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 | |
c1755c25 | 154 | struct kmem_cache *req_cachep; |
2b188cc1 | 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); | |
c1755c25 | 233 | kasan_poison_object_data(req_cachep, req); |
a8295b98 HX |
234 | } |
235 | ||
c072481d | 236 | static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) |
2b188cc1 JA |
237 | { |
238 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
239 | ||
0f158b4c | 240 | complete(&ctx->ref_comp); |
2b188cc1 JA |
241 | } |
242 | ||
c072481d | 243 | static __cold void io_fallback_req_func(struct work_struct *work) |
f56165e6 PB |
244 | { |
245 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
246 | fallback_work.work); | |
247 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
248 | struct io_kiocb *req, *tmp; | |
31f084b7 | 249 | bool locked = true; |
f56165e6 | 250 | |
31f084b7 | 251 | mutex_lock(&ctx->uring_lock); |
3218e5d3 | 252 | llist_for_each_entry_safe(req, tmp, node, io_task_work.node) |
f237c30a | 253 | req->io_task_work.func(req, &locked); |
31f084b7 PB |
254 | if (WARN_ON_ONCE(!locked)) |
255 | return; | |
256 | io_submit_flush_completions(ctx); | |
257 | mutex_unlock(&ctx->uring_lock); | |
f56165e6 PB |
258 | } |
259 | ||
e6f89be6 PB |
260 | static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) |
261 | { | |
262 | unsigned hash_buckets = 1U << bits; | |
263 | size_t hash_size = hash_buckets * sizeof(table->hbs[0]); | |
264 | ||
265 | table->hbs = kmalloc(hash_size, GFP_KERNEL); | |
266 | if (!table->hbs) | |
267 | return -ENOMEM; | |
268 | ||
269 | table->hash_bits = bits; | |
270 | init_hash_table(table, hash_buckets); | |
271 | return 0; | |
272 | } | |
273 | ||
c072481d | 274 | static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
2b188cc1 JA |
275 | { |
276 | struct io_ring_ctx *ctx; | |
9cfc7e94 | 277 | int hash_bits; |
2b188cc1 JA |
278 | |
279 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
280 | if (!ctx) | |
281 | return NULL; | |
282 | ||
9cfc7e94 JA |
283 | xa_init(&ctx->io_bl_xa); |
284 | ||
78076bb6 JA |
285 | /* |
286 | * Use 5 bits less than the max cq entries, that should give us around | |
4a07723f PB |
287 | * 32 entries per hash list if totally full and uniformly spread, but |
288 | * don't keep too many buckets to not overconsume memory. | |
78076bb6 | 289 | */ |
4a07723f PB |
290 | hash_bits = ilog2(p->cq_entries) - 5; |
291 | hash_bits = clamp(hash_bits, 1, 8); | |
e6f89be6 | 292 | if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) |
78076bb6 | 293 | goto err; |
9ca9fb24 PB |
294 | if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) |
295 | goto err; | |
38513c46 | 296 | |
6224843d PB |
297 | ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL); |
298 | if (!ctx->dummy_ubuf) | |
299 | goto err; | |
300 | /* set invalid range, so io_import_fixed() fails meeting it */ | |
301 | ctx->dummy_ubuf->ubuf = -1UL; | |
302 | ||
21482896 | 303 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
48904229 | 304 | 0, GFP_KERNEL)) |
206aefde | 305 | goto err; |
2b188cc1 JA |
306 | |
307 | ctx->flags = p->flags; | |
90554200 | 308 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 309 | INIT_LIST_HEAD(&ctx->sqd_list); |
1d7bb1d5 | 310 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
cc3cec83 | 311 | INIT_LIST_HEAD(&ctx->io_buffers_cache); |
9b797a37 | 312 | io_alloc_cache_init(&ctx->apoll_cache); |
43e0bbbd | 313 | io_alloc_cache_init(&ctx->netmsg_cache); |
0f158b4c | 314 | init_completion(&ctx->ref_comp); |
61cf9370 | 315 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 316 | mutex_init(&ctx->uring_lock); |
311997b3 | 317 | init_waitqueue_head(&ctx->cq_wait); |
7b235dd8 | 318 | init_waitqueue_head(&ctx->poll_wq); |
2b188cc1 | 319 | spin_lock_init(&ctx->completion_lock); |
89850fce | 320 | spin_lock_init(&ctx->timeout_lock); |
5eef4e87 | 321 | INIT_WQ_LIST(&ctx->iopoll_list); |
cc3cec83 JA |
322 | INIT_LIST_HEAD(&ctx->io_buffers_pages); |
323 | INIT_LIST_HEAD(&ctx->io_buffers_comp); | |
de0617e4 | 324 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 325 | INIT_LIST_HEAD(&ctx->timeout_list); |
ef9dd637 | 326 | INIT_LIST_HEAD(&ctx->ltimeout_list); |
d67d2263 BM |
327 | spin_lock_init(&ctx->rsrc_ref_lock); |
328 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); | |
269bbe5f | 329 | INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work); |
d34b1b0b | 330 | init_task_work(&ctx->rsrc_put_tw, io_rsrc_put_tw); |
269bbe5f | 331 | init_llist_head(&ctx->rsrc_put_llist); |
c0e0d6ba | 332 | init_llist_head(&ctx->work_llist); |
13bf43f5 | 333 | INIT_LIST_HEAD(&ctx->tctx_list); |
c2b6c6bc PB |
334 | ctx->submit_state.free_list.next = NULL; |
335 | INIT_WQ_LIST(&ctx->locked_free_list); | |
9011bf9a | 336 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
6f33b0bc | 337 | INIT_WQ_LIST(&ctx->submit_state.compl_reqs); |
2b188cc1 | 338 | return ctx; |
206aefde | 339 | err: |
6224843d | 340 | kfree(ctx->dummy_ubuf); |
e6f89be6 | 341 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 342 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 JA |
343 | kfree(ctx->io_bl); |
344 | xa_destroy(&ctx->io_bl_xa); | |
206aefde JA |
345 | kfree(ctx); |
346 | return NULL; | |
2b188cc1 JA |
347 | } |
348 | ||
8f6ed49a PB |
349 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
350 | { | |
351 | struct io_rings *r = ctx->rings; | |
352 | ||
353 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
354 | ctx->cq_extra--; | |
355 | } | |
356 | ||
9cf7c104 | 357 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 358 | { |
2bc9930e JA |
359 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
360 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 361 | |
8f6ed49a | 362 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 363 | } |
de0617e4 | 364 | |
9d858b21 | 365 | return false; |
de0617e4 JA |
366 | } |
367 | ||
9cae36a0 JA |
368 | static inline void io_req_track_inflight(struct io_kiocb *req) |
369 | { | |
370 | if (!(req->flags & REQ_F_INFLIGHT)) { | |
371 | req->flags |= REQ_F_INFLIGHT; | |
386e4fb6 | 372 | atomic_inc(&req->task->io_uring->inflight_tracked); |
9cae36a0 JA |
373 | } |
374 | } | |
375 | ||
fd08e530 PB |
376 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
377 | { | |
906c6caa PB |
378 | if (WARN_ON_ONCE(!req->link)) |
379 | return NULL; | |
380 | ||
4d13d1a4 PB |
381 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
382 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
383 | |
384 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 385 | io_req_set_refcount(req); |
4d13d1a4 PB |
386 | __io_req_set_refcount(req->link, 2); |
387 | return req->link; | |
fd08e530 PB |
388 | } |
389 | ||
390 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
391 | { | |
4d13d1a4 | 392 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
393 | return NULL; |
394 | return __io_prep_linked_timeout(req); | |
395 | } | |
396 | ||
cb2d344c PB |
397 | static noinline void __io_arm_ltimeout(struct io_kiocb *req) |
398 | { | |
399 | io_queue_linked_timeout(__io_prep_linked_timeout(req)); | |
400 | } | |
401 | ||
402 | static inline void io_arm_ltimeout(struct io_kiocb *req) | |
403 | { | |
404 | if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) | |
405 | __io_arm_ltimeout(req); | |
406 | } | |
407 | ||
1e6fa521 JA |
408 | static void io_prep_async_work(struct io_kiocb *req) |
409 | { | |
a7dd2782 | 410 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
1e6fa521 JA |
411 | struct io_ring_ctx *ctx = req->ctx; |
412 | ||
b8e64b53 PB |
413 | if (!(req->flags & REQ_F_CREDS)) { |
414 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 415 | req->creds = get_current_cred(); |
b8e64b53 | 416 | } |
003e8dcc | 417 | |
e1d675df PB |
418 | req->work.list.next = NULL; |
419 | req->work.flags = 0; | |
8e29da69 | 420 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); |
feaadc4f PB |
421 | if (req->flags & REQ_F_FORCE_ASYNC) |
422 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
423 | ||
f6b543fd JA |
424 | if (req->file && !io_req_ffs_set(req)) |
425 | req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT; | |
426 | ||
1e6fa521 JA |
427 | if (req->flags & REQ_F_ISREG) { |
428 | if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL)) | |
429 | io_wq_hash_work(&req->work, file_inode(req->file)); | |
4b982bd0 | 430 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
431 | if (def->unbound_nonreg_file) |
432 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
433 | } | |
561fb04a | 434 | } |
cccf0ee8 | 435 | |
cbdcb435 | 436 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 437 | { |
cbdcb435 | 438 | struct io_kiocb *cur; |
54a91f3b | 439 | |
44eff40a PB |
440 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
441 | struct io_ring_ctx *ctx = req->ctx; | |
442 | ||
674ee8e1 | 443 | spin_lock_irq(&ctx->timeout_lock); |
44eff40a PB |
444 | io_for_each_link(cur, req) |
445 | io_prep_async_work(cur); | |
674ee8e1 | 446 | spin_unlock_irq(&ctx->timeout_lock); |
44eff40a PB |
447 | } else { |
448 | io_for_each_link(cur, req) | |
449 | io_prep_async_work(cur); | |
450 | } | |
561fb04a JA |
451 | } |
452 | ||
f3b44f92 | 453 | void io_queue_iowq(struct io_kiocb *req, bool *dont_use) |
561fb04a | 454 | { |
cbdcb435 | 455 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 456 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 457 | |
3bfe6106 JA |
458 | BUG_ON(!tctx); |
459 | BUG_ON(!tctx->io_wq); | |
561fb04a | 460 | |
cbdcb435 PB |
461 | /* init ->work of the whole link before punting */ |
462 | io_prep_async_link(req); | |
991468dc JA |
463 | |
464 | /* | |
465 | * Not expected to happen, but if we do have a bug where this _can_ | |
466 | * happen, catch it here and ensure the request is marked as | |
467 | * canceled. That will make io-wq go through the usual work cancel | |
468 | * procedure rather than attempt to run this request (or create a new | |
469 | * worker for it). | |
470 | */ | |
471 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
472 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
473 | ||
48863ffd | 474 | trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); |
ebf93667 | 475 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
476 | if (link) |
477 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
478 | } |
479 | ||
c072481d | 480 | static __cold void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 481 | { |
441b8a78 | 482 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
483 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
484 | struct io_defer_entry, list); | |
de0617e4 | 485 | |
9cf7c104 | 486 | if (req_need_defer(de->req, de->seq)) |
04518945 | 487 | break; |
27dc8338 | 488 | list_del_init(&de->list); |
907d1df3 | 489 | io_req_task_queue(de->req); |
27dc8338 | 490 | kfree(de); |
441b8a78 | 491 | } |
04518945 PB |
492 | } |
493 | ||
21a091b9 DY |
494 | |
495 | static void io_eventfd_ops(struct rcu_head *rcu) | |
f2842ab5 | 496 | { |
d8e9214f | 497 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
21a091b9 | 498 | int ops = atomic_xchg(&ev_fd->ops, 0); |
305bef98 | 499 | |
21a091b9 | 500 | if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) |
44648532 | 501 | eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); |
d8e9214f | 502 | |
21a091b9 DY |
503 | /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback |
504 | * ordering in a race but if references are 0 we know we have to free | |
505 | * it regardless. | |
305bef98 | 506 | */ |
21a091b9 DY |
507 | if (atomic_dec_and_test(&ev_fd->refs)) { |
508 | eventfd_ctx_put(ev_fd->cq_ev_fd); | |
509 | kfree(ev_fd); | |
510 | } | |
d8e9214f DY |
511 | } |
512 | ||
77bc59b4 | 513 | static void io_eventfd_signal(struct io_ring_ctx *ctx) |
f2842ab5 | 514 | { |
21a091b9 | 515 | struct io_ev_fd *ev_fd = NULL; |
77bc59b4 | 516 | |
77bc59b4 UA |
517 | rcu_read_lock(); |
518 | /* | |
519 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
520 | * and eventfd_signal | |
521 | */ | |
522 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
523 | ||
524 | /* | |
525 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
526 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
527 | * the function and rcu_read_lock. | |
528 | */ | |
529 | if (unlikely(!ev_fd)) | |
530 | goto out; | |
7e55a19c | 531 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 | 532 | goto out; |
21a091b9 DY |
533 | if (ev_fd->eventfd_async && !io_wq_current_is_worker()) |
534 | goto out; | |
77bc59b4 | 535 | |
21a091b9 | 536 | if (likely(eventfd_signal_allowed())) { |
44648532 | 537 | eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); |
21a091b9 DY |
538 | } else { |
539 | atomic_inc(&ev_fd->refs); | |
540 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) | |
44a84da4 | 541 | call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops); |
21a091b9 DY |
542 | else |
543 | atomic_dec(&ev_fd->refs); | |
544 | } | |
545 | ||
77bc59b4 UA |
546 | out: |
547 | rcu_read_unlock(); | |
f2842ab5 JA |
548 | } |
549 | ||
21a091b9 DY |
550 | static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) |
551 | { | |
552 | bool skip; | |
553 | ||
554 | spin_lock(&ctx->completion_lock); | |
555 | ||
556 | /* | |
557 | * Eventfd should only get triggered when at least one event has been | |
558 | * posted. Some applications rely on the eventfd notification count | |
559 | * only changing IFF a new CQE has been added to the CQ ring. There's | |
560 | * no depedency on 1:1 relationship between how many times this | |
561 | * function is called (and hence the eventfd count) and number of CQEs | |
562 | * posted to the CQ ring. | |
563 | */ | |
564 | skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
565 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
566 | spin_unlock(&ctx->completion_lock); | |
567 | if (skip) | |
568 | return; | |
569 | ||
570 | io_eventfd_signal(ctx); | |
571 | } | |
572 | ||
a830ffd2 PB |
573 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
574 | { | |
bca39f39 PB |
575 | if (ctx->poll_activated) |
576 | io_poll_wq_wake(ctx); | |
e5f30f6f PB |
577 | if (ctx->off_timeout_used) |
578 | io_flush_timeouts(ctx); | |
579 | if (ctx->drain_active) { | |
a830ffd2 | 580 | spin_lock(&ctx->completion_lock); |
e5f30f6f | 581 | io_queue_deferred(ctx); |
a830ffd2 PB |
582 | spin_unlock(&ctx->completion_lock); |
583 | } | |
584 | if (ctx->has_evfd) | |
21a091b9 | 585 | io_eventfd_flush_signal(ctx); |
a830ffd2 PB |
586 | } |
587 | ||
f66f7342 PB |
588 | static inline void __io_cq_lock(struct io_ring_ctx *ctx) |
589 | __acquires(ctx->completion_lock) | |
590 | { | |
591 | if (!ctx->task_complete) | |
592 | spin_lock(&ctx->completion_lock); | |
593 | } | |
594 | ||
595 | static inline void __io_cq_unlock(struct io_ring_ctx *ctx) | |
596 | { | |
597 | if (!ctx->task_complete) | |
598 | spin_unlock(&ctx->completion_lock); | |
599 | } | |
600 | ||
6971253f PB |
601 | static inline void io_cq_lock(struct io_ring_ctx *ctx) |
602 | __acquires(ctx->completion_lock) | |
603 | { | |
604 | spin_lock(&ctx->completion_lock); | |
605 | } | |
606 | ||
607 | static inline void io_cq_unlock(struct io_ring_ctx *ctx) | |
608 | __releases(ctx->completion_lock) | |
609 | { | |
610 | spin_unlock(&ctx->completion_lock); | |
611 | } | |
612 | ||
5d772916 | 613 | /* keep it inlined for io_submit_flush_completions() */ |
f66f7342 | 614 | static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) |
25399321 PB |
615 | __releases(ctx->completion_lock) |
616 | { | |
617 | io_commit_cqring(ctx); | |
f66f7342 | 618 | __io_cq_unlock(ctx); |
6c16fe3c JA |
619 | io_commit_cqring_flush(ctx); |
620 | io_cqring_wake(ctx); | |
25399321 PB |
621 | } |
622 | ||
3181e22f PB |
623 | static inline void __io_cq_unlock_post_flush(struct io_ring_ctx *ctx) |
624 | __releases(ctx->completion_lock) | |
625 | { | |
626 | io_commit_cqring(ctx); | |
627 | __io_cq_unlock(ctx); | |
628 | io_commit_cqring_flush(ctx); | |
629 | ||
630 | /* | |
631 | * As ->task_complete implies that the ring is single tasked, cq_wait | |
632 | * may only be waited on by the current in io_cqring_wait(), but since | |
633 | * it will re-check the wakeup conditions once we return we can safely | |
634 | * skip waking it up. | |
635 | */ | |
636 | if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) { | |
637 | smp_mb(); | |
638 | __io_cqring_wake(ctx); | |
639 | } | |
640 | } | |
641 | ||
25399321 | 642 | void io_cq_unlock_post(struct io_ring_ctx *ctx) |
5d772916 | 643 | __releases(ctx->completion_lock) |
25399321 | 644 | { |
f66f7342 PB |
645 | io_commit_cqring(ctx); |
646 | spin_unlock(&ctx->completion_lock); | |
647 | io_commit_cqring_flush(ctx); | |
648 | io_cqring_wake(ctx); | |
25399321 PB |
649 | } |
650 | ||
c4a2ed72 | 651 | /* Returns true if there are no backlogged entries after the flush */ |
a85381d8 PB |
652 | static void io_cqring_overflow_kill(struct io_ring_ctx *ctx) |
653 | { | |
654 | struct io_overflow_cqe *ocqe; | |
655 | LIST_HEAD(list); | |
656 | ||
657 | io_cq_lock(ctx); | |
658 | list_splice_init(&ctx->cq_overflow_list, &list); | |
659 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); | |
660 | io_cq_unlock(ctx); | |
661 | ||
662 | while (!list_empty(&list)) { | |
663 | ocqe = list_first_entry(&list, struct io_overflow_cqe, list); | |
664 | list_del(&ocqe->list); | |
665 | kfree(ocqe); | |
666 | } | |
667 | } | |
668 | ||
1b346e4a | 669 | static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
1d7bb1d5 | 670 | { |
e45a3e05 | 671 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 672 | |
a85381d8 | 673 | if (__io_cqring_events(ctx) == ctx->cq_entries) |
1b346e4a | 674 | return; |
1d7bb1d5 | 675 | |
e45a3e05 SR |
676 | if (ctx->flags & IORING_SETUP_CQE32) |
677 | cqe_size <<= 1; | |
678 | ||
25399321 | 679 | io_cq_lock(ctx); |
6c2450ae | 680 | while (!list_empty(&ctx->cq_overflow_list)) { |
aa1df3a3 | 681 | struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true); |
6c2450ae | 682 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 683 | |
a85381d8 | 684 | if (!cqe) |
1d7bb1d5 | 685 | break; |
6c2450ae PB |
686 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
687 | struct io_overflow_cqe, list); | |
a85381d8 | 688 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae PB |
689 | list_del(&ocqe->list); |
690 | kfree(ocqe); | |
1d7bb1d5 JA |
691 | } |
692 | ||
1b346e4a | 693 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 694 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 695 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 696 | } |
25399321 | 697 | io_cq_unlock_post(ctx); |
1d7bb1d5 JA |
698 | } |
699 | ||
52ea806a JA |
700 | static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx) |
701 | { | |
702 | /* iopoll syncs against uring_lock, not completion_lock */ | |
703 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
704 | mutex_lock(&ctx->uring_lock); | |
705 | __io_cqring_overflow_flush(ctx); | |
706 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
707 | mutex_unlock(&ctx->uring_lock); | |
708 | } | |
709 | ||
1b346e4a | 710 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 711 | { |
52ea806a JA |
712 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) |
713 | io_cqring_do_overflow_flush(ctx); | |
6c503150 PB |
714 | } |
715 | ||
5afa4650 PB |
716 | /* can be called by any task */ |
717 | static void io_put_task_remote(struct task_struct *task, int nr) | |
6a290a14 PB |
718 | { |
719 | struct io_uring_task *tctx = task->io_uring; | |
720 | ||
9d170164 PB |
721 | percpu_counter_sub(&tctx->inflight, nr); |
722 | if (unlikely(atomic_read(&tctx->in_idle))) | |
723 | wake_up(&tctx->wait); | |
724 | put_task_struct_many(task, nr); | |
725 | } | |
726 | ||
5afa4650 PB |
727 | /* used by a task to put its own references */ |
728 | static void io_put_task_local(struct task_struct *task, int nr) | |
729 | { | |
730 | task->io_uring->cached_refs += nr; | |
731 | } | |
732 | ||
89800a2d PB |
733 | /* must to be called somewhat shortly after putting a request */ |
734 | static inline void io_put_task(struct task_struct *task, int nr) | |
735 | { | |
736 | if (likely(task == current)) | |
5afa4650 | 737 | io_put_task_local(task, nr); |
