Commit | Line | Data |
---|---|---|
2b188cc1 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Shared application/kernel submission and completion ring pairs, for | |
4 | * supporting fast/efficient IO. | |
5 | * | |
6 | * A note on the read/write ordering memory barriers that are matched between | |
1e84b97b SB |
7 | * the application and kernel side. |
8 | * | |
9 | * After the application reads the CQ ring tail, it must use an | |
10 | * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses | |
11 | * before writing the tail (using smp_load_acquire to read the tail will | |
12 | * do). It also needs a smp_mb() before updating CQ head (ordering the | |
13 | * entry load(s) with the head store), pairing with an implicit barrier | |
d068b506 | 14 | * through a control-dependency in io_get_cqe (smp_store_release to |
1e84b97b SB |
15 | * store head will do). Failure to do so could lead to reading invalid |
16 | * CQ entries. | |
17 | * | |
18 | * Likewise, the application must use an appropriate smp_wmb() before | |
19 | * writing the SQ tail (ordering SQ entry stores with the tail store), | |
20 | * which pairs with smp_load_acquire in io_get_sqring (smp_store_release | |
21 | * to store the tail will do). And it needs a barrier ordering the SQ | |
22 | * head load before writing new SQ entries (smp_load_acquire to read | |
23 | * head will do). | |
24 | * | |
25 | * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application | |
26 | * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after* | |
27 | * updating the SQ tail; a full memory barrier smp_mb() is needed | |
28 | * between. | |
2b188cc1 JA |
29 | * |
30 | * Also see the examples in the liburing library: | |
31 | * | |
32 | * git://git.kernel.dk/liburing | |
33 | * | |
34 | * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens | |
35 | * from data shared between the kernel and application. This is done both | |
36 | * for ordering purposes, but also to ensure that once a value is loaded from | |
37 | * data that the application could potentially modify, it remains stable. | |
38 | * | |
39 | * Copyright (C) 2018-2019 Jens Axboe | |
c992fe29 | 40 | * Copyright (c) 2018-2019 Christoph Hellwig |
2b188cc1 JA |
41 | */ |
42 | #include <linux/kernel.h> | |
43 | #include <linux/init.h> | |
44 | #include <linux/errno.h> | |
45 | #include <linux/syscalls.h> | |
52de1fe1 | 46 | #include <net/compat.h> |
2b188cc1 JA |
47 | #include <linux/refcount.h> |
48 | #include <linux/uio.h> | |
6b47ee6e | 49 | #include <linux/bits.h> |
2b188cc1 JA |
50 | |
51 | #include <linux/sched/signal.h> | |
52 | #include <linux/fs.h> | |
53 | #include <linux/file.h> | |
54 | #include <linux/fdtable.h> | |
55 | #include <linux/mm.h> | |
56 | #include <linux/mman.h> | |
2b188cc1 JA |
57 | #include <linux/percpu.h> |
58 | #include <linux/slab.h> | |
edafccee | 59 | #include <linux/bvec.h> |
2b188cc1 JA |
60 | #include <linux/net.h> |
61 | #include <net/sock.h> | |
2b188cc1 JA |
62 | #include <linux/anon_inodes.h> |
63 | #include <linux/sched/mm.h> | |
64 | #include <linux/uaccess.h> | |
65 | #include <linux/nospec.h> | |
15b71abe | 66 | #include <linux/fsnotify.h> |
4840e418 | 67 | #include <linux/fadvise.h> |
b41e9852 | 68 | #include <linux/task_work.h> |
0f212204 | 69 | #include <linux/io_uring.h> |
b66509b8 | 70 | #include <linux/io_uring/cmd.h> |
5bd2182d | 71 | #include <linux/audit.h> |
cdc1404a | 72 | #include <linux/security.h> |
d808459b | 73 | #include <asm/shmparam.h> |
2b188cc1 | 74 | |
c826bd7a DD |
75 | #define CREATE_TRACE_POINTS |
76 | #include <trace/events/io_uring.h> | |
77 | ||
2b188cc1 JA |
78 | #include <uapi/linux/io_uring.h> |
79 | ||
561fb04a | 80 | #include "io-wq.h" |
2b188cc1 | 81 | |
de23077e | 82 | #include "io_uring.h" |
329061d3 | 83 | #include "opdef.h" |
e418bbc9 | 84 | #include "refs.h" |
c9f06aa7 | 85 | #include "tctx.h" |
c4320315 | 86 | #include "register.h" |
17437f31 | 87 | #include "sqpoll.h" |
a4ad4f74 | 88 | #include "fdinfo.h" |
3b77495a | 89 | #include "kbuf.h" |
73572984 | 90 | #include "rsrc.h" |
38513c46 | 91 | #include "cancel.h" |
43e0bbbd | 92 | #include "net.h" |
eb42cebb | 93 | #include "notif.h" |
f31ecf67 | 94 | #include "waitid.h" |
194bb58c | 95 | #include "futex.h" |
8d0c12a8 | 96 | #include "napi.h" |
da12d9ab | 97 | #include "uring_cmd.h" |
e27f928e | 98 | |
59915143 | 99 | #include "timeout.h" |
329061d3 | 100 | #include "poll.h" |
c92fcfc2 | 101 | #include "rw.h" |
9b797a37 | 102 | #include "alloc_cache.h" |
5e2a18d9 | 103 | |
5277deaa | 104 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 105 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 | 106 | |
68fe256a PB |
107 | #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \ |
108 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) | |
109 | ||
5562a8d7 PB |
110 | #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \ |
111 | IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS) | |
68fe256a | 112 | |
c854357b | 113 | #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ |
9cae36a0 JA |
114 | REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \ |
115 | REQ_F_ASYNC_DATA) | |
b16fed66 | 116 | |
a538be5b PB |
117 | #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\ |
118 | IO_REQ_CLEAN_FLAGS) | |
119 | ||
09899b19 PB |
120 | #define IO_TCTX_REFS_CACHE_NR (1U << 10) |
121 | ||
6dd0be1e | 122 | #define IO_COMPL_BATCH 32 |
bf019da7 | 123 | #define IO_REQ_ALLOC_BATCH 8 |
258b29a9 | 124 | |
27dc8338 PB |
125 | struct io_defer_entry { |
126 | struct list_head list; | |
127 | struct io_kiocb *req; | |
9cf7c104 | 128 | u32 seq; |
2b188cc1 JA |
129 | }; |
130 | ||
0756a869 PB |
131 | /* requests with any of those set should undergo io_disarm_next() */ |
132 | #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) | |
da1a08c5 | 133 | #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK) |
0756a869 | 134 | |
b4bc35cf PB |
135 | /* |
136 | * No waiters. It's larger than any valid value of the tw counter | |
137 | * so that tests against ->cq_wait_nr would fail and skip wake_up(). | |
138 | */ | |
139 | #define IO_CQ_WAKE_INIT (-1U) | |
140 | /* Forced wake up if there is a waiter regardless of ->cq_wait_nr */ | |
141 | #define IO_CQ_WAKE_FORCE (IO_CQ_WAKE_INIT >> 1) | |
142 | ||
affa87db | 143 | static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
9936c7c2 | 144 | struct task_struct *task, |
3dd0c97a | 145 | bool cancel_all); |
1ffc5422 | 146 | |
cbc2e203 | 147 | static void io_queue_sqe(struct io_kiocb *req); |
de0617e4 | 148 | |
c1755c25 | 149 | struct kmem_cache *req_cachep; |
73eaa2b5 | 150 | static struct workqueue_struct *iou_wq __ro_after_init; |
2b188cc1 | 151 | |
76d3ccec MR |
152 | static int __read_mostly sysctl_io_uring_disabled; |
153 | static int __read_mostly sysctl_io_uring_group = -1; | |
154 | ||
155 | #ifdef CONFIG_SYSCTL | |
156 | static struct ctl_table kernel_io_uring_disabled_table[] = { | |
157 | { | |
158 | .procname = "io_uring_disabled", | |
159 | .data = &sysctl_io_uring_disabled, | |
160 | .maxlen = sizeof(sysctl_io_uring_disabled), | |
161 | .mode = 0644, | |
162 | .proc_handler = proc_dointvec_minmax, | |
163 | .extra1 = SYSCTL_ZERO, | |
164 | .extra2 = SYSCTL_TWO, | |
165 | }, | |
166 | { | |
167 | .procname = "io_uring_group", | |
168 | .data = &sysctl_io_uring_group, | |
169 | .maxlen = sizeof(gid_t), | |
170 | .mode = 0644, | |
171 | .proc_handler = proc_dointvec, | |
172 | }, | |
76d3ccec MR |
173 | }; |
174 | #endif | |
175 | ||
faf88dde PB |
176 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
177 | { | |
178 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
179 | } | |
180 | ||
0fc8c2ac DY |
181 | static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx) |
182 | { | |
183 | return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head); | |
184 | } | |
185 | ||
9cae36a0 JA |
186 | static bool io_match_linked(struct io_kiocb *head) |
187 | { | |
188 | struct io_kiocb *req; | |
189 | ||
190 | io_for_each_link(req, head) { | |
191 | if (req->flags & REQ_F_INFLIGHT) | |
192 | return true; | |
193 | } | |
194 | return false; | |
6af3f48b PB |
195 | } |
196 | ||
197 | /* | |
198 | * As io_match_task() but protected against racing with linked timeouts. | |
199 | * User must not hold timeout_lock. | |
200 | */ | |
329061d3 JA |
201 | bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, |
202 | bool cancel_all) | |
6af3f48b | 203 | { |
9cae36a0 JA |
204 | bool matched; |
205 | ||
6af3f48b PB |
206 | if (task && head->task != task) |
207 | return false; | |
9cae36a0 JA |
208 | if (cancel_all) |
209 | return true; | |
210 | ||
211 | if (head->flags & REQ_F_LINK_TIMEOUT) { | |
212 | struct io_ring_ctx *ctx = head->ctx; | |
213 | ||
214 | /* protect against races with linked timeouts */ | |
215 | spin_lock_irq(&ctx->timeout_lock); | |
216 | matched = io_match_linked(head); | |
217 | spin_unlock_irq(&ctx->timeout_lock); | |
218 | } else { | |
219 | matched = io_match_linked(head); | |
220 | } | |
221 | return matched; | |
6af3f48b PB |
222 | } |
223 | ||
a8295b98 HX |
224 | static inline void req_fail_link_node(struct io_kiocb *req, int res) |
225 | { | |
226 | req_set_fail(req); | |
97b388d7 | 227 | io_req_set_res(req, res, 0); |
a8295b98 HX |
228 | } |
229 | ||
fa05457a PB |
230 | static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) |
231 | { | |
232 | wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); | |
a8295b98 HX |
233 | } |
234 | ||
c072481d | 235 | static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) |
2b188cc1 JA |
236 | { |
237 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
238 | ||
0f158b4c | 239 | complete(&ctx->ref_comp); |
2b188cc1 JA |
240 | } |
241 | ||
c072481d | 242 | static __cold void io_fallback_req_func(struct work_struct *work) |
f56165e6 PB |
243 | { |
244 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
245 | fallback_work.work); | |
246 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
247 | struct io_kiocb *req, *tmp; | |
8e5b3b89 | 248 | struct io_tw_state ts = {}; |
f56165e6 | 249 | |
f7b32e78 | 250 | percpu_ref_get(&ctx->refs); |
31f084b7 | 251 | mutex_lock(&ctx->uring_lock); |
3218e5d3 | 252 | llist_for_each_entry_safe(req, tmp, node, io_task_work.node) |
a282967c | 253 | req->io_task_work.func(req, &ts); |
31f084b7 PB |
254 | io_submit_flush_completions(ctx); |
255 | mutex_unlock(&ctx->uring_lock); | |
f7b32e78 | 256 | percpu_ref_put(&ctx->refs); |
f56165e6 PB |
257 | } |
258 | ||
e6f89be6 PB |
259 | static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) |
260 | { | |
261 | unsigned hash_buckets = 1U << bits; | |
262 | size_t hash_size = hash_buckets * sizeof(table->hbs[0]); | |
263 | ||
264 | table->hbs = kmalloc(hash_size, GFP_KERNEL); | |
265 | if (!table->hbs) | |
266 | return -ENOMEM; | |
267 | ||
268 | table->hash_bits = bits; | |
269 | init_hash_table(table, hash_buckets); | |
270 | return 0; | |
271 | } | |
272 | ||
c072481d | 273 | static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
2b188cc1 JA |
274 | { |
275 | struct io_ring_ctx *ctx; | |
9cfc7e94 | 276 | int hash_bits; |
414d0f45 | 277 | bool ret; |
2b188cc1 JA |
278 | |
279 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
280 | if (!ctx) | |
281 | return NULL; | |
282 | ||
9cfc7e94 JA |
283 | xa_init(&ctx->io_bl_xa); |
284 | ||
78076bb6 JA |
285 | /* |
286 | * Use 5 bits less than the max cq entries, that should give us around | |
4a07723f PB |
287 | * 32 entries per hash list if totally full and uniformly spread, but |
288 | * don't keep too many buckets to not overconsume memory. | |
78076bb6 | 289 | */ |
4a07723f PB |
290 | hash_bits = ilog2(p->cq_entries) - 5; |
291 | hash_bits = clamp(hash_bits, 1, 8); | |
e6f89be6 | 292 | if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) |
78076bb6 | 293 | goto err; |
9ca9fb24 PB |
294 | if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) |
295 | goto err; | |
21482896 | 296 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
48904229 | 297 | 0, GFP_KERNEL)) |
206aefde | 298 | goto err; |
2b188cc1 JA |
299 | |
300 | ctx->flags = p->flags; | |
b4bc35cf | 301 | atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT); |
90554200 | 302 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 303 | INIT_LIST_HEAD(&ctx->sqd_list); |
1d7bb1d5 | 304 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
cc3cec83 | 305 | INIT_LIST_HEAD(&ctx->io_buffers_cache); |
414d0f45 | 306 | ret = io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX, |
69bbc6ad | 307 | sizeof(struct io_rsrc_node)); |
da22bdf3 | 308 | ret |= io_alloc_cache_init(&ctx->apoll_cache, IO_POLL_ALLOC_CACHE_MAX, |
69bbc6ad | 309 | sizeof(struct async_poll)); |
414d0f45 | 310 | ret |= io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX, |
69bbc6ad | 311 | sizeof(struct io_async_msghdr)); |
414d0f45 | 312 | ret |= io_alloc_cache_init(&ctx->rw_cache, IO_ALLOC_CACHE_MAX, |
a9165b83 | 313 | sizeof(struct io_async_rw)); |
414d0f45 | 314 | ret |= io_alloc_cache_init(&ctx->uring_cache, IO_ALLOC_CACHE_MAX, |
d10f19df | 315 | sizeof(struct uring_cache)); |
414d0f45 JA |
316 | ret |= io_futex_cache_init(ctx); |
317 | if (ret) | |
318 | goto err; | |
0f158b4c | 319 | init_completion(&ctx->ref_comp); |
61cf9370 | 320 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 321 | mutex_init(&ctx->uring_lock); |
311997b3 | 322 | init_waitqueue_head(&ctx->cq_wait); |
7b235dd8 | 323 | init_waitqueue_head(&ctx->poll_wq); |
4ea15b56 | 324 | init_waitqueue_head(&ctx->rsrc_quiesce_wq); |
2b188cc1 | 325 | spin_lock_init(&ctx->completion_lock); |
89850fce | 326 | spin_lock_init(&ctx->timeout_lock); |
5eef4e87 | 327 | INIT_WQ_LIST(&ctx->iopoll_list); |
cc3cec83 | 328 | INIT_LIST_HEAD(&ctx->io_buffers_comp); |
de0617e4 | 329 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 330 | INIT_LIST_HEAD(&ctx->timeout_list); |
ef9dd637 | 331 | INIT_LIST_HEAD(&ctx->ltimeout_list); |
d67d2263 | 332 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); |
c0e0d6ba | 333 | init_llist_head(&ctx->work_llist); |
13bf43f5 | 334 | INIT_LIST_HEAD(&ctx->tctx_list); |
c2b6c6bc PB |
335 | ctx->submit_state.free_list.next = NULL; |
336 | INIT_WQ_LIST(&ctx->locked_free_list); | |
f31ecf67 | 337 | INIT_HLIST_HEAD(&ctx->waitid_list); |
194bb58c JA |
338 | #ifdef CONFIG_FUTEX |
339 | INIT_HLIST_HEAD(&ctx->futex_list); | |
340 | #endif | |
9011bf9a | 341 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
6f33b0bc | 342 | INIT_WQ_LIST(&ctx->submit_state.compl_reqs); |
93b8cc60 | 343 | INIT_HLIST_HEAD(&ctx->cancelable_uring_cmd); |
8d0c12a8 SR |
344 | io_napi_init(ctx); |
345 | ||
2b188cc1 | 346 | return ctx; |
206aefde | 347 | err: |
414d0f45 JA |
348 | io_alloc_cache_free(&ctx->rsrc_node_cache, kfree); |
349 | io_alloc_cache_free(&ctx->apoll_cache, kfree); | |
350 | io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); | |
351 | io_alloc_cache_free(&ctx->rw_cache, io_rw_cache_free); | |
352 | io_alloc_cache_free(&ctx->uring_cache, kfree); | |
353 | io_futex_cache_free(ctx); | |
e6f89be6 | 354 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 355 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 | 356 | xa_destroy(&ctx->io_bl_xa); |
206aefde JA |
357 | kfree(ctx); |
358 | return NULL; | |
2b188cc1 JA |
359 | } |
360 | ||
8f6ed49a PB |
361 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
362 | { | |
363 | struct io_rings *r = ctx->rings; | |
364 | ||
365 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
366 | ctx->cq_extra--; | |
367 | } | |
368 | ||
9cf7c104 | 369 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 370 | { |
2bc9930e JA |
371 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
372 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 373 | |
8f6ed49a | 374 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 375 | } |
de0617e4 | 376 | |
9d858b21 | 377 | return false; |
de0617e4 JA |
378 | } |
379 | ||
5a754dea PB |
380 | static void io_clean_op(struct io_kiocb *req) |
381 | { | |
382 | if (req->flags & REQ_F_BUFFER_SELECTED) { | |
383 | spin_lock(&req->ctx->completion_lock); | |
384 | io_put_kbuf_comp(req); | |
385 | spin_unlock(&req->ctx->completion_lock); | |
386 | } | |
387 | ||
388 | if (req->flags & REQ_F_NEED_CLEANUP) { | |
389 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; | |
390 | ||
391 | if (def->cleanup) | |
392 | def->cleanup(req); | |
393 | } | |
394 | if ((req->flags & REQ_F_POLLED) && req->apoll) { | |
395 | kfree(req->apoll->double_poll); | |
396 | kfree(req->apoll); | |
397 | req->apoll = NULL; | |
398 | } | |
399 | if (req->flags & REQ_F_INFLIGHT) { | |
400 | struct io_uring_task *tctx = req->task->io_uring; | |
401 | ||
402 | atomic_dec(&tctx->inflight_tracked); | |
403 | } | |
404 | if (req->flags & REQ_F_CREDS) | |
405 | put_cred(req->creds); | |
406 | if (req->flags & REQ_F_ASYNC_DATA) { | |
407 | kfree(req->async_data); | |
408 | req->async_data = NULL; | |
409 | } | |
410 | req->flags &= ~IO_REQ_CLEAN_FLAGS; | |
411 | } | |
412 | ||
9cae36a0 JA |
413 | static inline void io_req_track_inflight(struct io_kiocb *req) |
414 | { | |
415 | if (!(req->flags & REQ_F_INFLIGHT)) { | |
416 | req->flags |= REQ_F_INFLIGHT; | |
386e4fb6 | 417 | atomic_inc(&req->task->io_uring->inflight_tracked); |
9cae36a0 JA |
418 | } |
419 | } | |
420 | ||
fd08e530 PB |
421 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
422 | { | |
906c6caa PB |
423 | if (WARN_ON_ONCE(!req->link)) |
424 | return NULL; | |
425 | ||
4d13d1a4 PB |
426 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
427 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
428 | |
429 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 430 | io_req_set_refcount(req); |
4d13d1a4 PB |
431 | __io_req_set_refcount(req->link, 2); |
432 | return req->link; | |
fd08e530 PB |
433 | } |
434 | ||
435 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
436 | { | |
4d13d1a4 | 437 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
438 | return NULL; |
439 | return __io_prep_linked_timeout(req); | |
440 | } | |
441 | ||
cb2d344c PB |
442 | static noinline void __io_arm_ltimeout(struct io_kiocb *req) |
443 | { | |
444 | io_queue_linked_timeout(__io_prep_linked_timeout(req)); | |
445 | } | |
446 | ||
447 | static inline void io_arm_ltimeout(struct io_kiocb *req) | |
448 | { | |
449 | if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) | |
450 | __io_arm_ltimeout(req); | |
451 | } | |
452 | ||
1e6fa521 JA |
453 | static void io_prep_async_work(struct io_kiocb *req) |
454 | { | |
a7dd2782 | 455 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
1e6fa521 JA |
456 | struct io_ring_ctx *ctx = req->ctx; |
457 | ||
b8e64b53 PB |
458 | if (!(req->flags & REQ_F_CREDS)) { |
459 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 460 | req->creds = get_current_cred(); |
b8e64b53 | 461 | } |
003e8dcc | 462 | |
e1d675df PB |
463 | req->work.list.next = NULL; |
464 | req->work.flags = 0; | |
feaadc4f PB |
465 | if (req->flags & REQ_F_FORCE_ASYNC) |
466 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
467 | ||
3beed235 | 468 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
8487f083 | 469 | req->flags |= io_file_get_flags(req->file); |
f6b543fd | 470 | |
8b1df11f | 471 | if (req->file && (req->flags & REQ_F_ISREG)) { |
d4755e15 JA |
472 | bool should_hash = def->hash_reg_file; |
473 | ||
474 | /* don't serialize this request if the fs doesn't need it */ | |
475 | if (should_hash && (req->file->f_flags & O_DIRECT) && | |
476 | (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE)) | |
477 | should_hash = false; | |
478 | if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL)) | |
1e6fa521 | 479 | io_wq_hash_work(&req->work, file_inode(req->file)); |
4b982bd0 | 480 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
481 | if (def->unbound_nonreg_file) |
482 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
483 | } | |
561fb04a | 484 | } |
cccf0ee8 | 485 | |
cbdcb435 | 486 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 487 | { |
cbdcb435 | 488 | struct io_kiocb *cur; |
54a91f3b | 489 | |
44eff40a PB |
490 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
491 | struct io_ring_ctx *ctx = req->ctx; | |
492 | ||
674ee8e1 | 493 | spin_lock_irq(&ctx->timeout_lock); |
44eff40a PB |
494 | io_for_each_link(cur, req) |
495 | io_prep_async_work(cur); | |
674ee8e1 | 496 | spin_unlock_irq(&ctx->timeout_lock); |
44eff40a PB |
497 | } else { |
498 | io_for_each_link(cur, req) | |
499 | io_prep_async_work(cur); | |
500 | } | |
561fb04a JA |
501 | } |
502 | ||
6e6b8c62 | 503 | static void io_queue_iowq(struct io_kiocb *req) |
561fb04a | 504 | { |
cbdcb435 | 505 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 506 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 507 | |
