182d | 73 | #include <linux/audit.h> |
cdc1404a | 74 | #include <linux/security.h> |
d808459b | 75 | #include <asm/shmparam.h> |
2b188cc1 | 76 | |
c826bd7a DD |
77 | #define CREATE_TRACE_POINTS |
78 | #include <trace/events/io_uring.h> | |
79 | ||
2b188cc1 JA |
80 | #include <uapi/linux/io_uring.h> |
81 | ||
561fb04a | 82 | #include "io-wq.h" |
2b188cc1 | 83 | |
de23077e | 84 | #include "io_uring.h" |
329061d3 | 85 | #include "opdef.h" |
e418bbc9 | 86 | #include "refs.h" |
c9f06aa7 | 87 | #include "tctx.h" |
17437f31 | 88 | #include "sqpoll.h" |
a4ad4f74 | 89 | #include "fdinfo.h" |
3b77495a | 90 | #include "kbuf.h" |
73572984 | 91 | #include "rsrc.h" |
38513c46 | 92 | #include "cancel.h" |
43e0bbbd | 93 | #include "net.h" |
eb42cebb | 94 | #include "notif.h" |
f31ecf67 | 95 | #include "waitid.h" |
194bb58c | 96 | #include "futex.h" |
e27f928e | 97 | |
59915143 | 98 | #include "timeout.h" |
329061d3 | 99 | #include "poll.h" |
c92fcfc2 | 100 | #include "rw.h" |
9b797a37 | 101 | #include "alloc_cache.h" |
5e2a18d9 | 102 | |
5277deaa | 103 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 104 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 | 105 | |
21b55dbc SG |
106 | #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \ |
107 | IORING_REGISTER_LAST + IORING_OP_LAST) | |
2b188cc1 | 108 | |
68fe256a PB |
109 | #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \ |
110 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) | |
111 | ||
5562a8d7 PB |
112 | #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \ |
113 | IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS) | |
68fe256a | 114 | |
c854357b | 115 | #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ |
9cae36a0 JA |
116 | REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \ |
117 | REQ_F_ASYNC_DATA) | |
b16fed66 | 118 | |
a538be5b PB |
119 | #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\ |
120 | IO_REQ_CLEAN_FLAGS) | |
121 | ||
09899b19 PB |
122 | #define IO_TCTX_REFS_CACHE_NR (1U << 10) |
123 | ||
6dd0be1e | 124 | #define IO_COMPL_BATCH 32 |
bf019da7 | 125 | #define IO_REQ_ALLOC_BATCH 8 |
258b29a9 | 126 | |
10988a0a DY |
127 | enum { |
128 | IO_CHECK_CQ_OVERFLOW_BIT, | |
155bc950 | 129 | IO_CHECK_CQ_DROPPED_BIT, |
10988a0a DY |
130 | }; |
131 | ||
21a091b9 DY |
132 | enum { |
133 | IO_EVENTFD_OP_SIGNAL_BIT, | |
134 | IO_EVENTFD_OP_FREE_BIT, | |
135 | }; | |
136 | ||
27dc8338 PB |
137 | struct io_defer_entry { |
138 | struct list_head list; | |
139 | struct io_kiocb *req; | |
9cf7c104 | 140 | u32 seq; |
2b188cc1 JA |
141 | }; |
142 | ||
0756a869 PB |
143 | /* requests with any of those set should undergo io_disarm_next() */ |
144 | #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) | |
da1a08c5 | 145 | #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK) |
0756a869 | 146 | |
affa87db | 147 | static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
9936c7c2 | 148 | struct task_struct *task, |
3dd0c97a | 149 | bool cancel_all); |
1ffc5422 | 150 | |
cbc2e203 | 151 | static void io_queue_sqe(struct io_kiocb *req); |
de0617e4 | 152 | |
c1755c25 | 153 | struct kmem_cache *req_cachep; |
2b188cc1 | 154 | |
76d3ccec MR |
155 | static int __read_mostly sysctl_io_uring_disabled; |
156 | static int __read_mostly sysctl_io_uring_group = -1; | |
157 | ||
158 | #ifdef CONFIG_SYSCTL | |
159 | static struct ctl_table kernel_io_uring_disabled_table[] = { | |
160 | { | |
161 | .procname = "io_uring_disabled", | |
162 | .data = &sysctl_io_uring_disabled, | |
163 | .maxlen = sizeof(sysctl_io_uring_disabled), | |
164 | .mode = 0644, | |
165 | .proc_handler = proc_dointvec_minmax, | |
166 | .extra1 = SYSCTL_ZERO, | |
167 | .extra2 = SYSCTL_TWO, | |
168 | }, | |
169 | { | |
170 | .procname = "io_uring_group", | |
171 | .data = &sysctl_io_uring_group, | |
172 | .maxlen = sizeof(gid_t), | |
173 | .mode = 0644, | |
174 | .proc_handler = proc_dointvec, | |
175 | }, | |
176 | {}, | |
177 | }; | |
178 | #endif | |
179 | ||
2b188cc1 JA |
180 | struct sock *io_uring_get_socket(struct file *file) |
181 | { | |
182 | #if defined(CONFIG_UNIX) | |
cd40cae2 | 183 | if (io_is_uring_fops(file)) { |
2b188cc1 JA |
184 | struct io_ring_ctx *ctx = file->private_data; |
185 | ||
186 | return ctx->ring_sock->sk; | |
187 | } | |
188 | #endif | |
189 | return NULL; | |
190 | } | |
191 | EXPORT_SYMBOL(io_uring_get_socket); | |
192 | ||
c450178d PB |
193 | static inline void io_submit_flush_completions(struct io_ring_ctx *ctx) |
194 | { | |
931147dd DY |
195 | if (!wq_list_empty(&ctx->submit_state.compl_reqs) || |
196 | ctx->submit_state.cqes_count) | |
c450178d PB |
197 | __io_submit_flush_completions(ctx); |
198 | } | |
199 | ||
faf88dde PB |
200 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
201 | { | |
202 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
203 | } | |
204 | ||
0fc8c2ac DY |
205 | static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx) |
206 | { | |
207 | return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head); | |
208 | } | |
209 | ||
9cae36a0 JA |
210 | static bool io_match_linked(struct io_kiocb *head) |
211 | { | |
212 | struct io_kiocb *req; | |
213 | ||
214 | io_for_each_link(req, head) { | |
215 | if (req->flags & REQ_F_INFLIGHT) | |
216 | return true; | |
217 | } | |
218 | return false; | |
6af3f48b PB |
219 | } |
220 | ||
221 | /* | |
222 | * As io_match_task() but protected against racing with linked timeouts. | |
223 | * User must not hold timeout_lock. | |
224 | */ | |
329061d3 JA |
225 | bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, |
226 | bool cancel_all) | |
6af3f48b | 227 | { |
9cae36a0 JA |
228 | bool matched; |
229 | ||
6af3f48b PB |
230 | if (task && head->task != task) |
231 | return false; | |
9cae36a0 JA |
232 | if (cancel_all) |
233 | return true; | |
234 | ||
235 | if (head->flags & REQ_F_LINK_TIMEOUT) { | |
236 | struct io_ring_ctx *ctx = head->ctx; | |
237 | ||
238 | /* protect against races with linked timeouts */ | |
239 | spin_lock_irq(&ctx->timeout_lock); | |
240 | matched = io_match_linked(head); | |
241 | spin_unlock_irq(&ctx->timeout_lock); | |
242 | } else { | |
243 | matched = io_match_linked(head); | |
244 | } | |
245 | return matched; | |
6af3f48b PB |
246 | } |
247 | ||
a8295b98 HX |
248 | static inline void req_fail_link_node(struct io_kiocb *req, int res) |
249 | { | |
250 | req_set_fail(req); | |
97b388d7 | 251 | io_req_set_res(req, res, 0); |
a8295b98 HX |
252 | } |
253 | ||
fa05457a PB |
254 | static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) |
255 | { | |
256 | wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); | |
a8295b98 HX |
257 | } |
258 | ||
c072481d | 259 | static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) |
2b188cc1 JA |
260 | { |
261 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
262 | ||
0f158b4c | 263 | complete(&ctx->ref_comp); |
2b188cc1 JA |
264 | } |
265 | ||
c072481d | 266 | static __cold void io_fallback_req_func(struct work_struct *work) |
f56165e6 PB |
267 | { |
268 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
269 | fallback_work.work); | |
270 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
271 | struct io_kiocb *req, *tmp; | |
a282967c | 272 | struct io_tw_state ts = { .locked = true, }; |
f56165e6 | 273 | |
31f084b7 | 274 | mutex_lock(&ctx->uring_lock); |
3218e5d3 | 275 | llist_for_each_entry_safe(req, tmp, node, io_task_work.node) |
a282967c PB |
276 | req->io_task_work.func(req, &ts); |
277 | if (WARN_ON_ONCE(!ts.locked)) | |
31f084b7 PB |
278 | return; |
279 | io_submit_flush_completions(ctx); | |
280 | mutex_unlock(&ctx->uring_lock); | |
f56165e6 PB |
281 | } |
282 | ||
e6f89be6 PB |
283 | static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) |
284 | { | |
285 | unsigned hash_buckets = 1U << bits; | |
286 | size_t hash_size = hash_buckets * sizeof(table->hbs[0]); | |
287 | ||
288 | table->hbs = kmalloc(hash_size, GFP_KERNEL); | |
289 | if (!table->hbs) | |
290 | return -ENOMEM; | |
291 | ||
292 | table->hash_bits = bits; | |
293 | init_hash_table(table, hash_buckets); | |
294 | return 0; | |
295 | } | |
296 | ||
c072481d | 297 | static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
2b188cc1 JA |
298 | { |
299 | struct io_ring_ctx *ctx; | |
9cfc7e94 | 300 | int hash_bits; |
2b188cc1 JA |
301 | |
302 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
303 | if (!ctx) | |
304 | return NULL; | |
305 | ||
9cfc7e94 JA |
306 | xa_init(&ctx->io_bl_xa); |
307 | ||
78076bb6 JA |
308 | /* |
309 | * Use 5 bits less than the max cq entries, that should give us around | |
4a07723f PB |
310 | * 32 entries per hash list if totally full and uniformly spread, but |
311 | * don't keep too many buckets to not overconsume memory. | |
78076bb6 | 312 | */ |
4a07723f PB |
313 | hash_bits = ilog2(p->cq_entries) - 5; |
314 | hash_bits = clamp(hash_bits, 1, 8); | |
e6f89be6 | 315 | if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) |
78076bb6 | 316 | goto err; |
9ca9fb24 PB |
317 | if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) |
318 | goto err; | |
21482896 | 319 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
48904229 | 320 | 0, GFP_KERNEL)) |
206aefde | 321 | goto err; |
2b188cc1 JA |
322 | |
323 | ctx->flags = p->flags; | |
90554200 | 324 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 325 | INIT_LIST_HEAD(&ctx->sqd_list); |
1d7bb1d5 | 326 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
cc3cec83 | 327 | INIT_LIST_HEAD(&ctx->io_buffers_cache); |
69bbc6ad PB |
328 | io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX, |
329 | sizeof(struct io_rsrc_node)); | |
330 | io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX, | |
331 | sizeof(struct async_poll)); | |
332 | io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX, | |
333 | sizeof(struct io_async_msghdr)); | |
194bb58c | 334 | io_futex_cache_init(ctx); |
0f158b4c | 335 | init_completion(&ctx->ref_comp); |
61cf9370 | 336 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 337 | mutex_init(&ctx->uring_lock); |
311997b3 | 338 | init_waitqueue_head(&ctx->cq_wait); |
7b235dd8 | 339 | init_waitqueue_head(&ctx->poll_wq); |
4ea15b56 | 340 | init_waitqueue_head(&ctx->rsrc_quiesce_wq); |
2b188cc1 | 341 | spin_lock_init(&ctx->completion_lock); |
89850fce | 342 | spin_lock_init(&ctx->timeout_lock); |
5eef4e87 | 343 | INIT_WQ_LIST(&ctx->iopoll_list); |
cc3cec83 | 344 | INIT_LIST_HEAD(&ctx->io_buffers_comp); |
de0617e4 | 345 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 346 | INIT_LIST_HEAD(&ctx->timeout_list); |
ef9dd637 | 347 | INIT_LIST_HEAD(&ctx->ltimeout_list); |
d67d2263 | 348 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); |
c0e0d6ba | 349 | init_llist_head(&ctx->work_llist); |
13bf43f5 | 350 | INIT_LIST_HEAD(&ctx->tctx_list); |
c2b6c6bc PB |
351 | ctx->submit_state.free_list.next = NULL; |
352 | INIT_WQ_LIST(&ctx->locked_free_list); | |
f31ecf67 | 353 | INIT_HLIST_HEAD(&ctx->waitid_list); |
194bb58c JA |
354 | #ifdef CONFIG_FUTEX |
355 | INIT_HLIST_HEAD(&ctx->futex_list); | |
356 | #endif | |
9011bf9a | 357 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
6f33b0bc | 358 | INIT_WQ_LIST(&ctx->submit_state.compl_reqs); |
93b8cc60 | 359 | INIT_HLIST_HEAD(&ctx->cancelable_uring_cmd); |
2b188cc1 | 360 | return ctx; |
206aefde | 361 | err: |
e6f89be6 | 362 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 363 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 JA |
364 | kfree(ctx->io_bl); |
365 | xa_destroy(&ctx->io_bl_xa); | |
206aefde JA |
366 | kfree(ctx); |
367 | return NULL; | |
2b188cc1 JA |
368 | } |
369 | ||
8f6ed49a PB |
370 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
371 | { | |
372 | struct io_rings *r = ctx->rings; | |
373 | ||
374 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
375 | ctx->cq_extra--; | |
376 | } | |
377 | ||
9cf7c104 | 378 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 379 | { |
2bc9930e JA |
380 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
381 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 382 | |
8f6ed49a | 383 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 384 | } |
de0617e4 | 385 | |
9d858b21 | 386 | return false; |
de0617e4 JA |
387 | } |
388 | ||
5a754dea PB |
389 | static void io_clean_op(struct io_kiocb *req) |
390 | { | |
391 | if (req->flags & REQ_F_BUFFER_SELECTED) { | |
392 | spin_lock(&req->ctx->completion_lock); | |
393 | io_put_kbuf_comp(req); | |
394 | spin_unlock(&req->ctx->completion_lock); | |
395 | } | |
396 | ||
397 | if (req->flags & REQ_F_NEED_CLEANUP) { | |
398 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; | |
399 | ||
400 | if (def->cleanup) | |
401 | def->cleanup(req); | |
402 | } | |
403 | if ((req->flags & REQ_F_POLLED) && req->apoll) { | |
404 | kfree(req->apoll->double_poll); | |
405 | kfree(req->apoll); | |
406 | req->apoll = NULL; | |
407 | } | |
408 | if (req->flags & REQ_F_INFLIGHT) { | |
409 | struct io_uring_task *tctx = req->task->io_uring; | |
410 | ||
411 | atomic_dec(&tctx->inflight_tracked); | |
412 | } | |
413 | if (req->flags & REQ_F_CREDS) | |
414 | put_cred(req->creds); | |
415 | if (req->flags & REQ_F_ASYNC_DATA) { | |
416 | kfree(req->async_data); | |
417 | req->async_data = NULL; | |
418 | } | |
419 | req->flags &= ~IO_REQ_CLEAN_FLAGS; | |
420 | } | |
421 | ||
9cae36a0 JA |
422 | static inline void io_req_track_inflight(struct io_kiocb *req) |
423 | { | |
424 | if (!(req->flags & REQ_F_INFLIGHT)) { | |
425 | req->flags |= REQ_F_INFLIGHT; | |
386e4fb6 | 426 | atomic_inc(&req->task->io_uring->inflight_tracked); |
9cae36a0 JA |
427 | } |
428 | } | |
429 | ||
fd08e530 PB |
430 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
431 | { | |
906c6caa PB |
432 | if (WARN_ON_ONCE(!req->link)) |
433 | return NULL; | |
434 | ||
4d13d1a4 PB |
435 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
436 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
437 | |
438 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 439 | io_req_set_refcount(req); |
4d13d1a4 PB |
440 | __io_req_set_refcount(req->link, 2); |
441 | return req->link; | |
fd08e530 PB |
442 | } |
443 | ||
444 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
445 | { | |
4d13d1a4 | 446 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
447 | return NULL; |
448 | return __io_prep_linked_timeout(req); | |
449 | } | |
450 | ||
cb2d344c PB |
451 | static noinline void __io_arm_ltimeout(struct io_kiocb *req) |
452 | { | |
453 | io_queue_linked_timeout(__io_prep_linked_timeout(req)); | |
454 | } | |
455 | ||
456 | static inline void io_arm_ltimeout(struct io_kiocb *req) | |
457 | { | |
458 | if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) | |
459 | __io_arm_ltimeout(req); | |
460 | } | |
461 | ||
1e6fa521 JA |
462 | static void io_prep_async_work(struct io_kiocb *req) |
463 | { | |
a7dd2782 | 464 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
1e6fa521 JA |
465 | struct io_ring_ctx *ctx = req->ctx; |
466 | ||
b8e64b53 PB |
467 | if (!(req->flags & REQ_F_CREDS)) { |
468 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 469 | req->creds = get_current_cred(); |
b8e64b53 | 470 | } |
003e8dcc | 471 | |
e1d675df PB |
472 | req->work.list.next = NULL; |
473 | req->work.flags = 0; | |
8e29da69 | 474 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); |
feaadc4f PB |
475 | if (req->flags & REQ_F_FORCE_ASYNC) |
476 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
477 | ||
3beed235 | 478 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
8487f083 | 479 | req->flags |= io_file_get_flags(req->file); |
f6b543fd | 480 | |
8b1df11f | 481 | if (req->file && (req->flags & REQ_F_ISREG)) { |
d4755e15 JA |
482 | bool should_hash = def->hash_reg_file; |
483 | ||
484 | /* don't serialize this request if the fs doesn't need it */ | |
485 | if (should_hash && (req->file->f_flags & O_DIRECT) && | |
486 | (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE)) | |
487 | should_hash = false; | |
488 | if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL)) | |
1e6fa521 | 489 | io_wq_hash_work(&req->work, file_inode(req->file)); |
4b982bd0 | 490 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
491 | if (def->unbound_nonreg_file) |
492 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
493 | } | |
561fb04a | 494 | } |
cccf0ee8 | 495 | |
cbdcb435 | 496 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 497 | { |
cbdcb435 | 498 | struct io_kiocb *cur; |
54a91f3b | 499 | |
44eff40a PB |
500 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
501 | struct io_ring_ctx *ctx = req->ctx; | |
502 | ||
674ee8e1 | 503 | spin_lock_irq(&ctx->timeout_lock); |
44eff40a PB |
504 | io_for_each_link(cur, req) |
505 | io_prep_async_work(cur); | |
674ee8e1 | 506 | spin_unlock_irq(&ctx->timeout_lock); |
44eff40a PB |
507 | } else { |
508 | io_for_each_link(cur, req) | |
509 | io_prep_async_work(cur); | |
510 | } | |
561fb04a JA |
511 | } |
512 | ||
a282967c | 513 | void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use) |
561fb04a | 514 | { |
cbdcb435 | 515 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 516 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 517 | |
3bfe6106 JA |
518 | BUG_ON(!tctx); |
519 | BUG_ON(!tctx->io_wq); | |
561fb04a | 520 | |
cbdcb435 PB |
521 | /* init ->work of the whole link before punting */ |
522 | io_prep_async_link(req); | |
991468dc JA |
523 | |
524 | /* | |
525 | * Not expected to happen, but if we do have a bug where this _can_ | |
526 | * happen, catch it here and ensure the request is marked as | |
527 | * canceled. That will make io-wq go through the usual work cancel | |
528 | * procedure rather than attempt to run this request (or create a new | |
529 | * worker for it). | |
530 | */ | |
531 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
532 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
533 | ||
48863ffd | 534 | trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); |
ebf93667 | 535 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
536 | if (link) |
537 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
538 | } |
539 | ||
c072481d | 540 | static __cold void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 541 | { |
441b8a78 | 542 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
543 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
544 | struct io_defer_entry, list); | |
de0617e4 | 545 | |
9cf7c104 | 546 | if (req_need_defer(de->req, de->seq)) |
04518945 | 547 | break; |
27dc8338 | 548 | list_del_init(&de->list); |
907d1df3 | 549 | io_req_task_queue(de->req); |
27dc8338 | 550 | kfree(de); |
441b8a78 | 551 | } |
04518945 PB |
552 | } |
553 | ||
21a091b9 DY |
554 | |
555 | static void io_eventfd_ops(struct rcu_head *rcu) | |
f2842ab5 | 556 | { |
d8e9214f | 557 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
21a091b9 | 558 | int ops = atomic_xchg(&ev_fd->ops, 0); |
305bef98 | 559 | |
21a091b9 | 560 | if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) |
44648532 | 561 | eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); |
d8e9214f | 562 | |
21a091b9 DY |
563 | /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback |
564 | * ordering in a race but if references are 0 we know we have to free | |
565 | * it regardless. | |
305bef98 | 566 | */ |
21a091b9 DY |
567 | if (atomic_dec_and_test(&ev_fd->refs)) { |
568 | eventfd_ctx_put(ev_fd->cq_ev_fd); | |
569 | kfree(ev_fd); | |
570 | } | |
d8e9214f DY |
571 | } |
572 | ||
77bc59b4 | 573 | static void io_eventfd_signal(struct io_ring_ctx *ctx) |
f2842ab5 | 574 | { |
21a091b9 | 575 | struct io_ev_fd *ev_fd = NULL; |
77bc59b4 | 576 | |
77bc59b4 UA |
577 | rcu_read_lock(); |
578 | /* | |
579 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
580 | * and eventfd_signal | |
581 | */ | |
582 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
583 | ||
584 | /* | |
585 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
586 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
587 | * the function and rcu_read_lock. | |
588 | */ | |
589 | if (unlikely(!ev_fd)) | |
590 | goto out; | |
7e55a19c | 591 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 | 592 | goto out; |
21a091b9 DY |
593 | if (ev_fd->eventfd_async && !io_wq_current_is_worker()) |
594 | goto out; | |
77bc59b4 | 595 | |
21a091b9 | 596 | if (likely(eventfd_signal_allowed())) { |
44648532 | 597 | eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); |
21a091b9 DY |
598 | } else { |
599 | atomic_inc(&ev_fd->refs); | |
600 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) | |
44a84da4 | 601 | call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops); |
21a091b9 DY |
602 | else |
603 | atomic_dec(&ev_fd->refs); | |
604 | } | |
605 | ||
77bc59b4 UA |
606 | out: |
607 | rcu_read_unlock(); | |
f2842ab5 JA |
608 | } |
609 | ||
21a091b9 DY |
610 | static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) |
611 | { | |
612 | bool skip; | |
613 | ||
614 | spin_lock(&ctx->completion_lock); | |
615 | ||
616 | /* | |
617 | * Eventfd should only get triggered when at least one event has been | |
618 | * posted. Some applications rely on the eventfd notification count | |
619 | * only changing IFF a new CQE has been added to the CQ ring. There's | |
620 | * no depedency on 1:1 relationship between how many times this | |
621 | * function is called (and hence the eventfd count) and number of CQEs | |
622 | * posted to the CQ ring. | |
623 | */ | |
624 | skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
625 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
626 | spin_unlock(&ctx->completion_lock); | |
627 | if (skip) | |
628 | return; | |
629 | ||
630 | io_eventfd_signal(ctx); | |
631 | } | |
632 | ||
a830ffd2 PB |
633 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
634 | { | |
bca39f39 PB |
635 | if (ctx->poll_activated) |
636 | io_poll_wq_wake(ctx); | |
e5f30f6f PB |
637 | if (ctx->off_timeout_used) |
638 | io_flush_timeouts(ctx); | |
639 | if (ctx->drain_active) { | |
a830ffd2 | 640 | spin_lock(&ctx->completion_lock); |
e5f30f6f | 641 | io_queue_deferred(ctx); |
a830ffd2 PB |
642 | spin_unlock(&ctx->completion_lock); |
643 | } | |
644 | if (ctx->has_evfd) | |
21a091b9 | 645 | io_eventfd_flush_signal(ctx); |
a830ffd2 PB |
646 | } |
647 | ||
f66f7342 | 648 | static inline void __io_cq_lock(struct io_ring_ctx *ctx) |
f66f7342 | 649 | { |
ec26c225 | 650 | if (!ctx->lockless_cq) |
f66f7342 PB |
651 | spin_lock(&ctx->completion_lock); |
652 | } | |
653 | ||
6971253f PB |
654 | static inline void io_cq_lock(struct io_ring_ctx *ctx) |
655 | __acquires(ctx->completion_lock) | |
656 | { | |
657 | spin_lock(&ctx->completion_lock); | |
658 | } | |
659 | ||
f66f7342 | 660 | static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) |
3181e22f PB |
661 | { |
662 | io_commit_cqring(ctx); | |