89800a2d | 738 | else |
5afa4650 | 739 | io_put_task_remote(task, nr); |
89800a2d PB |
740 | } |
741 | ||
63809137 | 742 | void io_task_refs_refill(struct io_uring_task *tctx) |
9a10867a PB |
743 | { |
744 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
745 | ||
746 | percpu_counter_add(&tctx->inflight, refill); | |
747 | refcount_add(refill, ¤t->usage); | |
748 | tctx->cached_refs += refill; | |
749 | } | |
750 | ||
3cc7fdb9 PB |
751 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
752 | { | |
753 | struct io_uring_task *tctx = task->io_uring; | |
754 | unsigned int refs = tctx->cached_refs; | |
755 | ||
756 | if (refs) { | |
757 | tctx->cached_refs = 0; | |
758 | percpu_counter_sub(&tctx->inflight, refs); | |
759 | put_task_struct_many(task, refs); | |
760 | } | |
761 | } | |
762 | ||
68494a65 PB |
763 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
764 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 765 | { |
cce4b8b0 | 766 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
767 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
768 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 769 | |
f26cc959 PB |
770 | lockdep_assert_held(&ctx->completion_lock); |
771 | ||
e45a3e05 SR |
772 | if (is_cqe32) |
773 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 774 | |
e45a3e05 | 775 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 776 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
777 | if (!ocqe) { |
778 | /* | |
779 | * If we're in ring overflow flush mode, or in task cancel mode, | |
780 | * or cannot allocate an overflow entry, then we need to drop it | |
781 | * on the floor. | |
782 | */ | |
8f6ed49a | 783 | io_account_cq_overflow(ctx); |
155bc950 | 784 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 785 | return false; |
2b188cc1 | 786 | } |
cce4b8b0 | 787 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 788 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 789 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 790 | |
cce4b8b0 | 791 | } |
d4d19c19 | 792 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
793 | ocqe->cqe.res = res; |
794 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
795 | if (is_cqe32) { |
796 | ocqe->cqe.big_cqe[0] = extra1; | |
797 | ocqe->cqe.big_cqe[1] = extra2; | |
798 | } | |
cce4b8b0 PB |
799 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
800 | return true; | |
2b188cc1 JA |
801 | } |
802 | ||
68494a65 PB |
803 | bool io_req_cqe_overflow(struct io_kiocb *req) |
804 | { | |
805 | if (!(req->flags & REQ_F_CQE32_INIT)) { | |
806 | req->extra1 = 0; | |
807 | req->extra2 = 0; | |
808 | } | |
809 | return io_cqring_event_overflow(req->ctx, req->cqe.user_data, | |
810 | req->cqe.res, req->cqe.flags, | |
811 | req->extra1, req->extra2); | |
812 | } | |
813 | ||
faf88dde PB |
814 | /* |
815 | * writes to the cq entry need to come after reading head; the | |
816 | * control dependency is enough as we're using WRITE_ONCE to | |
817 | * fill the cq entry | |
818 | */ | |
aa1df3a3 | 819 | struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow) |
faf88dde PB |
820 | { |
821 | struct io_rings *rings = ctx->rings; | |
822 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
823 | unsigned int free, queued, len; |
824 | ||
aa1df3a3 PB |
825 | /* |
826 | * Posting into the CQ when there are pending overflowed CQEs may break | |
827 | * ordering guarantees, which will affect links, F_MORE users and more. | |
828 | * Force overflow the completion. | |
829 | */ | |
830 | if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) | |
831 | return NULL; | |
faf88dde PB |
832 | |
833 | /* userspace may cheat modifying the tail, be safe and do min */ | |
834 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
835 | free = ctx->cq_entries - queued; | |
836 | /* we need a contiguous range, limit based on the current array offset */ | |
837 | len = min(free, ctx->cq_entries - off); | |
838 | if (!len) | |
839 | return NULL; | |
840 | ||
b3659a65 PB |
841 | if (ctx->flags & IORING_SETUP_CQE32) { |
842 | off <<= 1; | |
843 | len <<= 1; | |
844 | } | |
845 | ||
faf88dde PB |
846 | ctx->cqe_cached = &rings->cqes[off]; |
847 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
b3659a65 PB |
848 | |
849 | ctx->cached_cq_tail++; | |
faf88dde | 850 | ctx->cqe_cached++; |
b3659a65 PB |
851 | if (ctx->flags & IORING_SETUP_CQE32) |
852 | ctx->cqe_cached++; | |
853 | return &rings->cqes[off]; | |
faf88dde PB |
854 | } |
855 | ||
f66f7342 PB |
856 | static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, |
857 | u32 cflags) | |
bcda7baa | 858 | { |
cd94903d PB |
859 | struct io_uring_cqe *cqe; |
860 | ||
913a571a | 861 | ctx->cq_extra++; |
cd94903d PB |
862 | |
863 | /* | |
864 | * If we can't get a cq entry, userspace overflowed the | |
865 | * submission (by quite a lot). Increment the overflow count in | |
866 | * the ring. | |
867 | */ | |
868 | cqe = io_get_cqe(ctx); | |
869 | if (likely(cqe)) { | |
e0486f3f DY |
870 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
871 | ||
cd94903d PB |
872 | WRITE_ONCE(cqe->user_data, user_data); |
873 | WRITE_ONCE(cqe->res, res); | |
874 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
875 | |
876 | if (ctx->flags & IORING_SETUP_CQE32) { | |
877 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
878 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
879 | } | |
cd94903d PB |
880 | return true; |
881 | } | |
52120f0f | 882 | return false; |
bcda7baa JA |
883 | } |
884 | ||
931147dd DY |
885 | static void __io_flush_post_cqes(struct io_ring_ctx *ctx) |
886 | __must_hold(&ctx->uring_lock) | |
887 | { | |
888 | struct io_submit_state *state = &ctx->submit_state; | |
889 | unsigned int i; | |
890 | ||
891 | lockdep_assert_held(&ctx->uring_lock); | |
892 | for (i = 0; i < state->cqes_count; i++) { | |
893 | struct io_uring_cqe *cqe = &state->cqes[i]; | |
894 | ||
f66f7342 PB |
895 | if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) { |
896 | if (ctx->task_complete) { | |
897 | spin_lock(&ctx->completion_lock); | |
898 | io_cqring_event_overflow(ctx, cqe->user_data, | |
899 | cqe->res, cqe->flags, 0, 0); | |
900 | spin_unlock(&ctx->completion_lock); | |
901 | } else { | |
902 | io_cqring_event_overflow(ctx, cqe->user_data, | |
903 | cqe->res, cqe->flags, 0, 0); | |
904 | } | |
905 | } | |
931147dd DY |
906 | } |
907 | state->cqes_count = 0; | |
908 | } | |
909 | ||
b529c96a DY |
910 | static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, |
911 | bool allow_overflow) | |
d245bca6 PB |
912 | { |
913 | bool filled; | |
914 | ||
25399321 | 915 | io_cq_lock(ctx); |
f66f7342 PB |
916 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags); |
917 | if (!filled && allow_overflow) | |
918 | filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); | |
919 | ||
25399321 | 920 | io_cq_unlock_post(ctx); |
d245bca6 PB |
921 | return filled; |
922 | } | |
923 | ||
b529c96a DY |
924 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags) |
925 | { | |
926 | return __io_post_aux_cqe(ctx, user_data, res, cflags, true); | |
927 | } | |
928 | ||
9b8c5475 DY |
929 | bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags, |
930 | bool allow_overflow) | |
2b188cc1 | 931 | { |
9b8c5475 DY |
932 | struct io_uring_cqe *cqe; |
933 | unsigned int length; | |
934 | ||
935 | if (!defer) | |
b529c96a | 936 | return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow); |
9b8c5475 DY |
937 | |
938 | length = ARRAY_SIZE(ctx->submit_state.cqes); | |
939 | ||
940 | lockdep_assert_held(&ctx->uring_lock); | |
941 | ||
942 | if (ctx->submit_state.cqes_count == length) { | |
f66f7342 | 943 | __io_cq_lock(ctx); |
9b8c5475 DY |
944 | __io_flush_post_cqes(ctx); |
945 | /* no need to flush - flush is deferred */ | |
f66f7342 | 946 | __io_cq_unlock_post(ctx); |
9b8c5475 DY |
947 | } |
948 | ||
949 | /* For defered completions this is not as strict as it is otherwise, | |
950 | * however it's main job is to prevent unbounded posted completions, | |
951 | * and in that it works just as well. | |
952 | */ | |
953 | if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) | |
954 | return false; | |
955 | ||
956 | cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++]; | |
957 | cqe->user_data = user_data; | |
958 | cqe->res = res; | |
959 | cqe->flags = cflags; | |
960 | return true; | |
961 | } | |
962 | ||
1bec951c | 963 | static void __io_req_complete_post(struct io_kiocb *req) |
2b188cc1 | 964 | { |
fa18fa22 PB |
965 | struct io_ring_ctx *ctx = req->ctx; |
966 | ||
967 | io_cq_lock(ctx); | |
968 | if (!(req->flags & REQ_F_CQE_SKIP)) | |
a8cf95f9 | 969 | io_fill_cqe_req(ctx, req); |
fa18fa22 | 970 | |
c7dae4ba JA |
971 | /* |
972 | * If we're the last reference to this request, add to our locked | |
973 | * free_list cache. | |
974 | */ | |
de9b4cca | 975 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 976 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 977 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
978 | io_disarm_next(req); |
979 | if (req->link) { | |
980 | io_req_task_queue(req->link); | |
981 | req->link = NULL; | |
982 | } | |
983 | } | |
68a2cc1b PB |
984 | io_put_kbuf_comp(req); |
985 | io_dismantle_req(req); | |
7ac1edc4 | 986 | io_req_put_rsrc(req); |
8197b053 PB |
987 | /* |
988 | * Selected buffer deallocation in io_clean_op() assumes that | |
989 | * we don't hold ->completion_lock. Clean them here to avoid | |
990 | * deadlocks. | |
991 | */ | |
5afa4650 | 992 | io_put_task_remote(req->task, 1); |
c2b6c6bc | 993 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 994 | ctx->locked_free_nr++; |
180f829f | 995 | } |
25399321 | 996 | io_cq_unlock_post(ctx); |
4e3d9ff9 JA |
997 | } |
998 | ||
1bec951c | 999 | void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
bcda7baa | 1000 | { |
e6aeb272 PB |
1001 | if (req->ctx->task_complete && (issue_flags & IO_URING_F_IOWQ)) { |
1002 | req->io_task_work.func = io_req_task_complete; | |
1003 | io_req_task_work_add(req); | |
1004 | } else if (!(issue_flags & IO_URING_F_UNLOCKED) || | |
1005 | !(req->ctx->flags & IORING_SETUP_IOPOLL)) { | |
1bec951c PB |
1006 | __io_req_complete_post(req); |
1007 | } else { | |
1008 | struct io_ring_ctx *ctx = req->ctx; | |
1009 | ||
1010 | mutex_lock(&ctx->uring_lock); | |
1011 | __io_req_complete_post(req); | |
1012 | mutex_unlock(&ctx->uring_lock); | |
1013 | } | |
0ddf92e8 JA |
1014 | } |
1015 | ||
973fc83f | 1016 | void io_req_defer_failed(struct io_kiocb *req, s32 res) |
e276ae34 | 1017 | __must_hold(&ctx->uring_lock) |
f41db273 | 1018 | { |
f30bd4d0 | 1019 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; |
a47b255e | 1020 | |
e276ae34 PB |
1021 | lockdep_assert_held(&req->ctx->uring_lock); |
1022 | ||
93d2bcd2 | 1023 | req_set_fail(req); |
97b388d7 | 1024 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
a47b255e PB |
1025 | if (def->fail) |
1026 | def->fail(req); | |
973fc83f | 1027 | io_req_complete_defer(req); |
f41db273 PB |
1028 | } |
1029 | ||
864ea921 PB |
1030 | /* |
1031 | * Don't initialise the fields below on every allocation, but do that in | |
1032 | * advance and keep them valid across allocations. | |
1033 | */ | |
1034 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
1035 | { | |
1036 | req->ctx = ctx; | |
1037 | req->link = NULL; | |
1038 | req->async_data = NULL; | |
1039 | /* not necessary, but safer to zero */ | |
cef216fc | 1040 | req->cqe.res = 0; |
864ea921 PB |
1041 | } |
1042 | ||
dac7a098 | 1043 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 1044 | struct io_submit_state *state) |
dac7a098 | 1045 | { |
79ebeaee | 1046 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1047 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 1048 | ctx->locked_free_nr = 0; |
79ebeaee | 1049 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
1050 | } |
1051 | ||
5d5901a3 PB |
1052 | /* |
1053 | * A request might get retired back into the request caches even before opcode | |
1054 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
1055 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
1056 | * and with extra caution to not get a request that is still worked on. | |
1057 | */ | |
bd1a3783 | 1058 | __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 1059 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 1060 | { |
864ea921 | 1061 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 1062 | void *reqs[IO_REQ_ALLOC_BATCH]; |
864ea921 | 1063 | int ret, i; |
e5d1bc0a | 1064 | |
23a5c43b PB |
1065 | /* |
1066 | * If we have more than a batch's worth of requests in our IRQ side | |
1067 | * locked cache, grab the lock and move them over to our submission | |
1068 | * side cache. | |
1069 | */ | |
a6d97a8a | 1070 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 1071 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 1072 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
1073 | return true; |
1074 | } | |
e5d1bc0a | 1075 | |
3ab665b7 | 1076 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 1077 | |
864ea921 PB |
1078 | /* |
1079 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1080 | * retry single alloc to be on the safe side. | |
1081 | */ | |
1082 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
1083 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
1084 | if (!reqs[0]) | |
a33ae9ce | 1085 | return false; |
864ea921 | 1086 | ret = 1; |
2b188cc1 | 1087 | } |
864ea921 | 1088 | |
37f0e767 | 1089 | percpu_ref_get_many(&ctx->refs, ret); |
3ab665b7 | 1090 | for (i = 0; i < ret; i++) { |
23a5c43b | 1091 | struct io_kiocb *req = reqs[i]; |
3ab665b7 PB |
1092 | |
1093 | io_preinit_req(req, ctx); | |
fa05457a | 1094 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 1095 | } |
a33ae9ce PB |
1096 | return true; |
1097 | } | |
1098 | ||
6b639522 | 1099 | static inline void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 1100 | { |
094bae49 | 1101 | unsigned int flags = req->flags; |
929a3af9 | 1102 | |
867f8fa5 | 1103 | if (unlikely(flags & IO_REQ_CLEAN_FLAGS)) |
3a0a6902 | 1104 | io_clean_op(req); |
e1d767f0 PB |
1105 | if (!(flags & REQ_F_FIXED_FILE)) |
1106 | io_put_file(req->file); | |
e65ef56d JA |
1107 | } |
1108 | ||
59915143 | 1109 | __cold void io_free_req(struct io_kiocb *req) |
c6ca97b3 | 1110 | { |
51a4cc11 | 1111 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 1112 | |
7ac1edc4 | 1113 | io_req_put_rsrc(req); |
216578e5 | 1114 | io_dismantle_req(req); |
5afa4650 | 1115 | io_put_task_remote(req->task, 1); |
c6ca97b3 | 1116 | |
79ebeaee | 1117 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1118 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
c34b025f | 1119 | ctx->locked_free_nr++; |
79ebeaee | 1120 | spin_unlock(&ctx->completion_lock); |
e65ef56d JA |
1121 | } |
1122 | ||
d81499bf PB |
1123 | static void __io_req_find_next_prep(struct io_kiocb *req) |
1124 | { | |
1125 | struct io_ring_ctx *ctx = req->ctx; | |
d81499bf | 1126 | |
6971253f | 1127 | spin_lock(&ctx->completion_lock); |
305bef98 | 1128 | io_disarm_next(req); |
6971253f | 1129 | spin_unlock(&ctx->completion_lock); |
d81499bf PB |
1130 | } |
1131 | ||
1132 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 1133 | { |
33cc89a9 | 1134 | struct io_kiocb *nxt; |
944e58bf | 1135 | |
9e645e11 JA |
1136 | /* |
1137 | * If LINK is set, we have dependent requests in this chain. If we | |
1138 | * didn't fail this request, queue the first one up, moving any other | |
1139 | * dependencies to the next request. In case of failure, fail the rest | |
1140 | * of the chain. | |
1141 | */ | |
d81499bf PB |
1142 | if (unlikely(req->flags & IO_DISARM_MASK)) |
1143 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
1144 | nxt = req->link; |
1145 | req->link = NULL; | |
1146 | return nxt; | |
4d7dd462 | 1147 | } |
9e645e11 | 1148 | |
f237c30a | 1149 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked) |
2c32395d PB |
1150 | { |
1151 | if (!ctx) | |
1152 | return; | |
ef060ea9 JA |
1153 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1154 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
f237c30a | 1155 | if (*locked) { |
c450178d | 1156 | io_submit_flush_completions(ctx); |
2c32395d | 1157 | mutex_unlock(&ctx->uring_lock); |
f237c30a | 1158 | *locked = false; |
2c32395d PB |
1159 | } |
1160 | percpu_ref_put(&ctx->refs); | |
1161 | } | |
1162 | ||
c6dd763c DY |
1163 | static unsigned int handle_tw_list(struct llist_node *node, |
1164 | struct io_ring_ctx **ctx, bool *locked, | |
1165 | struct llist_node *last) | |
9f8d032a | 1166 | { |
c6dd763c DY |
1167 | unsigned int count = 0; |
1168 | ||
3a0c037b | 1169 | while (node != last) { |
f88262e6 | 1170 | struct llist_node *next = node->next; |
9f8d032a HX |
1171 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1172 | io_task_work.node); | |
1173 | ||
34d2bfe7 JA |
1174 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); |
1175 | ||
9f8d032a HX |
1176 | if (req->ctx != *ctx) { |
1177 | ctx_flush_and_put(*ctx, locked); | |
1178 | *ctx = req->ctx; | |
1179 | /* if not contended, grab and improve batching */ | |
1180 | *locked = mutex_trylock(&(*ctx)->uring_lock); | |
1181 | percpu_ref_get(&(*ctx)->refs); | |
1182 | } | |
1183 | req->io_task_work.func(req, locked); | |
1184 | node = next; | |
c6dd763c | 1185 | count++; |
3a0c037b | 1186 | } |
c6dd763c DY |
1187 | |
1188 | return count; | |
9f8d032a HX |
1189 | } |
1190 | ||
923d1592 DY |
1191 | /** |
1192 | * io_llist_xchg - swap all entries in a lock-less list | |
1193 | * @head: the head of lock-less list to delete all entries | |
1194 | * @new: new entry as the head of the list | |
1195 | * | |
1196 | * If list is empty, return NULL, otherwise, return the pointer to the first entry. | |
1197 | * The order of entries returned is from the newest to the oldest added one. | |
1198 | */ | |
1199 | static inline struct llist_node *io_llist_xchg(struct llist_head *head, | |
1200 | struct llist_node *new) | |
1201 | { | |
1202 | return xchg(&head->first, new); | |
1203 | } | |
1204 | ||
1205 | /** | |
1206 | * io_llist_cmpxchg - possibly swap all entries in a lock-less list | |
1207 | * @head: the head of lock-less list to delete all entries | |
1208 | * @old: expected old value of the first entry of the list | |
1209 | * @new: new entry as the head of the list | |
1210 | * | |
1211 | * perform a cmpxchg on the first entry of the list. | |
1212 | */ | |
1213 | ||
1214 | static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head, | |
1215 | struct llist_node *old, | |
1216 | struct llist_node *new) | |
1217 | { | |
1218 | return cmpxchg(&head->first, old, new); | |
1219 | } | |
1220 | ||
c9f06aa7 | 1221 | void tctx_task_work(struct callback_head *cb) |
c40f6379 | 1222 | { |
f28c240e | 1223 | bool uring_locked = false; |
ebd0df2e | 1224 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
1225 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
1226 | task_work); | |
3a0c037b | 1227 | struct llist_node fake = {}; |
77e443ab | 1228 | struct llist_node *node; |
c6dd763c | 1229 | unsigned int loops = 1; |
77e443ab | 1230 | unsigned int count; |
3a0c037b | 1231 | |
77e443ab PB |
1232 | if (unlikely(current->flags & PF_EXITING)) { |
1233 | io_fallback_tw(tctx); | |
1234 | return; | |
1235 | } | |
3a0c037b | 1236 | |
77e443ab PB |
1237 | node = io_llist_xchg(&tctx->task_list, &fake); |
1238 | count = handle_tw_list(node, &ctx, &uring_locked, NULL); | |
3a0c037b DY |
1239 | node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); |
1240 | while (node != &fake) { | |
c6dd763c | 1241 | loops++; |
3a0c037b | 1242 | node = io_llist_xchg(&tctx->task_list, &fake); |
c6dd763c | 1243 | count += handle_tw_list(node, &ctx, &uring_locked, &fake); |
3a0c037b | 1244 | node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); |
3f18407d | 1245 | } |
ebd0df2e | 1246 | |
f28c240e | 1247 | ctx_flush_and_put(ctx, &uring_locked); |
3cc7fdb9 PB |
1248 | |
1249 | /* relaxed read is enough as only the task itself sets ->in_idle */ | |
1250 | if (unlikely(atomic_read(&tctx->in_idle))) | |
1251 | io_uring_drop_tctx_refs(current); | |
c6dd763c DY |
1252 | |
1253 | trace_io_uring_task_work_run(tctx, count, loops); | |
7cbf1722 JA |
