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); |
de0617e4 | 152 | |
2b188cc1 JA |
153 | static struct kmem_cache *req_cachep; |
154 | ||
2b188cc1 JA |
155 | struct sock *io_uring_get_socket(struct file *file) |
156 | { | |
157 | #if defined(CONFIG_UNIX) | |
cd40cae2 | 158 | if (io_is_uring_fops(file)) { |
2b188cc1 JA |
159 | struct io_ring_ctx *ctx = file->private_data; |
160 | ||
161 | return ctx->ring_sock->sk; | |
162 | } | |
163 | #endif | |
164 | return NULL; | |
165 | } | |
166 | EXPORT_SYMBOL(io_uring_get_socket); | |
167 | ||
c450178d PB |
168 | static inline void io_submit_flush_completions(struct io_ring_ctx *ctx) |
169 | { | |
6f33b0bc | 170 | if (!wq_list_empty(&ctx->submit_state.compl_reqs)) |
c450178d PB |
171 | __io_submit_flush_completions(ctx); |
172 | } | |
173 | ||
faf88dde PB |
174 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
175 | { | |
176 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
177 | } | |
178 | ||
0fc8c2ac DY |
179 | static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx) |
180 | { | |
181 | return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head); | |
182 | } | |
183 | ||
9cae36a0 JA |
184 | static bool io_match_linked(struct io_kiocb *head) |
185 | { | |
186 | struct io_kiocb *req; | |
187 | ||
188 | io_for_each_link(req, head) { | |
189 | if (req->flags & REQ_F_INFLIGHT) | |
190 | return true; | |
191 | } | |
192 | return false; | |
6af3f48b PB |
193 | } |
194 | ||
195 | /* | |
196 | * As io_match_task() but protected against racing with linked timeouts. | |
197 | * User must not hold timeout_lock. | |
198 | */ | |
329061d3 JA |
199 | bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, |
200 | bool cancel_all) | |
6af3f48b | 201 | { |
9cae36a0 JA |
202 | bool matched; |
203 | ||
6af3f48b PB |
204 | if (task && head->task != task) |
205 | return false; | |
9cae36a0 JA |
206 | if (cancel_all) |
207 | return true; | |
208 | ||
209 | if (head->flags & REQ_F_LINK_TIMEOUT) { | |
210 | struct io_ring_ctx *ctx = head->ctx; | |
211 | ||
212 | /* protect against races with linked timeouts */ | |
213 | spin_lock_irq(&ctx->timeout_lock); | |
214 | matched = io_match_linked(head); | |
215 | spin_unlock_irq(&ctx->timeout_lock); | |
216 | } else { | |
217 | matched = io_match_linked(head); | |
218 | } | |
219 | return matched; | |
6af3f48b PB |
220 | } |
221 | ||
a8295b98 HX |
222 | static inline void req_fail_link_node(struct io_kiocb *req, int res) |
223 | { | |
224 | req_set_fail(req); | |
97b388d7 | 225 | io_req_set_res(req, res, 0); |
a8295b98 HX |
226 | } |
227 | ||
fa05457a PB |
228 | static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) |
229 | { | |
230 | wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); | |
a8295b98 HX |
231 | } |
232 | ||
c072481d | 233 | static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) |
2b188cc1 JA |
234 | { |
235 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
236 | ||
0f158b4c | 237 | complete(&ctx->ref_comp); |
2b188cc1 JA |
238 | } |
239 | ||
c072481d | 240 | static __cold void io_fallback_req_func(struct work_struct *work) |
f56165e6 PB |
241 | { |
242 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
243 | fallback_work.work); | |
244 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
245 | struct io_kiocb *req, *tmp; | |
f237c30a | 246 | bool locked = false; |
f56165e6 PB |
247 | |
248 | percpu_ref_get(&ctx->refs); | |
3218e5d3 | 249 | llist_for_each_entry_safe(req, tmp, node, io_task_work.node) |
f237c30a | 250 | req->io_task_work.func(req, &locked); |
5636c00d | 251 | |
f237c30a | 252 | if (locked) { |
c450178d | 253 | io_submit_flush_completions(ctx); |
f237c30a PB |
254 | mutex_unlock(&ctx->uring_lock); |
255 | } | |
f56165e6 PB |
256 | percpu_ref_put(&ctx->refs); |
257 | } | |
258 | ||
e6f89be6 PB |
259 | static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) |
260 | { | |
261 | unsigned hash_buckets = 1U << bits; | |
262 | size_t hash_size = hash_buckets * sizeof(table->hbs[0]); | |
263 | ||
264 | table->hbs = kmalloc(hash_size, GFP_KERNEL); | |
265 | if (!table->hbs) | |
266 | return -ENOMEM; | |
267 | ||
268 | table->hash_bits = bits; | |
269 | init_hash_table(table, hash_buckets); | |
270 | return 0; | |
271 | } | |
272 | ||
c072481d | 273 | static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
2b188cc1 JA |
274 | { |
275 | struct io_ring_ctx *ctx; | |
9cfc7e94 | 276 | int hash_bits; |
2b188cc1 JA |
277 | |
278 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
279 | if (!ctx) | |
280 | return NULL; | |
281 | ||
9cfc7e94 JA |
282 | xa_init(&ctx->io_bl_xa); |
283 | ||
78076bb6 JA |
284 | /* |
285 | * Use 5 bits less than the max cq entries, that should give us around | |
4a07723f PB |
286 | * 32 entries per hash list if totally full and uniformly spread, but |
287 | * don't keep too many buckets to not overconsume memory. | |
78076bb6 | 288 | */ |
4a07723f PB |
289 | hash_bits = ilog2(p->cq_entries) - 5; |
290 | hash_bits = clamp(hash_bits, 1, 8); | |
e6f89be6 | 291 | if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) |
78076bb6 | 292 | goto err; |
9ca9fb24 PB |
293 | if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) |
294 | goto err; | |
38513c46 | 295 | |
6224843d PB |
296 | ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL); |
297 | if (!ctx->dummy_ubuf) | |
298 | goto err; | |
299 | /* set invalid range, so io_import_fixed() fails meeting it */ | |
300 | ctx->dummy_ubuf->ubuf = -1UL; | |
301 | ||
21482896 | 302 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
48904229 | 303 | 0, GFP_KERNEL)) |
206aefde | 304 | goto err; |
2b188cc1 JA |
305 | |
306 | ctx->flags = p->flags; | |
90554200 | 307 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 308 | INIT_LIST_HEAD(&ctx->sqd_list); |
1d7bb1d5 | 309 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
cc3cec83 | 310 | INIT_LIST_HEAD(&ctx->io_buffers_cache); |
9b797a37 | 311 | io_alloc_cache_init(&ctx->apoll_cache); |
43e0bbbd | 312 | io_alloc_cache_init(&ctx->netmsg_cache); |
0f158b4c | 313 | init_completion(&ctx->ref_comp); |
61cf9370 | 314 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 315 | mutex_init(&ctx->uring_lock); |
311997b3 | 316 | init_waitqueue_head(&ctx->cq_wait); |
2b188cc1 | 317 | spin_lock_init(&ctx->completion_lock); |
89850fce | 318 | spin_lock_init(&ctx->timeout_lock); |
5eef4e87 | 319 | INIT_WQ_LIST(&ctx->iopoll_list); |
cc3cec83 JA |
320 | INIT_LIST_HEAD(&ctx->io_buffers_pages); |
321 | INIT_LIST_HEAD(&ctx->io_buffers_comp); | |
de0617e4 | 322 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 323 | INIT_LIST_HEAD(&ctx->timeout_list); |
ef9dd637 | 324 | INIT_LIST_HEAD(&ctx->ltimeout_list); |
d67d2263 BM |
325 | spin_lock_init(&ctx->rsrc_ref_lock); |
326 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); | |
269bbe5f BM |
327 | INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work); |
328 | init_llist_head(&ctx->rsrc_put_llist); | |
c0e0d6ba | 329 | init_llist_head(&ctx->work_llist); |
13bf43f5 | 330 | INIT_LIST_HEAD(&ctx->tctx_list); |
c2b6c6bc PB |
331 | ctx->submit_state.free_list.next = NULL; |
332 | INIT_WQ_LIST(&ctx->locked_free_list); | |
9011bf9a | 333 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
6f33b0bc | 334 | INIT_WQ_LIST(&ctx->submit_state.compl_reqs); |
2b188cc1 | 335 | return ctx; |
206aefde | 336 | err: |
6224843d | 337 | kfree(ctx->dummy_ubuf); |
e6f89be6 | 338 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 339 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 JA |
340 | kfree(ctx->io_bl); |
341 | xa_destroy(&ctx->io_bl_xa); | |
206aefde JA |
342 | kfree(ctx); |
343 | return NULL; | |
2b188cc1 JA |
344 | } |
345 | ||
8f6ed49a PB |
346 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
347 | { | |
348 | struct io_rings *r = ctx->rings; | |
349 | ||
350 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
351 | ctx->cq_extra--; | |
352 | } | |
353 | ||
9cf7c104 | 354 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 355 | { |
2bc9930e JA |
356 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
357 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 358 | |
8f6ed49a | 359 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 360 | } |
de0617e4 | 361 | |
9d858b21 | 362 | return false; |
de0617e4 JA |
363 | } |
364 | ||
9cae36a0 JA |
365 | static inline void io_req_track_inflight(struct io_kiocb *req) |
366 | { | |
367 | if (!(req->flags & REQ_F_INFLIGHT)) { | |
368 | req->flags |= REQ_F_INFLIGHT; | |
386e4fb6 | 369 | atomic_inc(&req->task->io_uring->inflight_tracked); |
9cae36a0 JA |
370 | } |
371 | } | |
372 | ||
fd08e530 PB |
373 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
374 | { | |
906c6caa PB |
375 | if (WARN_ON_ONCE(!req->link)) |
376 | return NULL; | |
377 | ||
4d13d1a4 PB |
378 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
379 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
380 | |
381 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 382 | io_req_set_refcount(req); |
4d13d1a4 PB |
383 | __io_req_set_refcount(req->link, 2); |
384 | return req->link; | |
fd08e530 PB |
385 | } |
386 | ||
387 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
388 | { | |
4d13d1a4 | 389 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
390 | return NULL; |
391 | return __io_prep_linked_timeout(req); | |
392 | } | |
393 | ||
cb2d344c PB |
394 | static noinline void __io_arm_ltimeout(struct io_kiocb *req) |
395 | { | |
396 | io_queue_linked_timeout(__io_prep_linked_timeout(req)); | |
397 | } | |
398 | ||
399 | static inline void io_arm_ltimeout(struct io_kiocb *req) | |
400 | { | |
401 | if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) | |
402 | __io_arm_ltimeout(req); | |
403 | } | |
404 | ||
1e6fa521 JA |
405 | static void io_prep_async_work(struct io_kiocb *req) |
406 | { | |
407 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
1e6fa521 JA |
408 | struct io_ring_ctx *ctx = req->ctx; |
409 | ||
b8e64b53 PB |
410 | if (!(req->flags & REQ_F_CREDS)) { |
411 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 412 | req->creds = get_current_cred(); |
b8e64b53 | 413 | } |
003e8dcc | 414 | |
e1d675df PB |
415 | req->work.list.next = NULL; |
416 | req->work.flags = 0; | |
8e29da69 | 417 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); |
feaadc4f PB |
418 | if (req->flags & REQ_F_FORCE_ASYNC) |
419 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
420 | ||
f6b543fd JA |
421 | if (req->file && !io_req_ffs_set(req)) |
422 | req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT; | |
423 | ||
1e6fa521 JA |
424 | if (req->flags & REQ_F_ISREG) { |
425 | if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL)) | |
426 | io_wq_hash_work(&req->work, file_inode(req->file)); | |
4b982bd0 | 427 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
428 | if (def->unbound_nonreg_file) |
429 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
430 | } | |
561fb04a | 431 | } |
cccf0ee8 | 432 | |
cbdcb435 | 433 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 434 | { |
cbdcb435 | 435 | struct io_kiocb *cur; |
54a91f3b | 436 | |
44eff40a PB |
437 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
438 | struct io_ring_ctx *ctx = req->ctx; | |
439 | ||
674ee8e1 | 440 | spin_lock_irq(&ctx->timeout_lock); |
44eff40a PB |
441 | io_for_each_link(cur, req) |
442 | io_prep_async_work(cur); | |
674ee8e1 | 443 | spin_unlock_irq(&ctx->timeout_lock); |
44eff40a PB |
444 | } else { |
445 | io_for_each_link(cur, req) | |
446 | io_prep_async_work(cur); | |
447 | } | |
561fb04a JA |
448 | } |
449 | ||
f3b44f92 | 450 | void io_queue_iowq(struct io_kiocb *req, bool *dont_use) |
561fb04a | 451 | { |
cbdcb435 | 452 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 453 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 454 | |
3bfe6106 JA |
455 | BUG_ON(!tctx); |
456 | BUG_ON(!tctx->io_wq); | |
561fb04a | 457 | |
cbdcb435 PB |
458 | /* init ->work of the whole link before punting */ |
459 | io_prep_async_link(req); | |
991468dc JA |
460 | |
461 | /* | |
462 | * Not expected to happen, but if we do have a bug where this _can_ | |
463 | * happen, catch it here and ensure the request is marked as | |
464 | * canceled. That will make io-wq go through the usual work cancel | |
465 | * procedure rather than attempt to run this request (or create a new | |
466 | * worker for it). | |
467 | */ | |
468 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
469 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
470 | ||
48863ffd | 471 | trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); |
ebf93667 | 472 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
473 | if (link) |
474 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
475 | } |
476 | ||
c072481d | 477 | static __cold void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 478 | { |
441b8a78 | 479 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
480 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
481 | struct io_defer_entry, list); | |
de0617e4 | 482 | |
9cf7c104 | 483 | if (req_need_defer(de->req, de->seq)) |
04518945 | 484 | break; |
27dc8338 | 485 | list_del_init(&de->list); |
907d1df3 | 486 | io_req_task_queue(de->req); |
27dc8338 | 487 | kfree(de); |
441b8a78 | 488 | } |
04518945 PB |
489 | } |
490 | ||
21a091b9 DY |
491 | |
492 | static void io_eventfd_ops(struct rcu_head *rcu) | |
f2842ab5 | 493 | { |
d8e9214f | 494 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
21a091b9 | 495 | int ops = atomic_xchg(&ev_fd->ops, 0); |
305bef98 | 496 | |
21a091b9 DY |
497 | if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) |
498 | eventfd_signal(ev_fd->cq_ev_fd, 1); | |
d8e9214f | 499 | |
21a091b9 DY |
500 | /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback |
501 | * ordering in a race but if references are 0 we know we have to free | |
502 | * it regardless. | |
305bef98 | 503 | */ |
21a091b9 DY |
504 | if (atomic_dec_and_test(&ev_fd->refs)) { |
505 | eventfd_ctx_put(ev_fd->cq_ev_fd); | |
506 | kfree(ev_fd); | |
507 | } | |
d8e9214f DY |
508 | } |
509 | ||
77bc59b4 | 510 | static void io_eventfd_signal(struct io_ring_ctx *ctx) |
f2842ab5 | 511 | { |
21a091b9 | 512 | struct io_ev_fd *ev_fd = NULL; |
77bc59b4 | 513 | |
77bc59b4 UA |
514 | rcu_read_lock(); |
515 | /* | |
516 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
517 | * and eventfd_signal | |
518 | */ | |
519 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
520 | ||
521 | /* | |
522 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
523 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
524 | * the function and rcu_read_lock. | |
525 | */ | |
526 | if (unlikely(!ev_fd)) | |
527 | goto out; | |
7e55a19c | 528 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 | 529 | goto out; |
21a091b9 DY |
530 | if (ev_fd->eventfd_async && !io_wq_current_is_worker()) |
531 | goto out; | |
77bc59b4 | 532 | |
21a091b9 | 533 | if (likely(eventfd_signal_allowed())) { |
77bc59b4 | 534 | eventfd_signal(ev_fd->cq_ev_fd, 1); |
21a091b9 DY |
535 | } else { |
536 | atomic_inc(&ev_fd->refs); | |
537 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) | |
538 | call_rcu(&ev_fd->rcu, io_eventfd_ops); | |
539 | else | |
540 | atomic_dec(&ev_fd->refs); | |
541 | } | |
542 | ||
77bc59b4 UA |
543 | out: |
544 | rcu_read_unlock(); | |
f2842ab5 JA |
545 | } |
546 | ||
21a091b9 DY |
547 | static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) |
548 | { | |
549 | bool skip; | |
550 | ||
551 | spin_lock(&ctx->completion_lock); | |
552 | ||
553 | /* | |
554 | * Eventfd should only get triggered when at least one event has been | |
555 | * posted. Some applications rely on the eventfd notification count | |
556 | * only changing IFF a new CQE has been added to the CQ ring. There's | |
557 | * no depedency on 1:1 relationship between how many times this | |
558 | * function is called (and hence the eventfd count) and number of CQEs | |
559 | * posted to the CQ ring. | |
560 | */ | |
561 | skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
562 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
563 | spin_unlock(&ctx->completion_lock); | |
564 | if (skip) | |
565 | return; | |
566 | ||
567 | io_eventfd_signal(ctx); | |
568 | } | |
569 | ||
a830ffd2 PB |
570 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
571 | { | |
572 | if (ctx->off_timeout_used || ctx->drain_active) { | |
573 | spin_lock(&ctx->completion_lock); | |
574 | if (ctx->off_timeout_used) | |
575 | io_flush_timeouts(ctx); | |
576 | if (ctx->drain_active) | |
577 | io_queue_deferred(ctx); | |
578 | spin_unlock(&ctx->completion_lock); | |
579 | } | |
580 | if (ctx->has_evfd) | |
21a091b9 | 581 | io_eventfd_flush_signal(ctx); |
a830ffd2 PB |
582 | } |
583 | ||
25399321 | 584 | static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 | 585 | { |
46929b08 | 586 | io_commit_cqring_flush(ctx); |
9aa8dfde | 587 | io_cqring_wake(ctx); |
1d7bb1d5 JA |
588 | } |
589 | ||
25399321 PB |
590 | static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) |
591 | __releases(ctx->completion_lock) | |
592 | { | |
593 | io_commit_cqring(ctx); | |
594 | spin_unlock(&ctx->completion_lock); | |
595 | io_cqring_ev_posted(ctx); | |
596 | } | |
597 | ||
598 | void io_cq_unlock_post(struct io_ring_ctx *ctx) | |
599 | { | |
600 | __io_cq_unlock_post(ctx); | |
601 | } | |
602 | ||
c4a2ed72 | 603 | /* Returns true if there are no backlogged entries after the flush */ |
6c2450ae | 604 | static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) |
1d7bb1d5 | 605 | { |
305bef98 | 606 | bool all_flushed; |
e45a3e05 | 607 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 608 | |
a566c556 | 609 | if (!force && __io_cqring_events(ctx) == ctx->cq_entries) |
e23de15f | 610 | return false; |
1d7bb1d5 | 611 | |
e45a3e05 SR |
612 | if (ctx->flags & IORING_SETUP_CQE32) |
613 | cqe_size <<= 1; | |
614 | ||
25399321 | 615 | io_cq_lock(ctx); |
6c2450ae | 616 | while (!list_empty(&ctx->cq_overflow_list)) { |
aa1df3a3 | 617 | struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true); |
6c2450ae | 618 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 619 | |
1d7bb1d5 JA |
620 | if (!cqe && !force) |
621 | break; | |
6c2450ae PB |
622 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
623 | struct io_overflow_cqe, list); | |
624 | if (cqe) | |
e45a3e05 | 625 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae | 626 | else |
8f6ed49a PB |
627 | io_account_cq_overflow(ctx); |
628 | ||
6c2450ae PB |
629 | list_del(&ocqe->list); |
630 | kfree(ocqe); | |
1d7bb1d5 JA |
631 | } |
632 | ||
09e88404 PB |
633 | all_flushed = list_empty(&ctx->cq_overflow_list); |
634 | if (all_flushed) { | |
10988a0a | 635 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 636 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 637 | } |
46930143 | 638 | |
25399321 | 639 | io_cq_unlock_post(ctx); |
09e88404 | 640 | return all_flushed; |
1d7bb1d5 JA |
641 | } |
642 | ||
90f67366 | 643 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 644 | { |
ca0a2651 JA |
645 | bool ret = true; |
646 | ||
10988a0a | 647 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
6c503150 PB |
648 | /* iopoll syncs against uring_lock, not completion_lock */ |
