Commit | Line | Data |
---|---|---|
2b188cc1 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Shared application/kernel submission and completion ring pairs, for | |
4 | * supporting fast/efficient IO. | |
5 | * | |
6 | * A note on the read/write ordering memory barriers that are matched between | |
1e84b97b SB |
7 | * the application and kernel side. |
8 | * | |
9 | * After the application reads the CQ ring tail, it must use an | |
10 | * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses | |
11 | * before writing the tail (using smp_load_acquire to read the tail will | |
12 | * do). It also needs a smp_mb() before updating CQ head (ordering the | |
13 | * entry load(s) with the head store), pairing with an implicit barrier | |
d068b506 | 14 | * through a control-dependency in io_get_cqe (smp_store_release to |
1e84b97b SB |
15 | * store head will do). Failure to do so could lead to reading invalid |
16 | * CQ entries. | |
17 | * | |
18 | * Likewise, the application must use an appropriate smp_wmb() before | |
19 | * writing the SQ tail (ordering SQ entry stores with the tail store), | |
20 | * which pairs with smp_load_acquire in io_get_sqring (smp_store_release | |
21 | * to store the tail will do). And it needs a barrier ordering the SQ | |
22 | * head load before writing new SQ entries (smp_load_acquire to read | |
23 | * head will do). | |
24 | * | |
25 | * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application | |
26 | * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after* | |
27 | * updating the SQ tail; a full memory barrier smp_mb() is needed | |
28 | * between. | |
2b188cc1 JA |
29 | * |
30 | * Also see the examples in the liburing library: | |
31 | * | |
32 | * git://git.kernel.dk/liburing | |
33 | * | |
34 | * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens | |
35 | * from data shared between the kernel and application. This is done both | |
36 | * for ordering purposes, but also to ensure that once a value is loaded from | |
37 | * data that the application could potentially modify, it remains stable. | |
38 | * | |
39 | * Copyright (C) 2018-2019 Jens Axboe | |
c992fe29 | 40 | * Copyright (c) 2018-2019 Christoph Hellwig |
2b188cc1 JA |
41 | */ |
42 | #include <linux/kernel.h> | |
43 | #include <linux/init.h> | |
44 | #include <linux/errno.h> | |
45 | #include <linux/syscalls.h> | |
52de1fe1 | 46 | #include <net/compat.h> |
2b188cc1 JA |
47 | #include <linux/refcount.h> |
48 | #include <linux/uio.h> | |
6b47ee6e | 49 | #include <linux/bits.h> |
2b188cc1 JA |
50 | |
51 | #include <linux/sched/signal.h> | |
52 | #include <linux/fs.h> | |
53 | #include <linux/file.h> | |
54 | #include <linux/fdtable.h> | |
55 | #include <linux/mm.h> | |
56 | #include <linux/mman.h> | |
2b188cc1 JA |
57 | #include <linux/percpu.h> |
58 | #include <linux/slab.h> | |
edafccee | 59 | #include <linux/bvec.h> |
2b188cc1 JA |
60 | #include <linux/net.h> |
61 | #include <net/sock.h> | |
62 | #include <net/af_unix.h> | |
6b06314c | 63 | #include <net/scm.h> |
2b188cc1 JA |
64 | #include <linux/anon_inodes.h> |
65 | #include <linux/sched/mm.h> | |
66 | #include <linux/uaccess.h> | |
67 | #include <linux/nospec.h> | |
aa4c3967 | 68 | #include <linux/highmem.h> |
15b71abe | 69 | #include <linux/fsnotify.h> |
4840e418 | 70 | #include <linux/fadvise.h> |
b41e9852 | 71 | #include <linux/task_work.h> |
0f212204 | 72 | #include <linux/io_uring.h> |
5bd2182d | 73 | #include <linux/audit.h> |
cdc1404a | 74 | #include <linux/security.h> |
2b188cc1 | 75 | |
c826bd7a DD |
76 | #define CREATE_TRACE_POINTS |
77 | #include <trace/events/io_uring.h> | |
78 | ||
2b188cc1 JA |
79 | #include <uapi/linux/io_uring.h> |
80 | ||
561fb04a | 81 | #include "io-wq.h" |
2b188cc1 | 82 | |
de23077e | 83 | #include "io_uring.h" |
329061d3 | 84 | #include "opdef.h" |
e418bbc9 | 85 | #include "refs.h" |
c9f06aa7 | 86 | #include "tctx.h" |
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" |
e27f928e | 94 | |
59915143 | 95 | #include "timeout.h" |
329061d3 | 96 | #include "poll.h" |
9b797a37 | 97 | #include "alloc_cache.h" |
5e2a18d9 | 98 | |
5277deaa | 99 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 100 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 | 101 | |
21b55dbc SG |
102 | #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \ |
103 | IORING_REGISTER_LAST + IORING_OP_LAST) | |
2b188cc1 | 104 | |
68fe256a PB |
105 | #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \ |
106 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) | |
107 | ||
5562a8d7 PB |
108 | #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \ |
109 | IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS) | |
68fe256a | 110 | |
c854357b | 111 | #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ |
9cae36a0 JA |
112 | REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \ |
113 | REQ_F_ASYNC_DATA) | |
b16fed66 | 114 | |
a538be5b PB |
115 | #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\ |
116 | IO_REQ_CLEAN_FLAGS) | |
117 | ||
09899b19 PB |
118 | #define IO_TCTX_REFS_CACHE_NR (1U << 10) |
119 | ||
6dd0be1e | 120 | #define IO_COMPL_BATCH 32 |
bf019da7 | 121 | #define IO_REQ_ALLOC_BATCH 8 |
258b29a9 | 122 | |
10988a0a DY |
123 | enum { |
124 | IO_CHECK_CQ_OVERFLOW_BIT, | |
155bc950 | 125 | IO_CHECK_CQ_DROPPED_BIT, |
10988a0a DY |
126 | }; |
127 | ||
21a091b9 DY |
128 | enum { |
129 | IO_EVENTFD_OP_SIGNAL_BIT, | |
130 | IO_EVENTFD_OP_FREE_BIT, | |
131 | }; | |
132 | ||
27dc8338 PB |
133 | struct io_defer_entry { |
134 | struct list_head list; | |
135 | struct io_kiocb *req; | |
9cf7c104 | 136 | u32 seq; |
2b188cc1 JA |
137 | }; |
138 | ||
0756a869 PB |
139 | /* requests with any of those set should undergo io_disarm_next() */ |
140 | #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) | |
da1a08c5 | 141 | #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK) |
0756a869 | 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 | |
c7dae4ba | 147 | static void io_dismantle_req(struct io_kiocb *req); |
68fb8979 | 148 | static void io_clean_op(struct io_kiocb *req); |
cbc2e203 | 149 | static void io_queue_sqe(struct io_kiocb *req); |
c0e0d6ba | 150 | static void io_move_task_work_from_local(struct io_ring_ctx *ctx); |
c450178d | 151 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx); |
77e443ab | 152 | static __cold void io_fallback_tw(struct io_uring_task *tctx); |
de0617e4 | 153 | |
2b188cc1 JA |
154 | static struct kmem_cache *req_cachep; |
155 | ||
2b188cc1 JA |
156 | struct sock *io_uring_get_socket(struct file *file) |
157 | { | |
158 | #if defined(CONFIG_UNIX) | |
cd40cae2 | 159 | if (io_is_uring_fops(file)) { |
2b188cc1 JA |
160 | struct io_ring_ctx *ctx = file->private_data; |
161 | ||
162 | return ctx->ring_sock->sk; | |
163 | } | |
164 | #endif | |
165 | return NULL; | |
166 | } | |
167 | EXPORT_SYMBOL(io_uring_get_socket); | |
168 | ||
c450178d PB |
169 | static inline void io_submit_flush_completions(struct io_ring_ctx *ctx) |
170 | { | |
931147dd DY |
171 | if (!wq_list_empty(&ctx->submit_state.compl_reqs) || |
172 | ctx->submit_state.cqes_count) | |
c450178d PB |
173 | __io_submit_flush_completions(ctx); |
174 | } | |
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; | |
f237c30a | 248 | bool locked = false; |
f56165e6 PB |
249 | |
250 | percpu_ref_get(&ctx->refs); | |
3218e5d3 | 251 | llist_for_each_entry_safe(req, tmp, node, io_task_work.node) |
f237c30a | 252 | req->io_task_work.func(req, &locked); |
5636c00d | 253 | |
f237c30a | 254 | if (locked) { |
c450178d | 255 | io_submit_flush_completions(ctx); |
f237c30a PB |
256 | mutex_unlock(&ctx->uring_lock); |
257 | } | |
f56165e6 PB |
258 | percpu_ref_put(&ctx->refs); |
259 | } | |
260 | ||
e6f89be6 PB |
261 | static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) |
262 | { | |
263 | unsigned hash_buckets = 1U << bits; | |
264 | size_t hash_size = hash_buckets * sizeof(table->hbs[0]); | |
265 | ||
266 | table->hbs = kmalloc(hash_size, GFP_KERNEL); | |
267 | if (!table->hbs) | |
268 | return -ENOMEM; | |
269 | ||
270 | table->hash_bits = bits; | |
271 | init_hash_table(table, hash_buckets); | |
272 | return 0; | |
273 | } | |
274 | ||
c072481d | 275 | static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
2b188cc1 JA |
276 | { |
277 | struct io_ring_ctx *ctx; | |
9cfc7e94 | 278 | int hash_bits; |
2b188cc1 JA |
279 | |
280 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
281 | if (!ctx) | |
282 | return NULL; | |
283 | ||
9cfc7e94 JA |
284 | xa_init(&ctx->io_bl_xa); |
285 | ||
78076bb6 JA |
286 | /* |
287 | * Use 5 bits less than the max cq entries, that should give us around | |
4a07723f PB |
288 | * 32 entries per hash list if totally full and uniformly spread, but |
289 | * don't keep too many buckets to not overconsume memory. | |
78076bb6 | 290 | */ |
4a07723f PB |
291 | hash_bits = ilog2(p->cq_entries) - 5; |
292 | hash_bits = clamp(hash_bits, 1, 8); | |
e6f89be6 | 293 | if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) |
78076bb6 | 294 | goto err; |
9ca9fb24 PB |
295 | if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) |
296 | goto err; | |
38513c46 | 297 | |
6224843d PB |
298 | ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL); |
299 | if (!ctx->dummy_ubuf) | |
300 | goto err; | |
301 | /* set invalid range, so io_import_fixed() fails meeting it */ | |
302 | ctx->dummy_ubuf->ubuf = -1UL; | |
303 | ||
21482896 | 304 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
48904229 | 305 | 0, GFP_KERNEL)) |
206aefde | 306 | goto err; |
2b188cc1 JA |
307 | |
308 | ctx->flags = p->flags; | |
90554200 | 309 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 310 | INIT_LIST_HEAD(&ctx->sqd_list); |
1d7bb1d5 | 311 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
cc3cec83 | 312 | INIT_LIST_HEAD(&ctx->io_buffers_cache); |
9b797a37 | 313 | io_alloc_cache_init(&ctx->apoll_cache); |
43e0bbbd | 314 | io_alloc_cache_init(&ctx->netmsg_cache); |
0f158b4c | 315 | init_completion(&ctx->ref_comp); |
61cf9370 | 316 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 317 | mutex_init(&ctx->uring_lock); |
311997b3 | 318 | init_waitqueue_head(&ctx->cq_wait); |
7b235dd8 | 319 | init_waitqueue_head(&ctx->poll_wq); |
2b188cc1 | 320 | spin_lock_init(&ctx->completion_lock); |
89850fce | 321 | spin_lock_init(&ctx->timeout_lock); |
5eef4e87 | 322 | INIT_WQ_LIST(&ctx->iopoll_list); |
cc3cec83 JA |
323 | INIT_LIST_HEAD(&ctx->io_buffers_pages); |
324 | INIT_LIST_HEAD(&ctx->io_buffers_comp); | |
de0617e4 | 325 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 326 | INIT_LIST_HEAD(&ctx->timeout_list); |
ef9dd637 | 327 | INIT_LIST_HEAD(&ctx->ltimeout_list); |
d67d2263 BM |
328 | spin_lock_init(&ctx->rsrc_ref_lock); |
329 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); | |
269bbe5f | 330 | INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work); |
d34b1b0b | 331 | init_task_work(&ctx->rsrc_put_tw, io_rsrc_put_tw); |
269bbe5f | 332 | init_llist_head(&ctx->rsrc_put_llist); |
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); | |
9011bf9a | 337 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
6f33b0bc | 338 | INIT_WQ_LIST(&ctx->submit_state.compl_reqs); |
2b188cc1 | 339 | return ctx; |
206aefde | 340 | err: |
6224843d | 341 | kfree(ctx->dummy_ubuf); |
e6f89be6 | 342 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 343 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 JA |
344 | kfree(ctx->io_bl); |
345 | xa_destroy(&ctx->io_bl_xa); | |
206aefde JA |
346 | kfree(ctx); |
347 | return NULL; | |
2b188cc1 JA |
348 | } |
349 | ||
8f6ed49a PB |
350 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
351 | { | |
352 | struct io_rings *r = ctx->rings; | |
353 | ||
354 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
355 | ctx->cq_extra--; | |
356 | } | |
357 | ||
9cf7c104 | 358 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 359 | { |
2bc9930e JA |
360 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
361 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 362 | |
8f6ed49a | 363 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 364 | } |
de0617e4 | 365 | |
9d858b21 | 366 | return false; |
de0617e4 JA |
367 | } |
368 | ||
9cae36a0 JA |
369 | static inline void io_req_track_inflight(struct io_kiocb *req) |
370 | { | |
371 | if (!(req->flags & REQ_F_INFLIGHT)) { | |
372 | req->flags |= REQ_F_INFLIGHT; | |
386e4fb6 | 373 | atomic_inc(&req->task->io_uring->inflight_tracked); |
9cae36a0 JA |
374 | } |
375 | } | |
376 | ||
fd08e530 PB |
377 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
378 | { | |
906c6caa PB |
379 | if (WARN_ON_ONCE(!req->link)) |
380 | return NULL; | |
381 | ||
4d13d1a4 PB |
382 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
383 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
384 | |
385 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 386 | io_req_set_refcount(req); |
4d13d1a4 PB |
387 | __io_req_set_refcount(req->link, 2); |
388 | return req->link; | |
fd08e530 PB |
389 | } |
390 | ||
391 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
392 | { | |
4d13d1a4 | 393 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
394 | return NULL; |
395 | return __io_prep_linked_timeout(req); | |
396 | } | |
397 | ||
cb2d344c PB |
398 | static noinline void __io_arm_ltimeout(struct io_kiocb *req) |
399 | { | |
400 | io_queue_linked_timeout(__io_prep_linked_timeout(req)); | |
401 | } | |
402 | ||
403 | static inline void io_arm_ltimeout(struct io_kiocb *req) | |
404 | { | |
405 | if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) | |
406 | __io_arm_ltimeout(req); | |
407 | } | |
408 | ||
1e6fa521 JA |
409 | static void io_prep_async_work(struct io_kiocb *req) |
410 | { | |
411 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
1e6fa521 JA |
412 | struct io_ring_ctx *ctx = req->ctx; |
413 | ||
b8e64b53 PB |
414 | if (!(req->flags & REQ_F_CREDS)) { |
415 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 416 | req->creds = get_current_cred(); |
b8e64b53 | 417 | } |
003e8dcc | 418 | |
e1d675df PB |
419 | req->work.list.next = NULL; |
420 | req->work.flags = 0; | |
8e29da69 | 421 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); |
feaadc4f PB |
422 | if (req->flags & REQ_F_FORCE_ASYNC) |
423 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
424 | ||
f6b543fd JA |
425 | if (req->file && !io_req_ffs_set(req)) |
426 | req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT; | |
427 | ||
1e6fa521 JA |
428 | if (req->flags & REQ_F_ISREG) { |
429 | if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL)) | |
430 | io_wq_hash_work(&req->work, file_inode(req->file)); | |
4b982bd0 | 431 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
432 | if (def->unbound_nonreg_file) |
433 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
434 | } | |
561fb04a | 435 | } |
cccf0ee8 | 436 | |
cbdcb435 | 437 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 438 | { |
cbdcb435 | 439 | struct io_kiocb *cur; |
54a91f3b | 440 | |
44eff40a PB |
441 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
442 | struct io_ring_ctx *ctx = req->ctx; | |
443 | ||
674ee8e1 | 444 | spin_lock_irq(&ctx->timeout_lock); |
44eff40a PB |
445 | io_for_each_link(cur, req) |
446 | io_prep_async_work(cur); | |
674ee8e1 | 447 | spin_unlock_irq(&ctx->timeout_lock); |
44eff40a PB |
448 | } else { |
449 | io_for_each_link(cur, req) | |
450 | io_prep_async_work(cur); | |
451 | } | |
561fb04a JA |
452 | } |
453 | ||
f3b44f92 | 454 | void io_queue_iowq(struct io_kiocb *req, bool *dont_use) |
561fb04a | 455 | { |
cbdcb435 | 456 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 457 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 458 | |
3bfe6106 JA |
459 | BUG_ON(!tctx); |
460 | BUG_ON(!tctx->io_wq); | |
561fb04a | 461 | |
cbdcb435 PB |
462 | /* init ->work of the whole link before punting */ |
463 | io_prep_async_link(req); | |
991468dc JA |
464 | |
465 | /* | |
466 | * Not expected to happen, but if we do have a bug where this _can_ | |
467 | * happen, catch it here and ensure the request is marked as | |
468 | * canceled. That will make io-wq go through the usual work cancel | |
469 | * procedure rather than attempt to run this request (or create a new | |
470 | * worker for it). | |
471 | */ | |
472 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
473 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
474 | ||
48863ffd | 475 | trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); |
ebf93667 | 476 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
477 | if (link) |
478 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
479 | } |
480 | ||
c072481d | 481 | static __cold void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 482 | { |
441b8a78 | 483 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
484 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
485 | struct io_defer_entry, list); | |
de0617e4 | 486 | |
9cf7c104 | 487 | if (req_need_defer(de->req, de->seq)) |
04518945 | 488 | break; |
27dc8338 | 489 | list_del_init(&de->list); |
907d1df3 | 490 | io_req_task_queue(de->req); |
27dc8338 | 491 | kfree(de); |
441b8a78 | 492 | } |
04518945 PB |
493 | } |
494 | ||
21a091b9 DY |
495 | |
496 | static void io_eventfd_ops(struct rcu_head *rcu) | |
f2842ab5 | 497 | { |
d8e9214f | 498 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
21a091b9 | 499 | int ops = atomic_xchg(&ev_fd->ops, 0); |
305bef98 | 500 | |
21a091b9 | 501 | if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) |
44648532 | 502 | eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); |
d8e9214f | 503 | |
21a091b9 DY |
504 | /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback |
505 | * ordering in a race but if references are 0 we know we have to free | |
506 | * it regardless. | |
305bef98 | 507 | */ |
21a091b9 DY |
508 | if (atomic_dec_and_test(&ev_fd->refs)) { |
509 | eventfd_ctx_put(ev_fd->cq_ev_fd); | |
510 | kfree(ev_fd); | |
511 | } | |
d8e9214f DY |
512 | } |
513 | ||
77bc59b4 | 514 | static void io_eventfd_signal(struct io_ring_ctx *ctx) |
f2842ab5 | 515 | { |
21a091b9 | 516 | struct io_ev_fd *ev_fd = NULL; |
77bc59b4 | 517 | |
77bc59b4 UA |
518 | rcu_read_lock(); |
519 | /* | |
520 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
521 | * and eventfd_signal | |
522 | */ | |
523 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
524 | ||
525 | /* | |
526 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
527 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
528 | * the function and rcu_read_lock. | |
529 | */ | |
530 | if (unlikely(!ev_fd)) | |
531 | goto out; | |
7e55a19c | 532 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 | 533 | goto out; |
21a091b9 DY |
534 | if (ev_fd->eventfd_async && !io_wq_current_is_worker()) |
535 | goto out; | |
77bc59b4 | 536 | |
21a091b9 | 537 | if (likely(eventfd_signal_allowed())) { |
44648532 | 538 | eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); |
21a091b9 DY |
539 | } else { |
540 | atomic_inc(&ev_fd->refs); | |
541 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) | |
44a84da4 | 542 | call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops); |
21a091b9 DY |
543 | else |
544 | atomic_dec(&ev_fd->refs); | |
545 | } | |
546 | ||
77bc59b4 UA |
547 | out: |
548 | rcu_read_unlock(); | |
f2842ab5 JA |
549 | } |
550 | ||
21a091b9 DY |
551 | static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) |
552 | { | |
553 | bool skip; | |
554 | ||
555 | spin_lock(&ctx->completion_lock); | |
556 | ||
557 | /* | |
558 | * Eventfd should only get triggered when at least one event has been | |
559 | * posted. Some applications rely on the eventfd notification count | |
560 | * only changing IFF a new CQE has been added to the CQ ring. There's | |
561 | * no depedency on 1:1 relationship between how many times this | |
562 | * function is called (and hence the eventfd count) and number of CQEs | |
563 | * posted to the CQ ring. | |
564 | */ | |
565 | skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
566 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
567 | spin_unlock(&ctx->completion_lock); | |
568 | if (skip) | |
569 | return; | |
570 | ||
571 | io_eventfd_signal(ctx); | |
572 | } | |
573 | ||
a830ffd2 PB |
574 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
575 | { | |
bca39f39 PB |
576 | if (ctx->poll_activated) |
577 | io_poll_wq_wake(ctx); | |
e5f30f6f PB |
578 | if (ctx->off_timeout_used) |
579 | io_flush_timeouts(ctx); | |
580 | if (ctx->drain_active) { | |
a830ffd2 | 581 | spin_lock(&ctx->completion_lock); |
e5f30f6f | 582 | io_queue_deferred(ctx); |
a830ffd2 PB |
583 | spin_unlock(&ctx->completion_lock); |
584 | } | |
585 | if (ctx->has_evfd) | |
21a091b9 | 586 | io_eventfd_flush_signal(ctx); |
a830ffd2 PB |
587 | } |
588 | ||
f66f7342 PB |
589 | static inline void __io_cq_lock(struct io_ring_ctx *ctx) |
590 | __acquires(ctx->completion_lock) | |