1254 | } |
1255 | ||
d7593606 PB |
1256 | static __cold void io_fallback_tw(struct io_uring_task *tctx) |
1257 | { | |
1258 | struct llist_node *node = llist_del_all(&tctx->task_list); | |
1259 | struct io_kiocb *req; | |
1260 | ||
1261 | while (node) { | |
1262 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1263 | node = node->next; | |
1264 | if (llist_add(&req->io_task_work.node, | |
1265 | &req->ctx->fallback_llist)) | |
1266 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1267 | } | |
1268 | } | |
1269 | ||
c0e0d6ba DY |
1270 | static void io_req_local_work_add(struct io_kiocb *req) |
1271 | { | |
1272 | struct io_ring_ctx *ctx = req->ctx; | |
1273 | ||
9ffa13ff PB |
1274 | percpu_ref_get(&ctx->refs); |
1275 | ||
1276 | if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) { | |
1277 | percpu_ref_put(&ctx->refs); | |
c0e0d6ba | 1278 | return; |
9ffa13ff | 1279 | } |
130bd686 | 1280 | /* needed for the following wake up */ |
fc86f9d3 | 1281 | smp_mb__after_atomic(); |
c0e0d6ba DY |
1282 | |
1283 | if (unlikely(atomic_read(&req->task->io_uring->in_idle))) { | |
1284 | io_move_task_work_from_local(ctx); | |
9ffa13ff | 1285 | percpu_ref_put(&ctx->refs); |
c0e0d6ba DY |
1286 | return; |
1287 | } | |
1288 | ||
1289 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) | |
1290 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
21a091b9 DY |
1291 | if (ctx->has_evfd) |
1292 | io_eventfd_signal(ctx); | |
d80c0f00 PB |
1293 | |
1294 | if (READ_ONCE(ctx->cq_waiting)) | |
1295 | wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE); | |
9ffa13ff | 1296 | percpu_ref_put(&ctx->refs); |
c0e0d6ba DY |
1297 | } |
1298 | ||
e52d2e58 | 1299 | void __io_req_task_work_add(struct io_kiocb *req, bool allow_local) |
7cbf1722 | 1300 | { |
c34398a8 | 1301 | struct io_uring_task *tctx = req->task->io_uring; |
9f010507 | 1302 | struct io_ring_ctx *ctx = req->ctx; |
7cbf1722 | 1303 | |
c0e0d6ba DY |
1304 | if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
1305 | io_req_local_work_add(req); | |
1306 | return; | |
1307 | } | |
7cbf1722 JA |
1308 | |
1309 | /* task_work already pending, we're done */ | |
32d91f05 | 1310 | if (!llist_add(&req->io_task_work.node, &tctx->task_list)) |
e09ee510 | 1311 | return; |
7cbf1722 | 1312 | |
ef060ea9 JA |
1313 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1314 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1315 | ||
3fe07bcd | 1316 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1317 | return; |
2215bed9 | 1318 | |
d7593606 | 1319 | io_fallback_tw(tctx); |
c0e0d6ba DY |
1320 | } |
1321 | ||
1322 | static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) | |
1323 | { | |
1324 | struct llist_node *node; | |
1325 | ||
1326 | node = llist_del_all(&ctx->work_llist); | |
1327 | while (node) { | |
1328 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1329 | io_task_work.node); | |
1330 | ||
1331 | node = node->next; | |
1332 | __io_req_task_work_add(req, false); | |
1333 | } | |
1334 | } | |
1335 | ||
360173ab | 1336 | static int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked) |
c0e0d6ba | 1337 | { |
c0e0d6ba | 1338 | struct llist_node *node; |
c3f4d39e | 1339 | unsigned int loops = 0; |
140102ae | 1340 | int ret = 0; |
c0e0d6ba | 1341 | |
140102ae | 1342 | if (WARN_ON_ONCE(ctx->submitter_task != current)) |
c0e0d6ba | 1343 | return -EEXIST; |
c3f4d39e PB |
1344 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1345 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
c0e0d6ba | 1346 | again: |
c3f4d39e PB |
1347 | node = io_llist_xchg(&ctx->work_llist, NULL); |
1348 | while (node) { | |
c0e0d6ba DY |
1349 | struct llist_node *next = node->next; |
1350 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1351 | io_task_work.node); | |
1352 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); | |
b3026767 | 1353 | req->io_task_work.func(req, locked); |
c0e0d6ba DY |
1354 | ret++; |
1355 | node = next; | |
1356 | } | |
c3f4d39e | 1357 | loops++; |
c0e0d6ba | 1358 | |
c3f4d39e | 1359 | if (!llist_empty(&ctx->work_llist)) |
c0e0d6ba | 1360 | goto again; |
b0b7a7d2 | 1361 | if (*locked) { |
c0e0d6ba | 1362 | io_submit_flush_completions(ctx); |
b0b7a7d2 PB |
1363 | if (!llist_empty(&ctx->work_llist)) |
1364 | goto again; | |
1365 | } | |
f75d5036 | 1366 | trace_io_uring_local_work_run(ctx, ret, loops); |
c0e0d6ba | 1367 | return ret; |
8ac5d85a JA |
1368 | } |
1369 | ||
360173ab PB |
1370 | static inline int io_run_local_work_locked(struct io_ring_ctx *ctx) |
1371 | { | |
1372 | bool locked; | |
1373 | int ret; | |
1374 | ||
1375 | if (llist_empty(&ctx->work_llist)) | |
1376 | return 0; | |
1377 | ||
1378 | locked = true; | |
1379 | ret = __io_run_local_work(ctx, &locked); | |
1380 | /* shouldn't happen! */ | |
1381 | if (WARN_ON_ONCE(!locked)) | |
1382 | mutex_lock(&ctx->uring_lock); | |
1383 | return ret; | |
1384 | } | |
1385 | ||
3e565555 | 1386 | static int io_run_local_work(struct io_ring_ctx *ctx) |
8ac5d85a | 1387 | { |
2f413956 | 1388 | bool locked = mutex_trylock(&ctx->uring_lock); |
8ac5d85a JA |
1389 | int ret; |
1390 | ||
b3026767 | 1391 | ret = __io_run_local_work(ctx, &locked); |
8ac5d85a JA |
1392 | if (locked) |
1393 | mutex_unlock(&ctx->uring_lock); | |
1394 | ||
1395 | return ret; | |
c0e0d6ba DY |
1396 | } |
1397 | ||
f237c30a | 1398 | static void io_req_task_cancel(struct io_kiocb *req, bool *locked) |
c40f6379 | 1399 | { |
971cf9c1 | 1400 | io_tw_lock(req->ctx, locked); |
973fc83f | 1401 | io_req_defer_failed(req, req->cqe.res); |
c40f6379 JA |
1402 | } |
1403 | ||
329061d3 | 1404 | void io_req_task_submit(struct io_kiocb *req, bool *locked) |
c40f6379 | 1405 | { |
971cf9c1 | 1406 | io_tw_lock(req->ctx, locked); |
316319e8 | 1407 | /* req->task == current here, checking PF_EXITING is safe */ |
af066f31 | 1408 | if (likely(!(req->task->flags & PF_EXITING))) |
cbc2e203 | 1409 | io_queue_sqe(req); |
81b6d05c | 1410 | else |
973fc83f | 1411 | io_req_defer_failed(req, -EFAULT); |
c40f6379 JA |
1412 | } |
1413 | ||
59915143 | 1414 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1415 | { |
97b388d7 | 1416 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1417 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1418 | io_req_task_work_add(req); |
c40f6379 JA |
1419 | } |
1420 | ||
f3b44f92 | 1421 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1422 | { |
5b0a6acc | 1423 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1424 | io_req_task_work_add(req); |
a3df7698 PB |
1425 | } |
1426 | ||
59915143 | 1427 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1428 | { |
9b5f7bd9 | 1429 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1430 | |
1431 | if (nxt) | |
906a8c3f | 1432 | io_req_task_queue(nxt); |
c69f8dbe JL |
1433 | } |
1434 | ||
f3b44f92 | 1435 | void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node) |
3aa83bfb | 1436 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1437 | { |
d4b7a5ef | 1438 | struct task_struct *task = NULL; |
37f0e767 | 1439 | int task_refs = 0; |
5af1d13e | 1440 | |
3aa83bfb PB |
1441 | do { |
1442 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1443 | comp_list); | |
2d6500d4 | 1444 | |
a538be5b PB |
1445 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1446 | if (req->flags & REQ_F_REFCOUNT) { | |
1447 | node = req->comp_list.next; | |
1448 | if (!req_ref_put_and_test(req)) | |
1449 | continue; | |
1450 | } | |
b605a7fa PB |
1451 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1452 | struct async_poll *apoll = req->apoll; | |
1453 | ||
1454 | if (apoll->double_poll) | |
1455 | kfree(apoll->double_poll); | |
9731bc98 JA |
1456 | if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) |
1457 | kfree(apoll); | |
b605a7fa PB |
1458 | req->flags &= ~REQ_F_POLLED; |
1459 | } | |
da1a08c5 | 1460 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1461 | io_queue_next(req); |
a538be5b PB |
1462 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1463 | io_clean_op(req); | |
c1e53a69 | 1464 | } |
a538be5b PB |
1465 | if (!(req->flags & REQ_F_FIXED_FILE)) |
1466 | io_put_file(req->file); | |
2d6500d4 | 1467 | |
ab409402 | 1468 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1469 | |
d4b7a5ef PB |
1470 | if (req->task != task) { |
1471 | if (task) | |
1472 | io_put_task(task, task_refs); | |
1473 | task = req->task; | |
1474 | task_refs = 0; | |
1475 | } | |
1476 | task_refs++; | |
c1e53a69 | 1477 | node = req->comp_list.next; |
fa05457a | 1478 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1479 | } while (node); |
d4b7a5ef | 1480 | |
d4b7a5ef PB |
1481 | if (task) |
1482 | io_put_task(task, task_refs); | |
7a743e22 PB |
1483 | } |
1484 | ||
c450178d | 1485 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1486 | __must_hold(&ctx->uring_lock) |
905c172f | 1487 | { |
6f33b0bc | 1488 | struct io_wq_work_node *node, *prev; |
cd0ca2e0 | 1489 | struct io_submit_state *state = &ctx->submit_state; |
905c172f | 1490 | |
f66f7342 | 1491 | __io_cq_lock(ctx); |
931147dd DY |
1492 | /* must come first to preserve CQE ordering in failure cases */ |
1493 | if (state->cqes_count) | |
1494 | __io_flush_post_cqes(ctx); | |
d9dee430 PB |
1495 | wq_list_for_each(node, prev, &state->compl_reqs) { |
1496 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1497 | comp_list); | |
3d4aeb9f | 1498 | |
f66f7342 PB |
1499 | if (!(req->flags & REQ_F_CQE_SKIP) && |
1500 | unlikely(!__io_fill_cqe_req(ctx, req))) { | |
1501 | if (ctx->task_complete) { | |
1502 | spin_lock(&ctx->completion_lock); | |
1503 | io_req_cqe_overflow(req); | |
1504 | spin_unlock(&ctx->completion_lock); | |
1505 | } else { | |
1506 | io_req_cqe_overflow(req); | |
1507 | } | |
1508 | } | |
905c172f | 1509 | } |
3181e22f | 1510 | __io_cq_unlock_post_flush(ctx); |
d9dee430 | 1511 | |
931147dd DY |
1512 | if (!wq_list_empty(&ctx->submit_state.compl_reqs)) { |
1513 | io_free_batch_list(ctx, state->compl_reqs.first); | |
1514 | INIT_WQ_LIST(&state->compl_reqs); | |
1515 | } | |
7a743e22 PB |
1516 | } |
1517 | ||
ba816ad6 JA |
1518 | /* |
1519 | * Drop reference to request, return next in chain (if there is one) if this | |
1520 | * was the last reference to this request. | |
1521 | */ | |
0d85035a | 1522 | static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 1523 | { |
9b5f7bd9 PB |
1524 | struct io_kiocb *nxt = NULL; |
1525 | ||
de9b4cca | 1526 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 1527 | if (unlikely(req->flags & IO_REQ_LINK_FLAGS)) |
7819a1f6 | 1528 | nxt = io_req_find_next(req); |
f5c6cf2a | 1529 | io_free_req(req); |
2a44f467 | 1530 | } |
9b5f7bd9 | 1531 | return nxt; |
2b188cc1 JA |
1532 | } |
1533 | ||
6c503150 | 1534 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1535 | { |
1536 | /* See comment at the top of this file */ | |
1537 | smp_rmb(); | |
e23de15f | 1538 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1539 | } |
1540 | ||
def596e9 JA |
1541 | /* |
1542 | * We can't just wait for polled events to come to us, we have to actively | |
1543 | * find and complete them. | |
1544 | */ | |
c072481d | 1545 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1546 | { |
1547 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1548 | return; | |
1549 | ||
1550 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1551 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1552 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1553 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1554 | break; |
08f5439f JA |
1555 | /* |
1556 | * Ensure we allow local-to-the-cpu processing to take place, | |
1557 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1558 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1559 | */ |
3fcee5a6 PB |
1560 | if (need_resched()) { |
1561 | mutex_unlock(&ctx->uring_lock); | |
1562 | cond_resched(); | |
1563 | mutex_lock(&ctx->uring_lock); | |
1564 | } | |
def596e9 JA |
1565 | } |
1566 | mutex_unlock(&ctx->uring_lock); | |
1567 | } | |
1568 | ||
7668b92a | 1569 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1570 | { |
7668b92a | 1571 | unsigned int nr_events = 0; |
e9979b36 | 1572 | int ret = 0; |
155bc950 | 1573 | unsigned long check_cq; |
500f9fba | 1574 | |
76de6749 PB |
1575 | if (!io_allowed_run_tw(ctx)) |
1576 | return -EEXIST; | |
1577 | ||
3a08576b PB |
1578 | check_cq = READ_ONCE(ctx->check_cq); |
1579 | if (unlikely(check_cq)) { | |
1580 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
a85381d8 | 1581 | __io_cqring_overflow_flush(ctx); |
3a08576b PB |
1582 | /* |
1583 | * Similarly do not spin if we have not informed the user of any | |
1584 | * dropped CQE. | |
1585 | */ | |
1586 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1587 | return -EBADR; | |
1588 | } | |
f39c8a5b PB |
1589 | /* |
1590 | * Don't enter poll loop if we already have events pending. | |
1591 | * If we do, we can potentially be spinning for commands that | |
1592 | * already triggered a CQE (eg in error). | |
1593 | */ | |
f39c8a5b | 1594 | if (io_cqring_events(ctx)) |
d487b43c | 1595 | return 0; |
155bc950 | 1596 | |
def596e9 | 1597 | do { |
500f9fba JA |
1598 | /* |
1599 | * If a submit got punted to a workqueue, we can have the | |
1600 | * application entering polling for a command before it gets | |
1601 | * issued. That app will hold the uring_lock for the duration | |
1602 | * of the poll right here, so we need to take a breather every | |
1603 | * now and then to ensure that the issue has a chance to add | |
1604 | * the poll to the issued list. Otherwise we can spin here | |
1605 | * forever, while the workqueue is stuck trying to acquire the | |
1606 | * very same mutex. | |
1607 | */ | |
dac6a0ea JA |
1608 | if (wq_list_empty(&ctx->iopoll_list) || |
1609 | io_task_work_pending(ctx)) { | |
8f487ef2 PB |
1610 | u32 tail = ctx->cached_cq_tail; |
1611 | ||
8de11cdc | 1612 | (void) io_run_local_work_locked(ctx); |
def596e9 | 1613 | |
dac6a0ea JA |
1614 | if (task_work_pending(current) || |
1615 | wq_list_empty(&ctx->iopoll_list)) { | |
dac6a0ea | 1616 | mutex_unlock(&ctx->uring_lock); |
9d54bd6a | 1617 | io_run_task_work(); |
dac6a0ea | 1618 | mutex_lock(&ctx->uring_lock); |
dac6a0ea | 1619 | } |
8f487ef2 PB |
1620 | /* some requests don't go through iopoll_list */ |
1621 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1622 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1623 | break; |
500f9fba | 1624 | } |
5ba3c874 PB |
1625 | ret = io_do_iopoll(ctx, !min); |
1626 | if (ret < 0) | |
1627 | break; | |
1628 | nr_events += ret; | |
1629 | ret = 0; | |
1630 | } while (nr_events < min && !need_resched()); | |
d487b43c | 1631 | |
def596e9 JA |
1632 | return ret; |
1633 | } | |
7012c815 PB |
1634 | |
1635 | void io_req_task_complete(struct io_kiocb *req, bool *locked) | |
8ef12efe | 1636 | { |
7012c815 | 1637 | if (*locked) |
9da070b1 | 1638 | io_req_complete_defer(req); |
7012c815 | 1639 | else |
27f35fe9 | 1640 | io_req_complete_post(req, IO_URING_F_UNLOCKED); |
8ef12efe JA |
1641 | } |
1642 | ||
def596e9 JA |
1643 | /* |
1644 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1645 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1646 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1647 | * accessing the kiocb cookie. |
1648 | */ | |
9882131c | 1649 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1650 | { |
1651 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1652 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1653 | |
1654 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1655 | if (unlikely(needs_lock)) |
cb3d8972 | 1656 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1657 | |
1658 | /* | |
1659 | * Track whether we have multiple files in our lists. This will impact | |
1660 | * how we do polling eventually, not spinning if we're on potentially | |
1661 | * different devices. | |
1662 | */ | |
5eef4e87 | 1663 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1664 | ctx->poll_multi_queue = false; |
1665 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1666 | struct io_kiocb *list_req; |
1667 | ||
5eef4e87 PB |
1668 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1669 | comp_list); | |
30da1b45 | 1670 | if (list_req->file != req->file) |
915b3dde | 1671 | ctx->poll_multi_queue = true; |
def596e9 JA |
1672 | } |
1673 | ||
1674 | /* | |
1675 | * For fast devices, IO may have already completed. If it has, add | |
1676 | * it to the front so we find it first. | |
1677 | */ | |
65a6543d | 1678 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1679 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1680 | else |
5eef4e87 | 1681 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1682 | |
3b44b371 | 1683 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1684 | /* |
1685 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1686 | * in sq thread task context or in io worker task context. If | |
1687 | * current task context is sq thread, we don't need to check | |
1688 | * whether should wake up sq thread. | |
1689 | */ | |
1690 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1691 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1692 | wake_up(&ctx->sq_data->wait); | |
1693 | ||
1694 | mutex_unlock(&ctx->uring_lock); | |
1695 | } | |
def596e9 JA |
1696 | } |
1697 | ||
4503b767 JA |
1698 | static bool io_bdev_nowait(struct block_device *bdev) |
1699 | { | |
568ec936 | 1700 | return !bdev || bdev_nowait(bdev); |
4503b767 JA |
1701 | } |
1702 | ||
2b188cc1 JA |
1703 | /* |
1704 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1705 | * any file. For now, just ensure that anything potentially problematic is done | |
1706 | * inline. | |
1707 | */ | |
88459b50 | 1708 | static bool __io_file_supports_nowait(struct file *file, umode_t mode) |
2b188cc1 | 1709 | { |
4503b767 | 1710 | if (S_ISBLK(mode)) { |
4e7b5671 CH |
1711 | if (IS_ENABLED(CONFIG_BLOCK) && |
1712 | io_bdev_nowait(I_BDEV(file->f_mapping->host))) | |
4503b767 JA |
1713 | return true; |
1714 | return false; | |
1715 | } | |
976517f1 | 1716 | if (S_ISSOCK(mode)) |
2b188cc1 | 1717 | return true; |
4503b767 | 1718 | if (S_ISREG(mode)) { |
4e7b5671 CH |
1719 | if (IS_ENABLED(CONFIG_BLOCK) && |
1720 | io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
e5550a14 | 1721 | !io_is_uring_fops(file)) |
4503b767 JA |
1722 | return true; |
1723 | return false; | |
1724 | } | |
2b188cc1 | 1725 | |
c5b85625 JA |
1726 | /* any ->read/write should understand O_NONBLOCK */ |
1727 | if (file->f_flags & O_NONBLOCK) | |
1728 | return true; | |
35645ac3 | 1729 | return file->f_mode & FMODE_NOWAIT; |
2b188cc1 | 1730 | } |
c5b85625 | 1731 | |
88459b50 PB |
1732 | /* |
1733 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1734 | * any file. For now, just ensure that anything potentially problematic is done | |
1735 | * inline. | |
1736 | */ | |
a4ad4f74 | 1737 | unsigned int io_file_get_flags(struct file *file) |
88459b50 PB |
1738 | { |
1739 | umode_t mode = file_inode(file)->i_mode; | |
1740 | unsigned int res = 0; | |
af197f50 | 1741 | |
88459b50 PB |
1742 | if (S_ISREG(mode)) |
1743 | res |= FFS_ISREG; | |
1744 | if (__io_file_supports_nowait(file, mode)) | |
1745 | res |= FFS_NOWAIT; | |
1746 | return res; | |
2b188cc1 JA |
1747 | } |
1748 | ||
99f15d8d | 1749 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1750 | { |
f30bd4d0 BL |
1751 | WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size); |
1752 | req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL); | |
d886e185 PB |
1753 | if (req->async_data) { |
1754 | req->flags |= REQ_F_ASYNC_DATA; | |
1755 | return false; | |
1756 | } | |
1757 | return true; | |
3d9932a8 XW |
1758 | } |
1759 | ||
f3b44f92 | 1760 | int io_req_prep_async(struct io_kiocb *req) |
f67676d1 | 1761 | { |
f30bd4d0 | 1762 | const struct io_cold_def *cdef = &io_cold_defs[req->opcode]; |
a7dd2782 | 1763 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
0702e536 JA |
1764 | |
1765 | /* assign early for deferred execution for non-fixed file */ | |
1766 | if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE)) | |
1767 | req->file = io_file_get_normal(req, req->cqe.fd); | |
f30bd4d0 | 1768 | if (!cdef->prep_async) |
0702e536 JA |
1769 | return 0; |