3bfe6106 JA |
508 | BUG_ON(!tctx); |
509 | BUG_ON(!tctx->io_wq); | |
561fb04a | 510 | |
cbdcb435 PB |
511 | /* init ->work of the whole link before punting */ |
512 | io_prep_async_link(req); | |
991468dc JA |
513 | |
514 | /* | |
515 | * Not expected to happen, but if we do have a bug where this _can_ | |
516 | * happen, catch it here and ensure the request is marked as | |
517 | * canceled. That will make io-wq go through the usual work cancel | |
518 | * procedure rather than attempt to run this request (or create a new | |
519 | * worker for it). | |
520 | */ | |
521 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
522 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
523 | ||
48863ffd | 524 | trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); |
ebf93667 | 525 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
526 | if (link) |
527 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
528 | } |
529 | ||
c072481d | 530 | static __cold void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 531 | { |
441b8a78 | 532 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
533 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
534 | struct io_defer_entry, list); | |
de0617e4 | 535 | |
9cf7c104 | 536 | if (req_need_defer(de->req, de->seq)) |
04518945 | 537 | break; |
27dc8338 | 538 | list_del_init(&de->list); |
907d1df3 | 539 | io_req_task_queue(de->req); |
27dc8338 | 540 | kfree(de); |
441b8a78 | 541 | } |
04518945 PB |
542 | } |
543 | ||
c4320315 | 544 | void io_eventfd_ops(struct rcu_head *rcu) |
f2842ab5 | 545 | { |
d8e9214f | 546 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
21a091b9 | 547 | int ops = atomic_xchg(&ev_fd->ops, 0); |
305bef98 | 548 | |
21a091b9 | 549 | if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) |
120ae585 | 550 | eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE); |
d8e9214f | 551 | |
21a091b9 DY |
552 | /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback |
553 | * ordering in a race but if references are 0 we know we have to free | |
554 | * it regardless. | |
305bef98 | 555 | */ |
21a091b9 DY |
556 | if (atomic_dec_and_test(&ev_fd->refs)) { |
557 | eventfd_ctx_put(ev_fd->cq_ev_fd); | |
558 | kfree(ev_fd); | |
559 | } | |
d8e9214f DY |
560 | } |
561 | ||
77bc59b4 | 562 | static void io_eventfd_signal(struct io_ring_ctx *ctx) |
f2842ab5 | 563 | { |
21a091b9 | 564 | struct io_ev_fd *ev_fd = NULL; |
77bc59b4 | 565 | |
77bc59b4 UA |
566 | rcu_read_lock(); |
567 | /* | |
568 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
569 | * and eventfd_signal | |
570 | */ | |
571 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
572 | ||
573 | /* | |
574 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
575 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
576 | * the function and rcu_read_lock. | |
577 | */ | |
578 | if (unlikely(!ev_fd)) | |
579 | goto out; | |
7e55a19c | 580 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 | 581 | goto out; |
21a091b9 DY |
582 | if (ev_fd->eventfd_async && !io_wq_current_is_worker()) |
583 | goto out; | |
77bc59b4 | 584 | |
21a091b9 | 585 | if (likely(eventfd_signal_allowed())) { |
120ae585 | 586 | eventfd_signal_mask(ev_fd->cq_ev_fd, EPOLL_URING_WAKE); |
21a091b9 DY |
587 | } else { |
588 | atomic_inc(&ev_fd->refs); | |
589 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) | |
44a84da4 | 590 | call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops); |
21a091b9 DY |
591 | else |
592 | atomic_dec(&ev_fd->refs); | |
593 | } | |
594 | ||
77bc59b4 UA |
595 | out: |
596 | rcu_read_unlock(); | |
f2842ab5 JA |
597 | } |
598 | ||
21a091b9 DY |
599 | static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) |
600 | { | |
601 | bool skip; | |
602 | ||
603 | spin_lock(&ctx->completion_lock); | |
604 | ||
605 | /* | |
606 | * Eventfd should only get triggered when at least one event has been | |
607 | * posted. Some applications rely on the eventfd notification count | |
608 | * only changing IFF a new CQE has been added to the CQ ring. There's | |
609 | * no depedency on 1:1 relationship between how many times this | |
610 | * function is called (and hence the eventfd count) and number of CQEs | |
611 | * posted to the CQ ring. | |
612 | */ | |
613 | skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
614 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
615 | spin_unlock(&ctx->completion_lock); | |
616 | if (skip) | |
617 | return; | |
618 | ||
619 | io_eventfd_signal(ctx); | |
620 | } | |
621 | ||
a830ffd2 PB |
622 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
623 | { | |
bca39f39 PB |
624 | if (ctx->poll_activated) |
625 | io_poll_wq_wake(ctx); | |
e5f30f6f PB |
626 | if (ctx->off_timeout_used) |
627 | io_flush_timeouts(ctx); | |
628 | if (ctx->drain_active) { | |
a830ffd2 | 629 | spin_lock(&ctx->completion_lock); |
e5f30f6f | 630 | io_queue_deferred(ctx); |
a830ffd2 PB |
631 | spin_unlock(&ctx->completion_lock); |
632 | } | |
633 | if (ctx->has_evfd) | |
21a091b9 | 634 | io_eventfd_flush_signal(ctx); |
a830ffd2 PB |
635 | } |
636 | ||
f66f7342 | 637 | static inline void __io_cq_lock(struct io_ring_ctx *ctx) |
f66f7342 | 638 | { |
ec26c225 | 639 | if (!ctx->lockless_cq) |
f66f7342 PB |
640 | spin_lock(&ctx->completion_lock); |
641 | } | |
642 | ||
6971253f PB |
643 | static inline void io_cq_lock(struct io_ring_ctx *ctx) |
644 | __acquires(ctx->completion_lock) | |
645 | { | |
646 | spin_lock(&ctx->completion_lock); | |
647 | } | |
648 | ||
f66f7342 | 649 | static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) |
3181e22f PB |
650 | { |
651 | io_commit_cqring(ctx); | |
54927baf | 652 | if (!ctx->task_complete) { |
ec26c225 PB |
653 | if (!ctx->lockless_cq) |
654 | spin_unlock(&ctx->completion_lock); | |
655 | /* IOPOLL rings only need to wake up if it's also SQPOLL */ | |
656 | if (!ctx->syscall_iopoll) | |
657 | io_cqring_wake(ctx); | |
3181e22f | 658 | } |
54927baf | 659 | io_commit_cqring_flush(ctx); |
3181e22f PB |
660 | } |
661 | ||
0fdb9a19 | 662 | static void io_cq_unlock_post(struct io_ring_ctx *ctx) |
5d772916 | 663 | __releases(ctx->completion_lock) |
25399321 | 664 | { |
f66f7342 PB |
665 | io_commit_cqring(ctx); |
666 | spin_unlock(&ctx->completion_lock); | |
f66f7342 | 667 | io_cqring_wake(ctx); |
54927baf | 668 | io_commit_cqring_flush(ctx); |
25399321 PB |
669 | } |
670 | ||
a85381d8 PB |
671 | static void io_cqring_overflow_kill(struct io_ring_ctx *ctx) |
672 | { | |
673 | struct io_overflow_cqe *ocqe; | |
674 | LIST_HEAD(list); | |
675 | ||
f432b76b | 676 | spin_lock(&ctx->completion_lock); |
a85381d8 PB |
677 | list_splice_init(&ctx->cq_overflow_list, &list); |
678 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); | |
f432b76b | 679 | spin_unlock(&ctx->completion_lock); |
a85381d8 PB |
680 | |
681 | while (!list_empty(&list)) { | |
682 | ocqe = list_first_entry(&list, struct io_overflow_cqe, list); | |
683 | list_del(&ocqe->list); | |
684 | kfree(ocqe); | |
685 | } | |
686 | } | |
687 | ||
1b346e4a | 688 | static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
1d7bb1d5 | 689 | { |
e45a3e05 | 690 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 691 | |
a85381d8 | 692 | if (__io_cqring_events(ctx) == ctx->cq_entries) |
1b346e4a | 693 | return; |
1d7bb1d5 | 694 | |
e45a3e05 SR |
695 | if (ctx->flags & IORING_SETUP_CQE32) |
696 | cqe_size <<= 1; | |
697 | ||
25399321 | 698 | io_cq_lock(ctx); |
6c2450ae | 699 | while (!list_empty(&ctx->cq_overflow_list)) { |
59fbc409 | 700 | struct io_uring_cqe *cqe; |
6c2450ae | 701 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 702 | |
59fbc409 | 703 | if (!io_get_cqe_overflow(ctx, &cqe, true)) |
1d7bb1d5 | 704 | break; |
6c2450ae PB |
705 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
706 | struct io_overflow_cqe, list); | |
a85381d8 | 707 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae PB |
708 | list_del(&ocqe->list); |
709 | kfree(ocqe); | |
1d7bb1d5 JA |
710 | } |
711 | ||
1b346e4a | 712 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 713 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 714 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 715 | } |
25399321 | 716 | io_cq_unlock_post(ctx); |
1d7bb1d5 JA |
717 | } |
718 | ||
52ea806a JA |
719 | static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx) |
720 | { | |
721 | /* iopoll syncs against uring_lock, not completion_lock */ | |
722 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
723 | mutex_lock(&ctx->uring_lock); | |
724 | __io_cqring_overflow_flush(ctx); | |
725 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
726 | mutex_unlock(&ctx->uring_lock); | |
727 | } | |
728 | ||
1b346e4a | 729 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 730 | { |
52ea806a JA |
731 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) |
732 | io_cqring_do_overflow_flush(ctx); | |
6c503150 PB |
733 | } |
734 | ||
5afa4650 | 735 | /* can be called by any task */ |
2fdd6fb5 | 736 | static void io_put_task_remote(struct task_struct *task) |
6a290a14 PB |
737 | { |
738 | struct io_uring_task *tctx = task->io_uring; | |
739 | ||
2fdd6fb5 | 740 | percpu_counter_sub(&tctx->inflight, 1); |
8d664282 | 741 | if (unlikely(atomic_read(&tctx->in_cancel))) |
9d170164 | 742 | wake_up(&tctx->wait); |
2fdd6fb5 | 743 | put_task_struct(task); |
9d170164 PB |
744 | } |
745 | ||
5afa4650 | 746 | /* used by a task to put its own references */ |
2fdd6fb5 | 747 | static void io_put_task_local(struct task_struct *task) |
5afa4650 | 748 | { |
2fdd6fb5 | 749 | task->io_uring->cached_refs++; |
5afa4650 PB |
750 | } |
751 | ||
89800a2d | 752 | /* must to be called somewhat shortly after putting a request */ |
2fdd6fb5 | 753 | static inline void io_put_task(struct task_struct *task) |
89800a2d PB |
754 | { |
755 | if (likely(task == current)) | |
2fdd6fb5 | 756 | io_put_task_local(task); |
89800a2d | 757 | else |
2fdd6fb5 | 758 | io_put_task_remote(task); |
89800a2d PB |
759 | } |
760 | ||
63809137 | 761 | void io_task_refs_refill(struct io_uring_task *tctx) |
9a10867a PB |
762 | { |
763 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
764 | ||
765 | percpu_counter_add(&tctx->inflight, refill); | |
766 | refcount_add(refill, ¤t->usage); | |
767 | tctx->cached_refs += refill; | |
768 | } | |
769 | ||
3cc7fdb9 PB |
770 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
771 | { | |
772 | struct io_uring_task *tctx = task->io_uring; | |
773 | unsigned int refs = tctx->cached_refs; | |
774 | ||
775 | if (refs) { | |
776 | tctx->cached_refs = 0; | |
777 | percpu_counter_sub(&tctx->inflight, refs); | |
778 | put_task_struct_many(task, refs); | |
779 | } | |
780 | } | |
781 | ||
68494a65 PB |
782 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
783 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 784 | { |
cce4b8b0 | 785 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
786 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
787 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 788 | |
f26cc959 PB |
789 | lockdep_assert_held(&ctx->completion_lock); |
790 | ||
e45a3e05 SR |
791 | if (is_cqe32) |
792 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 793 | |
e45a3e05 | 794 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 795 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
796 | if (!ocqe) { |
797 | /* | |
798 | * If we're in ring overflow flush mode, or in task cancel mode, | |
799 | * or cannot allocate an overflow entry, then we need to drop it | |
800 | * on the floor. | |
801 | */ | |
8f6ed49a | 802 | io_account_cq_overflow(ctx); |
155bc950 | 803 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 804 | return false; |
2b188cc1 | 805 | } |
cce4b8b0 | 806 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 807 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 808 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 809 | |
cce4b8b0 | 810 | } |
d4d19c19 | 811 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
812 | ocqe->cqe.res = res; |
813 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
814 | if (is_cqe32) { |
815 | ocqe->cqe.big_cqe[0] = extra1; | |
816 | ocqe->cqe.big_cqe[1] = extra2; | |
817 | } | |
cce4b8b0 PB |
818 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
819 | return true; | |
2b188cc1 JA |
820 | } |
821 | ||
056695bf | 822 | void io_req_cqe_overflow(struct io_kiocb *req) |
68494a65 | 823 | { |
056695bf PB |
824 | io_cqring_event_overflow(req->ctx, req->cqe.user_data, |
825 | req->cqe.res, req->cqe.flags, | |
b24c5d75 PB |
826 | req->big_cqe.extra1, req->big_cqe.extra2); |
827 | memset(&req->big_cqe, 0, sizeof(req->big_cqe)); | |
68494a65 PB |
828 | } |
829 | ||
faf88dde PB |
830 | /* |
831 | * writes to the cq entry need to come after reading head; the | |
832 | * control dependency is enough as we're using WRITE_ONCE to | |
833 | * fill the cq entry | |
834 | */ | |
20d6b633 | 835 | bool io_cqe_cache_refill(struct io_ring_ctx *ctx, bool overflow) |
faf88dde PB |
836 | { |
837 | struct io_rings *rings = ctx->rings; | |
838 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
839 | unsigned int free, queued, len; |
840 | ||
aa1df3a3 PB |
841 | /* |
842 | * Posting into the CQ when there are pending overflowed CQEs may break | |
843 | * ordering guarantees, which will affect links, F_MORE users and more. | |
844 | * Force overflow the completion. | |
845 | */ | |
846 | if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) | |
20d6b633 | 847 | return false; |
faf88dde PB |
848 | |
849 | /* userspace may cheat modifying the tail, be safe and do min */ | |
850 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
851 | free = ctx->cq_entries - queued; | |
852 | /* we need a contiguous range, limit based on the current array offset */ | |
853 | len = min(free, ctx->cq_entries - off); | |
854 | if (!len) | |
20d6b633 | 855 | return false; |
faf88dde | 856 | |
b3659a65 PB |
857 | if (ctx->flags & IORING_SETUP_CQE32) { |
858 | off <<= 1; | |
859 | len <<= 1; | |
860 | } | |
861 | ||
faf88dde PB |
862 | ctx->cqe_cached = &rings->cqes[off]; |
863 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
20d6b633 | 864 | return true; |
faf88dde PB |
865 | } |
866 | ||
f66f7342 PB |
867 | static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, |
868 | u32 cflags) | |
bcda7baa | 869 | { |
cd94903d PB |
870 | struct io_uring_cqe *cqe; |
871 | ||
913a571a | 872 | ctx->cq_extra++; |
cd94903d PB |
873 | |
874 | /* | |
875 | * If we can't get a cq entry, userspace overflowed the | |
876 | * submission (by quite a lot). Increment the overflow count in | |
877 | * the ring. | |
878 | */ | |
59fbc409 | 879 | if (likely(io_get_cqe(ctx, &cqe))) { |
e0486f3f DY |
880 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
881 | ||
cd94903d PB |
882 | WRITE_ONCE(cqe->user_data, user_data); |
883 | WRITE_ONCE(cqe->res, res); | |
884 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
885 | |
886 | if (ctx->flags & IORING_SETUP_CQE32) { | |
887 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
888 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
889 | } | |
cd94903d PB |
890 | return true; |
891 | } | |
52120f0f | 892 | return false; |
bcda7baa JA |
893 | } |
894 | ||
e5c12945 | 895 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags) |
d245bca6 PB |
896 | { |
897 | bool filled; | |
898 | ||
25399321 | 899 | io_cq_lock(ctx); |
f66f7342 | 900 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags); |
e5c12945 | 901 | if (!filled) |
f66f7342 PB |
902 | filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); |
903 | ||
25399321 | 904 | io_cq_unlock_post(ctx); |
d245bca6 PB |
905 | return filled; |
906 | } | |
907 | ||
b6b2bb58 PB |
908 | /* |
909 | * A helper for multishot requests posting additional CQEs. | |
910 | * Should only be used from a task_work including IO_URING_F_MULTISHOT. | |
911 | */ | |
e5c12945 | 912 | bool io_req_post_cqe(struct io_kiocb *req, s32 res, u32 cflags) |
2b188cc1 | 913 | { |
d86eaed1 | 914 | struct io_ring_ctx *ctx = req->ctx; |
902ce82c | 915 | bool posted; |
9b8c5475 | 916 | |
e0e4ab52 | 917 | lockdep_assert(!io_wq_current_is_worker()); |
9b8c5475 DY |
918 | lockdep_assert_held(&ctx->uring_lock); |
919 | ||
902ce82c PB |
920 | __io_cq_lock(ctx); |
921 | posted = io_fill_cqe_aux(ctx, req->cqe.user_data, res, cflags); | |
922 | ctx->submit_state.cq_flush = true; | |
923 | __io_cq_unlock_post(ctx); | |
924 | return posted; | |
9b8c5475 DY |
925 | } |
926 | ||
0667db14 | 927 | static void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
2b188cc1 | 928 | { |
fa18fa22 | 929 | struct io_ring_ctx *ctx = req->ctx; |
2ad4c6d0 | 930 | struct io_rsrc_node *rsrc_node = NULL; |
fa18fa22 | 931 | |
0667db14 PB |
932 | /* |
933 | * Handle special CQ sync cases via task_work. DEFER_TASKRUN requires | |
934 | * the submitter task context, IOPOLL protects with uring_lock. | |
935 | */ | |
936 | if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL)) { | |
937 | req->io_task_work.func = io_req_task_complete; | |
938 | io_req_task_work_add(req); | |
939 | return; | |
940 | } | |
941 | ||
fa18fa22 | 942 | io_cq_lock(ctx); |
00b0db56 PB |
943 | if (!(req->flags & REQ_F_CQE_SKIP)) { |
944 | if (!io_fill_cqe_req(ctx, req)) | |
945 | io_req_cqe_overflow(req); | |
946 | } | |
fa18fa22 | 947 | |
c7dae4ba JA |
948 | /* |
949 | * If we're the last reference to this request, add to our locked | |
950 | * free_list cache. | |
951 | */ | |
de9b4cca | 952 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 953 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 954 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
955 | io_disarm_next(req); |
956 | if (req->link) { | |
957 | io_req_task_queue(req->link); | |
958 | req->link = NULL; | |
959 | } | |
960 | } | |
68a2cc1b | 961 | io_put_kbuf_comp(req); |
3b7a612f PB |
962 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
963 | io_clean_op(req); | |
17bc2837 | 964 | io_put_file(req); |
3b7a612f | 965 | |
2ad4c6d0 | 966 | rsrc_node = req->rsrc_node; |
8197b053 PB |
967 | /* |
968 | * Selected buffer deallocation in io_clean_op() assumes that | |
969 | * we don't hold ->completion_lock. Clean them here to avoid | |
970 | * deadlocks. | |
971 | */ | |
2fdd6fb5 | 972 | io_put_task_remote(req->task); |
c2b6c6bc | 973 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 974 | ctx->locked_free_nr++; |
180f829f | 975 | } |
25399321 | 976 | io_cq_unlock_post(ctx); |
2ad4c6d0 | 977 | |
ef8ae64f PB |
978 | if (rsrc_node) { |
979 | io_ring_submit_lock(ctx, issue_flags); | |
1f2c8f61 | 980 | io_put_rsrc_node(ctx, rsrc_node); |
ef8ae64f PB |
981 | io_ring_submit_unlock(ctx, issue_flags); |
982 | } | |
4e3d9ff9 JA |
983 | } |
984 | ||
973fc83f | 985 | void io_req_defer_failed(struct io_kiocb *req, s32 res) |
e276ae34 | 986 | __must_hold(&ctx->uring_lock) |
f41db273 | 987 | { |
f30bd4d0 | 988 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; |
a47b255e | 989 | |
e276ae34 PB |
990 | lockdep_assert_held(&req->ctx->uring_lock); |
991 | ||
93d2bcd2 | 992 | req_set_fail(req); |
97b388d7 | 993 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
a47b255e PB |
994 | if (def->fail) |
995 | def->fail(req); | |
973fc83f | 996 | io_req_complete_defer(req); |
f41db273 PB |
997 | } |
998 | ||
864ea921 PB |
999 | /* |
1000 | * Don't initialise the fields below on every allocation, but do that in | |
1001 | * advance and keep them valid across allocations. | |
1002 | */ | |
1003 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
1004 | { | |
1005 | req->ctx = ctx; | |
1006 | req->link = NULL; | |
1007 | req->async_data = NULL; | |
1008 | /* not necessary, but safer to zero */ | |
31d3ba92 | 1009 | memset(&req->cqe, 0, sizeof(req->cqe)); |
b24c5d75 | 1010 | memset(&req->big_cqe, 0, sizeof(req->big_cqe)); |
864ea921 PB |
1011 | } |
1012 | ||
dac7a098 | 1013 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 1014 | struct io_submit_state *state) |
dac7a098 | 1015 | { |
79ebeaee | 1016 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1017 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 1018 | ctx->locked_free_nr = 0; |
79ebeaee | 1019 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
1020 | } |
1021 | ||
5d5901a3 PB |
1022 | /* |
1023 | * A request might get retired back into the request caches even before opcode | |
1024 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
1025 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
1026 | * and with extra caution to not get a request that is still worked on. | |
1027 | */ | |
bd1a3783 | 1028 | __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 1029 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 1030 | { |