649 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
650 | mutex_lock(&ctx->uring_lock); | |
90f67366 | 651 | ret = __io_cqring_overflow_flush(ctx, false); |
6c503150 PB |
652 | if (ctx->flags & IORING_SETUP_IOPOLL) |
653 | mutex_unlock(&ctx->uring_lock); | |
654 | } | |
ca0a2651 JA |
655 | |
656 | return ret; | |
6c503150 PB |
657 | } |
658 | ||
e70cb608 | 659 | void __io_put_task(struct task_struct *task, int nr) |
6a290a14 PB |
660 | { |
661 | struct io_uring_task *tctx = task->io_uring; | |
662 | ||
9d170164 PB |
663 | percpu_counter_sub(&tctx->inflight, nr); |
664 | if (unlikely(atomic_read(&tctx->in_idle))) | |
665 | wake_up(&tctx->wait); | |
666 | put_task_struct_many(task, nr); | |
667 | } | |
668 | ||
63809137 | 669 | void io_task_refs_refill(struct io_uring_task *tctx) |
9a10867a PB |
670 | { |
671 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
672 | ||
673 | percpu_counter_add(&tctx->inflight, refill); | |
674 | refcount_add(refill, ¤t->usage); | |
675 | tctx->cached_refs += refill; | |
676 | } | |
677 | ||
3cc7fdb9 PB |
678 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
679 | { | |
680 | struct io_uring_task *tctx = task->io_uring; | |
681 | unsigned int refs = tctx->cached_refs; | |
682 | ||
683 | if (refs) { | |
684 | tctx->cached_refs = 0; | |
685 | percpu_counter_sub(&tctx->inflight, refs); | |
686 | put_task_struct_many(task, refs); | |
687 | } | |
688 | } | |
689 | ||
68494a65 PB |
690 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
691 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 692 | { |
cce4b8b0 | 693 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
694 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
695 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 696 | |
e45a3e05 SR |
697 | if (is_cqe32) |
698 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 699 | |
e45a3e05 | 700 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 701 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
702 | if (!ocqe) { |
703 | /* | |
704 | * If we're in ring overflow flush mode, or in task cancel mode, | |
705 | * or cannot allocate an overflow entry, then we need to drop it | |
706 | * on the floor. | |
707 | */ | |
8f6ed49a | 708 | io_account_cq_overflow(ctx); |
155bc950 | 709 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 710 | return false; |
2b188cc1 | 711 | } |
cce4b8b0 | 712 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 713 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 714 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 715 | |
cce4b8b0 | 716 | } |
d4d19c19 | 717 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
718 | ocqe->cqe.res = res; |
719 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
720 | if (is_cqe32) { |
721 | ocqe->cqe.big_cqe[0] = extra1; | |
722 | ocqe->cqe.big_cqe[1] = extra2; | |
723 | } | |
cce4b8b0 PB |
724 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
725 | return true; | |
2b188cc1 JA |
726 | } |
727 | ||
68494a65 PB |
728 | bool io_req_cqe_overflow(struct io_kiocb *req) |
729 | { | |
730 | if (!(req->flags & REQ_F_CQE32_INIT)) { | |
731 | req->extra1 = 0; | |
732 | req->extra2 = 0; | |
733 | } | |
734 | return io_cqring_event_overflow(req->ctx, req->cqe.user_data, | |
735 | req->cqe.res, req->cqe.flags, | |
736 | req->extra1, req->extra2); | |
737 | } | |
738 | ||
faf88dde PB |
739 | /* |
740 | * writes to the cq entry need to come after reading head; the | |
741 | * control dependency is enough as we're using WRITE_ONCE to | |
742 | * fill the cq entry | |
743 | */ | |
aa1df3a3 | 744 | struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow) |
faf88dde PB |
745 | { |
746 | struct io_rings *rings = ctx->rings; | |
747 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
748 | unsigned int free, queued, len; |
749 | ||
aa1df3a3 PB |
750 | /* |
751 | * Posting into the CQ when there are pending overflowed CQEs may break | |
752 | * ordering guarantees, which will affect links, F_MORE users and more. | |
753 | * Force overflow the completion. | |
754 | */ | |
755 | if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) | |
756 | return NULL; | |
faf88dde PB |
757 | |
758 | /* userspace may cheat modifying the tail, be safe and do min */ | |
759 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
760 | free = ctx->cq_entries - queued; | |
761 | /* we need a contiguous range, limit based on the current array offset */ | |
762 | len = min(free, ctx->cq_entries - off); | |
763 | if (!len) | |
764 | return NULL; | |
765 | ||
b3659a65 PB |
766 | if (ctx->flags & IORING_SETUP_CQE32) { |
767 | off <<= 1; | |
768 | len <<= 1; | |
769 | } | |
770 | ||
faf88dde PB |
771 | ctx->cqe_cached = &rings->cqes[off]; |
772 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
b3659a65 PB |
773 | |
774 | ctx->cached_cq_tail++; | |
faf88dde | 775 | ctx->cqe_cached++; |
b3659a65 PB |
776 | if (ctx->flags & IORING_SETUP_CQE32) |
777 | ctx->cqe_cached++; | |
778 | return &rings->cqes[off]; | |
faf88dde PB |
779 | } |
780 | ||
eb42cebb PB |
781 | bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, |
782 | bool allow_overflow) | |
bcda7baa | 783 | { |
cd94903d PB |
784 | struct io_uring_cqe *cqe; |
785 | ||
913a571a | 786 | ctx->cq_extra++; |
cd94903d PB |
787 | |
788 | /* | |
789 | * If we can't get a cq entry, userspace overflowed the | |
790 | * submission (by quite a lot). Increment the overflow count in | |
791 | * the ring. | |
792 | */ | |
793 | cqe = io_get_cqe(ctx); | |
794 | if (likely(cqe)) { | |
e0486f3f DY |
795 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
796 | ||
cd94903d PB |
797 | WRITE_ONCE(cqe->user_data, user_data); |
798 | WRITE_ONCE(cqe->res, res); | |
799 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
800 | |
801 | if (ctx->flags & IORING_SETUP_CQE32) { | |
802 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
803 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
804 | } | |
cd94903d PB |
805 | return true; |
806 | } | |
52120f0f DY |
807 | |
808 | if (allow_overflow) | |
809 | return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); | |
810 | ||
811 | return false; | |
bcda7baa JA |
812 | } |
813 | ||
d245bca6 | 814 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, |
52120f0f DY |
815 | u64 user_data, s32 res, u32 cflags, |
816 | bool allow_overflow) | |
d245bca6 PB |
817 | { |
818 | bool filled; | |
819 | ||
25399321 | 820 | io_cq_lock(ctx); |
52120f0f | 821 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow); |
25399321 | 822 | io_cq_unlock_post(ctx); |
d245bca6 PB |
823 | return filled; |
824 | } | |
825 | ||
effcf8bd | 826 | static void __io_req_complete_put(struct io_kiocb *req) |
2b188cc1 | 827 | { |
c7dae4ba JA |
828 | /* |
829 | * If we're the last reference to this request, add to our locked | |
830 | * free_list cache. | |
831 | */ | |
de9b4cca | 832 | if (req_ref_put_and_test(req)) { |
effcf8bd SR |
833 | struct io_ring_ctx *ctx = req->ctx; |
834 | ||
da1a08c5 | 835 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 836 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
837 | io_disarm_next(req); |
838 | if (req->link) { | |
839 | io_req_task_queue(req->link); | |
840 | req->link = NULL; | |
841 | } | |
842 | } | |
7ac1edc4 | 843 | io_req_put_rsrc(req); |
8197b053 PB |
844 | /* |
845 | * Selected buffer deallocation in io_clean_op() assumes that | |
846 | * we don't hold ->completion_lock. Clean them here to avoid | |
847 | * deadlocks. | |
848 | */ | |
849 | io_put_kbuf_comp(req); | |
c7dae4ba JA |
850 | io_dismantle_req(req); |
851 | io_put_task(req->task, 1); | |
c2b6c6bc | 852 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 853 | ctx->locked_free_nr++; |
180f829f | 854 | } |
a37fae8a HX |
855 | } |
856 | ||
59915143 | 857 | void __io_req_complete_post(struct io_kiocb *req) |
effcf8bd | 858 | { |
97b388d7 | 859 | if (!(req->flags & REQ_F_CQE_SKIP)) |
91ef75a7 | 860 | __io_fill_cqe_req(req->ctx, req); |
effcf8bd SR |
861 | __io_req_complete_put(req); |
862 | } | |
863 | ||
59915143 | 864 | void io_req_complete_post(struct io_kiocb *req) |
a37fae8a HX |
865 | { |
866 | struct io_ring_ctx *ctx = req->ctx; | |
867 | ||
25399321 | 868 | io_cq_lock(ctx); |
97b388d7 | 869 | __io_req_complete_post(req); |
25399321 | 870 | io_cq_unlock_post(ctx); |
4e3d9ff9 JA |
871 | } |
872 | ||
99f15d8d | 873 | inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags) |
bcda7baa | 874 | { |
75d7b3ae | 875 | io_req_complete_post(req); |
0ddf92e8 JA |
876 | } |
877 | ||
329061d3 | 878 | void io_req_complete_failed(struct io_kiocb *req, s32 res) |
f41db273 | 879 | { |
a47b255e PB |
880 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
881 | ||
93d2bcd2 | 882 | req_set_fail(req); |
97b388d7 | 883 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
a47b255e PB |
884 | if (def->fail) |
885 | def->fail(req); | |
97b388d7 | 886 | io_req_complete_post(req); |
f41db273 PB |
887 | } |
888 | ||
864ea921 PB |
889 | /* |
890 | * Don't initialise the fields below on every allocation, but do that in | |
891 | * advance and keep them valid across allocations. | |
892 | */ | |
893 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
894 | { | |
895 | req->ctx = ctx; | |
896 | req->link = NULL; | |
897 | req->async_data = NULL; | |
898 | /* not necessary, but safer to zero */ | |
cef216fc | 899 | req->cqe.res = 0; |
864ea921 PB |
900 | } |
901 | ||
dac7a098 | 902 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 903 | struct io_submit_state *state) |
dac7a098 | 904 | { |
79ebeaee | 905 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 906 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 907 | ctx->locked_free_nr = 0; |
79ebeaee | 908 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
909 | } |
910 | ||
5d5901a3 PB |
911 | /* |
912 | * A request might get retired back into the request caches even before opcode | |
913 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
914 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
915 | * and with extra caution to not get a request that is still worked on. | |
916 | */ | |
bd1a3783 | 917 | __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 918 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 919 | { |
864ea921 | 920 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 921 | void *reqs[IO_REQ_ALLOC_BATCH]; |
864ea921 | 922 | int ret, i; |
e5d1bc0a | 923 | |
23a5c43b PB |
924 | /* |
925 | * If we have more than a batch's worth of requests in our IRQ side | |
926 | * locked cache, grab the lock and move them over to our submission | |
927 | * side cache. | |
928 | */ | |
a6d97a8a | 929 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 930 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 931 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
932 | return true; |
933 | } | |
e5d1bc0a | 934 | |
3ab665b7 | 935 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 936 | |
864ea921 PB |
937 | /* |
938 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
939 | * retry single alloc to be on the safe side. | |
940 | */ | |
941 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
942 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
943 | if (!reqs[0]) | |
a33ae9ce | 944 | return false; |
864ea921 | 945 | ret = 1; |
2b188cc1 | 946 | } |
864ea921 | 947 | |
37f0e767 | 948 | percpu_ref_get_many(&ctx->refs, ret); |
3ab665b7 | 949 | for (i = 0; i < ret; i++) { |
23a5c43b | 950 | struct io_kiocb *req = reqs[i]; |
3ab665b7 PB |
951 | |
952 | io_preinit_req(req, ctx); | |
fa05457a | 953 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 954 | } |
a33ae9ce PB |
955 | return true; |
956 | } | |
957 | ||
6b639522 | 958 | static inline void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 959 | { |
094bae49 | 960 | unsigned int flags = req->flags; |
929a3af9 | 961 | |
867f8fa5 | 962 | if (unlikely(flags & IO_REQ_CLEAN_FLAGS)) |
3a0a6902 | 963 | io_clean_op(req); |
e1d767f0 PB |
964 | if (!(flags & REQ_F_FIXED_FILE)) |
965 | io_put_file(req->file); | |
e65ef56d JA |
966 | } |
967 | ||
59915143 | 968 | __cold void io_free_req(struct io_kiocb *req) |
c6ca97b3 | 969 | { |
51a4cc11 | 970 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 971 | |
7ac1edc4 | 972 | io_req_put_rsrc(req); |
216578e5 | 973 | io_dismantle_req(req); |
7c660731 | 974 | io_put_task(req->task, 1); |
c6ca97b3 | 975 | |
79ebeaee | 976 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 977 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
c34b025f | 978 | ctx->locked_free_nr++; |
79ebeaee | 979 | spin_unlock(&ctx->completion_lock); |
e65ef56d JA |
980 | } |
981 | ||
d81499bf PB |
982 | static void __io_req_find_next_prep(struct io_kiocb *req) |
983 | { | |
984 | struct io_ring_ctx *ctx = req->ctx; | |
d81499bf | 985 | |
25399321 | 986 | io_cq_lock(ctx); |
305bef98 | 987 | io_disarm_next(req); |
25399321 | 988 | io_cq_unlock_post(ctx); |
d81499bf PB |
989 | } |
990 | ||
991 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 992 | { |
33cc89a9 | 993 | struct io_kiocb *nxt; |
944e58bf | 994 | |
9e645e11 JA |
995 | /* |
996 | * If LINK is set, we have dependent requests in this chain. If we | |
997 | * didn't fail this request, queue the first one up, moving any other | |
998 | * dependencies to the next request. In case of failure, fail the rest | |
999 | * of the chain. | |
1000 | */ | |
d81499bf PB |
1001 | if (unlikely(req->flags & IO_DISARM_MASK)) |
1002 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
1003 | nxt = req->link; |
1004 | req->link = NULL; | |
1005 | return nxt; | |
4d7dd462 | 1006 | } |
9e645e11 | 1007 | |
f237c30a | 1008 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked) |
2c32395d PB |
1009 | { |
1010 | if (!ctx) | |
1011 | return; | |
ef060ea9 JA |
1012 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1013 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
f237c30a | 1014 | if (*locked) { |
c450178d | 1015 | io_submit_flush_completions(ctx); |
2c32395d | 1016 | mutex_unlock(&ctx->uring_lock); |
f237c30a | 1017 | *locked = false; |
2c32395d PB |
1018 | } |
1019 | percpu_ref_put(&ctx->refs); | |
1020 | } | |
1021 | ||
c6dd763c DY |
1022 | static unsigned int handle_tw_list(struct llist_node *node, |
1023 | struct io_ring_ctx **ctx, bool *locked, | |
1024 | struct llist_node *last) | |
9f8d032a | 1025 | { |
c6dd763c DY |
1026 | unsigned int count = 0; |
1027 | ||
3a0c037b | 1028 | while (node != last) { |
f88262e6 | 1029 | struct llist_node *next = node->next; |
9f8d032a HX |
1030 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1031 | io_task_work.node); | |
1032 | ||
34d2bfe7 JA |
1033 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); |
1034 | ||
9f8d032a HX |
1035 | if (req->ctx != *ctx) { |
1036 | ctx_flush_and_put(*ctx, locked); | |
1037 | *ctx = req->ctx; | |
1038 | /* if not contended, grab and improve batching */ | |
1039 | *locked = mutex_trylock(&(*ctx)->uring_lock); | |
1040 | percpu_ref_get(&(*ctx)->refs); | |
1041 | } | |
1042 | req->io_task_work.func(req, locked); | |
1043 | node = next; | |
c6dd763c | 1044 | count++; |
3a0c037b | 1045 | } |
c6dd763c DY |
1046 | |
1047 | return count; | |
9f8d032a HX |
1048 | } |
1049 | ||
923d1592 DY |
1050 | /** |
1051 | * io_llist_xchg - swap all entries in a lock-less list | |
1052 | * @head: the head of lock-less list to delete all entries | |
1053 | * @new: new entry as the head of the list | |
1054 | * | |
1055 | * If list is empty, return NULL, otherwise, return the pointer to the first entry. | |
1056 | * The order of entries returned is from the newest to the oldest added one. | |
1057 | */ | |
1058 | static inline struct llist_node *io_llist_xchg(struct llist_head *head, | |
1059 | struct llist_node *new) | |
1060 | { | |
1061 | return xchg(&head->first, new); | |
1062 | } | |
1063 | ||
1064 | /** | |
1065 | * io_llist_cmpxchg - possibly swap all entries in a lock-less list | |
1066 | * @head: the head of lock-less list to delete all entries | |
1067 | * @old: expected old value of the first entry of the list | |
1068 | * @new: new entry as the head of the list | |
1069 | * | |
1070 | * perform a cmpxchg on the first entry of the list. | |
1071 | */ | |
1072 | ||
1073 | static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head, | |
1074 | struct llist_node *old, | |
1075 | struct llist_node *new) | |
1076 | { | |
1077 | return cmpxchg(&head->first, old, new); | |
1078 | } | |
1079 | ||
c9f06aa7 | 1080 | void tctx_task_work(struct callback_head *cb) |
c40f6379 | 1081 | { |
f28c240e | 1082 | bool uring_locked = false; |
ebd0df2e | 1083 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
1084 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
1085 | task_work); | |
3a0c037b DY |
1086 | struct llist_node fake = {}; |
1087 | struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake); | |
c6dd763c DY |
1088 | unsigned int loops = 1; |
1089 | unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL); | |
3a0c037b | 1090 | |
3a0c037b DY |
1091 | node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); |
1092 | while (node != &fake) { | |
c6dd763c | 1093 | loops++; |
3a0c037b | 1094 | node = io_llist_xchg(&tctx->task_list, &fake); |
c6dd763c | 1095 | count += handle_tw_list(node, &ctx, &uring_locked, &fake); |
3a0c037b | 1096 | node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); |
3f18407d | 1097 | } |
ebd0df2e | 1098 | |
f28c240e | 1099 | ctx_flush_and_put(ctx, &uring_locked); |
3cc7fdb9 PB |
1100 | |
1101 | /* relaxed read is enough as only the task itself sets ->in_idle */ | |
1102 | if (unlikely(atomic_read(&tctx->in_idle))) | |
1103 | io_uring_drop_tctx_refs(current); | |
c6dd763c DY |
1104 | |
1105 | trace_io_uring_task_work_run(tctx, count, loops); | |
7cbf1722 JA |
1106 | } |
1107 | ||
c0e0d6ba DY |
1108 | static void io_req_local_work_add(struct io_kiocb *req) |
1109 | { | |
1110 | struct io_ring_ctx *ctx = req->ctx; | |
1111 | ||
1112 | if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) | |
1113 | return; | |
fc86f9d3 PB |
1114 | /* need it for the following io_cqring_wake() */ |
1115 | smp_mb__after_atomic(); | |
c0e0d6ba DY |
1116 | |
1117 | if (unlikely(atomic_read(&req->task->io_uring->in_idle))) { | |
1118 | io_move_task_work_from_local(ctx); | |
1119 | return; | |
1120 | } | |
1121 | ||
1122 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) | |
1123 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1124 | ||
21a091b9 DY |
1125 | if (ctx->has_evfd) |
1126 | io_eventfd_signal(ctx); | |
fc86f9d3 | 1127 | __io_cqring_wake(ctx); |
c0e0d6ba DY |
1128 | } |
1129 | ||
1130 | static inline void __io_req_task_work_add(struct io_kiocb *req, bool allow_local) | |
7cbf1722 | 1131 | { |
c34398a8 | 1132 | struct io_uring_task *tctx = req->task->io_uring; |
9f010507 | 1133 | struct io_ring_ctx *ctx = req->ctx; |
f88262e6 | 1134 | struct llist_node *node; |
7cbf1722 | 1135 | |
c0e0d6ba DY |
1136 | if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
1137 | io_req_local_work_add(req); | |
1138 | return; | |
1139 | } | |
7cbf1722 JA |
1140 | |
1141 | /* task_work already pending, we're done */ | |
32d91f05 | 1142 | if (!llist_add(&req->io_task_work.node, &tctx->task_list)) |
e09ee510 | 1143 | return; |