1770 | if (WARN_ON_ONCE(req_has_async_data(req))) | |
1771 | return -EFAULT; | |
f30bd4d0 | 1772 | if (!def->manual_alloc) { |
59169439 PB |
1773 | if (io_alloc_async_data(req)) |
1774 | return -EAGAIN; | |
1775 | } | |
f30bd4d0 | 1776 | return cdef->prep_async(req); |
bfe76559 PB |
1777 | } |
1778 | ||
9cf7c104 PB |
1779 | static u32 io_get_sequence(struct io_kiocb *req) |
1780 | { | |
a3dbdf54 | 1781 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1782 | struct io_kiocb *cur; |
9cf7c104 | 1783 | |
a3dbdf54 | 1784 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1785 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1786 | seq--; |
1787 | return seq; | |
9cf7c104 PB |
1788 | } |
1789 | ||
c072481d | 1790 | static __cold void io_drain_req(struct io_kiocb *req) |
e276ae34 | 1791 | __must_hold(&ctx->uring_lock) |
de0617e4 | 1792 | { |
a197f664 | 1793 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1794 | struct io_defer_entry *de; |
f67676d1 | 1795 | int ret; |
e0eb71dc | 1796 | u32 seq = io_get_sequence(req); |
3c19966d | 1797 | |
9d858b21 | 1798 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1799 | spin_lock(&ctx->completion_lock); |
5e371265 | 1800 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1801 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1802 | queue: |
10c66904 | 1803 | ctx->drain_active = false; |
e0eb71dc PB |
1804 | io_req_task_queue(req); |
1805 | return; | |
10c66904 | 1806 | } |
e302f104 | 1807 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1808 | |
cbdcb435 | 1809 | io_prep_async_link(req); |
27dc8338 | 1810 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1811 | if (!de) { |
1b48773f | 1812 | ret = -ENOMEM; |
ef5c600a DY |
1813 | io_req_defer_failed(req, ret); |
1814 | return; | |
76cc33d7 | 1815 | } |
2d28390a | 1816 | |
79ebeaee | 1817 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1818 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1819 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1820 | kfree(de); |
e0eb71dc | 1821 | goto queue; |
de0617e4 JA |
1822 | } |
1823 | ||
48863ffd | 1824 | trace_io_uring_defer(req); |
27dc8338 | 1825 | de->req = req; |
9cf7c104 | 1826 | de->seq = seq; |
27dc8338 | 1827 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1828 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1829 | } |
1830 | ||
68fb8979 | 1831 | static void io_clean_op(struct io_kiocb *req) |
99bc4c38 | 1832 | { |
8197b053 PB |
1833 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
1834 | spin_lock(&req->ctx->completion_lock); | |
cc3cec83 | 1835 | io_put_kbuf_comp(req); |
8197b053 PB |
1836 | spin_unlock(&req->ctx->completion_lock); |
1837 | } | |
99bc4c38 | 1838 | |
0e1b6fe3 | 1839 | if (req->flags & REQ_F_NEED_CLEANUP) { |
f30bd4d0 | 1840 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; |
bb040a21 | 1841 | |
4d4c9cff JA |
1842 | if (def->cleanup) |
1843 | def->cleanup(req); | |
99bc4c38 | 1844 | } |
75652a30 JA |
1845 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1846 | kfree(req->apoll->double_poll); | |
1847 | kfree(req->apoll); | |
1848 | req->apoll = NULL; | |
1849 | } | |
9cae36a0 JA |
1850 | if (req->flags & REQ_F_INFLIGHT) { |
1851 | struct io_uring_task *tctx = req->task->io_uring; | |
1852 | ||
1853 | atomic_dec(&tctx->inflight_tracked); | |
1854 | } | |
c854357b | 1855 | if (req->flags & REQ_F_CREDS) |
b8e64b53 | 1856 | put_cred(req->creds); |
d886e185 PB |
1857 | if (req->flags & REQ_F_ASYNC_DATA) { |
1858 | kfree(req->async_data); | |
1859 | req->async_data = NULL; | |
1860 | } | |
c854357b | 1861 | req->flags &= ~IO_REQ_CLEAN_FLAGS; |
99bc4c38 PB |
1862 | } |
1863 | ||
f4992544 JA |
1864 | static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def, |
1865 | unsigned int issue_flags) | |
6bf9c47a | 1866 | { |
f4992544 | 1867 | if (req->file || !def->needs_file) |
6bf9c47a JA |
1868 | return true; |
1869 | ||
1870 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1871 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1872 | else |
cef216fc | 1873 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1874 | |
772f5e00 | 1875 | return !!req->file; |
6bf9c47a JA |
1876 | } |
1877 | ||
889fca73 | 1878 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1879 | { |
a7dd2782 | 1880 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
5730b27e | 1881 | const struct cred *creds = NULL; |
d625c6ee | 1882 | int ret; |
2b188cc1 | 1883 | |
f4992544 | 1884 | if (unlikely(!io_assign_file(req, def, issue_flags))) |
70152140 JA |
1885 | return -EBADF; |
1886 | ||
6878b40e | 1887 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1888 | creds = override_creds(req->creds); |
5730b27e | 1889 | |
fcde59fe | 1890 | if (!def->audit_skip) |
5bd2182d PM |
1891 | audit_uring_entry(req->opcode); |
1892 | ||
0702e536 | 1893 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1894 | |
fcde59fe | 1895 | if (!def->audit_skip) |
5bd2182d PM |
1896 | audit_uring_exit(!ret, ret); |
1897 | ||
5730b27e JA |
1898 | if (creds) |
1899 | revert_creds(creds); | |
97b388d7 | 1900 | |
75d7b3ae PB |
1901 | if (ret == IOU_OK) { |
1902 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
9da070b1 | 1903 | io_req_complete_defer(req); |
75d7b3ae | 1904 | else |
1bec951c | 1905 | io_req_complete_post(req, issue_flags); |
75d7b3ae | 1906 | } else if (ret != IOU_ISSUE_SKIP_COMPLETE) |
def596e9 | 1907 | return ret; |
97b388d7 | 1908 | |
b532576e | 1909 | /* If the op doesn't have a file, we're not polling for it */ |
ef0ec1ad | 1910 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue) |
9882131c | 1911 | io_iopoll_req_issued(req, issue_flags); |
def596e9 JA |
1912 | |
1913 | return 0; | |
2b188cc1 JA |
1914 | } |
1915 | ||
329061d3 JA |
1916 | int io_poll_issue(struct io_kiocb *req, bool *locked) |
1917 | { | |
1918 | io_tw_lock(req->ctx, locked); | |
9a692451 DY |
1919 | return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT| |
1920 | IO_URING_F_COMPLETE_DEFER); | |
329061d3 JA |
1921 | } |
1922 | ||
c9f06aa7 | 1923 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1924 | { |
1925 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1926 | ||
1927 | req = io_put_req_find_next(req); | |
1928 | return req ? &req->work : NULL; | |
1929 | } | |
1930 | ||
c9f06aa7 | 1931 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1932 | { |
1933 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
a7dd2782 | 1934 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
e6aeb272 | 1935 | unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ; |
d01905db | 1936 | bool needs_poll = false; |
6bf9c47a | 1937 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1938 | |
23a6c9ac | 1939 | /* one will be dropped by ->io_wq_free_work() after returning to io-wq */ |
48dcd38d PB |
1940 | if (!(req->flags & REQ_F_REFCOUNT)) |
1941 | __io_req_set_refcount(req, 2); | |
1942 | else | |
1943 | req_ref_get(req); | |
5d5901a3 | 1944 | |
cb2d344c | 1945 | io_arm_ltimeout(req); |
6bf9c47a | 1946 | |
dadebc35 | 1947 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1948 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1949 | fail: |
6bf9c47a | 1950 | io_req_task_queue_fail(req, err); |
d01905db PB |
1951 | return; |
1952 | } | |
f4992544 | 1953 | if (!io_assign_file(req, def, issue_flags)) { |
0f8da75b PB |
1954 | err = -EBADF; |
1955 | work->flags |= IO_WQ_WORK_CANCEL; | |
1956 | goto fail; | |
1957 | } | |
31b51510 | 1958 | |
d01905db | 1959 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
1960 | bool opcode_poll = def->pollin || def->pollout; |
1961 | ||
1962 | if (opcode_poll && file_can_poll(req->file)) { | |
1963 | needs_poll = true; | |
d01905db | 1964 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 1965 | } |
561fb04a | 1966 | } |
31b51510 | 1967 | |
d01905db PB |
1968 | do { |
1969 | ret = io_issue_sqe(req, issue_flags); | |
1970 | if (ret != -EAGAIN) | |
1971 | break; | |
1972 | /* | |
1973 | * We can get EAGAIN for iopolled IO even though we're | |
1974 | * forcing a sync submission from here, since we can't | |
1975 | * wait for request slots on the block side. | |
1976 | */ | |
1977 | if (!needs_poll) { | |
e0deb6a0 PB |
1978 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
1979 | break; | |
d01905db PB |
1980 | cond_resched(); |
1981 | continue; | |
90fa0288 HX |
1982 | } |
1983 | ||
4d9237e3 | 1984 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
1985 | return; |
1986 | /* aborted or ready, in either case retry blocking */ | |
1987 | needs_poll = false; | |
1988 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
1989 | } while (1); | |
31b51510 | 1990 | |
a3df7698 | 1991 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 1992 | if (ret < 0) |
a3df7698 | 1993 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
1994 | } |
1995 | ||
531113bb JA |
1996 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
1997 | unsigned int issue_flags) | |
09bb8394 | 1998 | { |
5106dd6e JA |
1999 | struct io_ring_ctx *ctx = req->ctx; |
2000 | struct file *file = NULL; | |
ac177053 | 2001 | unsigned long file_ptr; |
09bb8394 | 2002 | |
93f052cb | 2003 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 2004 | |
ac177053 | 2005 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 2006 | goto out; |
ac177053 PB |
2007 | fd = array_index_nospec(fd, ctx->nr_user_files); |
2008 | file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; | |
2009 | file = (struct file *) (file_ptr & FFS_MASK); | |
2010 | file_ptr &= ~FFS_MASK; | |
2011 | /* mask in overlapping REQ_F and FFS bits */ | |
35645ac3 | 2012 | req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT); |
5106dd6e JA |
2013 | io_req_set_rsrc_node(req, ctx, 0); |
2014 | out: | |
93f052cb | 2015 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
2016 | return file; |
2017 | } | |
d44f554e | 2018 | |
531113bb | 2019 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 2020 | { |
62906e89 | 2021 | struct file *file = fget(fd); |
ac177053 | 2022 | |
48863ffd | 2023 | trace_io_uring_file_get(req, fd); |
09bb8394 | 2024 | |
ac177053 | 2025 | /* we don't allow fixed io_uring files */ |
e5550a14 | 2026 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 2027 | io_req_track_inflight(req); |
8371adf5 | 2028 | return file; |
09bb8394 JA |
2029 | } |
2030 | ||
7bfa9bad | 2031 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
2032 | __must_hold(&req->ctx->uring_lock) |
2033 | { | |
7bfa9bad PB |
2034 | struct io_kiocb *linked_timeout; |
2035 | ||
2036 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
973fc83f | 2037 | io_req_defer_failed(req, ret); |
7bfa9bad PB |
2038 | return; |
2039 | } | |
2040 | ||
2041 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 2042 | |
4d9237e3 | 2043 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 2044 | case IO_APOLL_READY: |
336d28a8 | 2045 | io_kbuf_recycle(req, 0); |
d475a9a6 PB |
2046 | io_req_task_queue(req); |
2047 | break; | |
2048 | case IO_APOLL_ABORTED: | |
6436c770 | 2049 | io_kbuf_recycle(req, 0); |
77955efb | 2050 | io_queue_iowq(req, NULL); |
d475a9a6 | 2051 | break; |
b1c62645 | 2052 | case IO_APOLL_OK: |
b1c62645 | 2053 | break; |
d475a9a6 PB |
2054 | } |
2055 | ||
2056 | if (linked_timeout) | |
2057 | io_queue_linked_timeout(linked_timeout); | |
2058 | } | |
2059 | ||
cbc2e203 | 2060 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 2061 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 2062 | { |
e0c5c576 | 2063 | int ret; |
2b188cc1 | 2064 | |
c5eef2b9 | 2065 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 2066 | |
491381ce JA |
2067 | /* |
2068 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2069 | * doesn't support non-blocking read/write attempts | |
2070 | */ | |
7bfa9bad | 2071 | if (likely(!ret)) |
cb2d344c | 2072 | io_arm_ltimeout(req); |
7bfa9bad PB |
2073 | else |
2074 | io_queue_async(req, ret); | |
2b188cc1 JA |
2075 | } |
2076 | ||
4652fe3f | 2077 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 2078 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 2079 | { |
17b147f6 PB |
2080 | if (unlikely(req->flags & REQ_F_FAIL)) { |
2081 | /* | |
2082 | * We don't submit, fail them all, for that replace hardlinks | |
2083 | * with normal links. Extra REQ_F_LINK is tolerated. | |
2084 | */ | |
2085 | req->flags &= ~REQ_F_HARDLINK; | |
2086 | req->flags |= REQ_F_LINK; | |
973fc83f | 2087 | io_req_defer_failed(req, req->cqe.res); |
76cc33d7 PB |
2088 | } else { |
2089 | int ret = io_req_prep_async(req); | |
2090 | ||
ef5c600a | 2091 | if (unlikely(ret)) { |
973fc83f | 2092 | io_req_defer_failed(req, ret); |
ef5c600a DY |
2093 | return; |
2094 | } | |
2095 | ||
2096 | if (unlikely(req->ctx->drain_active)) | |
2097 | io_drain_req(req); | |
76cc33d7 | 2098 | else |
77955efb | 2099 | io_queue_iowq(req, NULL); |
ce35a47a | 2100 | } |
4fe2c963 JL |
2101 | } |
2102 | ||
b16fed66 PB |
2103 | /* |
2104 | * Check SQE restrictions (opcode and flags). | |
2105 | * | |
2106 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
2107 | */ | |
2108 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
2109 | struct io_kiocb *req, | |
2110 | unsigned int sqe_flags) | |
4fe2c963 | 2111 | { |
b16fed66 PB |
2112 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
2113 | return false; | |
2114 | ||
2115 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
2116 | ctx->restrictions.sqe_flags_required) | |
2117 | return false; | |
2118 | ||
2119 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
2120 | ctx->restrictions.sqe_flags_required)) | |
2121 | return false; | |
2122 | ||
2123 | return true; | |
4fe2c963 JL |
2124 | } |
2125 | ||
22b2ca31 PB |
2126 | static void io_init_req_drain(struct io_kiocb *req) |
2127 | { | |
2128 | struct io_ring_ctx *ctx = req->ctx; | |
2129 | struct io_kiocb *head = ctx->submit_state.link.head; | |
2130 | ||
2131 | ctx->drain_active = true; | |
2132 | if (head) { | |
2133 | /* | |
2134 | * If we need to drain a request in the middle of a link, drain | |
2135 | * the head request and the next request/link after the current | |
2136 | * link. Considering sequential execution of links, | |
b6c7db32 | 2137 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
2138 | * link. |
2139 | */ | |
b6c7db32 | 2140 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
2141 | ctx->drain_next = true; |
2142 | } | |
2143 | } | |
2144 | ||
b16fed66 PB |
2145 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2146 | const struct io_uring_sqe *sqe) | |
282cdc86 | 2147 | __must_hold(&ctx->uring_lock) |
b16fed66 | 2148 | { |
a7dd2782 | 2149 | const struct io_issue_def *def; |
b16fed66 | 2150 | unsigned int sqe_flags; |
fc0ae024 | 2151 | int personality; |
4a04d1d1 | 2152 | u8 opcode; |
b16fed66 | 2153 | |
864ea921 | 2154 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 2155 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 PB |
2156 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
2157 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
cef216fc | 2158 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 2159 | req->file = NULL; |
c1bdf8ed | 2160 | req->rsrc_node = NULL; |
b16fed66 | 2161 | req->task = current; |
b16fed66 | 2162 | |
4a04d1d1 PB |
2163 | if (unlikely(opcode >= IORING_OP_LAST)) { |
2164 | req->opcode = 0; | |
b16fed66 | 2165 | return -EINVAL; |
4a04d1d1 | 2166 | } |
a7dd2782 | 2167 | def = &io_issue_defs[opcode]; |
68fe256a PB |
2168 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
2169 | /* enforce forwards compatibility on users */ | |
2170 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
2171 | return -EINVAL; | |
4e906702 | 2172 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 2173 | if (!def->buffer_select) |
4e906702 JA |
2174 | return -EOPNOTSUPP; |
2175 | req->buf_index = READ_ONCE(sqe->buf_group); | |
2176 | } | |
5562a8d7 PB |
2177 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
2178 | ctx->drain_disabled = true; | |
2179 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
2180 | if (ctx->drain_disabled) | |
2181 | return -EOPNOTSUPP; | |
22b2ca31 | 2182 | io_init_req_drain(req); |
5562a8d7 | 2183 | } |
2a56a9bd PB |
2184 | } |
2185 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
2186 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
2187 | return -EACCES; | |
2188 | /* knock it to the slow queue path, will be drained there */ | |
2189 | if (ctx->drain_active) | |
2190 | req->flags |= REQ_F_FORCE_ASYNC; | |
2191 | /* if there is no link, we're at "next" request and need to drain */ | |
2192 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
2193 | ctx->drain_next = false; | |
2194 | ctx->drain_active = true; | |
b6c7db32 | 2195 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 2196 | } |
68fe256a | 2197 | } |
b16fed66 | 2198 | |
fcde59fe | 2199 | if (!def->ioprio && sqe->ioprio) |
73911426 | 2200 | return -EINVAL; |
fcde59fe | 2201 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
73911426 JA |
2202 | return -EINVAL; |
2203 | ||
fcde59fe | 2204 | if (def->needs_file) { |
6d63416d PB |
2205 | struct io_submit_state *state = &ctx->submit_state; |
2206 | ||
cef216fc | 2207 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 2208 | |
6d63416d PB |
2209 | /* |
2210 | * Plug now if we have more than 2 IO left after this, and the | |
2211 | * target is potentially a read/write to block based storage. | |
2212 | */ | |
fcde59fe | 2213 | if (state->need_plug && def->plug) { |
6d63416d PB |
2214 | state->plug_started = true; |
2215 | state->need_plug = false; | |
5ca7a8b3 | 2216 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 2217 | } |
b16fed66 | 2218 | } |
863e0560 | 2219 | |
003e8dcc JA |
2220 | personality = READ_ONCE(sqe->personality); |
2221 | if (personality) { | |
cdab10bf LT |
2222 | int ret; |
2223 | ||
c10d1f98 PB |
2224 | req->creds = xa_load(&ctx->personalities, personality); |
2225 | if (!req->creds) | |
003e8dcc | 2226 | return -EINVAL; |
c10d1f98 | 2227 | get_cred(req->creds); |
cdc1404a PM |
2228 | ret = security_uring_override_creds(req->creds); |
2229 | if (ret) { | |
2230 | put_cred(req->creds); | |
2231 | return ret; | |
2232 | } | |
b8e64b53 | 2233 | req->flags |= REQ_F_CREDS; |
003e8dcc | 2234 | } |
b16fed66 | 2235 | |
0702e536 | 2236 | return def->prep(req, sqe); |
b16fed66 PB |
2237 | } |
2238 | ||
df3becde PB |
2239 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
2240 | struct io_kiocb *req, int ret) | |
2241 | { | |
2242 | struct io_ring_ctx *ctx = req->ctx; | |
2243 | struct io_submit_link *link = &ctx->submit_state.link; | |
2244 | struct io_kiocb *head = link->head; | |
2245 | ||
48863ffd | 2246 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
2247 | |
2248 | /* | |
2249 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
2250 | * unusable. Instead of failing eagerly, continue assembling the link if | |
2251 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
2252 | * should find the flag and handle the rest. | |
2253 | */ | |
2254 | req_fail_link_node(req, ret); | |
2255 | if (head && !(head->flags & REQ_F_FAIL)) | |
2256 | req_fail_link_node(head, -ECANCELED); | |
2257 | ||
2258 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
2259 | if (head) { | |
2260 | link->last->link = req; | |
2261 | link->head = NULL; | |
2262 | req = head; | |
2263 | } | |
2264 | io_queue_sqe_fallback(req); | |
2265 | return ret; | |
2266 | } | |
2267 | ||
2268 | if (head) | |
2269 | link->last->link = req; | |
2270 | else | |
2271 | link->head = req; | |
2272 | link->last = req; | |
2273 | return 0; | |
2274 | } | |
2275 | ||
2276 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 2277 | const struct io_uring_sqe *sqe) |
282cdc86 | 2278 | __must_hold(&ctx->uring_lock) |
9e645e11 | 2279 | { |
a1ab7b35 | 2280 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 2281 | int ret; |
9e645e11 | 2282 | |
a6b8cadc | 2283 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
2284 | if (unlikely(ret)) |
2285 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 2286 | |
be7053b7 | 2287 | /* don't need @sqe from now on */ |
48863ffd | 2288 | trace_io_uring_submit_sqe(req, true); |
a6b8cadc | 2289 | |
9e645e11 JA |
2290 | /* |
2291 | * If we already have a head request, queue this one for async | |