54927baf | 663 | if (!ctx->task_complete) { |
ec26c225 PB |
664 | if (!ctx->lockless_cq) |
665 | spin_unlock(&ctx->completion_lock); | |
666 | /* IOPOLL rings only need to wake up if it's also SQPOLL */ | |
667 | if (!ctx->syscall_iopoll) | |
668 | io_cqring_wake(ctx); | |
3181e22f | 669 | } |
54927baf | 670 | io_commit_cqring_flush(ctx); |
3181e22f PB |
671 | } |
672 | ||
0fdb9a19 | 673 | static void io_cq_unlock_post(struct io_ring_ctx *ctx) |
5d772916 | 674 | __releases(ctx->completion_lock) |
25399321 | 675 | { |
f66f7342 PB |
676 | io_commit_cqring(ctx); |
677 | spin_unlock(&ctx->completion_lock); | |
f66f7342 | 678 | io_cqring_wake(ctx); |
54927baf | 679 | io_commit_cqring_flush(ctx); |
25399321 PB |
680 | } |
681 | ||
c4a2ed72 | 682 | /* Returns true if there are no backlogged entries after the flush */ |
a85381d8 PB |
683 | static void io_cqring_overflow_kill(struct io_ring_ctx *ctx) |
684 | { | |
685 | struct io_overflow_cqe *ocqe; | |
686 | LIST_HEAD(list); | |
687 | ||
f432b76b | 688 | spin_lock(&ctx->completion_lock); |
a85381d8 PB |
689 | list_splice_init(&ctx->cq_overflow_list, &list); |
690 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); | |
f432b76b | 691 | spin_unlock(&ctx->completion_lock); |
a85381d8 PB |
692 | |
693 | while (!list_empty(&list)) { | |
694 | ocqe = list_first_entry(&list, struct io_overflow_cqe, list); | |
695 | list_del(&ocqe->list); | |
696 | kfree(ocqe); | |
697 | } | |
698 | } | |
699 | ||
1b346e4a | 700 | static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
1d7bb1d5 | 701 | { |
e45a3e05 | 702 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 703 | |
a85381d8 | 704 | if (__io_cqring_events(ctx) == ctx->cq_entries) |
1b346e4a | 705 | return; |
1d7bb1d5 | 706 | |
e45a3e05 SR |
707 | if (ctx->flags & IORING_SETUP_CQE32) |
708 | cqe_size <<= 1; | |
709 | ||
25399321 | 710 | io_cq_lock(ctx); |
6c2450ae | 711 | while (!list_empty(&ctx->cq_overflow_list)) { |
59fbc409 | 712 | struct io_uring_cqe *cqe; |
6c2450ae | 713 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 714 | |
59fbc409 | 715 | if (!io_get_cqe_overflow(ctx, &cqe, true)) |
1d7bb1d5 | 716 | break; |
6c2450ae PB |
717 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
718 | struct io_overflow_cqe, list); | |
a85381d8 | 719 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae PB |
720 | list_del(&ocqe->list); |
721 | kfree(ocqe); | |
1d7bb1d5 JA |
722 | } |
723 | ||
1b346e4a | 724 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 725 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 726 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 727 | } |
25399321 | 728 | io_cq_unlock_post(ctx); |
1d7bb1d5 JA |
729 | } |
730 | ||
52ea806a JA |
731 | static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx) |
732 | { | |
733 | /* iopoll syncs against uring_lock, not completion_lock */ | |
734 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
735 | mutex_lock(&ctx->uring_lock); | |
736 | __io_cqring_overflow_flush(ctx); | |
737 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
738 | mutex_unlock(&ctx->uring_lock); | |
739 | } | |
740 | ||
1b346e4a | 741 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 742 | { |
52ea806a JA |
743 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) |
744 | io_cqring_do_overflow_flush(ctx); | |
6c503150 PB |
745 | } |
746 | ||
5afa4650 | 747 | /* can be called by any task */ |
2fdd6fb5 | 748 | static void io_put_task_remote(struct task_struct *task) |
6a290a14 PB |
749 | { |
750 | struct io_uring_task *tctx = task->io_uring; | |
751 | ||
2fdd6fb5 | 752 | percpu_counter_sub(&tctx->inflight, 1); |
8d664282 | 753 | if (unlikely(atomic_read(&tctx->in_cancel))) |
9d170164 | 754 | wake_up(&tctx->wait); |
2fdd6fb5 | 755 | put_task_struct(task); |
9d170164 PB |
756 | } |
757 | ||
5afa4650 | 758 | /* used by a task to put its own references */ |
2fdd6fb5 | 759 | static void io_put_task_local(struct task_struct *task) |
5afa4650 | 760 | { |
2fdd6fb5 | 761 | task->io_uring->cached_refs++; |
5afa4650 PB |
762 | } |
763 | ||
89800a2d | 764 | /* must to be called somewhat shortly after putting a request */ |
2fdd6fb5 | 765 | static inline void io_put_task(struct task_struct *task) |
89800a2d PB |
766 | { |
767 | if (likely(task == current)) | |
2fdd6fb5 | 768 | io_put_task_local(task); |
89800a2d | 769 | else |
2fdd6fb5 | 770 | io_put_task_remote(task); |
89800a2d PB |
771 | } |
772 | ||
63809137 | 773 | void io_task_refs_refill(struct io_uring_task *tctx) |
9a10867a PB |
774 | { |
775 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
776 | ||
777 | percpu_counter_add(&tctx->inflight, refill); | |
778 | refcount_add(refill, ¤t->usage); | |
779 | tctx->cached_refs += refill; | |
780 | } | |
781 | ||
3cc7fdb9 PB |
782 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
783 | { | |
784 | struct io_uring_task *tctx = task->io_uring; | |
785 | unsigned int refs = tctx->cached_refs; | |
786 | ||
787 | if (refs) { | |
788 | tctx->cached_refs = 0; | |
789 | percpu_counter_sub(&tctx->inflight, refs); | |
790 | put_task_struct_many(task, refs); | |
791 | } | |
792 | } | |
793 | ||
68494a65 PB |
794 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
795 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 796 | { |
cce4b8b0 | 797 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
798 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
799 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 800 | |
f26cc959 PB |
801 | lockdep_assert_held(&ctx->completion_lock); |
802 | ||
e45a3e05 SR |
803 | if (is_cqe32) |
804 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 805 | |
e45a3e05 | 806 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 807 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
808 | if (!ocqe) { |
809 | /* | |
810 | * If we're in ring overflow flush mode, or in task cancel mode, | |
811 | * or cannot allocate an overflow entry, then we need to drop it | |
812 | * on the floor. | |
813 | */ | |
8f6ed49a | 814 | io_account_cq_overflow(ctx); |
155bc950 | 815 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 816 | return false; |
2b188cc1 | 817 | } |
cce4b8b0 | 818 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 819 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 820 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 821 | |
cce4b8b0 | 822 | } |
d4d19c19 | 823 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
824 | ocqe->cqe.res = res; |
825 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
826 | if (is_cqe32) { |
827 | ocqe->cqe.big_cqe[0] = extra1; | |
828 | ocqe->cqe.big_cqe[1] = extra2; | |
829 | } | |
cce4b8b0 PB |
830 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
831 | return true; | |
2b188cc1 JA |
832 | } |
833 | ||
056695bf | 834 | void io_req_cqe_overflow(struct io_kiocb *req) |
68494a65 | 835 | { |
056695bf PB |
836 | io_cqring_event_overflow(req->ctx, req->cqe.user_data, |
837 | req->cqe.res, req->cqe.flags, | |
b24c5d75 PB |
838 | req->big_cqe.extra1, req->big_cqe.extra2); |
839 | memset(&req->big_cqe, 0, sizeof(req->big_cqe)); | |
68494a65 PB |
840 | } |
841 | ||
faf88dde PB |
842 | /* |
843 | * writes to the cq entry need to come after reading head; the | |
844 | * control dependency is enough as we're using WRITE_ONCE to | |
845 | * fill the cq entry | |
846 | */ | |
20d6b633 | 847 | bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow) |
faf88dde PB |
848 | { |
849 | struct io_rings *rings = ctx->rings; | |
850 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
851 | unsigned int free, queued, len; |
852 | ||
aa1df3a3 PB |
853 | /* |
854 | * Posting into the CQ when there are pending overflowed CQEs may break | |
855 | * ordering guarantees, which will affect links, F_MORE users and more. | |
856 | * Force overflow the completion. | |
857 | */ | |
858 | if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) | |
20d6b633 | 859 | return false; |
faf88dde PB |
860 | |
861 | /* userspace may cheat modifying the tail, be safe and do min */ | |
862 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
863 | free = ctx->cq_entries - queued; | |
864 | /* we need a contiguous range, limit based on the current array offset */ | |
865 | len = min(free, ctx->cq_entries - off); | |
866 | if (!len) | |
20d6b633 | 867 | return false; |
faf88dde | 868 | |
b3659a65 PB |
869 | if (ctx->flags & IORING_SETUP_CQE32) { |
870 | off <<= 1; | |
871 | len <<= 1; | |
872 | } | |
873 | ||
faf88dde PB |
874 | ctx->cqe_cached = &rings->cqes[off]; |
875 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
20d6b633 | 876 | return true; |
faf88dde PB |
877 | } |
878 | ||
f66f7342 PB |
879 | static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, |
880 | u32 cflags) | |
bcda7baa | 881 | { |
cd94903d PB |
882 | struct io_uring_cqe *cqe; |
883 | ||
913a571a | 884 | ctx->cq_extra++; |
cd94903d PB |
885 | |
886 | /* | |
887 | * If we can't get a cq entry, userspace overflowed the | |
888 | * submission (by quite a lot). Increment the overflow count in | |
889 | * the ring. | |
890 | */ | |
59fbc409 | 891 | if (likely(io_get_cqe(ctx, &cqe))) { |
e0486f3f DY |
892 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
893 | ||
cd94903d PB |
894 | WRITE_ONCE(cqe->user_data, user_data); |
895 | WRITE_ONCE(cqe->res, res); | |
896 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
897 | |
898 | if (ctx->flags & IORING_SETUP_CQE32) { | |
899 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
900 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
901 | } | |
cd94903d PB |
902 | return true; |
903 | } | |
52120f0f | 904 | return false; |
bcda7baa JA |
905 | } |
906 | ||
931147dd DY |
907 | static void __io_flush_post_cqes(struct io_ring_ctx *ctx) |
908 | __must_hold(&ctx->uring_lock) | |
909 | { | |
910 | struct io_submit_state *state = &ctx->submit_state; | |
911 | unsigned int i; | |
912 | ||
913 | lockdep_assert_held(&ctx->uring_lock); | |
914 | for (i = 0; i < state->cqes_count; i++) { | |
0aa7aa5f | 915 | struct io_uring_cqe *cqe = &ctx->completion_cqes[i]; |
931147dd | 916 | |
f66f7342 | 917 | if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) { |
27122c07 | 918 | if (ctx->lockless_cq) { |
f66f7342 PB |
919 | spin_lock(&ctx->completion_lock); |
920 | io_cqring_event_overflow(ctx, cqe->user_data, | |
921 | cqe->res, cqe->flags, 0, 0); | |
922 | spin_unlock(&ctx->completion_lock); | |
923 | } else { | |
924 | io_cqring_event_overflow(ctx, cqe->user_data, | |
925 | cqe->res, cqe->flags, 0, 0); | |
926 | } | |
927 | } | |
931147dd DY |
928 | } |
929 | state->cqes_count = 0; | |
930 | } | |
931 | ||
b529c96a DY |
932 | static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, |
933 | bool allow_overflow) | |
d245bca6 PB |
934 | { |
935 | bool filled; | |
936 | ||
25399321 | 937 | io_cq_lock(ctx); |
f66f7342 PB |
938 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags); |
939 | if (!filled && allow_overflow) | |
940 | filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); | |
941 | ||
25399321 | 942 | io_cq_unlock_post(ctx); |
d245bca6 PB |
943 | return filled; |
944 | } | |
945 | ||
b529c96a DY |
946 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags) |
947 | { | |
948 | return __io_post_aux_cqe(ctx, user_data, res, cflags, true); | |
949 | } | |
950 | ||
b6b2bb58 PB |
951 | /* |
952 | * A helper for multishot requests posting additional CQEs. | |
953 | * Should only be used from a task_work including IO_URING_F_MULTISHOT. | |
954 | */ | |
955 | bool io_fill_cqe_req_aux(struct io_kiocb *req, bool defer, s32 res, u32 cflags) | |
2b188cc1 | 956 | { |
d86eaed1 JA |
957 | struct io_ring_ctx *ctx = req->ctx; |
958 | u64 user_data = req->cqe.user_data; | |
9b8c5475 | 959 | struct io_uring_cqe *cqe; |
9b8c5475 DY |
960 | |
961 | if (!defer) | |
b6b2bb58 | 962 | return __io_post_aux_cqe(ctx, user_data, res, cflags, false); |
9b8c5475 | 963 | |
9b8c5475 DY |
964 | lockdep_assert_held(&ctx->uring_lock); |
965 | ||
0aa7aa5f | 966 | if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->completion_cqes)) { |
f66f7342 | 967 | __io_cq_lock(ctx); |
9b8c5475 DY |
968 | __io_flush_post_cqes(ctx); |
969 | /* no need to flush - flush is deferred */ | |
f66f7342 | 970 | __io_cq_unlock_post(ctx); |
9b8c5475 DY |
971 | } |
972 | ||
973 | /* For defered completions this is not as strict as it is otherwise, | |
974 | * however it's main job is to prevent unbounded posted completions, | |
975 | * and in that it works just as well. | |
976 | */ | |
b6b2bb58 | 977 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) |
9b8c5475 DY |
978 | return false; |
979 | ||
0aa7aa5f | 980 | cqe = &ctx->completion_cqes[ctx->submit_state.cqes_count++]; |
9b8c5475 DY |
981 | cqe->user_data = user_data; |
982 | cqe->res = res; | |
983 | cqe->flags = cflags; | |
984 | return true; | |
985 | } | |
986 | ||
ef8ae64f | 987 | static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
2b188cc1 | 988 | { |
fa18fa22 | 989 | struct io_ring_ctx *ctx = req->ctx; |
2ad4c6d0 | 990 | struct io_rsrc_node *rsrc_node = NULL; |
fa18fa22 PB |
991 | |
992 | io_cq_lock(ctx); | |
00b0db56 PB |
993 | if (!(req->flags & REQ_F_CQE_SKIP)) { |
994 | if (!io_fill_cqe_req(ctx, req)) | |
995 | io_req_cqe_overflow(req); | |
996 | } | |
fa18fa22 | 997 | |
c7dae4ba JA |
998 | /* |
999 | * If we're the last reference to this request, add to our locked | |
1000 | * free_list cache. | |
1001 | */ | |
de9b4cca | 1002 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 1003 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 1004 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
1005 | io_disarm_next(req); |
1006 | if (req->link) { | |
1007 | io_req_task_queue(req->link); | |
1008 | req->link = NULL; | |
1009 | } | |
1010 | } | |
68a2cc1b | 1011 | io_put_kbuf_comp(req); |
3b7a612f PB |
1012 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1013 | io_clean_op(req); | |
17bc2837 | 1014 | io_put_file(req); |
3b7a612f | 1015 | |
2ad4c6d0 | 1016 | rsrc_node = req->rsrc_node; |
8197b053 PB |
1017 | /* |
1018 | * Selected buffer deallocation in io_clean_op() assumes that | |
1019 | * we don't hold ->completion_lock. Clean them here to avoid | |
1020 | * deadlocks. | |
1021 | */ | |
2fdd6fb5 | 1022 | io_put_task_remote(req->task); |
c2b6c6bc | 1023 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 1024 | ctx->locked_free_nr++; |
180f829f | 1025 | } |
25399321 | 1026 | io_cq_unlock_post(ctx); |
2ad4c6d0 | 1027 | |
ef8ae64f PB |
1028 | if (rsrc_node) { |
1029 | io_ring_submit_lock(ctx, issue_flags); | |
1f2c8f61 | 1030 | io_put_rsrc_node(ctx, rsrc_node); |
ef8ae64f PB |
1031 | io_ring_submit_unlock(ctx, issue_flags); |
1032 | } | |
4e3d9ff9 JA |
1033 | } |
1034 | ||
1bec951c | 1035 | void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
bcda7baa | 1036 | { |
860e1c7f | 1037 | if (req->ctx->task_complete && req->ctx->submitter_task != current) { |
e6aeb272 PB |
1038 | req->io_task_work.func = io_req_task_complete; |
1039 | io_req_task_work_add(req); | |
1040 | } else if (!(issue_flags & IO_URING_F_UNLOCKED) || | |
1041 | !(req->ctx->flags & IORING_SETUP_IOPOLL)) { | |
ef8ae64f | 1042 | __io_req_complete_post(req, issue_flags); |
1bec951c PB |
1043 | } else { |
1044 | struct io_ring_ctx *ctx = req->ctx; | |
1045 | ||
1046 | mutex_lock(&ctx->uring_lock); | |
ef8ae64f | 1047 | __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED); |
1bec951c PB |
1048 | mutex_unlock(&ctx->uring_lock); |
1049 | } | |
0ddf92e8 JA |
1050 | } |
1051 | ||
973fc83f | 1052 | void io_req_defer_failed(struct io_kiocb *req, s32 res) |
e276ae34 | 1053 | __must_hold(&ctx->uring_lock) |
f41db273 | 1054 | { |
f30bd4d0 | 1055 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; |
a47b255e | 1056 | |
e276ae34 PB |
1057 | lockdep_assert_held(&req->ctx->uring_lock); |
1058 | ||
93d2bcd2 | 1059 | req_set_fail(req); |
97b388d7 | 1060 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
a47b255e PB |
1061 | if (def->fail) |
1062 | def->fail(req); | |
973fc83f | 1063 | io_req_complete_defer(req); |
f41db273 PB |
1064 | } |
1065 | ||
864ea921 PB |
1066 | /* |
1067 | * Don't initialise the fields below on every allocation, but do that in | |
1068 | * advance and keep them valid across allocations. | |
1069 | */ | |
1070 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
1071 | { | |
1072 | req->ctx = ctx; | |
1073 | req->link = NULL; | |
1074 | req->async_data = NULL; | |
1075 | /* not necessary, but safer to zero */ | |
31d3ba92 | 1076 | memset(&req->cqe, 0, sizeof(req->cqe)); |
b24c5d75 | 1077 | memset(&req->big_cqe, 0, sizeof(req->big_cqe)); |
864ea921 PB |
1078 | } |
1079 | ||
dac7a098 | 1080 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 1081 | struct io_submit_state *state) |
dac7a098 | 1082 | { |
79ebeaee | 1083 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1084 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 1085 | ctx->locked_free_nr = 0; |
79ebeaee | 1086 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
1087 | } |
1088 | ||
5d5901a3 PB |
1089 | /* |
1090 | * A request might get retired back into the request caches even before opcode | |
1091 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
1092 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
1093 | * and with extra caution to not get a request that is still worked on. | |
1094 | */ | |
bd1a3783 | 1095 | __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 1096 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 1097 | { |
864ea921 | 1098 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 1099 | void *reqs[IO_REQ_ALLOC_BATCH]; |
864ea921 | 1100 | int ret, i; |
e5d1bc0a | 1101 | |
23a5c43b PB |
1102 | /* |
1103 | * If we have more than a batch's worth of requests in our IRQ side | |
1104 | * locked cache, grab the lock and move them over to our submission | |
1105 | * side cache. | |
1106 | */ | |
a6d97a8a | 1107 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 1108 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 1109 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
1110 | return true; |
1111 | } | |
e5d1bc0a | 1112 | |
3ab665b7 | 1113 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 1114 | |
864ea921 PB |
1115 | /* |
1116 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1117 | * retry single alloc to be on the safe side. | |
1118 | */ | |
1119 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
1120 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
1121 | if (!reqs[0]) | |
a33ae9ce | 1122 | return false; |
864ea921 | 1123 | ret = 1; |
2b188cc1 | 1124 | } |
864ea921 | 1125 | |
37f0e767 | 1126 | percpu_ref_get_many(&ctx->refs, ret); |
3ab665b7 | 1127 | for (i = 0; i < ret; i++) { |
23a5c43b | 1128 | struct io_kiocb *req = reqs[i]; |
3ab665b7 PB |
1129 | |
1130 | io_preinit_req(req, ctx); | |
fa05457a | 1131 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 1132 | } |
a33ae9ce PB |
1133 | return true; |
1134 | } | |
1135 | ||
03adabe8 PB |
1136 | __cold void io_free_req(struct io_kiocb *req) |
1137 | { | |
6ec9afc7 PB |
1138 | /* refs were already put, restore them for io_req_task_complete() */ |
1139 | req->flags &= ~REQ_F_REFCOUNT; | |
1140 | /* we only want to free it, don't post CQEs */ | |
1141 | req->flags |= REQ_F_CQE_SKIP; | |
1142 | req->io_task_work.func = io_req_task_complete; | |
03adabe8 PB |
1143 | io_req_task_work_add(req); |
1144 | } | |
1145 | ||
d81499bf PB |
1146 | static void __io_req_find_next_prep(struct io_kiocb *req) |
1147 | { | |
1148 | struct io_ring_ctx *ctx = req->ctx; | |
d81499bf | 1149 | |
6971253f | 1150 | spin_lock(&ctx->completion_lock); |
305bef98 | 1151 | io_disarm_next(req); |
6971253f | 1152 | spin_unlock(&ctx->completion_lock); |
d81499bf PB |
1153 | } |
1154 | ||
1155 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 1156 | { |
33cc89a9 | 1157 | struct io_kiocb *nxt; |
944e58bf | 1158 | |
9e645e11 JA |
1159 | /* |
1160 | * If LINK is set, we have dependent requests in this chain. If we | |
1161 | * didn't fail this request, queue the first one up, moving any other | |
1162 | * dependencies to the next request. In case of failure, fail the rest | |
1163 | * of the chain. | |
1164 | */ | |
d81499bf PB |
1165 | if (unlikely(req->flags & IO_DISARM_MASK)) |
1166 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
1167 | nxt = req->link; |
1168 | req->link = NULL; | |
1169 | return nxt; | |
4d7dd462 | 1170 | } |
9e645e11 | 1171 | |
a282967c | 1172 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts) |
2c32395d PB |
1173 | { |
1174 | if (!ctx) | |
1175 | return; | |
ef060ea9 JA |
1176 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1177 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
a282967c | 1178 | if (ts->locked) { |
c450178d | 1179 | io_submit_flush_completions(ctx); |
2c32395d | 1180 | mutex_unlock(&ctx->uring_lock); |
a282967c | 1181 | ts->locked = false; |
2c32395d PB |
1182 | } |
1183 | percpu_ref_put(&ctx->refs); | |
1184 | } | |
1185 | ||
c6dd763c | 1186 | static unsigned int handle_tw_list(struct llist_node *node, |
a282967c PB |
1187 | struct io_ring_ctx **ctx, |
1188 | struct io_tw_state *ts, | |
c6dd763c | 1189 | struct llist_node *last) |
9f8d032a | 1190 | { |
c6dd763c DY |
1191 | unsigned int count = 0; |
1192 | ||
cb6bf7f2 | 1193 | while (node && node != last) { |
f88262e6 | 1194 | struct llist_node *next = node->next; |
9f8d032a HX |
1195 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1196 | io_task_work.node); | |
1197 | ||
34d2bfe7 JA |
1198 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); |
1199 | ||
9f8d032a | 1200 | if (req->ctx != *ctx) { |
a282967c | 1201 | ctx_flush_and_put(*ctx, ts); |
9f8d032a HX |
1202 | *ctx = req->ctx; |
1203 | /* if not contended, grab and improve batching */ | |
a282967c | 1204 | ts->locked = mutex_trylock(&(*ctx)->uring_lock); |
9f8d032a | 1205 | percpu_ref_get(&(*ctx)->refs); |
13bfa6f1 | 1206 | } |
c92fcfc2 JA |
1207 | INDIRECT_CALL_2(req->io_task_work.func, |
1208 | io_poll_task_func, io_req_rw_complete, | |
1209 | req, ts); | |
9f8d032a | 1210 | node = next; |
c6dd763c | 1211 | count++; |
f5868008 | 1212 | if (unlikely(need_resched())) { |
a282967c | 1213 | ctx_flush_and_put(*ctx, ts); |
f5868008 JA |
1214 | *ctx = NULL; |
1215 | cond_resched(); | |
1216 | } | |
3a0c037b | 1217 | } |
c6dd763c DY |
1218 | |
1219 | return count; | |
9f8d032a HX |
1220 | } |
1221 | ||
923d1592 DY |
1222 | /** |
1223 | * io_llist_xchg - swap all entries in a lock-less list | |
1224 | * @head: the head of lock-less list to delete all entries | |
1225 | * @new: new entry as the head of the list | |
1226 | * | |
1227 | * If list is empty, return NULL, otherwise, return the pointer to the first entry. | |
1228 | * The order of entries returned is from the newest to the oldest added one. | |