7cbf1722 | 1144 | |
ef060ea9 JA |
1145 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1146 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1147 | ||
3fe07bcd | 1148 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1149 | return; |
2215bed9 | 1150 | |
f88262e6 | 1151 | node = llist_del_all(&tctx->task_list); |
7cbf1722 | 1152 | |
e09ee510 PB |
1153 | while (node) { |
1154 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1155 | node = node->next; | |
3218e5d3 | 1156 | if (llist_add(&req->io_task_work.node, |
e09ee510 PB |
1157 | &req->ctx->fallback_llist)) |
1158 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1159 | } | |
eab30c4d PB |
1160 | } |
1161 | ||
c0e0d6ba DY |
1162 | void io_req_task_work_add(struct io_kiocb *req) |
1163 | { | |
1164 | __io_req_task_work_add(req, true); | |
1165 | } | |
1166 | ||
1167 | static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) | |
1168 | { | |
1169 | struct llist_node *node; | |
1170 | ||
1171 | node = llist_del_all(&ctx->work_llist); | |
1172 | while (node) { | |
1173 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1174 | io_task_work.node); | |
1175 | ||
1176 | node = node->next; | |
1177 | __io_req_task_work_add(req, false); | |
1178 | } | |
1179 | } | |
1180 | ||
b3026767 | 1181 | int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked) |
c0e0d6ba | 1182 | { |
c0e0d6ba DY |
1183 | struct llist_node *node; |
1184 | struct llist_node fake; | |
1185 | struct llist_node *current_final = NULL; | |
1186 | int ret; | |
f75d5036 | 1187 | unsigned int loops = 1; |
c0e0d6ba | 1188 | |
6567506b | 1189 | if (unlikely(ctx->submitter_task != current)) |
c0e0d6ba | 1190 | return -EEXIST; |
c0e0d6ba | 1191 | |
c0e0d6ba DY |
1192 | node = io_llist_xchg(&ctx->work_llist, &fake); |
1193 | ret = 0; | |
1194 | again: | |
1195 | while (node != current_final) { | |
1196 | struct llist_node *next = node->next; | |
1197 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1198 | io_task_work.node); | |
1199 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); | |
b3026767 | 1200 | req->io_task_work.func(req, locked); |
c0e0d6ba DY |
1201 | ret++; |
1202 | node = next; | |
1203 | } | |
1204 | ||
1205 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) | |
1206 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1207 | ||
1208 | node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL); | |
1209 | if (node != &fake) { | |
f75d5036 | 1210 | loops++; |
c0e0d6ba DY |
1211 | current_final = &fake; |
1212 | node = io_llist_xchg(&ctx->work_llist, &fake); | |
1213 | goto again; | |
1214 | } | |
1215 | ||
b3026767 | 1216 | if (*locked) |
c0e0d6ba | 1217 | io_submit_flush_completions(ctx); |
f75d5036 | 1218 | trace_io_uring_local_work_run(ctx, ret, loops); |
c0e0d6ba | 1219 | return ret; |
8ac5d85a JA |
1220 | |
1221 | } | |
1222 | ||
1223 | int io_run_local_work(struct io_ring_ctx *ctx) | |
1224 | { | |
1225 | bool locked; | |
1226 | int ret; | |
1227 | ||
7924fdfe PB |
1228 | if (llist_empty(&ctx->work_llist)) |
1229 | return 0; | |
1230 | ||
ec7fd256 | 1231 | __set_current_state(TASK_RUNNING); |
8ac5d85a | 1232 | locked = mutex_trylock(&ctx->uring_lock); |
b3026767 | 1233 | ret = __io_run_local_work(ctx, &locked); |
8ac5d85a JA |
1234 | if (locked) |
1235 | mutex_unlock(&ctx->uring_lock); | |
1236 | ||
1237 | return ret; | |
c0e0d6ba DY |
1238 | } |
1239 | ||
4e118cd9 | 1240 | static void io_req_tw_post(struct io_kiocb *req, bool *locked) |
c40f6379 | 1241 | { |
97b388d7 | 1242 | io_req_complete_post(req); |
4e118cd9 | 1243 | } |
c40f6379 | 1244 | |
59915143 | 1245 | void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags) |
4e118cd9 | 1246 | { |
97b388d7 | 1247 | io_req_set_res(req, res, cflags); |
4e118cd9 | 1248 | req->io_task_work.func = io_req_tw_post; |
3fe07bcd | 1249 | io_req_task_work_add(req); |
4e118cd9 PB |
1250 | } |
1251 | ||
f237c30a | 1252 | static void io_req_task_cancel(struct io_kiocb *req, bool *locked) |
c40f6379 | 1253 | { |
b18a1a45 | 1254 | /* not needed for normal modes, but SQPOLL depends on it */ |
971cf9c1 | 1255 | io_tw_lock(req->ctx, locked); |
cef216fc | 1256 | io_req_complete_failed(req, req->cqe.res); |
c40f6379 JA |
1257 | } |
1258 | ||
329061d3 | 1259 | void io_req_task_submit(struct io_kiocb *req, bool *locked) |
c40f6379 | 1260 | { |
971cf9c1 | 1261 | io_tw_lock(req->ctx, locked); |
316319e8 | 1262 | /* req->task == current here, checking PF_EXITING is safe */ |
af066f31 | 1263 | if (likely(!(req->task->flags & PF_EXITING))) |
cbc2e203 | 1264 | io_queue_sqe(req); |
81b6d05c | 1265 | else |
2593553a | 1266 | io_req_complete_failed(req, -EFAULT); |
c40f6379 JA |
1267 | } |
1268 | ||
59915143 | 1269 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1270 | { |
97b388d7 | 1271 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1272 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1273 | io_req_task_work_add(req); |
c40f6379 JA |
1274 | } |
1275 | ||
f3b44f92 | 1276 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1277 | { |
5b0a6acc | 1278 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1279 | io_req_task_work_add(req); |
a3df7698 PB |
1280 | } |
1281 | ||
59915143 | 1282 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1283 | { |
9b5f7bd9 | 1284 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1285 | |
1286 | if (nxt) | |
906a8c3f | 1287 | io_req_task_queue(nxt); |
c69f8dbe JL |
1288 | } |
1289 | ||
f3b44f92 | 1290 | void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node) |
3aa83bfb | 1291 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1292 | { |
d4b7a5ef | 1293 | struct task_struct *task = NULL; |
37f0e767 | 1294 | int task_refs = 0; |
5af1d13e | 1295 | |
3aa83bfb PB |
1296 | do { |
1297 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1298 | comp_list); | |
2d6500d4 | 1299 | |
a538be5b PB |
1300 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1301 | if (req->flags & REQ_F_REFCOUNT) { | |
1302 | node = req->comp_list.next; | |
1303 | if (!req_ref_put_and_test(req)) | |
1304 | continue; | |
1305 | } | |
b605a7fa PB |
1306 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1307 | struct async_poll *apoll = req->apoll; | |
1308 | ||
1309 | if (apoll->double_poll) | |
1310 | kfree(apoll->double_poll); | |
9731bc98 JA |
1311 | if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) |
1312 | kfree(apoll); | |
b605a7fa PB |
1313 | req->flags &= ~REQ_F_POLLED; |
1314 | } | |
da1a08c5 | 1315 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1316 | io_queue_next(req); |
a538be5b PB |
1317 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1318 | io_clean_op(req); | |
c1e53a69 | 1319 | } |
a538be5b PB |
1320 | if (!(req->flags & REQ_F_FIXED_FILE)) |
1321 | io_put_file(req->file); | |
2d6500d4 | 1322 | |
ab409402 | 1323 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1324 | |
d4b7a5ef PB |
1325 | if (req->task != task) { |
1326 | if (task) | |
1327 | io_put_task(task, task_refs); | |
1328 | task = req->task; | |
1329 | task_refs = 0; | |
1330 | } | |
1331 | task_refs++; | |
c1e53a69 | 1332 | node = req->comp_list.next; |
fa05457a | 1333 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1334 | } while (node); |
d4b7a5ef | 1335 | |
d4b7a5ef PB |
1336 | if (task) |
1337 | io_put_task(task, task_refs); | |
7a743e22 PB |
1338 | } |
1339 | ||
c450178d | 1340 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1341 | __must_hold(&ctx->uring_lock) |
905c172f | 1342 | { |
6f33b0bc | 1343 | struct io_wq_work_node *node, *prev; |
cd0ca2e0 | 1344 | struct io_submit_state *state = &ctx->submit_state; |
905c172f | 1345 | |
e9a88428 | 1346 | io_cq_lock(ctx); |
d9dee430 PB |
1347 | wq_list_for_each(node, prev, &state->compl_reqs) { |
1348 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1349 | comp_list); | |
3d4aeb9f | 1350 | |
d9dee430 PB |
1351 | if (!(req->flags & REQ_F_CQE_SKIP)) |
1352 | __io_fill_cqe_req(ctx, req); | |
905c172f | 1353 | } |
25399321 | 1354 | __io_cq_unlock_post(ctx); |
d9dee430 | 1355 | |
1cce17ac | 1356 | io_free_batch_list(ctx, state->compl_reqs.first); |
6f33b0bc | 1357 | INIT_WQ_LIST(&state->compl_reqs); |
7a743e22 PB |
1358 | } |
1359 | ||
ba816ad6 JA |
1360 | /* |
1361 | * Drop reference to request, return next in chain (if there is one) if this | |
1362 | * was the last reference to this request. | |
1363 | */ | |
0d85035a | 1364 | static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 1365 | { |
9b5f7bd9 PB |
1366 | struct io_kiocb *nxt = NULL; |
1367 | ||
de9b4cca | 1368 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 1369 | if (unlikely(req->flags & IO_REQ_LINK_FLAGS)) |
7819a1f6 | 1370 | nxt = io_req_find_next(req); |
f5c6cf2a | 1371 | io_free_req(req); |
2a44f467 | 1372 | } |
9b5f7bd9 | 1373 | return nxt; |
2b188cc1 JA |
1374 | } |
1375 | ||
6c503150 | 1376 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1377 | { |
1378 | /* See comment at the top of this file */ | |
1379 | smp_rmb(); | |
e23de15f | 1380 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1381 | } |
1382 | ||
def596e9 JA |
1383 | /* |
1384 | * We can't just wait for polled events to come to us, we have to actively | |
1385 | * find and complete them. | |
1386 | */ | |
c072481d | 1387 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1388 | { |
1389 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1390 | return; | |
1391 | ||
1392 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1393 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1394 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1395 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1396 | break; |
08f5439f JA |
1397 | /* |
1398 | * Ensure we allow local-to-the-cpu processing to take place, | |
1399 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1400 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1401 | */ |
3fcee5a6 PB |
1402 | if (need_resched()) { |
1403 | mutex_unlock(&ctx->uring_lock); | |
1404 | cond_resched(); | |
1405 | mutex_lock(&ctx->uring_lock); | |
1406 | } | |
def596e9 JA |
1407 | } |
1408 | mutex_unlock(&ctx->uring_lock); | |
1409 | } | |
1410 | ||
7668b92a | 1411 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1412 | { |
7668b92a | 1413 | unsigned int nr_events = 0; |
e9979b36 | 1414 | int ret = 0; |
155bc950 | 1415 | unsigned long check_cq; |
500f9fba | 1416 | |
76de6749 PB |
1417 | if (!io_allowed_run_tw(ctx)) |
1418 | return -EEXIST; | |
1419 | ||
3a08576b PB |
1420 | check_cq = READ_ONCE(ctx->check_cq); |
1421 | if (unlikely(check_cq)) { | |
1422 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
1423 | __io_cqring_overflow_flush(ctx, false); | |
1424 | /* | |
1425 | * Similarly do not spin if we have not informed the user of any | |
1426 | * dropped CQE. | |
1427 | */ | |
1428 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1429 | return -EBADR; | |
1430 | } | |
f39c8a5b PB |
1431 | /* |
1432 | * Don't enter poll loop if we already have events pending. | |
1433 | * If we do, we can potentially be spinning for commands that | |
1434 | * already triggered a CQE (eg in error). | |
1435 | */ | |
f39c8a5b | 1436 | if (io_cqring_events(ctx)) |
d487b43c | 1437 | return 0; |
155bc950 | 1438 | |
def596e9 | 1439 | do { |
500f9fba JA |
1440 | /* |
1441 | * If a submit got punted to a workqueue, we can have the | |
1442 | * application entering polling for a command before it gets | |
1443 | * issued. That app will hold the uring_lock for the duration | |
1444 | * of the poll right here, so we need to take a breather every | |
1445 | * now and then to ensure that the issue has a chance to add | |
1446 | * the poll to the issued list. Otherwise we can spin here | |
1447 | * forever, while the workqueue is stuck trying to acquire the | |
1448 | * very same mutex. | |
1449 | */ | |
dac6a0ea JA |
1450 | if (wq_list_empty(&ctx->iopoll_list) || |
1451 | io_task_work_pending(ctx)) { | |
8f487ef2 PB |
1452 | u32 tail = ctx->cached_cq_tail; |
1453 | ||
8de11cdc | 1454 | (void) io_run_local_work_locked(ctx); |
def596e9 | 1455 | |
dac6a0ea JA |
1456 | if (task_work_pending(current) || |
1457 | wq_list_empty(&ctx->iopoll_list)) { | |
dac6a0ea | 1458 | mutex_unlock(&ctx->uring_lock); |
9d54bd6a | 1459 | io_run_task_work(); |
dac6a0ea | 1460 | mutex_lock(&ctx->uring_lock); |
dac6a0ea | 1461 | } |
8f487ef2 PB |
1462 | /* some requests don't go through iopoll_list */ |
1463 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1464 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1465 | break; |
500f9fba | 1466 | } |
5ba3c874 PB |
1467 | ret = io_do_iopoll(ctx, !min); |
1468 | if (ret < 0) | |
1469 | break; | |
1470 | nr_events += ret; | |
1471 | ret = 0; | |
1472 | } while (nr_events < min && !need_resched()); | |
d487b43c | 1473 | |
def596e9 JA |
1474 | return ret; |
1475 | } | |
7012c815 PB |
1476 | |
1477 | void io_req_task_complete(struct io_kiocb *req, bool *locked) | |
8ef12efe | 1478 | { |
7012c815 | 1479 | if (*locked) |
9da070b1 | 1480 | io_req_complete_defer(req); |
7012c815 | 1481 | else |
97b388d7 | 1482 | io_req_complete_post(req); |
8ef12efe JA |
1483 | } |
1484 | ||
def596e9 JA |
1485 | /* |
1486 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1487 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1488 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1489 | * accessing the kiocb cookie. |
1490 | */ | |
9882131c | 1491 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1492 | { |
1493 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1494 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1495 | |
1496 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1497 | if (unlikely(needs_lock)) |
cb3d8972 | 1498 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1499 | |
1500 | /* | |
1501 | * Track whether we have multiple files in our lists. This will impact | |
1502 | * how we do polling eventually, not spinning if we're on potentially | |
1503 | * different devices. | |
1504 | */ | |
5eef4e87 | 1505 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1506 | ctx->poll_multi_queue = false; |
1507 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1508 | struct io_kiocb *list_req; |
1509 | ||
5eef4e87 PB |
1510 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1511 | comp_list); | |
30da1b45 | 1512 | if (list_req->file != req->file) |
915b3dde | 1513 | ctx->poll_multi_queue = true; |
def596e9 JA |
1514 | } |
1515 | ||
1516 | /* | |
1517 | * For fast devices, IO may have already completed. If it has, add | |
1518 | * it to the front so we find it first. | |
1519 | */ | |
65a6543d | 1520 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1521 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1522 | else |
5eef4e87 | 1523 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1524 | |
3b44b371 | 1525 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1526 | /* |
1527 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1528 | * in sq thread task context or in io worker task context. If | |
1529 | * current task context is sq thread, we don't need to check | |
1530 | * whether should wake up sq thread. | |
1531 | */ | |
1532 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1533 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1534 | wake_up(&ctx->sq_data->wait); | |
1535 | ||
1536 | mutex_unlock(&ctx->uring_lock); | |
1537 | } | |
def596e9 JA |
1538 | } |
1539 | ||
4503b767 JA |
1540 | static bool io_bdev_nowait(struct block_device *bdev) |
1541 | { | |
568ec936 | 1542 | return !bdev || bdev_nowait(bdev); |
4503b767 JA |
1543 | } |
1544 | ||
2b188cc1 JA |
1545 | /* |
1546 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1547 | * any file. For now, just ensure that anything potentially problematic is done | |
1548 | * inline. | |
1549 | */ | |
88459b50 | 1550 | static bool __io_file_supports_nowait(struct file *file, umode_t mode) |
2b188cc1 | 1551 | { |
4503b767 | 1552 | if (S_ISBLK(mode)) { |
4e7b5671 CH |
1553 | if (IS_ENABLED(CONFIG_BLOCK) && |
1554 | io_bdev_nowait(I_BDEV(file->f_mapping->host))) | |
4503b767 JA |
1555 | return true; |
1556 | return false; | |
1557 | } | |
976517f1 | 1558 | if (S_ISSOCK(mode)) |
2b188cc1 | 1559 | return true; |
4503b767 | 1560 | if (S_ISREG(mode)) { |
4e7b5671 CH |
1561 | if (IS_ENABLED(CONFIG_BLOCK) && |
1562 | io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
e5550a14 | 1563 | !io_is_uring_fops(file)) |
4503b767 JA |
1564 | return true; |
1565 | return false; | |
1566 | } | |
2b188cc1 | 1567 | |
c5b85625 JA |
1568 | /* any ->read/write should understand O_NONBLOCK */ |
1569 | if (file->f_flags & O_NONBLOCK) | |
1570 | return true; | |
35645ac3 | 1571 | return file->f_mode & FMODE_NOWAIT; |
2b188cc1 | 1572 | } |
c5b85625 | 1573 | |
88459b50 PB |
1574 | /* |
1575 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1576 | * any file. For now, just ensure that anything potentially problematic is done | |
1577 | * inline. | |
1578 | */ | |
a4ad4f74 | 1579 | unsigned int io_file_get_flags(struct file *file) |
88459b50 PB |
1580 | { |
1581 | umode_t mode = file_inode(file)->i_mode; | |
1582 | unsigned int res = 0; | |
af197f50 | 1583 | |
88459b50 PB |
1584 | if (S_ISREG(mode)) |
1585 | res |= FFS_ISREG; | |
1586 | if (__io_file_supports_nowait(file, mode)) | |
1587 | res |= FFS_NOWAIT; | |
1588 | return res; | |
2b188cc1 JA |
1589 | } |
1590 | ||
99f15d8d | 1591 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1592 | { |
e8c2bc1f JA |
1593 | WARN_ON_ONCE(!io_op_defs[req->opcode].async_size); |
1594 | req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL); | |
d886e185 PB |
1595 | if (req->async_data) { |
1596 | req->flags |= REQ_F_ASYNC_DATA; | |
1597 | return false; | |
1598 | } | |
1599 | return true; | |
3d9932a8 XW |
1600 | } |
1601 | ||
f3b44f92 | 1602 | int io_req_prep_async(struct io_kiocb *req) |
f67676d1 | 1603 | { |
0702e536 JA |
1604 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
1605 | ||
1606 | /* assign early for deferred execution for non-fixed file */ | |
1607 | if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE)) | |
1608 | req->file = io_file_get_normal(req, req->cqe.fd); | |
dc919caf | 1609 | if (!def->prep_async) |
0702e536 JA |
1610 | return 0; |
1611 | if (WARN_ON_ONCE(req_has_async_data(req))) | |
1612 | return -EFAULT; | |
59169439 PB |
1613 | if (!io_op_defs[req->opcode].manual_alloc) { |
1614 | if (io_alloc_async_data(req)) | |
1615 | return -EAGAIN; | |
1616 | } | |
dc919caf | 1617 | return def->prep_async(req); |
bfe76559 PB |
1618 | } |
1619 | ||
9cf7c104 PB |
1620 | static u32 io_get_sequence(struct io_kiocb *req) |
1621 | { | |
a3dbdf54 | 1622 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1623 | struct io_kiocb *cur; |
9cf7c104 | 1624 | |
a3dbdf54 | 1625 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1626 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1627 | seq--; |
1628 | return seq; | |
9cf7c104 PB |
1629 | } |
1630 | ||
c072481d | 1631 | static __cold void io_drain_req(struct io_kiocb *req) |
de0617e4 | 1632 | { |
a197f664 | 1633 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1634 | struct io_defer_entry *de; |
f67676d1 | 1635 | int ret; |
e0eb71dc | 1636 | u32 seq = io_get_sequence(req); |
3c19966d | 1637 | |