591 | { | |
592 | if (!ctx->task_complete) | |
593 | spin_lock(&ctx->completion_lock); | |
594 | } | |
595 | ||
596 | static inline void __io_cq_unlock(struct io_ring_ctx *ctx) | |
597 | { | |
598 | if (!ctx->task_complete) | |
599 | spin_unlock(&ctx->completion_lock); | |
600 | } | |
601 | ||
6971253f PB |
602 | static inline void io_cq_lock(struct io_ring_ctx *ctx) |
603 | __acquires(ctx->completion_lock) | |
604 | { | |
605 | spin_lock(&ctx->completion_lock); | |
606 | } | |
607 | ||
608 | static inline void io_cq_unlock(struct io_ring_ctx *ctx) | |
609 | __releases(ctx->completion_lock) | |
610 | { | |
611 | spin_unlock(&ctx->completion_lock); | |
612 | } | |
613 | ||
5d772916 | 614 | /* keep it inlined for io_submit_flush_completions() */ |
f66f7342 | 615 | static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) |
25399321 PB |
616 | __releases(ctx->completion_lock) |
617 | { | |
618 | io_commit_cqring(ctx); | |
f66f7342 | 619 | __io_cq_unlock(ctx); |
6c16fe3c JA |
620 | io_commit_cqring_flush(ctx); |
621 | io_cqring_wake(ctx); | |
25399321 PB |
622 | } |
623 | ||
3181e22f PB |
624 | static inline void __io_cq_unlock_post_flush(struct io_ring_ctx *ctx) |
625 | __releases(ctx->completion_lock) | |
626 | { | |
627 | io_commit_cqring(ctx); | |
628 | __io_cq_unlock(ctx); | |
629 | io_commit_cqring_flush(ctx); | |
630 | ||
631 | /* | |
632 | * As ->task_complete implies that the ring is single tasked, cq_wait | |
633 | * may only be waited on by the current in io_cqring_wait(), but since | |
634 | * it will re-check the wakeup conditions once we return we can safely | |
635 | * skip waking it up. | |
636 | */ | |
637 | if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) { | |
638 | smp_mb(); | |
639 | __io_cqring_wake(ctx); | |
640 | } | |
641 | } | |
642 | ||
25399321 | 643 | void io_cq_unlock_post(struct io_ring_ctx *ctx) |
5d772916 | 644 | __releases(ctx->completion_lock) |
25399321 | 645 | { |
f66f7342 PB |
646 | io_commit_cqring(ctx); |
647 | spin_unlock(&ctx->completion_lock); | |
648 | io_commit_cqring_flush(ctx); | |
649 | io_cqring_wake(ctx); | |
25399321 PB |
650 | } |
651 | ||
c4a2ed72 | 652 | /* Returns true if there are no backlogged entries after the flush */ |
a85381d8 PB |
653 | static void io_cqring_overflow_kill(struct io_ring_ctx *ctx) |
654 | { | |
655 | struct io_overflow_cqe *ocqe; | |
656 | LIST_HEAD(list); | |
657 | ||
658 | io_cq_lock(ctx); | |
659 | list_splice_init(&ctx->cq_overflow_list, &list); | |
660 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); | |
661 | io_cq_unlock(ctx); | |
662 | ||
663 | while (!list_empty(&list)) { | |
664 | ocqe = list_first_entry(&list, struct io_overflow_cqe, list); | |
665 | list_del(&ocqe->list); | |
666 | kfree(ocqe); | |
667 | } | |
668 | } | |
669 | ||
670 | /* Returns true if there are no backlogged entries after the flush */ | |
1b346e4a | 671 | static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
1d7bb1d5 | 672 | { |
e45a3e05 | 673 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 674 | |
a85381d8 | 675 | if (__io_cqring_events(ctx) == ctx->cq_entries) |
1b346e4a | 676 | return; |
1d7bb1d5 | 677 | |
e45a3e05 SR |
678 | if (ctx->flags & IORING_SETUP_CQE32) |
679 | cqe_size <<= 1; | |
680 | ||
25399321 | 681 | io_cq_lock(ctx); |
6c2450ae | 682 | while (!list_empty(&ctx->cq_overflow_list)) { |
aa1df3a3 | 683 | struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true); |
6c2450ae | 684 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 685 | |
a85381d8 | 686 | if (!cqe) |
1d7bb1d5 | 687 | break; |
6c2450ae PB |
688 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
689 | struct io_overflow_cqe, list); | |
a85381d8 | 690 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae PB |
691 | list_del(&ocqe->list); |
692 | kfree(ocqe); | |
1d7bb1d5 JA |
693 | } |
694 | ||
1b346e4a | 695 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 696 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 697 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 698 | } |
25399321 | 699 | io_cq_unlock_post(ctx); |
1d7bb1d5 JA |
700 | } |
701 | ||
52ea806a JA |
702 | static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx) |
703 | { | |
704 | /* iopoll syncs against uring_lock, not completion_lock */ | |
705 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
706 | mutex_lock(&ctx->uring_lock); | |
707 | __io_cqring_overflow_flush(ctx); | |
708 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
709 | mutex_unlock(&ctx->uring_lock); | |
710 | } | |
711 | ||
1b346e4a | 712 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 713 | { |
52ea806a JA |
714 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) |
715 | io_cqring_do_overflow_flush(ctx); | |
6c503150 PB |
716 | } |
717 | ||
e70cb608 | 718 | void __io_put_task(struct task_struct *task, int nr) |
6a290a14 PB |
719 | { |
720 | struct io_uring_task *tctx = task->io_uring; | |
721 | ||
9d170164 PB |
722 | percpu_counter_sub(&tctx->inflight, nr); |
723 | if (unlikely(atomic_read(&tctx->in_idle))) | |
724 | wake_up(&tctx->wait); | |
725 | put_task_struct_many(task, nr); | |
726 | } | |
727 | ||
63809137 | 728 | void io_task_refs_refill(struct io_uring_task *tctx) |
9a10867a PB |
729 | { |
730 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
731 | ||
732 | percpu_counter_add(&tctx->inflight, refill); | |
733 | refcount_add(refill, ¤t->usage); | |
734 | tctx->cached_refs += refill; | |
735 | } | |
736 | ||
3cc7fdb9 PB |
737 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
738 | { | |
739 | struct io_uring_task *tctx = task->io_uring; | |
740 | unsigned int refs = tctx->cached_refs; | |
741 | ||
742 | if (refs) { | |
743 | tctx->cached_refs = 0; | |
744 | percpu_counter_sub(&tctx->inflight, refs); | |
745 | put_task_struct_many(task, refs); | |
746 | } | |
747 | } | |
748 | ||
68494a65 PB |
749 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
750 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 751 | { |
cce4b8b0 | 752 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
753 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
754 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 755 | |
f26cc959 PB |
756 | lockdep_assert_held(&ctx->completion_lock); |
757 | ||
e45a3e05 SR |
758 | if (is_cqe32) |
759 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 760 | |
e45a3e05 | 761 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 762 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
763 | if (!ocqe) { |
764 | /* | |
765 | * If we're in ring overflow flush mode, or in task cancel mode, | |
766 | * or cannot allocate an overflow entry, then we need to drop it | |
767 | * on the floor. | |
768 | */ | |
8f6ed49a | 769 | io_account_cq_overflow(ctx); |
155bc950 | 770 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 771 | return false; |
2b188cc1 | 772 | } |
cce4b8b0 | 773 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 774 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 775 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 776 | |
cce4b8b0 | 777 | } |
d4d19c19 | 778 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
779 | ocqe->cqe.res = res; |
780 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
781 | if (is_cqe32) { |
782 | ocqe->cqe.big_cqe[0] = extra1; | |
783 | ocqe->cqe.big_cqe[1] = extra2; | |
784 | } | |
cce4b8b0 PB |
785 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
786 | return true; | |
2b188cc1 JA |
787 | } |
788 | ||
68494a65 PB |
789 | bool io_req_cqe_overflow(struct io_kiocb *req) |
790 | { | |
791 | if (!(req->flags & REQ_F_CQE32_INIT)) { | |
792 | req->extra1 = 0; | |
793 | req->extra2 = 0; | |
794 | } | |
795 | return io_cqring_event_overflow(req->ctx, req->cqe.user_data, | |
796 | req->cqe.res, req->cqe.flags, | |
797 | req->extra1, req->extra2); | |
798 | } | |
799 | ||
faf88dde PB |
800 | /* |
801 | * writes to the cq entry need to come after reading head; the | |
802 | * control dependency is enough as we're using WRITE_ONCE to | |
803 | * fill the cq entry | |
804 | */ | |
aa1df3a3 | 805 | struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow) |
faf88dde PB |
806 | { |
807 | struct io_rings *rings = ctx->rings; | |
808 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
809 | unsigned int free, queued, len; |
810 | ||
aa1df3a3 PB |
811 | /* |
812 | * Posting into the CQ when there are pending overflowed CQEs may break | |
813 | * ordering guarantees, which will affect links, F_MORE users and more. | |
814 | * Force overflow the completion. | |
815 | */ | |
816 | if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) | |
817 | return NULL; | |
faf88dde PB |
818 | |
819 | /* userspace may cheat modifying the tail, be safe and do min */ | |
820 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
821 | free = ctx->cq_entries - queued; | |
822 | /* we need a contiguous range, limit based on the current array offset */ | |
823 | len = min(free, ctx->cq_entries - off); | |
824 | if (!len) | |
825 | return NULL; | |
826 | ||
b3659a65 PB |
827 | if (ctx->flags & IORING_SETUP_CQE32) { |
828 | off <<= 1; | |
829 | len <<= 1; | |
830 | } | |
831 | ||
faf88dde PB |
832 | ctx->cqe_cached = &rings->cqes[off]; |
833 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
b3659a65 PB |
834 | |
835 | ctx->cached_cq_tail++; | |
faf88dde | 836 | ctx->cqe_cached++; |
b3659a65 PB |
837 | if (ctx->flags & IORING_SETUP_CQE32) |
838 | ctx->cqe_cached++; | |
839 | return &rings->cqes[off]; | |
faf88dde PB |
840 | } |
841 | ||
f66f7342 PB |
842 | static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, |
843 | u32 cflags) | |
bcda7baa | 844 | { |
cd94903d PB |
845 | struct io_uring_cqe *cqe; |
846 | ||
913a571a | 847 | ctx->cq_extra++; |
cd94903d PB |
848 | |
849 | /* | |
850 | * If we can't get a cq entry, userspace overflowed the | |
851 | * submission (by quite a lot). Increment the overflow count in | |
852 | * the ring. | |
853 | */ | |
854 | cqe = io_get_cqe(ctx); | |
855 | if (likely(cqe)) { | |
e0486f3f DY |
856 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
857 | ||
cd94903d PB |
858 | WRITE_ONCE(cqe->user_data, user_data); |
859 | WRITE_ONCE(cqe->res, res); | |
860 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
861 | |
862 | if (ctx->flags & IORING_SETUP_CQE32) { | |
863 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
864 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
865 | } | |
cd94903d PB |
866 | return true; |
867 | } | |
52120f0f | 868 | return false; |
bcda7baa JA |
869 | } |
870 | ||
931147dd DY |
871 | static void __io_flush_post_cqes(struct io_ring_ctx *ctx) |
872 | __must_hold(&ctx->uring_lock) | |
873 | { | |
874 | struct io_submit_state *state = &ctx->submit_state; | |
875 | unsigned int i; | |
876 | ||
877 | lockdep_assert_held(&ctx->uring_lock); | |
878 | for (i = 0; i < state->cqes_count; i++) { | |
879 | struct io_uring_cqe *cqe = &state->cqes[i]; | |
880 | ||
f66f7342 PB |
881 | if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) { |
882 | if (ctx->task_complete) { | |
883 | spin_lock(&ctx->completion_lock); | |
884 | io_cqring_event_overflow(ctx, cqe->user_data, | |
885 | cqe->res, cqe->flags, 0, 0); | |
886 | spin_unlock(&ctx->completion_lock); | |
887 | } else { | |
888 | io_cqring_event_overflow(ctx, cqe->user_data, | |
889 | cqe->res, cqe->flags, 0, 0); | |
890 | } | |
891 | } | |
931147dd DY |
892 | } |
893 | state->cqes_count = 0; | |
894 | } | |
895 | ||
b529c96a DY |
896 | static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, |
897 | bool allow_overflow) | |
d245bca6 PB |
898 | { |
899 | bool filled; | |
900 | ||
25399321 | 901 | io_cq_lock(ctx); |
f66f7342 PB |
902 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags); |
903 | if (!filled && allow_overflow) | |
904 | filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); | |
905 | ||
25399321 | 906 | io_cq_unlock_post(ctx); |
d245bca6 PB |
907 | return filled; |
908 | } | |
909 | ||
b529c96a DY |
910 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags) |
911 | { | |
912 | return __io_post_aux_cqe(ctx, user_data, res, cflags, true); | |
913 | } | |
914 | ||
9b8c5475 DY |
915 | bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags, |
916 | bool allow_overflow) | |
2b188cc1 | 917 | { |
9b8c5475 DY |
918 | struct io_uring_cqe *cqe; |
919 | unsigned int length; | |
920 | ||
921 | if (!defer) | |
b529c96a | 922 | return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow); |
9b8c5475 DY |
923 | |
924 | length = ARRAY_SIZE(ctx->submit_state.cqes); | |
925 | ||
926 | lockdep_assert_held(&ctx->uring_lock); | |
927 | ||
928 | if (ctx->submit_state.cqes_count == length) { | |
f66f7342 | 929 | __io_cq_lock(ctx); |
9b8c5475 DY |
930 | __io_flush_post_cqes(ctx); |
931 | /* no need to flush - flush is deferred */ | |
f66f7342 | 932 | __io_cq_unlock_post(ctx); |
9b8c5475 DY |
933 | } |
934 | ||
935 | /* For defered completions this is not as strict as it is otherwise, | |
936 | * however it's main job is to prevent unbounded posted completions, | |
937 | * and in that it works just as well. | |
938 | */ | |
939 | if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) | |
940 | return false; | |
941 | ||
942 | cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++]; | |
943 | cqe->user_data = user_data; | |
944 | cqe->res = res; | |
945 | cqe->flags = cflags; | |
946 | return true; | |
947 | } | |
948 | ||
1bec951c | 949 | static void __io_req_complete_post(struct io_kiocb *req) |
2b188cc1 | 950 | { |
fa18fa22 PB |
951 | struct io_ring_ctx *ctx = req->ctx; |
952 | ||
953 | io_cq_lock(ctx); | |
954 | if (!(req->flags & REQ_F_CQE_SKIP)) | |
a8cf95f9 | 955 | io_fill_cqe_req(ctx, req); |
fa18fa22 | 956 | |
c7dae4ba JA |
957 | /* |
958 | * If we're the last reference to this request, add to our locked | |
959 | * free_list cache. | |
960 | */ | |
de9b4cca | 961 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 962 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 963 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
964 | io_disarm_next(req); |
965 | if (req->link) { | |
966 | io_req_task_queue(req->link); | |
967 | req->link = NULL; | |
968 | } | |
969 | } | |
7ac1edc4 | 970 | io_req_put_rsrc(req); |
8197b053 PB |
971 | /* |
972 | * Selected buffer deallocation in io_clean_op() assumes that | |
973 | * we don't hold ->completion_lock. Clean them here to avoid | |
974 | * deadlocks. | |
975 | */ | |
976 | io_put_kbuf_comp(req); | |
c7dae4ba JA |
977 | io_dismantle_req(req); |
978 | io_put_task(req->task, 1); | |
c2b6c6bc | 979 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 980 | ctx->locked_free_nr++; |
180f829f | 981 | } |
25399321 | 982 | io_cq_unlock_post(ctx); |
4e3d9ff9 JA |
983 | } |
984 | ||
1bec951c | 985 | void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
bcda7baa | 986 | { |
e6aeb272 PB |
987 | if (req->ctx->task_complete && (issue_flags & IO_URING_F_IOWQ)) { |
988 | req->io_task_work.func = io_req_task_complete; | |
989 | io_req_task_work_add(req); | |
990 | } else if (!(issue_flags & IO_URING_F_UNLOCKED) || | |
991 | !(req->ctx->flags & IORING_SETUP_IOPOLL)) { | |
1bec951c PB |
992 | __io_req_complete_post(req); |
993 | } else { | |
994 | struct io_ring_ctx *ctx = req->ctx; | |
995 | ||
996 | mutex_lock(&ctx->uring_lock); | |
997 | __io_req_complete_post(req); | |
998 | mutex_unlock(&ctx->uring_lock); | |
999 | } | |
0ddf92e8 JA |
1000 | } |
1001 | ||
973fc83f | 1002 | void io_req_defer_failed(struct io_kiocb *req, s32 res) |
e276ae34 | 1003 | __must_hold(&ctx->uring_lock) |
f41db273 | 1004 | { |
a47b255e PB |
1005 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
1006 | ||
e276ae34 PB |
1007 | lockdep_assert_held(&req->ctx->uring_lock); |
1008 | ||
93d2bcd2 | 1009 | req_set_fail(req); |
97b388d7 | 1010 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
a47b255e PB |
1011 | if (def->fail) |
1012 | def->fail(req); | |
973fc83f | 1013 | io_req_complete_defer(req); |
f41db273 PB |
1014 | } |
1015 | ||
864ea921 PB |
1016 | /* |
1017 | * Don't initialise the fields below on every allocation, but do that in | |
1018 | * advance and keep them valid across allocations. | |
1019 | */ | |
1020 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
1021 | { | |
1022 | req->ctx = ctx; | |
1023 | req->link = NULL; | |
1024 | req->async_data = NULL; | |
1025 | /* not necessary, but safer to zero */ | |
cef216fc | 1026 | req->cqe.res = 0; |
864ea921 PB |
1027 | } |
1028 | ||
dac7a098 | 1029 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 1030 | struct io_submit_state *state) |
dac7a098 | 1031 | { |
79ebeaee | 1032 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1033 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 1034 | ctx->locked_free_nr = 0; |
79ebeaee | 1035 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
1036 | } |
1037 | ||
5d5901a3 PB |
1038 | /* |
1039 | * A request might get retired back into the request caches even before opcode | |
1040 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
1041 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
1042 | * and with extra caution to not get a request that is still worked on. | |
1043 | */ | |
bd1a3783 | 1044 | __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 1045 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 1046 | { |
864ea921 | 1047 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 1048 | void *reqs[IO_REQ_ALLOC_BATCH]; |
864ea921 | 1049 | int ret, i; |
e5d1bc0a | 1050 | |
23a5c43b PB |
1051 | /* |
1052 | * If we have more than a batch's worth of requests in our IRQ side | |
1053 | * locked cache, grab the lock and move them over to our submission | |
1054 | * side cache. | |
1055 | */ | |
a6d97a8a | 1056 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 1057 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 1058 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
1059 | return true; |
1060 | } | |
e5d1bc0a | 1061 | |
3ab665b7 | 1062 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 1063 | |
864ea921 PB |
1064 | /* |
1065 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1066 | * retry single alloc to be on the safe side. | |
1067 | */ | |
1068 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
1069 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
1070 | if (!reqs[0]) | |
a33ae9ce | 1071 | return false; |
864ea921 | 1072 | ret = 1; |
2b188cc1 | 1073 | } |
864ea921 | 1074 | |
37f0e767 | 1075 | percpu_ref_get_many(&ctx->refs, ret); |
3ab665b7 | 1076 | for (i = 0; i < ret; i++) { |
23a5c43b | 1077 | struct io_kiocb *req = reqs[i]; |
3ab665b7 PB |
1078 | |
1079 | io_preinit_req(req, ctx); | |
fa05457a | 1080 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 1081 | } |
a33ae9ce PB |
1082 | return true; |
1083 | } | |
1084 | ||
6b639522 | 1085 | static inline void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 1086 | { |
094bae49 | 1087 | unsigned int flags = req->flags; |
929a3af9 | 1088 | |
867f8fa5 | 1089 | if (unlikely(flags & IO_REQ_CLEAN_FLAGS)) |
3a0a6902 | 1090 | io_clean_op(req); |
e1d767f0 PB |
1091 | if (!(flags & REQ_F_FIXED_FILE)) |
1092 | io_put_file(req->file); | |
e65ef56d JA |
1093 | } |
1094 | ||
59915143 | 1095 | __cold void io_free_req(struct io_kiocb *req) |
c6ca97b3 | 1096 | { |
51a4cc11 | 1097 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 1098 | |
7ac1edc4 | 1099 | io_req_put_rsrc(req); |
216578e5 | 1100 | io_dismantle_req(req); |
7c660731 | 1101 | io_put_task(req->task, 1); |
c6ca97b3 | 1102 | |
79ebeaee | 1103 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1104 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
c34b025f | 1105 | ctx->locked_free_nr++; |
79ebeaee | 1106 | spin_unlock(&ctx->completion_lock); |
e65ef56d JA |
1107 | } |
1108 | ||
d81499bf PB |
1109 | static void __io_req_find_next_prep(struct io_kiocb *req) |
1110 | { | |
1111 | struct io_ring_ctx *ctx = req->ctx; | |
d81499bf | 1112 | |
6971253f | 1113 | spin_lock(&ctx->completion_lock); |
305bef98 | 1114 | io_disarm_next(req); |
6971253f | 1115 | spin_unlock(&ctx->completion_lock); |
d81499bf PB |
1116 | } |
1117 | ||
1118 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 1119 | { |
33cc89a9 | 1120 | struct io_kiocb *nxt; |
944e58bf | 1121 | |
9e645e11 JA |
1122 | /* |
1123 | * If LINK is set, we have dependent requests in this chain. If we | |
1124 | * didn't fail this request, queue the first one up, moving any other | |
1125 | * dependencies to the next request. In case of failure, fail the rest | |
1126 | * of the chain. | |
1127 | */ | |
d81499bf PB |
1128 | if (unlikely(req->flags & IO_DISARM_MASK)) |
1129 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
1130 | nxt = req->link; |
1131 | req->link = NULL; | |
1132 | return nxt; | |
4d7dd462 | 1133 | } |
9e645e11 | 1134 | |
f237c30a | 1135 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked) |