1229 | */ | |
1230 | static inline struct llist_node *io_llist_xchg(struct llist_head *head, | |
1231 | struct llist_node *new) | |
1232 | { | |
1233 | return xchg(&head->first, new); | |
1234 | } | |
1235 | ||
1236 | /** | |
1237 | * io_llist_cmpxchg - possibly swap all entries in a lock-less list | |
1238 | * @head: the head of lock-less list to delete all entries | |
1239 | * @old: expected old value of the first entry of the list | |
1240 | * @new: new entry as the head of the list | |
1241 | * | |
1242 | * perform a cmpxchg on the first entry of the list. | |
1243 | */ | |
1244 | ||
1245 | static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head, | |
1246 | struct llist_node *old, | |
1247 | struct llist_node *new) | |
1248 | { | |
1249 | return cmpxchg(&head->first, old, new); | |
1250 | } | |
1251 | ||
dfbe5561 | 1252 | static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync) |
10e1c0d5 JA |
1253 | { |
1254 | struct llist_node *node = llist_del_all(&tctx->task_list); | |
dfbe5561 | 1255 | struct io_ring_ctx *last_ctx = NULL; |
10e1c0d5 JA |
1256 | struct io_kiocb *req; |
1257 | ||
1258 | while (node) { | |
1259 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1260 | node = node->next; | |
dfbe5561 JA |
1261 | if (sync && last_ctx != req->ctx) { |
1262 | if (last_ctx) { | |
1263 | flush_delayed_work(&last_ctx->fallback_work); | |
1264 | percpu_ref_put(&last_ctx->refs); | |
1265 | } | |
1266 | last_ctx = req->ctx; | |
1267 | percpu_ref_get(&last_ctx->refs); | |
1268 | } | |
10e1c0d5 JA |
1269 | if (llist_add(&req->io_task_work.node, |
1270 | &req->ctx->fallback_llist)) | |
1271 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1272 | } | |
dfbe5561 JA |
1273 | |
1274 | if (last_ctx) { | |
1275 | flush_delayed_work(&last_ctx->fallback_work); | |
1276 | percpu_ref_put(&last_ctx->refs); | |
1277 | } | |
10e1c0d5 JA |
1278 | } |
1279 | ||
c9f06aa7 | 1280 | void tctx_task_work(struct callback_head *cb) |
c40f6379 | 1281 | { |
a282967c | 1282 | struct io_tw_state ts = {}; |
ebd0df2e | 1283 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
1284 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
1285 | task_work); | |
3a0c037b | 1286 | struct llist_node fake = {}; |
77e443ab | 1287 | struct llist_node *node; |
cb6bf7f2 PB |
1288 | unsigned int loops = 0; |
1289 | unsigned int count = 0; | |
3a0c037b | 1290 | |
77e443ab | 1291 | if (unlikely(current->flags & PF_EXITING)) { |
dfbe5561 | 1292 | io_fallback_tw(tctx, true); |
77e443ab PB |
1293 | return; |
1294 | } | |
3a0c037b | 1295 | |
cb6bf7f2 | 1296 | do { |
c6dd763c | 1297 | loops++; |
3a0c037b | 1298 | node = io_llist_xchg(&tctx->task_list, &fake); |
a282967c | 1299 | count += handle_tw_list(node, &ctx, &ts, &fake); |
50470fc5 PB |
1300 | |
1301 | /* skip expensive cmpxchg if there are items in the list */ | |
1302 | if (READ_ONCE(tctx->task_list.first) != &fake) | |
1303 | continue; | |
a282967c | 1304 | if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) { |
50470fc5 PB |
1305 | io_submit_flush_completions(ctx); |
1306 | if (READ_ONCE(tctx->task_list.first) != &fake) | |
1307 | continue; | |
1308 | } | |
3a0c037b | 1309 | node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); |
cb6bf7f2 | 1310 | } while (node != &fake); |
ebd0df2e | 1311 | |
a282967c | 1312 | ctx_flush_and_put(ctx, &ts); |
3cc7fdb9 | 1313 | |
8d664282 JA |
1314 | /* relaxed read is enough as only the task itself sets ->in_cancel */ |
1315 | if (unlikely(atomic_read(&tctx->in_cancel))) | |
3cc7fdb9 | 1316 | io_uring_drop_tctx_refs(current); |
c6dd763c DY |
1317 | |
1318 | trace_io_uring_task_work_run(tctx, count, loops); | |
7cbf1722 JA |
1319 | } |
1320 | ||
91c7884a | 1321 | static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags) |
c0e0d6ba DY |
1322 | { |
1323 | struct io_ring_ctx *ctx = req->ctx; | |
8751d154 | 1324 | unsigned nr_wait, nr_tw, nr_tw_prev; |
51509400 | 1325 | struct llist_node *first; |
c0e0d6ba | 1326 | |
8751d154 PB |
1327 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) |
1328 | flags &= ~IOU_F_TWQ_LAZY_WAKE; | |
ce8e04f6 | 1329 | |
51509400 PB |
1330 | first = READ_ONCE(ctx->work_llist.first); |
1331 | do { | |
8751d154 PB |
1332 | nr_tw_prev = 0; |
1333 | if (first) { | |
1334 | struct io_kiocb *first_req = container_of(first, | |
1335 | struct io_kiocb, | |
1336 | io_task_work.node); | |
1337 | /* | |
1338 | * Might be executed at any moment, rely on | |
1339 | * SLAB_TYPESAFE_BY_RCU to keep it alive. | |
1340 | */ | |
1341 | nr_tw_prev = READ_ONCE(first_req->nr_tw); | |
1342 | } | |
1343 | nr_tw = nr_tw_prev + 1; | |
1344 | /* Large enough to fail the nr_wait comparison below */ | |
1345 | if (!(flags & IOU_F_TWQ_LAZY_WAKE)) | |
1346 | nr_tw = -1U; | |
1347 | ||
1348 | req->nr_tw = nr_tw; | |
51509400 PB |
1349 | req->io_task_work.node.next = first; |
1350 | } while (!try_cmpxchg(&ctx->work_llist.first, &first, | |
1351 | &req->io_task_work.node)); | |
1352 | ||
8751d154 | 1353 | if (!first) { |
8751d154 PB |
1354 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1355 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1356 | if (ctx->has_evfd) | |
1357 | io_eventfd_signal(ctx); | |
c0e0d6ba DY |
1358 | } |
1359 | ||
8751d154 PB |
1360 | nr_wait = atomic_read(&ctx->cq_wait_nr); |
1361 | /* no one is waiting */ | |
1362 | if (!nr_wait) | |
1363 | return; | |
1364 | /* either not enough or the previous add has already woken it up */ | |
1365 | if (nr_wait > nr_tw || nr_tw_prev >= nr_wait) | |
1366 | return; | |
1367 | /* pairs with set_current_state() in io_cqring_wait() */ | |
1368 | smp_mb__after_atomic(); | |
1369 | wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE); | |
c0e0d6ba DY |
1370 | } |
1371 | ||
91c7884a | 1372 | static void io_req_normal_work_add(struct io_kiocb *req) |
7cbf1722 | 1373 | { |
c34398a8 | 1374 | struct io_uring_task *tctx = req->task->io_uring; |
9f010507 | 1375 | struct io_ring_ctx *ctx = req->ctx; |
7cbf1722 | 1376 | |
7cbf1722 | 1377 | /* task_work already pending, we're done */ |
32d91f05 | 1378 | if (!llist_add(&req->io_task_work.node, &tctx->task_list)) |
e09ee510 | 1379 | return; |
7cbf1722 | 1380 | |
ef060ea9 JA |
1381 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1382 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1383 | ||
3fe07bcd | 1384 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1385 | return; |
2215bed9 | 1386 | |
dfbe5561 | 1387 | io_fallback_tw(tctx, false); |
c0e0d6ba DY |
1388 | } |
1389 | ||
91c7884a PB |
1390 | void __io_req_task_work_add(struct io_kiocb *req, unsigned flags) |
1391 | { | |
1392 | if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) { | |
1393 | rcu_read_lock(); | |
1394 | io_req_local_work_add(req, flags); | |
1395 | rcu_read_unlock(); | |
1396 | } else { | |
1397 | io_req_normal_work_add(req); | |
1398 | } | |
1399 | } | |
1400 | ||
c0e0d6ba DY |
1401 | static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) |
1402 | { | |
1403 | struct llist_node *node; | |
1404 | ||
1405 | node = llist_del_all(&ctx->work_llist); | |
1406 | while (node) { | |
1407 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1408 | io_task_work.node); | |
1409 | ||
1410 | node = node->next; | |
91c7884a | 1411 | io_req_normal_work_add(req); |
c0e0d6ba DY |
1412 | } |
1413 | } | |
1414 | ||
a282967c | 1415 | static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts) |
c0e0d6ba | 1416 | { |
c0e0d6ba | 1417 | struct llist_node *node; |
c3f4d39e | 1418 | unsigned int loops = 0; |
140102ae | 1419 | int ret = 0; |
c0e0d6ba | 1420 | |
140102ae | 1421 | if (WARN_ON_ONCE(ctx->submitter_task != current)) |
c0e0d6ba | 1422 | return -EEXIST; |
c3f4d39e PB |
1423 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1424 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
c0e0d6ba | 1425 | again: |
3af0356c JA |
1426 | /* |
1427 | * llists are in reverse order, flip it back the right way before | |
1428 | * running the pending items. | |
1429 | */ | |
1430 | node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL)); | |
c3f4d39e | 1431 | while (node) { |
c0e0d6ba DY |
1432 | struct llist_node *next = node->next; |
1433 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1434 | io_task_work.node); | |
1435 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); | |
c92fcfc2 JA |
1436 | INDIRECT_CALL_2(req->io_task_work.func, |
1437 | io_poll_task_func, io_req_rw_complete, | |
1438 | req, ts); | |
c0e0d6ba DY |
1439 | ret++; |
1440 | node = next; | |
1441 | } | |
c3f4d39e | 1442 | loops++; |
c0e0d6ba | 1443 | |
c3f4d39e | 1444 | if (!llist_empty(&ctx->work_llist)) |
c0e0d6ba | 1445 | goto again; |
a282967c | 1446 | if (ts->locked) { |
c0e0d6ba | 1447 | io_submit_flush_completions(ctx); |
b0b7a7d2 PB |
1448 | if (!llist_empty(&ctx->work_llist)) |
1449 | goto again; | |
1450 | } | |
f75d5036 | 1451 | trace_io_uring_local_work_run(ctx, ret, loops); |
c0e0d6ba | 1452 | return ret; |
8ac5d85a JA |
1453 | } |
1454 | ||
360173ab PB |
1455 | static inline int io_run_local_work_locked(struct io_ring_ctx *ctx) |
1456 | { | |
a282967c | 1457 | struct io_tw_state ts = { .locked = true, }; |
360173ab PB |
1458 | int ret; |
1459 | ||
1460 | if (llist_empty(&ctx->work_llist)) | |
1461 | return 0; | |
1462 | ||
a282967c | 1463 | ret = __io_run_local_work(ctx, &ts); |
360173ab | 1464 | /* shouldn't happen! */ |
a282967c | 1465 | if (WARN_ON_ONCE(!ts.locked)) |
360173ab PB |
1466 | mutex_lock(&ctx->uring_lock); |
1467 | return ret; | |
1468 | } | |
1469 | ||
3e565555 | 1470 | static int io_run_local_work(struct io_ring_ctx *ctx) |
8ac5d85a | 1471 | { |
a282967c | 1472 | struct io_tw_state ts = {}; |
8ac5d85a JA |
1473 | int ret; |
1474 | ||
a282967c PB |
1475 | ts.locked = mutex_trylock(&ctx->uring_lock); |
1476 | ret = __io_run_local_work(ctx, &ts); | |
1477 | if (ts.locked) | |
8ac5d85a JA |
1478 | mutex_unlock(&ctx->uring_lock); |
1479 | ||
1480 | return ret; | |
c0e0d6ba DY |
1481 | } |
1482 | ||
a282967c | 1483 | static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts) |
c40f6379 | 1484 | { |
a282967c | 1485 | io_tw_lock(req->ctx, ts); |
973fc83f | 1486 | io_req_defer_failed(req, req->cqe.res); |
c40f6379 JA |
1487 | } |
1488 | ||
a282967c | 1489 | void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts) |
c40f6379 | 1490 | { |
a282967c | 1491 | io_tw_lock(req->ctx, ts); |
316319e8 | 1492 | /* req->task == current here, checking PF_EXITING is safe */ |
6bb30855 | 1493 | if (unlikely(req->task->flags & PF_EXITING)) |
973fc83f | 1494 | io_req_defer_failed(req, -EFAULT); |
6bb30855 | 1495 | else if (req->flags & REQ_F_FORCE_ASYNC) |
a282967c | 1496 | io_queue_iowq(req, ts); |
6bb30855 DY |
1497 | else |
1498 | io_queue_sqe(req); | |
c40f6379 JA |
1499 | } |
1500 | ||
59915143 | 1501 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1502 | { |
97b388d7 | 1503 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1504 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1505 | io_req_task_work_add(req); |
c40f6379 JA |
1506 | } |
1507 | ||
f3b44f92 | 1508 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1509 | { |
5b0a6acc | 1510 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1511 | io_req_task_work_add(req); |
a3df7698 PB |
1512 | } |
1513 | ||
59915143 | 1514 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1515 | { |
9b5f7bd9 | 1516 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1517 | |
1518 | if (nxt) | |
906a8c3f | 1519 | io_req_task_queue(nxt); |
c69f8dbe JL |
1520 | } |
1521 | ||
ec26c225 PB |
1522 | static void io_free_batch_list(struct io_ring_ctx *ctx, |
1523 | struct io_wq_work_node *node) | |
3aa83bfb | 1524 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1525 | { |
3aa83bfb PB |
1526 | do { |
1527 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1528 | comp_list); | |
2d6500d4 | 1529 | |
a538be5b PB |
1530 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1531 | if (req->flags & REQ_F_REFCOUNT) { | |
1532 | node = req->comp_list.next; | |
1533 | if (!req_ref_put_and_test(req)) | |
1534 | continue; | |
1535 | } | |
b605a7fa PB |
1536 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1537 | struct async_poll *apoll = req->apoll; | |
1538 | ||
1539 | if (apoll->double_poll) | |
1540 | kfree(apoll->double_poll); | |
9731bc98 JA |
1541 | if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) |
1542 | kfree(apoll); | |
b605a7fa PB |
1543 | req->flags &= ~REQ_F_POLLED; |
1544 | } | |
da1a08c5 | 1545 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1546 | io_queue_next(req); |
a538be5b PB |
1547 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1548 | io_clean_op(req); | |
c1e53a69 | 1549 | } |
17bc2837 | 1550 | io_put_file(req); |
2d6500d4 | 1551 | |
ab409402 | 1552 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1553 | |
2fdd6fb5 | 1554 | io_put_task(req->task); |
c1e53a69 | 1555 | node = req->comp_list.next; |
fa05457a | 1556 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1557 | } while (node); |
7a743e22 PB |
1558 | } |
1559 | ||
ec26c225 | 1560 | void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1561 | __must_hold(&ctx->uring_lock) |
905c172f | 1562 | { |
cd0ca2e0 | 1563 | struct io_submit_state *state = &ctx->submit_state; |
fa780334 | 1564 | struct io_wq_work_node *node; |
905c172f | 1565 | |
f66f7342 | 1566 | __io_cq_lock(ctx); |
931147dd DY |
1567 | /* must come first to preserve CQE ordering in failure cases */ |
1568 | if (state->cqes_count) | |
1569 | __io_flush_post_cqes(ctx); | |
fa780334 | 1570 | __wq_list_for_each(node, &state->compl_reqs) { |
d9dee430 PB |
1571 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1572 | comp_list); | |
3d4aeb9f | 1573 | |
f66f7342 | 1574 | if (!(req->flags & REQ_F_CQE_SKIP) && |
00b0db56 | 1575 | unlikely(!io_fill_cqe_req(ctx, req))) { |
27122c07 | 1576 | if (ctx->lockless_cq) { |
f66f7342 PB |
1577 | spin_lock(&ctx->completion_lock); |
1578 | io_req_cqe_overflow(req); | |
1579 | spin_unlock(&ctx->completion_lock); | |
1580 | } else { | |
1581 | io_req_cqe_overflow(req); | |
1582 | } | |
1583 | } | |
905c172f | 1584 | } |
c98c81a4 | 1585 | __io_cq_unlock_post(ctx); |
d9dee430 | 1586 | |
931147dd DY |
1587 | if (!wq_list_empty(&ctx->submit_state.compl_reqs)) { |
1588 | io_free_batch_list(ctx, state->compl_reqs.first); | |
1589 | INIT_WQ_LIST(&state->compl_reqs); | |
1590 | } | |
7a743e22 PB |
1591 | } |
1592 | ||
6c503150 | 1593 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1594 | { |
1595 | /* See comment at the top of this file */ | |
1596 | smp_rmb(); | |
e23de15f | 1597 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1598 | } |
1599 | ||
def596e9 JA |
1600 | /* |
1601 | * We can't just wait for polled events to come to us, we have to actively | |
1602 | * find and complete them. | |
1603 | */ | |
c072481d | 1604 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1605 | { |
1606 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1607 | return; | |
1608 | ||
1609 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1610 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1611 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1612 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1613 | break; |
08f5439f JA |
1614 | /* |
1615 | * Ensure we allow local-to-the-cpu processing to take place, | |
1616 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1617 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1618 | */ |
3fcee5a6 PB |
1619 | if (need_resched()) { |
1620 | mutex_unlock(&ctx->uring_lock); | |
1621 | cond_resched(); | |
1622 | mutex_lock(&ctx->uring_lock); | |
1623 | } | |
def596e9 JA |
1624 | } |
1625 | mutex_unlock(&ctx->uring_lock); | |
1626 | } | |
1627 | ||
7668b92a | 1628 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1629 | { |
7668b92a | 1630 | unsigned int nr_events = 0; |
155bc950 | 1631 | unsigned long check_cq; |
500f9fba | 1632 | |
76de6749 PB |
1633 | if (!io_allowed_run_tw(ctx)) |
1634 | return -EEXIST; | |
1635 | ||
3a08576b PB |
1636 | check_cq = READ_ONCE(ctx->check_cq); |
1637 | if (unlikely(check_cq)) { | |
1638 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
a85381d8 | 1639 | __io_cqring_overflow_flush(ctx); |
3a08576b PB |
1640 | /* |
1641 | * Similarly do not spin if we have not informed the user of any | |
1642 | * dropped CQE. | |
1643 | */ | |
1644 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1645 | return -EBADR; | |
1646 | } | |
f39c8a5b PB |
1647 | /* |
1648 | * Don't enter poll loop if we already have events pending. | |
1649 | * If we do, we can potentially be spinning for commands that | |
1650 | * already triggered a CQE (eg in error). | |
1651 | */ | |
f39c8a5b | 1652 | if (io_cqring_events(ctx)) |
d487b43c | 1653 | return 0; |
155bc950 | 1654 | |
def596e9 | 1655 | do { |
9e4bef2b JA |
1656 | int ret = 0; |
1657 | ||
500f9fba JA |
1658 | /* |
1659 | * If a submit got punted to a workqueue, we can have the | |
1660 | * application entering polling for a command before it gets | |
1661 | * issued. That app will hold the uring_lock for the duration | |
1662 | * of the poll right here, so we need to take a breather every | |
1663 | * now and then to ensure that the issue has a chance to add | |
1664 | * the poll to the issued list. Otherwise we can spin here | |
1665 | * forever, while the workqueue is stuck trying to acquire the | |
1666 | * very same mutex. | |
1667 | */ | |
dac6a0ea JA |
1668 | if (wq_list_empty(&ctx->iopoll_list) || |
1669 | io_task_work_pending(ctx)) { | |
8f487ef2 PB |
1670 | u32 tail = ctx->cached_cq_tail; |
1671 | ||
8de11cdc | 1672 | (void) io_run_local_work_locked(ctx); |
def596e9 | 1673 | |
dac6a0ea JA |
1674 | if (task_work_pending(current) || |
1675 | wq_list_empty(&ctx->iopoll_list)) { | |
dac6a0ea | 1676 | mutex_unlock(&ctx->uring_lock); |
9d54bd6a | 1677 | io_run_task_work(); |
dac6a0ea | 1678 | mutex_lock(&ctx->uring_lock); |
dac6a0ea | 1679 | } |
8f487ef2 PB |
1680 | /* some requests don't go through iopoll_list */ |
1681 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1682 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1683 | break; |
500f9fba | 1684 | } |
5ba3c874 | 1685 | ret = io_do_iopoll(ctx, !min); |
9e4bef2b JA |
1686 | if (unlikely(ret < 0)) |
1687 | return ret; | |
dc314886 PB |
1688 | |
1689 | if (task_sigpending(current)) | |
1690 | return -EINTR; | |
9e4bef2b | 1691 | if (need_resched()) |
5ba3c874 | 1692 | break; |
d487b43c | 1693 | |
5ba3c874 | 1694 | nr_events += ret; |
9e4bef2b | 1695 | } while (nr_events < min); |
d487b43c | 1696 | |
9e4bef2b | 1697 | return 0; |
def596e9 | 1698 | } |
7012c815 | 1699 | |
a282967c | 1700 | void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts) |
8ef12efe | 1701 | { |
a282967c | 1702 | if (ts->locked) |
9da070b1 | 1703 | io_req_complete_defer(req); |
7012c815 | 1704 | else |
27f35fe9 | 1705 | io_req_complete_post(req, IO_URING_F_UNLOCKED); |
8ef12efe JA |
1706 | } |
1707 | ||
def596e9 JA |
1708 | /* |
1709 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1710 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1711 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1712 | * accessing the kiocb cookie. |
1713 | */ | |
9882131c | 1714 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1715 | { |
1716 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1717 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1718 | |
1719 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1720 | if (unlikely(needs_lock)) |
cb3d8972 | 1721 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1722 | |
1723 | /* | |
1724 | * Track whether we have multiple files in our lists. This will impact | |
1725 | * how we do polling eventually, not spinning if we're on potentially | |
1726 | * different devices. | |
1727 | */ | |
5eef4e87 | 1728 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1729 | ctx->poll_multi_queue = false; |
1730 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1731 | struct io_kiocb *list_req; |
1732 | ||
5eef4e87 PB |
1733 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1734 | comp_list); | |
30da1b45 | 1735 | if (list_req->file != req->file) |
915b3dde | 1736 | ctx->poll_multi_queue = true; |
def596e9 JA |
1737 | } |
1738 | ||
1739 | /* | |
1740 | * For fast devices, IO may have already completed. If it has, add | |
1741 | * it to the front so we find it first. | |
1742 | */ | |
65a6543d | 1743 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1744 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1745 | else |
5eef4e87 | 1746 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1747 | |
3b44b371 | 1748 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1749 | /* |
1750 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1751 | * in sq thread task context or in io worker task context. If | |
1752 | * current task context is sq thread, we don't need to check | |
1753 | * whether should wake up sq thread. | |
1754 | */ | |
1755 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1756 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1757 | wake_up(&ctx->sq_data->wait); | |
1758 | ||
1759 | mutex_unlock(&ctx->uring_lock); | |
1760 | } | |
def596e9 JA |
1761 | } |
1762 | ||
a4ad4f74 | 1763 | unsigned int io_file_get_flags(struct file *file) |
88459b50 | 1764 | { |
88459b50 | 1765 | unsigned int res = 0; |
af197f50 | 1766 | |
53cfd5ce | 1767 | if (S_ISREG(file_inode(file)->i_mode)) |
8487f083 | 1768 | res |= REQ_F_ISREG; |
b9a6c945 | 1769 | if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT)) |
8487f083 | 1770 | res |= REQ_F_SUPPORT_NOWAIT; |
88459b50 | 1771 | return res; |
2b188cc1 JA |
1772 | } |
1773 | ||
99f15d8d | 1774 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1775 | { |
f30bd4d0 BL |
1776 | WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size); |
1777 | req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL); | |
d886e185 PB |
1778 | if (req->async_data) { |
1779 | req->flags |= REQ_F_ASYNC_DATA; | |
1780 | return false; | |
1781 | } | |
1782 | return true; | |
3d9932a8 XW |
1783 | } |
1784 | ||
f3b44f92 | 1785 | int io_req_prep_async(struct io_kiocb *req) |
f67676d1 | 1786 | { |
f30bd4d0 | 1787 | const struct io_cold_def *cdef = &io_cold_defs[req->opcode]; |
a7dd2782 | 1788 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
0702e536 JA |
1789 | |
1790 | /* assign early for deferred execution for non-fixed file */ | |
54aa7f23 | 1791 | if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file) |
0702e536 | 1792 | req->file = io_file_get_normal(req, req->cqe.fd); |
f30bd4d0 | 1793 | if (!cdef->prep_async) |
0702e536 JA |
1794 | return 0; |
1795 | if (WARN_ON_ONCE(req_has_async_data(req))) | |
1796 | return -EFAULT; | |
f30bd4d0 | 1797 | if (!def->manual_alloc) { |
59169439 PB |
1798 | if (io_alloc_async_data(req)) |