9d858b21 | 1638 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1639 | spin_lock(&ctx->completion_lock); |
5e371265 | 1640 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1641 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1642 | queue: |
10c66904 | 1643 | ctx->drain_active = false; |
e0eb71dc PB |
1644 | io_req_task_queue(req); |
1645 | return; | |
10c66904 | 1646 | } |
e302f104 | 1647 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1648 | |
b7e298d2 | 1649 | ret = io_req_prep_async(req); |
e0eb71dc PB |
1650 | if (ret) { |
1651 | fail: | |
1652 | io_req_complete_failed(req, ret); | |
1653 | return; | |
1654 | } | |
cbdcb435 | 1655 | io_prep_async_link(req); |
27dc8338 | 1656 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1657 | if (!de) { |
1b48773f | 1658 | ret = -ENOMEM; |
e0eb71dc | 1659 | goto fail; |
76cc33d7 | 1660 | } |
2d28390a | 1661 | |
79ebeaee | 1662 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1663 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1664 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1665 | kfree(de); |
e0eb71dc | 1666 | goto queue; |
de0617e4 JA |
1667 | } |
1668 | ||
48863ffd | 1669 | trace_io_uring_defer(req); |
27dc8338 | 1670 | de->req = req; |
9cf7c104 | 1671 | de->seq = seq; |
27dc8338 | 1672 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1673 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1674 | } |
1675 | ||
68fb8979 | 1676 | static void io_clean_op(struct io_kiocb *req) |
99bc4c38 | 1677 | { |
8197b053 PB |
1678 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
1679 | spin_lock(&req->ctx->completion_lock); | |
cc3cec83 | 1680 | io_put_kbuf_comp(req); |
8197b053 PB |
1681 | spin_unlock(&req->ctx->completion_lock); |
1682 | } | |
99bc4c38 | 1683 | |
0e1b6fe3 | 1684 | if (req->flags & REQ_F_NEED_CLEANUP) { |
4d4c9cff | 1685 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
bb040a21 | 1686 | |
4d4c9cff JA |
1687 | if (def->cleanup) |
1688 | def->cleanup(req); | |
99bc4c38 | 1689 | } |
75652a30 JA |
1690 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1691 | kfree(req->apoll->double_poll); | |
1692 | kfree(req->apoll); | |
1693 | req->apoll = NULL; | |
1694 | } | |
9cae36a0 JA |
1695 | if (req->flags & REQ_F_INFLIGHT) { |
1696 | struct io_uring_task *tctx = req->task->io_uring; | |
1697 | ||
1698 | atomic_dec(&tctx->inflight_tracked); | |
1699 | } | |
c854357b | 1700 | if (req->flags & REQ_F_CREDS) |
b8e64b53 | 1701 | put_cred(req->creds); |
d886e185 PB |
1702 | if (req->flags & REQ_F_ASYNC_DATA) { |
1703 | kfree(req->async_data); | |
1704 | req->async_data = NULL; | |
1705 | } | |
c854357b | 1706 | req->flags &= ~IO_REQ_CLEAN_FLAGS; |
99bc4c38 PB |
1707 | } |
1708 | ||
6bf9c47a JA |
1709 | static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags) |
1710 | { | |
1711 | if (req->file || !io_op_defs[req->opcode].needs_file) | |
1712 | return true; | |
1713 | ||
1714 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1715 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1716 | else |
cef216fc | 1717 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1718 | |
772f5e00 | 1719 | return !!req->file; |
6bf9c47a JA |
1720 | } |
1721 | ||
889fca73 | 1722 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1723 | { |
fcde59fe | 1724 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
5730b27e | 1725 | const struct cred *creds = NULL; |
d625c6ee | 1726 | int ret; |
2b188cc1 | 1727 | |
70152140 JA |
1728 | if (unlikely(!io_assign_file(req, issue_flags))) |
1729 | return -EBADF; | |
1730 | ||
6878b40e | 1731 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1732 | creds = override_creds(req->creds); |
5730b27e | 1733 | |
fcde59fe | 1734 | if (!def->audit_skip) |
5bd2182d PM |
1735 | audit_uring_entry(req->opcode); |
1736 | ||
0702e536 | 1737 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1738 | |
fcde59fe | 1739 | if (!def->audit_skip) |
5bd2182d PM |
1740 | audit_uring_exit(!ret, ret); |
1741 | ||
5730b27e JA |
1742 | if (creds) |
1743 | revert_creds(creds); | |
97b388d7 | 1744 | |
75d7b3ae PB |
1745 | if (ret == IOU_OK) { |
1746 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
9da070b1 | 1747 | io_req_complete_defer(req); |
75d7b3ae PB |
1748 | else |
1749 | io_req_complete_post(req); | |
1750 | } else if (ret != IOU_ISSUE_SKIP_COMPLETE) | |
def596e9 | 1751 | return ret; |
97b388d7 | 1752 | |
b532576e | 1753 | /* If the op doesn't have a file, we're not polling for it */ |
9983028e | 1754 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file) |
9882131c | 1755 | io_iopoll_req_issued(req, issue_flags); |
def596e9 JA |
1756 | |
1757 | return 0; | |
2b188cc1 JA |
1758 | } |
1759 | ||
329061d3 JA |
1760 | int io_poll_issue(struct io_kiocb *req, bool *locked) |
1761 | { | |
1762 | io_tw_lock(req->ctx, locked); | |
1763 | if (unlikely(req->task->flags & PF_EXITING)) | |
1764 | return -EFAULT; | |
91482864 | 1765 | return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT); |
329061d3 JA |
1766 | } |
1767 | ||
c9f06aa7 | 1768 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1769 | { |
1770 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1771 | ||
1772 | req = io_put_req_find_next(req); | |
1773 | return req ? &req->work : NULL; | |
1774 | } | |
1775 | ||
c9f06aa7 | 1776 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1777 | { |
1778 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6bf9c47a | 1779 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
d01905db PB |
1780 | unsigned int issue_flags = IO_URING_F_UNLOCKED; |
1781 | bool needs_poll = false; | |
6bf9c47a | 1782 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1783 | |
48dcd38d PB |
1784 | /* one will be dropped by ->io_free_work() after returning to io-wq */ |
1785 | if (!(req->flags & REQ_F_REFCOUNT)) | |
1786 | __io_req_set_refcount(req, 2); | |
1787 | else | |
1788 | req_ref_get(req); | |
5d5901a3 | 1789 | |
cb2d344c | 1790 | io_arm_ltimeout(req); |
6bf9c47a | 1791 | |
dadebc35 | 1792 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1793 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1794 | fail: |
6bf9c47a | 1795 | io_req_task_queue_fail(req, err); |
d01905db PB |
1796 | return; |
1797 | } | |
0f8da75b PB |
1798 | if (!io_assign_file(req, issue_flags)) { |
1799 | err = -EBADF; | |
1800 | work->flags |= IO_WQ_WORK_CANCEL; | |
1801 | goto fail; | |
1802 | } | |
31b51510 | 1803 | |
d01905db | 1804 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
1805 | bool opcode_poll = def->pollin || def->pollout; |
1806 | ||
1807 | if (opcode_poll && file_can_poll(req->file)) { | |
1808 | needs_poll = true; | |
d01905db | 1809 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 1810 | } |
561fb04a | 1811 | } |
31b51510 | 1812 | |
d01905db PB |
1813 | do { |
1814 | ret = io_issue_sqe(req, issue_flags); | |
1815 | if (ret != -EAGAIN) | |
1816 | break; | |
1817 | /* | |
1818 | * We can get EAGAIN for iopolled IO even though we're | |
1819 | * forcing a sync submission from here, since we can't | |
1820 | * wait for request slots on the block side. | |
1821 | */ | |
1822 | if (!needs_poll) { | |
e0deb6a0 PB |
1823 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
1824 | break; | |
d01905db PB |
1825 | cond_resched(); |
1826 | continue; | |
90fa0288 HX |
1827 | } |
1828 | ||
4d9237e3 | 1829 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
1830 | return; |
1831 | /* aborted or ready, in either case retry blocking */ | |
1832 | needs_poll = false; | |
1833 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
1834 | } while (1); | |
31b51510 | 1835 | |
a3df7698 | 1836 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 1837 | if (ret < 0) |
a3df7698 | 1838 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
1839 | } |
1840 | ||
531113bb JA |
1841 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
1842 | unsigned int issue_flags) | |
09bb8394 | 1843 | { |
5106dd6e JA |
1844 | struct io_ring_ctx *ctx = req->ctx; |
1845 | struct file *file = NULL; | |
ac177053 | 1846 | unsigned long file_ptr; |
09bb8394 | 1847 | |
93f052cb | 1848 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 1849 | |
ac177053 | 1850 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 1851 | goto out; |
ac177053 PB |
1852 | fd = array_index_nospec(fd, ctx->nr_user_files); |
1853 | file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; | |
1854 | file = (struct file *) (file_ptr & FFS_MASK); | |
1855 | file_ptr &= ~FFS_MASK; | |
1856 | /* mask in overlapping REQ_F and FFS bits */ | |
35645ac3 | 1857 | req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT); |
5106dd6e JA |
1858 | io_req_set_rsrc_node(req, ctx, 0); |
1859 | out: | |
93f052cb | 1860 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
1861 | return file; |
1862 | } | |
d44f554e | 1863 | |
531113bb | 1864 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 1865 | { |
62906e89 | 1866 | struct file *file = fget(fd); |
ac177053 | 1867 | |
48863ffd | 1868 | trace_io_uring_file_get(req, fd); |
09bb8394 | 1869 | |
ac177053 | 1870 | /* we don't allow fixed io_uring files */ |
e5550a14 | 1871 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 1872 | io_req_track_inflight(req); |
8371adf5 | 1873 | return file; |
09bb8394 JA |
1874 | } |
1875 | ||
7bfa9bad | 1876 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
1877 | __must_hold(&req->ctx->uring_lock) |
1878 | { | |
7bfa9bad PB |
1879 | struct io_kiocb *linked_timeout; |
1880 | ||
1881 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
1882 | io_req_complete_failed(req, ret); | |
1883 | return; | |
1884 | } | |
1885 | ||
1886 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 1887 | |
4d9237e3 | 1888 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 1889 | case IO_APOLL_READY: |
336d28a8 | 1890 | io_kbuf_recycle(req, 0); |
d475a9a6 PB |
1891 | io_req_task_queue(req); |
1892 | break; | |
1893 | case IO_APOLL_ABORTED: | |
6436c770 | 1894 | io_kbuf_recycle(req, 0); |
77955efb | 1895 | io_queue_iowq(req, NULL); |
d475a9a6 | 1896 | break; |
b1c62645 | 1897 | case IO_APOLL_OK: |
b1c62645 | 1898 | break; |
d475a9a6 PB |
1899 | } |
1900 | ||
1901 | if (linked_timeout) | |
1902 | io_queue_linked_timeout(linked_timeout); | |
1903 | } | |
1904 | ||
cbc2e203 | 1905 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 1906 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 1907 | { |
e0c5c576 | 1908 | int ret; |
2b188cc1 | 1909 | |
c5eef2b9 | 1910 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 1911 | |
491381ce JA |
1912 | /* |
1913 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
1914 | * doesn't support non-blocking read/write attempts | |
1915 | */ | |
7bfa9bad | 1916 | if (likely(!ret)) |
cb2d344c | 1917 | io_arm_ltimeout(req); |
7bfa9bad PB |
1918 | else |
1919 | io_queue_async(req, ret); | |
2b188cc1 JA |
1920 | } |
1921 | ||
4652fe3f | 1922 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 1923 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 1924 | { |
17b147f6 PB |
1925 | if (unlikely(req->flags & REQ_F_FAIL)) { |
1926 | /* | |
1927 | * We don't submit, fail them all, for that replace hardlinks | |
1928 | * with normal links. Extra REQ_F_LINK is tolerated. | |
1929 | */ | |
1930 | req->flags &= ~REQ_F_HARDLINK; | |
1931 | req->flags |= REQ_F_LINK; | |
1932 | io_req_complete_failed(req, req->cqe.res); | |
e0eb71dc PB |
1933 | } else if (unlikely(req->ctx->drain_active)) { |
1934 | io_drain_req(req); | |
76cc33d7 PB |
1935 | } else { |
1936 | int ret = io_req_prep_async(req); | |
1937 | ||
1938 | if (unlikely(ret)) | |
1939 | io_req_complete_failed(req, ret); | |
1940 | else | |
77955efb | 1941 | io_queue_iowq(req, NULL); |
ce35a47a | 1942 | } |
4fe2c963 JL |
1943 | } |
1944 | ||
b16fed66 PB |
1945 | /* |
1946 | * Check SQE restrictions (opcode and flags). | |
1947 | * | |
1948 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
1949 | */ | |
1950 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
1951 | struct io_kiocb *req, | |
1952 | unsigned int sqe_flags) | |
4fe2c963 | 1953 | { |
b16fed66 PB |
1954 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
1955 | return false; | |
1956 | ||
1957 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
1958 | ctx->restrictions.sqe_flags_required) | |
1959 | return false; | |
1960 | ||
1961 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
1962 | ctx->restrictions.sqe_flags_required)) | |
1963 | return false; | |
1964 | ||
1965 | return true; | |
4fe2c963 JL |
1966 | } |
1967 | ||
22b2ca31 PB |
1968 | static void io_init_req_drain(struct io_kiocb *req) |
1969 | { | |
1970 | struct io_ring_ctx *ctx = req->ctx; | |
1971 | struct io_kiocb *head = ctx->submit_state.link.head; | |
1972 | ||
1973 | ctx->drain_active = true; | |
1974 | if (head) { | |
1975 | /* | |
1976 | * If we need to drain a request in the middle of a link, drain | |
1977 | * the head request and the next request/link after the current | |
1978 | * link. Considering sequential execution of links, | |
b6c7db32 | 1979 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
1980 | * link. |
1981 | */ | |
b6c7db32 | 1982 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
1983 | ctx->drain_next = true; |
1984 | } | |
1985 | } | |
1986 | ||
b16fed66 PB |
1987 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1988 | const struct io_uring_sqe *sqe) | |
282cdc86 | 1989 | __must_hold(&ctx->uring_lock) |
b16fed66 | 1990 | { |
fcde59fe | 1991 | const struct io_op_def *def; |
b16fed66 | 1992 | unsigned int sqe_flags; |
fc0ae024 | 1993 | int personality; |
4a04d1d1 | 1994 | u8 opcode; |
b16fed66 | 1995 | |
864ea921 | 1996 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 1997 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 PB |
1998 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
1999 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
cef216fc | 2000 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 2001 | req->file = NULL; |
c1bdf8ed | 2002 | req->rsrc_node = NULL; |
b16fed66 | 2003 | req->task = current; |
b16fed66 | 2004 | |
4a04d1d1 PB |
2005 | if (unlikely(opcode >= IORING_OP_LAST)) { |
2006 | req->opcode = 0; | |
b16fed66 | 2007 | return -EINVAL; |
4a04d1d1 | 2008 | } |
fcde59fe | 2009 | def = &io_op_defs[opcode]; |
68fe256a PB |
2010 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
2011 | /* enforce forwards compatibility on users */ | |
2012 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
2013 | return -EINVAL; | |
4e906702 | 2014 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 2015 | if (!def->buffer_select) |
4e906702 JA |
2016 | return -EOPNOTSUPP; |
2017 | req->buf_index = READ_ONCE(sqe->buf_group); | |
2018 | } | |
5562a8d7 PB |
2019 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
2020 | ctx->drain_disabled = true; | |
2021 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
2022 | if (ctx->drain_disabled) | |
2023 | return -EOPNOTSUPP; | |
22b2ca31 | 2024 | io_init_req_drain(req); |
5562a8d7 | 2025 | } |
2a56a9bd PB |
2026 | } |
2027 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
2028 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
2029 | return -EACCES; | |
2030 | /* knock it to the slow queue path, will be drained there */ | |
2031 | if (ctx->drain_active) | |
2032 | req->flags |= REQ_F_FORCE_ASYNC; | |
2033 | /* if there is no link, we're at "next" request and need to drain */ | |
2034 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
2035 | ctx->drain_next = false; | |
2036 | ctx->drain_active = true; | |
b6c7db32 | 2037 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 2038 | } |
68fe256a | 2039 | } |
b16fed66 | 2040 | |
fcde59fe | 2041 | if (!def->ioprio && sqe->ioprio) |
73911426 | 2042 | return -EINVAL; |
fcde59fe | 2043 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
73911426 JA |
2044 | return -EINVAL; |
2045 | ||
fcde59fe | 2046 | if (def->needs_file) { |
6d63416d PB |
2047 | struct io_submit_state *state = &ctx->submit_state; |
2048 | ||
cef216fc | 2049 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 2050 | |
6d63416d PB |
2051 | /* |
2052 | * Plug now if we have more than 2 IO left after this, and the | |
2053 | * target is potentially a read/write to block based storage. | |
2054 | */ | |
fcde59fe | 2055 | if (state->need_plug && def->plug) { |
6d63416d PB |
2056 | state->plug_started = true; |
2057 | state->need_plug = false; | |
5ca7a8b3 | 2058 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 2059 | } |
b16fed66 | 2060 | } |
863e0560 | 2061 | |
003e8dcc JA |
2062 | personality = READ_ONCE(sqe->personality); |
2063 | if (personality) { | |
cdab10bf LT |
2064 | int ret; |
2065 | ||
c10d1f98 PB |
2066 | req->creds = xa_load(&ctx->personalities, personality); |
2067 | if (!req->creds) | |
003e8dcc | 2068 | return -EINVAL; |
c10d1f98 | 2069 | get_cred(req->creds); |
cdc1404a PM |
2070 | ret = security_uring_override_creds(req->creds); |
2071 | if (ret) { | |
2072 | put_cred(req->creds); | |
2073 | return ret; | |
2074 | } | |
b8e64b53 | 2075 | req->flags |= REQ_F_CREDS; |
003e8dcc | 2076 | } |
b16fed66 | 2077 | |
0702e536 | 2078 | return def->prep(req, sqe); |
b16fed66 PB |
2079 | } |
2080 | ||
df3becde PB |
2081 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
2082 | struct io_kiocb *req, int ret) | |
2083 | { | |
2084 | struct io_ring_ctx *ctx = req->ctx; | |
2085 | struct io_submit_link *link = &ctx->submit_state.link; | |
2086 | struct io_kiocb *head = link->head; | |
2087 | ||
48863ffd | 2088 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
2089 | |
2090 | /* | |
2091 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
2092 | * unusable. Instead of failing eagerly, continue assembling the link if | |
2093 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
2094 | * should find the flag and handle the rest. | |
2095 | */ | |
2096 | req_fail_link_node(req, ret); | |
2097 | if (head && !(head->flags & REQ_F_FAIL)) | |
2098 | req_fail_link_node(head, -ECANCELED); | |
2099 | ||
2100 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
2101 | if (head) { | |
2102 | link->last->link = req; | |
2103 | link->head = NULL; | |
2104 | req = head; | |
2105 | } | |
2106 | io_queue_sqe_fallback(req); | |
2107 | return ret; | |
2108 | } | |
2109 | ||
2110 | if (head) | |
2111 | link->last->link = req; | |
2112 | else | |
2113 | link->head = req; | |
2114 | link->last = req; | |
2115 | return 0; | |
2116 | } | |
2117 | ||
2118 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 2119 | const struct io_uring_sqe *sqe) |
282cdc86 | 2120 | __must_hold(&ctx->uring_lock) |
9e645e11 | 2121 | { |
a1ab7b35 | 2122 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 2123 | int ret; |
9e645e11 | 2124 | |
a6b8cadc | 2125 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
2126 | if (unlikely(ret)) |
2127 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 2128 | |
be7053b7 | 2129 | /* don't need @sqe from now on */ |
48863ffd | 2130 | trace_io_uring_submit_sqe(req, true); |
a6b8cadc | 2131 | |
9e645e11 JA |
2132 | /* |
2133 | * If we already have a head request, queue this one for async | |
2134 | * submittal once the head completes. If we don't have a head but | |