864ea921 | 1031 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 1032 | void *reqs[IO_REQ_ALLOC_BATCH]; |
05eb5fe2 | 1033 | int ret; |
e5d1bc0a | 1034 | |
23a5c43b PB |
1035 | /* |
1036 | * If we have more than a batch's worth of requests in our IRQ side | |
1037 | * locked cache, grab the lock and move them over to our submission | |
1038 | * side cache. | |
1039 | */ | |
a6d97a8a | 1040 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 1041 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 1042 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
1043 | return true; |
1044 | } | |
e5d1bc0a | 1045 | |
3ab665b7 | 1046 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 1047 | |
864ea921 PB |
1048 | /* |
1049 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1050 | * retry single alloc to be on the safe side. | |
1051 | */ | |
1052 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
1053 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
1054 | if (!reqs[0]) | |
a33ae9ce | 1055 | return false; |
864ea921 | 1056 | ret = 1; |
2b188cc1 | 1057 | } |
864ea921 | 1058 | |
37f0e767 | 1059 | percpu_ref_get_many(&ctx->refs, ret); |
05eb5fe2 JA |
1060 | while (ret--) { |
1061 | struct io_kiocb *req = reqs[ret]; | |
3ab665b7 PB |
1062 | |
1063 | io_preinit_req(req, ctx); | |
fa05457a | 1064 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 1065 | } |
a33ae9ce PB |
1066 | return true; |
1067 | } | |
1068 | ||
03adabe8 PB |
1069 | __cold void io_free_req(struct io_kiocb *req) |
1070 | { | |
6ec9afc7 PB |
1071 | /* refs were already put, restore them for io_req_task_complete() */ |
1072 | req->flags &= ~REQ_F_REFCOUNT; | |
1073 | /* we only want to free it, don't post CQEs */ | |
1074 | req->flags |= REQ_F_CQE_SKIP; | |
1075 | req->io_task_work.func = io_req_task_complete; | |
03adabe8 PB |
1076 | io_req_task_work_add(req); |
1077 | } | |
1078 | ||
d81499bf PB |
1079 | static void __io_req_find_next_prep(struct io_kiocb *req) |
1080 | { | |
1081 | struct io_ring_ctx *ctx = req->ctx; | |
d81499bf | 1082 | |
6971253f | 1083 | spin_lock(&ctx->completion_lock); |
305bef98 | 1084 | io_disarm_next(req); |
6971253f | 1085 | spin_unlock(&ctx->completion_lock); |
d81499bf PB |
1086 | } |
1087 | ||
1088 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 1089 | { |
33cc89a9 | 1090 | struct io_kiocb *nxt; |
944e58bf | 1091 | |
9e645e11 JA |
1092 | /* |
1093 | * If LINK is set, we have dependent requests in this chain. If we | |
1094 | * didn't fail this request, queue the first one up, moving any other | |
1095 | * dependencies to the next request. In case of failure, fail the rest | |
1096 | * of the chain. | |
1097 | */ | |
d81499bf PB |
1098 | if (unlikely(req->flags & IO_DISARM_MASK)) |
1099 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
1100 | nxt = req->link; |
1101 | req->link = NULL; | |
1102 | return nxt; | |
4d7dd462 | 1103 | } |
9e645e11 | 1104 | |
a282967c | 1105 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts) |
2c32395d PB |
1106 | { |
1107 | if (!ctx) | |
1108 | return; | |
ef060ea9 JA |
1109 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1110 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
8e5b3b89 PB |
1111 | |
1112 | io_submit_flush_completions(ctx); | |
1113 | mutex_unlock(&ctx->uring_lock); | |
2c32395d PB |
1114 | percpu_ref_put(&ctx->refs); |
1115 | } | |
1116 | ||
af5d68f8 JA |
1117 | /* |
1118 | * Run queued task_work, returning the number of entries processed in *count. | |
1119 | * If more entries than max_entries are available, stop processing once this | |
1120 | * is reached and return the rest of the list. | |
1121 | */ | |
1122 | struct llist_node *io_handle_tw_list(struct llist_node *node, | |
1123 | unsigned int *count, | |
1124 | unsigned int max_entries) | |
9f8d032a | 1125 | { |
42c0905f JA |
1126 | struct io_ring_ctx *ctx = NULL; |
1127 | struct io_tw_state ts = { }; | |
c6dd763c | 1128 | |
592b4805 | 1129 | do { |
f88262e6 | 1130 | struct llist_node *next = node->next; |
9f8d032a HX |
1131 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1132 | io_task_work.node); | |
1133 | ||
42c0905f JA |
1134 | if (req->ctx != ctx) { |
1135 | ctx_flush_and_put(ctx, &ts); | |
1136 | ctx = req->ctx; | |
92219afb | 1137 | mutex_lock(&ctx->uring_lock); |
42c0905f | 1138 | percpu_ref_get(&ctx->refs); |
13bfa6f1 | 1139 | } |
c92fcfc2 JA |
1140 | INDIRECT_CALL_2(req->io_task_work.func, |
1141 | io_poll_task_func, io_req_rw_complete, | |
42c0905f | 1142 | req, &ts); |
9f8d032a | 1143 | node = next; |
2708af1a | 1144 | (*count)++; |
f5868008 | 1145 | if (unlikely(need_resched())) { |
42c0905f JA |
1146 | ctx_flush_and_put(ctx, &ts); |
1147 | ctx = NULL; | |
f5868008 JA |
1148 | cond_resched(); |
1149 | } | |
af5d68f8 | 1150 | } while (node && *count < max_entries); |
c6dd763c | 1151 | |
42c0905f | 1152 | ctx_flush_and_put(ctx, &ts); |
af5d68f8 | 1153 | return node; |
9f8d032a HX |
1154 | } |
1155 | ||
923d1592 DY |
1156 | /** |
1157 | * io_llist_xchg - swap all entries in a lock-less list | |
1158 | * @head: the head of lock-less list to delete all entries | |
1159 | * @new: new entry as the head of the list | |
1160 | * | |
1161 | * If list is empty, return NULL, otherwise, return the pointer to the first entry. | |
1162 | * The order of entries returned is from the newest to the oldest added one. | |
1163 | */ | |
1164 | static inline struct llist_node *io_llist_xchg(struct llist_head *head, | |
1165 | struct llist_node *new) | |
1166 | { | |
1167 | return xchg(&head->first, new); | |
1168 | } | |
1169 | ||
dfbe5561 | 1170 | static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync) |
10e1c0d5 JA |
1171 | { |
1172 | struct llist_node *node = llist_del_all(&tctx->task_list); | |
dfbe5561 | 1173 | struct io_ring_ctx *last_ctx = NULL; |
10e1c0d5 JA |
1174 | struct io_kiocb *req; |
1175 | ||
1176 | while (node) { | |
1177 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1178 | node = node->next; | |
dfbe5561 JA |
1179 | if (sync && last_ctx != req->ctx) { |
1180 | if (last_ctx) { | |
1181 | flush_delayed_work(&last_ctx->fallback_work); | |
1182 | percpu_ref_put(&last_ctx->refs); | |
1183 | } | |
1184 | last_ctx = req->ctx; | |
1185 | percpu_ref_get(&last_ctx->refs); | |
1186 | } | |
10e1c0d5 JA |
1187 | if (llist_add(&req->io_task_work.node, |
1188 | &req->ctx->fallback_llist)) | |
1189 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1190 | } | |
dfbe5561 JA |
1191 | |
1192 | if (last_ctx) { | |
1193 | flush_delayed_work(&last_ctx->fallback_work); | |
1194 | percpu_ref_put(&last_ctx->refs); | |
1195 | } | |
10e1c0d5 JA |
1196 | } |
1197 | ||
af5d68f8 JA |
1198 | struct llist_node *tctx_task_work_run(struct io_uring_task *tctx, |
1199 | unsigned int max_entries, | |
1200 | unsigned int *count) | |
c40f6379 | 1201 | { |
77e443ab | 1202 | struct llist_node *node; |
3a0c037b | 1203 | |
77e443ab | 1204 | if (unlikely(current->flags & PF_EXITING)) { |
dfbe5561 | 1205 | io_fallback_tw(tctx, true); |
af5d68f8 | 1206 | return NULL; |
77e443ab | 1207 | } |
3a0c037b | 1208 | |
592b4805 | 1209 | node = llist_del_all(&tctx->task_list); |
af5d68f8 JA |
1210 | if (node) { |
1211 | node = llist_reverse_order(node); | |
1212 | node = io_handle_tw_list(node, count, max_entries); | |
1213 | } | |
3cc7fdb9 | 1214 | |
8d664282 JA |
1215 | /* relaxed read is enough as only the task itself sets ->in_cancel */ |
1216 | if (unlikely(atomic_read(&tctx->in_cancel))) | |
3cc7fdb9 | 1217 | io_uring_drop_tctx_refs(current); |
c6dd763c | 1218 | |
af5d68f8 JA |
1219 | trace_io_uring_task_work_run(tctx, *count); |
1220 | return node; | |
1221 | } | |
1222 | ||
1223 | void tctx_task_work(struct callback_head *cb) | |
1224 | { | |
1225 | struct io_uring_task *tctx; | |
1226 | struct llist_node *ret; | |
1227 | unsigned int count = 0; | |
1228 | ||
1229 | tctx = container_of(cb, struct io_uring_task, task_work); | |
1230 | ret = tctx_task_work_run(tctx, UINT_MAX, &count); | |
1231 | /* can't happen */ | |
1232 | WARN_ON_ONCE(ret); | |
7cbf1722 JA |
1233 | } |
1234 | ||
91c7884a | 1235 | static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags) |
c0e0d6ba DY |
1236 | { |
1237 | struct io_ring_ctx *ctx = req->ctx; | |
8751d154 | 1238 | unsigned nr_wait, nr_tw, nr_tw_prev; |
e8c40771 | 1239 | struct llist_node *head; |
c0e0d6ba | 1240 | |
b4bc35cf PB |
1241 | /* See comment above IO_CQ_WAKE_INIT */ |
1242 | BUILD_BUG_ON(IO_CQ_WAKE_FORCE <= IORING_MAX_CQ_ENTRIES); | |
c0e0d6ba | 1243 | |
b4bc35cf PB |
1244 | /* |
1245 | * We don't know how many reuqests is there in the link and whether | |
1246 | * they can even be queued lazily, fall back to non-lazy. | |
1247 | */ | |
8751d154 PB |
1248 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) |
1249 | flags &= ~IOU_F_TWQ_LAZY_WAKE; | |
ce8e04f6 | 1250 | |
e8c40771 | 1251 | head = READ_ONCE(ctx->work_llist.first); |
51509400 | 1252 | do { |
8751d154 | 1253 | nr_tw_prev = 0; |
e8c40771 PB |
1254 | if (head) { |
1255 | struct io_kiocb *first_req = container_of(head, | |
8751d154 PB |
1256 | struct io_kiocb, |
1257 | io_task_work.node); | |
1258 | /* | |
1259 | * Might be executed at any moment, rely on | |
1260 | * SLAB_TYPESAFE_BY_RCU to keep it alive. | |
1261 | */ | |
1262 | nr_tw_prev = READ_ONCE(first_req->nr_tw); | |
1263 | } | |
b4bc35cf PB |
1264 | |
1265 | /* | |
1266 | * Theoretically, it can overflow, but that's fine as one of | |
1267 | * previous adds should've tried to wake the task. | |
1268 | */ | |
8751d154 | 1269 | nr_tw = nr_tw_prev + 1; |
8751d154 | 1270 | if (!(flags & IOU_F_TWQ_LAZY_WAKE)) |
b4bc35cf | 1271 | nr_tw = IO_CQ_WAKE_FORCE; |
8751d154 PB |
1272 | |
1273 | req->nr_tw = nr_tw; | |
e8c40771 PB |
1274 | req->io_task_work.node.next = head; |
1275 | } while (!try_cmpxchg(&ctx->work_llist.first, &head, | |
51509400 PB |
1276 | &req->io_task_work.node)); |
1277 | ||
d381099f PB |
1278 | /* |
1279 | * cmpxchg implies a full barrier, which pairs with the barrier | |
1280 | * in set_current_state() on the io_cqring_wait() side. It's used | |
1281 | * to ensure that either we see updated ->cq_wait_nr, or waiters | |
1282 | * going to sleep will observe the work added to the list, which | |
1283 | * is similar to the wait/wawke task state sync. | |
1284 | */ | |
1285 | ||
e8c40771 | 1286 | if (!head) { |
8751d154 PB |
1287 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1288 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1289 | if (ctx->has_evfd) | |
1290 | io_eventfd_signal(ctx); | |
c0e0d6ba DY |
1291 | } |
1292 | ||
8751d154 | 1293 | nr_wait = atomic_read(&ctx->cq_wait_nr); |
b4bc35cf PB |
1294 | /* not enough or no one is waiting */ |
1295 | if (nr_tw < nr_wait) | |
8751d154 | 1296 | return; |
b4bc35cf PB |
1297 | /* the previous add has already woken it up */ |
1298 | if (nr_tw_prev >= nr_wait) | |
8751d154 | 1299 | return; |
8751d154 | 1300 | wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE); |
c0e0d6ba DY |
1301 | } |
1302 | ||
91c7884a | 1303 | static void io_req_normal_work_add(struct io_kiocb *req) |
7cbf1722 | 1304 | { |
c34398a8 | 1305 | struct io_uring_task *tctx = req->task->io_uring; |
9f010507 | 1306 | struct io_ring_ctx *ctx = req->ctx; |
7cbf1722 | 1307 | |
7cbf1722 | 1308 | /* task_work already pending, we're done */ |
32d91f05 | 1309 | if (!llist_add(&req->io_task_work.node, &tctx->task_list)) |
e09ee510 | 1310 | return; |
7cbf1722 | 1311 | |
ef060ea9 JA |
1312 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1313 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1314 | ||
af5d68f8 | 1315 | /* SQPOLL doesn't need the task_work added, it'll run it itself */ |
78f9b61b JA |
1316 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
1317 | struct io_sq_data *sqd = ctx->sq_data; | |
1318 | ||
1319 | if (wq_has_sleeper(&sqd->wait)) | |
1320 | wake_up(&sqd->wait); | |
af5d68f8 | 1321 | return; |
78f9b61b | 1322 | } |
af5d68f8 | 1323 | |
3fe07bcd | 1324 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1325 | return; |
2215bed9 | 1326 | |
dfbe5561 | 1327 | io_fallback_tw(tctx, false); |
c0e0d6ba DY |
1328 | } |
1329 | ||
91c7884a PB |
1330 | void __io_req_task_work_add(struct io_kiocb *req, unsigned flags) |
1331 | { | |
1332 | if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) { | |
1333 | rcu_read_lock(); | |
1334 | io_req_local_work_add(req, flags); | |
1335 | rcu_read_unlock(); | |
1336 | } else { | |
1337 | io_req_normal_work_add(req); | |
1338 | } | |
1339 | } | |
1340 | ||
c0e0d6ba DY |
1341 | static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) |
1342 | { | |
1343 | struct llist_node *node; | |
1344 | ||
1345 | node = llist_del_all(&ctx->work_llist); | |
1346 | while (node) { | |
1347 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1348 | io_task_work.node); | |
1349 | ||
1350 | node = node->next; | |
91c7884a | 1351 | io_req_normal_work_add(req); |
c0e0d6ba DY |
1352 | } |
1353 | } | |
1354 | ||
9fe3eaea JA |
1355 | static bool io_run_local_work_continue(struct io_ring_ctx *ctx, int events, |
1356 | int min_events) | |
1357 | { | |
1358 | if (llist_empty(&ctx->work_llist)) | |
1359 | return false; | |
1360 | if (events < min_events) | |
1361 | return true; | |
1362 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) | |
1363 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1364 | return false; | |
1365 | } | |
1366 | ||
1367 | static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts, | |
1368 | int min_events) | |
c0e0d6ba | 1369 | { |
c0e0d6ba | 1370 | struct llist_node *node; |
c3f4d39e | 1371 | unsigned int loops = 0; |
140102ae | 1372 | int ret = 0; |
c0e0d6ba | 1373 | |
140102ae | 1374 | if (WARN_ON_ONCE(ctx->submitter_task != current)) |
c0e0d6ba | 1375 | return -EEXIST; |
c3f4d39e PB |
1376 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1377 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
c0e0d6ba | 1378 | again: |
3af0356c JA |
1379 | /* |
1380 | * llists are in reverse order, flip it back the right way before | |
1381 | * running the pending items. | |
1382 | */ | |
1383 | node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL)); | |
c3f4d39e | 1384 | while (node) { |
c0e0d6ba DY |
1385 | struct llist_node *next = node->next; |
1386 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1387 | io_task_work.node); | |
c92fcfc2 JA |
1388 | INDIRECT_CALL_2(req->io_task_work.func, |
1389 | io_poll_task_func, io_req_rw_complete, | |
1390 | req, ts); | |
c0e0d6ba DY |
1391 | ret++; |
1392 | node = next; | |
1393 | } | |
c3f4d39e | 1394 | loops++; |
c0e0d6ba | 1395 | |
9fe3eaea | 1396 | if (io_run_local_work_continue(ctx, ret, min_events)) |
c0e0d6ba | 1397 | goto again; |
92219afb PB |
1398 | io_submit_flush_completions(ctx); |
1399 | if (io_run_local_work_continue(ctx, ret, min_events)) | |
1400 | goto again; | |
9fe3eaea | 1401 | |
f75d5036 | 1402 | trace_io_uring_local_work_run(ctx, ret, loops); |
c0e0d6ba | 1403 | return ret; |
8ac5d85a JA |
1404 | } |
1405 | ||
9fe3eaea JA |
1406 | static inline int io_run_local_work_locked(struct io_ring_ctx *ctx, |
1407 | int min_events) | |
360173ab | 1408 | { |
8e5b3b89 | 1409 | struct io_tw_state ts = {}; |
360173ab PB |
1410 | |
1411 | if (llist_empty(&ctx->work_llist)) | |
1412 | return 0; | |
8e5b3b89 | 1413 | return __io_run_local_work(ctx, &ts, min_events); |
360173ab PB |
1414 | } |
1415 | ||
9fe3eaea | 1416 | static int io_run_local_work(struct io_ring_ctx *ctx, int min_events) |
8ac5d85a | 1417 | { |
8e5b3b89 | 1418 | struct io_tw_state ts = {}; |
8ac5d85a JA |
1419 | int ret; |
1420 | ||
92219afb | 1421 | mutex_lock(&ctx->uring_lock); |
9fe3eaea | 1422 | ret = __io_run_local_work(ctx, &ts, min_events); |
92219afb | 1423 | mutex_unlock(&ctx->uring_lock); |
8ac5d85a | 1424 | return ret; |
c0e0d6ba DY |
1425 | } |
1426 | ||
a282967c | 1427 | static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts) |
c40f6379 | 1428 | { |
a282967c | 1429 | io_tw_lock(req->ctx, ts); |
973fc83f | 1430 | io_req_defer_failed(req, req->cqe.res); |
c40f6379 JA |
1431 | } |
1432 | ||
a282967c | 1433 | void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts) |
c40f6379 | 1434 | { |
a282967c | 1435 | io_tw_lock(req->ctx, ts); |
316319e8 | 1436 | /* req->task == current here, checking PF_EXITING is safe */ |
6bb30855 | 1437 | if (unlikely(req->task->flags & PF_EXITING)) |
973fc83f | 1438 | io_req_defer_failed(req, -EFAULT); |
6bb30855 | 1439 | else if (req->flags & REQ_F_FORCE_ASYNC) |
6e6b8c62 | 1440 | io_queue_iowq(req); |
6bb30855 DY |
1441 | else |
1442 | io_queue_sqe(req); | |
c40f6379 JA |
1443 | } |
1444 | ||
59915143 | 1445 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1446 | { |
97b388d7 | 1447 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1448 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1449 | io_req_task_work_add(req); |
c40f6379 JA |
1450 | } |
1451 | ||
f3b44f92 | 1452 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1453 | { |
5b0a6acc | 1454 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1455 | io_req_task_work_add(req); |
a3df7698 PB |
1456 | } |
1457 | ||
59915143 | 1458 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1459 | { |
9b5f7bd9 | 1460 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1461 | |
1462 | if (nxt) | |
906a8c3f | 1463 | io_req_task_queue(nxt); |
c69f8dbe JL |
1464 | } |
1465 | ||
ec26c225 PB |
1466 | static void io_free_batch_list(struct io_ring_ctx *ctx, |
1467 | struct io_wq_work_node *node) | |
3aa83bfb | 1468 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1469 | { |
3aa83bfb PB |
1470 | do { |
1471 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1472 | comp_list); | |
2d6500d4 | 1473 | |
a538be5b PB |
1474 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1475 | if (req->flags & REQ_F_REFCOUNT) { | |
1476 | node = req->comp_list.next; | |
1477 | if (!req_ref_put_and_test(req)) | |
1478 | continue; | |
1479 | } | |
b605a7fa PB |
1480 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1481 | struct async_poll *apoll = req->apoll; | |
1482 | ||
1483 | if (apoll->double_poll) | |
1484 | kfree(apoll->double_poll); | |
414d0f45 | 1485 | if (!io_alloc_cache_put(&ctx->apoll_cache, apoll)) |
9731bc98 | 1486 | kfree(apoll); |
b605a7fa PB |
1487 | req->flags &= ~REQ_F_POLLED; |
1488 | } | |
da1a08c5 | 1489 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1490 | io_queue_next(req); |
a538be5b PB |
1491 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1492 | io_clean_op(req); | |
c1e53a69 | 1493 | } |
17bc2837 | 1494 | io_put_file(req); |
2d6500d4 | 1495 | |
ab409402 | 1496 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1497 | |
2fdd6fb5 | 1498 | io_put_task(req->task); |
c1e53a69 | 1499 | node = req->comp_list.next; |
fa05457a | 1500 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1501 | } while (node); |
7a743e22 PB |
1502 | } |
1503 | ||
ec26c225 | 1504 | void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1505 | __must_hold(&ctx->uring_lock) |
905c172f | 1506 | { |
cd0ca2e0 | 1507 | struct io_submit_state *state = &ctx->submit_state; |
fa780334 | 1508 | struct io_wq_work_node *node; |
905c172f | 1509 | |
f66f7342 | 1510 | __io_cq_lock(ctx); |
fa780334 | 1511 | __wq_list_for_each(node, &state->compl_reqs) { |
d9dee430 PB |
1512 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1513 | comp_list); | |
3d4aeb9f | 1514 | |
f66f7342 | 1515 | if (!(req->flags & REQ_F_CQE_SKIP) && |
00b0db56 | 1516 | unlikely(!io_fill_cqe_req(ctx, req))) { |
27122c07 | 1517 | if (ctx->lockless_cq) { |
f66f7342 PB |
1518 | spin_lock(&ctx->completion_lock); |
1519 | io_req_cqe_overflow(req); | |
1520 | spin_unlock(&ctx->completion_lock); | |
1521 | } else { | |
1522 | io_req_cqe_overflow(req); | |
1523 | } | |
1524 | } | |
905c172f | 1525 | } |
c98c81a4 | 1526 | __io_cq_unlock_post(ctx); |
d9dee430 | 1527 | |
931147dd DY |
1528 | if (!wq_list_empty(&ctx->submit_state.compl_reqs)) { |
1529 | io_free_batch_list(ctx, state->compl_reqs.first); | |
1530 | INIT_WQ_LIST(&state->compl_reqs); | |
1531 | } | |
902ce82c | 1532 | ctx->submit_state.cq_flush = false; |
7a743e22 PB |
1533 | } |
1534 | ||
6c503150 | 1535 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1536 | { |
1537 | /* See comment at the top of this file */ | |
1538 | smp_rmb(); | |
e23de15f | 1539 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1540 | } |
1541 | ||
def596e9 JA |
1542 | /* |
1543 | * We can't just wait for polled events to come to us, we have to actively | |
1544 | * find and complete them. | |
1545 | */ | |
c072481d | 1546 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1547 | { |
1548 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1549 | return; | |
1550 | ||