1799 | return -EAGAIN; | |
1800 | } | |
f30bd4d0 | 1801 | return cdef->prep_async(req); |
bfe76559 PB |
1802 | } |
1803 | ||
9cf7c104 PB |
1804 | static u32 io_get_sequence(struct io_kiocb *req) |
1805 | { | |
a3dbdf54 | 1806 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1807 | struct io_kiocb *cur; |
9cf7c104 | 1808 | |
a3dbdf54 | 1809 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1810 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1811 | seq--; |
1812 | return seq; | |
9cf7c104 PB |
1813 | } |
1814 | ||
c072481d | 1815 | static __cold void io_drain_req(struct io_kiocb *req) |
e276ae34 | 1816 | __must_hold(&ctx->uring_lock) |
de0617e4 | 1817 | { |
a197f664 | 1818 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1819 | struct io_defer_entry *de; |
f67676d1 | 1820 | int ret; |
e0eb71dc | 1821 | u32 seq = io_get_sequence(req); |
3c19966d | 1822 | |
9d858b21 | 1823 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1824 | spin_lock(&ctx->completion_lock); |
5e371265 | 1825 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1826 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1827 | queue: |
10c66904 | 1828 | ctx->drain_active = false; |
e0eb71dc PB |
1829 | io_req_task_queue(req); |
1830 | return; | |
10c66904 | 1831 | } |
e302f104 | 1832 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1833 | |
cbdcb435 | 1834 | io_prep_async_link(req); |
27dc8338 | 1835 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1836 | if (!de) { |
1b48773f | 1837 | ret = -ENOMEM; |
ef5c600a DY |
1838 | io_req_defer_failed(req, ret); |
1839 | return; | |
76cc33d7 | 1840 | } |
2d28390a | 1841 | |
79ebeaee | 1842 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1843 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1844 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1845 | kfree(de); |
e0eb71dc | 1846 | goto queue; |
de0617e4 JA |
1847 | } |
1848 | ||
48863ffd | 1849 | trace_io_uring_defer(req); |
27dc8338 | 1850 | de->req = req; |
9cf7c104 | 1851 | de->seq = seq; |
27dc8338 | 1852 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1853 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1854 | } |
1855 | ||
f4992544 JA |
1856 | static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def, |
1857 | unsigned int issue_flags) | |
6bf9c47a | 1858 | { |
f4992544 | 1859 | if (req->file || !def->needs_file) |
6bf9c47a JA |
1860 | return true; |
1861 | ||
1862 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1863 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1864 | else |
cef216fc | 1865 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1866 | |
772f5e00 | 1867 | return !!req->file; |
6bf9c47a JA |
1868 | } |
1869 | ||
889fca73 | 1870 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1871 | { |
a7dd2782 | 1872 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
5730b27e | 1873 | const struct cred *creds = NULL; |
d625c6ee | 1874 | int ret; |
2b188cc1 | 1875 | |
f4992544 | 1876 | if (unlikely(!io_assign_file(req, def, issue_flags))) |
70152140 JA |
1877 | return -EBADF; |
1878 | ||
6878b40e | 1879 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1880 | creds = override_creds(req->creds); |
5730b27e | 1881 | |
fcde59fe | 1882 | if (!def->audit_skip) |
5bd2182d PM |
1883 | audit_uring_entry(req->opcode); |
1884 | ||
0702e536 | 1885 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1886 | |
fcde59fe | 1887 | if (!def->audit_skip) |
5bd2182d PM |
1888 | audit_uring_exit(!ret, ret); |
1889 | ||
5730b27e JA |
1890 | if (creds) |
1891 | revert_creds(creds); | |
97b388d7 | 1892 | |
75d7b3ae PB |
1893 | if (ret == IOU_OK) { |
1894 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
9da070b1 | 1895 | io_req_complete_defer(req); |
75d7b3ae | 1896 | else |
1bec951c | 1897 | io_req_complete_post(req, issue_flags); |
75d7b3ae | 1898 | } else if (ret != IOU_ISSUE_SKIP_COMPLETE) |
def596e9 | 1899 | return ret; |
97b388d7 | 1900 | |
b532576e | 1901 | /* If the op doesn't have a file, we're not polling for it */ |
ef0ec1ad | 1902 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue) |
9882131c | 1903 | io_iopoll_req_issued(req, issue_flags); |
def596e9 JA |
1904 | |
1905 | return 0; | |
2b188cc1 JA |
1906 | } |
1907 | ||
a282967c | 1908 | int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts) |
329061d3 | 1909 | { |
a282967c | 1910 | io_tw_lock(req->ctx, ts); |
9a692451 DY |
1911 | return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT| |
1912 | IO_URING_F_COMPLETE_DEFER); | |
329061d3 JA |
1913 | } |
1914 | ||
c9f06aa7 | 1915 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1916 | { |
1917 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
247f97a5 | 1918 | struct io_kiocb *nxt = NULL; |
ebc11b6c | 1919 | |
247f97a5 PB |
1920 | if (req_ref_put_and_test(req)) { |
1921 | if (req->flags & IO_REQ_LINK_FLAGS) | |
1922 | nxt = io_req_find_next(req); | |
1923 | io_free_req(req); | |
1924 | } | |
1925 | return nxt ? &nxt->work : NULL; | |
ebc11b6c PB |
1926 | } |
1927 | ||
c9f06aa7 | 1928 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1929 | { |
1930 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
a7dd2782 | 1931 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
e6aeb272 | 1932 | unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ; |
d01905db | 1933 | bool needs_poll = false; |
6bf9c47a | 1934 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1935 | |
23a6c9ac | 1936 | /* one will be dropped by ->io_wq_free_work() after returning to io-wq */ |
48dcd38d PB |
1937 | if (!(req->flags & REQ_F_REFCOUNT)) |
1938 | __io_req_set_refcount(req, 2); | |
1939 | else | |
1940 | req_ref_get(req); | |
5d5901a3 | 1941 | |
cb2d344c | 1942 | io_arm_ltimeout(req); |
6bf9c47a | 1943 | |
dadebc35 | 1944 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1945 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1946 | fail: |
6bf9c47a | 1947 | io_req_task_queue_fail(req, err); |
d01905db PB |
1948 | return; |
1949 | } | |
f4992544 | 1950 | if (!io_assign_file(req, def, issue_flags)) { |
0f8da75b PB |
1951 | err = -EBADF; |
1952 | work->flags |= IO_WQ_WORK_CANCEL; | |
1953 | goto fail; | |
1954 | } | |
31b51510 | 1955 | |
d01905db | 1956 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
1957 | bool opcode_poll = def->pollin || def->pollout; |
1958 | ||
1959 | if (opcode_poll && file_can_poll(req->file)) { | |
1960 | needs_poll = true; | |
d01905db | 1961 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 1962 | } |
561fb04a | 1963 | } |
31b51510 | 1964 | |
d01905db PB |
1965 | do { |
1966 | ret = io_issue_sqe(req, issue_flags); | |
1967 | if (ret != -EAGAIN) | |
1968 | break; | |
a9be2022 JA |
1969 | |
1970 | /* | |
1971 | * If REQ_F_NOWAIT is set, then don't wait or retry with | |
1972 | * poll. -EAGAIN is final for that case. | |
1973 | */ | |
1974 | if (req->flags & REQ_F_NOWAIT) | |
1975 | break; | |
1976 | ||
d01905db PB |
1977 | /* |
1978 | * We can get EAGAIN for iopolled IO even though we're | |
1979 | * forcing a sync submission from here, since we can't | |
1980 | * wait for request slots on the block side. | |
1981 | */ | |
1982 | if (!needs_poll) { | |
e0deb6a0 PB |
1983 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
1984 | break; | |
45500dc4 PB |
1985 | if (io_wq_worker_stopped()) |
1986 | break; | |
d01905db PB |
1987 | cond_resched(); |
1988 | continue; | |
90fa0288 HX |
1989 | } |
1990 | ||
4d9237e3 | 1991 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
1992 | return; |
1993 | /* aborted or ready, in either case retry blocking */ | |
1994 | needs_poll = false; | |
1995 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
1996 | } while (1); | |
31b51510 | 1997 | |
a3df7698 | 1998 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 1999 | if (ret < 0) |
a3df7698 | 2000 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
2001 | } |
2002 | ||
531113bb JA |
2003 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
2004 | unsigned int issue_flags) | |
09bb8394 | 2005 | { |
5106dd6e | 2006 | struct io_ring_ctx *ctx = req->ctx; |
4bfb0c9a | 2007 | struct io_fixed_file *slot; |
5106dd6e | 2008 | struct file *file = NULL; |
09bb8394 | 2009 | |
93f052cb | 2010 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 2011 | |
ac177053 | 2012 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 2013 | goto out; |
ac177053 | 2014 | fd = array_index_nospec(fd, ctx->nr_user_files); |
4bfb0c9a CH |
2015 | slot = io_fixed_file_slot(&ctx->file_table, fd); |
2016 | file = io_slot_file(slot); | |
2017 | req->flags |= io_slot_flags(slot); | |
5106dd6e JA |
2018 | io_req_set_rsrc_node(req, ctx, 0); |
2019 | out: | |
93f052cb | 2020 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
2021 | return file; |
2022 | } | |
d44f554e | 2023 | |
531113bb | 2024 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 2025 | { |
62906e89 | 2026 | struct file *file = fget(fd); |
ac177053 | 2027 | |
48863ffd | 2028 | trace_io_uring_file_get(req, fd); |
09bb8394 | 2029 | |
ac177053 | 2030 | /* we don't allow fixed io_uring files */ |
e5550a14 | 2031 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 2032 | io_req_track_inflight(req); |
8371adf5 | 2033 | return file; |
09bb8394 JA |
2034 | } |
2035 | ||
7bfa9bad | 2036 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
2037 | __must_hold(&req->ctx->uring_lock) |
2038 | { | |
7bfa9bad PB |
2039 | struct io_kiocb *linked_timeout; |
2040 | ||
2041 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
973fc83f | 2042 | io_req_defer_failed(req, ret); |
7bfa9bad PB |
2043 | return; |
2044 | } | |
2045 | ||
2046 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 2047 | |
4d9237e3 | 2048 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 2049 | case IO_APOLL_READY: |
336d28a8 | 2050 | io_kbuf_recycle(req, 0); |
d475a9a6 PB |
2051 | io_req_task_queue(req); |
2052 | break; | |
2053 | case IO_APOLL_ABORTED: | |
6436c770 | 2054 | io_kbuf_recycle(req, 0); |
77955efb | 2055 | io_queue_iowq(req, NULL); |
d475a9a6 | 2056 | break; |
b1c62645 | 2057 | case IO_APOLL_OK: |
b1c62645 | 2058 | break; |
d475a9a6 PB |
2059 | } |
2060 | ||
2061 | if (linked_timeout) | |
2062 | io_queue_linked_timeout(linked_timeout); | |
2063 | } | |
2064 | ||
cbc2e203 | 2065 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 2066 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 2067 | { |
e0c5c576 | 2068 | int ret; |
2b188cc1 | 2069 | |
c5eef2b9 | 2070 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 2071 | |
491381ce JA |
2072 | /* |
2073 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2074 | * doesn't support non-blocking read/write attempts | |
2075 | */ | |
7bfa9bad | 2076 | if (likely(!ret)) |
cb2d344c | 2077 | io_arm_ltimeout(req); |
7bfa9bad PB |
2078 | else |
2079 | io_queue_async(req, ret); | |
2b188cc1 JA |
2080 | } |
2081 | ||
4652fe3f | 2082 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 2083 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 2084 | { |
17b147f6 PB |
2085 | if (unlikely(req->flags & REQ_F_FAIL)) { |
2086 | /* | |
2087 | * We don't submit, fail them all, for that replace hardlinks | |
2088 | * with normal links. Extra REQ_F_LINK is tolerated. | |
2089 | */ | |
2090 | req->flags &= ~REQ_F_HARDLINK; | |
2091 | req->flags |= REQ_F_LINK; | |
973fc83f | 2092 | io_req_defer_failed(req, req->cqe.res); |
76cc33d7 PB |
2093 | } else { |
2094 | int ret = io_req_prep_async(req); | |
2095 | ||
ef5c600a | 2096 | if (unlikely(ret)) { |
973fc83f | 2097 | io_req_defer_failed(req, ret); |
ef5c600a DY |
2098 | return; |
2099 | } | |
2100 | ||
2101 | if (unlikely(req->ctx->drain_active)) | |
2102 | io_drain_req(req); | |
76cc33d7 | 2103 | else |
77955efb | 2104 | io_queue_iowq(req, NULL); |
ce35a47a | 2105 | } |
4fe2c963 JL |
2106 | } |
2107 | ||
b16fed66 PB |
2108 | /* |
2109 | * Check SQE restrictions (opcode and flags). | |
2110 | * | |
2111 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
2112 | */ | |
2113 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
2114 | struct io_kiocb *req, | |
2115 | unsigned int sqe_flags) | |
4fe2c963 | 2116 | { |
b16fed66 PB |
2117 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
2118 | return false; | |
2119 | ||
2120 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
2121 | ctx->restrictions.sqe_flags_required) | |
2122 | return false; | |
2123 | ||
2124 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
2125 | ctx->restrictions.sqe_flags_required)) | |
2126 | return false; | |
2127 | ||
2128 | return true; | |
4fe2c963 JL |
2129 | } |
2130 | ||
22b2ca31 PB |
2131 | static void io_init_req_drain(struct io_kiocb *req) |
2132 | { | |
2133 | struct io_ring_ctx *ctx = req->ctx; | |
2134 | struct io_kiocb *head = ctx->submit_state.link.head; | |
2135 | ||
2136 | ctx->drain_active = true; | |
2137 | if (head) { | |
2138 | /* | |
2139 | * If we need to drain a request in the middle of a link, drain | |
2140 | * the head request and the next request/link after the current | |
2141 | * link. Considering sequential execution of links, | |
b6c7db32 | 2142 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
2143 | * link. |
2144 | */ | |
b6c7db32 | 2145 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
2146 | ctx->drain_next = true; |
2147 | } | |
2148 | } | |
2149 | ||
b16fed66 PB |
2150 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2151 | const struct io_uring_sqe *sqe) | |
282cdc86 | 2152 | __must_hold(&ctx->uring_lock) |
b16fed66 | 2153 | { |
a7dd2782 | 2154 | const struct io_issue_def *def; |
b16fed66 | 2155 | unsigned int sqe_flags; |
fc0ae024 | 2156 | int personality; |
4a04d1d1 | 2157 | u8 opcode; |
b16fed66 | 2158 | |
864ea921 | 2159 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 2160 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 PB |
2161 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
2162 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
cef216fc | 2163 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 2164 | req->file = NULL; |
c1bdf8ed | 2165 | req->rsrc_node = NULL; |
b16fed66 | 2166 | req->task = current; |
b16fed66 | 2167 | |
4a04d1d1 PB |
2168 | if (unlikely(opcode >= IORING_OP_LAST)) { |
2169 | req->opcode = 0; | |
b16fed66 | 2170 | return -EINVAL; |
4a04d1d1 | 2171 | } |
a7dd2782 | 2172 | def = &io_issue_defs[opcode]; |
68fe256a PB |
2173 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
2174 | /* enforce forwards compatibility on users */ | |
2175 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
2176 | return -EINVAL; | |
4e906702 | 2177 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 2178 | if (!def->buffer_select) |
4e906702 JA |
2179 | return -EOPNOTSUPP; |
2180 | req->buf_index = READ_ONCE(sqe->buf_group); | |
2181 | } | |
5562a8d7 PB |
2182 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
2183 | ctx->drain_disabled = true; | |
2184 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
2185 | if (ctx->drain_disabled) | |
2186 | return -EOPNOTSUPP; | |
22b2ca31 | 2187 | io_init_req_drain(req); |
5562a8d7 | 2188 | } |
2a56a9bd PB |
2189 | } |
2190 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
2191 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
2192 | return -EACCES; | |
2193 | /* knock it to the slow queue path, will be drained there */ | |
2194 | if (ctx->drain_active) | |
2195 | req->flags |= REQ_F_FORCE_ASYNC; | |
2196 | /* if there is no link, we're at "next" request and need to drain */ | |
2197 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
2198 | ctx->drain_next = false; | |
2199 | ctx->drain_active = true; | |
b6c7db32 | 2200 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 2201 | } |
68fe256a | 2202 | } |
b16fed66 | 2203 | |
fcde59fe | 2204 | if (!def->ioprio && sqe->ioprio) |
73911426 | 2205 | return -EINVAL; |
fcde59fe | 2206 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
73911426 JA |
2207 | return -EINVAL; |
2208 | ||
fcde59fe | 2209 | if (def->needs_file) { |
6d63416d PB |
2210 | struct io_submit_state *state = &ctx->submit_state; |
2211 | ||
cef216fc | 2212 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 2213 | |
6d63416d PB |
2214 | /* |
2215 | * Plug now if we have more than 2 IO left after this, and the | |
2216 | * target is potentially a read/write to block based storage. | |
2217 | */ | |
fcde59fe | 2218 | if (state->need_plug && def->plug) { |
6d63416d PB |
2219 | state->plug_started = true; |
2220 | state->need_plug = false; | |
5ca7a8b3 | 2221 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 2222 | } |
b16fed66 | 2223 | } |
863e0560 | 2224 | |
003e8dcc JA |
2225 | personality = READ_ONCE(sqe->personality); |
2226 | if (personality) { | |
cdab10bf LT |
2227 | int ret; |
2228 | ||
c10d1f98 PB |
2229 | req->creds = xa_load(&ctx->personalities, personality); |
2230 | if (!req->creds) | |
003e8dcc | 2231 | return -EINVAL; |
c10d1f98 | 2232 | get_cred(req->creds); |
cdc1404a PM |
2233 | ret = security_uring_override_creds(req->creds); |
2234 | if (ret) { | |
2235 | put_cred(req->creds); | |
2236 | return ret; | |
2237 | } | |
b8e64b53 | 2238 | req->flags |= REQ_F_CREDS; |
003e8dcc | 2239 | } |
b16fed66 | 2240 | |
0702e536 | 2241 | return def->prep(req, sqe); |
b16fed66 PB |
2242 | } |
2243 | ||
df3becde PB |
2244 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
2245 | struct io_kiocb *req, int ret) | |
2246 | { | |
2247 | struct io_ring_ctx *ctx = req->ctx; | |
2248 | struct io_submit_link *link = &ctx->submit_state.link; | |
2249 | struct io_kiocb *head = link->head; | |
2250 | ||
48863ffd | 2251 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
2252 | |
2253 | /* | |
2254 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
2255 | * unusable. Instead of failing eagerly, continue assembling the link if | |
2256 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
2257 | * should find the flag and handle the rest. | |
2258 | */ | |
2259 | req_fail_link_node(req, ret); | |
2260 | if (head && !(head->flags & REQ_F_FAIL)) | |
2261 | req_fail_link_node(head, -ECANCELED); | |
2262 | ||
2263 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
2264 | if (head) { | |
2265 | link->last->link = req; | |
2266 | link->head = NULL; | |
2267 | req = head; | |
2268 | } | |
2269 | io_queue_sqe_fallback(req); | |
2270 | return ret; | |
2271 | } | |
2272 | ||
2273 | if (head) | |
2274 | link->last->link = req; | |
2275 | else | |
2276 | link->head = req; | |
2277 | link->last = req; | |
2278 | return 0; | |
2279 | } | |
2280 | ||
2281 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 2282 | const struct io_uring_sqe *sqe) |
282cdc86 | 2283 | __must_hold(&ctx->uring_lock) |
9e645e11 | 2284 | { |
a1ab7b35 | 2285 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 2286 | int ret; |
9e645e11 | 2287 | |
a6b8cadc | 2288 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
2289 | if (unlikely(ret)) |
2290 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 2291 | |
2ad57931 | 2292 | trace_io_uring_submit_req(req); |
a6b8cadc | 2293 | |
9e645e11 JA |
2294 | /* |
2295 | * If we already have a head request, queue this one for async | |
2296 | * submittal once the head completes. If we don't have a head but | |
2297 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2298 | * submitted sync once the chain is complete. If none of those | |
2299 | * conditions are true (normal request), then just queue it. | |
2300 | */ | |
924a07e4 | 2301 | if (unlikely(link->head)) { |
df3becde PB |
2302 | ret = io_req_prep_async(req); |
2303 | if (unlikely(ret)) | |
2304 | return io_submit_fail_init(sqe, req, ret); | |
2305 | ||
48863ffd | 2306 | trace_io_uring_link(req, link->head); |
f2f87370 | 2307 | link->last->link = req; |
863e0560 | 2308 | link->last = req; |
32fe525b | 2309 | |
da1a08c5 | 2310 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 2311 | return 0; |
df3becde PB |
2312 | /* last request of the link, flush it */ |
2313 | req = link->head; | |
f15a3431 | 2314 | link->head = NULL; |
924a07e4 PB |
2315 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
2316 | goto fallback; | |
2317 | ||
2318 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2319 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2320 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2321 | link->head = req; | |
2322 | link->last = req; | |
2323 | } else { | |
2324 | fallback: | |
2325 | io_queue_sqe_fallback(req); | |
2326 | } | |
f15a3431 | 2327 | return 0; |
9e645e11 | 2328 | } |
2e6e1fde | 2329 | |
f15a3431 | 2330 | io_queue_sqe(req); |
1d4240cc | 2331 | return 0; |
9e645e11 JA |
2332 | } |
2333 | ||
9a56a232 JA |
2334 | /* |
2335 | * Batched submission is done, ensure local IO is flushed out. | |
2336 | */ | |
553deffd | 2337 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2338 | { |
553deffd PB |
2339 | struct io_submit_state *state = &ctx->submit_state; |
2340 | ||
e126391c PB |
2341 | if (unlikely(state->link.head)) |
2342 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2343 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2344 | io_submit_flush_completions(ctx); |
27926b68 JA |
2345 | if (state->plug_started) |
2346 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2347 | } |
2348 | ||
2349 | /* | |
2350 | * Start submission side cache. | |
2351 | */ | |
2352 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2353 | unsigned int max_ios) |
9a56a232 | 2354 | { |
27926b68 | 2355 | state->plug_started = false; |
4b628aeb | 2356 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2357 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2358 | /* set only head, no need to init link_last in advance */ |
2359 | state->link.head = NULL; | |
9a56a232 JA |
2360 | } |
2361 | ||
2b188cc1 JA |
2362 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2363 | { | |
75b28aff | 2364 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2365 | |
caf582c6 PB |
2366 | /* |
2367 | * Ensure any loads from the SQEs are done at this point, | |
2368 | * since once we write the new head, the application could | |
2369 | * write new data to them. | |
2370 | */ | |
2371 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2372 | } |
2373 | ||
2b188cc1 | 2374 | /* |
dd9ae8a0 | 2375 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2376 | * that is mapped by userspace. This means that care needs to be taken to |
2377 | * ensure that reads are stable, as we cannot rely on userspace always | |
2378 | * being a good citizen. If members of the sqe are validated and then later | |
2379 | * used, it's important that those reads are done through READ_ONCE() to | |
2380 | * prevent a re-load down the line. | |
2381 | */ | |
b5083dfa | 2382 | static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe) |
2b188cc1 | 2383 | { |
2af89abd PB |
2384 | unsigned mask = ctx->sq_entries - 1; |