2c32395d PB |
1136 | { |
1137 | if (!ctx) | |
1138 | return; | |
ef060ea9 JA |
1139 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1140 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
f237c30a | 1141 | if (*locked) { |
c450178d | 1142 | io_submit_flush_completions(ctx); |
2c32395d | 1143 | mutex_unlock(&ctx->uring_lock); |
f237c30a | 1144 | *locked = false; |
2c32395d PB |
1145 | } |
1146 | percpu_ref_put(&ctx->refs); | |
1147 | } | |
1148 | ||
c6dd763c DY |
1149 | static unsigned int handle_tw_list(struct llist_node *node, |
1150 | struct io_ring_ctx **ctx, bool *locked, | |
1151 | struct llist_node *last) | |
9f8d032a | 1152 | { |
c6dd763c DY |
1153 | unsigned int count = 0; |
1154 | ||
3a0c037b | 1155 | while (node != last) { |
f88262e6 | 1156 | struct llist_node *next = node->next; |
9f8d032a HX |
1157 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1158 | io_task_work.node); | |
1159 | ||
34d2bfe7 JA |
1160 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); |
1161 | ||
9f8d032a HX |
1162 | if (req->ctx != *ctx) { |
1163 | ctx_flush_and_put(*ctx, locked); | |
1164 | *ctx = req->ctx; | |
1165 | /* if not contended, grab and improve batching */ | |
1166 | *locked = mutex_trylock(&(*ctx)->uring_lock); | |
1167 | percpu_ref_get(&(*ctx)->refs); | |
1168 | } | |
1169 | req->io_task_work.func(req, locked); | |
1170 | node = next; | |
c6dd763c | 1171 | count++; |
3a0c037b | 1172 | } |
c6dd763c DY |
1173 | |
1174 | return count; | |
9f8d032a HX |
1175 | } |
1176 | ||
923d1592 DY |
1177 | /** |
1178 | * io_llist_xchg - swap all entries in a lock-less list | |
1179 | * @head: the head of lock-less list to delete all entries | |
1180 | * @new: new entry as the head of the list | |
1181 | * | |
1182 | * If list is empty, return NULL, otherwise, return the pointer to the first entry. | |
1183 | * The order of entries returned is from the newest to the oldest added one. | |
1184 | */ | |
1185 | static inline struct llist_node *io_llist_xchg(struct llist_head *head, | |
1186 | struct llist_node *new) | |
1187 | { | |
1188 | return xchg(&head->first, new); | |
1189 | } | |
1190 | ||
1191 | /** | |
1192 | * io_llist_cmpxchg - possibly swap all entries in a lock-less list | |
1193 | * @head: the head of lock-less list to delete all entries | |
1194 | * @old: expected old value of the first entry of the list | |
1195 | * @new: new entry as the head of the list | |
1196 | * | |
1197 | * perform a cmpxchg on the first entry of the list. | |
1198 | */ | |
1199 | ||
1200 | static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head, | |
1201 | struct llist_node *old, | |
1202 | struct llist_node *new) | |
1203 | { | |
1204 | return cmpxchg(&head->first, old, new); | |
1205 | } | |
1206 | ||
c9f06aa7 | 1207 | void tctx_task_work(struct callback_head *cb) |
c40f6379 | 1208 | { |
f28c240e | 1209 | bool uring_locked = false; |
ebd0df2e | 1210 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
1211 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
1212 | task_work); | |
3a0c037b | 1213 | struct llist_node fake = {}; |
77e443ab | 1214 | struct llist_node *node; |
c6dd763c | 1215 | unsigned int loops = 1; |
77e443ab | 1216 | unsigned int count; |
3a0c037b | 1217 | |
77e443ab PB |
1218 | if (unlikely(current->flags & PF_EXITING)) { |
1219 | io_fallback_tw(tctx); | |
1220 | return; | |
1221 | } | |
3a0c037b | 1222 | |
77e443ab PB |
1223 | node = io_llist_xchg(&tctx->task_list, &fake); |
1224 | count = handle_tw_list(node, &ctx, &uring_locked, NULL); | |
3a0c037b DY |
1225 | node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); |
1226 | while (node != &fake) { | |
c6dd763c | 1227 | loops++; |
3a0c037b | 1228 | node = io_llist_xchg(&tctx->task_list, &fake); |
c6dd763c | 1229 | count += handle_tw_list(node, &ctx, &uring_locked, &fake); |
3a0c037b | 1230 | node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); |
3f18407d | 1231 | } |
ebd0df2e | 1232 | |
f28c240e | 1233 | ctx_flush_and_put(ctx, &uring_locked); |
3cc7fdb9 PB |
1234 | |
1235 | /* relaxed read is enough as only the task itself sets ->in_idle */ | |
1236 | if (unlikely(atomic_read(&tctx->in_idle))) | |
1237 | io_uring_drop_tctx_refs(current); | |
c6dd763c DY |
1238 | |
1239 | trace_io_uring_task_work_run(tctx, count, loops); | |
7cbf1722 JA |
1240 | } |
1241 | ||
d7593606 PB |
1242 | static __cold void io_fallback_tw(struct io_uring_task *tctx) |
1243 | { | |
1244 | struct llist_node *node = llist_del_all(&tctx->task_list); | |
1245 | struct io_kiocb *req; | |
1246 | ||
1247 | while (node) { | |
1248 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1249 | node = node->next; | |
1250 | if (llist_add(&req->io_task_work.node, | |
1251 | &req->ctx->fallback_llist)) | |
1252 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1253 | } | |
1254 | } | |
1255 | ||
c0e0d6ba DY |
1256 | static void io_req_local_work_add(struct io_kiocb *req) |
1257 | { | |
1258 | struct io_ring_ctx *ctx = req->ctx; | |
1259 | ||
9ffa13ff PB |
1260 | percpu_ref_get(&ctx->refs); |
1261 | ||
1262 | if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) { | |
1263 | percpu_ref_put(&ctx->refs); | |
c0e0d6ba | 1264 | return; |
9ffa13ff | 1265 | } |
130bd686 | 1266 | /* needed for the following wake up */ |
fc86f9d3 | 1267 | smp_mb__after_atomic(); |
c0e0d6ba DY |
1268 | |
1269 | if (unlikely(atomic_read(&req->task->io_uring->in_idle))) { | |
1270 | io_move_task_work_from_local(ctx); | |
9ffa13ff | 1271 | percpu_ref_put(&ctx->refs); |
c0e0d6ba DY |
1272 | return; |
1273 | } | |
1274 | ||
1275 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) | |
1276 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
21a091b9 DY |
1277 | if (ctx->has_evfd) |
1278 | io_eventfd_signal(ctx); | |
130bd686 | 1279 | wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE); |
9ffa13ff | 1280 | percpu_ref_put(&ctx->refs); |
c0e0d6ba DY |
1281 | } |
1282 | ||
e52d2e58 | 1283 | void __io_req_task_work_add(struct io_kiocb *req, bool allow_local) |
7cbf1722 | 1284 | { |
c34398a8 | 1285 | struct io_uring_task *tctx = req->task->io_uring; |
9f010507 | 1286 | struct io_ring_ctx *ctx = req->ctx; |
7cbf1722 | 1287 | |
c0e0d6ba DY |
1288 | if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
1289 | io_req_local_work_add(req); | |
1290 | return; | |
1291 | } | |
7cbf1722 JA |
1292 | |
1293 | /* task_work already pending, we're done */ | |
32d91f05 | 1294 | if (!llist_add(&req->io_task_work.node, &tctx->task_list)) |
e09ee510 | 1295 | return; |
7cbf1722 | 1296 | |
ef060ea9 JA |
1297 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1298 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1299 | ||
3fe07bcd | 1300 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1301 | return; |
2215bed9 | 1302 | |
d7593606 | 1303 | io_fallback_tw(tctx); |
c0e0d6ba DY |
1304 | } |
1305 | ||
1306 | static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) | |
1307 | { | |
1308 | struct llist_node *node; | |
1309 | ||
1310 | node = llist_del_all(&ctx->work_llist); | |
1311 | while (node) { | |
1312 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1313 | io_task_work.node); | |
1314 | ||
1315 | node = node->next; | |
1316 | __io_req_task_work_add(req, false); | |
1317 | } | |
1318 | } | |
1319 | ||
360173ab | 1320 | static int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked) |
c0e0d6ba | 1321 | { |
c0e0d6ba DY |
1322 | struct llist_node *node; |
1323 | struct llist_node fake; | |
1324 | struct llist_node *current_final = NULL; | |
140102ae | 1325 | int ret = 0; |
f75d5036 | 1326 | unsigned int loops = 1; |
c0e0d6ba | 1327 | |
140102ae | 1328 | if (WARN_ON_ONCE(ctx->submitter_task != current)) |
c0e0d6ba | 1329 | return -EEXIST; |
c0e0d6ba | 1330 | |
c0e0d6ba | 1331 | node = io_llist_xchg(&ctx->work_llist, &fake); |
c0e0d6ba DY |
1332 | again: |
1333 | while (node != current_final) { | |
1334 | struct llist_node *next = node->next; | |
1335 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1336 | io_task_work.node); | |
1337 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); | |
b3026767 | 1338 | req->io_task_work.func(req, locked); |
c0e0d6ba DY |
1339 | ret++; |
1340 | node = next; | |
1341 | } | |
1342 | ||
1343 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) | |
1344 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1345 | ||
1346 | node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL); | |
1347 | if (node != &fake) { | |
f75d5036 | 1348 | loops++; |
c0e0d6ba DY |
1349 | current_final = &fake; |
1350 | node = io_llist_xchg(&ctx->work_llist, &fake); | |
1351 | goto again; | |
1352 | } | |
1353 | ||
b3026767 | 1354 | if (*locked) |
c0e0d6ba | 1355 | io_submit_flush_completions(ctx); |
f75d5036 | 1356 | trace_io_uring_local_work_run(ctx, ret, loops); |
c0e0d6ba | 1357 | return ret; |
8ac5d85a JA |
1358 | |
1359 | } | |
1360 | ||
360173ab PB |
1361 | static inline int io_run_local_work_locked(struct io_ring_ctx *ctx) |
1362 | { | |
1363 | bool locked; | |
1364 | int ret; | |
1365 | ||
1366 | if (llist_empty(&ctx->work_llist)) | |
1367 | return 0; | |
1368 | ||
1369 | locked = true; | |
1370 | ret = __io_run_local_work(ctx, &locked); | |
1371 | /* shouldn't happen! */ | |
1372 | if (WARN_ON_ONCE(!locked)) | |
1373 | mutex_lock(&ctx->uring_lock); | |
1374 | return ret; | |
1375 | } | |
1376 | ||
3e565555 | 1377 | static int io_run_local_work(struct io_ring_ctx *ctx) |
8ac5d85a | 1378 | { |
2f413956 | 1379 | bool locked = mutex_trylock(&ctx->uring_lock); |
8ac5d85a JA |
1380 | int ret; |
1381 | ||
b3026767 | 1382 | ret = __io_run_local_work(ctx, &locked); |
8ac5d85a JA |
1383 | if (locked) |
1384 | mutex_unlock(&ctx->uring_lock); | |
1385 | ||
1386 | return ret; | |
c0e0d6ba DY |
1387 | } |
1388 | ||
f237c30a | 1389 | static void io_req_task_cancel(struct io_kiocb *req, bool *locked) |
c40f6379 | 1390 | { |
971cf9c1 | 1391 | io_tw_lock(req->ctx, locked); |
973fc83f | 1392 | io_req_defer_failed(req, req->cqe.res); |
c40f6379 JA |
1393 | } |
1394 | ||
329061d3 | 1395 | void io_req_task_submit(struct io_kiocb *req, bool *locked) |
c40f6379 | 1396 | { |
971cf9c1 | 1397 | io_tw_lock(req->ctx, locked); |
316319e8 | 1398 | /* req->task == current here, checking PF_EXITING is safe */ |
af066f31 | 1399 | if (likely(!(req->task->flags & PF_EXITING))) |
cbc2e203 | 1400 | io_queue_sqe(req); |
81b6d05c | 1401 | else |
973fc83f | 1402 | io_req_defer_failed(req, -EFAULT); |
c40f6379 JA |
1403 | } |
1404 | ||
59915143 | 1405 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1406 | { |
97b388d7 | 1407 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1408 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1409 | io_req_task_work_add(req); |
c40f6379 JA |
1410 | } |
1411 | ||
f3b44f92 | 1412 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1413 | { |
5b0a6acc | 1414 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1415 | io_req_task_work_add(req); |
a3df7698 PB |
1416 | } |
1417 | ||
59915143 | 1418 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1419 | { |
9b5f7bd9 | 1420 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1421 | |
1422 | if (nxt) | |
906a8c3f | 1423 | io_req_task_queue(nxt); |
c69f8dbe JL |
1424 | } |
1425 | ||
f3b44f92 | 1426 | void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node) |
3aa83bfb | 1427 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1428 | { |
d4b7a5ef | 1429 | struct task_struct *task = NULL; |
37f0e767 | 1430 | int task_refs = 0; |
5af1d13e | 1431 | |
3aa83bfb PB |
1432 | do { |
1433 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1434 | comp_list); | |
2d6500d4 | 1435 | |
a538be5b PB |
1436 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1437 | if (req->flags & REQ_F_REFCOUNT) { | |
1438 | node = req->comp_list.next; | |
1439 | if (!req_ref_put_and_test(req)) | |
1440 | continue; | |
1441 | } | |
b605a7fa PB |
1442 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1443 | struct async_poll *apoll = req->apoll; | |
1444 | ||
1445 | if (apoll->double_poll) | |
1446 | kfree(apoll->double_poll); | |
9731bc98 JA |
1447 | if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) |
1448 | kfree(apoll); | |
b605a7fa PB |
1449 | req->flags &= ~REQ_F_POLLED; |
1450 | } | |
da1a08c5 | 1451 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1452 | io_queue_next(req); |
a538be5b PB |
1453 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1454 | io_clean_op(req); | |
c1e53a69 | 1455 | } |
a538be5b PB |
1456 | if (!(req->flags & REQ_F_FIXED_FILE)) |
1457 | io_put_file(req->file); | |
2d6500d4 | 1458 | |
ab409402 | 1459 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1460 | |
d4b7a5ef PB |
1461 | if (req->task != task) { |
1462 | if (task) | |
1463 | io_put_task(task, task_refs); | |
1464 | task = req->task; | |
1465 | task_refs = 0; | |
1466 | } | |
1467 | task_refs++; | |
c1e53a69 | 1468 | node = req->comp_list.next; |
fa05457a | 1469 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1470 | } while (node); |
d4b7a5ef | 1471 | |
d4b7a5ef PB |
1472 | if (task) |
1473 | io_put_task(task, task_refs); | |
7a743e22 PB |
1474 | } |
1475 | ||
c450178d | 1476 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1477 | __must_hold(&ctx->uring_lock) |
905c172f | 1478 | { |
6f33b0bc | 1479 | struct io_wq_work_node *node, *prev; |
cd0ca2e0 | 1480 | struct io_submit_state *state = &ctx->submit_state; |
905c172f | 1481 | |
f66f7342 | 1482 | __io_cq_lock(ctx); |
931147dd DY |
1483 | /* must come first to preserve CQE ordering in failure cases */ |
1484 | if (state->cqes_count) | |
1485 | __io_flush_post_cqes(ctx); | |
d9dee430 PB |
1486 | wq_list_for_each(node, prev, &state->compl_reqs) { |
1487 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1488 | comp_list); | |
3d4aeb9f | 1489 | |
f66f7342 PB |
1490 | if (!(req->flags & REQ_F_CQE_SKIP) && |
1491 | unlikely(!__io_fill_cqe_req(ctx, req))) { | |
1492 | if (ctx->task_complete) { | |
1493 | spin_lock(&ctx->completion_lock); | |
1494 | io_req_cqe_overflow(req); | |
1495 | spin_unlock(&ctx->completion_lock); | |
1496 | } else { | |
1497 | io_req_cqe_overflow(req); | |
1498 | } | |
1499 | } | |
905c172f | 1500 | } |
3181e22f | 1501 | __io_cq_unlock_post_flush(ctx); |
d9dee430 | 1502 | |
931147dd DY |
1503 | if (!wq_list_empty(&ctx->submit_state.compl_reqs)) { |
1504 | io_free_batch_list(ctx, state->compl_reqs.first); | |
1505 | INIT_WQ_LIST(&state->compl_reqs); | |
1506 | } | |
7a743e22 PB |
1507 | } |
1508 | ||
ba816ad6 JA |
1509 | /* |
1510 | * Drop reference to request, return next in chain (if there is one) if this | |
1511 | * was the last reference to this request. | |
1512 | */ | |
0d85035a | 1513 | static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 1514 | { |
9b5f7bd9 PB |
1515 | struct io_kiocb *nxt = NULL; |
1516 | ||
de9b4cca | 1517 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 1518 | if (unlikely(req->flags & IO_REQ_LINK_FLAGS)) |
7819a1f6 | 1519 | nxt = io_req_find_next(req); |
f5c6cf2a | 1520 | io_free_req(req); |
2a44f467 | 1521 | } |
9b5f7bd9 | 1522 | return nxt; |
2b188cc1 JA |
1523 | } |
1524 | ||
6c503150 | 1525 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1526 | { |
1527 | /* See comment at the top of this file */ | |
1528 | smp_rmb(); | |
e23de15f | 1529 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1530 | } |
1531 | ||
def596e9 JA |
1532 | /* |
1533 | * We can't just wait for polled events to come to us, we have to actively | |
1534 | * find and complete them. | |
1535 | */ | |
c072481d | 1536 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1537 | { |
1538 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1539 | return; | |
1540 | ||
1541 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1542 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1543 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1544 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1545 | break; |
08f5439f JA |
1546 | /* |
1547 | * Ensure we allow local-to-the-cpu processing to take place, | |
1548 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1549 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1550 | */ |
3fcee5a6 PB |
1551 | if (need_resched()) { |
1552 | mutex_unlock(&ctx->uring_lock); | |
1553 | cond_resched(); | |
1554 | mutex_lock(&ctx->uring_lock); | |
1555 | } | |
def596e9 JA |
1556 | } |
1557 | mutex_unlock(&ctx->uring_lock); | |
1558 | } | |
1559 | ||
7668b92a | 1560 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1561 | { |
7668b92a | 1562 | unsigned int nr_events = 0; |
e9979b36 | 1563 | int ret = 0; |
155bc950 | 1564 | unsigned long check_cq; |
500f9fba | 1565 | |
76de6749 PB |
1566 | if (!io_allowed_run_tw(ctx)) |
1567 | return -EEXIST; | |
1568 | ||
3a08576b PB |
1569 | check_cq = READ_ONCE(ctx->check_cq); |
1570 | if (unlikely(check_cq)) { | |
1571 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
a85381d8 | 1572 | __io_cqring_overflow_flush(ctx); |
3a08576b PB |
1573 | /* |
1574 | * Similarly do not spin if we have not informed the user of any | |
1575 | * dropped CQE. | |
1576 | */ | |
1577 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1578 | return -EBADR; | |
1579 | } | |
f39c8a5b PB |
1580 | /* |
1581 | * Don't enter poll loop if we already have events pending. | |
1582 | * If we do, we can potentially be spinning for commands that | |
1583 | * already triggered a CQE (eg in error). | |
1584 | */ | |
f39c8a5b | 1585 | if (io_cqring_events(ctx)) |
d487b43c | 1586 | return 0; |
155bc950 | 1587 | |
def596e9 | 1588 | do { |
500f9fba JA |
1589 | /* |
1590 | * If a submit got punted to a workqueue, we can have the | |
1591 | * application entering polling for a command before it gets | |
1592 | * issued. That app will hold the uring_lock for the duration | |
1593 | * of the poll right here, so we need to take a breather every | |
1594 | * now and then to ensure that the issue has a chance to add | |
1595 | * the poll to the issued list. Otherwise we can spin here | |
1596 | * forever, while the workqueue is stuck trying to acquire the | |
1597 | * very same mutex. | |
1598 | */ | |
dac6a0ea JA |
1599 | if (wq_list_empty(&ctx->iopoll_list) || |
1600 | io_task_work_pending(ctx)) { | |
8f487ef2 PB |
1601 | u32 tail = ctx->cached_cq_tail; |
1602 | ||
8de11cdc | 1603 | (void) io_run_local_work_locked(ctx); |
def596e9 | 1604 | |
dac6a0ea JA |
1605 | if (task_work_pending(current) || |
1606 | wq_list_empty(&ctx->iopoll_list)) { | |
dac6a0ea | 1607 | mutex_unlock(&ctx->uring_lock); |
9d54bd6a | 1608 | io_run_task_work(); |
dac6a0ea | 1609 | mutex_lock(&ctx->uring_lock); |
dac6a0ea | 1610 | } |
8f487ef2 PB |
1611 | /* some requests don't go through iopoll_list */ |
1612 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1613 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1614 | break; |
500f9fba | 1615 | } |
5ba3c874 PB |
1616 | ret = io_do_iopoll(ctx, !min); |
1617 | if (ret < 0) | |
1618 | break; | |
1619 | nr_events += ret; | |
1620 | ret = 0; | |
1621 | } while (nr_events < min && !need_resched()); | |
d487b43c | 1622 | |
def596e9 JA |
1623 | return ret; |
1624 | } | |
7012c815 PB |
1625 | |
1626 | void io_req_task_complete(struct io_kiocb *req, bool *locked) | |
8ef12efe | 1627 | { |
7012c815 | 1628 | if (*locked) |
9da070b1 | 1629 | io_req_complete_defer(req); |
7012c815 | 1630 | else |
27f35fe9 | 1631 | io_req_complete_post(req, IO_URING_F_UNLOCKED); |
8ef12efe JA |
1632 | } |
1633 | ||
def596e9 JA |
1634 | /* |
1635 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1636 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1637 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1638 | * accessing the kiocb cookie. |
1639 | */ | |
9882131c | 1640 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1641 | { |
1642 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1643 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1644 | |
1645 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1646 | if (unlikely(needs_lock)) |
cb3d8972 | 1647 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1648 | |
1649 | /* | |
1650 | * Track whether we have multiple files in our lists. This will impact | |
1651 | * how we do polling eventually, not spinning if we're on potentially | |
1652 | * different devices. | |
1653 | */ | |
5eef4e87 | 1654 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1655 | ctx->poll_multi_queue = false; |
1656 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1657 | struct io_kiocb *list_req; |
1658 | ||
5eef4e87 PB |
1659 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1660 | comp_list); | |
30da1b45 | 1661 | if (list_req->file != req->file) |
915b3dde | 1662 | ctx->poll_multi_queue = true; |
def596e9 JA |
1663 | } |
1664 | ||
1665 | /* | |
1666 | * For fast devices, IO may have already completed. If it has, add | |
1667 | * it to the front so we find it first. | |
1668 | */ | |