1551 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1552 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1553 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1554 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1555 | break; |
08f5439f JA |
1556 | /* |
1557 | * Ensure we allow local-to-the-cpu processing to take place, | |
1558 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1559 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1560 | */ |
3fcee5a6 PB |
1561 | if (need_resched()) { |
1562 | mutex_unlock(&ctx->uring_lock); | |
1563 | cond_resched(); | |
1564 | mutex_lock(&ctx->uring_lock); | |
1565 | } | |
def596e9 JA |
1566 | } |
1567 | mutex_unlock(&ctx->uring_lock); | |
1568 | } | |
1569 | ||
7668b92a | 1570 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1571 | { |
7668b92a | 1572 | unsigned int nr_events = 0; |
155bc950 | 1573 | unsigned long check_cq; |
500f9fba | 1574 | |
76de6749 PB |
1575 | if (!io_allowed_run_tw(ctx)) |
1576 | return -EEXIST; | |
1577 | ||
3a08576b PB |
1578 | check_cq = READ_ONCE(ctx->check_cq); |
1579 | if (unlikely(check_cq)) { | |
1580 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
a85381d8 | 1581 | __io_cqring_overflow_flush(ctx); |
3a08576b PB |
1582 | /* |
1583 | * Similarly do not spin if we have not informed the user of any | |
1584 | * dropped CQE. | |
1585 | */ | |
1586 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1587 | return -EBADR; | |
1588 | } | |
f39c8a5b PB |
1589 | /* |
1590 | * Don't enter poll loop if we already have events pending. | |
1591 | * If we do, we can potentially be spinning for commands that | |
1592 | * already triggered a CQE (eg in error). | |
1593 | */ | |
f39c8a5b | 1594 | if (io_cqring_events(ctx)) |
d487b43c | 1595 | return 0; |
155bc950 | 1596 | |
def596e9 | 1597 | do { |
9e4bef2b JA |
1598 | int ret = 0; |
1599 | ||
500f9fba JA |
1600 | /* |
1601 | * If a submit got punted to a workqueue, we can have the | |
1602 | * application entering polling for a command before it gets | |
1603 | * issued. That app will hold the uring_lock for the duration | |
1604 | * of the poll right here, so we need to take a breather every | |
1605 | * now and then to ensure that the issue has a chance to add | |
1606 | * the poll to the issued list. Otherwise we can spin here | |
1607 | * forever, while the workqueue is stuck trying to acquire the | |
1608 | * very same mutex. | |
1609 | */ | |
dac6a0ea JA |
1610 | if (wq_list_empty(&ctx->iopoll_list) || |
1611 | io_task_work_pending(ctx)) { | |
8f487ef2 PB |
1612 | u32 tail = ctx->cached_cq_tail; |
1613 | ||
9fe3eaea | 1614 | (void) io_run_local_work_locked(ctx, min); |
def596e9 | 1615 | |
dac6a0ea JA |
1616 | if (task_work_pending(current) || |
1617 | wq_list_empty(&ctx->iopoll_list)) { | |
dac6a0ea | 1618 | mutex_unlock(&ctx->uring_lock); |
9d54bd6a | 1619 | io_run_task_work(); |
dac6a0ea | 1620 | mutex_lock(&ctx->uring_lock); |
dac6a0ea | 1621 | } |
8f487ef2 PB |
1622 | /* some requests don't go through iopoll_list */ |
1623 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1624 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1625 | break; |
500f9fba | 1626 | } |
5ba3c874 | 1627 | ret = io_do_iopoll(ctx, !min); |
9e4bef2b JA |
1628 | if (unlikely(ret < 0)) |
1629 | return ret; | |
dc314886 PB |
1630 | |
1631 | if (task_sigpending(current)) | |
1632 | return -EINTR; | |
9e4bef2b | 1633 | if (need_resched()) |
5ba3c874 | 1634 | break; |
d487b43c | 1635 | |
5ba3c874 | 1636 | nr_events += ret; |
9e4bef2b | 1637 | } while (nr_events < min); |
d487b43c | 1638 | |
9e4bef2b | 1639 | return 0; |
def596e9 | 1640 | } |
7012c815 | 1641 | |
a282967c | 1642 | void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts) |
8ef12efe | 1643 | { |
8e5b3b89 | 1644 | io_req_complete_defer(req); |
8ef12efe JA |
1645 | } |
1646 | ||
def596e9 JA |
1647 | /* |
1648 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1649 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1650 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1651 | * accessing the kiocb cookie. |
1652 | */ | |
9882131c | 1653 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1654 | { |
1655 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1656 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1657 | |
1658 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1659 | if (unlikely(needs_lock)) |
cb3d8972 | 1660 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1661 | |
1662 | /* | |
1663 | * Track whether we have multiple files in our lists. This will impact | |
1664 | * how we do polling eventually, not spinning if we're on potentially | |
1665 | * different devices. | |
1666 | */ | |
5eef4e87 | 1667 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1668 | ctx->poll_multi_queue = false; |
1669 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1670 | struct io_kiocb *list_req; |
1671 | ||
5eef4e87 PB |
1672 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1673 | comp_list); | |
30da1b45 | 1674 | if (list_req->file != req->file) |
915b3dde | 1675 | ctx->poll_multi_queue = true; |
def596e9 JA |
1676 | } |
1677 | ||
1678 | /* | |
1679 | * For fast devices, IO may have already completed. If it has, add | |
1680 | * it to the front so we find it first. | |
1681 | */ | |
65a6543d | 1682 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1683 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1684 | else |
5eef4e87 | 1685 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1686 | |
3b44b371 | 1687 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1688 | /* |
1689 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1690 | * in sq thread task context or in io worker task context. If | |
1691 | * current task context is sq thread, we don't need to check | |
1692 | * whether should wake up sq thread. | |
1693 | */ | |
1694 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1695 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1696 | wake_up(&ctx->sq_data->wait); | |
1697 | ||
1698 | mutex_unlock(&ctx->uring_lock); | |
1699 | } | |
def596e9 JA |
1700 | } |
1701 | ||
4bcb982c | 1702 | io_req_flags_t io_file_get_flags(struct file *file) |
88459b50 | 1703 | { |
4bcb982c | 1704 | io_req_flags_t res = 0; |
af197f50 | 1705 | |
53cfd5ce | 1706 | if (S_ISREG(file_inode(file)->i_mode)) |
8487f083 | 1707 | res |= REQ_F_ISREG; |
b9a6c945 | 1708 | if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT)) |
8487f083 | 1709 | res |= REQ_F_SUPPORT_NOWAIT; |
88459b50 | 1710 | return res; |
2b188cc1 JA |
1711 | } |
1712 | ||
99f15d8d | 1713 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1714 | { |
e10677a8 JA |
1715 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
1716 | ||
1717 | WARN_ON_ONCE(!def->async_size); | |
1718 | req->async_data = kmalloc(def->async_size, GFP_KERNEL); | |
d886e185 PB |
1719 | if (req->async_data) { |
1720 | req->flags |= REQ_F_ASYNC_DATA; | |
1721 | return false; | |
1722 | } | |
1723 | return true; | |
3d9932a8 XW |
1724 | } |
1725 | ||
9cf7c104 PB |
1726 | static u32 io_get_sequence(struct io_kiocb *req) |
1727 | { | |
a3dbdf54 | 1728 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1729 | struct io_kiocb *cur; |
9cf7c104 | 1730 | |
a3dbdf54 | 1731 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1732 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1733 | seq--; |
1734 | return seq; | |
9cf7c104 PB |
1735 | } |
1736 | ||
c072481d | 1737 | static __cold void io_drain_req(struct io_kiocb *req) |
e276ae34 | 1738 | __must_hold(&ctx->uring_lock) |
de0617e4 | 1739 | { |
a197f664 | 1740 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1741 | struct io_defer_entry *de; |
f67676d1 | 1742 | int ret; |
e0eb71dc | 1743 | u32 seq = io_get_sequence(req); |
3c19966d | 1744 | |
9d858b21 | 1745 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1746 | spin_lock(&ctx->completion_lock); |
5e371265 | 1747 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1748 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1749 | queue: |
10c66904 | 1750 | ctx->drain_active = false; |
e0eb71dc PB |
1751 | io_req_task_queue(req); |
1752 | return; | |
10c66904 | 1753 | } |
e302f104 | 1754 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1755 | |
cbdcb435 | 1756 | io_prep_async_link(req); |
27dc8338 | 1757 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1758 | if (!de) { |
1b48773f | 1759 | ret = -ENOMEM; |
ef5c600a DY |
1760 | io_req_defer_failed(req, ret); |
1761 | return; | |
76cc33d7 | 1762 | } |
2d28390a | 1763 | |
79ebeaee | 1764 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1765 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1766 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1767 | kfree(de); |
e0eb71dc | 1768 | goto queue; |
de0617e4 JA |
1769 | } |
1770 | ||
48863ffd | 1771 | trace_io_uring_defer(req); |
27dc8338 | 1772 | de->req = req; |
9cf7c104 | 1773 | de->seq = seq; |
27dc8338 | 1774 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1775 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1776 | } |
1777 | ||
f4992544 JA |
1778 | static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def, |
1779 | unsigned int issue_flags) | |
6bf9c47a | 1780 | { |
f4992544 | 1781 | if (req->file || !def->needs_file) |
6bf9c47a JA |
1782 | return true; |
1783 | ||
1784 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1785 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1786 | else |
cef216fc | 1787 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1788 | |
772f5e00 | 1789 | return !!req->file; |
6bf9c47a JA |
1790 | } |
1791 | ||
889fca73 | 1792 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1793 | { |
a7dd2782 | 1794 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
5730b27e | 1795 | const struct cred *creds = NULL; |
d625c6ee | 1796 | int ret; |
2b188cc1 | 1797 | |
f4992544 | 1798 | if (unlikely(!io_assign_file(req, def, issue_flags))) |
70152140 JA |
1799 | return -EBADF; |
1800 | ||
6878b40e | 1801 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1802 | creds = override_creds(req->creds); |
5730b27e | 1803 | |
fcde59fe | 1804 | if (!def->audit_skip) |
5bd2182d PM |
1805 | audit_uring_entry(req->opcode); |
1806 | ||
0702e536 | 1807 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1808 | |
fcde59fe | 1809 | if (!def->audit_skip) |
5bd2182d PM |
1810 | audit_uring_exit(!ret, ret); |
1811 | ||
5730b27e JA |
1812 | if (creds) |
1813 | revert_creds(creds); | |
97b388d7 | 1814 | |
75d7b3ae PB |
1815 | if (ret == IOU_OK) { |
1816 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
9da070b1 | 1817 | io_req_complete_defer(req); |
75d7b3ae | 1818 | else |
1bec951c | 1819 | io_req_complete_post(req, issue_flags); |
97b388d7 | 1820 | |
9b43ef3d PB |
1821 | return 0; |
1822 | } | |
def596e9 | 1823 | |
e0b23d99 PB |
1824 | if (ret == IOU_ISSUE_SKIP_COMPLETE) { |
1825 | ret = 0; | |
1826 | io_arm_ltimeout(req); | |
def596e9 | 1827 | |
e0b23d99 PB |
1828 | /* If the op doesn't have a file, we're not polling for it */ |
1829 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue) | |
1830 | io_iopoll_req_issued(req, issue_flags); | |
1831 | } | |
1832 | return ret; | |
2b188cc1 JA |
1833 | } |
1834 | ||
a282967c | 1835 | int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts) |
329061d3 | 1836 | { |
a282967c | 1837 | io_tw_lock(req->ctx, ts); |
9a692451 DY |
1838 | return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT| |
1839 | IO_URING_F_COMPLETE_DEFER); | |
329061d3 JA |
1840 | } |
1841 | ||
c9f06aa7 | 1842 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1843 | { |
1844 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
247f97a5 | 1845 | struct io_kiocb *nxt = NULL; |
ebc11b6c | 1846 | |
247f97a5 PB |
1847 | if (req_ref_put_and_test(req)) { |
1848 | if (req->flags & IO_REQ_LINK_FLAGS) | |
1849 | nxt = io_req_find_next(req); | |
1850 | io_free_req(req); | |
1851 | } | |
1852 | return nxt ? &nxt->work : NULL; | |
ebc11b6c PB |
1853 | } |
1854 | ||
c9f06aa7 | 1855 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1856 | { |
1857 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
a7dd2782 | 1858 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
e6aeb272 | 1859 | unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ; |
d01905db | 1860 | bool needs_poll = false; |
6bf9c47a | 1861 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1862 | |
23a6c9ac | 1863 | /* one will be dropped by ->io_wq_free_work() after returning to io-wq */ |
48dcd38d PB |
1864 | if (!(req->flags & REQ_F_REFCOUNT)) |
1865 | __io_req_set_refcount(req, 2); | |
1866 | else | |
1867 | req_ref_get(req); | |
5d5901a3 | 1868 | |
cb2d344c | 1869 | io_arm_ltimeout(req); |
6bf9c47a | 1870 | |
dadebc35 | 1871 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1872 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1873 | fail: |
6bf9c47a | 1874 | io_req_task_queue_fail(req, err); |
d01905db PB |
1875 | return; |
1876 | } | |
f4992544 | 1877 | if (!io_assign_file(req, def, issue_flags)) { |
0f8da75b PB |
1878 | err = -EBADF; |
1879 | work->flags |= IO_WQ_WORK_CANCEL; | |
1880 | goto fail; | |
1881 | } | |
31b51510 | 1882 | |
e0e4ab52 PB |
1883 | /* |
1884 | * If DEFER_TASKRUN is set, it's only allowed to post CQEs from the | |
1885 | * submitter task context. Final request completions are handed to the | |
1886 | * right context, however this is not the case of auxiliary CQEs, | |
1887 | * which is the main mean of operation for multishot requests. | |
1888 | * Don't allow any multishot execution from io-wq. It's more restrictive | |
1889 | * than necessary and also cleaner. | |
1890 | */ | |
1891 | if (req->flags & REQ_F_APOLL_MULTISHOT) { | |
1892 | err = -EBADFD; | |
1893 | if (!io_file_can_poll(req)) | |
1894 | goto fail; | |
bee1d5be JA |
1895 | if (req->file->f_flags & O_NONBLOCK || |
1896 | req->file->f_mode & FMODE_NOWAIT) { | |
1897 | err = -ECANCELED; | |
1898 | if (io_arm_poll_handler(req, issue_flags) != IO_APOLL_OK) | |
1899 | goto fail; | |
1900 | return; | |
1901 | } else { | |
1902 | req->flags &= ~REQ_F_APOLL_MULTISHOT; | |
1903 | } | |
e0e4ab52 PB |
1904 | } |
1905 | ||
d01905db | 1906 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
1907 | bool opcode_poll = def->pollin || def->pollout; |
1908 | ||
95041b93 | 1909 | if (opcode_poll && io_file_can_poll(req)) { |
afb7f56f | 1910 | needs_poll = true; |
d01905db | 1911 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 1912 | } |
561fb04a | 1913 | } |
31b51510 | 1914 | |
d01905db PB |
1915 | do { |
1916 | ret = io_issue_sqe(req, issue_flags); | |
1917 | if (ret != -EAGAIN) | |
1918 | break; | |
a9be2022 JA |
1919 | |
1920 | /* | |
1921 | * If REQ_F_NOWAIT is set, then don't wait or retry with | |
1922 | * poll. -EAGAIN is final for that case. | |
1923 | */ | |
1924 | if (req->flags & REQ_F_NOWAIT) | |
1925 | break; | |
1926 | ||
d01905db PB |
1927 | /* |
1928 | * We can get EAGAIN for iopolled IO even though we're | |
1929 | * forcing a sync submission from here, since we can't | |
1930 | * wait for request slots on the block side. | |
1931 | */ | |
1932 | if (!needs_poll) { | |
e0deb6a0 PB |
1933 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
1934 | break; | |
45500dc4 PB |
1935 | if (io_wq_worker_stopped()) |
1936 | break; | |
d01905db PB |
1937 | cond_resched(); |
1938 | continue; | |
90fa0288 HX |
1939 | } |
1940 | ||
4d9237e3 | 1941 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
1942 | return; |
1943 | /* aborted or ready, in either case retry blocking */ | |
1944 | needs_poll = false; | |
1945 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
1946 | } while (1); | |
31b51510 | 1947 | |
a3df7698 | 1948 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 1949 | if (ret < 0) |
a3df7698 | 1950 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
1951 | } |
1952 | ||
531113bb JA |
1953 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
1954 | unsigned int issue_flags) | |
09bb8394 | 1955 | { |
5106dd6e | 1956 | struct io_ring_ctx *ctx = req->ctx; |
4bfb0c9a | 1957 | struct io_fixed_file *slot; |
5106dd6e | 1958 | struct file *file = NULL; |
09bb8394 | 1959 | |
93f052cb | 1960 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 1961 | |
ac177053 | 1962 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 1963 | goto out; |
ac177053 | 1964 | fd = array_index_nospec(fd, ctx->nr_user_files); |
4bfb0c9a | 1965 | slot = io_fixed_file_slot(&ctx->file_table, fd); |
3f302388 JA |
1966 | if (!req->rsrc_node) |
1967 | __io_req_set_rsrc_node(req, ctx); | |
4bfb0c9a | 1968 | req->flags |= io_slot_flags(slot); |
3f302388 | 1969 | file = io_slot_file(slot); |
5106dd6e | 1970 | out: |
93f052cb | 1971 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
1972 | return file; |
1973 | } | |
d44f554e | 1974 | |
531113bb | 1975 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 1976 | { |
62906e89 | 1977 | struct file *file = fget(fd); |
ac177053 | 1978 | |
48863ffd | 1979 | trace_io_uring_file_get(req, fd); |
09bb8394 | 1980 | |
ac177053 | 1981 | /* we don't allow fixed io_uring files */ |
e5550a14 | 1982 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 1983 | io_req_track_inflight(req); |
8371adf5 | 1984 | return file; |
09bb8394 JA |
1985 | } |
1986 | ||
7bfa9bad | 1987 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
1988 | __must_hold(&req->ctx->uring_lock) |
1989 | { | |
7bfa9bad PB |
1990 | struct io_kiocb *linked_timeout; |
1991 | ||
1992 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
973fc83f | 1993 | io_req_defer_failed(req, ret); |
7bfa9bad PB |
1994 | return; |
1995 | } | |
1996 | ||
1997 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 1998 | |
4d9237e3 | 1999 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 2000 | case IO_APOLL_READY: |
336d28a8 | 2001 | io_kbuf_recycle(req, 0); |
d475a9a6 PB |
2002 | io_req_task_queue(req); |
2003 | break; | |
2004 | case IO_APOLL_ABORTED: | |
6436c770 | 2005 | io_kbuf_recycle(req, 0); |
6e6b8c62 | 2006 | io_queue_iowq(req); |
d475a9a6 | 2007 | break; |
b1c62645 | 2008 | case IO_APOLL_OK: |
b1c62645 | 2009 | break; |
d475a9a6 PB |
2010 | } |
2011 | ||
2012 | if (linked_timeout) | |
2013 | io_queue_linked_timeout(linked_timeout); | |
2014 | } | |
2015 | ||
cbc2e203 | 2016 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 2017 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 2018 | { |
e0c5c576 | 2019 | int ret; |
2b188cc1 | 2020 | |
c5eef2b9 | 2021 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 2022 | |
491381ce JA |
2023 | /* |
2024 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2025 | * doesn't support non-blocking read/write attempts | |
2026 | */ | |
e0b23d99 | 2027 | if (unlikely(ret)) |
7bfa9bad | 2028 | io_queue_async(req, ret); |
2b188cc1 JA |
2029 | } |
2030 | ||
4652fe3f | 2031 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 2032 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 2033 | { |
17b147f6 PB |
2034 | if (unlikely(req->flags & REQ_F_FAIL)) { |
2035 | /* | |
2036 | * We don't submit, fail them all, for that replace hardlinks | |
2037 | * with normal links. Extra REQ_F_LINK is tolerated. | |
2038 | */ | |
2039 | req->flags &= ~REQ_F_HARDLINK; | |
2040 | req->flags |= REQ_F_LINK; | |
973fc83f | 2041 | io_req_defer_failed(req, req->cqe.res); |
76cc33d7 | 2042 | } else { |
ef5c600a DY |
2043 | if (unlikely(req->ctx->drain_active)) |
2044 | io_drain_req(req); | |
76cc33d7 | 2045 | else |
6e6b8c62 | 2046 | io_queue_iowq(req); |
ce35a47a | 2047 | } |
4fe2c963 JL |
2048 | } |
2049 | ||
b16fed66 PB |
2050 | /* |
2051 | * Check SQE restrictions (opcode and flags). | |
2052 | * | |
2053 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
2054 | */ | |
2055 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
2056 | struct io_kiocb *req, | |
2057 | unsigned int sqe_flags) | |
4fe2c963 | 2058 | { |
b16fed66 PB |
2059 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
2060 | return false; | |
2061 | ||
2062 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
2063 | ctx->restrictions.sqe_flags_required) | |
2064 | return false; | |
2065 | ||
2066 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
2067 | ctx->restrictions.sqe_flags_required)) | |
2068 | return false; | |
2069 | ||
2070 | return true; | |
4fe2c963 JL |
2071 | } |
2072 | ||
22b2ca31 PB |
2073 | static void io_init_req_drain(struct io_kiocb *req) |
2074 | { | |
2075 | struct io_ring_ctx *ctx = req->ctx; | |
2076 | struct io_kiocb *head = ctx->submit_state.link.head; | |
2077 | ||
2078 | ctx->drain_active = true; | |
2079 | if (head) { | |
2080 | /* | |
2081 | * If we need to drain a request in the middle of a link, drain | |
2082 | * the head request and the next request/link after the current | |
2083 | * link. Considering sequential execution of links, | |