2385 | unsigned head = ctx->cached_sq_head++ & mask; | |
2386 | ||
2387 | if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) { | |
2388 | head = READ_ONCE(ctx->sq_array[head]); | |
2389 | if (unlikely(head >= ctx->sq_entries)) { | |
2390 | /* drop invalid entries */ | |
2391 | spin_lock(&ctx->completion_lock); | |
2392 | ctx->cq_extra--; | |
2393 | spin_unlock(&ctx->completion_lock); | |
2394 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2395 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
2396 | return false; | |
2397 | } | |
2398 | } | |
2b188cc1 JA |
2399 | |
2400 | /* | |
2401 | * The cached sq head (or cq tail) serves two purposes: | |
2402 | * | |
2403 | * 1) allows us to batch the cost of updating the user visible | |
2404 | * head updates. | |
2405 | * 2) allows the kernel side to track the head on its own, even | |
2406 | * though the application is the one updating it. | |
2407 | */ | |
2b188cc1 | 2408 | |
2af89abd PB |
2409 | /* double index for 128-byte SQEs, twice as long */ |
2410 | if (ctx->flags & IORING_SETUP_SQE128) | |
2411 | head <<= 1; | |
2412 | *sqe = &ctx->sq_sqes[head]; | |
2413 | return true; | |
709b302f PB |
2414 | } |
2415 | ||
17437f31 | 2416 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2417 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2418 | { |
69629809 | 2419 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2420 | unsigned int left; |
2421 | int ret; | |
6c271ce2 | 2422 | |
51d48dab | 2423 | if (unlikely(!entries)) |
69629809 | 2424 | return 0; |
ee7d46d9 | 2425 | /* make sure SQ entry isn't read before tail */ |
e3ef728f | 2426 | ret = left = min(nr, entries); |
8e6971a8 PB |
2427 | io_get_task_refs(left); |
2428 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2429 | |
69629809 | 2430 | do { |
3529d8c2 | 2431 | const struct io_uring_sqe *sqe; |
196be95c | 2432 | struct io_kiocb *req; |
fb5ccc98 | 2433 | |
c8576f3e | 2434 | if (unlikely(!io_alloc_req(ctx, &req))) |
fb5ccc98 | 2435 | break; |
b5083dfa | 2436 | if (unlikely(!io_get_sqe(ctx, &sqe))) { |
fa05457a | 2437 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2438 | break; |
2439 | } | |
6c271ce2 | 2440 | |
1cd15904 PB |
2441 | /* |
2442 | * Continue submitting even for sqe failure if the | |
2443 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2444 | */ | |
2445 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2446 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2447 | left--; | |
2448 | break; | |
bcbb7bf6 | 2449 | } |
1cd15904 | 2450 | } while (--left); |
9466f437 | 2451 | |
8e6971a8 PB |
2452 | if (unlikely(left)) { |
2453 | ret -= left; | |
2454 | /* try again if it submitted nothing and can't allocate a req */ | |
2455 | if (!ret && io_req_cache_empty(ctx)) | |
2456 | ret = -EAGAIN; | |
2457 | current->io_uring->cached_refs += left; | |
9466f437 | 2458 | } |
6c271ce2 | 2459 | |
553deffd | 2460 | io_submit_state_end(ctx); |
ae9428ca PB |
2461 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2462 | io_commit_sqring(ctx); | |
8e6971a8 | 2463 | return ret; |
6c271ce2 JA |
2464 | } |
2465 | ||
bda52162 JA |
2466 | struct io_wait_queue { |
2467 | struct wait_queue_entry wq; | |
2468 | struct io_ring_ctx *ctx; | |
5fd46178 | 2469 | unsigned cq_tail; |
bda52162 | 2470 | unsigned nr_timeouts; |
d33a39e5 | 2471 | ktime_t timeout; |
bda52162 JA |
2472 | }; |
2473 | ||
b4c98d59 DY |
2474 | static inline bool io_has_work(struct io_ring_ctx *ctx) |
2475 | { | |
c0e0d6ba | 2476 | return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) || |
490c00eb | 2477 | !llist_empty(&ctx->work_llist); |
b4c98d59 DY |
2478 | } |
2479 | ||
6c503150 | 2480 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
2481 | { |
2482 | struct io_ring_ctx *ctx = iowq->ctx; | |
0fc8c2ac | 2483 | int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail; |
bda52162 JA |
2484 | |
2485 | /* | |
d195a66e | 2486 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
2487 | * started waiting. For timeouts, we always want to return to userspace, |
2488 | * regardless of event count. | |
2489 | */ | |
5fd46178 | 2490 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
2491 | } |
2492 | ||
2493 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2494 | int wake_flags, void *key) | |
2495 | { | |
bd550173 | 2496 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq); |
bda52162 | 2497 | |
6c503150 PB |
2498 | /* |
2499 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2500 | * the task, and the next invocation will do it. | |
2501 | */ | |
bd550173 | 2502 | if (io_should_wake(iowq) || io_has_work(iowq->ctx)) |
6c503150 PB |
2503 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2504 | return -1; | |
bda52162 JA |
2505 | } |
2506 | ||
c0e0d6ba | 2507 | int io_run_task_work_sig(struct io_ring_ctx *ctx) |
af9c1a44 | 2508 | { |
1414d629 | 2509 | if (!llist_empty(&ctx->work_llist)) { |
2f413956 | 2510 | __set_current_state(TASK_RUNNING); |
1414d629 | 2511 | if (io_run_local_work(ctx) > 0) |
d246c759 | 2512 | return 0; |
1414d629 PB |
2513 | } |
2514 | if (io_run_task_work() > 0) | |
d246c759 | 2515 | return 0; |
c5020bc8 OL |
2516 | if (task_sigpending(current)) |
2517 | return -EINTR; | |
2518 | return 0; | |
af9c1a44 JA |
2519 | } |
2520 | ||
7b72d661 JA |
2521 | static bool current_pending_io(void) |
2522 | { | |
2523 | struct io_uring_task *tctx = current->io_uring; | |
2524 | ||
2525 | if (!tctx) | |
2526 | return false; | |
2527 | return percpu_counter_read_positive(&tctx->inflight); | |
2528 | } | |
2529 | ||
eeb60b9a PB |
2530 | /* when returns >0, the caller should retry */ |
2531 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
d33a39e5 | 2532 | struct io_wait_queue *iowq) |
eeb60b9a | 2533 | { |
7b72d661 | 2534 | int io_wait, ret; |
8a796565 | 2535 | |
3fcf19d5 PB |
2536 | if (unlikely(READ_ONCE(ctx->check_cq))) |
2537 | return 1; | |
846072f1 PB |
2538 | if (unlikely(!llist_empty(&ctx->work_llist))) |
2539 | return 1; | |
2540 | if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) | |
2541 | return 1; | |
2542 | if (unlikely(task_sigpending(current))) | |
2543 | return -EINTR; | |
2544 | if (unlikely(io_should_wake(iowq))) | |
2545 | return 0; | |
8a796565 AF |
2546 | |
2547 | /* | |
7b72d661 JA |
2548 | * Mark us as being in io_wait if we have pending requests, so cpufreq |
2549 | * can take into account that the task is waiting for IO - turns out | |
2550 | * to be important for low QD IO. | |
8a796565 | 2551 | */ |
7b72d661 JA |
2552 | io_wait = current->in_iowait; |
2553 | if (current_pending_io()) | |
2554 | current->in_iowait = 1; | |
8a796565 | 2555 | ret = 0; |
d33a39e5 | 2556 | if (iowq->timeout == KTIME_MAX) |
46ae7eef | 2557 | schedule(); |
d33a39e5 | 2558 | else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS)) |
8a796565 | 2559 | ret = -ETIME; |
7b72d661 | 2560 | current->in_iowait = io_wait; |
8a796565 | 2561 | return ret; |
eeb60b9a PB |
2562 | } |
2563 | ||
2b188cc1 JA |
2564 | /* |
2565 | * Wait until events become available, if we don't already have some. The | |
2566 | * application must reap them itself, as they reside on the shared cq ring. | |
2567 | */ | |
2568 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2569 | const sigset_t __user *sig, size_t sigsz, |
2570 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2571 | { |
90291099 | 2572 | struct io_wait_queue iowq; |
75b28aff | 2573 | struct io_rings *rings = ctx->rings; |
c1d5a224 | 2574 | int ret; |
2b188cc1 | 2575 | |
76de6749 PB |
2576 | if (!io_allowed_run_tw(ctx)) |
2577 | return -EEXIST; | |
140102ae PB |
2578 | if (!llist_empty(&ctx->work_llist)) |
2579 | io_run_local_work(ctx); | |
f36ba6cf PB |
2580 | io_run_task_work(); |
2581 | io_cqring_overflow_flush(ctx); | |
2582 | /* if user messes with these they will just get an early return */ | |
2583 | if (__io_cqring_events_user(ctx) >= min_events) | |
2584 | return 0; | |
2b188cc1 JA |
2585 | |
2586 | if (sig) { | |
9e75ad5d AB |
2587 | #ifdef CONFIG_COMPAT |
2588 | if (in_compat_syscall()) | |
2589 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2590 | sigsz); |
9e75ad5d AB |
2591 | else |
2592 | #endif | |
b772434b | 2593 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2594 | |
2b188cc1 JA |
2595 | if (ret) |
2596 | return ret; | |
2597 | } | |
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; | |
2612 | iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); | |
2613 | } | |
90291099 | 2614 | |
c826bd7a | 2615 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2616 | do { |
3fcf19d5 PB |
2617 | unsigned long check_cq; |
2618 | ||
130bd686 | 2619 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
8751d154 PB |
2620 | int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail); |
2621 | ||
2622 | atomic_set(&ctx->cq_wait_nr, nr_wait); | |
130bd686 PB |
2623 | set_current_state(TASK_INTERRUPTIBLE); |
2624 | } else { | |
2625 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, | |
2626 | TASK_INTERRUPTIBLE); | |
2627 | } | |
2628 | ||
d33a39e5 | 2629 | ret = io_cqring_wait_schedule(ctx, &iowq); |
130bd686 | 2630 | __set_current_state(TASK_RUNNING); |
8751d154 | 2631 | atomic_set(&ctx->cq_wait_nr, 0); |
d80c0f00 | 2632 | |
846072f1 PB |
2633 | if (ret < 0) |
2634 | break; | |
2635 | /* | |
2636 | * Run task_work after scheduling and before io_should_wake(). | |
2637 | * If we got woken because of task_work being processed, run it | |
2638 | * now rather than let the caller do another wait loop. | |
2639 | */ | |
2640 | io_run_task_work(); | |
2641 | if (!llist_empty(&ctx->work_llist)) | |
2642 | io_run_local_work(ctx); | |
3fcf19d5 PB |
2643 | |
2644 | check_cq = READ_ONCE(ctx->check_cq); | |
2645 | if (unlikely(check_cq)) { | |
2646 | /* let the caller flush overflows, retry */ | |
326a9e48 | 2647 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) |
3fcf19d5 | 2648 | io_cqring_do_overflow_flush(ctx); |
3fcf19d5 PB |
2649 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) { |
2650 | ret = -EBADR; | |
2651 | break; | |
2652 | } | |
2653 | } | |
2654 | ||
846072f1 PB |
2655 | if (io_should_wake(&iowq)) { |
2656 | ret = 0; | |
35d90f95 | 2657 | break; |
846072f1 | 2658 | } |
ca0a2651 | 2659 | cond_resched(); |
846072f1 | 2660 | } while (1); |
bda52162 | 2661 | |
130bd686 PB |
2662 | if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) |
2663 | finish_wait(&ctx->cq_wait, &iowq.wq); | |
b7db41c9 | 2664 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2665 | |
75b28aff | 2666 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2667 | } |
2668 | ||
73572984 | 2669 | static void io_mem_free(void *ptr) |
b9bd2bea | 2670 | { |
73572984 JA |
2671 | if (!ptr) |
2672 | return; | |
b9bd2bea | 2673 | |
99a9e0b8 | 2674 | folio_put(virt_to_folio(ptr)); |
b9bd2bea PB |
2675 | } |
2676 | ||
03d89a2d JA |
2677 | static void io_pages_free(struct page ***pages, int npages) |
2678 | { | |
2679 | struct page **page_array; | |
2680 | int i; | |
2681 | ||
2682 | if (!pages) | |
2683 | return; | |
8b51a395 | 2684 | |
03d89a2d | 2685 | page_array = *pages; |
8b51a395 JA |
2686 | if (!page_array) |
2687 | return; | |
2688 | ||
03d89a2d JA |
2689 | for (i = 0; i < npages; i++) |
2690 | unpin_user_page(page_array[i]); | |
2691 | kvfree(page_array); | |
2692 | *pages = NULL; | |
2693 | } | |
2694 | ||
2695 | static void *__io_uaddr_map(struct page ***pages, unsigned short *npages, | |
2696 | unsigned long uaddr, size_t size) | |
2697 | { | |
2698 | struct page **page_array; | |
2699 | unsigned int nr_pages; | |
223ef474 | 2700 | int ret, i; |
03d89a2d JA |
2701 | |
2702 | *npages = 0; | |
2703 | ||
2704 | if (uaddr & (PAGE_SIZE - 1) || !size) | |
2705 | return ERR_PTR(-EINVAL); | |
2706 | ||
2707 | nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2708 | if (nr_pages > USHRT_MAX) | |
2709 | return ERR_PTR(-EINVAL); | |
2710 | page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); | |
2711 | if (!page_array) | |
2712 | return ERR_PTR(-ENOMEM); | |
2713 | ||
2714 | ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM, | |
2715 | page_array); | |
2716 | if (ret != nr_pages) { | |
2717 | err: | |
2718 | io_pages_free(&page_array, ret > 0 ? ret : 0); | |
2719 | return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT); | |
2720 | } | |
2721 | /* | |
2722 | * Should be a single page. If the ring is small enough that we can | |
2723 | * use a normal page, that is fine. If we need multiple pages, then | |
2724 | * userspace should use a huge page. That's the only way to guarantee | |
2725 | * that we get contigious memory, outside of just being lucky or | |
2726 | * (currently) having low memory fragmentation. | |
2727 | */ | |
2728 | if (page_array[0] != page_array[ret - 1]) | |
2729 | goto err; | |
223ef474 JA |
2730 | |
2731 | /* | |
2732 | * Can't support mapping user allocated ring memory on 32-bit archs | |
2733 | * where it could potentially reside in highmem. Just fail those with | |
2734 | * -EINVAL, just like we did on kernels that didn't support this | |
2735 | * feature. | |
2736 | */ | |
2737 | for (i = 0; i < nr_pages; i++) { | |
2738 | if (PageHighMem(page_array[i])) { | |
2739 | ret = -EINVAL; | |
2740 | goto err; | |
2741 | } | |
2742 | } | |
2743 | ||
03d89a2d JA |
2744 | *pages = page_array; |
2745 | *npages = nr_pages; | |
2746 | return page_to_virt(page_array[0]); | |
2747 | } | |
2748 | ||
2749 | static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr, | |
2750 | size_t size) | |
2751 | { | |
2752 | return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr, | |
2753 | size); | |
2754 | } | |
2755 | ||
2756 | static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr, | |
2757 | size_t size) | |
2758 | { | |
2759 | return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr, | |
2760 | size); | |
2761 | } | |
2762 | ||
9c189eee JA |
2763 | static void io_rings_free(struct io_ring_ctx *ctx) |
2764 | { | |
03d89a2d JA |
2765 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) { |
2766 | io_mem_free(ctx->rings); | |
2767 | io_mem_free(ctx->sq_sqes); | |
2768 | ctx->rings = NULL; | |
2769 | ctx->sq_sqes = NULL; | |
2770 | } else { | |
2771 | io_pages_free(&ctx->ring_pages, ctx->n_ring_pages); | |
8b51a395 | 2772 | ctx->n_ring_pages = 0; |
03d89a2d | 2773 | io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages); |
8b51a395 | 2774 | ctx->n_sqe_pages = 0; |
03d89a2d | 2775 | } |
9c189eee JA |
2776 | } |
2777 | ||
73572984 | 2778 | static void *io_mem_alloc(size_t size) |
b9bd2bea | 2779 | { |
73572984 | 2780 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; |
e27cef86 | 2781 | void *ret; |
b9bd2bea | 2782 | |
e27cef86 JA |
2783 | ret = (void *) __get_free_pages(gfp, get_order(size)); |
2784 | if (ret) | |
2785 | return ret; | |
2786 | return ERR_PTR(-ENOMEM); | |
b9bd2bea PB |
2787 | } |
2788 | ||
73572984 JA |
2789 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2790 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2791 | { |
73572984 JA |
2792 | struct io_rings *rings; |
2793 | size_t off, sq_array_size; | |
6b06314c | 2794 | |
73572984 JA |
2795 | off = struct_size(rings, cqes, cq_entries); |
2796 | if (off == SIZE_MAX) | |
2797 | return SIZE_MAX; | |
2798 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2799 | if (check_shl_overflow(off, 1, &off)) | |
2800 | return SIZE_MAX; | |
2801 | } | |
ab409402 | 2802 | |
73572984 JA |
2803 | #ifdef CONFIG_SMP |
2804 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2805 | if (off == 0) | |
2806 | return SIZE_MAX; | |
2807 | #endif | |
82fbcfa9 | 2808 | |
2af89abd PB |
2809 | if (ctx->flags & IORING_SETUP_NO_SQARRAY) { |
2810 | if (sq_offset) | |
2811 | *sq_offset = SIZE_MAX; | |
2812 | return off; | |
2813 | } | |
2814 | ||
73572984 JA |
2815 | if (sq_offset) |
2816 | *sq_offset = off; | |
82fbcfa9 | 2817 | |
73572984 JA |
2818 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2819 | if (sq_array_size == SIZE_MAX) | |
2820 | return SIZE_MAX; | |
6b06314c | 2821 | |
73572984 JA |
2822 | if (check_add_overflow(off, sq_array_size, &off)) |
2823 | return SIZE_MAX; | |
8bad28d8 | 2824 | |
73572984 | 2825 | return off; |
8bad28d8 HX |
2826 | } |
2827 | ||
73572984 JA |
2828 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, |
2829 | unsigned int eventfd_async) | |
8bad28d8 | 2830 | { |
73572984 JA |
2831 | struct io_ev_fd *ev_fd; |
2832 | __s32 __user *fds = arg; | |
2833 | int fd; | |
f2303b1f | 2834 | |
73572984 JA |
2835 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, |
2836 | lockdep_is_held(&ctx->uring_lock)); | |
2837 | if (ev_fd) | |
2838 | return -EBUSY; | |
8bad28d8 | 2839 | |
73572984 JA |
2840 | if (copy_from_user(&fd, fds, sizeof(*fds))) |
2841 | return -EFAULT; | |
8dd03afe | 2842 | |
73572984 JA |
2843 | ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); |
2844 | if (!ev_fd) | |
2845 | return -ENOMEM; | |
05f3fb3c | 2846 | |
73572984 JA |
2847 | ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); |
2848 | if (IS_ERR(ev_fd->cq_ev_fd)) { | |
2849 | int ret = PTR_ERR(ev_fd->cq_ev_fd); | |
2850 | kfree(ev_fd); | |
2851 | return ret; | |
2852 | } | |
305bef98 PB |
2853 | |
2854 | spin_lock(&ctx->completion_lock); | |
2855 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
2856 | spin_unlock(&ctx->completion_lock); | |
2857 | ||
73572984 JA |
2858 | ev_fd->eventfd_async = eventfd_async; |
2859 | ctx->has_evfd = true; | |
2860 | rcu_assign_pointer(ctx->io_ev_fd, ev_fd); | |
21a091b9 DY |
2861 | atomic_set(&ev_fd->refs, 1); |
2862 | atomic_set(&ev_fd->ops, 0); | |
73572984 | 2863 | return 0; |
d7954b2b BM |
2864 | } |
2865 | ||
73572984 | 2866 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 2867 | { |
73572984 JA |
2868 | struct io_ev_fd *ev_fd; |
2869 | ||
2870 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, | |
2871 | lockdep_is_held(&ctx->uring_lock)); | |
2872 | if (ev_fd) { | |
2873 | ctx->has_evfd = false; | |
2874 | rcu_assign_pointer(ctx->io_ev_fd, NULL); | |
21a091b9 DY |
2875 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops)) |
2876 | call_rcu(&ev_fd->rcu, io_eventfd_ops); | |
73572984 JA |
2877 | return 0; |
2878 | } | |
2d091d62 | 2879 | |
73572984 | 2880 | return -ENXIO; |
44b31f2f PB |
2881 | } |
2882 | ||
73572984 | 2883 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2884 | { |
c8576f3e | 2885 | struct io_kiocb *req; |
37f0e767 | 2886 | int nr = 0; |
bf019da7 | 2887 | |
9a4fdbd8 | 2888 | mutex_lock(&ctx->uring_lock); |
34f0bc42 | 2889 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
9a4fdbd8 | 2890 | |
88ab95be | 2891 | while (!io_req_cache_empty(ctx)) { |
c8576f3e | 2892 | req = io_extract_req(ctx); |
c2b6c6bc | 2893 | kmem_cache_free(req_cachep, req); |
37f0e767 | 2894 | nr++; |
c2b6c6bc | 2895 | } |
37f0e767 PB |
2896 | if (nr) |
2897 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2898 | mutex_unlock(&ctx->uring_lock); |
2899 | } | |
2900 | ||
9eae8655 PB |
2901 | static void io_rsrc_node_cache_free(struct io_cache_entry *entry) |
2902 | { | |
2903 | kfree(container_of(entry, struct io_rsrc_node, cache)); | |
2904 | } | |
2905 | ||
c072481d | 2906 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2907 | { |
37d1e2e3 | 2908 | io_sq_thread_finish(ctx); |
43597aac | 2909 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
0b222eeb PB |
2910 | if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list))) |
2911 | return; | |
43597aac | 2912 | |
8bad28d8 | 2913 | mutex_lock(&ctx->uring_lock); |
43597aac | 2914 | if (ctx->buf_data) |
bd54b6fe | 2915 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2916 | if (ctx->file_data) |
08480400 | 2917 | __io_sqe_files_unregister(ctx); |
a85381d8 | 2918 | io_cqring_overflow_kill(ctx); |
9b402849 | 2919 | io_eventfd_unregister(ctx); |
9b797a37 | 2920 | io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); |
43e0bbbd | 2921 | io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); |
194bb58c | 2922 | io_futex_cache_free(ctx); |
5a2e745d | 2923 | io_destroy_buffers(ctx); |
b4a72c05 | 2924 | mutex_unlock(&ctx->uring_lock); |
07db298a PB |
2925 | if (ctx->sq_creds) |
2926 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2927 | if (ctx->submitter_task) |
2928 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2929 | |
a7f0ed5a PB |
2930 | /* there are no registered resources left, nobody uses it */ |
2931 | if (ctx->rsrc_node) | |
9eae8655 | 2932 | io_rsrc_node_destroy(ctx, ctx->rsrc_node); |
a7f0ed5a PB |
2933 | |
2934 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
def596e9 | 2935 | |
2b188cc1 | 2936 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
2937 | if (ctx->ring_sock) { |
2938 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 2939 | sock_release(ctx->ring_sock); |
355e8d26 | 2940 | } |
2b188cc1 | 2941 | #endif |
ef9dd637 | 2942 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
2b188cc1 | 2943 | |
9eae8655 | 2944 | io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free); |
42b6419d PB |
2945 | if (ctx->mm_account) { |
2946 | mmdrop(ctx->mm_account); | |
2947 | ctx->mm_account = NULL; | |
2948 | } | |
9c189eee | 2949 | io_rings_free(ctx); |
2b188cc1 JA |
2950 | |
2951 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2952 | free_uid(ctx->user); |
4010fec4 | 2953 | io_req_caches_free(ctx); |
e941894e JA |
2954 | if (ctx->hash_map) |
2955 | io_wq_put_hash(ctx->hash_map); | |
e6f89be6 | 2956 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2957 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 JA |
2958 | kfree(ctx->io_bl); |
2959 | xa_destroy(&ctx->io_bl_xa); | |
2b188cc1 JA |
2960 | kfree(ctx); |
2961 | } | |
2962 | ||
bca39f39 PB |
2963 | static __cold void io_activate_pollwq_cb(struct callback_head *cb) |
2964 | { | |
2965 | struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx, | |
2966 | poll_wq_task_work); | |
2967 | ||
2968 | mutex_lock(&ctx->uring_lock); | |
2969 | ctx->poll_activated = true; | |
2970 | mutex_unlock(&ctx->uring_lock); | |
2971 | ||
2972 | /* | |
2973 | * Wake ups for some events between start of polling and activation | |
2974 | * might've been lost due to loose synchronisation. | |
2975 | */ | |
2976 | wake_up_all(&ctx->poll_wq); | |
2977 | percpu_ref_put(&ctx->refs); | |
2978 | } | |
2979 | ||
2980 | static __cold void io_activate_pollwq(struct io_ring_ctx *ctx) | |
2981 | { | |
2982 | spin_lock(&ctx->completion_lock); | |
2983 | /* already activated or in progress */ | |
2984 | if (ctx->poll_activated || ctx->poll_wq_task_work.func) | |
2985 | goto out; | |
2986 | if (WARN_ON_ONCE(!ctx->task_complete)) | |
2987 | goto out; | |
2988 | if (!ctx->submitter_task) | |
2989 | goto out; | |
2990 | /* | |
2991 | * with ->submitter_task only the submitter task completes requests, we | |
2992 | * only need to sync with it, which is done by injecting a tw | |