65a6543d | 1669 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1670 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1671 | else |
5eef4e87 | 1672 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1673 | |
3b44b371 | 1674 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1675 | /* |
1676 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1677 | * in sq thread task context or in io worker task context. If | |
1678 | * current task context is sq thread, we don't need to check | |
1679 | * whether should wake up sq thread. | |
1680 | */ | |
1681 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1682 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1683 | wake_up(&ctx->sq_data->wait); | |
1684 | ||
1685 | mutex_unlock(&ctx->uring_lock); | |
1686 | } | |
def596e9 JA |
1687 | } |
1688 | ||
4503b767 JA |
1689 | static bool io_bdev_nowait(struct block_device *bdev) |
1690 | { | |
568ec936 | 1691 | return !bdev || bdev_nowait(bdev); |
4503b767 JA |
1692 | } |
1693 | ||
2b188cc1 JA |
1694 | /* |
1695 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1696 | * any file. For now, just ensure that anything potentially problematic is done | |
1697 | * inline. | |
1698 | */ | |
88459b50 | 1699 | static bool __io_file_supports_nowait(struct file *file, umode_t mode) |
2b188cc1 | 1700 | { |
4503b767 | 1701 | if (S_ISBLK(mode)) { |
4e7b5671 CH |
1702 | if (IS_ENABLED(CONFIG_BLOCK) && |
1703 | io_bdev_nowait(I_BDEV(file->f_mapping->host))) | |
4503b767 JA |
1704 | return true; |
1705 | return false; | |
1706 | } | |
976517f1 | 1707 | if (S_ISSOCK(mode)) |
2b188cc1 | 1708 | return true; |
4503b767 | 1709 | if (S_ISREG(mode)) { |
4e7b5671 CH |
1710 | if (IS_ENABLED(CONFIG_BLOCK) && |
1711 | io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
e5550a14 | 1712 | !io_is_uring_fops(file)) |
4503b767 JA |
1713 | return true; |
1714 | return false; | |
1715 | } | |
2b188cc1 | 1716 | |
c5b85625 JA |
1717 | /* any ->read/write should understand O_NONBLOCK */ |
1718 | if (file->f_flags & O_NONBLOCK) | |
1719 | return true; | |
35645ac3 | 1720 | return file->f_mode & FMODE_NOWAIT; |
2b188cc1 | 1721 | } |
c5b85625 | 1722 | |
88459b50 PB |
1723 | /* |
1724 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1725 | * any file. For now, just ensure that anything potentially problematic is done | |
1726 | * inline. | |
1727 | */ | |
a4ad4f74 | 1728 | unsigned int io_file_get_flags(struct file *file) |
88459b50 PB |
1729 | { |
1730 | umode_t mode = file_inode(file)->i_mode; | |
1731 | unsigned int res = 0; | |
af197f50 | 1732 | |
88459b50 PB |
1733 | if (S_ISREG(mode)) |
1734 | res |= FFS_ISREG; | |
1735 | if (__io_file_supports_nowait(file, mode)) | |
1736 | res |= FFS_NOWAIT; | |
1737 | return res; | |
2b188cc1 JA |
1738 | } |
1739 | ||
99f15d8d | 1740 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1741 | { |
e8c2bc1f JA |
1742 | WARN_ON_ONCE(!io_op_defs[req->opcode].async_size); |
1743 | req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL); | |
d886e185 PB |
1744 | if (req->async_data) { |
1745 | req->flags |= REQ_F_ASYNC_DATA; | |
1746 | return false; | |
1747 | } | |
1748 | return true; | |
3d9932a8 XW |
1749 | } |
1750 | ||
f3b44f92 | 1751 | int io_req_prep_async(struct io_kiocb *req) |
f67676d1 | 1752 | { |
0702e536 JA |
1753 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
1754 | ||
1755 | /* assign early for deferred execution for non-fixed file */ | |
1756 | if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE)) | |
1757 | req->file = io_file_get_normal(req, req->cqe.fd); | |
dc919caf | 1758 | if (!def->prep_async) |
0702e536 JA |
1759 | return 0; |
1760 | if (WARN_ON_ONCE(req_has_async_data(req))) | |
1761 | return -EFAULT; | |
59169439 PB |
1762 | if (!io_op_defs[req->opcode].manual_alloc) { |
1763 | if (io_alloc_async_data(req)) | |
1764 | return -EAGAIN; | |
1765 | } | |
dc919caf | 1766 | return def->prep_async(req); |
bfe76559 PB |
1767 | } |
1768 | ||
9cf7c104 PB |
1769 | static u32 io_get_sequence(struct io_kiocb *req) |
1770 | { | |
a3dbdf54 | 1771 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1772 | struct io_kiocb *cur; |
9cf7c104 | 1773 | |
a3dbdf54 | 1774 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1775 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1776 | seq--; |
1777 | return seq; | |
9cf7c104 PB |
1778 | } |
1779 | ||
c072481d | 1780 | static __cold void io_drain_req(struct io_kiocb *req) |
e276ae34 | 1781 | __must_hold(&ctx->uring_lock) |
de0617e4 | 1782 | { |
a197f664 | 1783 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1784 | struct io_defer_entry *de; |
f67676d1 | 1785 | int ret; |
e0eb71dc | 1786 | u32 seq = io_get_sequence(req); |
3c19966d | 1787 | |
9d858b21 | 1788 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1789 | spin_lock(&ctx->completion_lock); |
5e371265 | 1790 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1791 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1792 | queue: |
10c66904 | 1793 | ctx->drain_active = false; |
e0eb71dc PB |
1794 | io_req_task_queue(req); |
1795 | return; | |
10c66904 | 1796 | } |
e302f104 | 1797 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1798 | |
cbdcb435 | 1799 | io_prep_async_link(req); |
27dc8338 | 1800 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1801 | if (!de) { |
1b48773f | 1802 | ret = -ENOMEM; |
ef5c600a DY |
1803 | io_req_defer_failed(req, ret); |
1804 | return; | |
76cc33d7 | 1805 | } |
2d28390a | 1806 | |
79ebeaee | 1807 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1808 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1809 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1810 | kfree(de); |
e0eb71dc | 1811 | goto queue; |
de0617e4 JA |
1812 | } |
1813 | ||
48863ffd | 1814 | trace_io_uring_defer(req); |
27dc8338 | 1815 | de->req = req; |
9cf7c104 | 1816 | de->seq = seq; |
27dc8338 | 1817 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1818 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1819 | } |
1820 | ||
68fb8979 | 1821 | static void io_clean_op(struct io_kiocb *req) |
99bc4c38 | 1822 | { |
8197b053 PB |
1823 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
1824 | spin_lock(&req->ctx->completion_lock); | |
cc3cec83 | 1825 | io_put_kbuf_comp(req); |
8197b053 PB |
1826 | spin_unlock(&req->ctx->completion_lock); |
1827 | } | |
99bc4c38 | 1828 | |
0e1b6fe3 | 1829 | if (req->flags & REQ_F_NEED_CLEANUP) { |
4d4c9cff | 1830 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
bb040a21 | 1831 | |
4d4c9cff JA |
1832 | if (def->cleanup) |
1833 | def->cleanup(req); | |
99bc4c38 | 1834 | } |
75652a30 JA |
1835 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1836 | kfree(req->apoll->double_poll); | |
1837 | kfree(req->apoll); | |
1838 | req->apoll = NULL; | |
1839 | } | |
9cae36a0 JA |
1840 | if (req->flags & REQ_F_INFLIGHT) { |
1841 | struct io_uring_task *tctx = req->task->io_uring; | |
1842 | ||
1843 | atomic_dec(&tctx->inflight_tracked); | |
1844 | } | |
c854357b | 1845 | if (req->flags & REQ_F_CREDS) |
b8e64b53 | 1846 | put_cred(req->creds); |
d886e185 PB |
1847 | if (req->flags & REQ_F_ASYNC_DATA) { |
1848 | kfree(req->async_data); | |
1849 | req->async_data = NULL; | |
1850 | } | |
c854357b | 1851 | req->flags &= ~IO_REQ_CLEAN_FLAGS; |
99bc4c38 PB |
1852 | } |
1853 | ||
6bf9c47a JA |
1854 | static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags) |
1855 | { | |
1856 | if (req->file || !io_op_defs[req->opcode].needs_file) | |
1857 | return true; | |
1858 | ||
1859 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1860 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1861 | else |
cef216fc | 1862 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1863 | |
772f5e00 | 1864 | return !!req->file; |
6bf9c47a JA |
1865 | } |
1866 | ||
889fca73 | 1867 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1868 | { |
fcde59fe | 1869 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
5730b27e | 1870 | const struct cred *creds = NULL; |
d625c6ee | 1871 | int ret; |
2b188cc1 | 1872 | |
70152140 JA |
1873 | if (unlikely(!io_assign_file(req, issue_flags))) |
1874 | return -EBADF; | |
1875 | ||
6878b40e | 1876 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1877 | creds = override_creds(req->creds); |
5730b27e | 1878 | |
fcde59fe | 1879 | if (!def->audit_skip) |
5bd2182d PM |
1880 | audit_uring_entry(req->opcode); |
1881 | ||
0702e536 | 1882 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1883 | |
fcde59fe | 1884 | if (!def->audit_skip) |
5bd2182d PM |
1885 | audit_uring_exit(!ret, ret); |
1886 | ||
5730b27e JA |
1887 | if (creds) |
1888 | revert_creds(creds); | |
97b388d7 | 1889 | |
75d7b3ae PB |
1890 | if (ret == IOU_OK) { |
1891 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
9da070b1 | 1892 | io_req_complete_defer(req); |
75d7b3ae | 1893 | else |
1bec951c | 1894 | io_req_complete_post(req, issue_flags); |
75d7b3ae | 1895 | } else if (ret != IOU_ISSUE_SKIP_COMPLETE) |
def596e9 | 1896 | return ret; |
97b388d7 | 1897 | |
b532576e | 1898 | /* If the op doesn't have a file, we're not polling for it */ |
ef0ec1ad | 1899 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue) |
9882131c | 1900 | io_iopoll_req_issued(req, issue_flags); |
def596e9 JA |
1901 | |
1902 | return 0; | |
2b188cc1 JA |
1903 | } |
1904 | ||
329061d3 JA |
1905 | int io_poll_issue(struct io_kiocb *req, bool *locked) |
1906 | { | |
1907 | io_tw_lock(req->ctx, locked); | |
9a692451 DY |
1908 | return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT| |
1909 | IO_URING_F_COMPLETE_DEFER); | |
329061d3 JA |
1910 | } |
1911 | ||
c9f06aa7 | 1912 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1913 | { |
1914 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1915 | ||
1916 | req = io_put_req_find_next(req); | |
1917 | return req ? &req->work : NULL; | |
1918 | } | |
1919 | ||
c9f06aa7 | 1920 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1921 | { |
1922 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6bf9c47a | 1923 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
e6aeb272 | 1924 | unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ; |
d01905db | 1925 | bool needs_poll = false; |
6bf9c47a | 1926 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1927 | |
23a6c9ac | 1928 | /* one will be dropped by ->io_wq_free_work() after returning to io-wq */ |
48dcd38d PB |
1929 | if (!(req->flags & REQ_F_REFCOUNT)) |
1930 | __io_req_set_refcount(req, 2); | |
1931 | else | |
1932 | req_ref_get(req); | |
5d5901a3 | 1933 | |
cb2d344c | 1934 | io_arm_ltimeout(req); |
6bf9c47a | 1935 | |
dadebc35 | 1936 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1937 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1938 | fail: |
6bf9c47a | 1939 | io_req_task_queue_fail(req, err); |
d01905db PB |
1940 | return; |
1941 | } | |
0f8da75b PB |
1942 | if (!io_assign_file(req, issue_flags)) { |
1943 | err = -EBADF; | |
1944 | work->flags |= IO_WQ_WORK_CANCEL; | |
1945 | goto fail; | |
1946 | } | |
31b51510 | 1947 | |
d01905db | 1948 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
1949 | bool opcode_poll = def->pollin || def->pollout; |
1950 | ||
1951 | if (opcode_poll && file_can_poll(req->file)) { | |
1952 | needs_poll = true; | |
d01905db | 1953 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 1954 | } |
561fb04a | 1955 | } |
31b51510 | 1956 | |
d01905db PB |
1957 | do { |
1958 | ret = io_issue_sqe(req, issue_flags); | |
1959 | if (ret != -EAGAIN) | |
1960 | break; | |
1961 | /* | |
1962 | * We can get EAGAIN for iopolled IO even though we're | |
1963 | * forcing a sync submission from here, since we can't | |
1964 | * wait for request slots on the block side. | |
1965 | */ | |
1966 | if (!needs_poll) { | |
e0deb6a0 PB |
1967 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
1968 | break; | |
d01905db PB |
1969 | cond_resched(); |
1970 | continue; | |
90fa0288 HX |
1971 | } |
1972 | ||
4d9237e3 | 1973 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
1974 | return; |
1975 | /* aborted or ready, in either case retry blocking */ | |
1976 | needs_poll = false; | |
1977 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
1978 | } while (1); | |
31b51510 | 1979 | |
a3df7698 | 1980 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 1981 | if (ret < 0) |
a3df7698 | 1982 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
1983 | } |
1984 | ||
531113bb JA |
1985 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
1986 | unsigned int issue_flags) | |
09bb8394 | 1987 | { |
5106dd6e JA |
1988 | struct io_ring_ctx *ctx = req->ctx; |
1989 | struct file *file = NULL; | |
ac177053 | 1990 | unsigned long file_ptr; |
09bb8394 | 1991 | |
93f052cb | 1992 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 1993 | |
ac177053 | 1994 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 1995 | goto out; |
ac177053 PB |
1996 | fd = array_index_nospec(fd, ctx->nr_user_files); |
1997 | file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; | |
1998 | file = (struct file *) (file_ptr & FFS_MASK); | |
1999 | file_ptr &= ~FFS_MASK; | |
2000 | /* mask in overlapping REQ_F and FFS bits */ | |
35645ac3 | 2001 | req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT); |
5106dd6e JA |
2002 | io_req_set_rsrc_node(req, ctx, 0); |
2003 | out: | |
93f052cb | 2004 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
2005 | return file; |
2006 | } | |
d44f554e | 2007 | |
531113bb | 2008 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 2009 | { |
62906e89 | 2010 | struct file *file = fget(fd); |
ac177053 | 2011 | |
48863ffd | 2012 | trace_io_uring_file_get(req, fd); |
09bb8394 | 2013 | |
ac177053 | 2014 | /* we don't allow fixed io_uring files */ |
e5550a14 | 2015 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 2016 | io_req_track_inflight(req); |
8371adf5 | 2017 | return file; |
09bb8394 JA |
2018 | } |
2019 | ||
7bfa9bad | 2020 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
2021 | __must_hold(&req->ctx->uring_lock) |
2022 | { | |
7bfa9bad PB |
2023 | struct io_kiocb *linked_timeout; |
2024 | ||
2025 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
973fc83f | 2026 | io_req_defer_failed(req, ret); |
7bfa9bad PB |
2027 | return; |
2028 | } | |
2029 | ||
2030 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 2031 | |
4d9237e3 | 2032 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 2033 | case IO_APOLL_READY: |
336d28a8 | 2034 | io_kbuf_recycle(req, 0); |
d475a9a6 PB |
2035 | io_req_task_queue(req); |
2036 | break; | |
2037 | case IO_APOLL_ABORTED: | |
6436c770 | 2038 | io_kbuf_recycle(req, 0); |
77955efb | 2039 | io_queue_iowq(req, NULL); |
d475a9a6 | 2040 | break; |
b1c62645 | 2041 | case IO_APOLL_OK: |
b1c62645 | 2042 | break; |
d475a9a6 PB |
2043 | } |
2044 | ||
2045 | if (linked_timeout) | |
2046 | io_queue_linked_timeout(linked_timeout); | |
2047 | } | |
2048 | ||
cbc2e203 | 2049 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 2050 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 2051 | { |
e0c5c576 | 2052 | int ret; |
2b188cc1 | 2053 | |
c5eef2b9 | 2054 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 2055 | |
491381ce JA |
2056 | /* |
2057 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2058 | * doesn't support non-blocking read/write attempts | |
2059 | */ | |
7bfa9bad | 2060 | if (likely(!ret)) |
cb2d344c | 2061 | io_arm_ltimeout(req); |
7bfa9bad PB |
2062 | else |
2063 | io_queue_async(req, ret); | |
2b188cc1 JA |
2064 | } |
2065 | ||
4652fe3f | 2066 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 2067 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 2068 | { |
17b147f6 PB |
2069 | if (unlikely(req->flags & REQ_F_FAIL)) { |
2070 | /* | |
2071 | * We don't submit, fail them all, for that replace hardlinks | |
2072 | * with normal links. Extra REQ_F_LINK is tolerated. | |
2073 | */ | |
2074 | req->flags &= ~REQ_F_HARDLINK; | |
2075 | req->flags |= REQ_F_LINK; | |
973fc83f | 2076 | io_req_defer_failed(req, req->cqe.res); |
76cc33d7 PB |
2077 | } else { |
2078 | int ret = io_req_prep_async(req); | |
2079 | ||
ef5c600a | 2080 | if (unlikely(ret)) { |
973fc83f | 2081 | io_req_defer_failed(req, ret); |
ef5c600a DY |
2082 | return; |
2083 | } | |
2084 | ||
2085 | if (unlikely(req->ctx->drain_active)) | |
2086 | io_drain_req(req); | |
76cc33d7 | 2087 | else |
77955efb | 2088 | io_queue_iowq(req, NULL); |
ce35a47a | 2089 | } |
4fe2c963 JL |
2090 | } |
2091 | ||
b16fed66 PB |
2092 | /* |
2093 | * Check SQE restrictions (opcode and flags). | |
2094 | * | |
2095 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
2096 | */ | |
2097 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
2098 | struct io_kiocb *req, | |
2099 | unsigned int sqe_flags) | |
4fe2c963 | 2100 | { |
b16fed66 PB |
2101 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
2102 | return false; | |
2103 | ||
2104 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
2105 | ctx->restrictions.sqe_flags_required) | |
2106 | return false; | |
2107 | ||
2108 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
2109 | ctx->restrictions.sqe_flags_required)) | |
2110 | return false; | |
2111 | ||
2112 | return true; | |
4fe2c963 JL |
2113 | } |
2114 | ||
22b2ca31 PB |
2115 | static void io_init_req_drain(struct io_kiocb *req) |
2116 | { | |
2117 | struct io_ring_ctx *ctx = req->ctx; | |
2118 | struct io_kiocb *head = ctx->submit_state.link.head; | |
2119 | ||
2120 | ctx->drain_active = true; | |
2121 | if (head) { | |
2122 | /* | |
2123 | * If we need to drain a request in the middle of a link, drain | |
2124 | * the head request and the next request/link after the current | |
2125 | * link. Considering sequential execution of links, | |
b6c7db32 | 2126 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
2127 | * link. |
2128 | */ | |
b6c7db32 | 2129 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
2130 | ctx->drain_next = true; |
2131 | } | |
2132 | } | |
2133 | ||
b16fed66 PB |
2134 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2135 | const struct io_uring_sqe *sqe) | |
282cdc86 | 2136 | __must_hold(&ctx->uring_lock) |
b16fed66 | 2137 | { |
fcde59fe | 2138 | const struct io_op_def *def; |
b16fed66 | 2139 | unsigned int sqe_flags; |
fc0ae024 | 2140 | int personality; |
4a04d1d1 | 2141 | u8 opcode; |
b16fed66 | 2142 | |
864ea921 | 2143 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 2144 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 PB |
2145 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
2146 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
cef216fc | 2147 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 2148 | req->file = NULL; |
c1bdf8ed | 2149 | req->rsrc_node = NULL; |
b16fed66 | 2150 | req->task = current; |
b16fed66 | 2151 | |
4a04d1d1 PB |
2152 | if (unlikely(opcode >= IORING_OP_LAST)) { |
2153 | req->opcode = 0; | |
b16fed66 | 2154 | return -EINVAL; |
4a04d1d1 | 2155 | } |
fcde59fe | 2156 | def = &io_op_defs[opcode]; |
68fe256a PB |
2157 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
2158 | /* enforce forwards compatibility on users */ | |
2159 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
2160 | return -EINVAL; | |
4e906702 | 2161 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 2162 | if (!def->buffer_select) |
4e906702 JA |
2163 | return -EOPNOTSUPP; |
2164 | req->buf_index = READ_ONCE(sqe->buf_group); | |
2165 | } | |
5562a8d7 PB |
2166 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
2167 | ctx->drain_disabled = true; | |
2168 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
2169 | if (ctx->drain_disabled) | |
2170 | return -EOPNOTSUPP; | |
22b2ca31 | 2171 | io_init_req_drain(req); |
5562a8d7 | 2172 | } |
2a56a9bd PB |
2173 | } |
2174 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
2175 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
2176 | return -EACCES; | |
2177 | /* knock it to the slow queue path, will be drained there */ | |
2178 | if (ctx->drain_active) | |
2179 | req->flags |= REQ_F_FORCE_ASYNC; | |
2180 | /* if there is no link, we're at "next" request and need to drain */ | |
2181 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
2182 | ctx->drain_next = false; | |
2183 | ctx->drain_active = true; | |
b6c7db32 | 2184 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 2185 | } |
68fe256a | 2186 | } |
b16fed66 | 2187 | |
fcde59fe | 2188 | if (!def->ioprio && sqe->ioprio) |
73911426 | 2189 | return -EINVAL; |
fcde59fe | 2190 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
73911426 JA |
2191 | return -EINVAL; |
2192 | ||
fcde59fe | 2193 | if (def->needs_file) { |
6d63416d PB |
2194 | struct io_submit_state *state = &ctx->submit_state; |
2195 | ||
cef216fc | 2196 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 2197 | |
6d63416d PB |
2198 | /* |
2199 | * Plug now if we have more than 2 IO left after this, and the | |
2200 | * target is potentially a read/write to block based storage. | |