2135 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2136 | * submitted sync once the chain is complete. If none of those | |
2137 | * conditions are true (normal request), then just queue it. | |
2138 | */ | |
924a07e4 | 2139 | if (unlikely(link->head)) { |
df3becde PB |
2140 | ret = io_req_prep_async(req); |
2141 | if (unlikely(ret)) | |
2142 | return io_submit_fail_init(sqe, req, ret); | |
2143 | ||
48863ffd | 2144 | trace_io_uring_link(req, link->head); |
f2f87370 | 2145 | link->last->link = req; |
863e0560 | 2146 | link->last = req; |
32fe525b | 2147 | |
da1a08c5 | 2148 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 2149 | return 0; |
df3becde PB |
2150 | /* last request of the link, flush it */ |
2151 | req = link->head; | |
f15a3431 | 2152 | link->head = NULL; |
924a07e4 PB |
2153 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
2154 | goto fallback; | |
2155 | ||
2156 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2157 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2158 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2159 | link->head = req; | |
2160 | link->last = req; | |
2161 | } else { | |
2162 | fallback: | |
2163 | io_queue_sqe_fallback(req); | |
2164 | } | |
f15a3431 | 2165 | return 0; |
9e645e11 | 2166 | } |
2e6e1fde | 2167 | |
f15a3431 | 2168 | io_queue_sqe(req); |
1d4240cc | 2169 | return 0; |
9e645e11 JA |
2170 | } |
2171 | ||
9a56a232 JA |
2172 | /* |
2173 | * Batched submission is done, ensure local IO is flushed out. | |
2174 | */ | |
553deffd | 2175 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2176 | { |
553deffd PB |
2177 | struct io_submit_state *state = &ctx->submit_state; |
2178 | ||
e126391c PB |
2179 | if (unlikely(state->link.head)) |
2180 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2181 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2182 | io_submit_flush_completions(ctx); |
27926b68 JA |
2183 | if (state->plug_started) |
2184 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2185 | } |
2186 | ||
2187 | /* | |
2188 | * Start submission side cache. | |
2189 | */ | |
2190 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2191 | unsigned int max_ios) |
9a56a232 | 2192 | { |
27926b68 | 2193 | state->plug_started = false; |
4b628aeb | 2194 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2195 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2196 | /* set only head, no need to init link_last in advance */ |
2197 | state->link.head = NULL; | |
9a56a232 JA |
2198 | } |
2199 | ||
2b188cc1 JA |
2200 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2201 | { | |
75b28aff | 2202 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2203 | |
caf582c6 PB |
2204 | /* |
2205 | * Ensure any loads from the SQEs are done at this point, | |
2206 | * since once we write the new head, the application could | |
2207 | * write new data to them. | |
2208 | */ | |
2209 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2210 | } |
2211 | ||
2b188cc1 | 2212 | /* |
dd9ae8a0 | 2213 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2214 | * that is mapped by userspace. This means that care needs to be taken to |
2215 | * ensure that reads are stable, as we cannot rely on userspace always | |
2216 | * being a good citizen. If members of the sqe are validated and then later | |
2217 | * used, it's important that those reads are done through READ_ONCE() to | |
2218 | * prevent a re-load down the line. | |
2219 | */ | |
709b302f | 2220 | static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) |
2b188cc1 | 2221 | { |
ea5ab3b5 | 2222 | unsigned head, mask = ctx->sq_entries - 1; |
17d3aeb3 | 2223 | unsigned sq_idx = ctx->cached_sq_head++ & mask; |
2b188cc1 JA |
2224 | |
2225 | /* | |
2226 | * The cached sq head (or cq tail) serves two purposes: | |
2227 | * | |
2228 | * 1) allows us to batch the cost of updating the user visible | |
2229 | * head updates. | |
2230 | * 2) allows the kernel side to track the head on its own, even | |
2231 | * though the application is the one updating it. | |
2232 | */ | |
17d3aeb3 | 2233 | head = READ_ONCE(ctx->sq_array[sq_idx]); |
ebdeb7c0 JA |
2234 | if (likely(head < ctx->sq_entries)) { |
2235 | /* double index for 128-byte SQEs, twice as long */ | |
2236 | if (ctx->flags & IORING_SETUP_SQE128) | |
2237 | head <<= 1; | |
709b302f | 2238 | return &ctx->sq_sqes[head]; |
ebdeb7c0 | 2239 | } |
2b188cc1 JA |
2240 | |
2241 | /* drop invalid entries */ | |
15641e42 PB |
2242 | ctx->cq_extra--; |
2243 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2244 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
709b302f PB |
2245 | return NULL; |
2246 | } | |
2247 | ||
17437f31 | 2248 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2249 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2250 | { |
69629809 | 2251 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2252 | unsigned int left; |
2253 | int ret; | |
6c271ce2 | 2254 | |
51d48dab | 2255 | if (unlikely(!entries)) |
69629809 | 2256 | return 0; |
ee7d46d9 | 2257 | /* make sure SQ entry isn't read before tail */ |
8e6971a8 PB |
2258 | ret = left = min3(nr, ctx->sq_entries, entries); |
2259 | io_get_task_refs(left); | |
2260 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2261 | |
69629809 | 2262 | do { |
3529d8c2 | 2263 | const struct io_uring_sqe *sqe; |
196be95c | 2264 | struct io_kiocb *req; |
fb5ccc98 | 2265 | |
8e6971a8 | 2266 | if (unlikely(!io_alloc_req_refill(ctx))) |
fb5ccc98 | 2267 | break; |
a33ae9ce | 2268 | req = io_alloc_req(ctx); |
4fccfcbb PB |
2269 | sqe = io_get_sqe(ctx); |
2270 | if (unlikely(!sqe)) { | |
fa05457a | 2271 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2272 | break; |
2273 | } | |
6c271ce2 | 2274 | |
1cd15904 PB |
2275 | /* |
2276 | * Continue submitting even for sqe failure if the | |
2277 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2278 | */ | |
2279 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2280 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2281 | left--; | |
2282 | break; | |
bcbb7bf6 | 2283 | } |
1cd15904 | 2284 | } while (--left); |
9466f437 | 2285 | |
8e6971a8 PB |
2286 | if (unlikely(left)) { |
2287 | ret -= left; | |
2288 | /* try again if it submitted nothing and can't allocate a req */ | |
2289 | if (!ret && io_req_cache_empty(ctx)) | |
2290 | ret = -EAGAIN; | |
2291 | current->io_uring->cached_refs += left; | |
9466f437 | 2292 | } |
6c271ce2 | 2293 | |
553deffd | 2294 | io_submit_state_end(ctx); |
ae9428ca PB |
2295 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2296 | io_commit_sqring(ctx); | |
8e6971a8 | 2297 | return ret; |
6c271ce2 JA |
2298 | } |
2299 | ||
bda52162 JA |
2300 | struct io_wait_queue { |
2301 | struct wait_queue_entry wq; | |
2302 | struct io_ring_ctx *ctx; | |
5fd46178 | 2303 | unsigned cq_tail; |
bda52162 JA |
2304 | unsigned nr_timeouts; |
2305 | }; | |
2306 | ||
b4c98d59 DY |
2307 | static inline bool io_has_work(struct io_ring_ctx *ctx) |
2308 | { | |
c0e0d6ba DY |
2309 | return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) || |
2310 | ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && | |
2311 | !llist_empty(&ctx->work_llist)); | |
b4c98d59 DY |
2312 | } |
2313 | ||
6c503150 | 2314 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
2315 | { |
2316 | struct io_ring_ctx *ctx = iowq->ctx; | |
0fc8c2ac | 2317 | int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail; |
bda52162 JA |
2318 | |
2319 | /* | |
d195a66e | 2320 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
2321 | * started waiting. For timeouts, we always want to return to userspace, |
2322 | * regardless of event count. | |
2323 | */ | |
5fd46178 | 2324 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
2325 | } |
2326 | ||
2327 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2328 | int wake_flags, void *key) | |
2329 | { | |
2330 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
2331 | wq); | |
b4c98d59 | 2332 | struct io_ring_ctx *ctx = iowq->ctx; |
bda52162 | 2333 | |
6c503150 PB |
2334 | /* |
2335 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2336 | * the task, and the next invocation will do it. | |
2337 | */ | |
b4c98d59 | 2338 | if (io_should_wake(iowq) || io_has_work(ctx)) |
6c503150 PB |
2339 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2340 | return -1; | |
bda52162 JA |
2341 | } |
2342 | ||
c0e0d6ba | 2343 | int io_run_task_work_sig(struct io_ring_ctx *ctx) |
af9c1a44 | 2344 | { |
c0e0d6ba | 2345 | if (io_run_task_work_ctx(ctx) > 0) |
af9c1a44 | 2346 | return 1; |
c5020bc8 OL |
2347 | if (task_sigpending(current)) |
2348 | return -EINTR; | |
2349 | return 0; | |
af9c1a44 JA |
2350 | } |
2351 | ||
eeb60b9a PB |
2352 | /* when returns >0, the caller should retry */ |
2353 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
2354 | struct io_wait_queue *iowq, | |
22833966 | 2355 | ktime_t timeout) |
eeb60b9a PB |
2356 | { |
2357 | int ret; | |
155bc950 | 2358 | unsigned long check_cq; |
eeb60b9a PB |
2359 | |
2360 | /* make sure we run task_work before checking for signals */ | |
c0e0d6ba | 2361 | ret = io_run_task_work_sig(ctx); |
eeb60b9a PB |
2362 | if (ret || io_should_wake(iowq)) |
2363 | return ret; | |
3a08576b | 2364 | |
155bc950 | 2365 | check_cq = READ_ONCE(ctx->check_cq); |
3a08576b PB |
2366 | if (unlikely(check_cq)) { |
2367 | /* let the caller flush overflows, retry */ | |
2368 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
2369 | return 1; | |
2370 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
2371 | return -EBADR; | |
2372 | } | |
22833966 JA |
2373 | if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS)) |
2374 | return -ETIME; | |
2375 | return 1; | |
eeb60b9a PB |
2376 | } |
2377 | ||
2b188cc1 JA |
2378 | /* |
2379 | * Wait until events become available, if we don't already have some. The | |
2380 | * application must reap them itself, as they reside on the shared cq ring. | |
2381 | */ | |
2382 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2383 | const sigset_t __user *sig, size_t sigsz, |
2384 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2385 | { |
90291099 | 2386 | struct io_wait_queue iowq; |
75b28aff | 2387 | struct io_rings *rings = ctx->rings; |
22833966 | 2388 | ktime_t timeout = KTIME_MAX; |
c1d5a224 | 2389 | int ret; |
2b188cc1 | 2390 | |
76de6749 PB |
2391 | if (!io_allowed_run_tw(ctx)) |
2392 | return -EEXIST; | |
2393 | ||
b41e9852 | 2394 | do { |
c0e0d6ba DY |
2395 | /* always run at least 1 task work to process local work */ |
2396 | ret = io_run_task_work_ctx(ctx); | |
2397 | if (ret < 0) | |
2398 | return ret; | |
90f67366 | 2399 | io_cqring_overflow_flush(ctx); |
aa1df3a3 | 2400 | |
0fc8c2ac DY |
2401 | /* if user messes with these they will just get an early return */ |
2402 | if (__io_cqring_events_user(ctx) >= min_events) | |
b41e9852 | 2403 | return 0; |
c0e0d6ba | 2404 | } while (ret > 0); |
2b188cc1 JA |
2405 | |
2406 | if (sig) { | |
9e75ad5d AB |
2407 | #ifdef CONFIG_COMPAT |
2408 | if (in_compat_syscall()) | |
2409 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2410 | sigsz); |
9e75ad5d AB |
2411 | else |
2412 | #endif | |
b772434b | 2413 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2414 | |
2b188cc1 JA |
2415 | if (ret) |
2416 | return ret; | |
2417 | } | |
2418 | ||
950e79dd OL |
2419 | if (uts) { |
2420 | struct timespec64 ts; | |
2421 | ||
2422 | if (get_timespec64(&ts, uts)) | |
2423 | return -EFAULT; | |
2424 | timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); | |
2425 | } | |
2426 | ||
90291099 PB |
2427 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
2428 | iowq.wq.private = current; | |
2429 | INIT_LIST_HEAD(&iowq.wq.entry); | |
2430 | iowq.ctx = ctx; | |
bda52162 | 2431 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 2432 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
90291099 | 2433 | |
c826bd7a | 2434 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2435 | do { |
ca0a2651 | 2436 | /* if we can't even flush overflow, don't wait for more */ |
90f67366 | 2437 | if (!io_cqring_overflow_flush(ctx)) { |
ca0a2651 JA |
2438 | ret = -EBUSY; |
2439 | break; | |
2440 | } | |
311997b3 | 2441 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, |
bda52162 | 2442 | TASK_INTERRUPTIBLE); |
22833966 | 2443 | ret = io_cqring_wait_schedule(ctx, &iowq, timeout); |
ca0a2651 | 2444 | cond_resched(); |
eeb60b9a | 2445 | } while (ret > 0); |
bda52162 | 2446 | |
b4f20bb4 | 2447 | finish_wait(&ctx->cq_wait, &iowq.wq); |
b7db41c9 | 2448 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2449 | |
75b28aff | 2450 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2451 | } |
2452 | ||
73572984 | 2453 | static void io_mem_free(void *ptr) |
b9bd2bea | 2454 | { |
73572984 | 2455 | struct page *page; |
b36a2050 | 2456 | |
73572984 JA |
2457 | if (!ptr) |
2458 | return; | |
b9bd2bea | 2459 | |
73572984 JA |
2460 | page = virt_to_head_page(ptr); |
2461 | if (put_page_testzero(page)) | |
2462 | free_compound_page(page); | |
b9bd2bea PB |
2463 | } |
2464 | ||
73572984 | 2465 | static void *io_mem_alloc(size_t size) |
b9bd2bea | 2466 | { |
73572984 | 2467 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; |
b9bd2bea | 2468 | |
73572984 | 2469 | return (void *) __get_free_pages(gfp, get_order(size)); |
b9bd2bea PB |
2470 | } |
2471 | ||
73572984 JA |
2472 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2473 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2474 | { |
73572984 JA |
2475 | struct io_rings *rings; |
2476 | size_t off, sq_array_size; | |
6b06314c | 2477 | |
73572984 JA |
2478 | off = struct_size(rings, cqes, cq_entries); |
2479 | if (off == SIZE_MAX) | |
2480 | return SIZE_MAX; | |
2481 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2482 | if (check_shl_overflow(off, 1, &off)) | |
2483 | return SIZE_MAX; | |
2484 | } | |
ab409402 | 2485 | |
73572984 JA |
2486 | #ifdef CONFIG_SMP |
2487 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2488 | if (off == 0) | |
2489 | return SIZE_MAX; | |
2490 | #endif | |
82fbcfa9 | 2491 | |
73572984 JA |
2492 | if (sq_offset) |
2493 | *sq_offset = off; | |
82fbcfa9 | 2494 | |
73572984 JA |
2495 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2496 | if (sq_array_size == SIZE_MAX) | |
2497 | return SIZE_MAX; | |
6b06314c | 2498 | |
73572984 JA |
2499 | if (check_add_overflow(off, sq_array_size, &off)) |
2500 | return SIZE_MAX; | |
8bad28d8 | 2501 | |
73572984 | 2502 | return off; |
8bad28d8 HX |
2503 | } |
2504 | ||
73572984 JA |
2505 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, |
2506 | unsigned int eventfd_async) | |
8bad28d8 | 2507 | { |
73572984 JA |
2508 | struct io_ev_fd *ev_fd; |
2509 | __s32 __user *fds = arg; | |
2510 | int fd; | |
f2303b1f | 2511 | |
73572984 JA |
2512 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, |
2513 | lockdep_is_held(&ctx->uring_lock)); | |
2514 | if (ev_fd) | |
2515 | return -EBUSY; | |
8bad28d8 | 2516 | |
73572984 JA |
2517 | if (copy_from_user(&fd, fds, sizeof(*fds))) |
2518 | return -EFAULT; | |
8dd03afe | 2519 | |
73572984 JA |
2520 | ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); |
2521 | if (!ev_fd) | |
2522 | return -ENOMEM; | |
05f3fb3c | 2523 | |
73572984 JA |
2524 | ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); |
2525 | if (IS_ERR(ev_fd->cq_ev_fd)) { | |
2526 | int ret = PTR_ERR(ev_fd->cq_ev_fd); | |
2527 | kfree(ev_fd); | |
2528 | return ret; | |
2529 | } | |
305bef98 PB |
2530 | |
2531 | spin_lock(&ctx->completion_lock); | |
2532 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
2533 | spin_unlock(&ctx->completion_lock); | |
2534 | ||
73572984 JA |
2535 | ev_fd->eventfd_async = eventfd_async; |
2536 | ctx->has_evfd = true; | |
2537 | rcu_assign_pointer(ctx->io_ev_fd, ev_fd); | |
21a091b9 DY |
2538 | atomic_set(&ev_fd->refs, 1); |
2539 | atomic_set(&ev_fd->ops, 0); | |
73572984 | 2540 | return 0; |
d7954b2b BM |
2541 | } |
2542 | ||
73572984 | 2543 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 2544 | { |
73572984 JA |
2545 | struct io_ev_fd *ev_fd; |
2546 | ||
2547 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, | |
2548 | lockdep_is_held(&ctx->uring_lock)); | |
2549 | if (ev_fd) { | |
2550 | ctx->has_evfd = false; | |
2551 | rcu_assign_pointer(ctx->io_ev_fd, NULL); | |
21a091b9 DY |
2552 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops)) |
2553 | call_rcu(&ev_fd->rcu, io_eventfd_ops); | |
73572984 JA |
2554 | return 0; |
2555 | } | |
2d091d62 | 2556 | |
73572984 | 2557 | return -ENXIO; |
44b31f2f PB |
2558 | } |
2559 | ||
73572984 | 2560 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2561 | { |
37f0e767 | 2562 | int nr = 0; |
bf019da7 | 2563 | |
9a4fdbd8 | 2564 | mutex_lock(&ctx->uring_lock); |
34f0bc42 | 2565 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
9a4fdbd8 | 2566 | |
88ab95be | 2567 | while (!io_req_cache_empty(ctx)) { |
34f0bc42 | 2568 | struct io_kiocb *req = io_alloc_req(ctx); |
9a4fdbd8 | 2569 | |
c2b6c6bc | 2570 | kmem_cache_free(req_cachep, req); |
37f0e767 | 2571 | nr++; |
c2b6c6bc | 2572 | } |
37f0e767 PB |
2573 | if (nr) |
2574 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2575 | mutex_unlock(&ctx->uring_lock); |
2576 | } | |
2577 | ||
c072481d | 2578 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2579 | { |
37d1e2e3 | 2580 | io_sq_thread_finish(ctx); |
ab409402 | 2581 | io_rsrc_refs_drop(ctx); |
43597aac PB |
2582 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
2583 | io_wait_rsrc_data(ctx->buf_data); | |
2584 | io_wait_rsrc_data(ctx->file_data); | |
2585 | ||
8bad28d8 | 2586 | mutex_lock(&ctx->uring_lock); |
43597aac | 2587 | if (ctx->buf_data) |
bd54b6fe | 2588 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2589 | if (ctx->file_data) |
08480400 | 2590 | __io_sqe_files_unregister(ctx); |
c4ea060e PB |
2591 | if (ctx->rings) |
2592 | __io_cqring_overflow_flush(ctx, true); | |
9b402849 | 2593 | io_eventfd_unregister(ctx); |
9b797a37 | 2594 | io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); |
43e0bbbd | 2595 | io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); |
77bc59b4 | 2596 | mutex_unlock(&ctx->uring_lock); |
5a2e745d | 2597 | io_destroy_buffers(ctx); |
07db298a PB |
2598 | if (ctx->sq_creds) |
2599 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2600 | if (ctx->submitter_task) |
2601 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2602 | |
a7f0ed5a PB |
2603 | /* there are no registered resources left, nobody uses it */ |
2604 | if (ctx->rsrc_node) | |
2605 | io_rsrc_node_destroy(ctx->rsrc_node); | |
8dd03afe | 2606 | if (ctx->rsrc_backup_node) |
b895c9a6 | 2607 | io_rsrc_node_destroy(ctx->rsrc_backup_node); |
a7f0ed5a | 2608 | flush_delayed_work(&ctx->rsrc_put_work); |
756ab7c0 | 2609 | flush_delayed_work(&ctx->fallback_work); |
a7f0ed5a PB |
2610 | |
2611 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
2612 | WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist)); | |
def596e9 | 2613 | |
2b188cc1 | 2614 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
2615 | if (ctx->ring_sock) { |
2616 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 2617 | sock_release(ctx->ring_sock); |
355e8d26 | 2618 | } |
2b188cc1 | 2619 | #endif |
ef9dd637 | 2620 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
2b188cc1 | 2621 | |
42b6419d PB |