2993 | */ | |
2994 | init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb); | |
2995 | percpu_ref_get(&ctx->refs); | |
2996 | if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL)) | |
2997 | percpu_ref_put(&ctx->refs); | |
2998 | out: | |
2999 | spin_unlock(&ctx->completion_lock); | |
3000 | } | |
3001 | ||
2b188cc1 JA |
3002 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) |
3003 | { | |
3004 | struct io_ring_ctx *ctx = file->private_data; | |
3005 | __poll_t mask = 0; | |
3006 | ||
bca39f39 PB |
3007 | if (unlikely(!ctx->poll_activated)) |
3008 | io_activate_pollwq(ctx); | |
3009 | ||
7b235dd8 | 3010 | poll_wait(file, &ctx->poll_wq, wait); |
4f7067c3 SB |
3011 | /* |
3012 | * synchronizes with barrier from wq_has_sleeper call in | |
3013 | * io_commit_cqring | |
3014 | */ | |
2b188cc1 | 3015 | smp_rmb(); |
90554200 | 3016 | if (!io_sqring_full(ctx)) |
2b188cc1 | 3017 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
3018 | |
3019 | /* | |
3020 | * Don't flush cqring overflow list here, just do a simple check. | |
3021 | * Otherwise there could possible be ABBA deadlock: | |
3022 | * CPU0 CPU1 | |
3023 | * ---- ---- | |
3024 | * lock(&ctx->uring_lock); | |
3025 | * lock(&ep->mtx); | |
3026 | * lock(&ctx->uring_lock); | |
3027 | * lock(&ep->mtx); | |
3028 | * | |
3029 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
10d8bc35 | 3030 | * pushes them to do the flush. |
ed670c3f | 3031 | */ |
b4c98d59 | 3032 | |
c10bb646 | 3033 | if (__io_cqring_events_user(ctx) || io_has_work(ctx)) |
2b188cc1 JA |
3034 | mask |= EPOLLIN | EPOLLRDNORM; |
3035 | ||
3036 | return mask; | |
3037 | } | |
3038 | ||
0bead8cd | 3039 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 3040 | { |
4379bf8b | 3041 | const struct cred *creds; |
071698e1 | 3042 | |
61cf9370 | 3043 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
3044 | if (creds) { |
3045 | put_cred(creds); | |
0bead8cd | 3046 | return 0; |
1e6fa521 | 3047 | } |
0bead8cd YD |
3048 | |
3049 | return -EINVAL; | |
3050 | } | |
3051 | ||
d56d938b PB |
3052 | struct io_tctx_exit { |
3053 | struct callback_head task_work; | |
3054 | struct completion completion; | |
baf186c4 | 3055 | struct io_ring_ctx *ctx; |
d56d938b PB |
3056 | }; |
3057 | ||
c072481d | 3058 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
3059 | { |
3060 | struct io_uring_task *tctx = current->io_uring; | |
3061 | struct io_tctx_exit *work; | |
3062 | ||
3063 | work = container_of(cb, struct io_tctx_exit, task_work); | |
3064 | /* | |
8d664282 | 3065 | * When @in_cancel, we're in cancellation and it's racy to remove the |
d56d938b | 3066 | * node. It'll be removed by the end of cancellation, just ignore it. |
998b30c3 HM |
3067 | * tctx can be NULL if the queueing of this task_work raced with |
3068 | * work cancelation off the exec path. | |
d56d938b | 3069 | */ |
8d664282 | 3070 | if (tctx && !atomic_read(&tctx->in_cancel)) |
eef51daa | 3071 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
3072 | complete(&work->completion); |
3073 | } | |
3074 | ||
c072481d | 3075 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
3076 | { |
3077 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
3078 | ||
3079 | return req->ctx == data; | |
3080 | } | |
3081 | ||
c072481d | 3082 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 3083 | { |
d56d938b | 3084 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 3085 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 3086 | unsigned long interval = HZ / 20; |
d56d938b PB |
3087 | struct io_tctx_exit exit; |
3088 | struct io_tctx_node *node; | |
3089 | int ret; | |
85faa7b8 | 3090 | |
56952e91 JA |
3091 | /* |
3092 | * If we're doing polled IO and end up having requests being | |
3093 | * submitted async (out-of-line), then completions can come in while | |
3094 | * we're waiting for refs to drop. We need to reap these manually, | |
3095 | * as nobody else will be looking for them. | |
3096 | */ | |
b2edc0a7 | 3097 | do { |
a85381d8 PB |
3098 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
3099 | mutex_lock(&ctx->uring_lock); | |
3100 | io_cqring_overflow_kill(ctx); | |
3101 | mutex_unlock(&ctx->uring_lock); | |
3102 | } | |
3103 | ||
c0e0d6ba DY |
3104 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) |
3105 | io_move_task_work_from_local(ctx); | |
3106 | ||
affa87db PB |
3107 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
3108 | cond_resched(); | |
3109 | ||
28090c13 PB |
3110 | if (ctx->sq_data) { |
3111 | struct io_sq_data *sqd = ctx->sq_data; | |
3112 | struct task_struct *tsk; | |
3113 | ||
3114 | io_sq_thread_park(sqd); | |
3115 | tsk = sqd->thread; | |
3116 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
3117 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
3118 | io_cancel_ctx_cb, ctx, true); | |
3119 | io_sq_thread_unpark(sqd); | |
3120 | } | |
b5bb3a24 | 3121 | |
37f0e767 PB |
3122 | io_req_caches_free(ctx); |
3123 | ||
58d3be2c PB |
3124 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
3125 | /* there is little hope left, don't run it too often */ | |
3126 | interval = HZ * 60; | |
3127 | } | |
4826c594 JA |
3128 | /* |
3129 | * This is really an uninterruptible wait, as it has to be | |
3130 | * complete. But it's also run from a kworker, which doesn't | |
3131 | * take signals, so it's fine to make it interruptible. This | |
3132 | * avoids scenarios where we knowingly can wait much longer | |
3133 | * on completions, for example if someone does a SIGSTOP on | |
3134 | * a task that needs to finish task_work to make this loop | |
3135 | * complete. That's a synthetic situation that should not | |
3136 | * cause a stuck task backtrace, and hence a potential panic | |
3137 | * on stuck tasks if that is enabled. | |
3138 | */ | |
3139 | } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 3140 | |
7f00651a PB |
3141 | init_completion(&exit.completion); |
3142 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
3143 | exit.ctx = ctx; | |
89b5066e PB |
3144 | /* |
3145 | * Some may use context even when all refs and requests have been put, | |
3146 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 3147 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
3148 | * this lock/unlock section also waits them to finish. |
3149 | */ | |
d56d938b PB |
3150 | mutex_lock(&ctx->uring_lock); |
3151 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
3152 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
3153 | ||
d56d938b PB |
3154 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
3155 | ctx_node); | |
7f00651a PB |
3156 | /* don't spin on a single task if cancellation failed */ |
3157 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
3158 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
3159 | if (WARN_ON_ONCE(ret)) | |
3160 | continue; | |
d56d938b PB |
3161 | |
3162 | mutex_unlock(&ctx->uring_lock); | |
4826c594 JA |
3163 | /* |
3164 | * See comment above for | |
3165 | * wait_for_completion_interruptible_timeout() on why this | |
3166 | * wait is marked as interruptible. | |
3167 | */ | |
3168 | wait_for_completion_interruptible(&exit.completion); | |
d56d938b PB |
3169 | mutex_lock(&ctx->uring_lock); |
3170 | } | |
3171 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
3172 | spin_lock(&ctx->completion_lock); |
3173 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 3174 | |
d73a572d PB |
3175 | /* pairs with RCU read section in io_req_local_work_add() */ |
3176 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3177 | synchronize_rcu(); | |
3178 | ||
85faa7b8 JA |
3179 | io_ring_ctx_free(ctx); |
3180 | } | |
3181 | ||
c072481d | 3182 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 3183 | { |
61cf9370 MWO |
3184 | unsigned long index; |
3185 | struct creds *creds; | |
3186 | ||
2b188cc1 JA |
3187 | mutex_lock(&ctx->uring_lock); |
3188 | percpu_ref_kill(&ctx->refs); | |
61cf9370 MWO |
3189 | xa_for_each(&ctx->personalities, index, creds) |
3190 | io_unregister_personality(ctx, index); | |
9ca9fb24 PB |
3191 | if (ctx->rings) |
3192 | io_poll_remove_all(ctx, NULL, true); | |
2b188cc1 JA |
3193 | mutex_unlock(&ctx->uring_lock); |
3194 | ||
02bac94b PB |
3195 | /* |
3196 | * If we failed setting up the ctx, we might not have any rings | |
3197 | * and therefore did not submit any requests | |
3198 | */ | |
3199 | if (ctx->rings) | |
60053be8 | 3200 | io_kill_timeouts(ctx, NULL, true); |
309fc03a | 3201 | |
dfbe5561 JA |
3202 | flush_delayed_work(&ctx->fallback_work); |
3203 | ||
85faa7b8 | 3204 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
3205 | /* |
3206 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
3207 | * if we're exiting a ton of rings at the same time. It just adds | |
3208 | * noise and overhead, there's no discernable change in runtime | |
3209 | * over using system_wq. | |
3210 | */ | |
3211 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
3212 | } |
3213 | ||
3214 | static int io_uring_release(struct inode *inode, struct file *file) | |
3215 | { | |
3216 | struct io_ring_ctx *ctx = file->private_data; | |
3217 | ||
3218 | file->private_data = NULL; | |
3219 | io_ring_ctx_wait_and_kill(ctx); | |
3220 | return 0; | |
3221 | } | |
3222 | ||
f6edbabb PB |
3223 | struct io_task_cancel { |
3224 | struct task_struct *task; | |
3dd0c97a | 3225 | bool all; |
f6edbabb | 3226 | }; |
f254ac04 | 3227 | |
f6edbabb | 3228 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 3229 | { |
9a472ef7 | 3230 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 3231 | struct io_task_cancel *cancel = data; |
9a472ef7 | 3232 | |
6af3f48b | 3233 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
3234 | } |
3235 | ||
c072481d PB |
3236 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
3237 | struct task_struct *task, | |
3238 | bool cancel_all) | |
b7ddce3c | 3239 | { |
e1915f76 | 3240 | struct io_defer_entry *de; |
b7ddce3c PB |
3241 | LIST_HEAD(list); |
3242 | ||
79ebeaee | 3243 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 3244 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 3245 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
3246 | list_cut_position(&list, &ctx->defer_list, &de->list); |
3247 | break; | |
3248 | } | |
3249 | } | |
79ebeaee | 3250 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
3251 | if (list_empty(&list)) |
3252 | return false; | |
b7ddce3c PB |
3253 | |
3254 | while (!list_empty(&list)) { | |
3255 | de = list_first_entry(&list, struct io_defer_entry, list); | |
3256 | list_del_init(&de->list); | |
e276ae34 | 3257 | io_req_task_queue_fail(de->req, -ECANCELED); |
b7ddce3c PB |
3258 | kfree(de); |
3259 | } | |
e1915f76 | 3260 | return true; |
b7ddce3c PB |
3261 | } |
3262 | ||
c072481d | 3263 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
3264 | { |
3265 | struct io_tctx_node *node; | |
3266 | enum io_wq_cancel cret; | |
3267 | bool ret = false; | |
3268 | ||
3269 | mutex_lock(&ctx->uring_lock); | |
3270 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3271 | struct io_uring_task *tctx = node->task->io_uring; | |
3272 | ||
3273 | /* | |
3274 | * io_wq will stay alive while we hold uring_lock, because it's | |
3275 | * killed after ctx nodes, which requires to take the lock. | |
3276 | */ | |
3277 | if (!tctx || !tctx->io_wq) | |
3278 | continue; | |
3279 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
3280 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3281 | } | |
3282 | mutex_unlock(&ctx->uring_lock); | |
3283 | ||
3284 | return ret; | |
3285 | } | |
3286 | ||
93b8cc60 ML |
3287 | static bool io_uring_try_cancel_uring_cmd(struct io_ring_ctx *ctx, |
3288 | struct task_struct *task, bool cancel_all) | |
3289 | { | |
3290 | struct hlist_node *tmp; | |
3291 | struct io_kiocb *req; | |
3292 | bool ret = false; | |
3293 | ||
3294 | lockdep_assert_held(&ctx->uring_lock); | |
3295 | ||
3296 | hlist_for_each_entry_safe(req, tmp, &ctx->cancelable_uring_cmd, | |
3297 | hash_node) { | |
3298 | struct io_uring_cmd *cmd = io_kiocb_to_cmd(req, | |
3299 | struct io_uring_cmd); | |
3300 | struct file *file = req->file; | |
3301 | ||
3302 | if (!cancel_all && req->task != task) | |
3303 | continue; | |
3304 | ||
3305 | if (cmd->flags & IORING_URING_CMD_CANCELABLE) { | |
3306 | /* ->sqe isn't available if no async data */ | |
3307 | if (!req_has_async_data(req)) | |
3308 | cmd->sqe = NULL; | |
3309 | file->f_op->uring_cmd(cmd, IO_URING_F_CANCEL); | |
3310 | ret = true; | |
3311 | } | |
3312 | } | |
3313 | io_submit_flush_completions(ctx); | |
3314 | ||
3315 | return ret; | |
3316 | } | |
3317 | ||
affa87db | 3318 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
3319 | struct task_struct *task, |
3320 | bool cancel_all) | |
9936c7c2 | 3321 | { |
3dd0c97a | 3322 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 3323 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
3324 | enum io_wq_cancel cret; |
3325 | bool ret = false; | |
9936c7c2 | 3326 | |
360cd42c PB |
3327 | /* set it so io_req_local_work_add() would wake us up */ |
3328 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { | |
3329 | atomic_set(&ctx->cq_wait_nr, 1); | |
3330 | smp_mb(); | |
3331 | } | |
3332 | ||
60053be8 PB |
3333 | /* failed during ring init, it couldn't have issued any requests */ |
3334 | if (!ctx->rings) | |
affa87db | 3335 | return false; |
60053be8 | 3336 | |
affa87db PB |
3337 | if (!task) { |
3338 | ret |= io_uring_try_cancel_iowq(ctx); | |
3339 | } else if (tctx && tctx->io_wq) { | |
3340 | /* | |
3341 | * Cancels requests of all rings, not only @ctx, but | |
3342 | * it's fine as the task is in exit/exec. | |
3343 | */ | |
3344 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
3345 | &cancel, true); | |
3346 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3347 | } | |
9936c7c2 | 3348 | |
affa87db PB |
3349 | /* SQPOLL thread does its own polling */ |
3350 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
3351 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
3352 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
3353 | io_iopoll_try_reap_events(ctx); | |
3354 | ret = true; | |
fcc926bb | 3355 | cond_resched(); |
9936c7c2 | 3356 | } |
9936c7c2 | 3357 | } |
affa87db | 3358 | |
140102ae PB |
3359 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3360 | io_allowed_defer_tw_run(ctx)) | |
c0e0d6ba | 3361 | ret |= io_run_local_work(ctx) > 0; |
affa87db PB |
3362 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
3363 | mutex_lock(&ctx->uring_lock); | |
3364 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
f31ecf67 | 3365 | ret |= io_waitid_remove_all(ctx, task, cancel_all); |
194bb58c | 3366 | ret |= io_futex_remove_all(ctx, task, cancel_all); |
93b8cc60 | 3367 | ret |= io_uring_try_cancel_uring_cmd(ctx, task, cancel_all); |
affa87db PB |
3368 | mutex_unlock(&ctx->uring_lock); |
3369 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
3370 | if (task) | |
c0e0d6ba | 3371 | ret |= io_run_task_work() > 0; |
affa87db | 3372 | return ret; |
9936c7c2 PB |
3373 | } |
3374 | ||
3f48cf18 | 3375 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 3376 | { |
3f48cf18 | 3377 | if (tracked) |
9cae36a0 | 3378 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
3379 | return percpu_counter_sum(&tctx->inflight); |
3380 | } | |
3381 | ||
78cc687b PB |
3382 | /* |
3383 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 3384 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 3385 | */ |
17437f31 | 3386 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 3387 | { |
521d6a73 | 3388 | struct io_uring_task *tctx = current->io_uring; |
734551df | 3389 | struct io_ring_ctx *ctx; |
360cd42c PB |
3390 | struct io_tctx_node *node; |
3391 | unsigned long index; | |
0e9ddb39 PB |
3392 | s64 inflight; |
3393 | DEFINE_WAIT(wait); | |
fdaf083c | 3394 | |
78cc687b PB |
3395 | WARN_ON_ONCE(sqd && sqd->thread != current); |
3396 | ||
6d042ffb PO |
3397 | if (!current->io_uring) |
3398 | return; | |
17a91051 PB |
3399 | if (tctx->io_wq) |
3400 | io_wq_exit_start(tctx->io_wq); | |
3401 | ||
8d664282 | 3402 | atomic_inc(&tctx->in_cancel); |
0e9ddb39 | 3403 | do { |
affa87db PB |
3404 | bool loop = false; |
3405 | ||
e9dbe221 | 3406 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 3407 | /* read completions before cancelations */ |
78cc687b | 3408 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
3409 | if (!inflight) |
3410 | break; | |
fdaf083c | 3411 | |
78cc687b | 3412 | if (!sqd) { |
78cc687b PB |
3413 | xa_for_each(&tctx->xa, index, node) { |
3414 | /* sqpoll task will cancel all its requests */ | |
3415 | if (node->ctx->sq_data) | |
3416 | continue; | |
affa87db PB |
3417 | loop |= io_uring_try_cancel_requests(node->ctx, |
3418 | current, cancel_all); | |
78cc687b PB |
3419 | } |
3420 | } else { | |
3421 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
3422 | loop |= io_uring_try_cancel_requests(ctx, |
3423 | current, | |
3424 | cancel_all); | |
3425 | } | |
3426 | ||
3427 | if (loop) { | |
3428 | cond_resched(); | |
3429 | continue; | |
78cc687b | 3430 | } |
17a91051 | 3431 | |
78a78060 JA |
3432 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
3433 | io_run_task_work(); | |
e9dbe221 | 3434 | io_uring_drop_tctx_refs(current); |
360cd42c PB |
3435 | xa_for_each(&tctx->xa, index, node) { |
3436 | if (!llist_empty(&node->ctx->work_llist)) { | |
3437 | WARN_ON_ONCE(node->ctx->submitter_task && | |
3438 | node->ctx->submitter_task != current); | |
3439 | goto end_wait; | |
3440 | } | |
3441 | } | |
0f212204 | 3442 | /* |
a1bb3cd5 PB |
3443 | * If we've seen completions, retry without waiting. This |
3444 | * avoids a race where a completion comes in before we did | |
3445 | * prepare_to_wait(). | |
0f212204 | 3446 | */ |
3dd0c97a | 3447 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 3448 | schedule(); |
360cd42c | 3449 | end_wait: |
f57555ed | 3450 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 3451 | } while (1); |
de7f1d9e | 3452 | |
8452d4a6 | 3453 | io_uring_clean_tctx(tctx); |
3dd0c97a | 3454 | if (cancel_all) { |
3cc7fdb9 PB |
3455 | /* |
3456 | * We shouldn't run task_works after cancel, so just leave | |
8d664282 | 3457 | * ->in_cancel set for normal exit. |
3cc7fdb9 | 3458 | */ |
8d664282 | 3459 | atomic_dec(&tctx->in_cancel); |
3f48cf18 PB |
3460 | /* for exec all current's requests should be gone, kill tctx */ |
3461 | __io_uring_free(current); | |
3462 | } | |
44e728b8 PB |
3463 | } |
3464 | ||
f552a27a | 3465 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 3466 | { |
f552a27a | 3467 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
3468 | } |
3469 | ||
6c5c240e RP |
3470 | static void *io_uring_validate_mmap_request(struct file *file, |
3471 | loff_t pgoff, size_t sz) | |
2b188cc1 | 3472 | { |
2b188cc1 | 3473 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 3474 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
3475 | struct page *page; |
3476 | void *ptr; | |
3477 | ||
03d89a2d JA |
3478 | /* Don't allow mmap if the ring was setup without it */ |
3479 | if (ctx->flags & IORING_SETUP_NO_MMAP) | |
3480 | return ERR_PTR(-EINVAL); | |
3481 | ||
c56e022c | 3482 | switch (offset & IORING_OFF_MMAP_MASK) { |
2b188cc1 | 3483 | case IORING_OFF_SQ_RING: |
75b28aff HV |
3484 | case IORING_OFF_CQ_RING: |
3485 | ptr = ctx->rings; | |
2b188cc1 JA |
3486 | break; |
3487 | case IORING_OFF_SQES: | |
3488 | ptr = ctx->sq_sqes; | |
3489 | break; | |
c56e022c JA |
3490 | case IORING_OFF_PBUF_RING: { |
3491 | unsigned int bgid; | |
3492 | ||
3493 | bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT; | |
3494 | mutex_lock(&ctx->uring_lock); | |
3495 | ptr = io_pbuf_get_address(ctx, bgid); | |
3496 | mutex_unlock(&ctx->uring_lock); | |
3497 | if (!ptr) | |
3498 | return ERR_PTR(-EINVAL); | |
3499 | break; | |
3500 | } | |
2b188cc1 | 3501 | default: |
6c5c240e | 3502 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
3503 | } |
3504 | ||
3505 | page = virt_to_head_page(ptr); | |
a50b854e | 3506 | if (sz > page_size(page)) |
6c5c240e RP |
3507 | return ERR_PTR(-EINVAL); |
3508 | ||
3509 | return ptr; | |
3510 | } | |
3511 | ||
3512 | #ifdef CONFIG_MMU | |
3513 | ||
c072481d | 3514 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e RP |
3515 | { |
3516 | size_t sz = vma->vm_end - vma->vm_start; | |
3517 | unsigned long pfn; | |
3518 | void *ptr; | |
3519 | ||
3520 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
3521 | if (IS_ERR(ptr)) | |
3522 | return PTR_ERR(ptr); | |
2b188cc1 JA |
3523 | |
3524 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3525 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3526 | } | |
3527 | ||
d808459b HD |
3528 | static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp, |
3529 | unsigned long addr, unsigned long len, | |
3530 | unsigned long pgoff, unsigned long flags) | |
3531 | { | |
d808459b HD |
3532 | void *ptr; |
3533 | ||
3534 | /* | |
3535 | * Do not allow to map to user-provided address to avoid breaking the | |
3536 | * aliasing rules. Userspace is not able to guess the offset address of | |
3537 | * kernel kmalloc()ed memory area. | |
3538 | */ | |
3539 | if (addr) | |
3540 | return -EINVAL; | |
3541 | ||
3542 | ptr = io_uring_validate_mmap_request(filp, pgoff, len); | |
3543 | if (IS_ERR(ptr)) | |
3544 | return -ENOMEM; | |
3545 | ||
32832a40 HD |
3546 | /* |
3547 | * Some architectures have strong cache aliasing requirements. | |
3548 | * For such architectures we need a coherent mapping which aliases | |
3549 | * kernel memory *and* userspace memory. To achieve that: | |
3550 | * - use a NULL file pointer to reference physical memory, and | |
3551 | * - use the kernel virtual address of the shared io_uring context | |
3552 | * (instead of the userspace-provided address, which has to be 0UL | |
3553 | * anyway). | |
56675f8b HD |
3554 | * - use the same pgoff which the get_unmapped_area() uses to |
3555 | * calculate the page colouring. | |
32832a40 HD |
3556 | * For architectures without such aliasing requirements, the |
3557 | * architecture will return any suitable mapping because addr is 0. | |
3558 | */ | |
3559 | filp = NULL; | |
3560 | flags |= MAP_SHARED; | |
3561 | pgoff = 0; /* has been translated to ptr above */ | |
d808459b | 3562 | #ifdef SHM_COLOUR |
32832a40 | 3563 | addr = (uintptr_t) ptr; |
56675f8b | 3564 | pgoff = addr >> PAGE_SHIFT; |
d808459b | 3565 | #else |
32832a40 | 3566 | addr = 0UL; |
d808459b | 3567 | #endif |
32832a40 | 3568 | return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); |
d808459b HD |
3569 | } |
3570 | ||
6c5c240e RP |
3571 | #else /* !CONFIG_MMU */ |
3572 | ||
3573 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3574 | { | |
fc4f4be9 | 3575 | return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL; |
6c5c240e RP |
3576 | } |
3577 | ||