b6c7db32 | 2084 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
2085 | * link. |
2086 | */ | |
b6c7db32 | 2087 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
2088 | ctx->drain_next = true; |
2089 | } | |
2090 | } | |
2091 | ||
e21e1c45 JA |
2092 | static __cold int io_init_fail_req(struct io_kiocb *req, int err) |
2093 | { | |
2094 | /* ensure per-opcode data is cleared if we fail before prep */ | |
2095 | memset(&req->cmd.data, 0, sizeof(req->cmd.data)); | |
2096 | return err; | |
2097 | } | |
2098 | ||
b16fed66 PB |
2099 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2100 | const struct io_uring_sqe *sqe) | |
282cdc86 | 2101 | __must_hold(&ctx->uring_lock) |
b16fed66 | 2102 | { |
a7dd2782 | 2103 | const struct io_issue_def *def; |
b16fed66 | 2104 | unsigned int sqe_flags; |
fc0ae024 | 2105 | int personality; |
4a04d1d1 | 2106 | u8 opcode; |
b16fed66 | 2107 | |
864ea921 | 2108 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 2109 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 | 2110 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
4bcb982c JA |
2111 | sqe_flags = READ_ONCE(sqe->flags); |
2112 | req->flags = (io_req_flags_t) sqe_flags; | |
cef216fc | 2113 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 2114 | req->file = NULL; |
c1bdf8ed | 2115 | req->rsrc_node = NULL; |
b16fed66 | 2116 | req->task = current; |
b16fed66 | 2117 | |
4a04d1d1 PB |
2118 | if (unlikely(opcode >= IORING_OP_LAST)) { |
2119 | req->opcode = 0; | |
e21e1c45 | 2120 | return io_init_fail_req(req, -EINVAL); |
4a04d1d1 | 2121 | } |
a7dd2782 | 2122 | def = &io_issue_defs[opcode]; |
68fe256a PB |
2123 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
2124 | /* enforce forwards compatibility on users */ | |
2125 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
e21e1c45 | 2126 | return io_init_fail_req(req, -EINVAL); |
4e906702 | 2127 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 2128 | if (!def->buffer_select) |
e21e1c45 | 2129 | return io_init_fail_req(req, -EOPNOTSUPP); |
4e906702 JA |
2130 | req->buf_index = READ_ONCE(sqe->buf_group); |
2131 | } | |
5562a8d7 PB |
2132 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
2133 | ctx->drain_disabled = true; | |
2134 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
2135 | if (ctx->drain_disabled) | |
e21e1c45 | 2136 | return io_init_fail_req(req, -EOPNOTSUPP); |
22b2ca31 | 2137 | io_init_req_drain(req); |
5562a8d7 | 2138 | } |
2a56a9bd PB |
2139 | } |
2140 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
2141 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
e21e1c45 | 2142 | return io_init_fail_req(req, -EACCES); |
2a56a9bd PB |
2143 | /* knock it to the slow queue path, will be drained there */ |
2144 | if (ctx->drain_active) | |
2145 | req->flags |= REQ_F_FORCE_ASYNC; | |
2146 | /* if there is no link, we're at "next" request and need to drain */ | |
2147 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
2148 | ctx->drain_next = false; | |
2149 | ctx->drain_active = true; | |
b6c7db32 | 2150 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 2151 | } |
68fe256a | 2152 | } |
b16fed66 | 2153 | |
fcde59fe | 2154 | if (!def->ioprio && sqe->ioprio) |
e21e1c45 | 2155 | return io_init_fail_req(req, -EINVAL); |
fcde59fe | 2156 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
e21e1c45 | 2157 | return io_init_fail_req(req, -EINVAL); |
73911426 | 2158 | |
fcde59fe | 2159 | if (def->needs_file) { |
6d63416d PB |
2160 | struct io_submit_state *state = &ctx->submit_state; |
2161 | ||
cef216fc | 2162 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 2163 | |
6d63416d PB |
2164 | /* |
2165 | * Plug now if we have more than 2 IO left after this, and the | |
2166 | * target is potentially a read/write to block based storage. | |
2167 | */ | |
fcde59fe | 2168 | if (state->need_plug && def->plug) { |
6d63416d PB |
2169 | state->plug_started = true; |
2170 | state->need_plug = false; | |
5ca7a8b3 | 2171 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 2172 | } |
b16fed66 | 2173 | } |
863e0560 | 2174 | |
003e8dcc JA |
2175 | personality = READ_ONCE(sqe->personality); |
2176 | if (personality) { | |
cdab10bf LT |
2177 | int ret; |
2178 | ||
c10d1f98 PB |
2179 | req->creds = xa_load(&ctx->personalities, personality); |
2180 | if (!req->creds) | |
e21e1c45 | 2181 | return io_init_fail_req(req, -EINVAL); |
c10d1f98 | 2182 | get_cred(req->creds); |
cdc1404a PM |
2183 | ret = security_uring_override_creds(req->creds); |
2184 | if (ret) { | |
2185 | put_cred(req->creds); | |
e21e1c45 | 2186 | return io_init_fail_req(req, ret); |
cdc1404a | 2187 | } |
b8e64b53 | 2188 | req->flags |= REQ_F_CREDS; |
003e8dcc | 2189 | } |
b16fed66 | 2190 | |
0702e536 | 2191 | return def->prep(req, sqe); |
b16fed66 PB |
2192 | } |
2193 | ||
df3becde PB |
2194 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
2195 | struct io_kiocb *req, int ret) | |
2196 | { | |
2197 | struct io_ring_ctx *ctx = req->ctx; | |
2198 | struct io_submit_link *link = &ctx->submit_state.link; | |
2199 | struct io_kiocb *head = link->head; | |
2200 | ||
48863ffd | 2201 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
2202 | |
2203 | /* | |
2204 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
2205 | * unusable. Instead of failing eagerly, continue assembling the link if | |
2206 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
2207 | * should find the flag and handle the rest. | |
2208 | */ | |
2209 | req_fail_link_node(req, ret); | |
2210 | if (head && !(head->flags & REQ_F_FAIL)) | |
2211 | req_fail_link_node(head, -ECANCELED); | |
2212 | ||
2213 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
2214 | if (head) { | |
2215 | link->last->link = req; | |
2216 | link->head = NULL; | |
2217 | req = head; | |
2218 | } | |
2219 | io_queue_sqe_fallback(req); | |
2220 | return ret; | |
2221 | } | |
2222 | ||
2223 | if (head) | |
2224 | link->last->link = req; | |
2225 | else | |
2226 | link->head = req; | |
2227 | link->last = req; | |
2228 | return 0; | |
2229 | } | |
2230 | ||
2231 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 2232 | const struct io_uring_sqe *sqe) |
282cdc86 | 2233 | __must_hold(&ctx->uring_lock) |
9e645e11 | 2234 | { |
a1ab7b35 | 2235 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 2236 | int ret; |
9e645e11 | 2237 | |
a6b8cadc | 2238 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
2239 | if (unlikely(ret)) |
2240 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 2241 | |
2ad57931 | 2242 | trace_io_uring_submit_req(req); |
a6b8cadc | 2243 | |
9e645e11 JA |
2244 | /* |
2245 | * If we already have a head request, queue this one for async | |
2246 | * submittal once the head completes. If we don't have a head but | |
2247 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2248 | * submitted sync once the chain is complete. If none of those | |
2249 | * conditions are true (normal request), then just queue it. | |
2250 | */ | |
924a07e4 | 2251 | if (unlikely(link->head)) { |
48863ffd | 2252 | trace_io_uring_link(req, link->head); |
f2f87370 | 2253 | link->last->link = req; |
863e0560 | 2254 | link->last = req; |
32fe525b | 2255 | |
da1a08c5 | 2256 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 2257 | return 0; |
df3becde PB |
2258 | /* last request of the link, flush it */ |
2259 | req = link->head; | |
f15a3431 | 2260 | link->head = NULL; |
924a07e4 PB |
2261 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
2262 | goto fallback; | |
2263 | ||
2264 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2265 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2266 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2267 | link->head = req; | |
2268 | link->last = req; | |
2269 | } else { | |
2270 | fallback: | |
2271 | io_queue_sqe_fallback(req); | |
2272 | } | |
f15a3431 | 2273 | return 0; |
9e645e11 | 2274 | } |
2e6e1fde | 2275 | |
f15a3431 | 2276 | io_queue_sqe(req); |
1d4240cc | 2277 | return 0; |
9e645e11 JA |
2278 | } |
2279 | ||
9a56a232 JA |
2280 | /* |
2281 | * Batched submission is done, ensure local IO is flushed out. | |
2282 | */ | |
553deffd | 2283 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2284 | { |
553deffd PB |
2285 | struct io_submit_state *state = &ctx->submit_state; |
2286 | ||
e126391c PB |
2287 | if (unlikely(state->link.head)) |
2288 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2289 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2290 | io_submit_flush_completions(ctx); |
27926b68 JA |
2291 | if (state->plug_started) |
2292 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2293 | } |
2294 | ||
2295 | /* | |
2296 | * Start submission side cache. | |
2297 | */ | |
2298 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2299 | unsigned int max_ios) |
9a56a232 | 2300 | { |
27926b68 | 2301 | state->plug_started = false; |
4b628aeb | 2302 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2303 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2304 | /* set only head, no need to init link_last in advance */ |
2305 | state->link.head = NULL; | |
9a56a232 JA |
2306 | } |
2307 | ||
2b188cc1 JA |
2308 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2309 | { | |
75b28aff | 2310 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2311 | |
caf582c6 PB |
2312 | /* |
2313 | * Ensure any loads from the SQEs are done at this point, | |
2314 | * since once we write the new head, the application could | |
2315 | * write new data to them. | |
2316 | */ | |
2317 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2318 | } |
2319 | ||
2b188cc1 | 2320 | /* |
dd9ae8a0 | 2321 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2322 | * that is mapped by userspace. This means that care needs to be taken to |
2323 | * ensure that reads are stable, as we cannot rely on userspace always | |
2324 | * being a good citizen. If members of the sqe are validated and then later | |
2325 | * used, it's important that those reads are done through READ_ONCE() to | |
2326 | * prevent a re-load down the line. | |
2327 | */ | |
b5083dfa | 2328 | static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe) |
2b188cc1 | 2329 | { |
2af89abd PB |
2330 | unsigned mask = ctx->sq_entries - 1; |
2331 | unsigned head = ctx->cached_sq_head++ & mask; | |
2332 | ||
2333 | if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) { | |
2334 | head = READ_ONCE(ctx->sq_array[head]); | |
2335 | if (unlikely(head >= ctx->sq_entries)) { | |
2336 | /* drop invalid entries */ | |
2337 | spin_lock(&ctx->completion_lock); | |
2338 | ctx->cq_extra--; | |
2339 | spin_unlock(&ctx->completion_lock); | |
2340 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2341 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
2342 | return false; | |
2343 | } | |
2344 | } | |
2b188cc1 JA |
2345 | |
2346 | /* | |
2347 | * The cached sq head (or cq tail) serves two purposes: | |
2348 | * | |
2349 | * 1) allows us to batch the cost of updating the user visible | |
2350 | * head updates. | |
2351 | * 2) allows the kernel side to track the head on its own, even | |
2352 | * though the application is the one updating it. | |
2353 | */ | |
2b188cc1 | 2354 | |
2af89abd PB |
2355 | /* double index for 128-byte SQEs, twice as long */ |
2356 | if (ctx->flags & IORING_SETUP_SQE128) | |
2357 | head <<= 1; | |
2358 | *sqe = &ctx->sq_sqes[head]; | |
2359 | return true; | |
709b302f PB |
2360 | } |
2361 | ||
17437f31 | 2362 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2363 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2364 | { |
69629809 | 2365 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2366 | unsigned int left; |
2367 | int ret; | |
6c271ce2 | 2368 | |
51d48dab | 2369 | if (unlikely(!entries)) |
69629809 | 2370 | return 0; |
ee7d46d9 | 2371 | /* make sure SQ entry isn't read before tail */ |
e3ef728f | 2372 | ret = left = min(nr, entries); |
8e6971a8 PB |
2373 | io_get_task_refs(left); |
2374 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2375 | |
69629809 | 2376 | do { |
3529d8c2 | 2377 | const struct io_uring_sqe *sqe; |
196be95c | 2378 | struct io_kiocb *req; |
fb5ccc98 | 2379 | |
c8576f3e | 2380 | if (unlikely(!io_alloc_req(ctx, &req))) |
fb5ccc98 | 2381 | break; |
b5083dfa | 2382 | if (unlikely(!io_get_sqe(ctx, &sqe))) { |
fa05457a | 2383 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2384 | break; |
2385 | } | |
6c271ce2 | 2386 | |
1cd15904 PB |
2387 | /* |
2388 | * Continue submitting even for sqe failure if the | |
2389 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2390 | */ | |
2391 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2392 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2393 | left--; | |
2394 | break; | |
bcbb7bf6 | 2395 | } |
1cd15904 | 2396 | } while (--left); |
9466f437 | 2397 | |
8e6971a8 PB |
2398 | if (unlikely(left)) { |
2399 | ret -= left; | |
2400 | /* try again if it submitted nothing and can't allocate a req */ | |
2401 | if (!ret && io_req_cache_empty(ctx)) | |
2402 | ret = -EAGAIN; | |
2403 | current->io_uring->cached_refs += left; | |
9466f437 | 2404 | } |
6c271ce2 | 2405 | |
553deffd | 2406 | io_submit_state_end(ctx); |
ae9428ca PB |
2407 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2408 | io_commit_sqring(ctx); | |
8e6971a8 | 2409 | return ret; |
6c271ce2 JA |
2410 | } |
2411 | ||
bda52162 JA |
2412 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, |
2413 | int wake_flags, void *key) | |
2414 | { | |
bd550173 | 2415 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq); |
bda52162 | 2416 | |
6c503150 PB |
2417 | /* |
2418 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2419 | * the task, and the next invocation will do it. | |
2420 | */ | |
bd550173 | 2421 | if (io_should_wake(iowq) || io_has_work(iowq->ctx)) |
6c503150 PB |
2422 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2423 | return -1; | |
bda52162 JA |
2424 | } |
2425 | ||
c0e0d6ba | 2426 | int io_run_task_work_sig(struct io_ring_ctx *ctx) |
af9c1a44 | 2427 | { |
1414d629 | 2428 | if (!llist_empty(&ctx->work_llist)) { |
2f413956 | 2429 | __set_current_state(TASK_RUNNING); |
9fe3eaea | 2430 | if (io_run_local_work(ctx, INT_MAX) > 0) |
d246c759 | 2431 | return 0; |
1414d629 PB |
2432 | } |
2433 | if (io_run_task_work() > 0) | |
d246c759 | 2434 | return 0; |
c5020bc8 OL |
2435 | if (task_sigpending(current)) |
2436 | return -EINTR; | |
2437 | return 0; | |
af9c1a44 JA |
2438 | } |
2439 | ||
7b72d661 JA |
2440 | static bool current_pending_io(void) |
2441 | { | |
2442 | struct io_uring_task *tctx = current->io_uring; | |
2443 | ||
2444 | if (!tctx) | |
2445 | return false; | |
2446 | return percpu_counter_read_positive(&tctx->inflight); | |
2447 | } | |
2448 | ||
eeb60b9a PB |
2449 | /* when returns >0, the caller should retry */ |
2450 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
d33a39e5 | 2451 | struct io_wait_queue *iowq) |
eeb60b9a | 2452 | { |
6f0974ec | 2453 | int ret; |
8a796565 | 2454 | |
3fcf19d5 PB |
2455 | if (unlikely(READ_ONCE(ctx->check_cq))) |
2456 | return 1; | |
846072f1 PB |
2457 | if (unlikely(!llist_empty(&ctx->work_llist))) |
2458 | return 1; | |
2459 | if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) | |
2460 | return 1; | |
2461 | if (unlikely(task_sigpending(current))) | |
2462 | return -EINTR; | |
2463 | if (unlikely(io_should_wake(iowq))) | |
2464 | return 0; | |
8a796565 AF |
2465 | |
2466 | /* | |
7b72d661 JA |
2467 | * Mark us as being in io_wait if we have pending requests, so cpufreq |
2468 | * can take into account that the task is waiting for IO - turns out | |
2469 | * to be important for low QD IO. | |
8a796565 | 2470 | */ |
7b72d661 JA |
2471 | if (current_pending_io()) |
2472 | current->in_iowait = 1; | |
8a796565 | 2473 | ret = 0; |
d33a39e5 | 2474 | if (iowq->timeout == KTIME_MAX) |
46ae7eef | 2475 | schedule(); |
d33a39e5 | 2476 | else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS)) |
8a796565 | 2477 | ret = -ETIME; |
6f0974ec | 2478 | current->in_iowait = 0; |
8a796565 | 2479 | return ret; |
eeb60b9a PB |
2480 | } |
2481 | ||
2b188cc1 JA |
2482 | /* |
2483 | * Wait until events become available, if we don't already have some. The | |
2484 | * application must reap them itself, as they reside on the shared cq ring. | |
2485 | */ | |
2486 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2487 | const sigset_t __user *sig, size_t sigsz, |
2488 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2489 | { |
90291099 | 2490 | struct io_wait_queue iowq; |
75b28aff | 2491 | struct io_rings *rings = ctx->rings; |
c1d5a224 | 2492 | int ret; |
2b188cc1 | 2493 | |
76de6749 PB |
2494 | if (!io_allowed_run_tw(ctx)) |
2495 | return -EEXIST; | |
140102ae | 2496 | if (!llist_empty(&ctx->work_llist)) |
9fe3eaea | 2497 | io_run_local_work(ctx, min_events); |
f36ba6cf PB |
2498 | io_run_task_work(); |
2499 | io_cqring_overflow_flush(ctx); | |
2500 | /* if user messes with these they will just get an early return */ | |
2501 | if (__io_cqring_events_user(ctx) >= min_events) | |
2502 | return 0; | |
2b188cc1 | 2503 | |
90291099 PB |
2504 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
2505 | iowq.wq.private = current; | |
2506 | INIT_LIST_HEAD(&iowq.wq.entry); | |
2507 | iowq.ctx = ctx; | |
bda52162 | 2508 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 2509 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
d33a39e5 PB |
2510 | iowq.timeout = KTIME_MAX; |
2511 | ||
2512 | if (uts) { | |
2513 | struct timespec64 ts; | |
2514 | ||
2515 | if (get_timespec64(&ts, uts)) | |
2516 | return -EFAULT; | |
8d0c12a8 | 2517 | |
d33a39e5 | 2518 | iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); |
8d0c12a8 | 2519 | io_napi_adjust_timeout(ctx, &iowq, &ts); |
d33a39e5 | 2520 | } |
90291099 | 2521 | |
978e5c19 AI |
2522 | if (sig) { |
2523 | #ifdef CONFIG_COMPAT | |
2524 | if (in_compat_syscall()) | |
2525 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
2526 | sigsz); | |
2527 | else | |
2528 | #endif | |
2529 | ret = set_user_sigmask(sig, sigsz); | |
2530 | ||
2531 | if (ret) | |
2532 | return ret; | |
2533 | } | |
2534 | ||
8d0c12a8 SR |
2535 | io_napi_busy_loop(ctx, &iowq); |
2536 | ||
c826bd7a | 2537 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2538 | do { |
9fe3eaea | 2539 | int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail); |
3fcf19d5 PB |
2540 | unsigned long check_cq; |
2541 | ||
130bd686 | 2542 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
8751d154 | 2543 | atomic_set(&ctx->cq_wait_nr, nr_wait); |
130bd686 PB |
2544 | set_current_state(TASK_INTERRUPTIBLE); |
2545 | } else { | |
2546 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, | |
2547 | TASK_INTERRUPTIBLE); | |
2548 | } | |
2549 | ||
d33a39e5 | 2550 | ret = io_cqring_wait_schedule(ctx, &iowq); |
130bd686 | 2551 | __set_current_state(TASK_RUNNING); |
b4bc35cf | 2552 | atomic_set(&ctx->cq_wait_nr, IO_CQ_WAKE_INIT); |
d80c0f00 | 2553 | |
846072f1 PB |
2554 | /* |
2555 | * Run task_work after scheduling and before io_should_wake(). | |
2556 | * If we got woken because of task_work being processed, run it | |
2557 | * now rather than let the caller do another wait loop. | |
2558 | */ | |
2559 | io_run_task_work(); | |
2560 | if (!llist_empty(&ctx->work_llist)) | |
9fe3eaea | 2561 | io_run_local_work(ctx, nr_wait); |
3fcf19d5 | 2562 | |
6ff1407e JA |
2563 | /* |
2564 | * Non-local task_work will be run on exit to userspace, but | |
2565 | * if we're using DEFER_TASKRUN, then we could have waited | |
2566 | * with a timeout for a number of requests. If the timeout | |
2567 | * hits, we could have some requests ready to process. Ensure | |
2568 | * this break is _after_ we have run task_work, to avoid | |
2569 | * deferring running potentially pending requests until the | |
2570 | * next time we wait for events. | |
2571 | */ | |
2572 | if (ret < 0) | |
2573 | break; | |
2574 | ||
3fcf19d5 PB |
2575 | check_cq = READ_ONCE(ctx->check_cq); |
2576 | if (unlikely(check_cq)) { | |
2577 | /* let the caller flush overflows, retry */ | |
326a9e48 | 2578 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) |
3fcf19d5 | 2579 | io_cqring_do_overflow_flush(ctx); |
3fcf19d5 PB |
2580 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) { |
2581 | ret = -EBADR; | |
2582 | break; | |
2583 | } | |
2584 | } | |
2585 | ||
846072f1 PB |
2586 | if (io_should_wake(&iowq)) { |
2587 | ret = 0; | |
35d90f95 | 2588 | break; |
846072f1 | 2589 | } |
ca0a2651 | 2590 | cond_resched(); |
846072f1 | 2591 | } while (1); |
bda52162 | 2592 | |
130bd686 PB |
2593 | if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) |
2594 | finish_wait(&ctx->cq_wait, &iowq.wq); | |
b7db41c9 | 2595 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2596 | |