2622 | if (ctx->mm_account) { |
2623 | mmdrop(ctx->mm_account); | |
2624 | ctx->mm_account = NULL; | |
2625 | } | |
75b28aff | 2626 | io_mem_free(ctx->rings); |
2b188cc1 | 2627 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
2628 | |
2629 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2630 | free_uid(ctx->user); |
4010fec4 | 2631 | io_req_caches_free(ctx); |
e941894e JA |
2632 | if (ctx->hash_map) |
2633 | io_wq_put_hash(ctx->hash_map); | |
e6f89be6 | 2634 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2635 | kfree(ctx->cancel_table_locked.hbs); |
6224843d | 2636 | kfree(ctx->dummy_ubuf); |
9cfc7e94 JA |
2637 | kfree(ctx->io_bl); |
2638 | xa_destroy(&ctx->io_bl_xa); | |
2b188cc1 JA |
2639 | kfree(ctx); |
2640 | } | |
2641 | ||
2642 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
2643 | { | |
2644 | struct io_ring_ctx *ctx = file->private_data; | |
2645 | __poll_t mask = 0; | |
2646 | ||
d60aa65b | 2647 | poll_wait(file, &ctx->cq_wait, wait); |
4f7067c3 SB |
2648 | /* |
2649 | * synchronizes with barrier from wq_has_sleeper call in | |
2650 | * io_commit_cqring | |
2651 | */ | |
2b188cc1 | 2652 | smp_rmb(); |
90554200 | 2653 | if (!io_sqring_full(ctx)) |
2b188cc1 | 2654 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
2655 | |
2656 | /* | |
2657 | * Don't flush cqring overflow list here, just do a simple check. | |
2658 | * Otherwise there could possible be ABBA deadlock: | |
2659 | * CPU0 CPU1 | |
2660 | * ---- ---- | |
2661 | * lock(&ctx->uring_lock); | |
2662 | * lock(&ep->mtx); | |
2663 | * lock(&ctx->uring_lock); | |
2664 | * lock(&ep->mtx); | |
2665 | * | |
2666 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
2667 | * pushs them to do the flush. | |
2668 | */ | |
b4c98d59 DY |
2669 | |
2670 | if (io_cqring_events(ctx) || io_has_work(ctx)) | |
2b188cc1 JA |
2671 | mask |= EPOLLIN | EPOLLRDNORM; |
2672 | ||
2673 | return mask; | |
2674 | } | |
2675 | ||
0bead8cd | 2676 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 2677 | { |
4379bf8b | 2678 | const struct cred *creds; |
071698e1 | 2679 | |
61cf9370 | 2680 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
2681 | if (creds) { |
2682 | put_cred(creds); | |
0bead8cd | 2683 | return 0; |
1e6fa521 | 2684 | } |
0bead8cd YD |
2685 | |
2686 | return -EINVAL; | |
2687 | } | |
2688 | ||
d56d938b PB |
2689 | struct io_tctx_exit { |
2690 | struct callback_head task_work; | |
2691 | struct completion completion; | |
baf186c4 | 2692 | struct io_ring_ctx *ctx; |
d56d938b PB |
2693 | }; |
2694 | ||
c072481d | 2695 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
2696 | { |
2697 | struct io_uring_task *tctx = current->io_uring; | |
2698 | struct io_tctx_exit *work; | |
2699 | ||
2700 | work = container_of(cb, struct io_tctx_exit, task_work); | |
2701 | /* | |
2702 | * When @in_idle, we're in cancellation and it's racy to remove the | |
2703 | * node. It'll be removed by the end of cancellation, just ignore it. | |
2704 | */ | |
2705 | if (!atomic_read(&tctx->in_idle)) | |
eef51daa | 2706 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
2707 | complete(&work->completion); |
2708 | } | |
2709 | ||
c072481d | 2710 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
2711 | { |
2712 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2713 | ||
2714 | return req->ctx == data; | |
2715 | } | |
2716 | ||
c072481d | 2717 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 2718 | { |
d56d938b | 2719 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 2720 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 2721 | unsigned long interval = HZ / 20; |
d56d938b PB |
2722 | struct io_tctx_exit exit; |
2723 | struct io_tctx_node *node; | |
2724 | int ret; | |
85faa7b8 | 2725 | |
56952e91 JA |
2726 | /* |
2727 | * If we're doing polled IO and end up having requests being | |
2728 | * submitted async (out-of-line), then completions can come in while | |
2729 | * we're waiting for refs to drop. We need to reap these manually, | |
2730 | * as nobody else will be looking for them. | |
2731 | */ | |
b2edc0a7 | 2732 | do { |
c0e0d6ba DY |
2733 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) |
2734 | io_move_task_work_from_local(ctx); | |
2735 | ||
affa87db PB |
2736 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
2737 | cond_resched(); | |
2738 | ||
28090c13 PB |
2739 | if (ctx->sq_data) { |
2740 | struct io_sq_data *sqd = ctx->sq_data; | |
2741 | struct task_struct *tsk; | |
2742 | ||
2743 | io_sq_thread_park(sqd); | |
2744 | tsk = sqd->thread; | |
2745 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
2746 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
2747 | io_cancel_ctx_cb, ctx, true); | |
2748 | io_sq_thread_unpark(sqd); | |
2749 | } | |
b5bb3a24 | 2750 | |
37f0e767 PB |
2751 | io_req_caches_free(ctx); |
2752 | ||
58d3be2c PB |
2753 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
2754 | /* there is little hope left, don't run it too often */ | |
2755 | interval = HZ * 60; | |
2756 | } | |
2757 | } while (!wait_for_completion_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 2758 | |
7f00651a PB |
2759 | init_completion(&exit.completion); |
2760 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
2761 | exit.ctx = ctx; | |
89b5066e PB |
2762 | /* |
2763 | * Some may use context even when all refs and requests have been put, | |
2764 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 2765 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
2766 | * this lock/unlock section also waits them to finish. |
2767 | */ | |
d56d938b PB |
2768 | mutex_lock(&ctx->uring_lock); |
2769 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
2770 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
2771 | ||
d56d938b PB |
2772 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
2773 | ctx_node); | |
7f00651a PB |
2774 | /* don't spin on a single task if cancellation failed */ |
2775 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
2776 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
2777 | if (WARN_ON_ONCE(ret)) | |
2778 | continue; | |
d56d938b PB |
2779 | |
2780 | mutex_unlock(&ctx->uring_lock); | |
2781 | wait_for_completion(&exit.completion); | |
d56d938b PB |
2782 | mutex_lock(&ctx->uring_lock); |
2783 | } | |
2784 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
2785 | spin_lock(&ctx->completion_lock); |
2786 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 2787 | |
85faa7b8 JA |
2788 | io_ring_ctx_free(ctx); |
2789 | } | |
2790 | ||
c072481d | 2791 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 2792 | { |
61cf9370 MWO |
2793 | unsigned long index; |
2794 | struct creds *creds; | |
2795 | ||
2b188cc1 JA |
2796 | mutex_lock(&ctx->uring_lock); |
2797 | percpu_ref_kill(&ctx->refs); | |
634578f8 | 2798 | if (ctx->rings) |
6c2450ae | 2799 | __io_cqring_overflow_flush(ctx, true); |
61cf9370 MWO |
2800 | xa_for_each(&ctx->personalities, index, creds) |
2801 | io_unregister_personality(ctx, index); | |
9ca9fb24 PB |
2802 | if (ctx->rings) |
2803 | io_poll_remove_all(ctx, NULL, true); | |
2b188cc1 JA |
2804 | mutex_unlock(&ctx->uring_lock); |
2805 | ||
02bac94b PB |
2806 | /* |
2807 | * If we failed setting up the ctx, we might not have any rings | |
2808 | * and therefore did not submit any requests | |
2809 | */ | |
2810 | if (ctx->rings) | |
60053be8 | 2811 | io_kill_timeouts(ctx, NULL, true); |
309fc03a | 2812 | |
85faa7b8 | 2813 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
2814 | /* |
2815 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
2816 | * if we're exiting a ton of rings at the same time. It just adds | |
2817 | * noise and overhead, there's no discernable change in runtime | |
2818 | * over using system_wq. | |
2819 | */ | |
2820 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
2821 | } |
2822 | ||
2823 | static int io_uring_release(struct inode *inode, struct file *file) | |
2824 | { | |
2825 | struct io_ring_ctx *ctx = file->private_data; | |
2826 | ||
2827 | file->private_data = NULL; | |
2828 | io_ring_ctx_wait_and_kill(ctx); | |
2829 | return 0; | |
2830 | } | |
2831 | ||
f6edbabb PB |
2832 | struct io_task_cancel { |
2833 | struct task_struct *task; | |
3dd0c97a | 2834 | bool all; |
f6edbabb | 2835 | }; |
f254ac04 | 2836 | |
f6edbabb | 2837 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 2838 | { |
9a472ef7 | 2839 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 2840 | struct io_task_cancel *cancel = data; |
9a472ef7 | 2841 | |
6af3f48b | 2842 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
2843 | } |
2844 | ||
c072481d PB |
2845 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
2846 | struct task_struct *task, | |
2847 | bool cancel_all) | |
b7ddce3c | 2848 | { |
e1915f76 | 2849 | struct io_defer_entry *de; |
b7ddce3c PB |
2850 | LIST_HEAD(list); |
2851 | ||
79ebeaee | 2852 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 2853 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 2854 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
2855 | list_cut_position(&list, &ctx->defer_list, &de->list); |
2856 | break; | |
2857 | } | |
2858 | } | |
79ebeaee | 2859 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
2860 | if (list_empty(&list)) |
2861 | return false; | |
b7ddce3c PB |
2862 | |
2863 | while (!list_empty(&list)) { | |
2864 | de = list_first_entry(&list, struct io_defer_entry, list); | |
2865 | list_del_init(&de->list); | |
f41db273 | 2866 | io_req_complete_failed(de->req, -ECANCELED); |
b7ddce3c PB |
2867 | kfree(de); |
2868 | } | |
e1915f76 | 2869 | return true; |
b7ddce3c PB |
2870 | } |
2871 | ||
c072481d | 2872 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
2873 | { |
2874 | struct io_tctx_node *node; | |
2875 | enum io_wq_cancel cret; | |
2876 | bool ret = false; | |
2877 | ||
2878 | mutex_lock(&ctx->uring_lock); | |
2879 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
2880 | struct io_uring_task *tctx = node->task->io_uring; | |
2881 | ||
2882 | /* | |
2883 | * io_wq will stay alive while we hold uring_lock, because it's | |
2884 | * killed after ctx nodes, which requires to take the lock. | |
2885 | */ | |
2886 | if (!tctx || !tctx->io_wq) | |
2887 | continue; | |
2888 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
2889 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
2890 | } | |
2891 | mutex_unlock(&ctx->uring_lock); | |
2892 | ||
2893 | return ret; | |
2894 | } | |
2895 | ||
affa87db | 2896 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
2897 | struct task_struct *task, |
2898 | bool cancel_all) | |
9936c7c2 | 2899 | { |
3dd0c97a | 2900 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 2901 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
2902 | enum io_wq_cancel cret; |
2903 | bool ret = false; | |
9936c7c2 | 2904 | |
60053be8 PB |
2905 | /* failed during ring init, it couldn't have issued any requests */ |
2906 | if (!ctx->rings) | |
affa87db | 2907 | return false; |
60053be8 | 2908 | |
affa87db PB |
2909 | if (!task) { |
2910 | ret |= io_uring_try_cancel_iowq(ctx); | |
2911 | } else if (tctx && tctx->io_wq) { | |
2912 | /* | |
2913 | * Cancels requests of all rings, not only @ctx, but | |
2914 | * it's fine as the task is in exit/exec. | |
2915 | */ | |
2916 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
2917 | &cancel, true); | |
2918 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
2919 | } | |
9936c7c2 | 2920 | |
affa87db PB |
2921 | /* SQPOLL thread does its own polling */ |
2922 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
2923 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
2924 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
2925 | io_iopoll_try_reap_events(ctx); | |
2926 | ret = true; | |
9936c7c2 | 2927 | } |
9936c7c2 | 2928 | } |
affa87db | 2929 | |
c0e0d6ba DY |
2930 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) |
2931 | ret |= io_run_local_work(ctx) > 0; | |
affa87db PB |
2932 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
2933 | mutex_lock(&ctx->uring_lock); | |
2934 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
2935 | mutex_unlock(&ctx->uring_lock); | |
2936 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
2937 | if (task) | |
c0e0d6ba | 2938 | ret |= io_run_task_work() > 0; |
affa87db | 2939 | return ret; |
9936c7c2 PB |
2940 | } |
2941 | ||
3f48cf18 | 2942 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 2943 | { |
3f48cf18 | 2944 | if (tracked) |
9cae36a0 | 2945 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
2946 | return percpu_counter_sum(&tctx->inflight); |
2947 | } | |
2948 | ||
78cc687b PB |
2949 | /* |
2950 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 2951 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 2952 | */ |
17437f31 | 2953 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 2954 | { |
521d6a73 | 2955 | struct io_uring_task *tctx = current->io_uring; |
734551df | 2956 | struct io_ring_ctx *ctx; |
0e9ddb39 PB |
2957 | s64 inflight; |
2958 | DEFINE_WAIT(wait); | |
fdaf083c | 2959 | |
78cc687b PB |
2960 | WARN_ON_ONCE(sqd && sqd->thread != current); |
2961 | ||
6d042ffb PO |
2962 | if (!current->io_uring) |
2963 | return; | |
17a91051 PB |
2964 | if (tctx->io_wq) |
2965 | io_wq_exit_start(tctx->io_wq); | |
2966 | ||
0e9ddb39 PB |
2967 | atomic_inc(&tctx->in_idle); |
2968 | do { | |
affa87db PB |
2969 | bool loop = false; |
2970 | ||
e9dbe221 | 2971 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 2972 | /* read completions before cancelations */ |
78cc687b | 2973 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
2974 | if (!inflight) |
2975 | break; | |
fdaf083c | 2976 | |
78cc687b PB |
2977 | if (!sqd) { |
2978 | struct io_tctx_node *node; | |
2979 | unsigned long index; | |
0f212204 | 2980 | |
78cc687b PB |
2981 | xa_for_each(&tctx->xa, index, node) { |
2982 | /* sqpoll task will cancel all its requests */ | |
2983 | if (node->ctx->sq_data) | |
2984 | continue; | |
affa87db PB |
2985 | loop |= io_uring_try_cancel_requests(node->ctx, |
2986 | current, cancel_all); | |
78cc687b PB |
2987 | } |
2988 | } else { | |
2989 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
2990 | loop |= io_uring_try_cancel_requests(ctx, |
2991 | current, | |
2992 | cancel_all); | |
2993 | } | |
2994 | ||
2995 | if (loop) { | |
2996 | cond_resched(); | |
2997 | continue; | |
78cc687b | 2998 | } |
17a91051 | 2999 | |
78a78060 JA |
3000 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
3001 | io_run_task_work(); | |
e9dbe221 | 3002 | io_uring_drop_tctx_refs(current); |
78a78060 | 3003 | |
0f212204 | 3004 | /* |
a1bb3cd5 PB |
3005 | * If we've seen completions, retry without waiting. This |
3006 | * avoids a race where a completion comes in before we did | |
3007 | * prepare_to_wait(). | |
0f212204 | 3008 | */ |
3dd0c97a | 3009 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 3010 | schedule(); |
f57555ed | 3011 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 3012 | } while (1); |
de7f1d9e | 3013 | |
8452d4a6 | 3014 | io_uring_clean_tctx(tctx); |
3dd0c97a | 3015 | if (cancel_all) { |
3cc7fdb9 PB |
3016 | /* |
3017 | * We shouldn't run task_works after cancel, so just leave | |
3018 | * ->in_idle set for normal exit. | |
3019 | */ | |
3020 | atomic_dec(&tctx->in_idle); | |
3f48cf18 PB |
3021 | /* for exec all current's requests should be gone, kill tctx */ |
3022 | __io_uring_free(current); | |
3023 | } | |
44e728b8 PB |
3024 | } |
3025 | ||
f552a27a | 3026 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 3027 | { |
f552a27a | 3028 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
3029 | } |
3030 | ||
6c5c240e RP |
3031 | static void *io_uring_validate_mmap_request(struct file *file, |
3032 | loff_t pgoff, size_t sz) | |
2b188cc1 | 3033 | { |
2b188cc1 | 3034 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 3035 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
3036 | struct page *page; |
3037 | void *ptr; | |
3038 | ||
3039 | switch (offset) { | |
3040 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
3041 | case IORING_OFF_CQ_RING: |
3042 | ptr = ctx->rings; | |
2b188cc1 JA |
3043 | break; |
3044 | case IORING_OFF_SQES: | |
3045 | ptr = ctx->sq_sqes; | |
3046 | break; | |
2b188cc1 | 3047 | default: |
6c5c240e | 3048 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
3049 | } |
3050 | ||
3051 | page = virt_to_head_page(ptr); | |
a50b854e | 3052 | if (sz > page_size(page)) |
6c5c240e RP |
3053 | return ERR_PTR(-EINVAL); |
3054 | ||
3055 | return ptr; | |
3056 | } | |
3057 | ||
3058 | #ifdef CONFIG_MMU | |
3059 | ||
c072481d | 3060 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e RP |
3061 | { |
3062 | size_t sz = vma->vm_end - vma->vm_start; | |
3063 | unsigned long pfn; | |
3064 | void *ptr; | |
3065 | ||
3066 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
3067 | if (IS_ERR(ptr)) | |
3068 | return PTR_ERR(ptr); | |
2b188cc1 JA |
3069 | |
3070 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3071 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3072 | } | |
3073 | ||
6c5c240e RP |
3074 | #else /* !CONFIG_MMU */ |
3075 | ||
3076 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3077 | { | |
3078 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
3079 | } | |
3080 | ||
3081 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
3082 | { | |
3083 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
3084 | } | |
3085 | ||
3086 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
3087 | unsigned long addr, unsigned long len, | |
3088 | unsigned long pgoff, unsigned long flags) | |
3089 | { | |
3090 | void *ptr; | |
3091 | ||
3092 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
3093 | if (IS_ERR(ptr)) | |
3094 | return PTR_ERR(ptr); | |
3095 | ||
3096 | return (unsigned long) ptr; | |
3097 | } | |
3098 | ||
3099 | #endif /* !CONFIG_MMU */ | |
3100 | ||
f81440d3 PB |
3101 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
3102 | { | |
3103 | if (flags & IORING_ENTER_EXT_ARG) { | |
3104 | struct io_uring_getevents_arg arg; | |
3105 | ||
3106 | if (argsz != sizeof(arg)) | |
3107 | return -EINVAL; | |
3108 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3109 | return -EFAULT; | |
3110 | } | |
3111 | return 0; | |
3112 | } | |
3113 | ||
c73ebb68 HX |
3114 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
3115 | struct __kernel_timespec __user **ts, | |
3116 | const sigset_t __user **sig) | |
3117 | { | |
3118 | struct io_uring_getevents_arg arg; | |
3119 | ||
3120 | /* | |
3121 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
3122 | * is just a pointer to the sigset_t. | |
3123 | */ | |
3124 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
3125 | *sig = (const sigset_t __user *) argp; | |
3126 | *ts = NULL; | |
3127 | return 0; | |
3128 | } | |
3129 | ||
3130 | /* | |
3131 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
3132 | * timespec and sigset_t pointers if good. | |
3133 | */ | |
3134 | if (*argsz != sizeof(arg)) | |
3135 | return -EINVAL; | |
3136 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3137 | return -EFAULT; | |
d2347b96 DY |
3138 | if (arg.pad) |
3139 | return -EINVAL; | |
c73ebb68 HX |
3140 | *sig = u64_to_user_ptr(arg.sigmask); |
3141 | *argsz = arg.sigmask_sz; | |