2292 | * submittal once the head completes. If we don't have a head but | |
2293 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2294 | * submitted sync once the chain is complete. If none of those | |
2295 | * conditions are true (normal request), then just queue it. | |
2296 | */ | |
924a07e4 | 2297 | if (unlikely(link->head)) { |
df3becde PB |
2298 | ret = io_req_prep_async(req); |
2299 | if (unlikely(ret)) | |
2300 | return io_submit_fail_init(sqe, req, ret); | |
2301 | ||
48863ffd | 2302 | trace_io_uring_link(req, link->head); |
f2f87370 | 2303 | link->last->link = req; |
863e0560 | 2304 | link->last = req; |
32fe525b | 2305 | |
da1a08c5 | 2306 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 2307 | return 0; |
df3becde PB |
2308 | /* last request of the link, flush it */ |
2309 | req = link->head; | |
f15a3431 | 2310 | link->head = NULL; |
924a07e4 PB |
2311 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
2312 | goto fallback; | |
2313 | ||
2314 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2315 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2316 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2317 | link->head = req; | |
2318 | link->last = req; | |
2319 | } else { | |
2320 | fallback: | |
2321 | io_queue_sqe_fallback(req); | |
2322 | } | |
f15a3431 | 2323 | return 0; |
9e645e11 | 2324 | } |
2e6e1fde | 2325 | |
f15a3431 | 2326 | io_queue_sqe(req); |
1d4240cc | 2327 | return 0; |
9e645e11 JA |
2328 | } |
2329 | ||
9a56a232 JA |
2330 | /* |
2331 | * Batched submission is done, ensure local IO is flushed out. | |
2332 | */ | |
553deffd | 2333 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2334 | { |
553deffd PB |
2335 | struct io_submit_state *state = &ctx->submit_state; |
2336 | ||
e126391c PB |
2337 | if (unlikely(state->link.head)) |
2338 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2339 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2340 | io_submit_flush_completions(ctx); |
27926b68 JA |
2341 | if (state->plug_started) |
2342 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2343 | } |
2344 | ||
2345 | /* | |
2346 | * Start submission side cache. | |
2347 | */ | |
2348 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2349 | unsigned int max_ios) |
9a56a232 | 2350 | { |
27926b68 | 2351 | state->plug_started = false; |
4b628aeb | 2352 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2353 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2354 | /* set only head, no need to init link_last in advance */ |
2355 | state->link.head = NULL; | |
9a56a232 JA |
2356 | } |
2357 | ||
2b188cc1 JA |
2358 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2359 | { | |
75b28aff | 2360 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2361 | |
caf582c6 PB |
2362 | /* |
2363 | * Ensure any loads from the SQEs are done at this point, | |
2364 | * since once we write the new head, the application could | |
2365 | * write new data to them. | |
2366 | */ | |
2367 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2368 | } |
2369 | ||
2b188cc1 | 2370 | /* |
dd9ae8a0 | 2371 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2372 | * that is mapped by userspace. This means that care needs to be taken to |
2373 | * ensure that reads are stable, as we cannot rely on userspace always | |
2374 | * being a good citizen. If members of the sqe are validated and then later | |
2375 | * used, it's important that those reads are done through READ_ONCE() to | |
2376 | * prevent a re-load down the line. | |
2377 | */ | |
b5083dfa | 2378 | static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe) |
2b188cc1 | 2379 | { |
ea5ab3b5 | 2380 | unsigned head, mask = ctx->sq_entries - 1; |
17d3aeb3 | 2381 | unsigned sq_idx = ctx->cached_sq_head++ & mask; |
2b188cc1 JA |
2382 | |
2383 | /* | |
2384 | * The cached sq head (or cq tail) serves two purposes: | |
2385 | * | |
2386 | * 1) allows us to batch the cost of updating the user visible | |
2387 | * head updates. | |
2388 | * 2) allows the kernel side to track the head on its own, even | |
2389 | * though the application is the one updating it. | |
2390 | */ | |
17d3aeb3 | 2391 | head = READ_ONCE(ctx->sq_array[sq_idx]); |
ebdeb7c0 JA |
2392 | if (likely(head < ctx->sq_entries)) { |
2393 | /* double index for 128-byte SQEs, twice as long */ | |
2394 | if (ctx->flags & IORING_SETUP_SQE128) | |
2395 | head <<= 1; | |
b5083dfa PB |
2396 | *sqe = &ctx->sq_sqes[head]; |
2397 | return true; | |
ebdeb7c0 | 2398 | } |
2b188cc1 JA |
2399 | |
2400 | /* drop invalid entries */ | |
15641e42 PB |
2401 | ctx->cq_extra--; |
2402 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2403 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
b5083dfa | 2404 | return false; |
709b302f PB |
2405 | } |
2406 | ||
17437f31 | 2407 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2408 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2409 | { |
69629809 | 2410 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2411 | unsigned int left; |
2412 | int ret; | |
6c271ce2 | 2413 | |
51d48dab | 2414 | if (unlikely(!entries)) |
69629809 | 2415 | return 0; |
ee7d46d9 | 2416 | /* make sure SQ entry isn't read before tail */ |
8e6971a8 PB |
2417 | ret = left = min3(nr, ctx->sq_entries, entries); |
2418 | io_get_task_refs(left); | |
2419 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2420 | |
69629809 | 2421 | do { |
3529d8c2 | 2422 | const struct io_uring_sqe *sqe; |
196be95c | 2423 | struct io_kiocb *req; |
fb5ccc98 | 2424 | |
c8576f3e | 2425 | if (unlikely(!io_alloc_req(ctx, &req))) |
fb5ccc98 | 2426 | break; |
b5083dfa | 2427 | if (unlikely(!io_get_sqe(ctx, &sqe))) { |
fa05457a | 2428 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2429 | break; |
2430 | } | |
6c271ce2 | 2431 | |
1cd15904 PB |
2432 | /* |
2433 | * Continue submitting even for sqe failure if the | |
2434 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2435 | */ | |
2436 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2437 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2438 | left--; | |
2439 | break; | |
bcbb7bf6 | 2440 | } |
1cd15904 | 2441 | } while (--left); |
9466f437 | 2442 | |
8e6971a8 PB |
2443 | if (unlikely(left)) { |
2444 | ret -= left; | |
2445 | /* try again if it submitted nothing and can't allocate a req */ | |
2446 | if (!ret && io_req_cache_empty(ctx)) | |
2447 | ret = -EAGAIN; | |
2448 | current->io_uring->cached_refs += left; | |
9466f437 | 2449 | } |
6c271ce2 | 2450 | |
553deffd | 2451 | io_submit_state_end(ctx); |
ae9428ca PB |
2452 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2453 | io_commit_sqring(ctx); | |
8e6971a8 | 2454 | return ret; |
6c271ce2 JA |
2455 | } |
2456 | ||
bda52162 JA |
2457 | struct io_wait_queue { |
2458 | struct wait_queue_entry wq; | |
2459 | struct io_ring_ctx *ctx; | |
5fd46178 | 2460 | unsigned cq_tail; |
bda52162 | 2461 | unsigned nr_timeouts; |
d33a39e5 | 2462 | ktime_t timeout; |
bda52162 JA |
2463 | }; |
2464 | ||
b4c98d59 DY |
2465 | static inline bool io_has_work(struct io_ring_ctx *ctx) |
2466 | { | |
c0e0d6ba | 2467 | return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) || |
490c00eb | 2468 | !llist_empty(&ctx->work_llist); |
b4c98d59 DY |
2469 | } |
2470 | ||
6c503150 | 2471 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
2472 | { |
2473 | struct io_ring_ctx *ctx = iowq->ctx; | |
0fc8c2ac | 2474 | int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail; |
bda52162 JA |
2475 | |
2476 | /* | |
d195a66e | 2477 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
2478 | * started waiting. For timeouts, we always want to return to userspace, |
2479 | * regardless of event count. | |
2480 | */ | |
5fd46178 | 2481 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
2482 | } |
2483 | ||
2484 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2485 | int wake_flags, void *key) | |
2486 | { | |
bd550173 | 2487 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq); |
bda52162 | 2488 | |
6c503150 PB |
2489 | /* |
2490 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2491 | * the task, and the next invocation will do it. | |
2492 | */ | |
bd550173 | 2493 | if (io_should_wake(iowq) || io_has_work(iowq->ctx)) |
6c503150 PB |
2494 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2495 | return -1; | |
bda52162 JA |
2496 | } |
2497 | ||
c0e0d6ba | 2498 | int io_run_task_work_sig(struct io_ring_ctx *ctx) |
af9c1a44 | 2499 | { |
1414d629 | 2500 | if (!llist_empty(&ctx->work_llist)) { |
2f413956 | 2501 | __set_current_state(TASK_RUNNING); |
1414d629 PB |
2502 | if (io_run_local_work(ctx) > 0) |
2503 | return 1; | |
2504 | } | |
2505 | if (io_run_task_work() > 0) | |
af9c1a44 | 2506 | return 1; |
c5020bc8 OL |
2507 | if (task_sigpending(current)) |
2508 | return -EINTR; | |
2509 | return 0; | |
af9c1a44 JA |
2510 | } |
2511 | ||
eeb60b9a PB |
2512 | /* when returns >0, the caller should retry */ |
2513 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
d33a39e5 | 2514 | struct io_wait_queue *iowq) |
eeb60b9a | 2515 | { |
3fcf19d5 PB |
2516 | if (unlikely(READ_ONCE(ctx->check_cq))) |
2517 | return 1; | |
846072f1 PB |
2518 | if (unlikely(!llist_empty(&ctx->work_llist))) |
2519 | return 1; | |
2520 | if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) | |
2521 | return 1; | |
2522 | if (unlikely(task_sigpending(current))) | |
2523 | return -EINTR; | |
2524 | if (unlikely(io_should_wake(iowq))) | |
2525 | return 0; | |
d33a39e5 | 2526 | if (iowq->timeout == KTIME_MAX) |
46ae7eef | 2527 | schedule(); |
d33a39e5 | 2528 | else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS)) |
22833966 | 2529 | return -ETIME; |
846072f1 | 2530 | return 0; |
eeb60b9a PB |
2531 | } |
2532 | ||
2b188cc1 JA |
2533 | /* |
2534 | * Wait until events become available, if we don't already have some. The | |
2535 | * application must reap them itself, as they reside on the shared cq ring. | |
2536 | */ | |
2537 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2538 | const sigset_t __user *sig, size_t sigsz, |
2539 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2540 | { |
90291099 | 2541 | struct io_wait_queue iowq; |
75b28aff | 2542 | struct io_rings *rings = ctx->rings; |
c1d5a224 | 2543 | int ret; |
2b188cc1 | 2544 | |
76de6749 PB |
2545 | if (!io_allowed_run_tw(ctx)) |
2546 | return -EEXIST; | |
140102ae PB |
2547 | if (!llist_empty(&ctx->work_llist)) |
2548 | io_run_local_work(ctx); | |
f36ba6cf PB |
2549 | io_run_task_work(); |
2550 | io_cqring_overflow_flush(ctx); | |
2551 | /* if user messes with these they will just get an early return */ | |
2552 | if (__io_cqring_events_user(ctx) >= min_events) | |
2553 | return 0; | |
2b188cc1 JA |
2554 | |
2555 | if (sig) { | |
9e75ad5d AB |
2556 | #ifdef CONFIG_COMPAT |
2557 | if (in_compat_syscall()) | |
2558 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2559 | sigsz); |
9e75ad5d AB |
2560 | else |
2561 | #endif | |
b772434b | 2562 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2563 | |
2b188cc1 JA |
2564 | if (ret) |
2565 | return ret; | |
2566 | } | |
2567 | ||
90291099 PB |
2568 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
2569 | iowq.wq.private = current; | |
2570 | INIT_LIST_HEAD(&iowq.wq.entry); | |
2571 | iowq.ctx = ctx; | |
bda52162 | 2572 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 2573 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
d33a39e5 PB |
2574 | iowq.timeout = KTIME_MAX; |
2575 | ||
2576 | if (uts) { | |
2577 | struct timespec64 ts; | |
2578 | ||
2579 | if (get_timespec64(&ts, uts)) | |
2580 | return -EFAULT; | |
2581 | iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); | |
2582 | } | |
90291099 | 2583 | |
c826bd7a | 2584 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2585 | do { |
3fcf19d5 PB |
2586 | unsigned long check_cq; |
2587 | ||
130bd686 | 2588 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
d80c0f00 | 2589 | WRITE_ONCE(ctx->cq_waiting, 1); |
130bd686 PB |
2590 | set_current_state(TASK_INTERRUPTIBLE); |
2591 | } else { | |
2592 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, | |
2593 | TASK_INTERRUPTIBLE); | |
2594 | } | |
2595 | ||
d33a39e5 | 2596 | ret = io_cqring_wait_schedule(ctx, &iowq); |
130bd686 | 2597 | __set_current_state(TASK_RUNNING); |
d80c0f00 PB |
2598 | WRITE_ONCE(ctx->cq_waiting, 0); |
2599 | ||
846072f1 PB |
2600 | if (ret < 0) |
2601 | break; | |
2602 | /* | |
2603 | * Run task_work after scheduling and before io_should_wake(). | |
2604 | * If we got woken because of task_work being processed, run it | |
2605 | * now rather than let the caller do another wait loop. | |
2606 | */ | |
2607 | io_run_task_work(); | |
2608 | if (!llist_empty(&ctx->work_llist)) | |
2609 | io_run_local_work(ctx); | |
3fcf19d5 PB |
2610 | |
2611 | check_cq = READ_ONCE(ctx->check_cq); | |
2612 | if (unlikely(check_cq)) { | |
2613 | /* let the caller flush overflows, retry */ | |
326a9e48 | 2614 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) |
3fcf19d5 | 2615 | io_cqring_do_overflow_flush(ctx); |
3fcf19d5 PB |
2616 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) { |
2617 | ret = -EBADR; | |
2618 | break; | |
2619 | } | |
2620 | } | |
2621 | ||
846072f1 PB |
2622 | if (io_should_wake(&iowq)) { |
2623 | ret = 0; | |
35d90f95 | 2624 | break; |
846072f1 | 2625 | } |
ca0a2651 | 2626 | cond_resched(); |
846072f1 | 2627 | } while (1); |
bda52162 | 2628 | |
130bd686 PB |
2629 | if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) |
2630 | finish_wait(&ctx->cq_wait, &iowq.wq); | |
b7db41c9 | 2631 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2632 | |
75b28aff | 2633 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2634 | } |
2635 | ||
73572984 | 2636 | static void io_mem_free(void *ptr) |
b9bd2bea | 2637 | { |
73572984 | 2638 | struct page *page; |
b36a2050 | 2639 | |
73572984 JA |
2640 | if (!ptr) |
2641 | return; | |
b9bd2bea | 2642 | |
73572984 JA |
2643 | page = virt_to_head_page(ptr); |
2644 | if (put_page_testzero(page)) | |
2645 | free_compound_page(page); | |
b9bd2bea PB |
2646 | } |
2647 | ||
73572984 | 2648 | static void *io_mem_alloc(size_t size) |
b9bd2bea | 2649 | { |
73572984 | 2650 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; |
b9bd2bea | 2651 | |
73572984 | 2652 | return (void *) __get_free_pages(gfp, get_order(size)); |
b9bd2bea PB |
2653 | } |
2654 | ||
73572984 JA |
2655 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2656 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2657 | { |
73572984 JA |
2658 | struct io_rings *rings; |
2659 | size_t off, sq_array_size; | |
6b06314c | 2660 | |
73572984 JA |
2661 | off = struct_size(rings, cqes, cq_entries); |
2662 | if (off == SIZE_MAX) | |
2663 | return SIZE_MAX; | |
2664 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2665 | if (check_shl_overflow(off, 1, &off)) | |
2666 | return SIZE_MAX; | |
2667 | } | |
ab409402 | 2668 | |
73572984 JA |
2669 | #ifdef CONFIG_SMP |
2670 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2671 | if (off == 0) | |
2672 | return SIZE_MAX; | |
2673 | #endif | |
82fbcfa9 | 2674 | |
73572984 JA |
2675 | if (sq_offset) |
2676 | *sq_offset = off; | |
82fbcfa9 | 2677 | |
73572984 JA |
2678 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2679 | if (sq_array_size == SIZE_MAX) | |
2680 | return SIZE_MAX; | |
6b06314c | 2681 | |
73572984 JA |
2682 | if (check_add_overflow(off, sq_array_size, &off)) |
2683 | return SIZE_MAX; | |
8bad28d8 | 2684 | |
73572984 | 2685 | return off; |
8bad28d8 HX |
2686 | } |
2687 | ||
73572984 JA |
2688 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, |
2689 | unsigned int eventfd_async) | |
8bad28d8 | 2690 | { |
73572984 JA |
2691 | struct io_ev_fd *ev_fd; |
2692 | __s32 __user *fds = arg; | |
2693 | int fd; | |
f2303b1f | 2694 | |
73572984 JA |
2695 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, |
2696 | lockdep_is_held(&ctx->uring_lock)); | |
2697 | if (ev_fd) | |
2698 | return -EBUSY; | |
8bad28d8 | 2699 | |
73572984 JA |
2700 | if (copy_from_user(&fd, fds, sizeof(*fds))) |
2701 | return -EFAULT; | |
8dd03afe | 2702 | |
73572984 JA |
2703 | ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); |
2704 | if (!ev_fd) | |
2705 | return -ENOMEM; | |
05f3fb3c | 2706 | |
73572984 JA |
2707 | ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); |
2708 | if (IS_ERR(ev_fd->cq_ev_fd)) { | |
2709 | int ret = PTR_ERR(ev_fd->cq_ev_fd); | |
2710 | kfree(ev_fd); | |
2711 | return ret; | |
2712 | } | |
305bef98 PB |
2713 | |
2714 | spin_lock(&ctx->completion_lock); | |
2715 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
2716 | spin_unlock(&ctx->completion_lock); | |
2717 | ||
73572984 JA |
2718 | ev_fd->eventfd_async = eventfd_async; |
2719 | ctx->has_evfd = true; | |
2720 | rcu_assign_pointer(ctx->io_ev_fd, ev_fd); | |
21a091b9 DY |
2721 | atomic_set(&ev_fd->refs, 1); |
2722 | atomic_set(&ev_fd->ops, 0); | |
73572984 | 2723 | return 0; |
d7954b2b BM |
2724 | } |
2725 | ||
73572984 | 2726 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 2727 | { |
73572984 JA |
2728 | struct io_ev_fd *ev_fd; |
2729 | ||
2730 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, | |
2731 | lockdep_is_held(&ctx->uring_lock)); | |
2732 | if (ev_fd) { | |
2733 | ctx->has_evfd = false; | |
2734 | rcu_assign_pointer(ctx->io_ev_fd, NULL); | |
21a091b9 DY |
2735 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops)) |
2736 | call_rcu(&ev_fd->rcu, io_eventfd_ops); | |
73572984 JA |
2737 | return 0; |
2738 | } | |
2d091d62 | 2739 | |
73572984 | 2740 | return -ENXIO; |
44b31f2f PB |
2741 | } |
2742 | ||
73572984 | 2743 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2744 | { |
c8576f3e | 2745 | struct io_kiocb *req; |
37f0e767 | 2746 | int nr = 0; |
bf019da7 | 2747 | |
9a4fdbd8 | 2748 | mutex_lock(&ctx->uring_lock); |
34f0bc42 | 2749 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
9a4fdbd8 | 2750 | |
88ab95be | 2751 | while (!io_req_cache_empty(ctx)) { |
c8576f3e | 2752 | req = io_extract_req(ctx); |
c2b6c6bc | 2753 | kmem_cache_free(req_cachep, req); |
37f0e767 | 2754 | nr++; |
c2b6c6bc | 2755 | } |
37f0e767 PB |
2756 | if (nr) |
2757 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2758 | mutex_unlock(&ctx->uring_lock); |
2759 | } | |
2760 | ||
c072481d | 2761 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2762 | { |
37d1e2e3 | 2763 | io_sq_thread_finish(ctx); |
ab409402 | 2764 | io_rsrc_refs_drop(ctx); |
43597aac PB |
2765 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
2766 | io_wait_rsrc_data(ctx->buf_data); | |
2767 | io_wait_rsrc_data(ctx->file_data); | |
2768 | ||
8bad28d8 | 2769 | mutex_lock(&ctx->uring_lock); |
43597aac | 2770 | if (ctx->buf_data) |
bd54b6fe | 2771 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2772 | if (ctx->file_data) |
08480400 | 2773 | __io_sqe_files_unregister(ctx); |
a85381d8 | 2774 | io_cqring_overflow_kill(ctx); |
9b402849 | 2775 | io_eventfd_unregister(ctx); |
9b797a37 | 2776 | io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); |
43e0bbbd | 2777 | io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); |
77bc59b4 | 2778 | mutex_unlock(&ctx->uring_lock); |
5a2e745d | 2779 | io_destroy_buffers(ctx); |
07db298a PB |
2780 | if (ctx->sq_creds) |
2781 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2782 | if (ctx->submitter_task) |
2783 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2784 | |
a7f0ed5a PB |
2785 | /* there are no registered resources left, nobody uses it */ |
2786 | if (ctx->rsrc_node) | |
2787 | io_rsrc_node_destroy(ctx->rsrc_node); | |
8dd03afe | 2788 | if (ctx->rsrc_backup_node) |
b895c9a6 | 2789 | io_rsrc_node_destroy(ctx->rsrc_backup_node); |
a7f0ed5a | 2790 | flush_delayed_work(&ctx->rsrc_put_work); |
756ab7c0 | 2791 | flush_delayed_work(&ctx->fallback_work); |
a7f0ed5a PB |
2792 | |
2793 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
2794 | WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist)); | |
def596e9 | 2795 | |
2b188cc1 | 2796 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
2797 | if (ctx->ring_sock) { |
2798 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 2799 | sock_release(ctx->ring_sock); |
355e8d26 | 2800 | } |
2b188cc1 | 2801 | #endif |