3578 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
3579 | { | |
3580 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
3581 | } | |
3582 | ||
3583 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
3584 | unsigned long addr, unsigned long len, | |
3585 | unsigned long pgoff, unsigned long flags) | |
3586 | { | |
3587 | void *ptr; | |
3588 | ||
3589 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
3590 | if (IS_ERR(ptr)) | |
3591 | return PTR_ERR(ptr); | |
3592 | ||
3593 | return (unsigned long) ptr; | |
3594 | } | |
3595 | ||
3596 | #endif /* !CONFIG_MMU */ | |
3597 | ||
f81440d3 PB |
3598 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
3599 | { | |
3600 | if (flags & IORING_ENTER_EXT_ARG) { | |
3601 | struct io_uring_getevents_arg arg; | |
3602 | ||
3603 | if (argsz != sizeof(arg)) | |
3604 | return -EINVAL; | |
3605 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3606 | return -EFAULT; | |
3607 | } | |
3608 | return 0; | |
3609 | } | |
3610 | ||
c73ebb68 HX |
3611 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
3612 | struct __kernel_timespec __user **ts, | |
3613 | const sigset_t __user **sig) | |
3614 | { | |
3615 | struct io_uring_getevents_arg arg; | |
3616 | ||
3617 | /* | |
3618 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
3619 | * is just a pointer to the sigset_t. | |
3620 | */ | |
3621 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
3622 | *sig = (const sigset_t __user *) argp; | |
3623 | *ts = NULL; | |
3624 | return 0; | |
3625 | } | |
3626 | ||
3627 | /* | |
3628 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
3629 | * timespec and sigset_t pointers if good. | |
3630 | */ | |
3631 | if (*argsz != sizeof(arg)) | |
3632 | return -EINVAL; | |
3633 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3634 | return -EFAULT; | |
d2347b96 DY |
3635 | if (arg.pad) |
3636 | return -EINVAL; | |
c73ebb68 HX |
3637 | *sig = u64_to_user_ptr(arg.sigmask); |
3638 | *argsz = arg.sigmask_sz; | |
3639 | *ts = u64_to_user_ptr(arg.ts); | |
3640 | return 0; | |
3641 | } | |
3642 | ||
2b188cc1 | 3643 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
3644 | u32, min_complete, u32, flags, const void __user *, argp, |
3645 | size_t, argsz) | |
2b188cc1 JA |
3646 | { |
3647 | struct io_ring_ctx *ctx; | |
2b188cc1 | 3648 | struct fd f; |
33f993da | 3649 | long ret; |
2b188cc1 | 3650 | |
33f993da | 3651 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
3652 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
3653 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
3654 | return -EINVAL; |
3655 | ||
e7a6c00d JA |
3656 | /* |
3657 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3658 | * need only dereference our task private array to find it. | |
3659 | */ | |
3660 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3661 | struct io_uring_task *tctx = current->io_uring; | |
3662 | ||
3273c440 | 3663 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
e7a6c00d JA |
3664 | return -EINVAL; |
3665 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
3666 | f.file = tctx->registered_rings[fd]; | |
4329490a | 3667 | f.flags = 0; |
3273c440 PB |
3668 | if (unlikely(!f.file)) |
3669 | return -EBADF; | |
e7a6c00d JA |
3670 | } else { |
3671 | f = fdget(fd); | |
3273c440 PB |
3672 | if (unlikely(!f.file)) |
3673 | return -EBADF; | |
3674 | ret = -EOPNOTSUPP; | |
3675 | if (unlikely(!io_is_uring_fops(f.file))) | |
fbb8bb02 | 3676 | goto out; |
e7a6c00d | 3677 | } |
2b188cc1 | 3678 | |
2b188cc1 | 3679 | ctx = f.file->private_data; |
7e84e1c7 | 3680 | ret = -EBADFD; |
33f993da | 3681 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3682 | goto out; |
3683 | ||
6c271ce2 JA |
3684 | /* |
3685 | * For SQ polling, the thread will do all submissions and completions. | |
3686 | * Just return the requested submit count, and wake the thread if | |
3687 | * we were asked to. | |
3688 | */ | |
b2a9eada | 3689 | ret = 0; |
6c271ce2 | 3690 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3691 | io_cqring_overflow_flush(ctx); |
89448c47 | 3692 | |
21f96522 JA |
3693 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3694 | ret = -EOWNERDEAD; | |
04147488 | 3695 | goto out; |
21f96522 | 3696 | } |
6c271ce2 | 3697 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3698 | wake_up(&ctx->sq_data->wait); |
88b80534 QF |
3699 | if (flags & IORING_ENTER_SQ_WAIT) |
3700 | io_sqpoll_wait_sq(ctx); | |
3701 | ||
3e813c90 | 3702 | ret = to_submit; |
b2a9eada | 3703 | } else if (to_submit) { |
eef51daa | 3704 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3705 | if (unlikely(ret)) |
3706 | goto out; | |
7c504e65 | 3707 | |
2b188cc1 | 3708 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3709 | ret = io_submit_sqes(ctx, to_submit); |
3710 | if (ret != to_submit) { | |
d487b43c | 3711 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3712 | goto out; |
d487b43c | 3713 | } |
44f87745 PB |
3714 | if (flags & IORING_ENTER_GETEVENTS) { |
3715 | if (ctx->syscall_iopoll) | |
3716 | goto iopoll_locked; | |
3717 | /* | |
3718 | * Ignore errors, we'll soon call io_cqring_wait() and | |
3719 | * it should handle ownership problems if any. | |
3720 | */ | |
3721 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3722 | (void)io_run_local_work_locked(ctx); | |
3723 | } | |
d487b43c | 3724 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 | 3725 | } |
c0e0d6ba | 3726 | |
2b188cc1 | 3727 | if (flags & IORING_ENTER_GETEVENTS) { |
3e813c90 | 3728 | int ret2; |
c0e0d6ba | 3729 | |
773697b6 | 3730 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3731 | /* |
3732 | * We disallow the app entering submit/complete with | |
3733 | * polling, but we still need to lock the ring to | |
3734 | * prevent racing with polled issue that got punted to | |
3735 | * a workqueue. | |
3736 | */ | |
3737 | mutex_lock(&ctx->uring_lock); | |
3738 | iopoll_locked: | |
3e813c90 DY |
3739 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3740 | if (likely(!ret2)) { | |
3741 | min_complete = min(min_complete, | |
3742 | ctx->cq_entries); | |
3743 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3744 | } |
3745 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3746 | } else { |
f81440d3 PB |
3747 | const sigset_t __user *sig; |
3748 | struct __kernel_timespec __user *ts; | |
3749 | ||
3e813c90 DY |
3750 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3751 | if (likely(!ret2)) { | |
3752 | min_complete = min(min_complete, | |
3753 | ctx->cq_entries); | |
3754 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3755 | argsz, ts); | |
3756 | } | |
def596e9 | 3757 | } |
c73ebb68 | 3758 | |
155bc950 | 3759 | if (!ret) { |
3e813c90 | 3760 | ret = ret2; |
2b188cc1 | 3761 | |
155bc950 DY |
3762 | /* |
3763 | * EBADR indicates that one or more CQE were dropped. | |
3764 | * Once the user has been informed we can clear the bit | |
3765 | * as they are obviously ok with those drops. | |
3766 | */ | |
3767 | if (unlikely(ret2 == -EBADR)) | |
3768 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3769 | &ctx->check_cq); | |
def596e9 | 3770 | } |
2b188cc1 | 3771 | } |
7c504e65 | 3772 | out: |
4329490a | 3773 | fdput(f); |
3e813c90 | 3774 | return ret; |
2b188cc1 JA |
3775 | } |
3776 | ||
3777 | static const struct file_operations io_uring_fops = { | |
3778 | .release = io_uring_release, | |
3779 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3780 | #ifndef CONFIG_MMU |
3781 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3782 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
d808459b HD |
3783 | #else |
3784 | .get_unmapped_area = io_uring_mmu_get_unmapped_area, | |
6c5c240e | 3785 | #endif |
2b188cc1 | 3786 | .poll = io_uring_poll, |
bebdb65e | 3787 | #ifdef CONFIG_PROC_FS |
87ce955b | 3788 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3789 | #endif |
2b188cc1 JA |
3790 | }; |
3791 | ||
92ac8bea JA |
3792 | bool io_is_uring_fops(struct file *file) |
3793 | { | |
3794 | return file->f_op == &io_uring_fops; | |
3795 | } | |
3796 | ||
c072481d PB |
3797 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3798 | struct io_uring_params *p) | |
2b188cc1 | 3799 | { |
75b28aff HV |
3800 | struct io_rings *rings; |
3801 | size_t size, sq_array_offset; | |
e27cef86 | 3802 | void *ptr; |
2b188cc1 | 3803 | |
bd740481 JA |
3804 | /* make sure these are sane, as we already accounted them */ |
3805 | ctx->sq_entries = p->sq_entries; | |
3806 | ctx->cq_entries = p->cq_entries; | |
3807 | ||
baf9cb64 | 3808 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3809 | if (size == SIZE_MAX) |
3810 | return -EOVERFLOW; | |
3811 | ||
03d89a2d JA |
3812 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3813 | rings = io_mem_alloc(size); | |
3814 | else | |
3815 | rings = io_rings_map(ctx, p->cq_off.user_addr, size); | |
3816 | ||
e27cef86 JA |
3817 | if (IS_ERR(rings)) |
3818 | return PTR_ERR(rings); | |
2b188cc1 | 3819 | |
75b28aff | 3820 | ctx->rings = rings; |
2af89abd PB |
3821 | if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) |
3822 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
75b28aff HV |
3823 | rings->sq_ring_mask = p->sq_entries - 1; |
3824 | rings->cq_ring_mask = p->cq_entries - 1; | |
3825 | rings->sq_ring_entries = p->sq_entries; | |
3826 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3827 | |
ebdeb7c0 JA |
3828 | if (p->flags & IORING_SETUP_SQE128) |
3829 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3830 | else | |
3831 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 | 3832 | if (size == SIZE_MAX) { |
9c189eee | 3833 | io_rings_free(ctx); |
2b188cc1 | 3834 | return -EOVERFLOW; |
eb065d30 | 3835 | } |
2b188cc1 | 3836 | |
03d89a2d JA |
3837 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3838 | ptr = io_mem_alloc(size); | |
3839 | else | |
3840 | ptr = io_sqes_map(ctx, p->sq_off.user_addr, size); | |
3841 | ||
e27cef86 | 3842 | if (IS_ERR(ptr)) { |
9c189eee | 3843 | io_rings_free(ctx); |
e27cef86 | 3844 | return PTR_ERR(ptr); |
eb065d30 | 3845 | } |
2b188cc1 | 3846 | |
e27cef86 | 3847 | ctx->sq_sqes = ptr; |
2b188cc1 JA |
3848 | return 0; |
3849 | } | |
3850 | ||
6e76ac59 | 3851 | static int io_uring_install_fd(struct file *file) |
9faadcc8 | 3852 | { |
6e76ac59 | 3853 | int fd; |
9faadcc8 PB |
3854 | |
3855 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3856 | if (fd < 0) | |
3857 | return fd; | |
9faadcc8 PB |
3858 | fd_install(fd, file); |
3859 | return fd; | |
3860 | } | |
3861 | ||
2b188cc1 JA |
3862 | /* |
3863 | * Allocate an anonymous fd, this is what constitutes the application | |
3864 | * visible backing of an io_uring instance. The application mmaps this | |
3865 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3866 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3867 | */ | |
9faadcc8 | 3868 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
3869 | { |
3870 | struct file *file; | |
9faadcc8 | 3871 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
3872 | int ret; |
3873 | ||
2b188cc1 JA |
3874 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
3875 | &ctx->ring_sock); | |
3876 | if (ret) | |
9faadcc8 | 3877 | return ERR_PTR(ret); |
2b188cc1 JA |
3878 | #endif |
3879 | ||
91a9ab7c PM |
3880 | file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, |
3881 | O_RDWR | O_CLOEXEC, NULL); | |
2b188cc1 | 3882 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
3883 | if (IS_ERR(file)) { |
3884 | sock_release(ctx->ring_sock); | |
3885 | ctx->ring_sock = NULL; | |
3886 | } else { | |
3887 | ctx->ring_sock->file = file; | |
0f212204 | 3888 | } |
2b188cc1 | 3889 | #endif |
9faadcc8 | 3890 | return file; |
2b188cc1 JA |
3891 | } |
3892 | ||
c072481d PB |
3893 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3894 | struct io_uring_params __user *params) | |
2b188cc1 | 3895 | { |
2b188cc1 | 3896 | struct io_ring_ctx *ctx; |
6e76ac59 | 3897 | struct io_uring_task *tctx; |
9faadcc8 | 3898 | struct file *file; |
2b188cc1 JA |
3899 | int ret; |
3900 | ||
8110c1a6 | 3901 | if (!entries) |
2b188cc1 | 3902 | return -EINVAL; |
8110c1a6 JA |
3903 | if (entries > IORING_MAX_ENTRIES) { |
3904 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3905 | return -EINVAL; | |
3906 | entries = IORING_MAX_ENTRIES; | |
3907 | } | |
2b188cc1 | 3908 | |
6e76ac59 JT |
3909 | if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY) |
3910 | && !(p->flags & IORING_SETUP_NO_MMAP)) | |
3911 | return -EINVAL; | |
3912 | ||
2b188cc1 JA |
3913 | /* |
3914 | * Use twice as many entries for the CQ ring. It's possible for the | |
3915 | * application to drive a higher depth than the size of the SQ ring, | |
3916 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3917 | * some flexibility in overcommitting a bit. If the application has |
3918 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3919 | * of CQ ring entries manually. | |
2b188cc1 JA |
3920 | */ |
3921 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3922 | if (p->flags & IORING_SETUP_CQSIZE) { |
3923 | /* | |
3924 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3925 | * to a power-of-two, if it isn't already. We do NOT impose | |
3926 | * any cq vs sq ring sizing. | |
3927 | */ | |
eb2667b3 | 3928 | if (!p->cq_entries) |
33a107f0 | 3929 | return -EINVAL; |
8110c1a6 JA |
3930 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3931 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3932 | return -EINVAL; | |
3933 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3934 | } | |
eb2667b3 JQ |
3935 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3936 | if (p->cq_entries < p->sq_entries) | |
3937 | return -EINVAL; | |
33a107f0 JA |
3938 | } else { |
3939 | p->cq_entries = 2 * p->sq_entries; | |
3940 | } | |
2b188cc1 | 3941 | |
2b188cc1 | 3942 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3943 | if (!ctx) |
2b188cc1 | 3944 | return -ENOMEM; |
773697b6 | 3945 | |
e6aeb272 PB |
3946 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3947 | !(ctx->flags & IORING_SETUP_IOPOLL) && | |
3948 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3949 | ctx->task_complete = true; | |
3950 | ||
ec26c225 PB |
3951 | if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL)) |
3952 | ctx->lockless_cq = true; | |
3953 | ||
bca39f39 PB |
3954 | /* |
3955 | * lazy poll_wq activation relies on ->task_complete for synchronisation | |
3956 | * purposes, see io_activate_pollwq() | |
3957 | */ | |
3958 | if (!ctx->task_complete) | |
3959 | ctx->poll_activated = true; | |
3960 | ||
773697b6 PB |
3961 | /* |
3962 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3963 | * space applications don't need to do io completion events | |
3964 | * polling again, they can rely on io_sq_thread to do polling | |
3965 | * work, which can reduce cpu usage and uring_lock contention. | |
3966 | */ | |
3967 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3968 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3969 | ctx->syscall_iopoll = 1; | |
3970 | ||
2b188cc1 | 3971 | ctx->compat = in_compat_syscall(); |
6adc2272 | 3972 | if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK)) |
62e398be | 3973 | ctx->user = get_uid(current_user()); |
2aede0e4 | 3974 | |
9f010507 | 3975 | /* |
e1169f06 JA |
3976 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3977 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3978 | */ |
e1169f06 JA |
3979 | ret = -EINVAL; |
3980 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3981 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 | 3982 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
c0e0d6ba DY |
3983 | IORING_SETUP_TASKRUN_FLAG | |
3984 | IORING_SETUP_DEFER_TASKRUN)) | |
e1169f06 | 3985 | goto err; |
9f010507 | 3986 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3987 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3988 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3989 | } else { | |
c0e0d6ba DY |
3990 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && |
3991 | !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) | |
ef060ea9 | 3992 | goto err; |
9f010507 | 3993 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3994 | } |
9f010507 | 3995 | |
c0e0d6ba DY |
3996 | /* |
3997 | * For DEFER_TASKRUN we require the completion task to be the same as the | |
3998 | * submission task. This implies that there is only one submitter, so enforce | |
3999 | * that. | |
4000 | */ | |
4001 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && | |
4002 | !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { | |
4003 | goto err; | |
4004 | } | |
4005 | ||
2aede0e4 JA |
4006 | /* |
4007 | * This is just grabbed for accounting purposes. When a process exits, | |
4008 | * the mm is exited and dropped before the files, hence we need to hang | |
4009 | * on to this mm purely for the purposes of being able to unaccount | |
4010 | * memory (locked/pinned vm). It's not used for anything else. | |
4011 | */ | |
6b7898eb | 4012 | mmgrab(current->mm); |
2aede0e4 | 4013 | ctx->mm_account = current->mm; |
6b7898eb | 4014 | |
2b188cc1 JA |
4015 | ret = io_allocate_scq_urings(ctx, p); |
4016 | if (ret) | |
4017 | goto err; | |
4018 | ||
7e84e1c7 | 4019 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
4020 | if (ret) |
4021 | goto err; | |
2933ae6e PB |
4022 | |
4023 | ret = io_rsrc_init(ctx); | |
47b228ce PB |
4024 | if (ret) |
4025 | goto err; | |
2b188cc1 | 4026 | |
75b28aff HV |
4027 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
4028 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
4029 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
4030 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
4031 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
4032 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
2af89abd PB |
4033 | if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) |
4034 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
9b1b58ca | 4035 | p->sq_off.resv1 = 0; |
03d89a2d JA |
4036 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
4037 | p->sq_off.user_addr = 0; | |
2b188cc1 | 4038 | |
75b28aff HV |
4039 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
4040 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
4041 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
4042 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
4043 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
4044 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 4045 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
9b1b58ca | 4046 | p->cq_off.resv1 = 0; |
03d89a2d JA |
4047 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
4048 | p->cq_off.user_addr = 0; | |
ac90f249 | 4049 | |
7f13657d XW |
4050 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
4051 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 4052 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 4053 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 4054 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e | 4055 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
7d3fd88d | 4056 | IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING; |
7f13657d XW |
4057 | |
4058 | if (copy_to_user(params, p, sizeof(*p))) { | |
4059 | ret = -EFAULT; | |
4060 | goto err; | |
4061 | } | |
d1719f70 | 4062 | |
7cae596b DY |
4063 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER |
4064 | && !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
8579538c | 4065 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
7cae596b | 4066 | |
9faadcc8 PB |
4067 | file = io_uring_get_file(ctx); |
4068 | if (IS_ERR(file)) { | |
4069 | ret = PTR_ERR(file); | |
4070 | goto err; | |
4071 | } | |
4072 | ||
6e76ac59 JT |
4073 | ret = __io_uring_add_tctx_node(ctx); |
4074 | if (ret) | |
4075 | goto err_fput; | |
4076 | tctx = current->io_uring; | |
4077 | ||
044c1ab3 JA |
4078 | /* |
4079 | * Install ring fd as the very last thing, so we don't risk someone | |
4080 | * having closed it before we finish setup | |
4081 | */ | |
6e76ac59 JT |
4082 | if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY) |
4083 | ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX); | |
4084 | else | |
4085 | ret = io_uring_install_fd(file); | |
4086 | if (ret < 0) | |
4087 | goto err_fput; | |
044c1ab3 | 4088 | |
c826bd7a | 4089 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
4090 | return ret; |
4091 | err: | |
4092 | io_ring_ctx_wait_and_kill(ctx); | |
4093 | return ret; | |
6e76ac59 JT |
4094 | err_fput: |
4095 | fput(file); | |
4096 | return ret; | |
2b188cc1 JA |
4097 | } |
4098 | ||
4099 | /* | |
4100 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
4101 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
4102 | * params structure passed in. | |
4103 | */ | |
4104 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
4105 | { | |
4106 | struct io_uring_params p; | |
2b188cc1 JA |
4107 | int i; |
4108 | ||
4109 | if (copy_from_user(&p, params, sizeof(p))) | |
4110 | return -EFAULT; | |
4111 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
4112 | if (p.resv[i]) | |
4113 | return -EINVAL; | |
4114 | } | |
4115 | ||
6c271ce2 | 4116 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 4117 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 4118 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 4119 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 4120 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 | 4121 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
03d89a2d | 4122 | IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN | |
2af89abd PB |
4123 | IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY | |
4124 | IORING_SETUP_NO_SQARRAY)) | |
2b188cc1 JA |
4125 | return -EINVAL; |
4126 | ||
ef060ea9 | 4127 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
4128 | } |
4129 | ||
76d3ccec MR |
4130 | static inline bool io_uring_allowed(void) |
4131 | { | |
4132 | int disabled = READ_ONCE(sysctl_io_uring_disabled); | |
4133 | kgid_t io_uring_group; | |
4134 | ||
4135 | if (disabled == 2) | |
4136 | return false; | |
4137 | ||
4138 | if (disabled == 0 || capable(CAP_SYS_ADMIN)) | |
4139 | return true; | |
4140 | ||
4141 | io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group); | |
4142 | if (!gid_valid(io_uring_group)) | |
4143 | return false; | |
4144 | ||
4145 | return in_group_p(io_uring_group); | |
4146 | } | |
4147 | ||
2b188cc1 JA |
4148 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, |
4149 | struct io_uring_params __user *, params) | |
4150 | { | |
76d3ccec MR |
4151 | if (!io_uring_allowed()) |
4152 | return -EPERM; | |
4153 | ||
2b188cc1 JA |
4154 | return io_uring_setup(entries, params); |
4155 | } | |
4156 | ||
c072481d PB |
4157 | static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, |
4158 | unsigned nr_args) | |