75b28aff | 2597 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2598 | } |
2599 | ||
87585b05 JA |
2600 | void io_pages_unmap(void *ptr, struct page ***pages, unsigned short *npages, |
2601 | bool put_pages) | |
3ab1db3c JA |
2602 | { |
2603 | bool do_vunmap = false; | |
2604 | ||
2605 | if (!ptr) | |
2606 | return; | |
2607 | ||
87585b05 | 2608 | if (put_pages && *npages) { |
3ab1db3c JA |
2609 | struct page **to_free = *pages; |
2610 | int i; | |
2611 | ||
2612 | /* | |
2613 | * Only did vmap for the non-compound multiple page case. | |
2614 | * For the compound page, we just need to put the head. | |
2615 | */ | |
2616 | if (PageCompound(to_free[0])) | |
2617 | *npages = 1; | |
2618 | else if (*npages > 1) | |
2619 | do_vunmap = true; | |
2620 | for (i = 0; i < *npages; i++) | |
2621 | put_page(to_free[i]); | |
2622 | } | |
2623 | if (do_vunmap) | |
2624 | vunmap(ptr); | |
2625 | kvfree(*pages); | |
2626 | *pages = NULL; | |
2627 | *npages = 0; | |
2628 | } | |
2629 | ||
03d89a2d JA |
2630 | static void io_pages_free(struct page ***pages, int npages) |
2631 | { | |
2c5c0ba1 | 2632 | struct page **page_array = *pages; |
03d89a2d | 2633 | |
8b51a395 JA |
2634 | if (!page_array) |
2635 | return; | |
2636 | ||
18595c0a | 2637 | unpin_user_pages(page_array, npages); |
03d89a2d JA |
2638 | kvfree(page_array); |
2639 | *pages = NULL; | |
2640 | } | |
2641 | ||
1943f96b JA |
2642 | struct page **io_pin_pages(unsigned long uaddr, unsigned long len, int *npages) |
2643 | { | |
2644 | unsigned long start, end, nr_pages; | |
2645 | struct page **pages; | |
2646 | int ret; | |
2647 | ||
2648 | end = (uaddr + len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2649 | start = uaddr >> PAGE_SHIFT; | |
2650 | nr_pages = end - start; | |
2651 | if (WARN_ON_ONCE(!nr_pages)) | |
2652 | return ERR_PTR(-EINVAL); | |
2653 | ||
2654 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); | |
2655 | if (!pages) | |
2656 | return ERR_PTR(-ENOMEM); | |
2657 | ||
2658 | ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM, | |
2659 | pages); | |
2660 | /* success, mapped all pages */ | |
2661 | if (ret == nr_pages) { | |
2662 | *npages = nr_pages; | |
2663 | return pages; | |
2664 | } | |
2665 | ||
2666 | /* partial map, or didn't map anything */ | |
2667 | if (ret >= 0) { | |
2668 | /* if we did partial map, release any pages we did get */ | |
2669 | if (ret) | |
2670 | unpin_user_pages(pages, ret); | |
2671 | ret = -EFAULT; | |
2672 | } | |
2673 | kvfree(pages); | |
2674 | return ERR_PTR(ret); | |
2675 | } | |
2676 | ||
03d89a2d JA |
2677 | static void *__io_uaddr_map(struct page ***pages, unsigned short *npages, |
2678 | unsigned long uaddr, size_t size) | |
2679 | { | |
2680 | struct page **page_array; | |
2681 | unsigned int nr_pages; | |
820d070f | 2682 | void *page_addr; |
03d89a2d JA |
2683 | |
2684 | *npages = 0; | |
2685 | ||
2686 | if (uaddr & (PAGE_SIZE - 1) || !size) | |
2687 | return ERR_PTR(-EINVAL); | |
2688 | ||
1943f96b JA |
2689 | nr_pages = 0; |
2690 | page_array = io_pin_pages(uaddr, size, &nr_pages); | |
2691 | if (IS_ERR(page_array)) | |
2692 | return page_array; | |
223ef474 | 2693 | |
09fc75e0 JA |
2694 | page_addr = vmap(page_array, nr_pages, VM_MAP, PAGE_KERNEL); |
2695 | if (page_addr) { | |
2696 | *pages = page_array; | |
2697 | *npages = nr_pages; | |
2698 | return page_addr; | |
223ef474 | 2699 | } |
1943f96b JA |
2700 | |
2701 | io_pages_free(&page_array, nr_pages); | |
2702 | return ERR_PTR(-ENOMEM); | |
03d89a2d JA |
2703 | } |
2704 | ||
2705 | static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr, | |
2706 | size_t size) | |
2707 | { | |
2708 | return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr, | |
2709 | size); | |
2710 | } | |
2711 | ||
2712 | static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr, | |
2713 | size_t size) | |
2714 | { | |
2715 | return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr, | |
2716 | size); | |
2717 | } | |
2718 | ||
9c189eee JA |
2719 | static void io_rings_free(struct io_ring_ctx *ctx) |
2720 | { | |
03d89a2d | 2721 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) { |
87585b05 JA |
2722 | io_pages_unmap(ctx->rings, &ctx->ring_pages, &ctx->n_ring_pages, |
2723 | true); | |
2724 | io_pages_unmap(ctx->sq_sqes, &ctx->sqe_pages, &ctx->n_sqe_pages, | |
2725 | true); | |
03d89a2d JA |
2726 | } else { |
2727 | io_pages_free(&ctx->ring_pages, ctx->n_ring_pages); | |
8b51a395 | 2728 | ctx->n_ring_pages = 0; |
03d89a2d | 2729 | io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages); |
8b51a395 | 2730 | ctx->n_sqe_pages = 0; |
09fc75e0 JA |
2731 | vunmap(ctx->rings); |
2732 | vunmap(ctx->sq_sqes); | |
03d89a2d | 2733 | } |
cef59d1e PB |
2734 | |
2735 | ctx->rings = NULL; | |
2736 | ctx->sq_sqes = NULL; | |
9c189eee JA |
2737 | } |
2738 | ||
3ab1db3c JA |
2739 | static void *io_mem_alloc_compound(struct page **pages, int nr_pages, |
2740 | size_t size, gfp_t gfp) | |
2741 | { | |
2742 | struct page *page; | |
2743 | int i, order; | |
2744 | ||
2745 | order = get_order(size); | |
2746 | if (order > MAX_PAGE_ORDER) | |
2747 | return ERR_PTR(-ENOMEM); | |
2748 | else if (order) | |
2749 | gfp |= __GFP_COMP; | |
2750 | ||
2751 | page = alloc_pages(gfp, order); | |
2752 | if (!page) | |
2753 | return ERR_PTR(-ENOMEM); | |
2754 | ||
2755 | for (i = 0; i < nr_pages; i++) | |
2756 | pages[i] = page + i; | |
2757 | ||
2758 | return page_address(page); | |
2759 | } | |
2760 | ||
2761 | static void *io_mem_alloc_single(struct page **pages, int nr_pages, size_t size, | |
2762 | gfp_t gfp) | |
2763 | { | |
2764 | void *ret; | |
2765 | int i; | |
2766 | ||
2767 | for (i = 0; i < nr_pages; i++) { | |
2768 | pages[i] = alloc_page(gfp); | |
2769 | if (!pages[i]) | |
2770 | goto err; | |
2771 | } | |
2772 | ||
2773 | ret = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL); | |
2774 | if (ret) | |
2775 | return ret; | |
2776 | err: | |
2777 | while (i--) | |
2778 | put_page(pages[i]); | |
2779 | return ERR_PTR(-ENOMEM); | |
2780 | } | |
2781 | ||
87585b05 JA |
2782 | void *io_pages_map(struct page ***out_pages, unsigned short *npages, |
2783 | size_t size) | |
3ab1db3c JA |
2784 | { |
2785 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN; | |
2786 | struct page **pages; | |
2787 | int nr_pages; | |
2788 | void *ret; | |
2789 | ||
2790 | nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2791 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), gfp); | |
2792 | if (!pages) | |
2793 | return ERR_PTR(-ENOMEM); | |
2794 | ||
2795 | ret = io_mem_alloc_compound(pages, nr_pages, size, gfp); | |
2796 | if (!IS_ERR(ret)) | |
2797 | goto done; | |
2798 | ||
2799 | ret = io_mem_alloc_single(pages, nr_pages, size, gfp); | |
2800 | if (!IS_ERR(ret)) { | |
2801 | done: | |
2802 | *out_pages = pages; | |
2803 | *npages = nr_pages; | |
2804 | return ret; | |
2805 | } | |
2806 | ||
2807 | kvfree(pages); | |
2808 | *out_pages = NULL; | |
2809 | *npages = 0; | |
2810 | return ret; | |
2811 | } | |
2812 | ||
73572984 JA |
2813 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2814 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2815 | { |
73572984 JA |
2816 | struct io_rings *rings; |
2817 | size_t off, sq_array_size; | |
6b06314c | 2818 | |
73572984 JA |
2819 | off = struct_size(rings, cqes, cq_entries); |
2820 | if (off == SIZE_MAX) | |
2821 | return SIZE_MAX; | |
2822 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2823 | if (check_shl_overflow(off, 1, &off)) | |
2824 | return SIZE_MAX; | |
2825 | } | |
ab409402 | 2826 | |
73572984 JA |
2827 | #ifdef CONFIG_SMP |
2828 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2829 | if (off == 0) | |
2830 | return SIZE_MAX; | |
2831 | #endif | |
82fbcfa9 | 2832 | |
2af89abd PB |
2833 | if (ctx->flags & IORING_SETUP_NO_SQARRAY) { |
2834 | if (sq_offset) | |
2835 | *sq_offset = SIZE_MAX; | |
2836 | return off; | |
2837 | } | |
2838 | ||
73572984 JA |
2839 | if (sq_offset) |
2840 | *sq_offset = off; | |
82fbcfa9 | 2841 | |
73572984 JA |
2842 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2843 | if (sq_array_size == SIZE_MAX) | |
2844 | return SIZE_MAX; | |
6b06314c | 2845 | |
73572984 JA |
2846 | if (check_add_overflow(off, sq_array_size, &off)) |
2847 | return SIZE_MAX; | |
8bad28d8 | 2848 | |
73572984 | 2849 | return off; |
8bad28d8 HX |
2850 | } |
2851 | ||
73572984 | 2852 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2853 | { |
c8576f3e | 2854 | struct io_kiocb *req; |
37f0e767 | 2855 | int nr = 0; |
bf019da7 | 2856 | |
9a4fdbd8 | 2857 | mutex_lock(&ctx->uring_lock); |
34f0bc42 | 2858 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
9a4fdbd8 | 2859 | |
88ab95be | 2860 | while (!io_req_cache_empty(ctx)) { |
c8576f3e | 2861 | req = io_extract_req(ctx); |
c2b6c6bc | 2862 | kmem_cache_free(req_cachep, req); |
37f0e767 | 2863 | nr++; |
c2b6c6bc | 2864 | } |
37f0e767 PB |
2865 | if (nr) |
2866 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2867 | mutex_unlock(&ctx->uring_lock); |
2868 | } | |
2869 | ||
c072481d | 2870 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2871 | { |
37d1e2e3 | 2872 | io_sq_thread_finish(ctx); |
43597aac | 2873 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
0b222eeb PB |
2874 | if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list))) |
2875 | return; | |
43597aac | 2876 | |
8bad28d8 | 2877 | mutex_lock(&ctx->uring_lock); |
43597aac | 2878 | if (ctx->buf_data) |
bd54b6fe | 2879 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2880 | if (ctx->file_data) |
08480400 | 2881 | __io_sqe_files_unregister(ctx); |
a85381d8 | 2882 | io_cqring_overflow_kill(ctx); |
9b402849 | 2883 | io_eventfd_unregister(ctx); |
414d0f45 | 2884 | io_alloc_cache_free(&ctx->apoll_cache, kfree); |
43e0bbbd | 2885 | io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); |
a9165b83 | 2886 | io_alloc_cache_free(&ctx->rw_cache, io_rw_cache_free); |
414d0f45 | 2887 | io_alloc_cache_free(&ctx->uring_cache, kfree); |
194bb58c | 2888 | io_futex_cache_free(ctx); |
5a2e745d | 2889 | io_destroy_buffers(ctx); |
b4a72c05 | 2890 | mutex_unlock(&ctx->uring_lock); |
07db298a PB |
2891 | if (ctx->sq_creds) |
2892 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2893 | if (ctx->submitter_task) |
2894 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2895 | |
a7f0ed5a PB |
2896 | /* there are no registered resources left, nobody uses it */ |
2897 | if (ctx->rsrc_node) | |
9eae8655 | 2898 | io_rsrc_node_destroy(ctx, ctx->rsrc_node); |
a7f0ed5a PB |
2899 | |
2900 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
ef9dd637 | 2901 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
2b188cc1 | 2902 | |
414d0f45 | 2903 | io_alloc_cache_free(&ctx->rsrc_node_cache, kfree); |
42b6419d PB |
2904 | if (ctx->mm_account) { |
2905 | mmdrop(ctx->mm_account); | |
2906 | ctx->mm_account = NULL; | |
2907 | } | |
9c189eee | 2908 | io_rings_free(ctx); |
2b188cc1 JA |
2909 | |
2910 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2911 | free_uid(ctx->user); |
4010fec4 | 2912 | io_req_caches_free(ctx); |
e941894e JA |
2913 | if (ctx->hash_map) |
2914 | io_wq_put_hash(ctx->hash_map); | |
8d0c12a8 | 2915 | io_napi_free(ctx); |
e6f89be6 | 2916 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2917 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 | 2918 | xa_destroy(&ctx->io_bl_xa); |
2b188cc1 JA |
2919 | kfree(ctx); |
2920 | } | |
2921 | ||
bca39f39 PB |
2922 | static __cold void io_activate_pollwq_cb(struct callback_head *cb) |
2923 | { | |
2924 | struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx, | |
2925 | poll_wq_task_work); | |
2926 | ||
2927 | mutex_lock(&ctx->uring_lock); | |
2928 | ctx->poll_activated = true; | |
2929 | mutex_unlock(&ctx->uring_lock); | |
2930 | ||
2931 | /* | |
2932 | * Wake ups for some events between start of polling and activation | |
2933 | * might've been lost due to loose synchronisation. | |
2934 | */ | |
2935 | wake_up_all(&ctx->poll_wq); | |
2936 | percpu_ref_put(&ctx->refs); | |
2937 | } | |
2938 | ||
c4320315 | 2939 | __cold void io_activate_pollwq(struct io_ring_ctx *ctx) |
bca39f39 PB |
2940 | { |
2941 | spin_lock(&ctx->completion_lock); | |
2942 | /* already activated or in progress */ | |
2943 | if (ctx->poll_activated || ctx->poll_wq_task_work.func) | |
2944 | goto out; | |
2945 | if (WARN_ON_ONCE(!ctx->task_complete)) | |
2946 | goto out; | |
2947 | if (!ctx->submitter_task) | |
2948 | goto out; | |
2949 | /* | |
2950 | * with ->submitter_task only the submitter task completes requests, we | |
2951 | * only need to sync with it, which is done by injecting a tw | |
2952 | */ | |
2953 | init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb); | |
2954 | percpu_ref_get(&ctx->refs); | |
2955 | if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL)) | |
2956 | percpu_ref_put(&ctx->refs); | |
2957 | out: | |
2958 | spin_unlock(&ctx->completion_lock); | |
2959 | } | |
2960 | ||
2b188cc1 JA |
2961 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) |
2962 | { | |
2963 | struct io_ring_ctx *ctx = file->private_data; | |
2964 | __poll_t mask = 0; | |
2965 | ||
bca39f39 PB |
2966 | if (unlikely(!ctx->poll_activated)) |
2967 | io_activate_pollwq(ctx); | |
2968 | ||
7b235dd8 | 2969 | poll_wait(file, &ctx->poll_wq, wait); |
4f7067c3 SB |
2970 | /* |
2971 | * synchronizes with barrier from wq_has_sleeper call in | |
2972 | * io_commit_cqring | |
2973 | */ | |
2b188cc1 | 2974 | smp_rmb(); |
90554200 | 2975 | if (!io_sqring_full(ctx)) |
2b188cc1 | 2976 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
2977 | |
2978 | /* | |
2979 | * Don't flush cqring overflow list here, just do a simple check. | |
2980 | * Otherwise there could possible be ABBA deadlock: | |
2981 | * CPU0 CPU1 | |
2982 | * ---- ---- | |
2983 | * lock(&ctx->uring_lock); | |
2984 | * lock(&ep->mtx); | |
2985 | * lock(&ctx->uring_lock); | |
2986 | * lock(&ep->mtx); | |
2987 | * | |
2988 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
10d8bc35 | 2989 | * pushes them to do the flush. |
ed670c3f | 2990 | */ |
b4c98d59 | 2991 | |
c10bb646 | 2992 | if (__io_cqring_events_user(ctx) || io_has_work(ctx)) |
2b188cc1 JA |
2993 | mask |= EPOLLIN | EPOLLRDNORM; |
2994 | ||
2995 | return mask; | |
2996 | } | |
2997 | ||
d56d938b PB |
2998 | struct io_tctx_exit { |
2999 | struct callback_head task_work; | |
3000 | struct completion completion; | |
baf186c4 | 3001 | struct io_ring_ctx *ctx; |
d56d938b PB |
3002 | }; |
3003 | ||
c072481d | 3004 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
3005 | { |
3006 | struct io_uring_task *tctx = current->io_uring; | |
3007 | struct io_tctx_exit *work; | |
3008 | ||
3009 | work = container_of(cb, struct io_tctx_exit, task_work); | |
3010 | /* | |
8d664282 | 3011 | * When @in_cancel, we're in cancellation and it's racy to remove the |
d56d938b | 3012 | * node. It'll be removed by the end of cancellation, just ignore it. |
998b30c3 HM |
3013 | * tctx can be NULL if the queueing of this task_work raced with |
3014 | * work cancelation off the exec path. | |
d56d938b | 3015 | */ |
8d664282 | 3016 | if (tctx && !atomic_read(&tctx->in_cancel)) |
eef51daa | 3017 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
3018 | complete(&work->completion); |
3019 | } | |
3020 | ||
c072481d | 3021 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
3022 | { |
3023 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
3024 | ||
3025 | return req->ctx == data; | |
3026 | } | |
3027 | ||
c072481d | 3028 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 3029 | { |
d56d938b | 3030 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 3031 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 3032 | unsigned long interval = HZ / 20; |
d56d938b PB |
3033 | struct io_tctx_exit exit; |
3034 | struct io_tctx_node *node; | |
3035 | int ret; | |
85faa7b8 | 3036 | |
56952e91 JA |
3037 | /* |
3038 | * If we're doing polled IO and end up having requests being | |
3039 | * submitted async (out-of-line), then completions can come in while | |
3040 | * we're waiting for refs to drop. We need to reap these manually, | |
3041 | * as nobody else will be looking for them. | |
3042 | */ | |
b2edc0a7 | 3043 | do { |
a85381d8 PB |
3044 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
3045 | mutex_lock(&ctx->uring_lock); | |
3046 | io_cqring_overflow_kill(ctx); | |
3047 | mutex_unlock(&ctx->uring_lock); | |
3048 | } | |
3049 | ||
c0e0d6ba DY |
3050 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) |
3051 | io_move_task_work_from_local(ctx); | |
3052 | ||
affa87db PB |
3053 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
3054 | cond_resched(); | |
3055 | ||
28090c13 PB |
3056 | if (ctx->sq_data) { |
3057 | struct io_sq_data *sqd = ctx->sq_data; | |
3058 | struct task_struct *tsk; | |
3059 | ||
3060 | io_sq_thread_park(sqd); | |
3061 | tsk = sqd->thread; | |
3062 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
3063 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
3064 | io_cancel_ctx_cb, ctx, true); | |
3065 | io_sq_thread_unpark(sqd); | |
3066 | } | |
b5bb3a24 | 3067 | |
37f0e767 PB |
3068 | io_req_caches_free(ctx); |
3069 | ||
58d3be2c PB |
3070 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
3071 | /* there is little hope left, don't run it too often */ | |
3072 | interval = HZ * 60; | |
3073 | } | |
4826c594 JA |
3074 | /* |
3075 | * This is really an uninterruptible wait, as it has to be | |
3076 | * complete. But it's also run from a kworker, which doesn't | |
3077 | * take signals, so it's fine to make it interruptible. This | |
3078 | * avoids scenarios where we knowingly can wait much longer | |
3079 | * on completions, for example if someone does a SIGSTOP on | |
3080 | * a task that needs to finish task_work to make this loop | |
3081 | * complete. That's a synthetic situation that should not | |
3082 | * cause a stuck task backtrace, and hence a potential panic | |
3083 | * on stuck tasks if that is enabled. | |
3084 | */ | |
3085 | } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 3086 | |
7f00651a PB |
3087 | init_completion(&exit.completion); |
3088 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
3089 | exit.ctx = ctx; | |
f7b32e78 | 3090 | |
d56d938b PB |
3091 | mutex_lock(&ctx->uring_lock); |
3092 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
3093 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
3094 | ||
d56d938b PB |
3095 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
3096 | ctx_node); | |
7f00651a PB |
3097 | /* don't spin on a single task if cancellation failed */ |
3098 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
3099 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
3100 | if (WARN_ON_ONCE(ret)) | |
3101 | continue; | |
d56d938b PB |
3102 | |
3103 | mutex_unlock(&ctx->uring_lock); | |
4826c594 JA |
3104 | /* |
3105 | * See comment above for | |
3106 | * wait_for_completion_interruptible_timeout() on why this | |
3107 | * wait is marked as interruptible. | |
3108 | */ | |
3109 | wait_for_completion_interruptible(&exit.completion); | |
d56d938b PB |
3110 | mutex_lock(&ctx->uring_lock); |
3111 | } | |
3112 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
3113 | spin_lock(&ctx->completion_lock); |
3114 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 3115 | |
d73a572d PB |
3116 | /* pairs with RCU read section in io_req_local_work_add() */ |
3117 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3118 | synchronize_rcu(); | |
3119 | ||
85faa7b8 JA |
3120 | io_ring_ctx_free(ctx); |
3121 | } | |
3122 | ||
c072481d | 3123 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 3124 | { |
61cf9370 MWO |
3125 | unsigned long index; |
3126 | struct creds *creds; | |
3127 | ||
2b188cc1 JA |
3128 | mutex_lock(&ctx->uring_lock); |
3129 | percpu_ref_kill(&ctx->refs); | |
61cf9370 MWO |
3130 | xa_for_each(&ctx->personalities, index, creds) |
3131 | io_unregister_personality(ctx, index); | |
2b188cc1 JA |
3132 | mutex_unlock(&ctx->uring_lock); |
3133 | ||
dfbe5561 JA |
3134 | flush_delayed_work(&ctx->fallback_work); |
3135 | ||
85faa7b8 | 3136 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
3137 | /* |
3138 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
3139 | * if we're exiting a ton of rings at the same time. It just adds | |
3140 | * noise and overhead, there's no discernable change in runtime | |
3141 | * over using system_wq. | |
3142 | */ | |
73eaa2b5 | 3143 | queue_work(iou_wq, &ctx->exit_work); |
2b188cc1 JA |
3144 | } |
3145 | ||
3146 | static int io_uring_release(struct inode *inode, struct file *file) | |
3147 | { | |
3148 | struct io_ring_ctx *ctx = file->private_data; | |
3149 | ||
3150 | file->private_data = NULL; | |
3151 | io_ring_ctx_wait_and_kill(ctx); | |