3142 | *ts = u64_to_user_ptr(arg.ts); | |
3143 | return 0; | |
3144 | } | |
3145 | ||
2b188cc1 | 3146 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
3147 | u32, min_complete, u32, flags, const void __user *, argp, |
3148 | size_t, argsz) | |
2b188cc1 JA |
3149 | { |
3150 | struct io_ring_ctx *ctx; | |
2b188cc1 | 3151 | struct fd f; |
33f993da | 3152 | long ret; |
2b188cc1 | 3153 | |
33f993da | 3154 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
3155 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
3156 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
3157 | return -EINVAL; |
3158 | ||
e7a6c00d JA |
3159 | /* |
3160 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3161 | * need only dereference our task private array to find it. | |
3162 | */ | |
3163 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3164 | struct io_uring_task *tctx = current->io_uring; | |
3165 | ||
3273c440 | 3166 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
e7a6c00d JA |
3167 | return -EINVAL; |
3168 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
3169 | f.file = tctx->registered_rings[fd]; | |
4329490a | 3170 | f.flags = 0; |
3273c440 PB |
3171 | if (unlikely(!f.file)) |
3172 | return -EBADF; | |
e7a6c00d JA |
3173 | } else { |
3174 | f = fdget(fd); | |
3273c440 PB |
3175 | if (unlikely(!f.file)) |
3176 | return -EBADF; | |
3177 | ret = -EOPNOTSUPP; | |
3178 | if (unlikely(!io_is_uring_fops(f.file))) | |
fbb8bb02 | 3179 | goto out; |
e7a6c00d | 3180 | } |
2b188cc1 | 3181 | |
2b188cc1 | 3182 | ctx = f.file->private_data; |
7e84e1c7 | 3183 | ret = -EBADFD; |
33f993da | 3184 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3185 | goto out; |
3186 | ||
6c271ce2 JA |
3187 | /* |
3188 | * For SQ polling, the thread will do all submissions and completions. | |
3189 | * Just return the requested submit count, and wake the thread if | |
3190 | * we were asked to. | |
3191 | */ | |
b2a9eada | 3192 | ret = 0; |
6c271ce2 | 3193 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3194 | io_cqring_overflow_flush(ctx); |
89448c47 | 3195 | |
21f96522 JA |
3196 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3197 | ret = -EOWNERDEAD; | |
04147488 | 3198 | goto out; |
21f96522 | 3199 | } |
6c271ce2 | 3200 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3201 | wake_up(&ctx->sq_data->wait); |
d9d05217 PB |
3202 | if (flags & IORING_ENTER_SQ_WAIT) { |
3203 | ret = io_sqpoll_wait_sq(ctx); | |
3204 | if (ret) | |
3205 | goto out; | |
3206 | } | |
3e813c90 | 3207 | ret = to_submit; |
b2a9eada | 3208 | } else if (to_submit) { |
eef51daa | 3209 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3210 | if (unlikely(ret)) |
3211 | goto out; | |
7c504e65 | 3212 | |
2b188cc1 | 3213 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3214 | ret = io_submit_sqes(ctx, to_submit); |
3215 | if (ret != to_submit) { | |
d487b43c | 3216 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3217 | goto out; |
d487b43c | 3218 | } |
44f87745 PB |
3219 | if (flags & IORING_ENTER_GETEVENTS) { |
3220 | if (ctx->syscall_iopoll) | |
3221 | goto iopoll_locked; | |
3222 | /* | |
3223 | * Ignore errors, we'll soon call io_cqring_wait() and | |
3224 | * it should handle ownership problems if any. | |
3225 | */ | |
3226 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3227 | (void)io_run_local_work_locked(ctx); | |
3228 | } | |
d487b43c | 3229 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 | 3230 | } |
c0e0d6ba | 3231 | |
2b188cc1 | 3232 | if (flags & IORING_ENTER_GETEVENTS) { |
3e813c90 | 3233 | int ret2; |
c0e0d6ba | 3234 | |
773697b6 | 3235 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3236 | /* |
3237 | * We disallow the app entering submit/complete with | |
3238 | * polling, but we still need to lock the ring to | |
3239 | * prevent racing with polled issue that got punted to | |
3240 | * a workqueue. | |
3241 | */ | |
3242 | mutex_lock(&ctx->uring_lock); | |
3243 | iopoll_locked: | |
3e813c90 DY |
3244 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3245 | if (likely(!ret2)) { | |
3246 | min_complete = min(min_complete, | |
3247 | ctx->cq_entries); | |
3248 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3249 | } |
3250 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3251 | } else { |
f81440d3 PB |
3252 | const sigset_t __user *sig; |
3253 | struct __kernel_timespec __user *ts; | |
3254 | ||
3e813c90 DY |
3255 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3256 | if (likely(!ret2)) { | |
3257 | min_complete = min(min_complete, | |
3258 | ctx->cq_entries); | |
3259 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3260 | argsz, ts); | |
3261 | } | |
def596e9 | 3262 | } |
c73ebb68 | 3263 | |
155bc950 | 3264 | if (!ret) { |
3e813c90 | 3265 | ret = ret2; |
2b188cc1 | 3266 | |
155bc950 DY |
3267 | /* |
3268 | * EBADR indicates that one or more CQE were dropped. | |
3269 | * Once the user has been informed we can clear the bit | |
3270 | * as they are obviously ok with those drops. | |
3271 | */ | |
3272 | if (unlikely(ret2 == -EBADR)) | |
3273 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3274 | &ctx->check_cq); | |
def596e9 | 3275 | } |
2b188cc1 | 3276 | } |
7c504e65 | 3277 | out: |
4329490a | 3278 | fdput(f); |
3e813c90 | 3279 | return ret; |
2b188cc1 JA |
3280 | } |
3281 | ||
3282 | static const struct file_operations io_uring_fops = { | |
3283 | .release = io_uring_release, | |
3284 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3285 | #ifndef CONFIG_MMU |
3286 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3287 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
3288 | #endif | |
2b188cc1 | 3289 | .poll = io_uring_poll, |
bebdb65e | 3290 | #ifdef CONFIG_PROC_FS |
87ce955b | 3291 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3292 | #endif |
2b188cc1 JA |
3293 | }; |
3294 | ||
92ac8bea JA |
3295 | bool io_is_uring_fops(struct file *file) |
3296 | { | |
3297 | return file->f_op == &io_uring_fops; | |
3298 | } | |
3299 | ||
c072481d PB |
3300 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3301 | struct io_uring_params *p) | |
2b188cc1 | 3302 | { |
75b28aff HV |
3303 | struct io_rings *rings; |
3304 | size_t size, sq_array_offset; | |
2b188cc1 | 3305 | |
bd740481 JA |
3306 | /* make sure these are sane, as we already accounted them */ |
3307 | ctx->sq_entries = p->sq_entries; | |
3308 | ctx->cq_entries = p->cq_entries; | |
3309 | ||
baf9cb64 | 3310 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3311 | if (size == SIZE_MAX) |
3312 | return -EOVERFLOW; | |
3313 | ||
3314 | rings = io_mem_alloc(size); | |
3315 | if (!rings) | |
2b188cc1 JA |
3316 | return -ENOMEM; |
3317 | ||
75b28aff HV |
3318 | ctx->rings = rings; |
3319 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
3320 | rings->sq_ring_mask = p->sq_entries - 1; | |
3321 | rings->cq_ring_mask = p->cq_entries - 1; | |
3322 | rings->sq_ring_entries = p->sq_entries; | |
3323 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3324 | |
ebdeb7c0 JA |
3325 | if (p->flags & IORING_SETUP_SQE128) |
3326 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3327 | else | |
3328 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
3329 | if (size == SIZE_MAX) { |
3330 | io_mem_free(ctx->rings); | |
3331 | ctx->rings = NULL; | |
2b188cc1 | 3332 | return -EOVERFLOW; |
eb065d30 | 3333 | } |
2b188cc1 JA |
3334 | |
3335 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
3336 | if (!ctx->sq_sqes) { |
3337 | io_mem_free(ctx->rings); | |
3338 | ctx->rings = NULL; | |
2b188cc1 | 3339 | return -ENOMEM; |
eb065d30 | 3340 | } |
2b188cc1 | 3341 | |
2b188cc1 JA |
3342 | return 0; |
3343 | } | |
3344 | ||
9faadcc8 PB |
3345 | static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) |
3346 | { | |
3347 | int ret, fd; | |
3348 | ||
3349 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3350 | if (fd < 0) | |
3351 | return fd; | |
3352 | ||
97c96e9f | 3353 | ret = __io_uring_add_tctx_node(ctx); |
9faadcc8 PB |
3354 | if (ret) { |
3355 | put_unused_fd(fd); | |
3356 | return ret; | |
3357 | } | |
3358 | fd_install(fd, file); | |
3359 | return fd; | |
3360 | } | |
3361 | ||
2b188cc1 JA |
3362 | /* |
3363 | * Allocate an anonymous fd, this is what constitutes the application | |
3364 | * visible backing of an io_uring instance. The application mmaps this | |
3365 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3366 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3367 | */ | |
9faadcc8 | 3368 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
3369 | { |
3370 | struct file *file; | |
9faadcc8 | 3371 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
3372 | int ret; |
3373 | ||
2b188cc1 JA |
3374 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
3375 | &ctx->ring_sock); | |
3376 | if (ret) | |
9faadcc8 | 3377 | return ERR_PTR(ret); |
2b188cc1 JA |
3378 | #endif |
3379 | ||
91a9ab7c PM |
3380 | file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, |
3381 | O_RDWR | O_CLOEXEC, NULL); | |
2b188cc1 | 3382 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
3383 | if (IS_ERR(file)) { |
3384 | sock_release(ctx->ring_sock); | |
3385 | ctx->ring_sock = NULL; | |
3386 | } else { | |
3387 | ctx->ring_sock->file = file; | |
0f212204 | 3388 | } |
2b188cc1 | 3389 | #endif |
9faadcc8 | 3390 | return file; |
2b188cc1 JA |
3391 | } |
3392 | ||
c072481d PB |
3393 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3394 | struct io_uring_params __user *params) | |
2b188cc1 | 3395 | { |
2b188cc1 | 3396 | struct io_ring_ctx *ctx; |
9faadcc8 | 3397 | struct file *file; |
2b188cc1 JA |
3398 | int ret; |
3399 | ||
8110c1a6 | 3400 | if (!entries) |
2b188cc1 | 3401 | return -EINVAL; |
8110c1a6 JA |
3402 | if (entries > IORING_MAX_ENTRIES) { |
3403 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3404 | return -EINVAL; | |
3405 | entries = IORING_MAX_ENTRIES; | |
3406 | } | |
2b188cc1 JA |
3407 | |
3408 | /* | |
3409 | * Use twice as many entries for the CQ ring. It's possible for the | |
3410 | * application to drive a higher depth than the size of the SQ ring, | |
3411 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3412 | * some flexibility in overcommitting a bit. If the application has |
3413 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3414 | * of CQ ring entries manually. | |
2b188cc1 JA |
3415 | */ |
3416 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3417 | if (p->flags & IORING_SETUP_CQSIZE) { |
3418 | /* | |
3419 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3420 | * to a power-of-two, if it isn't already. We do NOT impose | |
3421 | * any cq vs sq ring sizing. | |
3422 | */ | |
eb2667b3 | 3423 | if (!p->cq_entries) |
33a107f0 | 3424 | return -EINVAL; |
8110c1a6 JA |
3425 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3426 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3427 | return -EINVAL; | |
3428 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3429 | } | |
eb2667b3 JQ |
3430 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3431 | if (p->cq_entries < p->sq_entries) | |
3432 | return -EINVAL; | |
33a107f0 JA |
3433 | } else { |
3434 | p->cq_entries = 2 * p->sq_entries; | |
3435 | } | |
2b188cc1 | 3436 | |
2b188cc1 | 3437 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3438 | if (!ctx) |
2b188cc1 | 3439 | return -ENOMEM; |
773697b6 PB |
3440 | |
3441 | /* | |
3442 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3443 | * space applications don't need to do io completion events | |
3444 | * polling again, they can rely on io_sq_thread to do polling | |
3445 | * work, which can reduce cpu usage and uring_lock contention. | |
3446 | */ | |
3447 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3448 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3449 | ctx->syscall_iopoll = 1; | |
3450 | ||
2b188cc1 | 3451 | ctx->compat = in_compat_syscall(); |
62e398be JA |
3452 | if (!capable(CAP_IPC_LOCK)) |
3453 | ctx->user = get_uid(current_user()); | |
2aede0e4 | 3454 | |
9f010507 | 3455 | /* |
e1169f06 JA |
3456 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3457 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3458 | */ |
e1169f06 JA |
3459 | ret = -EINVAL; |
3460 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3461 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 | 3462 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
c0e0d6ba DY |
3463 | IORING_SETUP_TASKRUN_FLAG | |
3464 | IORING_SETUP_DEFER_TASKRUN)) | |
e1169f06 | 3465 | goto err; |
9f010507 | 3466 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3467 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3468 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3469 | } else { | |
c0e0d6ba DY |
3470 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && |
3471 | !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) | |
ef060ea9 | 3472 | goto err; |
9f010507 | 3473 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3474 | } |
9f010507 | 3475 | |
c0e0d6ba DY |
3476 | /* |
3477 | * For DEFER_TASKRUN we require the completion task to be the same as the | |
3478 | * submission task. This implies that there is only one submitter, so enforce | |
3479 | * that. | |
3480 | */ | |
3481 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && | |
3482 | !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { | |
3483 | goto err; | |
3484 | } | |
3485 | ||
2aede0e4 JA |
3486 | /* |
3487 | * This is just grabbed for accounting purposes. When a process exits, | |
3488 | * the mm is exited and dropped before the files, hence we need to hang | |
3489 | * on to this mm purely for the purposes of being able to unaccount | |
3490 | * memory (locked/pinned vm). It's not used for anything else. | |
3491 | */ | |
6b7898eb | 3492 | mmgrab(current->mm); |
2aede0e4 | 3493 | ctx->mm_account = current->mm; |
6b7898eb | 3494 | |
2b188cc1 JA |
3495 | ret = io_allocate_scq_urings(ctx, p); |
3496 | if (ret) | |
3497 | goto err; | |
3498 | ||
7e84e1c7 | 3499 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
3500 | if (ret) |
3501 | goto err; | |
eae071c9 | 3502 | /* always set a rsrc node */ |
47b228ce PB |
3503 | ret = io_rsrc_node_switch_start(ctx); |
3504 | if (ret) | |
3505 | goto err; | |
eae071c9 | 3506 | io_rsrc_node_switch(ctx, NULL); |
2b188cc1 | 3507 | |
2b188cc1 | 3508 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
3509 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3510 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3511 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3512 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3513 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3514 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
3515 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
3516 | |
3517 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
3518 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3519 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3520 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3521 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3522 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3523 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 3524 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 3525 | |
7f13657d XW |
3526 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
3527 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 3528 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 3529 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 3530 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e JA |
3531 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
3532 | IORING_FEAT_LINKED_FILE; | |
7f13657d XW |
3533 | |
3534 | if (copy_to_user(params, p, sizeof(*p))) { | |
3535 | ret = -EFAULT; | |
3536 | goto err; | |
3537 | } | |
d1719f70 | 3538 | |
7cae596b DY |
3539 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER |
3540 | && !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3541 | ctx->submitter_task = get_task_struct(current); | |
3542 | ||
9faadcc8 PB |
3543 | file = io_uring_get_file(ctx); |
3544 | if (IS_ERR(file)) { | |
3545 | ret = PTR_ERR(file); | |
3546 | goto err; | |
3547 | } | |
3548 | ||
044c1ab3 JA |
3549 | /* |
3550 | * Install ring fd as the very last thing, so we don't risk someone | |
3551 | * having closed it before we finish setup | |
3552 | */ | |
9faadcc8 PB |
3553 | ret = io_uring_install_fd(ctx, file); |
3554 | if (ret < 0) { | |
3555 | /* fput will clean it up */ | |
3556 | fput(file); | |
3557 | return ret; | |
3558 | } | |
044c1ab3 | 3559 | |
c826bd7a | 3560 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
3561 | return ret; |
3562 | err: | |
3563 | io_ring_ctx_wait_and_kill(ctx); | |
3564 | return ret; | |
3565 | } | |
3566 | ||
3567 | /* | |
3568 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3569 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3570 | * params structure passed in. | |
3571 | */ | |
3572 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3573 | { | |
3574 | struct io_uring_params p; | |
2b188cc1 JA |
3575 | int i; |
3576 | ||
3577 | if (copy_from_user(&p, params, sizeof(p))) | |
3578 | return -EFAULT; | |
3579 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3580 | if (p.resv[i]) | |
3581 | return -EINVAL; | |
3582 | } | |
3583 | ||
6c271ce2 | 3584 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 3585 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 3586 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 3587 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 3588 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 | 3589 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
c0e0d6ba | 3590 | IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN)) |
2b188cc1 JA |
3591 | return -EINVAL; |
3592 | ||
ef060ea9 | 3593 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
3594 | } |
3595 | ||
3596 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3597 | struct io_uring_params __user *, params) | |
3598 | { | |
3599 | return io_uring_setup(entries, params); | |
3600 | } | |
3601 | ||
c072481d PB |
3602 | static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, |
3603 | unsigned nr_args) | |
66f4af93 JA |
3604 | { |
3605 | struct io_uring_probe *p; | |
3606 | size_t size; | |
3607 | int i, ret; | |
3608 | ||
3609 | size = struct_size(p, ops, nr_args); | |
3610 | if (size == SIZE_MAX) | |
3611 | return -EOVERFLOW; | |
3612 | p = kzalloc(size, GFP_KERNEL); | |
3613 | if (!p) | |
3614 | return -ENOMEM; | |
3615 | ||
3616 | ret = -EFAULT; | |
3617 | if (copy_from_user(p, arg, size)) | |
3618 | goto out; | |
3619 | ret = -EINVAL; | |
3620 | if (memchr_inv(p, 0, size)) | |
3621 | goto out; | |
3622 | ||
3623 | p->last_op = IORING_OP_LAST - 1; | |
3624 | if (nr_args > IORING_OP_LAST) | |
3625 | nr_args = IORING_OP_LAST; | |
3626 | ||
3627 | for (i = 0; i < nr_args; i++) { | |
3628 | p->ops[i].op = i; | |
3629 | if (!io_op_defs[i].not_supported) | |
3630 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
3631 | } | |
3632 | p->ops_len = i; | |
3633 | ||
3634 | ret = 0; | |
3635 | if (copy_to_user(arg, p, size)) | |
3636 | ret = -EFAULT; | |
3637 | out: | |
3638 | kfree(p); | |
3639 | return ret; | |
3640 | } | |
3641 | ||
071698e1 JA |