2201 | */ | |
fcde59fe | 2202 | if (state->need_plug && def->plug) { |
6d63416d PB |
2203 | state->plug_started = true; |
2204 | state->need_plug = false; | |
5ca7a8b3 | 2205 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 2206 | } |
b16fed66 | 2207 | } |
863e0560 | 2208 | |
003e8dcc JA |
2209 | personality = READ_ONCE(sqe->personality); |
2210 | if (personality) { | |
cdab10bf LT |
2211 | int ret; |
2212 | ||
c10d1f98 PB |
2213 | req->creds = xa_load(&ctx->personalities, personality); |
2214 | if (!req->creds) | |
003e8dcc | 2215 | return -EINVAL; |
c10d1f98 | 2216 | get_cred(req->creds); |
cdc1404a PM |
2217 | ret = security_uring_override_creds(req->creds); |
2218 | if (ret) { | |
2219 | put_cred(req->creds); | |
2220 | return ret; | |
2221 | } | |
b8e64b53 | 2222 | req->flags |= REQ_F_CREDS; |
003e8dcc | 2223 | } |
b16fed66 | 2224 | |
0702e536 | 2225 | return def->prep(req, sqe); |
b16fed66 PB |
2226 | } |
2227 | ||
df3becde PB |
2228 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
2229 | struct io_kiocb *req, int ret) | |
2230 | { | |
2231 | struct io_ring_ctx *ctx = req->ctx; | |
2232 | struct io_submit_link *link = &ctx->submit_state.link; | |
2233 | struct io_kiocb *head = link->head; | |
2234 | ||
48863ffd | 2235 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
2236 | |
2237 | /* | |
2238 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
2239 | * unusable. Instead of failing eagerly, continue assembling the link if | |
2240 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
2241 | * should find the flag and handle the rest. | |
2242 | */ | |
2243 | req_fail_link_node(req, ret); | |
2244 | if (head && !(head->flags & REQ_F_FAIL)) | |
2245 | req_fail_link_node(head, -ECANCELED); | |
2246 | ||
2247 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
2248 | if (head) { | |
2249 | link->last->link = req; | |
2250 | link->head = NULL; | |
2251 | req = head; | |
2252 | } | |
2253 | io_queue_sqe_fallback(req); | |
2254 | return ret; | |
2255 | } | |
2256 | ||
2257 | if (head) | |
2258 | link->last->link = req; | |
2259 | else | |
2260 | link->head = req; | |
2261 | link->last = req; | |
2262 | return 0; | |
2263 | } | |
2264 | ||
2265 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 2266 | const struct io_uring_sqe *sqe) |
282cdc86 | 2267 | __must_hold(&ctx->uring_lock) |
9e645e11 | 2268 | { |
a1ab7b35 | 2269 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 2270 | int ret; |
9e645e11 | 2271 | |
a6b8cadc | 2272 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
2273 | if (unlikely(ret)) |
2274 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 2275 | |
be7053b7 | 2276 | /* don't need @sqe from now on */ |
48863ffd | 2277 | trace_io_uring_submit_sqe(req, true); |
a6b8cadc | 2278 | |
9e645e11 JA |
2279 | /* |
2280 | * If we already have a head request, queue this one for async | |
2281 | * submittal once the head completes. If we don't have a head but | |
2282 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2283 | * submitted sync once the chain is complete. If none of those | |
2284 | * conditions are true (normal request), then just queue it. | |
2285 | */ | |
924a07e4 | 2286 | if (unlikely(link->head)) { |
df3becde PB |
2287 | ret = io_req_prep_async(req); |
2288 | if (unlikely(ret)) | |
2289 | return io_submit_fail_init(sqe, req, ret); | |
2290 | ||
48863ffd | 2291 | trace_io_uring_link(req, link->head); |
f2f87370 | 2292 | link->last->link = req; |
863e0560 | 2293 | link->last = req; |
32fe525b | 2294 | |
da1a08c5 | 2295 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 2296 | return 0; |
df3becde PB |
2297 | /* last request of the link, flush it */ |
2298 | req = link->head; | |
f15a3431 | 2299 | link->head = NULL; |
924a07e4 PB |
2300 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
2301 | goto fallback; | |
2302 | ||
2303 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2304 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2305 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2306 | link->head = req; | |
2307 | link->last = req; | |
2308 | } else { | |
2309 | fallback: | |
2310 | io_queue_sqe_fallback(req); | |
2311 | } | |
f15a3431 | 2312 | return 0; |
9e645e11 | 2313 | } |
2e6e1fde | 2314 | |
f15a3431 | 2315 | io_queue_sqe(req); |
1d4240cc | 2316 | return 0; |
9e645e11 JA |
2317 | } |
2318 | ||
9a56a232 JA |
2319 | /* |
2320 | * Batched submission is done, ensure local IO is flushed out. | |
2321 | */ | |
553deffd | 2322 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2323 | { |
553deffd PB |
2324 | struct io_submit_state *state = &ctx->submit_state; |
2325 | ||
e126391c PB |
2326 | if (unlikely(state->link.head)) |
2327 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2328 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2329 | io_submit_flush_completions(ctx); |
27926b68 JA |
2330 | if (state->plug_started) |
2331 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2332 | } |
2333 | ||
2334 | /* | |
2335 | * Start submission side cache. | |
2336 | */ | |
2337 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2338 | unsigned int max_ios) |
9a56a232 | 2339 | { |
27926b68 | 2340 | state->plug_started = false; |
4b628aeb | 2341 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2342 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2343 | /* set only head, no need to init link_last in advance */ |
2344 | state->link.head = NULL; | |
9a56a232 JA |
2345 | } |
2346 | ||
2b188cc1 JA |
2347 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2348 | { | |
75b28aff | 2349 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2350 | |
caf582c6 PB |
2351 | /* |
2352 | * Ensure any loads from the SQEs are done at this point, | |
2353 | * since once we write the new head, the application could | |
2354 | * write new data to them. | |
2355 | */ | |
2356 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2357 | } |
2358 | ||
2b188cc1 | 2359 | /* |
dd9ae8a0 | 2360 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2361 | * that is mapped by userspace. This means that care needs to be taken to |
2362 | * ensure that reads are stable, as we cannot rely on userspace always | |
2363 | * being a good citizen. If members of the sqe are validated and then later | |
2364 | * used, it's important that those reads are done through READ_ONCE() to | |
2365 | * prevent a re-load down the line. | |
2366 | */ | |
709b302f | 2367 | static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) |
2b188cc1 | 2368 | { |
ea5ab3b5 | 2369 | unsigned head, mask = ctx->sq_entries - 1; |
17d3aeb3 | 2370 | unsigned sq_idx = ctx->cached_sq_head++ & mask; |
2b188cc1 JA |
2371 | |
2372 | /* | |
2373 | * The cached sq head (or cq tail) serves two purposes: | |
2374 | * | |
2375 | * 1) allows us to batch the cost of updating the user visible | |
2376 | * head updates. | |
2377 | * 2) allows the kernel side to track the head on its own, even | |
2378 | * though the application is the one updating it. | |
2379 | */ | |
17d3aeb3 | 2380 | head = READ_ONCE(ctx->sq_array[sq_idx]); |
ebdeb7c0 JA |
2381 | if (likely(head < ctx->sq_entries)) { |
2382 | /* double index for 128-byte SQEs, twice as long */ | |
2383 | if (ctx->flags & IORING_SETUP_SQE128) | |
2384 | head <<= 1; | |
709b302f | 2385 | return &ctx->sq_sqes[head]; |
ebdeb7c0 | 2386 | } |
2b188cc1 JA |
2387 | |
2388 | /* drop invalid entries */ | |
15641e42 PB |
2389 | ctx->cq_extra--; |
2390 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2391 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
709b302f PB |
2392 | return NULL; |
2393 | } | |
2394 | ||
17437f31 | 2395 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2396 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2397 | { |
69629809 | 2398 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2399 | unsigned int left; |
2400 | int ret; | |
6c271ce2 | 2401 | |
51d48dab | 2402 | if (unlikely(!entries)) |
69629809 | 2403 | return 0; |
ee7d46d9 | 2404 | /* make sure SQ entry isn't read before tail */ |
8e6971a8 PB |
2405 | ret = left = min3(nr, ctx->sq_entries, entries); |
2406 | io_get_task_refs(left); | |
2407 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2408 | |
69629809 | 2409 | do { |
3529d8c2 | 2410 | const struct io_uring_sqe *sqe; |
196be95c | 2411 | struct io_kiocb *req; |
fb5ccc98 | 2412 | |
8e6971a8 | 2413 | if (unlikely(!io_alloc_req_refill(ctx))) |
fb5ccc98 | 2414 | break; |
a33ae9ce | 2415 | req = io_alloc_req(ctx); |
4fccfcbb PB |
2416 | sqe = io_get_sqe(ctx); |
2417 | if (unlikely(!sqe)) { | |
fa05457a | 2418 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2419 | break; |
2420 | } | |
6c271ce2 | 2421 | |
1cd15904 PB |
2422 | /* |
2423 | * Continue submitting even for sqe failure if the | |
2424 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2425 | */ | |
2426 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2427 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2428 | left--; | |
2429 | break; | |
bcbb7bf6 | 2430 | } |
1cd15904 | 2431 | } while (--left); |
9466f437 | 2432 | |
8e6971a8 PB |
2433 | if (unlikely(left)) { |
2434 | ret -= left; | |
2435 | /* try again if it submitted nothing and can't allocate a req */ | |
2436 | if (!ret && io_req_cache_empty(ctx)) | |
2437 | ret = -EAGAIN; | |
2438 | current->io_uring->cached_refs += left; | |
9466f437 | 2439 | } |
6c271ce2 | 2440 | |
553deffd | 2441 | io_submit_state_end(ctx); |
ae9428ca PB |
2442 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2443 | io_commit_sqring(ctx); | |
8e6971a8 | 2444 | return ret; |
6c271ce2 JA |
2445 | } |
2446 | ||
bda52162 JA |
2447 | struct io_wait_queue { |
2448 | struct wait_queue_entry wq; | |
2449 | struct io_ring_ctx *ctx; | |
5fd46178 | 2450 | unsigned cq_tail; |
bda52162 | 2451 | unsigned nr_timeouts; |
d33a39e5 | 2452 | ktime_t timeout; |
bda52162 JA |
2453 | }; |
2454 | ||
b4c98d59 DY |
2455 | static inline bool io_has_work(struct io_ring_ctx *ctx) |
2456 | { | |
c0e0d6ba | 2457 | return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) || |
490c00eb | 2458 | !llist_empty(&ctx->work_llist); |
b4c98d59 DY |
2459 | } |
2460 | ||
6c503150 | 2461 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
2462 | { |
2463 | struct io_ring_ctx *ctx = iowq->ctx; | |
0fc8c2ac | 2464 | int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail; |
bda52162 JA |
2465 | |
2466 | /* | |
d195a66e | 2467 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
2468 | * started waiting. For timeouts, we always want to return to userspace, |
2469 | * regardless of event count. | |
2470 | */ | |
5fd46178 | 2471 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
2472 | } |
2473 | ||
2474 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2475 | int wake_flags, void *key) | |
2476 | { | |
bd550173 | 2477 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq); |
bda52162 | 2478 | |
6c503150 PB |
2479 | /* |
2480 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2481 | * the task, and the next invocation will do it. | |
2482 | */ | |
bd550173 | 2483 | if (io_should_wake(iowq) || io_has_work(iowq->ctx)) |
6c503150 PB |
2484 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2485 | return -1; | |
bda52162 JA |
2486 | } |
2487 | ||
c0e0d6ba | 2488 | int io_run_task_work_sig(struct io_ring_ctx *ctx) |
af9c1a44 | 2489 | { |
1414d629 | 2490 | if (!llist_empty(&ctx->work_llist)) { |
2f413956 | 2491 | __set_current_state(TASK_RUNNING); |
1414d629 PB |
2492 | if (io_run_local_work(ctx) > 0) |
2493 | return 1; | |
2494 | } | |
2495 | if (io_run_task_work() > 0) | |
af9c1a44 | 2496 | return 1; |
c5020bc8 OL |
2497 | if (task_sigpending(current)) |
2498 | return -EINTR; | |
2499 | return 0; | |
af9c1a44 JA |
2500 | } |
2501 | ||
eeb60b9a PB |
2502 | /* when returns >0, the caller should retry */ |
2503 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
d33a39e5 | 2504 | struct io_wait_queue *iowq) |
eeb60b9a | 2505 | { |
3fcf19d5 PB |
2506 | if (unlikely(READ_ONCE(ctx->check_cq))) |
2507 | return 1; | |
846072f1 PB |
2508 | if (unlikely(!llist_empty(&ctx->work_llist))) |
2509 | return 1; | |
2510 | if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) | |
2511 | return 1; | |
2512 | if (unlikely(task_sigpending(current))) | |
2513 | return -EINTR; | |
2514 | if (unlikely(io_should_wake(iowq))) | |
2515 | return 0; | |
d33a39e5 | 2516 | if (iowq->timeout == KTIME_MAX) |
46ae7eef | 2517 | schedule(); |
d33a39e5 | 2518 | else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS)) |
22833966 | 2519 | return -ETIME; |
846072f1 | 2520 | return 0; |
eeb60b9a PB |
2521 | } |
2522 | ||
2b188cc1 JA |
2523 | /* |
2524 | * Wait until events become available, if we don't already have some. The | |
2525 | * application must reap them itself, as they reside on the shared cq ring. | |
2526 | */ | |
2527 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2528 | const sigset_t __user *sig, size_t sigsz, |
2529 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2530 | { |
90291099 | 2531 | struct io_wait_queue iowq; |
75b28aff | 2532 | struct io_rings *rings = ctx->rings; |
c1d5a224 | 2533 | int ret; |
2b188cc1 | 2534 | |
76de6749 PB |
2535 | if (!io_allowed_run_tw(ctx)) |
2536 | return -EEXIST; | |
140102ae PB |
2537 | if (!llist_empty(&ctx->work_llist)) |
2538 | io_run_local_work(ctx); | |
f36ba6cf PB |
2539 | io_run_task_work(); |
2540 | io_cqring_overflow_flush(ctx); | |
2541 | /* if user messes with these they will just get an early return */ | |
2542 | if (__io_cqring_events_user(ctx) >= min_events) | |
2543 | return 0; | |
2b188cc1 JA |
2544 | |
2545 | if (sig) { | |
9e75ad5d AB |
2546 | #ifdef CONFIG_COMPAT |
2547 | if (in_compat_syscall()) | |
2548 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2549 | sigsz); |
9e75ad5d AB |
2550 | else |
2551 | #endif | |
b772434b | 2552 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2553 | |
2b188cc1 JA |
2554 | if (ret) |
2555 | return ret; | |
2556 | } | |
2557 | ||
90291099 PB |
2558 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
2559 | iowq.wq.private = current; | |
2560 | INIT_LIST_HEAD(&iowq.wq.entry); | |
2561 | iowq.ctx = ctx; | |
bda52162 | 2562 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 2563 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
d33a39e5 PB |
2564 | iowq.timeout = KTIME_MAX; |
2565 | ||
2566 | if (uts) { | |
2567 | struct timespec64 ts; | |
2568 | ||
2569 | if (get_timespec64(&ts, uts)) | |
2570 | return -EFAULT; | |
2571 | iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); | |
2572 | } | |
90291099 | 2573 | |
c826bd7a | 2574 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2575 | do { |
3fcf19d5 PB |
2576 | unsigned long check_cq; |
2577 | ||
130bd686 PB |
2578 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
2579 | set_current_state(TASK_INTERRUPTIBLE); | |
2580 | } else { | |
2581 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, | |
2582 | TASK_INTERRUPTIBLE); | |
2583 | } | |
2584 | ||
d33a39e5 | 2585 | ret = io_cqring_wait_schedule(ctx, &iowq); |
130bd686 | 2586 | __set_current_state(TASK_RUNNING); |
846072f1 PB |
2587 | if (ret < 0) |
2588 | break; | |
2589 | /* | |
2590 | * Run task_work after scheduling and before io_should_wake(). | |
2591 | * If we got woken because of task_work being processed, run it | |
2592 | * now rather than let the caller do another wait loop. | |
2593 | */ | |
2594 | io_run_task_work(); | |
2595 | if (!llist_empty(&ctx->work_llist)) | |
2596 | io_run_local_work(ctx); | |
3fcf19d5 PB |
2597 | |
2598 | check_cq = READ_ONCE(ctx->check_cq); | |
2599 | if (unlikely(check_cq)) { | |
2600 | /* let the caller flush overflows, retry */ | |
326a9e48 | 2601 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) |
3fcf19d5 | 2602 | io_cqring_do_overflow_flush(ctx); |
3fcf19d5 PB |
2603 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) { |
2604 | ret = -EBADR; | |
2605 | break; | |
2606 | } | |
2607 | } | |
2608 | ||
846072f1 PB |
2609 | if (io_should_wake(&iowq)) { |
2610 | ret = 0; | |
35d90f95 | 2611 | break; |
846072f1 | 2612 | } |
ca0a2651 | 2613 | cond_resched(); |
846072f1 | 2614 | } while (1); |
bda52162 | 2615 | |
130bd686 PB |
2616 | if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) |
2617 | finish_wait(&ctx->cq_wait, &iowq.wq); | |
b7db41c9 | 2618 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2619 | |
75b28aff | 2620 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2621 | } |
2622 | ||
73572984 | 2623 | static void io_mem_free(void *ptr) |
b9bd2bea | 2624 | { |
73572984 | 2625 | struct page *page; |
b36a2050 | 2626 | |
73572984 JA |
2627 | if (!ptr) |
2628 | return; | |
b9bd2bea | 2629 | |
73572984 JA |
2630 | page = virt_to_head_page(ptr); |
2631 | if (put_page_testzero(page)) | |
2632 | free_compound_page(page); | |
b9bd2bea PB |
2633 | } |
2634 | ||
73572984 | 2635 | static void *io_mem_alloc(size_t size) |
b9bd2bea | 2636 | { |
73572984 | 2637 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; |
b9bd2bea | 2638 | |
73572984 | 2639 | return (void *) __get_free_pages(gfp, get_order(size)); |
b9bd2bea PB |
2640 | } |
2641 | ||
73572984 JA |
2642 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2643 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2644 | { |
73572984 JA |
2645 | struct io_rings *rings; |
2646 | size_t off, sq_array_size; | |
6b06314c | 2647 | |
73572984 JA |
2648 | off = struct_size(rings, cqes, cq_entries); |
2649 | if (off == SIZE_MAX) | |
2650 | return SIZE_MAX; | |
2651 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2652 | if (check_shl_overflow(off, 1, &off)) | |
2653 | return SIZE_MAX; | |
2654 | } | |
ab409402 | 2655 | |
73572984 JA |
2656 | #ifdef CONFIG_SMP |
2657 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2658 | if (off == 0) | |
2659 | return SIZE_MAX; | |
2660 | #endif | |
82fbcfa9 | 2661 | |
73572984 JA |
2662 | if (sq_offset) |
2663 | *sq_offset = off; | |
82fbcfa9 | 2664 | |
73572984 JA |
2665 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2666 | if (sq_array_size == SIZE_MAX) | |
2667 | return SIZE_MAX; | |
6b06314c | 2668 | |
73572984 JA |
2669 | if (check_add_overflow(off, sq_array_size, &off)) |
2670 | return SIZE_MAX; | |
8bad28d8 | 2671 | |
73572984 | 2672 | return off; |
8bad28d8 HX |
2673 | } |
2674 | ||
73572984 JA |
2675 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, |
2676 | unsigned int eventfd_async) | |
8bad28d8 | 2677 | { |
73572984 JA |
2678 | struct io_ev_fd *ev_fd; |
2679 | __s32 __user *fds = arg; | |
2680 | int fd; | |
f2303b1f | 2681 | |
73572984 JA |
2682 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, |
2683 | lockdep_is_held(&ctx->uring_lock)); | |
2684 | if (ev_fd) | |
2685 | return -EBUSY; | |
8bad28d8 | 2686 | |
73572984 JA |
2687 | if (copy_from_user(&fd, fds, sizeof(*fds))) |
2688 | return -EFAULT; | |
8dd03afe | 2689 | |
73572984 JA |
2690 | ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); |
2691 | if (!ev_fd) | |
2692 | return -ENOMEM; | |
05f3fb3c | 2693 | |
73572984 JA |
2694 | ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); |
2695 | if (IS_ERR(ev_fd->cq_ev_fd)) { | |
2696 | int ret = PTR_ERR(ev_fd->cq_ev_fd); | |
2697 | kfree(ev_fd); | |
2698 | return ret; | |
2699 | } | |
305bef98 PB |
2700 | |
2701 | spin_lock(&ctx->completion_lock); | |
2702 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
2703 | spin_unlock(&ctx->completion_lock); | |
2704 | ||
73572984 JA |
2705 | ev_fd->eventfd_async = eventfd_async; |
2706 | ctx->has_evfd = true; | |
2707 | rcu_assign_pointer(ctx->io_ev_fd, ev_fd); | |
21a091b9 DY |
2708 | atomic_set(&ev_fd->refs, 1); |
2709 | atomic_set(&ev_fd->ops, 0); | |
73572984 | 2710 | return 0; |
d7954b2b BM |
2711 | } |
2712 | ||
73572984 | 2713 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 2714 | { |
73572984 JA |
2715 | struct io_ev_fd *ev_fd; |
2716 | ||
2717 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, | |
2718 | lockdep_is_held(&ctx->uring_lock)); | |
2719 | if (ev_fd) { | |
2720 | ctx->has_evfd = false; | |
2721 | rcu_assign_pointer(ctx->io_ev_fd, NULL); | |
21a091b9 DY |
2722 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops)) |
2723 | call_rcu(&ev_fd->rcu, io_eventfd_ops); | |
73572984 JA |
2724 | return 0; |
2725 | } | |
2d091d62 | 2726 | |
73572984 | 2727 | return -ENXIO; |
44b31f2f PB |
2728 | } |
2729 | ||
73572984 | 2730 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2731 | { |
37f0e767 | 2732 | int nr = 0; |
bf019da7 | 2733 | |
9a4fdbd8 | 2734 | mutex_lock(&ctx->uring_lock); |
34f0bc42 | 2735 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
9a4fdbd8 | 2736 | |
88ab95be | 2737 | while (!io_req_cache_empty(ctx)) { |
34f0bc42 | 2738 | struct io_kiocb *req = io_alloc_req(ctx); |
9a4fdbd8 | 2739 | |
c2b6c6bc | 2740 | kmem_cache_free(req_cachep, req); |
37f0e767 | 2741 | nr++; |
c2b6c6bc | 2742 | } |
37f0e767 PB |
2743 | if (nr) |