ef9dd637 | 2802 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
2b188cc1 | 2803 | |
42b6419d PB |
2804 | if (ctx->mm_account) { |
2805 | mmdrop(ctx->mm_account); | |
2806 | ctx->mm_account = NULL; | |
2807 | } | |
75b28aff | 2808 | io_mem_free(ctx->rings); |
2b188cc1 | 2809 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
2810 | |
2811 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2812 | free_uid(ctx->user); |
4010fec4 | 2813 | io_req_caches_free(ctx); |
e941894e JA |
2814 | if (ctx->hash_map) |
2815 | io_wq_put_hash(ctx->hash_map); | |
e6f89be6 | 2816 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2817 | kfree(ctx->cancel_table_locked.hbs); |
6224843d | 2818 | kfree(ctx->dummy_ubuf); |
9cfc7e94 JA |
2819 | kfree(ctx->io_bl); |
2820 | xa_destroy(&ctx->io_bl_xa); | |
2b188cc1 JA |
2821 | kfree(ctx); |
2822 | } | |
2823 | ||
bca39f39 PB |
2824 | static __cold void io_activate_pollwq_cb(struct callback_head *cb) |
2825 | { | |
2826 | struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx, | |
2827 | poll_wq_task_work); | |
2828 | ||
2829 | mutex_lock(&ctx->uring_lock); | |
2830 | ctx->poll_activated = true; | |
2831 | mutex_unlock(&ctx->uring_lock); | |
2832 | ||
2833 | /* | |
2834 | * Wake ups for some events between start of polling and activation | |
2835 | * might've been lost due to loose synchronisation. | |
2836 | */ | |
2837 | wake_up_all(&ctx->poll_wq); | |
2838 | percpu_ref_put(&ctx->refs); | |
2839 | } | |
2840 | ||
2841 | static __cold void io_activate_pollwq(struct io_ring_ctx *ctx) | |
2842 | { | |
2843 | spin_lock(&ctx->completion_lock); | |
2844 | /* already activated or in progress */ | |
2845 | if (ctx->poll_activated || ctx->poll_wq_task_work.func) | |
2846 | goto out; | |
2847 | if (WARN_ON_ONCE(!ctx->task_complete)) | |
2848 | goto out; | |
2849 | if (!ctx->submitter_task) | |
2850 | goto out; | |
2851 | /* | |
2852 | * with ->submitter_task only the submitter task completes requests, we | |
2853 | * only need to sync with it, which is done by injecting a tw | |
2854 | */ | |
2855 | init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb); | |
2856 | percpu_ref_get(&ctx->refs); | |
2857 | if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL)) | |
2858 | percpu_ref_put(&ctx->refs); | |
2859 | out: | |
2860 | spin_unlock(&ctx->completion_lock); | |
2861 | } | |
2862 | ||
2b188cc1 JA |
2863 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) |
2864 | { | |
2865 | struct io_ring_ctx *ctx = file->private_data; | |
2866 | __poll_t mask = 0; | |
2867 | ||
bca39f39 PB |
2868 | if (unlikely(!ctx->poll_activated)) |
2869 | io_activate_pollwq(ctx); | |
2870 | ||
7b235dd8 | 2871 | poll_wait(file, &ctx->poll_wq, wait); |
4f7067c3 SB |
2872 | /* |
2873 | * synchronizes with barrier from wq_has_sleeper call in | |
2874 | * io_commit_cqring | |
2875 | */ | |
2b188cc1 | 2876 | smp_rmb(); |
90554200 | 2877 | if (!io_sqring_full(ctx)) |
2b188cc1 | 2878 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
2879 | |
2880 | /* | |
2881 | * Don't flush cqring overflow list here, just do a simple check. | |
2882 | * Otherwise there could possible be ABBA deadlock: | |
2883 | * CPU0 CPU1 | |
2884 | * ---- ---- | |
2885 | * lock(&ctx->uring_lock); | |
2886 | * lock(&ep->mtx); | |
2887 | * lock(&ctx->uring_lock); | |
2888 | * lock(&ep->mtx); | |
2889 | * | |
2890 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
10d8bc35 | 2891 | * pushes them to do the flush. |
ed670c3f | 2892 | */ |
b4c98d59 | 2893 | |
c10bb646 | 2894 | if (__io_cqring_events_user(ctx) || io_has_work(ctx)) |
2b188cc1 JA |
2895 | mask |= EPOLLIN | EPOLLRDNORM; |
2896 | ||
2897 | return mask; | |
2898 | } | |
2899 | ||
0bead8cd | 2900 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 2901 | { |
4379bf8b | 2902 | const struct cred *creds; |
071698e1 | 2903 | |
61cf9370 | 2904 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
2905 | if (creds) { |
2906 | put_cred(creds); | |
0bead8cd | 2907 | return 0; |
1e6fa521 | 2908 | } |
0bead8cd YD |
2909 | |
2910 | return -EINVAL; | |
2911 | } | |
2912 | ||
d56d938b PB |
2913 | struct io_tctx_exit { |
2914 | struct callback_head task_work; | |
2915 | struct completion completion; | |
baf186c4 | 2916 | struct io_ring_ctx *ctx; |
d56d938b PB |
2917 | }; |
2918 | ||
c072481d | 2919 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
2920 | { |
2921 | struct io_uring_task *tctx = current->io_uring; | |
2922 | struct io_tctx_exit *work; | |
2923 | ||
2924 | work = container_of(cb, struct io_tctx_exit, task_work); | |
2925 | /* | |
2926 | * When @in_idle, we're in cancellation and it's racy to remove the | |
2927 | * node. It'll be removed by the end of cancellation, just ignore it. | |
998b30c3 HM |
2928 | * tctx can be NULL if the queueing of this task_work raced with |
2929 | * work cancelation off the exec path. | |
d56d938b | 2930 | */ |
998b30c3 | 2931 | if (tctx && !atomic_read(&tctx->in_idle)) |
eef51daa | 2932 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
2933 | complete(&work->completion); |
2934 | } | |
2935 | ||
c072481d | 2936 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
2937 | { |
2938 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2939 | ||
2940 | return req->ctx == data; | |
2941 | } | |
2942 | ||
c072481d | 2943 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 2944 | { |
d56d938b | 2945 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 2946 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 2947 | unsigned long interval = HZ / 20; |
d56d938b PB |
2948 | struct io_tctx_exit exit; |
2949 | struct io_tctx_node *node; | |
2950 | int ret; | |
85faa7b8 | 2951 | |
56952e91 JA |
2952 | /* |
2953 | * If we're doing polled IO and end up having requests being | |
2954 | * submitted async (out-of-line), then completions can come in while | |
2955 | * we're waiting for refs to drop. We need to reap these manually, | |
2956 | * as nobody else will be looking for them. | |
2957 | */ | |
b2edc0a7 | 2958 | do { |
a85381d8 PB |
2959 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
2960 | mutex_lock(&ctx->uring_lock); | |
2961 | io_cqring_overflow_kill(ctx); | |
2962 | mutex_unlock(&ctx->uring_lock); | |
2963 | } | |
2964 | ||
c0e0d6ba DY |
2965 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) |
2966 | io_move_task_work_from_local(ctx); | |
2967 | ||
affa87db PB |
2968 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
2969 | cond_resched(); | |
2970 | ||
28090c13 PB |
2971 | if (ctx->sq_data) { |
2972 | struct io_sq_data *sqd = ctx->sq_data; | |
2973 | struct task_struct *tsk; | |
2974 | ||
2975 | io_sq_thread_park(sqd); | |
2976 | tsk = sqd->thread; | |
2977 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
2978 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
2979 | io_cancel_ctx_cb, ctx, true); | |
2980 | io_sq_thread_unpark(sqd); | |
2981 | } | |
b5bb3a24 | 2982 | |
37f0e767 PB |
2983 | io_req_caches_free(ctx); |
2984 | ||
58d3be2c PB |
2985 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
2986 | /* there is little hope left, don't run it too often */ | |
2987 | interval = HZ * 60; | |
2988 | } | |
2989 | } while (!wait_for_completion_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 2990 | |
7f00651a PB |
2991 | init_completion(&exit.completion); |
2992 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
2993 | exit.ctx = ctx; | |
89b5066e PB |
2994 | /* |
2995 | * Some may use context even when all refs and requests have been put, | |
2996 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 2997 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
2998 | * this lock/unlock section also waits them to finish. |
2999 | */ | |
d56d938b PB |
3000 | mutex_lock(&ctx->uring_lock); |
3001 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
3002 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
3003 | ||
d56d938b PB |
3004 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
3005 | ctx_node); | |
7f00651a PB |
3006 | /* don't spin on a single task if cancellation failed */ |
3007 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
3008 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
3009 | if (WARN_ON_ONCE(ret)) | |
3010 | continue; | |
d56d938b PB |
3011 | |
3012 | mutex_unlock(&ctx->uring_lock); | |
3013 | wait_for_completion(&exit.completion); | |
d56d938b PB |
3014 | mutex_lock(&ctx->uring_lock); |
3015 | } | |
3016 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
3017 | spin_lock(&ctx->completion_lock); |
3018 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 3019 | |
85faa7b8 JA |
3020 | io_ring_ctx_free(ctx); |
3021 | } | |
3022 | ||
c072481d | 3023 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 3024 | { |
61cf9370 MWO |
3025 | unsigned long index; |
3026 | struct creds *creds; | |
3027 | ||
2b188cc1 JA |
3028 | mutex_lock(&ctx->uring_lock); |
3029 | percpu_ref_kill(&ctx->refs); | |
61cf9370 MWO |
3030 | xa_for_each(&ctx->personalities, index, creds) |
3031 | io_unregister_personality(ctx, index); | |
9ca9fb24 PB |
3032 | if (ctx->rings) |
3033 | io_poll_remove_all(ctx, NULL, true); | |
2b188cc1 JA |
3034 | mutex_unlock(&ctx->uring_lock); |
3035 | ||
02bac94b PB |
3036 | /* |
3037 | * If we failed setting up the ctx, we might not have any rings | |
3038 | * and therefore did not submit any requests | |
3039 | */ | |
3040 | if (ctx->rings) | |
60053be8 | 3041 | io_kill_timeouts(ctx, NULL, true); |
309fc03a | 3042 | |
85faa7b8 | 3043 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
3044 | /* |
3045 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
3046 | * if we're exiting a ton of rings at the same time. It just adds | |
3047 | * noise and overhead, there's no discernable change in runtime | |
3048 | * over using system_wq. | |
3049 | */ | |
3050 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
3051 | } |
3052 | ||
3053 | static int io_uring_release(struct inode *inode, struct file *file) | |
3054 | { | |
3055 | struct io_ring_ctx *ctx = file->private_data; | |
3056 | ||
3057 | file->private_data = NULL; | |
3058 | io_ring_ctx_wait_and_kill(ctx); | |
3059 | return 0; | |
3060 | } | |
3061 | ||
f6edbabb PB |
3062 | struct io_task_cancel { |
3063 | struct task_struct *task; | |
3dd0c97a | 3064 | bool all; |
f6edbabb | 3065 | }; |
f254ac04 | 3066 | |
f6edbabb | 3067 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 3068 | { |
9a472ef7 | 3069 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 3070 | struct io_task_cancel *cancel = data; |
9a472ef7 | 3071 | |
6af3f48b | 3072 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
3073 | } |
3074 | ||
c072481d PB |
3075 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
3076 | struct task_struct *task, | |
3077 | bool cancel_all) | |
b7ddce3c | 3078 | { |
e1915f76 | 3079 | struct io_defer_entry *de; |
b7ddce3c PB |
3080 | LIST_HEAD(list); |
3081 | ||
79ebeaee | 3082 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 3083 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 3084 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
3085 | list_cut_position(&list, &ctx->defer_list, &de->list); |
3086 | break; | |
3087 | } | |
3088 | } | |
79ebeaee | 3089 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
3090 | if (list_empty(&list)) |
3091 | return false; | |
b7ddce3c PB |
3092 | |
3093 | while (!list_empty(&list)) { | |
3094 | de = list_first_entry(&list, struct io_defer_entry, list); | |
3095 | list_del_init(&de->list); | |
e276ae34 | 3096 | io_req_task_queue_fail(de->req, -ECANCELED); |
b7ddce3c PB |
3097 | kfree(de); |
3098 | } | |
e1915f76 | 3099 | return true; |
b7ddce3c PB |
3100 | } |
3101 | ||
c072481d | 3102 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
3103 | { |
3104 | struct io_tctx_node *node; | |
3105 | enum io_wq_cancel cret; | |
3106 | bool ret = false; | |
3107 | ||
3108 | mutex_lock(&ctx->uring_lock); | |
3109 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3110 | struct io_uring_task *tctx = node->task->io_uring; | |
3111 | ||
3112 | /* | |
3113 | * io_wq will stay alive while we hold uring_lock, because it's | |
3114 | * killed after ctx nodes, which requires to take the lock. | |
3115 | */ | |
3116 | if (!tctx || !tctx->io_wq) | |
3117 | continue; | |
3118 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
3119 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3120 | } | |
3121 | mutex_unlock(&ctx->uring_lock); | |
3122 | ||
3123 | return ret; | |
3124 | } | |
3125 | ||
affa87db | 3126 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
3127 | struct task_struct *task, |
3128 | bool cancel_all) | |
9936c7c2 | 3129 | { |
3dd0c97a | 3130 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 3131 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
3132 | enum io_wq_cancel cret; |
3133 | bool ret = false; | |
9936c7c2 | 3134 | |
60053be8 PB |
3135 | /* failed during ring init, it couldn't have issued any requests */ |
3136 | if (!ctx->rings) | |
affa87db | 3137 | return false; |
60053be8 | 3138 | |
affa87db PB |
3139 | if (!task) { |
3140 | ret |= io_uring_try_cancel_iowq(ctx); | |
3141 | } else if (tctx && tctx->io_wq) { | |
3142 | /* | |
3143 | * Cancels requests of all rings, not only @ctx, but | |
3144 | * it's fine as the task is in exit/exec. | |
3145 | */ | |
3146 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
3147 | &cancel, true); | |
3148 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3149 | } | |
9936c7c2 | 3150 | |
affa87db PB |
3151 | /* SQPOLL thread does its own polling */ |
3152 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
3153 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
3154 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
3155 | io_iopoll_try_reap_events(ctx); | |
3156 | ret = true; | |
9936c7c2 | 3157 | } |
9936c7c2 | 3158 | } |
affa87db | 3159 | |
140102ae PB |
3160 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3161 | io_allowed_defer_tw_run(ctx)) | |
c0e0d6ba | 3162 | ret |= io_run_local_work(ctx) > 0; |
affa87db PB |
3163 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
3164 | mutex_lock(&ctx->uring_lock); | |
3165 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
3166 | mutex_unlock(&ctx->uring_lock); | |
3167 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
3168 | if (task) | |
c0e0d6ba | 3169 | ret |= io_run_task_work() > 0; |
affa87db | 3170 | return ret; |
9936c7c2 PB |
3171 | } |
3172 | ||
3f48cf18 | 3173 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 3174 | { |
3f48cf18 | 3175 | if (tracked) |
9cae36a0 | 3176 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
3177 | return percpu_counter_sum(&tctx->inflight); |
3178 | } | |
3179 | ||
78cc687b PB |
3180 | /* |
3181 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 3182 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 3183 | */ |
17437f31 | 3184 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 3185 | { |
521d6a73 | 3186 | struct io_uring_task *tctx = current->io_uring; |
734551df | 3187 | struct io_ring_ctx *ctx; |
0e9ddb39 PB |
3188 | s64 inflight; |
3189 | DEFINE_WAIT(wait); | |
fdaf083c | 3190 | |
78cc687b PB |
3191 | WARN_ON_ONCE(sqd && sqd->thread != current); |
3192 | ||
6d042ffb PO |
3193 | if (!current->io_uring) |
3194 | return; | |
17a91051 PB |
3195 | if (tctx->io_wq) |
3196 | io_wq_exit_start(tctx->io_wq); | |
3197 | ||
0e9ddb39 PB |
3198 | atomic_inc(&tctx->in_idle); |
3199 | do { | |
affa87db PB |
3200 | bool loop = false; |
3201 | ||
e9dbe221 | 3202 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 3203 | /* read completions before cancelations */ |
78cc687b | 3204 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
3205 | if (!inflight) |
3206 | break; | |
fdaf083c | 3207 | |
78cc687b PB |
3208 | if (!sqd) { |
3209 | struct io_tctx_node *node; | |
3210 | unsigned long index; | |
0f212204 | 3211 | |
78cc687b PB |
3212 | xa_for_each(&tctx->xa, index, node) { |
3213 | /* sqpoll task will cancel all its requests */ | |
3214 | if (node->ctx->sq_data) | |
3215 | continue; | |
affa87db PB |
3216 | loop |= io_uring_try_cancel_requests(node->ctx, |
3217 | current, cancel_all); | |
78cc687b PB |
3218 | } |
3219 | } else { | |
3220 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
3221 | loop |= io_uring_try_cancel_requests(ctx, |
3222 | current, | |
3223 | cancel_all); | |
3224 | } | |
3225 | ||
3226 | if (loop) { | |
3227 | cond_resched(); | |
3228 | continue; | |
78cc687b | 3229 | } |
17a91051 | 3230 | |
78a78060 JA |
3231 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
3232 | io_run_task_work(); | |
e9dbe221 | 3233 | io_uring_drop_tctx_refs(current); |
78a78060 | 3234 | |
0f212204 | 3235 | /* |
a1bb3cd5 PB |
3236 | * If we've seen completions, retry without waiting. This |
3237 | * avoids a race where a completion comes in before we did | |
3238 | * prepare_to_wait(). | |
0f212204 | 3239 | */ |
3dd0c97a | 3240 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 3241 | schedule(); |
f57555ed | 3242 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 3243 | } while (1); |
de7f1d9e | 3244 | |
8452d4a6 | 3245 | io_uring_clean_tctx(tctx); |
3dd0c97a | 3246 | if (cancel_all) { |
3cc7fdb9 PB |
3247 | /* |
3248 | * We shouldn't run task_works after cancel, so just leave | |
3249 | * ->in_idle set for normal exit. | |
3250 | */ | |
3251 | atomic_dec(&tctx->in_idle); | |
3f48cf18 PB |
3252 | /* for exec all current's requests should be gone, kill tctx */ |
3253 | __io_uring_free(current); | |
3254 | } | |
44e728b8 PB |
3255 | } |
3256 | ||
f552a27a | 3257 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 3258 | { |
f552a27a | 3259 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
3260 | } |
3261 | ||
6c5c240e RP |
3262 | static void *io_uring_validate_mmap_request(struct file *file, |
3263 | loff_t pgoff, size_t sz) | |
2b188cc1 | 3264 | { |
2b188cc1 | 3265 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 3266 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
3267 | struct page *page; |
3268 | void *ptr; | |
3269 | ||
3270 | switch (offset) { | |
3271 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
3272 | case IORING_OFF_CQ_RING: |
3273 | ptr = ctx->rings; | |
2b188cc1 JA |
3274 | break; |
3275 | case IORING_OFF_SQES: | |
3276 | ptr = ctx->sq_sqes; | |
3277 | break; | |
2b188cc1 | 3278 | default: |
6c5c240e | 3279 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
3280 | } |
3281 | ||
3282 | page = virt_to_head_page(ptr); | |
a50b854e | 3283 | if (sz > page_size(page)) |
6c5c240e RP |
3284 | return ERR_PTR(-EINVAL); |
3285 | ||
3286 | return ptr; | |
3287 | } | |
3288 | ||
3289 | #ifdef CONFIG_MMU | |
3290 | ||
c072481d | 3291 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e RP |
3292 | { |
3293 | size_t sz = vma->vm_end - vma->vm_start; | |
3294 | unsigned long pfn; | |
3295 | void *ptr; | |
3296 | ||
3297 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
3298 | if (IS_ERR(ptr)) | |
3299 | return PTR_ERR(ptr); | |
2b188cc1 JA |
3300 | |
3301 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3302 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3303 | } | |
3304 | ||
6c5c240e RP |
3305 | #else /* !CONFIG_MMU */ |
3306 | ||
3307 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3308 | { | |
3309 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
3310 | } | |
3311 | ||
3312 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
3313 | { | |
3314 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
3315 | } | |
3316 | ||
3317 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
3318 | unsigned long addr, unsigned long len, | |
3319 | unsigned long pgoff, unsigned long flags) | |
3320 | { | |
3321 | void *ptr; | |
3322 | ||
3323 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
3324 | if (IS_ERR(ptr)) | |
3325 | return PTR_ERR(ptr); | |
3326 | ||
3327 | return (unsigned long) ptr; | |
3328 | } | |
3329 | ||
3330 | #endif /* !CONFIG_MMU */ | |
3331 | ||
f81440d3 PB |
3332 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
3333 | { | |
3334 | if (flags & IORING_ENTER_EXT_ARG) { | |
3335 | struct io_uring_getevents_arg arg; | |
3336 | ||
3337 | if (argsz != sizeof(arg)) | |
3338 | return -EINVAL; | |