3642 | static int io_register_personality(struct io_ring_ctx *ctx) |
3643 | { | |
4379bf8b | 3644 | const struct cred *creds; |
61cf9370 | 3645 | u32 id; |
1e6fa521 | 3646 | int ret; |
071698e1 | 3647 | |
4379bf8b | 3648 | creds = get_current_cred(); |
1e6fa521 | 3649 | |
61cf9370 MWO |
3650 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
3651 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
3652 | if (ret < 0) { |
3653 | put_cred(creds); | |
3654 | return ret; | |
3655 | } | |
3656 | return id; | |
071698e1 JA |
3657 | } |
3658 | ||
c072481d PB |
3659 | static __cold int io_register_restrictions(struct io_ring_ctx *ctx, |
3660 | void __user *arg, unsigned int nr_args) | |
21b55dbc SG |
3661 | { |
3662 | struct io_uring_restriction *res; | |
3663 | size_t size; | |
3664 | int i, ret; | |
3665 | ||
7e84e1c7 SG |
3666 | /* Restrictions allowed only if rings started disabled */ |
3667 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3668 | return -EBADFD; | |
3669 | ||
21b55dbc | 3670 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 3671 | if (ctx->restrictions.registered) |
21b55dbc SG |
3672 | return -EBUSY; |
3673 | ||
3674 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
3675 | return -EINVAL; | |
3676 | ||
3677 | size = array_size(nr_args, sizeof(*res)); | |
3678 | if (size == SIZE_MAX) | |
3679 | return -EOVERFLOW; | |
3680 | ||
3681 | res = memdup_user(arg, size); | |
3682 | if (IS_ERR(res)) | |
3683 | return PTR_ERR(res); | |
3684 | ||
3685 | ret = 0; | |
3686 | ||
3687 | for (i = 0; i < nr_args; i++) { | |
3688 | switch (res[i].opcode) { | |
3689 | case IORING_RESTRICTION_REGISTER_OP: | |
3690 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
3691 | ret = -EINVAL; | |
3692 | goto out; | |
3693 | } | |
3694 | ||
3695 | __set_bit(res[i].register_op, | |
3696 | ctx->restrictions.register_op); | |
3697 | break; | |
3698 | case IORING_RESTRICTION_SQE_OP: | |
3699 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
3700 | ret = -EINVAL; | |
3701 | goto out; | |
3702 | } | |
3703 | ||
3704 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
3705 | break; | |
3706 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
3707 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
3708 | break; | |
3709 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
3710 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
3711 | break; | |
3712 | default: | |
3713 | ret = -EINVAL; | |
3714 | goto out; | |
3715 | } | |
3716 | } | |
3717 | ||
3718 | out: | |
3719 | /* Reset all restrictions if an error happened */ | |
3720 | if (ret != 0) | |
3721 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
3722 | else | |
7e84e1c7 | 3723 | ctx->restrictions.registered = true; |
21b55dbc SG |
3724 | |
3725 | kfree(res); | |
3726 | return ret; | |
3727 | } | |
3728 | ||
7e84e1c7 SG |
3729 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
3730 | { | |
3731 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3732 | return -EBADFD; | |
3733 | ||
7cae596b DY |
3734 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) |
3735 | ctx->submitter_task = get_task_struct(current); | |
3736 | ||
7e84e1c7 SG |
3737 | if (ctx->restrictions.registered) |
3738 | ctx->restricted = 1; | |
3739 | ||
0298ef96 PB |
3740 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
3741 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
3742 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
3743 | return 0; |
3744 | } | |
3745 | ||
c072481d PB |
3746 | static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, |
3747 | void __user *arg, unsigned len) | |
fe76421d JA |
3748 | { |
3749 | struct io_uring_task *tctx = current->io_uring; | |
3750 | cpumask_var_t new_mask; | |
3751 | int ret; | |
3752 | ||
3753 | if (!tctx || !tctx->io_wq) | |
3754 | return -EINVAL; | |
3755 | ||
3756 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
3757 | return -ENOMEM; | |
3758 | ||
3759 | cpumask_clear(new_mask); | |
3760 | if (len > cpumask_size()) | |
3761 | len = cpumask_size(); | |
3762 | ||
0f5e4b83 ES |
3763 | if (in_compat_syscall()) { |
3764 | ret = compat_get_bitmap(cpumask_bits(new_mask), | |
3765 | (const compat_ulong_t __user *)arg, | |
3766 | len * 8 /* CHAR_BIT */); | |
3767 | } else { | |
3768 | ret = copy_from_user(new_mask, arg, len); | |
3769 | } | |
3770 | ||
3771 | if (ret) { | |
fe76421d JA |
3772 | free_cpumask_var(new_mask); |
3773 | return -EFAULT; | |
3774 | } | |
3775 | ||
3776 | ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); | |
3777 | free_cpumask_var(new_mask); | |
3778 | return ret; | |
3779 | } | |
3780 | ||
c072481d | 3781 | static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) |
fe76421d JA |
3782 | { |
3783 | struct io_uring_task *tctx = current->io_uring; | |
3784 | ||
3785 | if (!tctx || !tctx->io_wq) | |
3786 | return -EINVAL; | |
3787 | ||
3788 | return io_wq_cpu_affinity(tctx->io_wq, NULL); | |
3789 | } | |
3790 | ||
c072481d PB |
3791 | static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, |
3792 | void __user *arg) | |
b22fa62a | 3793 | __must_hold(&ctx->uring_lock) |
2e480058 | 3794 | { |
b22fa62a | 3795 | struct io_tctx_node *node; |
fa84693b JA |
3796 | struct io_uring_task *tctx = NULL; |
3797 | struct io_sq_data *sqd = NULL; | |
2e480058 JA |
3798 | __u32 new_count[2]; |
3799 | int i, ret; | |
3800 | ||
2e480058 JA |
3801 | if (copy_from_user(new_count, arg, sizeof(new_count))) |
3802 | return -EFAULT; | |
3803 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
3804 | if (new_count[i] > INT_MAX) | |
3805 | return -EINVAL; | |
3806 | ||
fa84693b JA |
3807 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3808 | sqd = ctx->sq_data; | |
3809 | if (sqd) { | |
009ad9f0 JA |
3810 | /* |
3811 | * Observe the correct sqd->lock -> ctx->uring_lock | |
3812 | * ordering. Fine to drop uring_lock here, we hold | |
3813 | * a ref to the ctx. | |
3814 | */ | |
41d3a6bd | 3815 | refcount_inc(&sqd->refs); |
009ad9f0 | 3816 | mutex_unlock(&ctx->uring_lock); |
fa84693b | 3817 | mutex_lock(&sqd->lock); |
009ad9f0 | 3818 | mutex_lock(&ctx->uring_lock); |
41d3a6bd JA |
3819 | if (sqd->thread) |
3820 | tctx = sqd->thread->io_uring; | |
fa84693b JA |
3821 | } |
3822 | } else { | |
3823 | tctx = current->io_uring; | |
3824 | } | |
3825 | ||
e139a1ec | 3826 | BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); |
fa84693b | 3827 | |
bad119b9 PB |
3828 | for (i = 0; i < ARRAY_SIZE(new_count); i++) |
3829 | if (new_count[i]) | |
3830 | ctx->iowq_limits[i] = new_count[i]; | |
e139a1ec PB |
3831 | ctx->iowq_limits_set = true; |
3832 | ||
e139a1ec PB |
3833 | if (tctx && tctx->io_wq) { |
3834 | ret = io_wq_max_workers(tctx->io_wq, new_count); | |
3835 | if (ret) | |
3836 | goto err; | |
3837 | } else { | |
3838 | memset(new_count, 0, sizeof(new_count)); | |
3839 | } | |
fa84693b | 3840 | |
41d3a6bd | 3841 | if (sqd) { |
fa84693b | 3842 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
3843 | io_put_sq_data(sqd); |
3844 | } | |
2e480058 JA |
3845 | |
3846 | if (copy_to_user(arg, new_count, sizeof(new_count))) | |
3847 | return -EFAULT; | |
3848 | ||
b22fa62a PB |
3849 | /* that's it for SQPOLL, only the SQPOLL task creates requests */ |
3850 | if (sqd) | |
3851 | return 0; | |
3852 | ||
3853 | /* now propagate the restriction to all registered users */ | |
3854 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3855 | struct io_uring_task *tctx = node->task->io_uring; | |
3856 | ||
3857 | if (WARN_ON_ONCE(!tctx->io_wq)) | |
3858 | continue; | |
3859 | ||
3860 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
3861 | new_count[i] = ctx->iowq_limits[i]; | |
3862 | /* ignore errors, it always returns zero anyway */ | |
3863 | (void)io_wq_max_workers(tctx->io_wq, new_count); | |
3864 | } | |
2e480058 | 3865 | return 0; |
fa84693b | 3866 | err: |
41d3a6bd | 3867 | if (sqd) { |
fa84693b | 3868 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
3869 | io_put_sq_data(sqd); |
3870 | } | |
fa84693b | 3871 | return ret; |
2e480058 JA |
3872 | } |
3873 | ||
edafccee JA |
3874 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
3875 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
3876 | __releases(ctx->uring_lock) |
3877 | __acquires(ctx->uring_lock) | |
edafccee JA |
3878 | { |
3879 | int ret; | |
3880 | ||
35fa71a0 | 3881 | /* |
fbb8bb02 PB |
3882 | * We don't quiesce the refs for register anymore and so it can't be |
3883 | * dying as we're holding a file ref here. | |
35fa71a0 | 3884 | */ |
fbb8bb02 | 3885 | if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs))) |
35fa71a0 JA |
3886 | return -ENXIO; |
3887 | ||
d7cce96c PB |
3888 | if (ctx->submitter_task && ctx->submitter_task != current) |
3889 | return -EEXIST; | |
3890 | ||
75c4021a PB |
3891 | if (ctx->restricted) { |
3892 | if (opcode >= IORING_REGISTER_LAST) | |
3893 | return -EINVAL; | |
3894 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); | |
3895 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
3896 | return -EACCES; | |
3897 | } | |
3898 | ||
edafccee JA |
3899 | switch (opcode) { |
3900 | case IORING_REGISTER_BUFFERS: | |
0184f08e PB |
3901 | ret = -EFAULT; |
3902 | if (!arg) | |
3903 | break; | |
634d00df | 3904 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
3905 | break; |
3906 | case IORING_UNREGISTER_BUFFERS: | |
3907 | ret = -EINVAL; | |
3908 | if (arg || nr_args) | |
3909 | break; | |
0a96bbe4 | 3910 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 3911 | break; |
6b06314c | 3912 | case IORING_REGISTER_FILES: |
a8da73a3 JA |
3913 | ret = -EFAULT; |
3914 | if (!arg) | |
3915 | break; | |
792e3582 | 3916 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
3917 | break; |
3918 | case IORING_UNREGISTER_FILES: | |
3919 | ret = -EINVAL; | |
3920 | if (arg || nr_args) | |
3921 | break; | |
3922 | ret = io_sqe_files_unregister(ctx); | |
3923 | break; | |
c3a31e60 | 3924 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 3925 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 3926 | break; |
9b402849 JA |
3927 | case IORING_REGISTER_EVENTFD: |
3928 | ret = -EINVAL; | |
3929 | if (nr_args != 1) | |
3930 | break; | |
c75312dd UA |
3931 | ret = io_eventfd_register(ctx, arg, 0); |
3932 | break; | |
3933 | case IORING_REGISTER_EVENTFD_ASYNC: | |
3934 | ret = -EINVAL; | |
3935 | if (nr_args != 1) | |
f2842ab5 | 3936 | break; |
c75312dd | 3937 | ret = io_eventfd_register(ctx, arg, 1); |
9b402849 JA |
3938 | break; |
3939 | case IORING_UNREGISTER_EVENTFD: | |
3940 | ret = -EINVAL; | |
3941 | if (arg || nr_args) | |
3942 | break; | |
3943 | ret = io_eventfd_unregister(ctx); | |
3944 | break; | |
66f4af93 JA |
3945 | case IORING_REGISTER_PROBE: |
3946 | ret = -EINVAL; | |
3947 | if (!arg || nr_args > 256) | |
3948 | break; | |
3949 | ret = io_probe(ctx, arg, nr_args); | |
3950 | break; | |
071698e1 JA |
3951 | case IORING_REGISTER_PERSONALITY: |
3952 | ret = -EINVAL; | |
3953 | if (arg || nr_args) | |
3954 | break; | |
3955 | ret = io_register_personality(ctx); | |
3956 | break; | |
3957 | case IORING_UNREGISTER_PERSONALITY: | |
3958 | ret = -EINVAL; | |
3959 | if (arg) | |
3960 | break; | |
3961 | ret = io_unregister_personality(ctx, nr_args); | |
3962 | break; | |
7e84e1c7 SG |
3963 | case IORING_REGISTER_ENABLE_RINGS: |
3964 | ret = -EINVAL; | |
3965 | if (arg || nr_args) | |
3966 | break; | |
3967 | ret = io_register_enable_rings(ctx); | |
3968 | break; | |
21b55dbc SG |
3969 | case IORING_REGISTER_RESTRICTIONS: |
3970 | ret = io_register_restrictions(ctx, arg, nr_args); | |
3971 | break; | |
992da01a PB |
3972 | case IORING_REGISTER_FILES2: |
3973 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
3974 | break; | |
3975 | case IORING_REGISTER_FILES_UPDATE2: | |
3976 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
3977 | IORING_RSRC_FILE); | |
3978 | break; | |
3979 | case IORING_REGISTER_BUFFERS2: | |
3980 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 3981 | break; |
992da01a PB |
3982 | case IORING_REGISTER_BUFFERS_UPDATE: |
3983 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
3984 | IORING_RSRC_BUFFER); | |
c3bdad02 | 3985 | break; |
fe76421d JA |
3986 | case IORING_REGISTER_IOWQ_AFF: |
3987 | ret = -EINVAL; | |
3988 | if (!arg || !nr_args) | |
3989 | break; | |
3990 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
3991 | break; | |
3992 | case IORING_UNREGISTER_IOWQ_AFF: | |
3993 | ret = -EINVAL; | |
3994 | if (arg || nr_args) | |
3995 | break; | |
3996 | ret = io_unregister_iowq_aff(ctx); | |
3997 | break; | |
2e480058 JA |
3998 | case IORING_REGISTER_IOWQ_MAX_WORKERS: |
3999 | ret = -EINVAL; | |
4000 | if (!arg || nr_args != 2) | |
4001 | break; | |
4002 | ret = io_register_iowq_max_workers(ctx, arg); | |
4003 | break; | |
e7a6c00d JA |
4004 | case IORING_REGISTER_RING_FDS: |
4005 | ret = io_ringfd_register(ctx, arg, nr_args); | |
4006 | break; | |
4007 | case IORING_UNREGISTER_RING_FDS: | |
4008 | ret = io_ringfd_unregister(ctx, arg, nr_args); | |
4009 | break; | |
c7fb1942 JA |
4010 | case IORING_REGISTER_PBUF_RING: |
4011 | ret = -EINVAL; | |
4012 | if (!arg || nr_args != 1) | |
4013 | break; | |
4014 | ret = io_register_pbuf_ring(ctx, arg); | |
4015 | break; | |
4016 | case IORING_UNREGISTER_PBUF_RING: | |
4017 | ret = -EINVAL; | |
4018 | if (!arg || nr_args != 1) | |
4019 | break; | |
4020 | ret = io_unregister_pbuf_ring(ctx, arg); | |
4021 | break; | |
78a861b9 JA |
4022 | case IORING_REGISTER_SYNC_CANCEL: |
4023 | ret = -EINVAL; | |
4024 | if (!arg || nr_args != 1) | |
4025 | break; | |
4026 | ret = io_sync_cancel(ctx, arg); | |
4027 | break; | |
6e73dffb PB |
4028 | case IORING_REGISTER_FILE_ALLOC_RANGE: |
4029 | ret = -EINVAL; | |
4030 | if (!arg || nr_args) | |
4031 | break; | |
4032 | ret = io_register_file_alloc_range(ctx, arg); | |
4033 | break; | |
edafccee JA |
4034 | default: |
4035 | ret = -EINVAL; | |
4036 | break; | |
4037 | } | |
4038 | ||
edafccee JA |
4039 | return ret; |
4040 | } | |
4041 | ||
4042 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
4043 | void __user *, arg, unsigned int, nr_args) | |
4044 | { | |
4045 | struct io_ring_ctx *ctx; | |
4046 | long ret = -EBADF; | |
4047 | struct fd f; | |
4048 | ||
4049 | f = fdget(fd); | |
4050 | if (!f.file) | |
4051 | return -EBADF; | |
4052 | ||
4053 | ret = -EOPNOTSUPP; | |
e5550a14 | 4054 | if (!io_is_uring_fops(f.file)) |
edafccee JA |
4055 | goto out_fput; |
4056 | ||
4057 | ctx = f.file->private_data; | |
4058 | ||
4059 | mutex_lock(&ctx->uring_lock); | |
4060 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
4061 | mutex_unlock(&ctx->uring_lock); | |
2757be22 | 4062 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); |
edafccee JA |
4063 | out_fput: |
4064 | fdput(f); | |
4065 | return ret; | |
4066 | } | |
4067 | ||
2b188cc1 JA |
4068 | static int __init io_uring_init(void) |
4069 | { | |
9c71d39a | 4070 | #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ |
d7f62e82 | 4071 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ |
9c71d39a | 4072 | BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ |
d7f62e82 SM |
4073 | } while (0) |
4074 | ||
4075 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9c71d39a SM |
4076 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) |
4077 | #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ | |
4078 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) | |
d7f62e82 SM |
4079 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); |
4080 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
4081 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
4082 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
4083 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
4084 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
4085 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9c71d39a SM |
4086 | BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); |
4087 | BUILD_BUG_SQE_ELEM(12, __u32, __pad1); | |
d7f62e82 | 4088 | BUILD_BUG_SQE_ELEM(16, __u64, addr); |
7d67af2c | 4089 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
4090 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
4091 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
4092 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
4093 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
4094 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
4095 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
4096 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
4097 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
4098 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
4099 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
4100 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
4101 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
4102 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
4103 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
4104 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 4105 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
9c71d39a SM |
4106 | BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); |
4107 | BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); | |
4108 | BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); | |
4109 | BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); | |
4110 | BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); | |
d7f62e82 SM |
4111 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
4112 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 4113 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 4114 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 4115 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 4116 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
b48c312b PB |
4117 | BUILD_BUG_SQE_ELEM(44, __u16, addr_len); |
4118 | BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); | |
e9621e2b | 4119 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
9c71d39a SM |
4120 | BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); |
4121 | BUILD_BUG_SQE_ELEM(56, __u64, __pad2); | |
d7f62e82 | 4122 | |
b0d658ec PB |
4123 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
4124 | sizeof(struct io_uring_rsrc_update)); | |
4125 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
4126 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
4127 | |
4128 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
4129 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
4130 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
4131 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 4132 | |
b0d658ec PB |
4133 | /* should fit into one byte */ |
4134 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
4135 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
4136 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 4137 | |
32c2d33e | 4138 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); |
16340eab | 4139 | |
3a4b89a2 JA |
4140 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
4141 | ||
d9b57aa3 | 4142 | io_uring_optable_init(); |
0702e536 | 4143 | |
91f245d5 JA |
4144 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | |
4145 | SLAB_ACCOUNT); | |
2b188cc1 JA |
4146 | return 0; |
4147 | }; | |
4148 | __initcall(io_uring_init); |