2744 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2745 | mutex_unlock(&ctx->uring_lock); |
2746 | } | |
2747 | ||
c072481d | 2748 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2749 | { |
37d1e2e3 | 2750 | io_sq_thread_finish(ctx); |
ab409402 | 2751 | io_rsrc_refs_drop(ctx); |
43597aac PB |
2752 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
2753 | io_wait_rsrc_data(ctx->buf_data); | |
2754 | io_wait_rsrc_data(ctx->file_data); | |
2755 | ||
8bad28d8 | 2756 | mutex_lock(&ctx->uring_lock); |
43597aac | 2757 | if (ctx->buf_data) |
bd54b6fe | 2758 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2759 | if (ctx->file_data) |
08480400 | 2760 | __io_sqe_files_unregister(ctx); |
a85381d8 | 2761 | io_cqring_overflow_kill(ctx); |
9b402849 | 2762 | io_eventfd_unregister(ctx); |
9b797a37 | 2763 | io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); |
43e0bbbd | 2764 | io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); |
77bc59b4 | 2765 | mutex_unlock(&ctx->uring_lock); |
5a2e745d | 2766 | io_destroy_buffers(ctx); |
07db298a PB |
2767 | if (ctx->sq_creds) |
2768 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2769 | if (ctx->submitter_task) |
2770 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2771 | |
a7f0ed5a PB |
2772 | /* there are no registered resources left, nobody uses it */ |
2773 | if (ctx->rsrc_node) | |
2774 | io_rsrc_node_destroy(ctx->rsrc_node); | |
8dd03afe | 2775 | if (ctx->rsrc_backup_node) |
b895c9a6 | 2776 | io_rsrc_node_destroy(ctx->rsrc_backup_node); |
a7f0ed5a | 2777 | flush_delayed_work(&ctx->rsrc_put_work); |
756ab7c0 | 2778 | flush_delayed_work(&ctx->fallback_work); |
a7f0ed5a PB |
2779 | |
2780 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
2781 | WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist)); | |
def596e9 | 2782 | |
2b188cc1 | 2783 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
2784 | if (ctx->ring_sock) { |
2785 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 2786 | sock_release(ctx->ring_sock); |
355e8d26 | 2787 | } |
2b188cc1 | 2788 | #endif |
ef9dd637 | 2789 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
2b188cc1 | 2790 | |
42b6419d PB |
2791 | if (ctx->mm_account) { |
2792 | mmdrop(ctx->mm_account); | |
2793 | ctx->mm_account = NULL; | |
2794 | } | |
75b28aff | 2795 | io_mem_free(ctx->rings); |
2b188cc1 | 2796 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
2797 | |
2798 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2799 | free_uid(ctx->user); |
4010fec4 | 2800 | io_req_caches_free(ctx); |
e941894e JA |
2801 | if (ctx->hash_map) |
2802 | io_wq_put_hash(ctx->hash_map); | |
e6f89be6 | 2803 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2804 | kfree(ctx->cancel_table_locked.hbs); |
6224843d | 2805 | kfree(ctx->dummy_ubuf); |
9cfc7e94 JA |
2806 | kfree(ctx->io_bl); |
2807 | xa_destroy(&ctx->io_bl_xa); | |
2b188cc1 JA |
2808 | kfree(ctx); |
2809 | } | |
2810 | ||
bca39f39 PB |
2811 | static __cold void io_activate_pollwq_cb(struct callback_head *cb) |
2812 | { | |
2813 | struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx, | |
2814 | poll_wq_task_work); | |
2815 | ||
2816 | mutex_lock(&ctx->uring_lock); | |
2817 | ctx->poll_activated = true; | |
2818 | mutex_unlock(&ctx->uring_lock); | |
2819 | ||
2820 | /* | |
2821 | * Wake ups for some events between start of polling and activation | |
2822 | * might've been lost due to loose synchronisation. | |
2823 | */ | |
2824 | wake_up_all(&ctx->poll_wq); | |
2825 | percpu_ref_put(&ctx->refs); | |
2826 | } | |
2827 | ||
2828 | static __cold void io_activate_pollwq(struct io_ring_ctx *ctx) | |
2829 | { | |
2830 | spin_lock(&ctx->completion_lock); | |
2831 | /* already activated or in progress */ | |
2832 | if (ctx->poll_activated || ctx->poll_wq_task_work.func) | |
2833 | goto out; | |
2834 | if (WARN_ON_ONCE(!ctx->task_complete)) | |
2835 | goto out; | |
2836 | if (!ctx->submitter_task) | |
2837 | goto out; | |
2838 | /* | |
2839 | * with ->submitter_task only the submitter task completes requests, we | |
2840 | * only need to sync with it, which is done by injecting a tw | |
2841 | */ | |
2842 | init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb); | |
2843 | percpu_ref_get(&ctx->refs); | |
2844 | if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL)) | |
2845 | percpu_ref_put(&ctx->refs); | |
2846 | out: | |
2847 | spin_unlock(&ctx->completion_lock); | |
2848 | } | |
2849 | ||
2b188cc1 JA |
2850 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) |
2851 | { | |
2852 | struct io_ring_ctx *ctx = file->private_data; | |
2853 | __poll_t mask = 0; | |
2854 | ||
bca39f39 PB |
2855 | if (unlikely(!ctx->poll_activated)) |
2856 | io_activate_pollwq(ctx); | |
2857 | ||
7b235dd8 | 2858 | poll_wait(file, &ctx->poll_wq, wait); |
4f7067c3 SB |
2859 | /* |
2860 | * synchronizes with barrier from wq_has_sleeper call in | |
2861 | * io_commit_cqring | |
2862 | */ | |
2b188cc1 | 2863 | smp_rmb(); |
90554200 | 2864 | if (!io_sqring_full(ctx)) |
2b188cc1 | 2865 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
2866 | |
2867 | /* | |
2868 | * Don't flush cqring overflow list here, just do a simple check. | |
2869 | * Otherwise there could possible be ABBA deadlock: | |
2870 | * CPU0 CPU1 | |
2871 | * ---- ---- | |
2872 | * lock(&ctx->uring_lock); | |
2873 | * lock(&ep->mtx); | |
2874 | * lock(&ctx->uring_lock); | |
2875 | * lock(&ep->mtx); | |
2876 | * | |
2877 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
10d8bc35 | 2878 | * pushes them to do the flush. |
ed670c3f | 2879 | */ |
b4c98d59 DY |
2880 | |
2881 | if (io_cqring_events(ctx) || io_has_work(ctx)) | |
2b188cc1 JA |
2882 | mask |= EPOLLIN | EPOLLRDNORM; |
2883 | ||
2884 | return mask; | |
2885 | } | |
2886 | ||
0bead8cd | 2887 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 2888 | { |
4379bf8b | 2889 | const struct cred *creds; |
071698e1 | 2890 | |
61cf9370 | 2891 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
2892 | if (creds) { |
2893 | put_cred(creds); | |
0bead8cd | 2894 | return 0; |
1e6fa521 | 2895 | } |
0bead8cd YD |
2896 | |
2897 | return -EINVAL; | |
2898 | } | |
2899 | ||
d56d938b PB |
2900 | struct io_tctx_exit { |
2901 | struct callback_head task_work; | |
2902 | struct completion completion; | |
baf186c4 | 2903 | struct io_ring_ctx *ctx; |
d56d938b PB |
2904 | }; |
2905 | ||
c072481d | 2906 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
2907 | { |
2908 | struct io_uring_task *tctx = current->io_uring; | |
2909 | struct io_tctx_exit *work; | |
2910 | ||
2911 | work = container_of(cb, struct io_tctx_exit, task_work); | |
2912 | /* | |
2913 | * When @in_idle, we're in cancellation and it's racy to remove the | |
2914 | * node. It'll be removed by the end of cancellation, just ignore it. | |
998b30c3 HM |
2915 | * tctx can be NULL if the queueing of this task_work raced with |
2916 | * work cancelation off the exec path. | |
d56d938b | 2917 | */ |
998b30c3 | 2918 | if (tctx && !atomic_read(&tctx->in_idle)) |
eef51daa | 2919 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
2920 | complete(&work->completion); |
2921 | } | |
2922 | ||
c072481d | 2923 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
2924 | { |
2925 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2926 | ||
2927 | return req->ctx == data; | |
2928 | } | |
2929 | ||
c072481d | 2930 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 2931 | { |
d56d938b | 2932 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 2933 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 2934 | unsigned long interval = HZ / 20; |
d56d938b PB |
2935 | struct io_tctx_exit exit; |
2936 | struct io_tctx_node *node; | |
2937 | int ret; | |
85faa7b8 | 2938 | |
56952e91 JA |
2939 | /* |
2940 | * If we're doing polled IO and end up having requests being | |
2941 | * submitted async (out-of-line), then completions can come in while | |
2942 | * we're waiting for refs to drop. We need to reap these manually, | |
2943 | * as nobody else will be looking for them. | |
2944 | */ | |
b2edc0a7 | 2945 | do { |
a85381d8 PB |
2946 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
2947 | mutex_lock(&ctx->uring_lock); | |
2948 | io_cqring_overflow_kill(ctx); | |
2949 | mutex_unlock(&ctx->uring_lock); | |
2950 | } | |
2951 | ||
c0e0d6ba DY |
2952 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) |
2953 | io_move_task_work_from_local(ctx); | |
2954 | ||
affa87db PB |
2955 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
2956 | cond_resched(); | |
2957 | ||
28090c13 PB |
2958 | if (ctx->sq_data) { |
2959 | struct io_sq_data *sqd = ctx->sq_data; | |
2960 | struct task_struct *tsk; | |
2961 | ||
2962 | io_sq_thread_park(sqd); | |
2963 | tsk = sqd->thread; | |
2964 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
2965 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
2966 | io_cancel_ctx_cb, ctx, true); | |
2967 | io_sq_thread_unpark(sqd); | |
2968 | } | |
b5bb3a24 | 2969 | |
37f0e767 PB |
2970 | io_req_caches_free(ctx); |
2971 | ||
58d3be2c PB |
2972 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
2973 | /* there is little hope left, don't run it too often */ | |
2974 | interval = HZ * 60; | |
2975 | } | |
2976 | } while (!wait_for_completion_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 2977 | |
7f00651a PB |
2978 | init_completion(&exit.completion); |
2979 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
2980 | exit.ctx = ctx; | |
89b5066e PB |
2981 | /* |
2982 | * Some may use context even when all refs and requests have been put, | |
2983 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 2984 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
2985 | * this lock/unlock section also waits them to finish. |
2986 | */ | |
d56d938b PB |
2987 | mutex_lock(&ctx->uring_lock); |
2988 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
2989 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
2990 | ||
d56d938b PB |
2991 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
2992 | ctx_node); | |
7f00651a PB |
2993 | /* don't spin on a single task if cancellation failed */ |
2994 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
2995 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
2996 | if (WARN_ON_ONCE(ret)) | |
2997 | continue; | |
d56d938b PB |
2998 | |
2999 | mutex_unlock(&ctx->uring_lock); | |
3000 | wait_for_completion(&exit.completion); | |
d56d938b PB |
3001 | mutex_lock(&ctx->uring_lock); |
3002 | } | |
3003 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
3004 | spin_lock(&ctx->completion_lock); |
3005 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 3006 | |
85faa7b8 JA |
3007 | io_ring_ctx_free(ctx); |
3008 | } | |
3009 | ||
c072481d | 3010 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 3011 | { |
61cf9370 MWO |
3012 | unsigned long index; |
3013 | struct creds *creds; | |
3014 | ||
2b188cc1 JA |
3015 | mutex_lock(&ctx->uring_lock); |
3016 | percpu_ref_kill(&ctx->refs); | |
61cf9370 MWO |
3017 | xa_for_each(&ctx->personalities, index, creds) |
3018 | io_unregister_personality(ctx, index); | |
9ca9fb24 PB |
3019 | if (ctx->rings) |
3020 | io_poll_remove_all(ctx, NULL, true); | |
2b188cc1 JA |
3021 | mutex_unlock(&ctx->uring_lock); |
3022 | ||
02bac94b PB |
3023 | /* |
3024 | * If we failed setting up the ctx, we might not have any rings | |
3025 | * and therefore did not submit any requests | |
3026 | */ | |
3027 | if (ctx->rings) | |
60053be8 | 3028 | io_kill_timeouts(ctx, NULL, true); |
309fc03a | 3029 | |
85faa7b8 | 3030 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
3031 | /* |
3032 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
3033 | * if we're exiting a ton of rings at the same time. It just adds | |
3034 | * noise and overhead, there's no discernable change in runtime | |
3035 | * over using system_wq. | |
3036 | */ | |
3037 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
3038 | } |
3039 | ||
3040 | static int io_uring_release(struct inode *inode, struct file *file) | |
3041 | { | |
3042 | struct io_ring_ctx *ctx = file->private_data; | |
3043 | ||
3044 | file->private_data = NULL; | |
3045 | io_ring_ctx_wait_and_kill(ctx); | |
3046 | return 0; | |
3047 | } | |
3048 | ||
f6edbabb PB |
3049 | struct io_task_cancel { |
3050 | struct task_struct *task; | |
3dd0c97a | 3051 | bool all; |
f6edbabb | 3052 | }; |
f254ac04 | 3053 | |
f6edbabb | 3054 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 3055 | { |
9a472ef7 | 3056 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 3057 | struct io_task_cancel *cancel = data; |
9a472ef7 | 3058 | |
6af3f48b | 3059 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
3060 | } |
3061 | ||
c072481d PB |
3062 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
3063 | struct task_struct *task, | |
3064 | bool cancel_all) | |
b7ddce3c | 3065 | { |
e1915f76 | 3066 | struct io_defer_entry *de; |
b7ddce3c PB |
3067 | LIST_HEAD(list); |
3068 | ||
79ebeaee | 3069 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 3070 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 3071 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
3072 | list_cut_position(&list, &ctx->defer_list, &de->list); |
3073 | break; | |
3074 | } | |
3075 | } | |
79ebeaee | 3076 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
3077 | if (list_empty(&list)) |
3078 | return false; | |
b7ddce3c PB |
3079 | |
3080 | while (!list_empty(&list)) { | |
3081 | de = list_first_entry(&list, struct io_defer_entry, list); | |
3082 | list_del_init(&de->list); | |
e276ae34 | 3083 | io_req_task_queue_fail(de->req, -ECANCELED); |
b7ddce3c PB |
3084 | kfree(de); |
3085 | } | |
e1915f76 | 3086 | return true; |
b7ddce3c PB |
3087 | } |
3088 | ||
c072481d | 3089 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
3090 | { |
3091 | struct io_tctx_node *node; | |
3092 | enum io_wq_cancel cret; | |
3093 | bool ret = false; | |
3094 | ||
3095 | mutex_lock(&ctx->uring_lock); | |
3096 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3097 | struct io_uring_task *tctx = node->task->io_uring; | |
3098 | ||
3099 | /* | |
3100 | * io_wq will stay alive while we hold uring_lock, because it's | |
3101 | * killed after ctx nodes, which requires to take the lock. | |
3102 | */ | |
3103 | if (!tctx || !tctx->io_wq) | |
3104 | continue; | |
3105 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
3106 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3107 | } | |
3108 | mutex_unlock(&ctx->uring_lock); | |
3109 | ||
3110 | return ret; | |
3111 | } | |
3112 | ||
affa87db | 3113 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
3114 | struct task_struct *task, |
3115 | bool cancel_all) | |
9936c7c2 | 3116 | { |
3dd0c97a | 3117 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 3118 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
3119 | enum io_wq_cancel cret; |
3120 | bool ret = false; | |
9936c7c2 | 3121 | |
60053be8 PB |
3122 | /* failed during ring init, it couldn't have issued any requests */ |
3123 | if (!ctx->rings) | |
affa87db | 3124 | return false; |
60053be8 | 3125 | |
affa87db PB |
3126 | if (!task) { |
3127 | ret |= io_uring_try_cancel_iowq(ctx); | |
3128 | } else if (tctx && tctx->io_wq) { | |
3129 | /* | |
3130 | * Cancels requests of all rings, not only @ctx, but | |
3131 | * it's fine as the task is in exit/exec. | |
3132 | */ | |
3133 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
3134 | &cancel, true); | |
3135 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3136 | } | |
9936c7c2 | 3137 | |
affa87db PB |
3138 | /* SQPOLL thread does its own polling */ |
3139 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
3140 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
3141 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
3142 | io_iopoll_try_reap_events(ctx); | |
3143 | ret = true; | |
9936c7c2 | 3144 | } |
9936c7c2 | 3145 | } |
affa87db | 3146 | |
140102ae PB |
3147 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3148 | io_allowed_defer_tw_run(ctx)) | |
c0e0d6ba | 3149 | ret |= io_run_local_work(ctx) > 0; |
affa87db PB |
3150 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
3151 | mutex_lock(&ctx->uring_lock); | |
3152 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
3153 | mutex_unlock(&ctx->uring_lock); | |
3154 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
3155 | if (task) | |
c0e0d6ba | 3156 | ret |= io_run_task_work() > 0; |
affa87db | 3157 | return ret; |
9936c7c2 PB |
3158 | } |
3159 | ||
3f48cf18 | 3160 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 3161 | { |
3f48cf18 | 3162 | if (tracked) |
9cae36a0 | 3163 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
3164 | return percpu_counter_sum(&tctx->inflight); |
3165 | } | |
3166 | ||
78cc687b PB |
3167 | /* |
3168 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 3169 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 3170 | */ |
17437f31 | 3171 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 3172 | { |
521d6a73 | 3173 | struct io_uring_task *tctx = current->io_uring; |
734551df | 3174 | struct io_ring_ctx *ctx; |
0e9ddb39 PB |
3175 | s64 inflight; |
3176 | DEFINE_WAIT(wait); | |
fdaf083c | 3177 | |
78cc687b PB |
3178 | WARN_ON_ONCE(sqd && sqd->thread != current); |
3179 | ||
6d042ffb PO |
3180 | if (!current->io_uring) |
3181 | return; | |
17a91051 PB |
3182 | if (tctx->io_wq) |
3183 | io_wq_exit_start(tctx->io_wq); | |
3184 | ||
0e9ddb39 PB |
3185 | atomic_inc(&tctx->in_idle); |
3186 | do { | |
affa87db PB |
3187 | bool loop = false; |
3188 | ||
e9dbe221 | 3189 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 3190 | /* read completions before cancelations */ |
78cc687b | 3191 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
3192 | if (!inflight) |
3193 | break; | |
fdaf083c | 3194 | |
78cc687b PB |
3195 | if (!sqd) { |
3196 | struct io_tctx_node *node; | |
3197 | unsigned long index; | |
0f212204 | 3198 | |
78cc687b PB |
3199 | xa_for_each(&tctx->xa, index, node) { |
3200 | /* sqpoll task will cancel all its requests */ | |
3201 | if (node->ctx->sq_data) | |
3202 | continue; | |
affa87db PB |
3203 | loop |= io_uring_try_cancel_requests(node->ctx, |
3204 | current, cancel_all); | |
78cc687b PB |
3205 | } |
3206 | } else { | |
3207 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
3208 | loop |= io_uring_try_cancel_requests(ctx, |
3209 | current, | |
3210 | cancel_all); | |
3211 | } | |
3212 | ||
3213 | if (loop) { | |
3214 | cond_resched(); | |
3215 | continue; | |
78cc687b | 3216 | } |
17a91051 | 3217 | |
78a78060 JA |
3218 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
3219 | io_run_task_work(); | |
e9dbe221 | 3220 | io_uring_drop_tctx_refs(current); |
78a78060 | 3221 | |
0f212204 | 3222 | /* |
a1bb3cd5 PB |
3223 | * If we've seen completions, retry without waiting. This |
3224 | * avoids a race where a completion comes in before we did | |
3225 | * prepare_to_wait(). | |
0f212204 | 3226 | */ |
3dd0c97a | 3227 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 3228 | schedule(); |
f57555ed | 3229 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 3230 | } while (1); |
de7f1d9e | 3231 | |
8452d4a6 | 3232 | io_uring_clean_tctx(tctx); |
3dd0c97a | 3233 | if (cancel_all) { |
3cc7fdb9 PB |
3234 | /* |
3235 | * We shouldn't run task_works after cancel, so just leave | |
3236 | * ->in_idle set for normal exit. | |
3237 | */ | |
3238 | atomic_dec(&tctx->in_idle); | |
3f48cf18 PB |
3239 | /* for exec all current's requests should be gone, kill tctx */ |
3240 | __io_uring_free(current); | |
3241 | } | |
44e728b8 PB |
3242 | } |
3243 | ||
f552a27a | 3244 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 3245 | { |
f552a27a | 3246 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
3247 | } |
3248 | ||
6c5c240e RP |
3249 | static void *io_uring_validate_mmap_request(struct file *file, |
3250 | loff_t pgoff, size_t sz) | |
2b188cc1 | 3251 | { |
2b188cc1 | 3252 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 3253 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
3254 | struct page *page; |
3255 | void *ptr; | |
3256 | ||
3257 | switch (offset) { | |
3258 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
3259 | case IORING_OFF_CQ_RING: |
3260 | ptr = ctx->rings; | |
2b188cc1 JA |
3261 | break; |
3262 | case IORING_OFF_SQES: | |
3263 | ptr = ctx->sq_sqes; | |
3264 | break; | |
2b188cc1 | 3265 | default: |
6c5c240e | 3266 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
3267 | } |
3268 | ||
3269 | page = virt_to_head_page(ptr); | |
a50b854e | 3270 | if (sz > page_size(page)) |
6c5c240e RP |
3271 | return ERR_PTR(-EINVAL); |
3272 | ||
3273 | return ptr; | |
3274 | } | |
3275 | ||
3276 | #ifdef CONFIG_MMU | |
3277 | ||
c072481d | 3278 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e RP |
3279 | { |
3280 | size_t sz = vma->vm_end - vma->vm_start; | |
3281 | unsigned long pfn; | |
3282 | void *ptr; | |
3283 | ||