3339 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3340 | return -EFAULT; | |
3341 | } | |
3342 | return 0; | |
3343 | } | |
3344 | ||
c73ebb68 HX |
3345 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
3346 | struct __kernel_timespec __user **ts, | |
3347 | const sigset_t __user **sig) | |
3348 | { | |
3349 | struct io_uring_getevents_arg arg; | |
3350 | ||
3351 | /* | |
3352 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
3353 | * is just a pointer to the sigset_t. | |
3354 | */ | |
3355 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
3356 | *sig = (const sigset_t __user *) argp; | |
3357 | *ts = NULL; | |
3358 | return 0; | |
3359 | } | |
3360 | ||
3361 | /* | |
3362 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
3363 | * timespec and sigset_t pointers if good. | |
3364 | */ | |
3365 | if (*argsz != sizeof(arg)) | |
3366 | return -EINVAL; | |
3367 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3368 | return -EFAULT; | |
d2347b96 DY |
3369 | if (arg.pad) |
3370 | return -EINVAL; | |
c73ebb68 HX |
3371 | *sig = u64_to_user_ptr(arg.sigmask); |
3372 | *argsz = arg.sigmask_sz; | |
3373 | *ts = u64_to_user_ptr(arg.ts); | |
3374 | return 0; | |
3375 | } | |
3376 | ||
2b188cc1 | 3377 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
3378 | u32, min_complete, u32, flags, const void __user *, argp, |
3379 | size_t, argsz) | |
2b188cc1 JA |
3380 | { |
3381 | struct io_ring_ctx *ctx; | |
2b188cc1 | 3382 | struct fd f; |
33f993da | 3383 | long ret; |
2b188cc1 | 3384 | |
33f993da | 3385 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
3386 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
3387 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
3388 | return -EINVAL; |
3389 | ||
e7a6c00d JA |
3390 | /* |
3391 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3392 | * need only dereference our task private array to find it. | |
3393 | */ | |
3394 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3395 | struct io_uring_task *tctx = current->io_uring; | |
3396 | ||
3273c440 | 3397 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
e7a6c00d JA |
3398 | return -EINVAL; |
3399 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
3400 | f.file = tctx->registered_rings[fd]; | |
4329490a | 3401 | f.flags = 0; |
3273c440 PB |
3402 | if (unlikely(!f.file)) |
3403 | return -EBADF; | |
e7a6c00d JA |
3404 | } else { |
3405 | f = fdget(fd); | |
3273c440 PB |
3406 | if (unlikely(!f.file)) |
3407 | return -EBADF; | |
3408 | ret = -EOPNOTSUPP; | |
3409 | if (unlikely(!io_is_uring_fops(f.file))) | |
fbb8bb02 | 3410 | goto out; |
e7a6c00d | 3411 | } |
2b188cc1 | 3412 | |
2b188cc1 | 3413 | ctx = f.file->private_data; |
7e84e1c7 | 3414 | ret = -EBADFD; |
33f993da | 3415 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3416 | goto out; |
3417 | ||
6c271ce2 JA |
3418 | /* |
3419 | * For SQ polling, the thread will do all submissions and completions. | |
3420 | * Just return the requested submit count, and wake the thread if | |
3421 | * we were asked to. | |
3422 | */ | |
b2a9eada | 3423 | ret = 0; |
6c271ce2 | 3424 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3425 | io_cqring_overflow_flush(ctx); |
89448c47 | 3426 | |
21f96522 JA |
3427 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3428 | ret = -EOWNERDEAD; | |
04147488 | 3429 | goto out; |
21f96522 | 3430 | } |
6c271ce2 | 3431 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3432 | wake_up(&ctx->sq_data->wait); |
88b80534 QF |
3433 | if (flags & IORING_ENTER_SQ_WAIT) |
3434 | io_sqpoll_wait_sq(ctx); | |
3435 | ||
3e813c90 | 3436 | ret = to_submit; |
b2a9eada | 3437 | } else if (to_submit) { |
eef51daa | 3438 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3439 | if (unlikely(ret)) |
3440 | goto out; | |
7c504e65 | 3441 | |
2b188cc1 | 3442 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3443 | ret = io_submit_sqes(ctx, to_submit); |
3444 | if (ret != to_submit) { | |
d487b43c | 3445 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3446 | goto out; |
d487b43c | 3447 | } |
44f87745 PB |
3448 | if (flags & IORING_ENTER_GETEVENTS) { |
3449 | if (ctx->syscall_iopoll) | |
3450 | goto iopoll_locked; | |
3451 | /* | |
3452 | * Ignore errors, we'll soon call io_cqring_wait() and | |
3453 | * it should handle ownership problems if any. | |
3454 | */ | |
3455 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3456 | (void)io_run_local_work_locked(ctx); | |
3457 | } | |
d487b43c | 3458 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 | 3459 | } |
c0e0d6ba | 3460 | |
2b188cc1 | 3461 | if (flags & IORING_ENTER_GETEVENTS) { |
3e813c90 | 3462 | int ret2; |
c0e0d6ba | 3463 | |
773697b6 | 3464 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3465 | /* |
3466 | * We disallow the app entering submit/complete with | |
3467 | * polling, but we still need to lock the ring to | |
3468 | * prevent racing with polled issue that got punted to | |
3469 | * a workqueue. | |
3470 | */ | |
3471 | mutex_lock(&ctx->uring_lock); | |
3472 | iopoll_locked: | |
3e813c90 DY |
3473 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3474 | if (likely(!ret2)) { | |
3475 | min_complete = min(min_complete, | |
3476 | ctx->cq_entries); | |
3477 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3478 | } |
3479 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3480 | } else { |
f81440d3 PB |
3481 | const sigset_t __user *sig; |
3482 | struct __kernel_timespec __user *ts; | |
3483 | ||
3e813c90 DY |
3484 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3485 | if (likely(!ret2)) { | |
3486 | min_complete = min(min_complete, | |
3487 | ctx->cq_entries); | |
3488 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3489 | argsz, ts); | |
3490 | } | |
def596e9 | 3491 | } |
c73ebb68 | 3492 | |
155bc950 | 3493 | if (!ret) { |
3e813c90 | 3494 | ret = ret2; |
2b188cc1 | 3495 | |
155bc950 DY |
3496 | /* |
3497 | * EBADR indicates that one or more CQE were dropped. | |
3498 | * Once the user has been informed we can clear the bit | |
3499 | * as they are obviously ok with those drops. | |
3500 | */ | |
3501 | if (unlikely(ret2 == -EBADR)) | |
3502 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3503 | &ctx->check_cq); | |
def596e9 | 3504 | } |
2b188cc1 | 3505 | } |
7c504e65 | 3506 | out: |
4329490a | 3507 | fdput(f); |
3e813c90 | 3508 | return ret; |
2b188cc1 JA |
3509 | } |
3510 | ||
3511 | static const struct file_operations io_uring_fops = { | |
3512 | .release = io_uring_release, | |
3513 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3514 | #ifndef CONFIG_MMU |
3515 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3516 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
3517 | #endif | |
2b188cc1 | 3518 | .poll = io_uring_poll, |
bebdb65e | 3519 | #ifdef CONFIG_PROC_FS |
87ce955b | 3520 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3521 | #endif |
2b188cc1 JA |
3522 | }; |
3523 | ||
92ac8bea JA |
3524 | bool io_is_uring_fops(struct file *file) |
3525 | { | |
3526 | return file->f_op == &io_uring_fops; | |
3527 | } | |
3528 | ||
c072481d PB |
3529 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3530 | struct io_uring_params *p) | |
2b188cc1 | 3531 | { |
75b28aff HV |
3532 | struct io_rings *rings; |
3533 | size_t size, sq_array_offset; | |
2b188cc1 | 3534 | |
bd740481 JA |
3535 | /* make sure these are sane, as we already accounted them */ |
3536 | ctx->sq_entries = p->sq_entries; | |
3537 | ctx->cq_entries = p->cq_entries; | |
3538 | ||
baf9cb64 | 3539 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3540 | if (size == SIZE_MAX) |
3541 | return -EOVERFLOW; | |
3542 | ||
3543 | rings = io_mem_alloc(size); | |
3544 | if (!rings) | |
2b188cc1 JA |
3545 | return -ENOMEM; |
3546 | ||
75b28aff HV |
3547 | ctx->rings = rings; |
3548 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
3549 | rings->sq_ring_mask = p->sq_entries - 1; | |
3550 | rings->cq_ring_mask = p->cq_entries - 1; | |
3551 | rings->sq_ring_entries = p->sq_entries; | |
3552 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3553 | |
ebdeb7c0 JA |
3554 | if (p->flags & IORING_SETUP_SQE128) |
3555 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3556 | else | |
3557 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
3558 | if (size == SIZE_MAX) { |
3559 | io_mem_free(ctx->rings); | |
3560 | ctx->rings = NULL; | |
2b188cc1 | 3561 | return -EOVERFLOW; |
eb065d30 | 3562 | } |
2b188cc1 JA |
3563 | |
3564 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
3565 | if (!ctx->sq_sqes) { |
3566 | io_mem_free(ctx->rings); | |
3567 | ctx->rings = NULL; | |
2b188cc1 | 3568 | return -ENOMEM; |
eb065d30 | 3569 | } |
2b188cc1 | 3570 | |
2b188cc1 JA |
3571 | return 0; |
3572 | } | |
3573 | ||
9faadcc8 PB |
3574 | static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) |
3575 | { | |
3576 | int ret, fd; | |
3577 | ||
3578 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3579 | if (fd < 0) | |
3580 | return fd; | |
3581 | ||
97c96e9f | 3582 | ret = __io_uring_add_tctx_node(ctx); |
9faadcc8 PB |
3583 | if (ret) { |
3584 | put_unused_fd(fd); | |
3585 | return ret; | |
3586 | } | |
3587 | fd_install(fd, file); | |
3588 | return fd; | |
3589 | } | |
3590 | ||
2b188cc1 JA |
3591 | /* |
3592 | * Allocate an anonymous fd, this is what constitutes the application | |
3593 | * visible backing of an io_uring instance. The application mmaps this | |
3594 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3595 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3596 | */ | |
9faadcc8 | 3597 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
3598 | { |
3599 | struct file *file; | |
9faadcc8 | 3600 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
3601 | int ret; |
3602 | ||
2b188cc1 JA |
3603 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
3604 | &ctx->ring_sock); | |
3605 | if (ret) | |
9faadcc8 | 3606 | return ERR_PTR(ret); |
2b188cc1 JA |
3607 | #endif |
3608 | ||
91a9ab7c PM |
3609 | file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, |
3610 | O_RDWR | O_CLOEXEC, NULL); | |
2b188cc1 | 3611 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
3612 | if (IS_ERR(file)) { |
3613 | sock_release(ctx->ring_sock); | |
3614 | ctx->ring_sock = NULL; | |
3615 | } else { | |
3616 | ctx->ring_sock->file = file; | |
0f212204 | 3617 | } |
2b188cc1 | 3618 | #endif |
9faadcc8 | 3619 | return file; |
2b188cc1 JA |
3620 | } |
3621 | ||
c072481d PB |
3622 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3623 | struct io_uring_params __user *params) | |
2b188cc1 | 3624 | { |
2b188cc1 | 3625 | struct io_ring_ctx *ctx; |
9faadcc8 | 3626 | struct file *file; |
2b188cc1 JA |
3627 | int ret; |
3628 | ||
8110c1a6 | 3629 | if (!entries) |
2b188cc1 | 3630 | return -EINVAL; |
8110c1a6 JA |
3631 | if (entries > IORING_MAX_ENTRIES) { |
3632 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3633 | return -EINVAL; | |
3634 | entries = IORING_MAX_ENTRIES; | |
3635 | } | |
2b188cc1 JA |
3636 | |
3637 | /* | |
3638 | * Use twice as many entries for the CQ ring. It's possible for the | |
3639 | * application to drive a higher depth than the size of the SQ ring, | |
3640 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3641 | * some flexibility in overcommitting a bit. If the application has |
3642 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3643 | * of CQ ring entries manually. | |
2b188cc1 JA |
3644 | */ |
3645 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3646 | if (p->flags & IORING_SETUP_CQSIZE) { |
3647 | /* | |
3648 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3649 | * to a power-of-two, if it isn't already. We do NOT impose | |
3650 | * any cq vs sq ring sizing. | |
3651 | */ | |
eb2667b3 | 3652 | if (!p->cq_entries) |
33a107f0 | 3653 | return -EINVAL; |
8110c1a6 JA |
3654 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3655 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3656 | return -EINVAL; | |
3657 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3658 | } | |
eb2667b3 JQ |
3659 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3660 | if (p->cq_entries < p->sq_entries) | |
3661 | return -EINVAL; | |
33a107f0 JA |
3662 | } else { |
3663 | p->cq_entries = 2 * p->sq_entries; | |
3664 | } | |
2b188cc1 | 3665 | |
2b188cc1 | 3666 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3667 | if (!ctx) |
2b188cc1 | 3668 | return -ENOMEM; |
773697b6 | 3669 | |
e6aeb272 PB |
3670 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3671 | !(ctx->flags & IORING_SETUP_IOPOLL) && | |
3672 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3673 | ctx->task_complete = true; | |
3674 | ||
bca39f39 PB |
3675 | /* |
3676 | * lazy poll_wq activation relies on ->task_complete for synchronisation | |
3677 | * purposes, see io_activate_pollwq() | |
3678 | */ | |
3679 | if (!ctx->task_complete) | |
3680 | ctx->poll_activated = true; | |
3681 | ||
773697b6 PB |
3682 | /* |
3683 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3684 | * space applications don't need to do io completion events | |
3685 | * polling again, they can rely on io_sq_thread to do polling | |
3686 | * work, which can reduce cpu usage and uring_lock contention. | |
3687 | */ | |
3688 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3689 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3690 | ctx->syscall_iopoll = 1; | |
3691 | ||
2b188cc1 | 3692 | ctx->compat = in_compat_syscall(); |
62e398be JA |
3693 | if (!capable(CAP_IPC_LOCK)) |
3694 | ctx->user = get_uid(current_user()); | |
2aede0e4 | 3695 | |
9f010507 | 3696 | /* |
e1169f06 JA |
3697 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3698 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3699 | */ |
e1169f06 JA |
3700 | ret = -EINVAL; |
3701 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3702 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 | 3703 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
c0e0d6ba DY |
3704 | IORING_SETUP_TASKRUN_FLAG | |
3705 | IORING_SETUP_DEFER_TASKRUN)) | |
e1169f06 | 3706 | goto err; |
9f010507 | 3707 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3708 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3709 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3710 | } else { | |
c0e0d6ba DY |
3711 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && |
3712 | !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) | |
ef060ea9 | 3713 | goto err; |
9f010507 | 3714 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3715 | } |
9f010507 | 3716 | |
c0e0d6ba DY |
3717 | /* |
3718 | * For DEFER_TASKRUN we require the completion task to be the same as the | |
3719 | * submission task. This implies that there is only one submitter, so enforce | |
3720 | * that. | |
3721 | */ | |
3722 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && | |
3723 | !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { | |
3724 | goto err; | |
3725 | } | |
3726 | ||
2aede0e4 JA |
3727 | /* |
3728 | * This is just grabbed for accounting purposes. When a process exits, | |
3729 | * the mm is exited and dropped before the files, hence we need to hang | |
3730 | * on to this mm purely for the purposes of being able to unaccount | |
3731 | * memory (locked/pinned vm). It's not used for anything else. | |
3732 | */ | |
6b7898eb | 3733 | mmgrab(current->mm); |
2aede0e4 | 3734 | ctx->mm_account = current->mm; |
6b7898eb | 3735 | |
2b188cc1 JA |
3736 | ret = io_allocate_scq_urings(ctx, p); |
3737 | if (ret) | |
3738 | goto err; | |
3739 | ||
7e84e1c7 | 3740 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
3741 | if (ret) |
3742 | goto err; | |
eae071c9 | 3743 | /* always set a rsrc node */ |
47b228ce PB |
3744 | ret = io_rsrc_node_switch_start(ctx); |
3745 | if (ret) | |
3746 | goto err; | |
eae071c9 | 3747 | io_rsrc_node_switch(ctx, NULL); |
2b188cc1 | 3748 | |
2b188cc1 | 3749 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
3750 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3751 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3752 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3753 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3754 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3755 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
3756 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
3757 | |
3758 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
3759 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3760 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3761 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3762 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3763 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3764 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 3765 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 3766 | |
7f13657d XW |
3767 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
3768 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 3769 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 3770 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 3771 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e JA |
3772 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
3773 | IORING_FEAT_LINKED_FILE; | |
7f13657d XW |
3774 | |
3775 | if (copy_to_user(params, p, sizeof(*p))) { | |
3776 | ret = -EFAULT; | |
3777 | goto err; | |
3778 | } | |
d1719f70 | 3779 | |
7cae596b DY |
3780 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER |
3781 | && !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
8579538c | 3782 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
7cae596b | 3783 | |
9faadcc8 PB |
3784 | file = io_uring_get_file(ctx); |
3785 | if (IS_ERR(file)) { | |
3786 | ret = PTR_ERR(file); | |
3787 | goto err; | |
3788 | } | |
3789 | ||
044c1ab3 JA |
3790 | /* |
3791 | * Install ring fd as the very last thing, so we don't risk someone | |
3792 | * having closed it before we finish setup | |
3793 | */ | |
9faadcc8 PB |
3794 | ret = io_uring_install_fd(ctx, file); |
3795 | if (ret < 0) { | |
3796 | /* fput will clean it up */ | |
3797 | fput(file); | |
3798 | return ret; | |
3799 | } | |
044c1ab3 | 3800 | |
c826bd7a | 3801 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
3802 | return ret; |
3803 | err: | |
3804 | io_ring_ctx_wait_and_kill(ctx); | |
3805 | return ret; | |
3806 | } | |
3807 | ||
3808 | /* | |
3809 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3810 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3811 | * params structure passed in. | |
3812 | */ | |
3813 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3814 | { | |
3815 | struct io_uring_params p; | |
2b188cc1 JA |
3816 | int i; |
3817 | ||
3818 | if (copy_from_user(&p, params, sizeof(p))) | |
3819 | return -EFAULT; | |
3820 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3821 | if (p.resv[i]) | |
3822 | return -EINVAL; | |
3823 | } | |
3824 | ||
6c271ce2 | 3825 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 3826 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 3827 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 3828 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 3829 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 | 3830 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
c0e0d6ba | 3831 | IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN)) |
2b188cc1 JA |
3832 | return -EINVAL; |
3833 | ||
ef060ea9 | 3834 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
3835 | } |
3836 | ||
3837 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3838 | struct io_uring_params __user *, params) | |
3839 | { | |
3840 | return io_uring_setup(entries, params); | |
3841 | } | |
3842 | ||
c072481d PB |
3843 | static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, |
3844 | unsigned nr_args) | |
66f4af93 JA |
3845 | { |
3846 | struct io_uring_probe *p; | |
3847 | size_t size; | |
3848 | int i, ret; | |
3849 | ||
3850 | size = struct_size(p, ops, nr_args); | |
3851 | if (size == SIZE_MAX) | |
3852 | return -EOVERFLOW; | |
3853 | p = kzalloc(size, GFP_KERNEL); | |
3854 | if (!p) | |
3855 | return -ENOMEM; | |
3856 | ||
3857 | ret = -EFAULT; | |
3858 | if (copy_from_user(p, arg, size)) | |
3859 | goto out; | |