3152 | return 0; | |
3153 | } | |
3154 | ||
f6edbabb PB |
3155 | struct io_task_cancel { |
3156 | struct task_struct *task; | |
3dd0c97a | 3157 | bool all; |
f6edbabb | 3158 | }; |
f254ac04 | 3159 | |
f6edbabb | 3160 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 3161 | { |
9a472ef7 | 3162 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 3163 | struct io_task_cancel *cancel = data; |
9a472ef7 | 3164 | |
6af3f48b | 3165 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
3166 | } |
3167 | ||
c072481d PB |
3168 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
3169 | struct task_struct *task, | |
3170 | bool cancel_all) | |
b7ddce3c | 3171 | { |
e1915f76 | 3172 | struct io_defer_entry *de; |
b7ddce3c PB |
3173 | LIST_HEAD(list); |
3174 | ||
79ebeaee | 3175 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 3176 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 3177 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
3178 | list_cut_position(&list, &ctx->defer_list, &de->list); |
3179 | break; | |
3180 | } | |
3181 | } | |
79ebeaee | 3182 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
3183 | if (list_empty(&list)) |
3184 | return false; | |
b7ddce3c PB |
3185 | |
3186 | while (!list_empty(&list)) { | |
3187 | de = list_first_entry(&list, struct io_defer_entry, list); | |
3188 | list_del_init(&de->list); | |
e276ae34 | 3189 | io_req_task_queue_fail(de->req, -ECANCELED); |
b7ddce3c PB |
3190 | kfree(de); |
3191 | } | |
e1915f76 | 3192 | return true; |
b7ddce3c PB |
3193 | } |
3194 | ||
c072481d | 3195 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
3196 | { |
3197 | struct io_tctx_node *node; | |
3198 | enum io_wq_cancel cret; | |
3199 | bool ret = false; | |
3200 | ||
3201 | mutex_lock(&ctx->uring_lock); | |
3202 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3203 | struct io_uring_task *tctx = node->task->io_uring; | |
3204 | ||
3205 | /* | |
3206 | * io_wq will stay alive while we hold uring_lock, because it's | |
3207 | * killed after ctx nodes, which requires to take the lock. | |
3208 | */ | |
3209 | if (!tctx || !tctx->io_wq) | |
3210 | continue; | |
3211 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
3212 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3213 | } | |
3214 | mutex_unlock(&ctx->uring_lock); | |
3215 | ||
3216 | return ret; | |
3217 | } | |
3218 | ||
affa87db | 3219 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
3220 | struct task_struct *task, |
3221 | bool cancel_all) | |
9936c7c2 | 3222 | { |
3dd0c97a | 3223 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 3224 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
3225 | enum io_wq_cancel cret; |
3226 | bool ret = false; | |
9936c7c2 | 3227 | |
360cd42c PB |
3228 | /* set it so io_req_local_work_add() would wake us up */ |
3229 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { | |
3230 | atomic_set(&ctx->cq_wait_nr, 1); | |
3231 | smp_mb(); | |
3232 | } | |
3233 | ||
60053be8 PB |
3234 | /* failed during ring init, it couldn't have issued any requests */ |
3235 | if (!ctx->rings) | |
affa87db | 3236 | return false; |
60053be8 | 3237 | |
affa87db PB |
3238 | if (!task) { |
3239 | ret |= io_uring_try_cancel_iowq(ctx); | |
3240 | } else if (tctx && tctx->io_wq) { | |
3241 | /* | |
3242 | * Cancels requests of all rings, not only @ctx, but | |
3243 | * it's fine as the task is in exit/exec. | |
3244 | */ | |
3245 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
3246 | &cancel, true); | |
3247 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3248 | } | |
9936c7c2 | 3249 | |
affa87db PB |
3250 | /* SQPOLL thread does its own polling */ |
3251 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
3252 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
3253 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
3254 | io_iopoll_try_reap_events(ctx); | |
3255 | ret = true; | |
fcc926bb | 3256 | cond_resched(); |
9936c7c2 | 3257 | } |
9936c7c2 | 3258 | } |
affa87db | 3259 | |
140102ae PB |
3260 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3261 | io_allowed_defer_tw_run(ctx)) | |
9fe3eaea | 3262 | ret |= io_run_local_work(ctx, INT_MAX) > 0; |
affa87db PB |
3263 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
3264 | mutex_lock(&ctx->uring_lock); | |
3265 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
f31ecf67 | 3266 | ret |= io_waitid_remove_all(ctx, task, cancel_all); |
194bb58c | 3267 | ret |= io_futex_remove_all(ctx, task, cancel_all); |
93b8cc60 | 3268 | ret |= io_uring_try_cancel_uring_cmd(ctx, task, cancel_all); |
affa87db PB |
3269 | mutex_unlock(&ctx->uring_lock); |
3270 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
3271 | if (task) | |
c0e0d6ba | 3272 | ret |= io_run_task_work() > 0; |
25417623 JA |
3273 | else |
3274 | ret |= flush_delayed_work(&ctx->fallback_work); | |
affa87db | 3275 | return ret; |
9936c7c2 PB |
3276 | } |
3277 | ||
3f48cf18 | 3278 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 3279 | { |
3f48cf18 | 3280 | if (tracked) |
9cae36a0 | 3281 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
3282 | return percpu_counter_sum(&tctx->inflight); |
3283 | } | |
3284 | ||
78cc687b PB |
3285 | /* |
3286 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 3287 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 3288 | */ |
17437f31 | 3289 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 3290 | { |
521d6a73 | 3291 | struct io_uring_task *tctx = current->io_uring; |
734551df | 3292 | struct io_ring_ctx *ctx; |
360cd42c PB |
3293 | struct io_tctx_node *node; |
3294 | unsigned long index; | |
0e9ddb39 PB |
3295 | s64 inflight; |
3296 | DEFINE_WAIT(wait); | |
fdaf083c | 3297 | |
78cc687b PB |
3298 | WARN_ON_ONCE(sqd && sqd->thread != current); |
3299 | ||
6d042ffb PO |
3300 | if (!current->io_uring) |
3301 | return; | |
17a91051 PB |
3302 | if (tctx->io_wq) |
3303 | io_wq_exit_start(tctx->io_wq); | |
3304 | ||
8d664282 | 3305 | atomic_inc(&tctx->in_cancel); |
0e9ddb39 | 3306 | do { |
affa87db PB |
3307 | bool loop = false; |
3308 | ||
e9dbe221 | 3309 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 3310 | /* read completions before cancelations */ |
78cc687b | 3311 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
3312 | if (!inflight) |
3313 | break; | |
fdaf083c | 3314 | |
78cc687b | 3315 | if (!sqd) { |
78cc687b PB |
3316 | xa_for_each(&tctx->xa, index, node) { |
3317 | /* sqpoll task will cancel all its requests */ | |
3318 | if (node->ctx->sq_data) | |
3319 | continue; | |
affa87db PB |
3320 | loop |= io_uring_try_cancel_requests(node->ctx, |
3321 | current, cancel_all); | |
78cc687b PB |
3322 | } |
3323 | } else { | |
3324 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
3325 | loop |= io_uring_try_cancel_requests(ctx, |
3326 | current, | |
3327 | cancel_all); | |
3328 | } | |
3329 | ||
3330 | if (loop) { | |
3331 | cond_resched(); | |
3332 | continue; | |
78cc687b | 3333 | } |
17a91051 | 3334 | |
78a78060 JA |
3335 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
3336 | io_run_task_work(); | |
e9dbe221 | 3337 | io_uring_drop_tctx_refs(current); |
360cd42c PB |
3338 | xa_for_each(&tctx->xa, index, node) { |
3339 | if (!llist_empty(&node->ctx->work_llist)) { | |
3340 | WARN_ON_ONCE(node->ctx->submitter_task && | |
3341 | node->ctx->submitter_task != current); | |
3342 | goto end_wait; | |
3343 | } | |
3344 | } | |
0f212204 | 3345 | /* |
a1bb3cd5 PB |
3346 | * If we've seen completions, retry without waiting. This |
3347 | * avoids a race where a completion comes in before we did | |
3348 | * prepare_to_wait(). | |
0f212204 | 3349 | */ |
3dd0c97a | 3350 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 3351 | schedule(); |
360cd42c | 3352 | end_wait: |
f57555ed | 3353 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 3354 | } while (1); |
de7f1d9e | 3355 | |
8452d4a6 | 3356 | io_uring_clean_tctx(tctx); |
3dd0c97a | 3357 | if (cancel_all) { |
3cc7fdb9 PB |
3358 | /* |
3359 | * We shouldn't run task_works after cancel, so just leave | |
8d664282 | 3360 | * ->in_cancel set for normal exit. |
3cc7fdb9 | 3361 | */ |
8d664282 | 3362 | atomic_dec(&tctx->in_cancel); |
3f48cf18 PB |
3363 | /* for exec all current's requests should be gone, kill tctx */ |
3364 | __io_uring_free(current); | |
3365 | } | |
44e728b8 PB |
3366 | } |
3367 | ||
f552a27a | 3368 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 3369 | { |
f552a27a | 3370 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
3371 | } |
3372 | ||
6c5c240e RP |
3373 | static void *io_uring_validate_mmap_request(struct file *file, |
3374 | loff_t pgoff, size_t sz) | |
2b188cc1 | 3375 | { |
2b188cc1 | 3376 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 3377 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 | 3378 | |
87585b05 | 3379 | switch ((pgoff << PAGE_SHIFT) & IORING_OFF_MMAP_MASK) { |
2b188cc1 | 3380 | case IORING_OFF_SQ_RING: |
75b28aff | 3381 | case IORING_OFF_CQ_RING: |
6f007b14 JA |
3382 | /* Don't allow mmap if the ring was setup without it */ |
3383 | if (ctx->flags & IORING_SETUP_NO_MMAP) | |
3384 | return ERR_PTR(-EINVAL); | |
3ab1db3c | 3385 | return ctx->rings; |
2b188cc1 | 3386 | case IORING_OFF_SQES: |
6f007b14 JA |
3387 | /* Don't allow mmap if the ring was setup without it */ |
3388 | if (ctx->flags & IORING_SETUP_NO_MMAP) | |
3389 | return ERR_PTR(-EINVAL); | |
3ab1db3c | 3390 | return ctx->sq_sqes; |
c56e022c | 3391 | case IORING_OFF_PBUF_RING: { |
561e4f94 | 3392 | struct io_buffer_list *bl; |
c56e022c | 3393 | unsigned int bgid; |
87585b05 | 3394 | void *ptr; |
c56e022c JA |
3395 | |
3396 | bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT; | |
561e4f94 JA |
3397 | bl = io_pbuf_get_bl(ctx, bgid); |
3398 | if (IS_ERR(bl)) | |
3399 | return bl; | |
3400 | ptr = bl->buf_ring; | |
3401 | io_put_bl(ctx, bl); | |
87585b05 | 3402 | return ptr; |
c56e022c | 3403 | } |
2b188cc1 JA |
3404 | } |
3405 | ||
87585b05 | 3406 | return ERR_PTR(-EINVAL); |
6c5c240e RP |
3407 | } |
3408 | ||
3ab1db3c JA |
3409 | int io_uring_mmap_pages(struct io_ring_ctx *ctx, struct vm_area_struct *vma, |
3410 | struct page **pages, int npages) | |
3411 | { | |
3412 | unsigned long nr_pages = npages; | |
3413 | ||
3414 | vm_flags_set(vma, VM_DONTEXPAND); | |
3415 | return vm_insert_pages(vma, vma->vm_start, pages, &nr_pages); | |
3416 | } | |
3417 | ||
6c5c240e RP |
3418 | #ifdef CONFIG_MMU |
3419 | ||
c072481d | 3420 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e | 3421 | { |
3ab1db3c | 3422 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 3423 | size_t sz = vma->vm_end - vma->vm_start; |
3ab1db3c | 3424 | long offset = vma->vm_pgoff << PAGE_SHIFT; |
6c5c240e RP |
3425 | void *ptr; |
3426 | ||
3427 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
3428 | if (IS_ERR(ptr)) | |
3429 | return PTR_ERR(ptr); | |
2b188cc1 | 3430 | |
3ab1db3c JA |
3431 | switch (offset & IORING_OFF_MMAP_MASK) { |
3432 | case IORING_OFF_SQ_RING: | |
3433 | case IORING_OFF_CQ_RING: | |
3434 | return io_uring_mmap_pages(ctx, vma, ctx->ring_pages, | |
3435 | ctx->n_ring_pages); | |
3436 | case IORING_OFF_SQES: | |
3437 | return io_uring_mmap_pages(ctx, vma, ctx->sqe_pages, | |
3438 | ctx->n_sqe_pages); | |
87585b05 JA |
3439 | case IORING_OFF_PBUF_RING: |
3440 | return io_pbuf_mmap(file, vma); | |
3ab1db3c JA |
3441 | } |
3442 | ||
87585b05 | 3443 | return -EINVAL; |
2b188cc1 JA |
3444 | } |
3445 | ||
d808459b HD |
3446 | static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp, |
3447 | unsigned long addr, unsigned long len, | |
3448 | unsigned long pgoff, unsigned long flags) | |
3449 | { | |
d808459b HD |
3450 | void *ptr; |
3451 | ||
3452 | /* | |
3453 | * Do not allow to map to user-provided address to avoid breaking the | |
3454 | * aliasing rules. Userspace is not able to guess the offset address of | |
3455 | * kernel kmalloc()ed memory area. | |
3456 | */ | |
3457 | if (addr) | |
3458 | return -EINVAL; | |
3459 | ||
3460 | ptr = io_uring_validate_mmap_request(filp, pgoff, len); | |
3461 | if (IS_ERR(ptr)) | |
3462 | return -ENOMEM; | |
3463 | ||
32832a40 HD |
3464 | /* |
3465 | * Some architectures have strong cache aliasing requirements. | |
3466 | * For such architectures we need a coherent mapping which aliases | |
3467 | * kernel memory *and* userspace memory. To achieve that: | |
3468 | * - use a NULL file pointer to reference physical memory, and | |
3469 | * - use the kernel virtual address of the shared io_uring context | |
3470 | * (instead of the userspace-provided address, which has to be 0UL | |
3471 | * anyway). | |
56675f8b HD |
3472 | * - use the same pgoff which the get_unmapped_area() uses to |
3473 | * calculate the page colouring. | |
32832a40 HD |
3474 | * For architectures without such aliasing requirements, the |
3475 | * architecture will return any suitable mapping because addr is 0. | |
3476 | */ | |
3477 | filp = NULL; | |
3478 | flags |= MAP_SHARED; | |
3479 | pgoff = 0; /* has been translated to ptr above */ | |
d808459b | 3480 | #ifdef SHM_COLOUR |
32832a40 | 3481 | addr = (uintptr_t) ptr; |
56675f8b | 3482 | pgoff = addr >> PAGE_SHIFT; |
d808459b | 3483 | #else |
32832a40 | 3484 | addr = 0UL; |
d808459b | 3485 | #endif |
32832a40 | 3486 | return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags); |
d808459b HD |
3487 | } |
3488 | ||
6c5c240e RP |
3489 | #else /* !CONFIG_MMU */ |
3490 | ||
3491 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3492 | { | |
fc4f4be9 | 3493 | return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL; |
6c5c240e RP |
3494 | } |
3495 | ||
3496 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
3497 | { | |
3498 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
3499 | } | |
3500 | ||
3501 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
3502 | unsigned long addr, unsigned long len, | |
3503 | unsigned long pgoff, unsigned long flags) | |
3504 | { | |
3505 | void *ptr; | |
3506 | ||
3507 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
3508 | if (IS_ERR(ptr)) | |
3509 | return PTR_ERR(ptr); | |
3510 | ||
3511 | return (unsigned long) ptr; | |
3512 | } | |
3513 | ||
3514 | #endif /* !CONFIG_MMU */ | |
3515 | ||
f81440d3 PB |
3516 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
3517 | { | |
3518 | if (flags & IORING_ENTER_EXT_ARG) { | |
3519 | struct io_uring_getevents_arg arg; | |
3520 | ||
3521 | if (argsz != sizeof(arg)) | |
3522 | return -EINVAL; | |
3523 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3524 | return -EFAULT; | |
3525 | } | |
3526 | return 0; | |
3527 | } | |
3528 | ||
c73ebb68 HX |
3529 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
3530 | struct __kernel_timespec __user **ts, | |
3531 | const sigset_t __user **sig) | |
3532 | { | |
3533 | struct io_uring_getevents_arg arg; | |
3534 | ||
3535 | /* | |
3536 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
3537 | * is just a pointer to the sigset_t. | |
3538 | */ | |
3539 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
3540 | *sig = (const sigset_t __user *) argp; | |
3541 | *ts = NULL; | |
3542 | return 0; | |
3543 | } | |
3544 | ||
3545 | /* | |
3546 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
3547 | * timespec and sigset_t pointers if good. | |
3548 | */ | |
3549 | if (*argsz != sizeof(arg)) | |
3550 | return -EINVAL; | |
3551 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3552 | return -EFAULT; | |
d2347b96 DY |
3553 | if (arg.pad) |
3554 | return -EINVAL; | |
c73ebb68 HX |
3555 | *sig = u64_to_user_ptr(arg.sigmask); |
3556 | *argsz = arg.sigmask_sz; | |
3557 | *ts = u64_to_user_ptr(arg.ts); | |
3558 | return 0; | |
3559 | } | |
3560 | ||
2b188cc1 | 3561 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
3562 | u32, min_complete, u32, flags, const void __user *, argp, |
3563 | size_t, argsz) | |
2b188cc1 JA |
3564 | { |
3565 | struct io_ring_ctx *ctx; | |
73363c26 | 3566 | struct file *file; |
33f993da | 3567 | long ret; |
2b188cc1 | 3568 | |
33f993da | 3569 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
3570 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
3571 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
3572 | return -EINVAL; |
3573 | ||
e7a6c00d JA |
3574 | /* |
3575 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3576 | * need only dereference our task private array to find it. | |
3577 | */ | |
3578 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3579 | struct io_uring_task *tctx = current->io_uring; | |
3580 | ||
3273c440 | 3581 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
e7a6c00d JA |
3582 | return -EINVAL; |
3583 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
73363c26 JA |
3584 | file = tctx->registered_rings[fd]; |
3585 | if (unlikely(!file)) | |
3273c440 | 3586 | return -EBADF; |
e7a6c00d | 3587 | } else { |
73363c26 JA |
3588 | file = fget(fd); |
3589 | if (unlikely(!file)) | |
3273c440 PB |
3590 | return -EBADF; |
3591 | ret = -EOPNOTSUPP; | |
73363c26 | 3592 | if (unlikely(!io_is_uring_fops(file))) |
fbb8bb02 | 3593 | goto out; |
e7a6c00d | 3594 | } |
2b188cc1 | 3595 | |
73363c26 | 3596 | ctx = file->private_data; |
7e84e1c7 | 3597 | ret = -EBADFD; |
33f993da | 3598 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3599 | goto out; |
3600 | ||
6c271ce2 JA |
3601 | /* |
3602 | * For SQ polling, the thread will do all submissions and completions. | |
3603 | * Just return the requested submit count, and wake the thread if | |
3604 | * we were asked to. | |
3605 | */ | |
b2a9eada | 3606 | ret = 0; |
6c271ce2 | 3607 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3608 | io_cqring_overflow_flush(ctx); |
89448c47 | 3609 | |
21f96522 JA |
3610 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3611 | ret = -EOWNERDEAD; | |
04147488 | 3612 | goto out; |
21f96522 | 3613 | } |
6c271ce2 | 3614 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3615 | wake_up(&ctx->sq_data->wait); |
88b80534 QF |
3616 | if (flags & IORING_ENTER_SQ_WAIT) |
3617 | io_sqpoll_wait_sq(ctx); | |
3618 | ||
3e813c90 | 3619 | ret = to_submit; |
b2a9eada | 3620 | } else if (to_submit) { |
eef51daa | 3621 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3622 | if (unlikely(ret)) |
3623 | goto out; | |
7c504e65 | 3624 | |
2b188cc1 | 3625 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3626 | ret = io_submit_sqes(ctx, to_submit); |
3627 | if (ret != to_submit) { | |
d487b43c | 3628 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3629 | goto out; |
d487b43c | 3630 | } |
44f87745 PB |
3631 | if (flags & IORING_ENTER_GETEVENTS) { |
3632 | if (ctx->syscall_iopoll) | |
3633 | goto iopoll_locked; | |
3634 | /* | |
3635 | * Ignore errors, we'll soon call io_cqring_wait() and | |
3636 | * it should handle ownership problems if any. | |
3637 | */ | |
3638 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
9fe3eaea | 3639 | (void)io_run_local_work_locked(ctx, min_complete); |
44f87745 | 3640 | } |
d487b43c | 3641 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 | 3642 | } |
c0e0d6ba | 3643 | |
2b188cc1 | 3644 | if (flags & IORING_ENTER_GETEVENTS) { |
3e813c90 | 3645 | int ret2; |
c0e0d6ba | 3646 | |
773697b6 | 3647 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3648 | /* |
3649 | * We disallow the app entering submit/complete with | |
3650 | * polling, but we still need to lock the ring to | |
3651 | * prevent racing with polled issue that got punted to | |
3652 | * a workqueue. | |
3653 | */ | |
3654 | mutex_lock(&ctx->uring_lock); | |
3655 | iopoll_locked: | |
3e813c90 DY |
3656 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3657 | if (likely(!ret2)) { | |
3658 | min_complete = min(min_complete, | |
3659 | ctx->cq_entries); | |
3660 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3661 | } |
3662 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3663 | } else { |
f81440d3 PB |
3664 | const sigset_t __user *sig; |
3665 | struct __kernel_timespec __user *ts; | |
3666 | ||
3e813c90 DY |
3667 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3668 | if (likely(!ret2)) { | |
3669 | min_complete = min(min_complete, | |
3670 | ctx->cq_entries); | |
3671 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3672 | argsz, ts); | |
3673 | } | |
def596e9 | 3674 | } |
c73ebb68 | 3675 | |
155bc950 | 3676 | if (!ret) { |
3e813c90 | 3677 | ret = ret2; |
2b188cc1 | 3678 | |
155bc950 DY |
3679 | /* |
3680 | * EBADR indicates that one or more CQE were dropped. | |
3681 | * Once the user has been informed we can clear the bit | |
3682 | * as they are obviously ok with those drops. | |
3683 | */ | |
3684 | if (unlikely(ret2 == -EBADR)) | |
3685 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3686 | &ctx->check_cq); | |
def596e9 | 3687 | } |
2b188cc1 | 3688 | } |
7c504e65 | 3689 | out: |