3284 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
3285 | if (IS_ERR(ptr)) | |
3286 | return PTR_ERR(ptr); | |
2b188cc1 JA |
3287 | |
3288 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3289 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3290 | } | |
3291 | ||
6c5c240e RP |
3292 | #else /* !CONFIG_MMU */ |
3293 | ||
3294 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3295 | { | |
3296 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
3297 | } | |
3298 | ||
3299 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
3300 | { | |
3301 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
3302 | } | |
3303 | ||
3304 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
3305 | unsigned long addr, unsigned long len, | |
3306 | unsigned long pgoff, unsigned long flags) | |
3307 | { | |
3308 | void *ptr; | |
3309 | ||
3310 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
3311 | if (IS_ERR(ptr)) | |
3312 | return PTR_ERR(ptr); | |
3313 | ||
3314 | return (unsigned long) ptr; | |
3315 | } | |
3316 | ||
3317 | #endif /* !CONFIG_MMU */ | |
3318 | ||
f81440d3 PB |
3319 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
3320 | { | |
3321 | if (flags & IORING_ENTER_EXT_ARG) { | |
3322 | struct io_uring_getevents_arg arg; | |
3323 | ||
3324 | if (argsz != sizeof(arg)) | |
3325 | return -EINVAL; | |
3326 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3327 | return -EFAULT; | |
3328 | } | |
3329 | return 0; | |
3330 | } | |
3331 | ||
c73ebb68 HX |
3332 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
3333 | struct __kernel_timespec __user **ts, | |
3334 | const sigset_t __user **sig) | |
3335 | { | |
3336 | struct io_uring_getevents_arg arg; | |
3337 | ||
3338 | /* | |
3339 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
3340 | * is just a pointer to the sigset_t. | |
3341 | */ | |
3342 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
3343 | *sig = (const sigset_t __user *) argp; | |
3344 | *ts = NULL; | |
3345 | return 0; | |
3346 | } | |
3347 | ||
3348 | /* | |
3349 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
3350 | * timespec and sigset_t pointers if good. | |
3351 | */ | |
3352 | if (*argsz != sizeof(arg)) | |
3353 | return -EINVAL; | |
3354 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3355 | return -EFAULT; | |
d2347b96 DY |
3356 | if (arg.pad) |
3357 | return -EINVAL; | |
c73ebb68 HX |
3358 | *sig = u64_to_user_ptr(arg.sigmask); |
3359 | *argsz = arg.sigmask_sz; | |
3360 | *ts = u64_to_user_ptr(arg.ts); | |
3361 | return 0; | |
3362 | } | |
3363 | ||
2b188cc1 | 3364 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
3365 | u32, min_complete, u32, flags, const void __user *, argp, |
3366 | size_t, argsz) | |
2b188cc1 JA |
3367 | { |
3368 | struct io_ring_ctx *ctx; | |
2b188cc1 | 3369 | struct fd f; |
33f993da | 3370 | long ret; |
2b188cc1 | 3371 | |
33f993da | 3372 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
3373 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
3374 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
3375 | return -EINVAL; |
3376 | ||
e7a6c00d JA |
3377 | /* |
3378 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3379 | * need only dereference our task private array to find it. | |
3380 | */ | |
3381 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3382 | struct io_uring_task *tctx = current->io_uring; | |
3383 | ||
3273c440 | 3384 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
e7a6c00d JA |
3385 | return -EINVAL; |
3386 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
3387 | f.file = tctx->registered_rings[fd]; | |
4329490a | 3388 | f.flags = 0; |
3273c440 PB |
3389 | if (unlikely(!f.file)) |
3390 | return -EBADF; | |
e7a6c00d JA |
3391 | } else { |
3392 | f = fdget(fd); | |
3273c440 PB |
3393 | if (unlikely(!f.file)) |
3394 | return -EBADF; | |
3395 | ret = -EOPNOTSUPP; | |
3396 | if (unlikely(!io_is_uring_fops(f.file))) | |
fbb8bb02 | 3397 | goto out; |
e7a6c00d | 3398 | } |
2b188cc1 | 3399 | |
2b188cc1 | 3400 | ctx = f.file->private_data; |
7e84e1c7 | 3401 | ret = -EBADFD; |
33f993da | 3402 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3403 | goto out; |
3404 | ||
6c271ce2 JA |
3405 | /* |
3406 | * For SQ polling, the thread will do all submissions and completions. | |
3407 | * Just return the requested submit count, and wake the thread if | |
3408 | * we were asked to. | |
3409 | */ | |
b2a9eada | 3410 | ret = 0; |
6c271ce2 | 3411 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3412 | io_cqring_overflow_flush(ctx); |
89448c47 | 3413 | |
21f96522 JA |
3414 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3415 | ret = -EOWNERDEAD; | |
04147488 | 3416 | goto out; |
21f96522 | 3417 | } |
6c271ce2 | 3418 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3419 | wake_up(&ctx->sq_data->wait); |
d9d05217 PB |
3420 | if (flags & IORING_ENTER_SQ_WAIT) { |
3421 | ret = io_sqpoll_wait_sq(ctx); | |
3422 | if (ret) | |
3423 | goto out; | |
3424 | } | |
3e813c90 | 3425 | ret = to_submit; |
b2a9eada | 3426 | } else if (to_submit) { |
eef51daa | 3427 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3428 | if (unlikely(ret)) |
3429 | goto out; | |
7c504e65 | 3430 | |
2b188cc1 | 3431 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3432 | ret = io_submit_sqes(ctx, to_submit); |
3433 | if (ret != to_submit) { | |
d487b43c | 3434 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3435 | goto out; |
d487b43c | 3436 | } |
44f87745 PB |
3437 | if (flags & IORING_ENTER_GETEVENTS) { |
3438 | if (ctx->syscall_iopoll) | |
3439 | goto iopoll_locked; | |
3440 | /* | |
3441 | * Ignore errors, we'll soon call io_cqring_wait() and | |
3442 | * it should handle ownership problems if any. | |
3443 | */ | |
3444 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3445 | (void)io_run_local_work_locked(ctx); | |
3446 | } | |
d487b43c | 3447 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 | 3448 | } |
c0e0d6ba | 3449 | |
2b188cc1 | 3450 | if (flags & IORING_ENTER_GETEVENTS) { |
3e813c90 | 3451 | int ret2; |
c0e0d6ba | 3452 | |
773697b6 | 3453 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3454 | /* |
3455 | * We disallow the app entering submit/complete with | |
3456 | * polling, but we still need to lock the ring to | |
3457 | * prevent racing with polled issue that got punted to | |
3458 | * a workqueue. | |
3459 | */ | |
3460 | mutex_lock(&ctx->uring_lock); | |
3461 | iopoll_locked: | |
3e813c90 DY |
3462 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3463 | if (likely(!ret2)) { | |
3464 | min_complete = min(min_complete, | |
3465 | ctx->cq_entries); | |
3466 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3467 | } |
3468 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3469 | } else { |
f81440d3 PB |
3470 | const sigset_t __user *sig; |
3471 | struct __kernel_timespec __user *ts; | |
3472 | ||
3e813c90 DY |
3473 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3474 | if (likely(!ret2)) { | |
3475 | min_complete = min(min_complete, | |
3476 | ctx->cq_entries); | |
3477 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3478 | argsz, ts); | |
3479 | } | |
def596e9 | 3480 | } |
c73ebb68 | 3481 | |
155bc950 | 3482 | if (!ret) { |
3e813c90 | 3483 | ret = ret2; |
2b188cc1 | 3484 | |
155bc950 DY |
3485 | /* |
3486 | * EBADR indicates that one or more CQE were dropped. | |
3487 | * Once the user has been informed we can clear the bit | |
3488 | * as they are obviously ok with those drops. | |
3489 | */ | |
3490 | if (unlikely(ret2 == -EBADR)) | |
3491 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3492 | &ctx->check_cq); | |
def596e9 | 3493 | } |
2b188cc1 | 3494 | } |
7c504e65 | 3495 | out: |
4329490a | 3496 | fdput(f); |
3e813c90 | 3497 | return ret; |
2b188cc1 JA |
3498 | } |
3499 | ||
3500 | static const struct file_operations io_uring_fops = { | |
3501 | .release = io_uring_release, | |
3502 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3503 | #ifndef CONFIG_MMU |
3504 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3505 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
3506 | #endif | |
2b188cc1 | 3507 | .poll = io_uring_poll, |
bebdb65e | 3508 | #ifdef CONFIG_PROC_FS |
87ce955b | 3509 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3510 | #endif |
2b188cc1 JA |
3511 | }; |
3512 | ||
92ac8bea JA |
3513 | bool io_is_uring_fops(struct file *file) |
3514 | { | |
3515 | return file->f_op == &io_uring_fops; | |
3516 | } | |
3517 | ||
c072481d PB |
3518 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3519 | struct io_uring_params *p) | |
2b188cc1 | 3520 | { |
75b28aff HV |
3521 | struct io_rings *rings; |
3522 | size_t size, sq_array_offset; | |
2b188cc1 | 3523 | |
bd740481 JA |
3524 | /* make sure these are sane, as we already accounted them */ |
3525 | ctx->sq_entries = p->sq_entries; | |
3526 | ctx->cq_entries = p->cq_entries; | |
3527 | ||
baf9cb64 | 3528 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3529 | if (size == SIZE_MAX) |
3530 | return -EOVERFLOW; | |
3531 | ||
3532 | rings = io_mem_alloc(size); | |
3533 | if (!rings) | |
2b188cc1 JA |
3534 | return -ENOMEM; |
3535 | ||
75b28aff HV |
3536 | ctx->rings = rings; |
3537 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
3538 | rings->sq_ring_mask = p->sq_entries - 1; | |
3539 | rings->cq_ring_mask = p->cq_entries - 1; | |
3540 | rings->sq_ring_entries = p->sq_entries; | |
3541 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3542 | |
ebdeb7c0 JA |
3543 | if (p->flags & IORING_SETUP_SQE128) |
3544 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3545 | else | |
3546 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
3547 | if (size == SIZE_MAX) { |
3548 | io_mem_free(ctx->rings); | |
3549 | ctx->rings = NULL; | |
2b188cc1 | 3550 | return -EOVERFLOW; |
eb065d30 | 3551 | } |
2b188cc1 JA |
3552 | |
3553 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
3554 | if (!ctx->sq_sqes) { |
3555 | io_mem_free(ctx->rings); | |
3556 | ctx->rings = NULL; | |
2b188cc1 | 3557 | return -ENOMEM; |
eb065d30 | 3558 | } |
2b188cc1 | 3559 | |
2b188cc1 JA |
3560 | return 0; |
3561 | } | |
3562 | ||
9faadcc8 PB |
3563 | static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) |
3564 | { | |
3565 | int ret, fd; | |
3566 | ||
3567 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3568 | if (fd < 0) | |
3569 | return fd; | |
3570 | ||
97c96e9f | 3571 | ret = __io_uring_add_tctx_node(ctx); |
9faadcc8 PB |
3572 | if (ret) { |
3573 | put_unused_fd(fd); | |
3574 | return ret; | |
3575 | } | |
3576 | fd_install(fd, file); | |
3577 | return fd; | |
3578 | } | |
3579 | ||
2b188cc1 JA |
3580 | /* |
3581 | * Allocate an anonymous fd, this is what constitutes the application | |
3582 | * visible backing of an io_uring instance. The application mmaps this | |
3583 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3584 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3585 | */ | |
9faadcc8 | 3586 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
3587 | { |
3588 | struct file *file; | |
9faadcc8 | 3589 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
3590 | int ret; |
3591 | ||
2b188cc1 JA |
3592 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
3593 | &ctx->ring_sock); | |
3594 | if (ret) | |
9faadcc8 | 3595 | return ERR_PTR(ret); |
2b188cc1 JA |
3596 | #endif |
3597 | ||
91a9ab7c PM |
3598 | file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, |
3599 | O_RDWR | O_CLOEXEC, NULL); | |
2b188cc1 | 3600 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
3601 | if (IS_ERR(file)) { |
3602 | sock_release(ctx->ring_sock); | |
3603 | ctx->ring_sock = NULL; | |
3604 | } else { | |
3605 | ctx->ring_sock->file = file; | |
0f212204 | 3606 | } |
2b188cc1 | 3607 | #endif |
9faadcc8 | 3608 | return file; |
2b188cc1 JA |
3609 | } |
3610 | ||
c072481d PB |
3611 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3612 | struct io_uring_params __user *params) | |
2b188cc1 | 3613 | { |
2b188cc1 | 3614 | struct io_ring_ctx *ctx; |
9faadcc8 | 3615 | struct file *file; |
2b188cc1 JA |
3616 | int ret; |
3617 | ||
8110c1a6 | 3618 | if (!entries) |
2b188cc1 | 3619 | return -EINVAL; |
8110c1a6 JA |
3620 | if (entries > IORING_MAX_ENTRIES) { |
3621 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3622 | return -EINVAL; | |
3623 | entries = IORING_MAX_ENTRIES; | |
3624 | } | |
2b188cc1 JA |
3625 | |
3626 | /* | |
3627 | * Use twice as many entries for the CQ ring. It's possible for the | |
3628 | * application to drive a higher depth than the size of the SQ ring, | |
3629 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3630 | * some flexibility in overcommitting a bit. If the application has |
3631 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3632 | * of CQ ring entries manually. | |
2b188cc1 JA |
3633 | */ |
3634 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3635 | if (p->flags & IORING_SETUP_CQSIZE) { |
3636 | /* | |
3637 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3638 | * to a power-of-two, if it isn't already. We do NOT impose | |
3639 | * any cq vs sq ring sizing. | |
3640 | */ | |
eb2667b3 | 3641 | if (!p->cq_entries) |
33a107f0 | 3642 | return -EINVAL; |
8110c1a6 JA |
3643 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3644 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3645 | return -EINVAL; | |
3646 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3647 | } | |
eb2667b3 JQ |
3648 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3649 | if (p->cq_entries < p->sq_entries) | |
3650 | return -EINVAL; | |
33a107f0 JA |
3651 | } else { |
3652 | p->cq_entries = 2 * p->sq_entries; | |
3653 | } | |
2b188cc1 | 3654 | |
2b188cc1 | 3655 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3656 | if (!ctx) |
2b188cc1 | 3657 | return -ENOMEM; |
773697b6 | 3658 | |
e6aeb272 PB |
3659 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3660 | !(ctx->flags & IORING_SETUP_IOPOLL) && | |
3661 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3662 | ctx->task_complete = true; | |
3663 | ||
bca39f39 PB |
3664 | /* |
3665 | * lazy poll_wq activation relies on ->task_complete for synchronisation | |
3666 | * purposes, see io_activate_pollwq() | |
3667 | */ | |
3668 | if (!ctx->task_complete) | |
3669 | ctx->poll_activated = true; | |
3670 | ||
773697b6 PB |
3671 | /* |
3672 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3673 | * space applications don't need to do io completion events | |
3674 | * polling again, they can rely on io_sq_thread to do polling | |
3675 | * work, which can reduce cpu usage and uring_lock contention. | |
3676 | */ | |
3677 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3678 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3679 | ctx->syscall_iopoll = 1; | |
3680 | ||
2b188cc1 | 3681 | ctx->compat = in_compat_syscall(); |
62e398be JA |
3682 | if (!capable(CAP_IPC_LOCK)) |
3683 | ctx->user = get_uid(current_user()); | |
2aede0e4 | 3684 | |
9f010507 | 3685 | /* |
e1169f06 JA |
3686 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3687 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3688 | */ |
e1169f06 JA |
3689 | ret = -EINVAL; |
3690 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3691 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 | 3692 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
c0e0d6ba DY |
3693 | IORING_SETUP_TASKRUN_FLAG | |
3694 | IORING_SETUP_DEFER_TASKRUN)) | |
e1169f06 | 3695 | goto err; |
9f010507 | 3696 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3697 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3698 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3699 | } else { | |
c0e0d6ba DY |
3700 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && |
3701 | !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) | |
ef060ea9 | 3702 | goto err; |
9f010507 | 3703 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3704 | } |
9f010507 | 3705 | |
c0e0d6ba DY |
3706 | /* |
3707 | * For DEFER_TASKRUN we require the completion task to be the same as the | |
3708 | * submission task. This implies that there is only one submitter, so enforce | |
3709 | * that. | |
3710 | */ | |
3711 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && | |
3712 | !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { | |
3713 | goto err; | |
3714 | } | |
3715 | ||
2aede0e4 JA |
3716 | /* |
3717 | * This is just grabbed for accounting purposes. When a process exits, | |
3718 | * the mm is exited and dropped before the files, hence we need to hang | |
3719 | * on to this mm purely for the purposes of being able to unaccount | |
3720 | * memory (locked/pinned vm). It's not used for anything else. | |
3721 | */ | |
6b7898eb | 3722 | mmgrab(current->mm); |
2aede0e4 | 3723 | ctx->mm_account = current->mm; |
6b7898eb | 3724 | |
2b188cc1 JA |
3725 | ret = io_allocate_scq_urings(ctx, p); |
3726 | if (ret) | |
3727 | goto err; | |
3728 | ||
7e84e1c7 | 3729 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
3730 | if (ret) |
3731 | goto err; | |
eae071c9 | 3732 | /* always set a rsrc node */ |
47b228ce PB |
3733 | ret = io_rsrc_node_switch_start(ctx); |
3734 | if (ret) | |
3735 | goto err; | |
eae071c9 | 3736 | io_rsrc_node_switch(ctx, NULL); |
2b188cc1 | 3737 | |
2b188cc1 | 3738 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
3739 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3740 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3741 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3742 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3743 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3744 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
3745 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
3746 | |
3747 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
3748 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3749 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3750 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3751 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3752 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3753 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 3754 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 3755 | |
7f13657d XW |
3756 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
3757 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 3758 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 3759 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 3760 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e JA |
3761 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
3762 | IORING_FEAT_LINKED_FILE; | |
7f13657d XW |
3763 | |
3764 | if (copy_to_user(params, p, sizeof(*p))) { | |
3765 | ret = -EFAULT; | |
3766 | goto err; | |
3767 | } | |
d1719f70 | 3768 | |
7cae596b DY |
3769 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER |
3770 | && !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
8579538c | 3771 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
7cae596b | 3772 | |
9faadcc8 PB |
3773 | file = io_uring_get_file(ctx); |
3774 | if (IS_ERR(file)) { | |
3775 | ret = PTR_ERR(file); | |
3776 | goto err; | |
3777 | } | |
3778 | ||
044c1ab3 JA |
3779 | /* |
3780 | * Install ring fd as the very last thing, so we don't risk someone | |
3781 | * having closed it before we finish setup | |
3782 | */ | |
9faadcc8 PB |
3783 | ret = io_uring_install_fd(ctx, file); |
3784 | if (ret < 0) { | |
3785 | /* fput will clean it up */ | |
3786 | fput(file); | |
3787 | return ret; | |
3788 | } | |
044c1ab3 | 3789 | |
c826bd7a | 3790 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
3791 | return ret; |
3792 | err: | |
3793 | io_ring_ctx_wait_and_kill(ctx); | |
3794 | return ret; | |
3795 | } | |
3796 | ||
3797 | /* | |
3798 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3799 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3800 | * params structure passed in. | |
3801 | */ | |
3802 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3803 | { | |
3804 | struct io_uring_params p; | |
2b188cc1 JA |
3805 | int i; |
3806 | ||
3807 | if (copy_from_user(&p, params, sizeof(p))) | |
3808 | return -EFAULT; | |
3809 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3810 | if (p.resv[i]) | |
3811 | return -EINVAL; | |
3812 | } | |
3813 | ||
6c271ce2 | 3814 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 3815 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 3816 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 3817 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 3818 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 | 3819 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
c0e0d6ba | 3820 | IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN)) |
2b188cc1 JA |
3821 | return -EINVAL; |
3822 | ||
ef060ea9 | 3823 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
3824 | } |
3825 | ||