3860 | ret = -EINVAL; | |
3861 | if (memchr_inv(p, 0, size)) | |
3862 | goto out; | |
3863 | ||
3864 | p->last_op = IORING_OP_LAST - 1; | |
3865 | if (nr_args > IORING_OP_LAST) | |
3866 | nr_args = IORING_OP_LAST; | |
3867 | ||
3868 | for (i = 0; i < nr_args; i++) { | |
3869 | p->ops[i].op = i; | |
a7dd2782 | 3870 | if (!io_issue_defs[i].not_supported) |
66f4af93 JA |
3871 | p->ops[i].flags = IO_URING_OP_SUPPORTED; |
3872 | } | |
3873 | p->ops_len = i; | |
3874 | ||
3875 | ret = 0; | |
3876 | if (copy_to_user(arg, p, size)) | |
3877 | ret = -EFAULT; | |
3878 | out: | |
3879 | kfree(p); | |
3880 | return ret; | |
3881 | } | |
3882 | ||
071698e1 JA |
3883 | static int io_register_personality(struct io_ring_ctx *ctx) |
3884 | { | |
4379bf8b | 3885 | const struct cred *creds; |
61cf9370 | 3886 | u32 id; |
1e6fa521 | 3887 | int ret; |
071698e1 | 3888 | |
4379bf8b | 3889 | creds = get_current_cred(); |
1e6fa521 | 3890 | |
61cf9370 MWO |
3891 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
3892 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
3893 | if (ret < 0) { |
3894 | put_cred(creds); | |
3895 | return ret; | |
3896 | } | |
3897 | return id; | |
071698e1 JA |
3898 | } |
3899 | ||
c072481d PB |
3900 | static __cold int io_register_restrictions(struct io_ring_ctx *ctx, |
3901 | void __user *arg, unsigned int nr_args) | |
21b55dbc SG |
3902 | { |
3903 | struct io_uring_restriction *res; | |
3904 | size_t size; | |
3905 | int i, ret; | |
3906 | ||
7e84e1c7 SG |
3907 | /* Restrictions allowed only if rings started disabled */ |
3908 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3909 | return -EBADFD; | |
3910 | ||
21b55dbc | 3911 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 3912 | if (ctx->restrictions.registered) |
21b55dbc SG |
3913 | return -EBUSY; |
3914 | ||
3915 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
3916 | return -EINVAL; | |
3917 | ||
3918 | size = array_size(nr_args, sizeof(*res)); | |
3919 | if (size == SIZE_MAX) | |
3920 | return -EOVERFLOW; | |
3921 | ||
3922 | res = memdup_user(arg, size); | |
3923 | if (IS_ERR(res)) | |
3924 | return PTR_ERR(res); | |
3925 | ||
3926 | ret = 0; | |
3927 | ||
3928 | for (i = 0; i < nr_args; i++) { | |
3929 | switch (res[i].opcode) { | |
3930 | case IORING_RESTRICTION_REGISTER_OP: | |
3931 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
3932 | ret = -EINVAL; | |
3933 | goto out; | |
3934 | } | |
3935 | ||
3936 | __set_bit(res[i].register_op, | |
3937 | ctx->restrictions.register_op); | |
3938 | break; | |
3939 | case IORING_RESTRICTION_SQE_OP: | |
3940 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
3941 | ret = -EINVAL; | |
3942 | goto out; | |
3943 | } | |
3944 | ||
3945 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
3946 | break; | |
3947 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
3948 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
3949 | break; | |
3950 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
3951 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
3952 | break; | |
3953 | default: | |
3954 | ret = -EINVAL; | |
3955 | goto out; | |
3956 | } | |
3957 | } | |
3958 | ||
3959 | out: | |
3960 | /* Reset all restrictions if an error happened */ | |
3961 | if (ret != 0) | |
3962 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
3963 | else | |
7e84e1c7 | 3964 | ctx->restrictions.registered = true; |
21b55dbc SG |
3965 | |
3966 | kfree(res); | |
3967 | return ret; | |
3968 | } | |
3969 | ||
7e84e1c7 SG |
3970 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
3971 | { | |
3972 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3973 | return -EBADFD; | |
3974 | ||
bca39f39 | 3975 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) { |
8579538c | 3976 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
bca39f39 PB |
3977 | /* |
3978 | * Lazy activation attempts would fail if it was polled before | |
3979 | * submitter_task is set. | |
3980 | */ | |
3981 | if (wq_has_sleeper(&ctx->poll_wq)) | |
3982 | io_activate_pollwq(ctx); | |
3983 | } | |
7cae596b | 3984 | |
7e84e1c7 SG |
3985 | if (ctx->restrictions.registered) |
3986 | ctx->restricted = 1; | |
3987 | ||
0298ef96 PB |
3988 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
3989 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
3990 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
3991 | return 0; |
3992 | } | |
3993 | ||
c072481d PB |
3994 | static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, |
3995 | void __user *arg, unsigned len) | |
fe76421d JA |
3996 | { |
3997 | struct io_uring_task *tctx = current->io_uring; | |
3998 | cpumask_var_t new_mask; | |
3999 | int ret; | |
4000 | ||
4001 | if (!tctx || !tctx->io_wq) | |
4002 | return -EINVAL; | |
4003 | ||
4004 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
4005 | return -ENOMEM; | |
4006 | ||
4007 | cpumask_clear(new_mask); | |
4008 | if (len > cpumask_size()) | |
4009 | len = cpumask_size(); | |
4010 | ||
0f5e4b83 ES |
4011 | if (in_compat_syscall()) { |
4012 | ret = compat_get_bitmap(cpumask_bits(new_mask), | |
4013 | (const compat_ulong_t __user *)arg, | |
4014 | len * 8 /* CHAR_BIT */); | |
4015 | } else { | |
4016 | ret = copy_from_user(new_mask, arg, len); | |
4017 | } | |
4018 | ||
4019 | if (ret) { | |
fe76421d JA |
4020 | free_cpumask_var(new_mask); |
4021 | return -EFAULT; | |
4022 | } | |
4023 | ||
4024 | ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); | |
4025 | free_cpumask_var(new_mask); | |
4026 | return ret; | |
4027 | } | |
4028 | ||
c072481d | 4029 | static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) |
fe76421d JA |
4030 | { |
4031 | struct io_uring_task *tctx = current->io_uring; | |
4032 | ||
4033 | if (!tctx || !tctx->io_wq) | |
4034 | return -EINVAL; | |
4035 | ||
4036 | return io_wq_cpu_affinity(tctx->io_wq, NULL); | |
4037 | } | |
4038 | ||
c072481d PB |
4039 | static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, |
4040 | void __user *arg) | |
b22fa62a | 4041 | __must_hold(&ctx->uring_lock) |
2e480058 | 4042 | { |
b22fa62a | 4043 | struct io_tctx_node *node; |
fa84693b JA |
4044 | struct io_uring_task *tctx = NULL; |
4045 | struct io_sq_data *sqd = NULL; | |
2e480058 JA |
4046 | __u32 new_count[2]; |
4047 | int i, ret; | |
4048 | ||
2e480058 JA |
4049 | if (copy_from_user(new_count, arg, sizeof(new_count))) |
4050 | return -EFAULT; | |
4051 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
4052 | if (new_count[i] > INT_MAX) | |
4053 | return -EINVAL; | |
4054 | ||
fa84693b JA |
4055 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
4056 | sqd = ctx->sq_data; | |
4057 | if (sqd) { | |
009ad9f0 JA |
4058 | /* |
4059 | * Observe the correct sqd->lock -> ctx->uring_lock | |
4060 | * ordering. Fine to drop uring_lock here, we hold | |
4061 | * a ref to the ctx. | |
4062 | */ | |
41d3a6bd | 4063 | refcount_inc(&sqd->refs); |
009ad9f0 | 4064 | mutex_unlock(&ctx->uring_lock); |
fa84693b | 4065 | mutex_lock(&sqd->lock); |
009ad9f0 | 4066 | mutex_lock(&ctx->uring_lock); |
41d3a6bd JA |
4067 | if (sqd->thread) |
4068 | tctx = sqd->thread->io_uring; | |
fa84693b JA |
4069 | } |
4070 | } else { | |
4071 | tctx = current->io_uring; | |
4072 | } | |
4073 | ||
e139a1ec | 4074 | BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); |
fa84693b | 4075 | |
bad119b9 PB |
4076 | for (i = 0; i < ARRAY_SIZE(new_count); i++) |
4077 | if (new_count[i]) | |
4078 | ctx->iowq_limits[i] = new_count[i]; | |
e139a1ec PB |
4079 | ctx->iowq_limits_set = true; |
4080 | ||
e139a1ec PB |
4081 | if (tctx && tctx->io_wq) { |
4082 | ret = io_wq_max_workers(tctx->io_wq, new_count); | |
4083 | if (ret) | |
4084 | goto err; | |
4085 | } else { | |
4086 | memset(new_count, 0, sizeof(new_count)); | |
4087 | } | |
fa84693b | 4088 | |
41d3a6bd | 4089 | if (sqd) { |
fa84693b | 4090 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
4091 | io_put_sq_data(sqd); |
4092 | } | |
2e480058 JA |
4093 | |
4094 | if (copy_to_user(arg, new_count, sizeof(new_count))) | |
4095 | return -EFAULT; | |
4096 | ||
b22fa62a PB |
4097 | /* that's it for SQPOLL, only the SQPOLL task creates requests */ |
4098 | if (sqd) | |
4099 | return 0; | |
4100 | ||
4101 | /* now propagate the restriction to all registered users */ | |
4102 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
4103 | struct io_uring_task *tctx = node->task->io_uring; | |
4104 | ||
4105 | if (WARN_ON_ONCE(!tctx->io_wq)) | |
4106 | continue; | |
4107 | ||
4108 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
4109 | new_count[i] = ctx->iowq_limits[i]; | |
4110 | /* ignore errors, it always returns zero anyway */ | |
4111 | (void)io_wq_max_workers(tctx->io_wq, new_count); | |
4112 | } | |
2e480058 | 4113 | return 0; |
fa84693b | 4114 | err: |
41d3a6bd | 4115 | if (sqd) { |
fa84693b | 4116 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
4117 | io_put_sq_data(sqd); |
4118 | } | |
fa84693b | 4119 | return ret; |
2e480058 JA |
4120 | } |
4121 | ||
edafccee JA |
4122 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
4123 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
4124 | __releases(ctx->uring_lock) |
4125 | __acquires(ctx->uring_lock) | |
edafccee JA |
4126 | { |
4127 | int ret; | |
4128 | ||
35fa71a0 | 4129 | /* |
fbb8bb02 PB |
4130 | * We don't quiesce the refs for register anymore and so it can't be |
4131 | * dying as we're holding a file ref here. | |
35fa71a0 | 4132 | */ |
fbb8bb02 | 4133 | if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs))) |
35fa71a0 JA |
4134 | return -ENXIO; |
4135 | ||
d7cce96c PB |
4136 | if (ctx->submitter_task && ctx->submitter_task != current) |
4137 | return -EEXIST; | |
4138 | ||
75c4021a | 4139 | if (ctx->restricted) { |
75c4021a PB |
4140 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); |
4141 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
4142 | return -EACCES; | |
4143 | } | |
4144 | ||
edafccee JA |
4145 | switch (opcode) { |
4146 | case IORING_REGISTER_BUFFERS: | |
0184f08e PB |
4147 | ret = -EFAULT; |
4148 | if (!arg) | |
4149 | break; | |
634d00df | 4150 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
4151 | break; |
4152 | case IORING_UNREGISTER_BUFFERS: | |
4153 | ret = -EINVAL; | |
4154 | if (arg || nr_args) | |
4155 | break; | |
0a96bbe4 | 4156 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 4157 | break; |
6b06314c | 4158 | case IORING_REGISTER_FILES: |
a8da73a3 JA |
4159 | ret = -EFAULT; |
4160 | if (!arg) | |
4161 | break; | |
792e3582 | 4162 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
4163 | break; |
4164 | case IORING_UNREGISTER_FILES: | |
4165 | ret = -EINVAL; | |
4166 | if (arg || nr_args) | |
4167 | break; | |
4168 | ret = io_sqe_files_unregister(ctx); | |
4169 | break; | |
c3a31e60 | 4170 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 4171 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 4172 | break; |
9b402849 JA |
4173 | case IORING_REGISTER_EVENTFD: |
4174 | ret = -EINVAL; | |
4175 | if (nr_args != 1) | |
4176 | break; | |
c75312dd UA |
4177 | ret = io_eventfd_register(ctx, arg, 0); |
4178 | break; | |
4179 | case IORING_REGISTER_EVENTFD_ASYNC: | |
4180 | ret = -EINVAL; | |
4181 | if (nr_args != 1) | |
f2842ab5 | 4182 | break; |
c75312dd | 4183 | ret = io_eventfd_register(ctx, arg, 1); |
9b402849 JA |
4184 | break; |
4185 | case IORING_UNREGISTER_EVENTFD: | |
4186 | ret = -EINVAL; | |
4187 | if (arg || nr_args) | |
4188 | break; | |
4189 | ret = io_eventfd_unregister(ctx); | |
4190 | break; | |
66f4af93 JA |
4191 | case IORING_REGISTER_PROBE: |
4192 | ret = -EINVAL; | |
4193 | if (!arg || nr_args > 256) | |
4194 | break; | |
4195 | ret = io_probe(ctx, arg, nr_args); | |
4196 | break; | |
071698e1 JA |
4197 | case IORING_REGISTER_PERSONALITY: |
4198 | ret = -EINVAL; | |
4199 | if (arg || nr_args) | |
4200 | break; | |
4201 | ret = io_register_personality(ctx); | |
4202 | break; | |
4203 | case IORING_UNREGISTER_PERSONALITY: | |
4204 | ret = -EINVAL; | |
4205 | if (arg) | |
4206 | break; | |
4207 | ret = io_unregister_personality(ctx, nr_args); | |
4208 | break; | |
7e84e1c7 SG |
4209 | case IORING_REGISTER_ENABLE_RINGS: |
4210 | ret = -EINVAL; | |
4211 | if (arg || nr_args) | |
4212 | break; | |
4213 | ret = io_register_enable_rings(ctx); | |
4214 | break; | |
21b55dbc SG |
4215 | case IORING_REGISTER_RESTRICTIONS: |
4216 | ret = io_register_restrictions(ctx, arg, nr_args); | |
4217 | break; | |
992da01a PB |
4218 | case IORING_REGISTER_FILES2: |
4219 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
4220 | break; | |
4221 | case IORING_REGISTER_FILES_UPDATE2: | |
4222 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
4223 | IORING_RSRC_FILE); | |
4224 | break; | |
4225 | case IORING_REGISTER_BUFFERS2: | |
4226 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 4227 | break; |
992da01a PB |
4228 | case IORING_REGISTER_BUFFERS_UPDATE: |
4229 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
4230 | IORING_RSRC_BUFFER); | |
c3bdad02 | 4231 | break; |
fe76421d JA |
4232 | case IORING_REGISTER_IOWQ_AFF: |
4233 | ret = -EINVAL; | |
4234 | if (!arg || !nr_args) | |
4235 | break; | |
4236 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
4237 | break; | |
4238 | case IORING_UNREGISTER_IOWQ_AFF: | |
4239 | ret = -EINVAL; | |
4240 | if (arg || nr_args) | |
4241 | break; | |
4242 | ret = io_unregister_iowq_aff(ctx); | |
4243 | break; | |
2e480058 JA |
4244 | case IORING_REGISTER_IOWQ_MAX_WORKERS: |
4245 | ret = -EINVAL; | |
4246 | if (!arg || nr_args != 2) | |
4247 | break; | |
4248 | ret = io_register_iowq_max_workers(ctx, arg); | |
4249 | break; | |
e7a6c00d JA |
4250 | case IORING_REGISTER_RING_FDS: |
4251 | ret = io_ringfd_register(ctx, arg, nr_args); | |
4252 | break; | |
4253 | case IORING_UNREGISTER_RING_FDS: | |
4254 | ret = io_ringfd_unregister(ctx, arg, nr_args); | |
4255 | break; | |
c7fb1942 JA |
4256 | case IORING_REGISTER_PBUF_RING: |
4257 | ret = -EINVAL; | |
4258 | if (!arg || nr_args != 1) | |
4259 | break; | |
4260 | ret = io_register_pbuf_ring(ctx, arg); | |
4261 | break; | |
4262 | case IORING_UNREGISTER_PBUF_RING: | |
4263 | ret = -EINVAL; | |
4264 | if (!arg || nr_args != 1) | |
4265 | break; | |
4266 | ret = io_unregister_pbuf_ring(ctx, arg); | |
4267 | break; | |
78a861b9 JA |
4268 | case IORING_REGISTER_SYNC_CANCEL: |
4269 | ret = -EINVAL; | |
4270 | if (!arg || nr_args != 1) | |
4271 | break; | |
4272 | ret = io_sync_cancel(ctx, arg); | |
4273 | break; | |
6e73dffb PB |
4274 | case IORING_REGISTER_FILE_ALLOC_RANGE: |
4275 | ret = -EINVAL; | |
4276 | if (!arg || nr_args) | |
4277 | break; | |
4278 | ret = io_register_file_alloc_range(ctx, arg); | |
4279 | break; | |
edafccee JA |
4280 | default: |
4281 | ret = -EINVAL; | |
4282 | break; | |
4283 | } | |
4284 | ||
edafccee JA |
4285 | return ret; |
4286 | } | |
4287 | ||
4288 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
4289 | void __user *, arg, unsigned int, nr_args) | |
4290 | { | |
4291 | struct io_ring_ctx *ctx; | |
4292 | long ret = -EBADF; | |
4293 | struct fd f; | |
4294 | ||
34319084 JA |
4295 | if (opcode >= IORING_REGISTER_LAST) |
4296 | return -EINVAL; | |
4297 | ||
edafccee JA |
4298 | f = fdget(fd); |
4299 | if (!f.file) | |
4300 | return -EBADF; | |
4301 | ||
4302 | ret = -EOPNOTSUPP; | |
e5550a14 | 4303 | if (!io_is_uring_fops(f.file)) |
edafccee JA |
4304 | goto out_fput; |
4305 | ||
4306 | ctx = f.file->private_data; | |
4307 | ||
4308 | mutex_lock(&ctx->uring_lock); | |
4309 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
4310 | mutex_unlock(&ctx->uring_lock); | |
2757be22 | 4311 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); |
edafccee JA |
4312 | out_fput: |
4313 | fdput(f); | |
4314 | return ret; | |
4315 | } | |
4316 | ||
2b188cc1 JA |
4317 | static int __init io_uring_init(void) |
4318 | { | |
9c71d39a | 4319 | #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ |
d7f62e82 | 4320 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ |
9c71d39a | 4321 | BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ |
d7f62e82 SM |
4322 | } while (0) |
4323 | ||
4324 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9c71d39a SM |
4325 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) |
4326 | #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ | |
4327 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) | |
d7f62e82 SM |
4328 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); |
4329 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
4330 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
4331 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
4332 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
4333 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
4334 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9c71d39a SM |
4335 | BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); |
4336 | BUILD_BUG_SQE_ELEM(12, __u32, __pad1); | |
d7f62e82 | 4337 | BUILD_BUG_SQE_ELEM(16, __u64, addr); |
7d67af2c | 4338 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
4339 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
4340 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
4341 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
4342 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
4343 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
4344 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
4345 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
4346 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
4347 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
4348 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
4349 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
4350 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
4351 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
4352 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
4353 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 4354 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
9c71d39a SM |
4355 | BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); |
4356 | BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); | |
4357 | BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); | |
4358 | BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); | |
4359 | BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); | |
d7f62e82 SM |
4360 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
4361 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 4362 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 4363 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 4364 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 4365 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
b48c312b PB |
4366 | BUILD_BUG_SQE_ELEM(44, __u16, addr_len); |
4367 | BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); | |
e9621e2b | 4368 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
9c71d39a SM |
4369 | BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); |
4370 | BUILD_BUG_SQE_ELEM(56, __u64, __pad2); | |
d7f62e82 | 4371 | |
b0d658ec PB |
4372 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
4373 | sizeof(struct io_uring_rsrc_update)); | |
4374 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
4375 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
4376 | |
4377 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
4378 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
4379 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
4380 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 4381 | |
b0d658ec PB |
4382 | /* should fit into one byte */ |
4383 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
4384 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
4385 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 4386 | |
32c2d33e | 4387 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); |
16340eab | 4388 | |
3a4b89a2 JA |
4389 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
4390 | ||
d9b57aa3 | 4391 | io_uring_optable_init(); |
0702e536 | 4392 | |
91f245d5 JA |
4393 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | |
4394 | SLAB_ACCOUNT); | |
2b188cc1 JA |
4395 | return 0; |
4396 | }; | |
4397 | __initcall(io_uring_init); |