73363c26 JA |
3690 | if (!(flags & IORING_ENTER_REGISTERED_RING)) |
3691 | fput(file); | |
3e813c90 | 3692 | return ret; |
2b188cc1 JA |
3693 | } |
3694 | ||
3695 | static const struct file_operations io_uring_fops = { | |
3696 | .release = io_uring_release, | |
3697 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3698 | #ifndef CONFIG_MMU |
3699 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3700 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
d808459b HD |
3701 | #else |
3702 | .get_unmapped_area = io_uring_mmu_get_unmapped_area, | |
6c5c240e | 3703 | #endif |
2b188cc1 | 3704 | .poll = io_uring_poll, |
bebdb65e | 3705 | #ifdef CONFIG_PROC_FS |
87ce955b | 3706 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3707 | #endif |
2b188cc1 JA |
3708 | }; |
3709 | ||
92ac8bea JA |
3710 | bool io_is_uring_fops(struct file *file) |
3711 | { | |
3712 | return file->f_op == &io_uring_fops; | |
3713 | } | |
3714 | ||
c072481d PB |
3715 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3716 | struct io_uring_params *p) | |
2b188cc1 | 3717 | { |
75b28aff HV |
3718 | struct io_rings *rings; |
3719 | size_t size, sq_array_offset; | |
e27cef86 | 3720 | void *ptr; |
2b188cc1 | 3721 | |
bd740481 JA |
3722 | /* make sure these are sane, as we already accounted them */ |
3723 | ctx->sq_entries = p->sq_entries; | |
3724 | ctx->cq_entries = p->cq_entries; | |
3725 | ||
baf9cb64 | 3726 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3727 | if (size == SIZE_MAX) |
3728 | return -EOVERFLOW; | |
3729 | ||
03d89a2d | 3730 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3ab1db3c | 3731 | rings = io_pages_map(&ctx->ring_pages, &ctx->n_ring_pages, size); |
03d89a2d JA |
3732 | else |
3733 | rings = io_rings_map(ctx, p->cq_off.user_addr, size); | |
3734 | ||
e27cef86 JA |
3735 | if (IS_ERR(rings)) |
3736 | return PTR_ERR(rings); | |
2b188cc1 | 3737 | |
75b28aff | 3738 | ctx->rings = rings; |
2af89abd PB |
3739 | if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) |
3740 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
75b28aff HV |
3741 | rings->sq_ring_mask = p->sq_entries - 1; |
3742 | rings->cq_ring_mask = p->cq_entries - 1; | |
3743 | rings->sq_ring_entries = p->sq_entries; | |
3744 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3745 | |
ebdeb7c0 JA |
3746 | if (p->flags & IORING_SETUP_SQE128) |
3747 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3748 | else | |
3749 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 | 3750 | if (size == SIZE_MAX) { |
9c189eee | 3751 | io_rings_free(ctx); |
2b188cc1 | 3752 | return -EOVERFLOW; |
eb065d30 | 3753 | } |
2b188cc1 | 3754 | |
03d89a2d | 3755 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3ab1db3c | 3756 | ptr = io_pages_map(&ctx->sqe_pages, &ctx->n_sqe_pages, size); |
03d89a2d JA |
3757 | else |
3758 | ptr = io_sqes_map(ctx, p->sq_off.user_addr, size); | |
3759 | ||
e27cef86 | 3760 | if (IS_ERR(ptr)) { |
9c189eee | 3761 | io_rings_free(ctx); |
e27cef86 | 3762 | return PTR_ERR(ptr); |
eb065d30 | 3763 | } |
2b188cc1 | 3764 | |
e27cef86 | 3765 | ctx->sq_sqes = ptr; |
2b188cc1 JA |
3766 | return 0; |
3767 | } | |
3768 | ||
6e76ac59 | 3769 | static int io_uring_install_fd(struct file *file) |
9faadcc8 | 3770 | { |
6e76ac59 | 3771 | int fd; |
9faadcc8 PB |
3772 | |
3773 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3774 | if (fd < 0) | |
3775 | return fd; | |
9faadcc8 PB |
3776 | fd_install(fd, file); |
3777 | return fd; | |
3778 | } | |
3779 | ||
2b188cc1 JA |
3780 | /* |
3781 | * Allocate an anonymous fd, this is what constitutes the application | |
3782 | * visible backing of an io_uring instance. The application mmaps this | |
6e5e6d27 | 3783 | * fd to gain access to the SQ/CQ ring details. |
2b188cc1 | 3784 | */ |
9faadcc8 | 3785 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 | 3786 | { |
4f0b9194 | 3787 | /* Create a new inode so that the LSM can block the creation. */ |
09d1c6a8 | 3788 | return anon_inode_create_getfile("[io_uring]", &io_uring_fops, ctx, |
91a9ab7c | 3789 | O_RDWR | O_CLOEXEC, NULL); |
2b188cc1 JA |
3790 | } |
3791 | ||
c072481d PB |
3792 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3793 | struct io_uring_params __user *params) | |
2b188cc1 | 3794 | { |
2b188cc1 | 3795 | struct io_ring_ctx *ctx; |
6e76ac59 | 3796 | struct io_uring_task *tctx; |
9faadcc8 | 3797 | struct file *file; |
2b188cc1 JA |
3798 | int ret; |
3799 | ||
8110c1a6 | 3800 | if (!entries) |
2b188cc1 | 3801 | return -EINVAL; |
8110c1a6 JA |
3802 | if (entries > IORING_MAX_ENTRIES) { |
3803 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3804 | return -EINVAL; | |
3805 | entries = IORING_MAX_ENTRIES; | |
3806 | } | |
2b188cc1 | 3807 | |
6e76ac59 JT |
3808 | if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY) |
3809 | && !(p->flags & IORING_SETUP_NO_MMAP)) | |
3810 | return -EINVAL; | |
3811 | ||
2b188cc1 JA |
3812 | /* |
3813 | * Use twice as many entries for the CQ ring. It's possible for the | |
3814 | * application to drive a higher depth than the size of the SQ ring, | |
3815 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3816 | * some flexibility in overcommitting a bit. If the application has |
3817 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3818 | * of CQ ring entries manually. | |
2b188cc1 JA |
3819 | */ |
3820 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3821 | if (p->flags & IORING_SETUP_CQSIZE) { |
3822 | /* | |
3823 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3824 | * to a power-of-two, if it isn't already. We do NOT impose | |
3825 | * any cq vs sq ring sizing. | |
3826 | */ | |
eb2667b3 | 3827 | if (!p->cq_entries) |
33a107f0 | 3828 | return -EINVAL; |
8110c1a6 JA |
3829 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3830 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3831 | return -EINVAL; | |
3832 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3833 | } | |
eb2667b3 JQ |
3834 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3835 | if (p->cq_entries < p->sq_entries) | |
3836 | return -EINVAL; | |
33a107f0 JA |
3837 | } else { |
3838 | p->cq_entries = 2 * p->sq_entries; | |
3839 | } | |
2b188cc1 | 3840 | |
2b188cc1 | 3841 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3842 | if (!ctx) |
2b188cc1 | 3843 | return -ENOMEM; |
773697b6 | 3844 | |
e6aeb272 PB |
3845 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3846 | !(ctx->flags & IORING_SETUP_IOPOLL) && | |
3847 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3848 | ctx->task_complete = true; | |
3849 | ||
ec26c225 PB |
3850 | if (ctx->task_complete || (ctx->flags & IORING_SETUP_IOPOLL)) |
3851 | ctx->lockless_cq = true; | |
3852 | ||
bca39f39 PB |
3853 | /* |
3854 | * lazy poll_wq activation relies on ->task_complete for synchronisation | |
3855 | * purposes, see io_activate_pollwq() | |
3856 | */ | |
3857 | if (!ctx->task_complete) | |
3858 | ctx->poll_activated = true; | |
3859 | ||
773697b6 PB |
3860 | /* |
3861 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3862 | * space applications don't need to do io completion events | |
3863 | * polling again, they can rely on io_sq_thread to do polling | |
3864 | * work, which can reduce cpu usage and uring_lock contention. | |
3865 | */ | |
3866 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3867 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3868 | ctx->syscall_iopoll = 1; | |
3869 | ||
2b188cc1 | 3870 | ctx->compat = in_compat_syscall(); |
6adc2272 | 3871 | if (!ns_capable_noaudit(&init_user_ns, CAP_IPC_LOCK)) |
62e398be | 3872 | ctx->user = get_uid(current_user()); |
2aede0e4 | 3873 | |
9f010507 | 3874 | /* |
e1169f06 JA |
3875 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3876 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3877 | */ |
e1169f06 JA |
3878 | ret = -EINVAL; |
3879 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3880 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 | 3881 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
c0e0d6ba DY |
3882 | IORING_SETUP_TASKRUN_FLAG | |
3883 | IORING_SETUP_DEFER_TASKRUN)) | |
e1169f06 | 3884 | goto err; |
9f010507 | 3885 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3886 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3887 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3888 | } else { | |
c0e0d6ba DY |
3889 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && |
3890 | !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) | |
ef060ea9 | 3891 | goto err; |
9f010507 | 3892 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3893 | } |
9f010507 | 3894 | |
c0e0d6ba DY |
3895 | /* |
3896 | * For DEFER_TASKRUN we require the completion task to be the same as the | |
3897 | * submission task. This implies that there is only one submitter, so enforce | |
3898 | * that. | |
3899 | */ | |
3900 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && | |
3901 | !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { | |
3902 | goto err; | |
3903 | } | |
3904 | ||
2aede0e4 JA |
3905 | /* |
3906 | * This is just grabbed for accounting purposes. When a process exits, | |
3907 | * the mm is exited and dropped before the files, hence we need to hang | |
3908 | * on to this mm purely for the purposes of being able to unaccount | |
3909 | * memory (locked/pinned vm). It's not used for anything else. | |
3910 | */ | |
6b7898eb | 3911 | mmgrab(current->mm); |
2aede0e4 | 3912 | ctx->mm_account = current->mm; |
6b7898eb | 3913 | |
2b188cc1 JA |
3914 | ret = io_allocate_scq_urings(ctx, p); |
3915 | if (ret) | |
3916 | goto err; | |
3917 | ||
7e84e1c7 | 3918 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
3919 | if (ret) |
3920 | goto err; | |
2933ae6e PB |
3921 | |
3922 | ret = io_rsrc_init(ctx); | |
47b228ce PB |
3923 | if (ret) |
3924 | goto err; | |
2b188cc1 | 3925 | |
75b28aff HV |
3926 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3927 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3928 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3929 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3930 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3931 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
2af89abd PB |
3932 | if (!(ctx->flags & IORING_SETUP_NO_SQARRAY)) |
3933 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
9b1b58ca | 3934 | p->sq_off.resv1 = 0; |
03d89a2d JA |
3935 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3936 | p->sq_off.user_addr = 0; | |
2b188cc1 | 3937 | |
75b28aff HV |
3938 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3939 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3940 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3941 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3942 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3943 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 3944 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
9b1b58ca | 3945 | p->cq_off.resv1 = 0; |
03d89a2d JA |
3946 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3947 | p->cq_off.user_addr = 0; | |
ac90f249 | 3948 | |
7f13657d XW |
3949 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
3950 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 3951 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 3952 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 3953 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e | 3954 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
7d3fd88d | 3955 | IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING; |
7f13657d XW |
3956 | |
3957 | if (copy_to_user(params, p, sizeof(*p))) { | |
3958 | ret = -EFAULT; | |
3959 | goto err; | |
3960 | } | |
d1719f70 | 3961 | |
7cae596b DY |
3962 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER |
3963 | && !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
8579538c | 3964 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
7cae596b | 3965 | |
9faadcc8 PB |
3966 | file = io_uring_get_file(ctx); |
3967 | if (IS_ERR(file)) { | |
3968 | ret = PTR_ERR(file); | |
3969 | goto err; | |
3970 | } | |
3971 | ||
6e76ac59 JT |
3972 | ret = __io_uring_add_tctx_node(ctx); |
3973 | if (ret) | |
3974 | goto err_fput; | |
3975 | tctx = current->io_uring; | |
3976 | ||
044c1ab3 JA |
3977 | /* |
3978 | * Install ring fd as the very last thing, so we don't risk someone | |
3979 | * having closed it before we finish setup | |
3980 | */ | |
6e76ac59 JT |
3981 | if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY) |
3982 | ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX); | |
3983 | else | |
3984 | ret = io_uring_install_fd(file); | |
3985 | if (ret < 0) | |
3986 | goto err_fput; | |
044c1ab3 | 3987 | |
c826bd7a | 3988 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
3989 | return ret; |
3990 | err: | |
3991 | io_ring_ctx_wait_and_kill(ctx); | |
3992 | return ret; | |
6e76ac59 JT |
3993 | err_fput: |
3994 | fput(file); | |
3995 | return ret; | |
2b188cc1 JA |
3996 | } |
3997 | ||
3998 | /* | |
3999 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
4000 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
4001 | * params structure passed in. | |
4002 | */ | |
4003 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
4004 | { | |
4005 | struct io_uring_params p; | |
2b188cc1 JA |
4006 | int i; |
4007 | ||
4008 | if (copy_from_user(&p, params, sizeof(p))) | |
4009 | return -EFAULT; | |
4010 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
4011 | if (p.resv[i]) | |
4012 | return -EINVAL; | |
4013 | } | |
4014 | ||
6c271ce2 | 4015 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 4016 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 4017 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 4018 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 4019 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 | 4020 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
03d89a2d | 4021 | IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN | |
2af89abd PB |
4022 | IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY | |
4023 | IORING_SETUP_NO_SQARRAY)) | |
2b188cc1 JA |
4024 | return -EINVAL; |
4025 | ||
ef060ea9 | 4026 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
4027 | } |
4028 | ||
76d3ccec MR |
4029 | static inline bool io_uring_allowed(void) |
4030 | { | |
4031 | int disabled = READ_ONCE(sysctl_io_uring_disabled); | |
4032 | kgid_t io_uring_group; | |
4033 | ||
4034 | if (disabled == 2) | |
4035 | return false; | |
4036 | ||
4037 | if (disabled == 0 || capable(CAP_SYS_ADMIN)) | |
4038 | return true; | |
4039 | ||
4040 | io_uring_group = make_kgid(&init_user_ns, sysctl_io_uring_group); | |
4041 | if (!gid_valid(io_uring_group)) | |
4042 | return false; | |
4043 | ||
4044 | return in_group_p(io_uring_group); | |
4045 | } | |
4046 | ||
2b188cc1 JA |
4047 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, |
4048 | struct io_uring_params __user *, params) | |
4049 | { | |
76d3ccec MR |
4050 | if (!io_uring_allowed()) |
4051 | return -EPERM; | |
4052 | ||
2b188cc1 JA |
4053 | return io_uring_setup(entries, params); |
4054 | } | |
4055 | ||
4056 | static int __init io_uring_init(void) | |
4057 | { | |
9c71d39a | 4058 | #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ |
d7f62e82 | 4059 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ |
9c71d39a | 4060 | BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ |
d7f62e82 SM |
4061 | } while (0) |
4062 | ||
4063 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9c71d39a SM |
4064 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) |
4065 | #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ | |
4066 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) | |
d7f62e82 SM |
4067 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); |
4068 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
4069 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
4070 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
4071 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
4072 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
4073 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9c71d39a SM |
4074 | BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); |
4075 | BUILD_BUG_SQE_ELEM(12, __u32, __pad1); | |
d7f62e82 | 4076 | BUILD_BUG_SQE_ELEM(16, __u64, addr); |
7d67af2c | 4077 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
4078 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
4079 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
4080 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
4081 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
4082 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
4083 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
4084 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
4085 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
4086 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
4087 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
4088 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
4089 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
4090 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
4091 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
4092 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 4093 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
9c71d39a SM |
4094 | BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); |
4095 | BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); | |
4096 | BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); | |
4097 | BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); | |
4098 | BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); | |
d7f62e82 SM |
4099 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
4100 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 4101 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 4102 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 4103 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 4104 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
b48c312b PB |
4105 | BUILD_BUG_SQE_ELEM(44, __u16, addr_len); |
4106 | BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); | |
e9621e2b | 4107 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
9c71d39a SM |
4108 | BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); |
4109 | BUILD_BUG_SQE_ELEM(56, __u64, __pad2); | |
d7f62e82 | 4110 | |
b0d658ec PB |
4111 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
4112 | sizeof(struct io_uring_rsrc_update)); | |
4113 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
4114 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
4115 | |
4116 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
4117 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
4118 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
4119 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 4120 | |
b0d658ec PB |
4121 | /* should fit into one byte */ |
4122 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
4123 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
4124 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 4125 | |
4bcb982c | 4126 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof_field(struct io_kiocb, flags)); |
16340eab | 4127 | |
3a4b89a2 JA |
4128 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
4129 | ||
528ce678 ML |
4130 | /* top 8bits are for internal use */ |
4131 | BUILD_BUG_ON((IORING_URING_CMD_MASK & 0xff000000) != 0); | |
4132 | ||
d9b57aa3 | 4133 | io_uring_optable_init(); |
0702e536 | 4134 | |
b97f96e2 JA |
4135 | /* |
4136 | * Allow user copy in the per-command field, which starts after the | |
4137 | * file in io_kiocb and until the opcode field. The openat2 handling | |
4138 | * requires copying in user memory into the io_kiocb object in that | |
4139 | * range, and HARDENED_USERCOPY will complain if we haven't | |
4140 | * correctly annotated this range. | |
4141 | */ | |
4142 | req_cachep = kmem_cache_create_usercopy("io_kiocb", | |
4143 | sizeof(struct io_kiocb), 0, | |
4144 | SLAB_HWCACHE_ALIGN | SLAB_PANIC | | |
4145 | SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU, | |
4146 | offsetof(struct io_kiocb, cmd.data), | |
4147 | sizeof_field(struct io_kiocb, cmd.data), NULL); | |
a6e959bd KC |
4148 | io_buf_cachep = KMEM_CACHE(io_buffer, |
4149 | SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT); | |
b97f96e2 | 4150 | |
73eaa2b5 JA |
4151 | iou_wq = alloc_workqueue("iou_exit", WQ_UNBOUND, 64); |
4152 | ||
76d3ccec MR |
4153 | #ifdef CONFIG_SYSCTL |
4154 | register_sysctl_init("kernel", kernel_io_uring_disabled_table); | |
4155 | #endif | |
4156 | ||
2b188cc1 JA |
4157 | return 0; |
4158 | }; | |
4159 | __initcall(io_uring_init); |