3826 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3827 | struct io_uring_params __user *, params) | |
3828 | { | |
3829 | return io_uring_setup(entries, params); | |
3830 | } | |
3831 | ||
c072481d PB |
3832 | static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, |
3833 | unsigned nr_args) | |
66f4af93 JA |
3834 | { |
3835 | struct io_uring_probe *p; | |
3836 | size_t size; | |
3837 | int i, ret; | |
3838 | ||
3839 | size = struct_size(p, ops, nr_args); | |
3840 | if (size == SIZE_MAX) | |
3841 | return -EOVERFLOW; | |
3842 | p = kzalloc(size, GFP_KERNEL); | |
3843 | if (!p) | |
3844 | return -ENOMEM; | |
3845 | ||
3846 | ret = -EFAULT; | |
3847 | if (copy_from_user(p, arg, size)) | |
3848 | goto out; | |
3849 | ret = -EINVAL; | |
3850 | if (memchr_inv(p, 0, size)) | |
3851 | goto out; | |
3852 | ||
3853 | p->last_op = IORING_OP_LAST - 1; | |
3854 | if (nr_args > IORING_OP_LAST) | |
3855 | nr_args = IORING_OP_LAST; | |
3856 | ||
3857 | for (i = 0; i < nr_args; i++) { | |
3858 | p->ops[i].op = i; | |
3859 | if (!io_op_defs[i].not_supported) | |
3860 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
3861 | } | |
3862 | p->ops_len = i; | |
3863 | ||
3864 | ret = 0; | |
3865 | if (copy_to_user(arg, p, size)) | |
3866 | ret = -EFAULT; | |
3867 | out: | |
3868 | kfree(p); | |
3869 | return ret; | |
3870 | } | |
3871 | ||
071698e1 JA |
3872 | static int io_register_personality(struct io_ring_ctx *ctx) |
3873 | { | |
4379bf8b | 3874 | const struct cred *creds; |
61cf9370 | 3875 | u32 id; |
1e6fa521 | 3876 | int ret; |
071698e1 | 3877 | |
4379bf8b | 3878 | creds = get_current_cred(); |
1e6fa521 | 3879 | |
61cf9370 MWO |
3880 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
3881 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
3882 | if (ret < 0) { |
3883 | put_cred(creds); | |
3884 | return ret; | |
3885 | } | |
3886 | return id; | |
071698e1 JA |
3887 | } |
3888 | ||
c072481d PB |
3889 | static __cold int io_register_restrictions(struct io_ring_ctx *ctx, |
3890 | void __user *arg, unsigned int nr_args) | |
21b55dbc SG |
3891 | { |
3892 | struct io_uring_restriction *res; | |
3893 | size_t size; | |
3894 | int i, ret; | |
3895 | ||
7e84e1c7 SG |
3896 | /* Restrictions allowed only if rings started disabled */ |
3897 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3898 | return -EBADFD; | |
3899 | ||
21b55dbc | 3900 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 3901 | if (ctx->restrictions.registered) |
21b55dbc SG |
3902 | return -EBUSY; |
3903 | ||
3904 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
3905 | return -EINVAL; | |
3906 | ||
3907 | size = array_size(nr_args, sizeof(*res)); | |
3908 | if (size == SIZE_MAX) | |
3909 | return -EOVERFLOW; | |
3910 | ||
3911 | res = memdup_user(arg, size); | |
3912 | if (IS_ERR(res)) | |
3913 | return PTR_ERR(res); | |
3914 | ||
3915 | ret = 0; | |
3916 | ||
3917 | for (i = 0; i < nr_args; i++) { | |
3918 | switch (res[i].opcode) { | |
3919 | case IORING_RESTRICTION_REGISTER_OP: | |
3920 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
3921 | ret = -EINVAL; | |
3922 | goto out; | |
3923 | } | |
3924 | ||
3925 | __set_bit(res[i].register_op, | |
3926 | ctx->restrictions.register_op); | |
3927 | break; | |
3928 | case IORING_RESTRICTION_SQE_OP: | |
3929 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
3930 | ret = -EINVAL; | |
3931 | goto out; | |
3932 | } | |
3933 | ||
3934 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
3935 | break; | |
3936 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
3937 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
3938 | break; | |
3939 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
3940 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
3941 | break; | |
3942 | default: | |
3943 | ret = -EINVAL; | |
3944 | goto out; | |
3945 | } | |
3946 | } | |
3947 | ||
3948 | out: | |
3949 | /* Reset all restrictions if an error happened */ | |
3950 | if (ret != 0) | |
3951 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
3952 | else | |
7e84e1c7 | 3953 | ctx->restrictions.registered = true; |
21b55dbc SG |
3954 | |
3955 | kfree(res); | |
3956 | return ret; | |
3957 | } | |
3958 | ||
7e84e1c7 SG |
3959 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
3960 | { | |
3961 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3962 | return -EBADFD; | |
3963 | ||
bca39f39 | 3964 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) { |
8579538c | 3965 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
bca39f39 PB |
3966 | /* |
3967 | * Lazy activation attempts would fail if it was polled before | |
3968 | * submitter_task is set. | |
3969 | */ | |
3970 | if (wq_has_sleeper(&ctx->poll_wq)) | |
3971 | io_activate_pollwq(ctx); | |
3972 | } | |
7cae596b | 3973 | |
7e84e1c7 SG |
3974 | if (ctx->restrictions.registered) |
3975 | ctx->restricted = 1; | |
3976 | ||
0298ef96 PB |
3977 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
3978 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
3979 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
3980 | return 0; |
3981 | } | |
3982 | ||
c072481d PB |
3983 | static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, |
3984 | void __user *arg, unsigned len) | |
fe76421d JA |
3985 | { |
3986 | struct io_uring_task *tctx = current->io_uring; | |
3987 | cpumask_var_t new_mask; | |
3988 | int ret; | |
3989 | ||
3990 | if (!tctx || !tctx->io_wq) | |
3991 | return -EINVAL; | |
3992 | ||
3993 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
3994 | return -ENOMEM; | |
3995 | ||
3996 | cpumask_clear(new_mask); | |
3997 | if (len > cpumask_size()) | |
3998 | len = cpumask_size(); | |
3999 | ||
0f5e4b83 ES |
4000 | if (in_compat_syscall()) { |
4001 | ret = compat_get_bitmap(cpumask_bits(new_mask), | |
4002 | (const compat_ulong_t __user *)arg, | |
4003 | len * 8 /* CHAR_BIT */); | |
4004 | } else { | |
4005 | ret = copy_from_user(new_mask, arg, len); | |
4006 | } | |
4007 | ||
4008 | if (ret) { | |
fe76421d JA |
4009 | free_cpumask_var(new_mask); |
4010 | return -EFAULT; | |
4011 | } | |
4012 | ||
4013 | ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); | |
4014 | free_cpumask_var(new_mask); | |
4015 | return ret; | |
4016 | } | |
4017 | ||
c072481d | 4018 | static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) |
fe76421d JA |
4019 | { |
4020 | struct io_uring_task *tctx = current->io_uring; | |
4021 | ||
4022 | if (!tctx || !tctx->io_wq) | |
4023 | return -EINVAL; | |
4024 | ||
4025 | return io_wq_cpu_affinity(tctx->io_wq, NULL); | |
4026 | } | |
4027 | ||
c072481d PB |
4028 | static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, |
4029 | void __user *arg) | |
b22fa62a | 4030 | __must_hold(&ctx->uring_lock) |
2e480058 | 4031 | { |
b22fa62a | 4032 | struct io_tctx_node *node; |
fa84693b JA |
4033 | struct io_uring_task *tctx = NULL; |
4034 | struct io_sq_data *sqd = NULL; | |
2e480058 JA |
4035 | __u32 new_count[2]; |
4036 | int i, ret; | |
4037 | ||
2e480058 JA |
4038 | if (copy_from_user(new_count, arg, sizeof(new_count))) |
4039 | return -EFAULT; | |
4040 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
4041 | if (new_count[i] > INT_MAX) | |
4042 | return -EINVAL; | |
4043 | ||
fa84693b JA |
4044 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
4045 | sqd = ctx->sq_data; | |
4046 | if (sqd) { | |
009ad9f0 JA |
4047 | /* |
4048 | * Observe the correct sqd->lock -> ctx->uring_lock | |
4049 | * ordering. Fine to drop uring_lock here, we hold | |
4050 | * a ref to the ctx. | |
4051 | */ | |
41d3a6bd | 4052 | refcount_inc(&sqd->refs); |
009ad9f0 | 4053 | mutex_unlock(&ctx->uring_lock); |
fa84693b | 4054 | mutex_lock(&sqd->lock); |
009ad9f0 | 4055 | mutex_lock(&ctx->uring_lock); |
41d3a6bd JA |
4056 | if (sqd->thread) |
4057 | tctx = sqd->thread->io_uring; | |
fa84693b JA |
4058 | } |
4059 | } else { | |
4060 | tctx = current->io_uring; | |
4061 | } | |
4062 | ||
e139a1ec | 4063 | BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); |
fa84693b | 4064 | |
bad119b9 PB |
4065 | for (i = 0; i < ARRAY_SIZE(new_count); i++) |
4066 | if (new_count[i]) | |
4067 | ctx->iowq_limits[i] = new_count[i]; | |
e139a1ec PB |
4068 | ctx->iowq_limits_set = true; |
4069 | ||
e139a1ec PB |
4070 | if (tctx && tctx->io_wq) { |
4071 | ret = io_wq_max_workers(tctx->io_wq, new_count); | |
4072 | if (ret) | |
4073 | goto err; | |
4074 | } else { | |
4075 | memset(new_count, 0, sizeof(new_count)); | |
4076 | } | |
fa84693b | 4077 | |
41d3a6bd | 4078 | if (sqd) { |
fa84693b | 4079 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
4080 | io_put_sq_data(sqd); |
4081 | } | |
2e480058 JA |
4082 | |
4083 | if (copy_to_user(arg, new_count, sizeof(new_count))) | |
4084 | return -EFAULT; | |
4085 | ||
b22fa62a PB |
4086 | /* that's it for SQPOLL, only the SQPOLL task creates requests */ |
4087 | if (sqd) | |
4088 | return 0; | |
4089 | ||
4090 | /* now propagate the restriction to all registered users */ | |
4091 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
4092 | struct io_uring_task *tctx = node->task->io_uring; | |
4093 | ||
4094 | if (WARN_ON_ONCE(!tctx->io_wq)) | |
4095 | continue; | |
4096 | ||
4097 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
4098 | new_count[i] = ctx->iowq_limits[i]; | |
4099 | /* ignore errors, it always returns zero anyway */ | |
4100 | (void)io_wq_max_workers(tctx->io_wq, new_count); | |
4101 | } | |
2e480058 | 4102 | return 0; |
fa84693b | 4103 | err: |
41d3a6bd | 4104 | if (sqd) { |
fa84693b | 4105 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
4106 | io_put_sq_data(sqd); |
4107 | } | |
fa84693b | 4108 | return ret; |
2e480058 JA |
4109 | } |
4110 | ||
edafccee JA |
4111 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
4112 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
4113 | __releases(ctx->uring_lock) |
4114 | __acquires(ctx->uring_lock) | |
edafccee JA |
4115 | { |
4116 | int ret; | |
4117 | ||
35fa71a0 | 4118 | /* |
fbb8bb02 PB |
4119 | * We don't quiesce the refs for register anymore and so it can't be |
4120 | * dying as we're holding a file ref here. | |
35fa71a0 | 4121 | */ |
fbb8bb02 | 4122 | if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs))) |
35fa71a0 JA |
4123 | return -ENXIO; |
4124 | ||
d7cce96c PB |
4125 | if (ctx->submitter_task && ctx->submitter_task != current) |
4126 | return -EEXIST; | |
4127 | ||
75c4021a | 4128 | if (ctx->restricted) { |
75c4021a PB |
4129 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); |
4130 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
4131 | return -EACCES; | |
4132 | } | |
4133 | ||
edafccee JA |
4134 | switch (opcode) { |
4135 | case IORING_REGISTER_BUFFERS: | |
0184f08e PB |
4136 | ret = -EFAULT; |
4137 | if (!arg) | |
4138 | break; | |
634d00df | 4139 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
4140 | break; |
4141 | case IORING_UNREGISTER_BUFFERS: | |
4142 | ret = -EINVAL; | |
4143 | if (arg || nr_args) | |
4144 | break; | |
0a96bbe4 | 4145 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 4146 | break; |
6b06314c | 4147 | case IORING_REGISTER_FILES: |
a8da73a3 JA |
4148 | ret = -EFAULT; |
4149 | if (!arg) | |
4150 | break; | |
792e3582 | 4151 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
4152 | break; |
4153 | case IORING_UNREGISTER_FILES: | |
4154 | ret = -EINVAL; | |
4155 | if (arg || nr_args) | |
4156 | break; | |
4157 | ret = io_sqe_files_unregister(ctx); | |
4158 | break; | |
c3a31e60 | 4159 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 4160 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 4161 | break; |
9b402849 JA |
4162 | case IORING_REGISTER_EVENTFD: |
4163 | ret = -EINVAL; | |
4164 | if (nr_args != 1) | |
4165 | break; | |
c75312dd UA |
4166 | ret = io_eventfd_register(ctx, arg, 0); |
4167 | break; | |
4168 | case IORING_REGISTER_EVENTFD_ASYNC: | |
4169 | ret = -EINVAL; | |
4170 | if (nr_args != 1) | |
f2842ab5 | 4171 | break; |
c75312dd | 4172 | ret = io_eventfd_register(ctx, arg, 1); |
9b402849 JA |
4173 | break; |
4174 | case IORING_UNREGISTER_EVENTFD: | |
4175 | ret = -EINVAL; | |
4176 | if (arg || nr_args) | |
4177 | break; | |
4178 | ret = io_eventfd_unregister(ctx); | |
4179 | break; | |
66f4af93 JA |
4180 | case IORING_REGISTER_PROBE: |
4181 | ret = -EINVAL; | |
4182 | if (!arg || nr_args > 256) | |
4183 | break; | |
4184 | ret = io_probe(ctx, arg, nr_args); | |
4185 | break; | |
071698e1 JA |
4186 | case IORING_REGISTER_PERSONALITY: |
4187 | ret = -EINVAL; | |
4188 | if (arg || nr_args) | |
4189 | break; | |
4190 | ret = io_register_personality(ctx); | |
4191 | break; | |
4192 | case IORING_UNREGISTER_PERSONALITY: | |
4193 | ret = -EINVAL; | |
4194 | if (arg) | |
4195 | break; | |
4196 | ret = io_unregister_personality(ctx, nr_args); | |
4197 | break; | |
7e84e1c7 SG |
4198 | case IORING_REGISTER_ENABLE_RINGS: |
4199 | ret = -EINVAL; | |
4200 | if (arg || nr_args) | |
4201 | break; | |
4202 | ret = io_register_enable_rings(ctx); | |
4203 | break; | |
21b55dbc SG |
4204 | case IORING_REGISTER_RESTRICTIONS: |
4205 | ret = io_register_restrictions(ctx, arg, nr_args); | |
4206 | break; | |
992da01a PB |
4207 | case IORING_REGISTER_FILES2: |
4208 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
4209 | break; | |
4210 | case IORING_REGISTER_FILES_UPDATE2: | |
4211 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
4212 | IORING_RSRC_FILE); | |
4213 | break; | |
4214 | case IORING_REGISTER_BUFFERS2: | |
4215 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 4216 | break; |
992da01a PB |
4217 | case IORING_REGISTER_BUFFERS_UPDATE: |
4218 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
4219 | IORING_RSRC_BUFFER); | |
c3bdad02 | 4220 | break; |
fe76421d JA |
4221 | case IORING_REGISTER_IOWQ_AFF: |
4222 | ret = -EINVAL; | |
4223 | if (!arg || !nr_args) | |
4224 | break; | |
4225 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
4226 | break; | |
4227 | case IORING_UNREGISTER_IOWQ_AFF: | |
4228 | ret = -EINVAL; | |
4229 | if (arg || nr_args) | |
4230 | break; | |
4231 | ret = io_unregister_iowq_aff(ctx); | |
4232 | break; | |
2e480058 JA |
4233 | case IORING_REGISTER_IOWQ_MAX_WORKERS: |
4234 | ret = -EINVAL; | |
4235 | if (!arg || nr_args != 2) | |
4236 | break; | |
4237 | ret = io_register_iowq_max_workers(ctx, arg); | |
4238 | break; | |
e7a6c00d JA |
4239 | case IORING_REGISTER_RING_FDS: |
4240 | ret = io_ringfd_register(ctx, arg, nr_args); | |
4241 | break; | |
4242 | case IORING_UNREGISTER_RING_FDS: | |
4243 | ret = io_ringfd_unregister(ctx, arg, nr_args); | |
4244 | break; | |
c7fb1942 JA |
4245 | case IORING_REGISTER_PBUF_RING: |
4246 | ret = -EINVAL; | |
4247 | if (!arg || nr_args != 1) | |
4248 | break; | |
4249 | ret = io_register_pbuf_ring(ctx, arg); | |
4250 | break; | |
4251 | case IORING_UNREGISTER_PBUF_RING: | |
4252 | ret = -EINVAL; | |
4253 | if (!arg || nr_args != 1) | |
4254 | break; | |
4255 | ret = io_unregister_pbuf_ring(ctx, arg); | |
4256 | break; | |
78a861b9 JA |
4257 | case IORING_REGISTER_SYNC_CANCEL: |
4258 | ret = -EINVAL; | |
4259 | if (!arg || nr_args != 1) | |
4260 | break; | |
4261 | ret = io_sync_cancel(ctx, arg); | |
4262 | break; | |
6e73dffb PB |
4263 | case IORING_REGISTER_FILE_ALLOC_RANGE: |
4264 | ret = -EINVAL; | |
4265 | if (!arg || nr_args) | |
4266 | break; | |
4267 | ret = io_register_file_alloc_range(ctx, arg); | |
4268 | break; | |
edafccee JA |
4269 | default: |
4270 | ret = -EINVAL; | |
4271 | break; | |
4272 | } | |
4273 | ||
edafccee JA |
4274 | return ret; |
4275 | } | |
4276 | ||
4277 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
4278 | void __user *, arg, unsigned int, nr_args) | |
4279 | { | |
4280 | struct io_ring_ctx *ctx; | |
4281 | long ret = -EBADF; | |
4282 | struct fd f; | |
4283 | ||
34319084 JA |
4284 | if (opcode >= IORING_REGISTER_LAST) |
4285 | return -EINVAL; | |
4286 | ||
edafccee JA |
4287 | f = fdget(fd); |
4288 | if (!f.file) | |
4289 | return -EBADF; | |
4290 | ||
4291 | ret = -EOPNOTSUPP; | |
e5550a14 | 4292 | if (!io_is_uring_fops(f.file)) |
edafccee JA |
4293 | goto out_fput; |
4294 | ||
4295 | ctx = f.file->private_data; | |
4296 | ||
4297 | mutex_lock(&ctx->uring_lock); | |
4298 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
4299 | mutex_unlock(&ctx->uring_lock); | |
2757be22 | 4300 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); |
edafccee JA |
4301 | out_fput: |
4302 | fdput(f); | |
4303 | return ret; | |
4304 | } | |
4305 | ||
2b188cc1 JA |
4306 | static int __init io_uring_init(void) |
4307 | { | |
9c71d39a | 4308 | #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ |
d7f62e82 | 4309 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ |
9c71d39a | 4310 | BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ |
d7f62e82 SM |
4311 | } while (0) |
4312 | ||
4313 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9c71d39a SM |
4314 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) |
4315 | #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ | |
4316 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) | |
d7f62e82 SM |
4317 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); |
4318 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
4319 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
4320 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
4321 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
4322 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
4323 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9c71d39a SM |
4324 | BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); |
4325 | BUILD_BUG_SQE_ELEM(12, __u32, __pad1); | |
d7f62e82 | 4326 | BUILD_BUG_SQE_ELEM(16, __u64, addr); |
7d67af2c | 4327 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
4328 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
4329 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
4330 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
4331 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
4332 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
4333 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
4334 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
4335 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
4336 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
4337 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
4338 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
4339 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
4340 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
4341 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
4342 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 4343 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
9c71d39a SM |
4344 | BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); |
4345 | BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); | |
4346 | BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); | |
4347 | BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); | |
4348 | BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); | |
d7f62e82 SM |
4349 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
4350 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 4351 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 4352 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 4353 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 4354 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
b48c312b PB |
4355 | BUILD_BUG_SQE_ELEM(44, __u16, addr_len); |
4356 | BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); | |
e9621e2b | 4357 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
9c71d39a SM |
4358 | BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); |
4359 | BUILD_BUG_SQE_ELEM(56, __u64, __pad2); | |
d7f62e82 | 4360 | |
b0d658ec PB |
4361 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
4362 | sizeof(struct io_uring_rsrc_update)); | |
4363 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
4364 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
4365 | |
4366 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
4367 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
4368 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
4369 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 4370 | |
b0d658ec PB |
4371 | /* should fit into one byte */ |
4372 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
4373 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
4374 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 4375 | |
32c2d33e | 4376 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); |
16340eab | 4377 | |
3a4b89a2 JA |
4378 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
4379 | ||
d9b57aa3 | 4380 | io_uring_optable_init(); |
0702e536 | 4381 | |
91f245d5 JA |
4382 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | |
4383 | SLAB_ACCOUNT); | |
2b188cc1 JA |
4384 | return 0; |
4385 | }; | |
4386 | __initcall(io_uring_init); |