66f4af93 JA |
4159 | { |
4160 | struct io_uring_probe *p; | |
4161 | size_t size; | |
4162 | int i, ret; | |
4163 | ||
4164 | size = struct_size(p, ops, nr_args); | |
4165 | if (size == SIZE_MAX) | |
4166 | return -EOVERFLOW; | |
4167 | p = kzalloc(size, GFP_KERNEL); | |
4168 | if (!p) | |
4169 | return -ENOMEM; | |
4170 | ||
4171 | ret = -EFAULT; | |
4172 | if (copy_from_user(p, arg, size)) | |
4173 | goto out; | |
4174 | ret = -EINVAL; | |
4175 | if (memchr_inv(p, 0, size)) | |
4176 | goto out; | |
4177 | ||
4178 | p->last_op = IORING_OP_LAST - 1; | |
4179 | if (nr_args > IORING_OP_LAST) | |
4180 | nr_args = IORING_OP_LAST; | |
4181 | ||
4182 | for (i = 0; i < nr_args; i++) { | |
4183 | p->ops[i].op = i; | |
a7dd2782 | 4184 | if (!io_issue_defs[i].not_supported) |
66f4af93 JA |
4185 | p->ops[i].flags = IO_URING_OP_SUPPORTED; |
4186 | } | |
4187 | p->ops_len = i; | |
4188 | ||
4189 | ret = 0; | |
4190 | if (copy_to_user(arg, p, size)) | |
4191 | ret = -EFAULT; | |
4192 | out: | |
4193 | kfree(p); | |
4194 | return ret; | |
4195 | } | |
4196 | ||
071698e1 JA |
4197 | static int io_register_personality(struct io_ring_ctx *ctx) |
4198 | { | |
4379bf8b | 4199 | const struct cred *creds; |
61cf9370 | 4200 | u32 id; |
1e6fa521 | 4201 | int ret; |
071698e1 | 4202 | |
4379bf8b | 4203 | creds = get_current_cred(); |
1e6fa521 | 4204 | |
61cf9370 MWO |
4205 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
4206 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
4207 | if (ret < 0) { |
4208 | put_cred(creds); | |
4209 | return ret; | |
4210 | } | |
4211 | return id; | |
071698e1 JA |
4212 | } |
4213 | ||
c072481d PB |
4214 | static __cold int io_register_restrictions(struct io_ring_ctx *ctx, |
4215 | void __user *arg, unsigned int nr_args) | |
21b55dbc SG |
4216 | { |
4217 | struct io_uring_restriction *res; | |
4218 | size_t size; | |
4219 | int i, ret; | |
4220 | ||
7e84e1c7 SG |
4221 | /* Restrictions allowed only if rings started disabled */ |
4222 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
4223 | return -EBADFD; | |
4224 | ||
21b55dbc | 4225 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 4226 | if (ctx->restrictions.registered) |
21b55dbc SG |
4227 | return -EBUSY; |
4228 | ||
4229 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
4230 | return -EINVAL; | |
4231 | ||
4232 | size = array_size(nr_args, sizeof(*res)); | |
4233 | if (size == SIZE_MAX) | |
4234 | return -EOVERFLOW; | |
4235 | ||
4236 | res = memdup_user(arg, size); | |
4237 | if (IS_ERR(res)) | |
4238 | return PTR_ERR(res); | |
4239 | ||
4240 | ret = 0; | |
4241 | ||
4242 | for (i = 0; i < nr_args; i++) { | |
4243 | switch (res[i].opcode) { | |
4244 | case IORING_RESTRICTION_REGISTER_OP: | |
4245 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
4246 | ret = -EINVAL; | |
4247 | goto out; | |
4248 | } | |
4249 | ||
4250 | __set_bit(res[i].register_op, | |
4251 | ctx->restrictions.register_op); | |
4252 | break; | |
4253 | case IORING_RESTRICTION_SQE_OP: | |
4254 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
4255 | ret = -EINVAL; | |
4256 | goto out; | |
4257 | } | |
4258 | ||
4259 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
4260 | break; | |
4261 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
4262 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
4263 | break; | |
4264 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
4265 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
4266 | break; | |
4267 | default: | |
4268 | ret = -EINVAL; | |
4269 | goto out; | |
4270 | } | |
4271 | } | |
4272 | ||
4273 | out: | |
4274 | /* Reset all restrictions if an error happened */ | |
4275 | if (ret != 0) | |
4276 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
4277 | else | |
7e84e1c7 | 4278 | ctx->restrictions.registered = true; |
21b55dbc SG |
4279 | |
4280 | kfree(res); | |
4281 | return ret; | |
4282 | } | |
4283 | ||
7e84e1c7 SG |
4284 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
4285 | { | |
4286 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
4287 | return -EBADFD; | |
4288 | ||
bca39f39 | 4289 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) { |
8579538c | 4290 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
bca39f39 PB |
4291 | /* |
4292 | * Lazy activation attempts would fail if it was polled before | |
4293 | * submitter_task is set. | |
4294 | */ | |
4295 | if (wq_has_sleeper(&ctx->poll_wq)) | |
4296 | io_activate_pollwq(ctx); | |
4297 | } | |
7cae596b | 4298 | |
7e84e1c7 SG |
4299 | if (ctx->restrictions.registered) |
4300 | ctx->restricted = 1; | |
4301 | ||
0298ef96 PB |
4302 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
4303 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
4304 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
4305 | return 0; |
4306 | } | |
4307 | ||
ebdfefc0 JA |
4308 | static __cold int __io_register_iowq_aff(struct io_ring_ctx *ctx, |
4309 | cpumask_var_t new_mask) | |
4310 | { | |
4311 | int ret; | |
4312 | ||
4313 | if (!(ctx->flags & IORING_SETUP_SQPOLL)) { | |
4314 | ret = io_wq_cpu_affinity(current->io_uring, new_mask); | |
4315 | } else { | |
4316 | mutex_unlock(&ctx->uring_lock); | |
4317 | ret = io_sqpoll_wq_cpu_affinity(ctx, new_mask); | |
4318 | mutex_lock(&ctx->uring_lock); | |
4319 | } | |
4320 | ||
4321 | return ret; | |
4322 | } | |
4323 | ||
c072481d PB |
4324 | static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, |
4325 | void __user *arg, unsigned len) | |
fe76421d | 4326 | { |
fe76421d JA |
4327 | cpumask_var_t new_mask; |
4328 | int ret; | |
4329 | ||
fe76421d JA |
4330 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) |
4331 | return -ENOMEM; | |
4332 | ||
4333 | cpumask_clear(new_mask); | |
4334 | if (len > cpumask_size()) | |
4335 | len = cpumask_size(); | |
4336 | ||
0f5e4b83 ES |
4337 | if (in_compat_syscall()) { |
4338 | ret = compat_get_bitmap(cpumask_bits(new_mask), | |
4339 | (const compat_ulong_t __user *)arg, | |
4340 | len * 8 /* CHAR_BIT */); | |
4341 | } else { | |
4342 | ret = copy_from_user(new_mask, arg, len); | |
4343 | } | |
4344 | ||
4345 | if (ret) { | |
fe76421d JA |
4346 | free_cpumask_var(new_mask); |
4347 | return -EFAULT; | |
4348 | } | |
4349 | ||
ebdfefc0 | 4350 | ret = __io_register_iowq_aff(ctx, new_mask); |
fe76421d JA |
4351 | free_cpumask_var(new_mask); |
4352 | return ret; | |
4353 | } | |
4354 | ||
c072481d | 4355 | static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) |
fe76421d | 4356 | { |
ebdfefc0 | 4357 | return __io_register_iowq_aff(ctx, NULL); |
fe76421d JA |
4358 | } |
4359 | ||
c072481d PB |
4360 | static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, |
4361 | void __user *arg) | |
b22fa62a | 4362 | __must_hold(&ctx->uring_lock) |
2e480058 | 4363 | { |
b22fa62a | 4364 | struct io_tctx_node *node; |
fa84693b JA |
4365 | struct io_uring_task *tctx = NULL; |
4366 | struct io_sq_data *sqd = NULL; | |
2e480058 JA |
4367 | __u32 new_count[2]; |
4368 | int i, ret; | |
4369 | ||
2e480058 JA |
4370 | if (copy_from_user(new_count, arg, sizeof(new_count))) |
4371 | return -EFAULT; | |
4372 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
4373 | if (new_count[i] > INT_MAX) | |
4374 | return -EINVAL; | |
4375 | ||
fa84693b JA |
4376 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
4377 | sqd = ctx->sq_data; | |
4378 | if (sqd) { | |
009ad9f0 JA |
4379 | /* |
4380 | * Observe the correct sqd->lock -> ctx->uring_lock | |
4381 | * ordering. Fine to drop uring_lock here, we hold | |
4382 | * a ref to the ctx. | |
4383 | */ | |
41d3a6bd | 4384 | refcount_inc(&sqd->refs); |
009ad9f0 | 4385 | mutex_unlock(&ctx->uring_lock); |
fa84693b | 4386 | mutex_lock(&sqd->lock); |
009ad9f0 | 4387 | mutex_lock(&ctx->uring_lock); |
41d3a6bd JA |
4388 | if (sqd->thread) |
4389 | tctx = sqd->thread->io_uring; | |
fa84693b JA |
4390 | } |
4391 | } else { | |
4392 | tctx = current->io_uring; | |
4393 | } | |
4394 | ||
e139a1ec | 4395 | BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); |
fa84693b | 4396 | |
bad119b9 PB |
4397 | for (i = 0; i < ARRAY_SIZE(new_count); i++) |
4398 | if (new_count[i]) | |
4399 | ctx->iowq_limits[i] = new_count[i]; | |
e139a1ec PB |
4400 | ctx->iowq_limits_set = true; |
4401 | ||
e139a1ec PB |
4402 | if (tctx && tctx->io_wq) { |
4403 | ret = io_wq_max_workers(tctx->io_wq, new_count); | |
4404 | if (ret) | |
4405 | goto err; | |
4406 | } else { | |
4407 | memset(new_count, 0, sizeof(new_count)); | |
4408 | } | |
fa84693b | 4409 | |
41d3a6bd | 4410 | if (sqd) { |
fa84693b | 4411 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
4412 | io_put_sq_data(sqd); |
4413 | } | |
2e480058 JA |
4414 | |
4415 | if (copy_to_user(arg, new_count, sizeof(new_count))) | |
4416 | return -EFAULT; | |
4417 | ||
b22fa62a PB |
4418 | /* that's it for SQPOLL, only the SQPOLL task creates requests */ |
4419 | if (sqd) | |
4420 | return 0; | |
4421 | ||
4422 | /* now propagate the restriction to all registered users */ | |
4423 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
4424 | struct io_uring_task *tctx = node->task->io_uring; | |
4425 | ||
4426 | if (WARN_ON_ONCE(!tctx->io_wq)) | |
4427 | continue; | |
4428 | ||
4429 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
4430 | new_count[i] = ctx->iowq_limits[i]; | |
4431 | /* ignore errors, it always returns zero anyway */ | |
4432 | (void)io_wq_max_workers(tctx->io_wq, new_count); | |
4433 | } | |
2e480058 | 4434 | return 0; |
fa84693b | 4435 | err: |
41d3a6bd | 4436 | if (sqd) { |
fa84693b | 4437 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
4438 | io_put_sq_data(sqd); |
4439 | } | |
fa84693b | 4440 | return ret; |
2e480058 JA |
4441 | } |
4442 | ||
edafccee JA |
4443 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
4444 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
4445 | __releases(ctx->uring_lock) |
4446 | __acquires(ctx->uring_lock) | |
edafccee JA |
4447 | { |
4448 | int ret; | |
4449 | ||
35fa71a0 | 4450 | /* |
fbb8bb02 PB |
4451 | * We don't quiesce the refs for register anymore and so it can't be |
4452 | * dying as we're holding a file ref here. | |
35fa71a0 | 4453 | */ |
fbb8bb02 | 4454 | if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs))) |
35fa71a0 JA |
4455 | return -ENXIO; |
4456 | ||
d7cce96c PB |
4457 | if (ctx->submitter_task && ctx->submitter_task != current) |
4458 | return -EEXIST; | |
4459 | ||
75c4021a | 4460 | if (ctx->restricted) { |
75c4021a PB |
4461 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); |
4462 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
4463 | return -EACCES; | |
4464 | } | |
4465 | ||
edafccee JA |
4466 | switch (opcode) { |
4467 | case IORING_REGISTER_BUFFERS: | |
0184f08e PB |
4468 | ret = -EFAULT; |
4469 | if (!arg) | |
4470 | break; | |
634d00df | 4471 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
4472 | break; |
4473 | case IORING_UNREGISTER_BUFFERS: | |
4474 | ret = -EINVAL; | |
4475 | if (arg || nr_args) | |
4476 | break; | |
0a96bbe4 | 4477 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 4478 | break; |
6b06314c | 4479 | case IORING_REGISTER_FILES: |
a8da73a3 JA |
4480 | ret = -EFAULT; |
4481 | if (!arg) | |
4482 | break; | |
792e3582 | 4483 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
4484 | break; |
4485 | case IORING_UNREGISTER_FILES: | |
4486 | ret = -EINVAL; | |
4487 | if (arg || nr_args) | |
4488 | break; | |
4489 | ret = io_sqe_files_unregister(ctx); | |
4490 | break; | |
c3a31e60 | 4491 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 4492 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 4493 | break; |
9b402849 JA |
4494 | case IORING_REGISTER_EVENTFD: |
4495 | ret = -EINVAL; | |
4496 | if (nr_args != 1) | |
4497 | break; | |
c75312dd UA |
4498 | ret = io_eventfd_register(ctx, arg, 0); |
4499 | break; | |
4500 | case IORING_REGISTER_EVENTFD_ASYNC: | |
4501 | ret = -EINVAL; | |
4502 | if (nr_args != 1) | |
f2842ab5 | 4503 | break; |
c75312dd | 4504 | ret = io_eventfd_register(ctx, arg, 1); |
9b402849 JA |
4505 | break; |
4506 | case IORING_UNREGISTER_EVENTFD: | |
4507 | ret = -EINVAL; | |
4508 | if (arg || nr_args) | |
4509 | break; | |
4510 | ret = io_eventfd_unregister(ctx); | |
4511 | break; | |
66f4af93 JA |
4512 | case IORING_REGISTER_PROBE: |
4513 | ret = -EINVAL; | |
4514 | if (!arg || nr_args > 256) | |
4515 | break; | |
4516 | ret = io_probe(ctx, arg, nr_args); | |
4517 | break; | |
071698e1 JA |
4518 | case IORING_REGISTER_PERSONALITY: |
4519 | ret = -EINVAL; | |
4520 | if (arg || nr_args) | |
4521 | break; | |
4522 | ret = io_register_personality(ctx); | |
4523 | break; | |
4524 | case IORING_UNREGISTER_PERSONALITY: | |
4525 | ret = -EINVAL; | |
4526 | if (arg) | |
4527 | break; | |
4528 | ret = io_unregister_personality(ctx, nr_args); | |
4529 | break; | |
7e84e1c7 SG |
4530 | case IORING_REGISTER_ENABLE_RINGS: |
4531 | ret = -EINVAL; | |
4532 | if (arg || nr_args) | |
4533 | break; | |
4534 | ret = io_register_enable_rings(ctx); | |
4535 | break; | |
21b55dbc SG |
4536 | case IORING_REGISTER_RESTRICTIONS: |
4537 | ret = io_register_restrictions(ctx, arg, nr_args); | |
4538 | break; | |
992da01a PB |
4539 | case IORING_REGISTER_FILES2: |
4540 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
4541 | break; | |
4542 | case IORING_REGISTER_FILES_UPDATE2: | |
4543 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
4544 | IORING_RSRC_FILE); | |
4545 | break; | |
4546 | case IORING_REGISTER_BUFFERS2: | |
4547 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 4548 | break; |
992da01a PB |
4549 | case IORING_REGISTER_BUFFERS_UPDATE: |
4550 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
4551 | IORING_RSRC_BUFFER); | |
c3bdad02 | 4552 | break; |
fe76421d JA |
4553 | case IORING_REGISTER_IOWQ_AFF: |
4554 | ret = -EINVAL; | |
4555 | if (!arg || !nr_args) | |
4556 | break; | |
4557 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
4558 | break; | |
4559 | case IORING_UNREGISTER_IOWQ_AFF: | |
4560 | ret = -EINVAL; | |
4561 | if (arg || nr_args) | |
4562 | break; | |
4563 | ret = io_unregister_iowq_aff(ctx); | |
4564 | break; | |
2e480058 JA |
4565 | case IORING_REGISTER_IOWQ_MAX_WORKERS: |
4566 | ret = -EINVAL; | |
4567 | if (!arg || nr_args != 2) | |
4568 | break; | |
4569 | ret = io_register_iowq_max_workers(ctx, arg); | |
4570 | break; | |
e7a6c00d JA |
4571 | case IORING_REGISTER_RING_FDS: |
4572 | ret = io_ringfd_register(ctx, arg, nr_args); | |
4573 | break; | |
4574 | case IORING_UNREGISTER_RING_FDS: | |
4575 | ret = io_ringfd_unregister(ctx, arg, nr_args); | |
4576 | break; | |
c7fb1942 JA |
4577 | case IORING_REGISTER_PBUF_RING: |
4578 | ret = -EINVAL; | |
4579 | if (!arg || nr_args != 1) | |
4580 | break; | |
4581 | ret = io_register_pbuf_ring(ctx, arg); | |
4582 | break; | |
4583 | case IORING_UNREGISTER_PBUF_RING: | |
4584 | ret = -EINVAL; | |
4585 | if (!arg || nr_args != 1) | |
4586 | break; | |
4587 | ret = io_unregister_pbuf_ring(ctx, arg); | |
4588 | break; | |
78a861b9 JA |
4589 | case IORING_REGISTER_SYNC_CANCEL: |
4590 | ret = -EINVAL; | |
4591 | if (!arg || nr_args != 1) | |
4592 | break; | |
4593 | ret = io_sync_cancel(ctx, arg); | |
4594 | break; | |
6e73dffb PB |
4595 | case IORING_REGISTER_FILE_ALLOC_RANGE: |
4596 | ret = -EINVAL; | |
4597 | if (!arg || nr_args) | |
4598 | break; | |
4599 | ret = io_register_file_alloc_range(ctx, arg); | |
4600 | break; | |
edafccee JA |
4601 | default: |
4602 | ret = -EINVAL; | |
4603 | break; | |
4604 | } | |
4605 | ||
edafccee JA |
4606 | return ret; |
4607 | } | |
4608 | ||
4609 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
4610 | void __user *, arg, unsigned int, nr_args) | |
4611 | { | |
4612 | struct io_ring_ctx *ctx; | |
4613 | long ret = -EBADF; | |
4614 | struct fd f; | |
7d3fd88d JT |
4615 | bool use_registered_ring; |
4616 | ||
4617 | use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING); | |
4618 | opcode &= ~IORING_REGISTER_USE_REGISTERED_RING; | |
edafccee | 4619 | |
34319084 JA |
4620 | if (opcode >= IORING_REGISTER_LAST) |
4621 | return -EINVAL; | |
4622 | ||
7d3fd88d JT |
4623 | if (use_registered_ring) { |
4624 | /* | |
4625 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
4626 | * need only dereference our task private array to find it. | |
4627 | */ | |
4628 | struct io_uring_task *tctx = current->io_uring; | |
edafccee | 4629 | |
7d3fd88d JT |
4630 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
4631 | return -EINVAL; | |
4632 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
4633 | f.file = tctx->registered_rings[fd]; | |
4634 | f.flags = 0; | |
4635 | if (unlikely(!f.file)) | |
4636 | return -EBADF; | |
4637 | } else { | |
4638 | f = fdget(fd); | |
4639 | if (unlikely(!f.file)) | |
4640 | return -EBADF; | |
4641 | ret = -EOPNOTSUPP; | |
4642 | if (!io_is_uring_fops(f.file)) | |
4643 | goto out_fput; | |
4644 | } | |
edafccee JA |
4645 | |
4646 | ctx = f.file->private_data; | |
4647 | ||
4648 | mutex_lock(&ctx->uring_lock); | |
4649 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
4650 | mutex_unlock(&ctx->uring_lock); | |
2757be22 | 4651 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); |
edafccee JA |
4652 | out_fput: |
4653 | fdput(f); | |
4654 | return ret; | |
4655 | } | |
4656 | ||
2b188cc1 JA |
4657 | static int __init io_uring_init(void) |
4658 | { | |
9c71d39a | 4659 | #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ |
d7f62e82 | 4660 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ |
9c71d39a | 4661 | BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ |
d7f62e82 SM |
4662 | } while (0) |
4663 | ||
4664 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9c71d39a SM |
4665 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) |
4666 | #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ | |
4667 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) | |
d7f62e82 SM |
4668 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); |
4669 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
4670 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
4671 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
4672 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
4673 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
4674 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9c71d39a SM |
4675 | BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); |
4676 | BUILD_BUG_SQE_ELEM(12, __u32, __pad1); | |
d7f62e82 | 4677 | BUILD_BUG_SQE_ELEM(16, __u64, addr); |
7d67af2c | 4678 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
4679 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
4680 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
4681 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
4682 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
4683 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
4684 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
4685 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
4686 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
4687 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
4688 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
4689 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
4690 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
4691 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
4692 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
4693 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 4694 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
9c71d39a SM |
4695 | BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); |
4696 | BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); | |
4697 | BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); | |
4698 | BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); | |
4699 | BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); | |
d7f62e82 SM |
4700 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
4701 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 4702 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 4703 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 4704 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 4705 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
b48c312b PB |
4706 | BUILD_BUG_SQE_ELEM(44, __u16, addr_len); |
4707 | BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); | |
e9621e2b | 4708 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
9c71d39a SM |
4709 | BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); |
4710 | BUILD_BUG_SQE_ELEM(56, __u64, __pad2); | |
d7f62e82 | 4711 | |
b0d658ec PB |
4712 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
4713 | sizeof(struct io_uring_rsrc_update)); | |
4714 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
4715 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
4716 | |
4717 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
4718 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
4719 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
4720 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 4721 | |
b0d658ec PB |
4722 | /* should fit into one byte */ |
4723 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
4724 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
4725 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 4726 | |
32c2d33e | 4727 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); |
16340eab | 4728 | |
3a4b89a2 JA |
4729 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
4730 | ||
528ce678 ML |
4731 | /* top 8bits are for internal use */ |
4732 | BUILD_BUG_ON((IORING_URING_CMD_MASK & 0xff000000) != 0); | |
4733 | ||
d9b57aa3 | 4734 | io_uring_optable_init(); |
0702e536 | 4735 | |
b97f96e2 JA |
4736 | /* |
4737 | * Allow user copy in the per-command field, which starts after the | |
4738 | * file in io_kiocb and until the opcode field. The openat2 handling | |
4739 | * requires copying in user memory into the io_kiocb object in that | |
4740 | * range, and HARDENED_USERCOPY will complain if we haven't | |
4741 | * correctly annotated this range. | |
4742 | */ | |
4743 | req_cachep = kmem_cache_create_usercopy("io_kiocb", | |
4744 | sizeof(struct io_kiocb), 0, | |
4745 | SLAB_HWCACHE_ALIGN | SLAB_PANIC | | |
4746 | SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU, | |
4747 | offsetof(struct io_kiocb, cmd.data), | |
4748 | sizeof_field(struct io_kiocb, cmd.data), NULL); | |
b3a4dbc8 GKB |
4749 | io_buf_cachep = kmem_cache_create("io_buffer", sizeof(struct io_buffer), 0, |
4750 | SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT, | |
4751 | NULL); | |
b97f96e2 | 4752 | |
76d3ccec MR |
4753 | #ifdef CONFIG_SYSCTL |
4754 | register_sysctl_init("kernel", kernel_io_uring_disabled_table); | |
4755 | #endif | |
4756 | ||
2b188cc1 JA |
4757 | return 0; |
4758 | }; | |
4759 | __initcall(io_uring_init); |