Commit | Line | Data |
---|---|---|
2b188cc1 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Shared application/kernel submission and completion ring pairs, for | |
4 | * supporting fast/efficient IO. | |
5 | * | |
6 | * A note on the read/write ordering memory barriers that are matched between | |
1e84b97b SB |
7 | * the application and kernel side. |
8 | * | |
9 | * After the application reads the CQ ring tail, it must use an | |
10 | * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses | |
11 | * before writing the tail (using smp_load_acquire to read the tail will | |
12 | * do). It also needs a smp_mb() before updating CQ head (ordering the | |
13 | * entry load(s) with the head store), pairing with an implicit barrier | |
d068b506 | 14 | * through a control-dependency in io_get_cqe (smp_store_release to |
1e84b97b SB |
15 | * store head will do). Failure to do so could lead to reading invalid |
16 | * CQ entries. | |
17 | * | |
18 | * Likewise, the application must use an appropriate smp_wmb() before | |
19 | * writing the SQ tail (ordering SQ entry stores with the tail store), | |
20 | * which pairs with smp_load_acquire in io_get_sqring (smp_store_release | |
21 | * to store the tail will do). And it needs a barrier ordering the SQ | |
22 | * head load before writing new SQ entries (smp_load_acquire to read | |
23 | * head will do). | |
24 | * | |
25 | * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application | |
26 | * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after* | |
27 | * updating the SQ tail; a full memory barrier smp_mb() is needed | |
28 | * between. | |
2b188cc1 JA |
29 | * |
30 | * Also see the examples in the liburing library: | |
31 | * | |
32 | * git://git.kernel.dk/liburing | |
33 | * | |
34 | * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens | |
35 | * from data shared between the kernel and application. This is done both | |
36 | * for ordering purposes, but also to ensure that once a value is loaded from | |
37 | * data that the application could potentially modify, it remains stable. | |
38 | * | |
39 | * Copyright (C) 2018-2019 Jens Axboe | |
c992fe29 | 40 | * Copyright (c) 2018-2019 Christoph Hellwig |
2b188cc1 JA |
41 | */ |
42 | #include <linux/kernel.h> | |
43 | #include <linux/init.h> | |
44 | #include <linux/errno.h> | |
45 | #include <linux/syscalls.h> | |
52de1fe1 | 46 | #include <net/compat.h> |
2b188cc1 JA |
47 | #include <linux/refcount.h> |
48 | #include <linux/uio.h> | |
6b47ee6e | 49 | #include <linux/bits.h> |
2b188cc1 JA |
50 | |
51 | #include <linux/sched/signal.h> | |
52 | #include <linux/fs.h> | |
53 | #include <linux/file.h> | |
54 | #include <linux/fdtable.h> | |
55 | #include <linux/mm.h> | |
56 | #include <linux/mman.h> | |
2b188cc1 JA |
57 | #include <linux/percpu.h> |
58 | #include <linux/slab.h> | |
edafccee | 59 | #include <linux/bvec.h> |
2b188cc1 JA |
60 | #include <linux/net.h> |
61 | #include <net/sock.h> | |
62 | #include <net/af_unix.h> | |
6b06314c | 63 | #include <net/scm.h> |
2b188cc1 JA |
64 | #include <linux/anon_inodes.h> |
65 | #include <linux/sched/mm.h> | |
66 | #include <linux/uaccess.h> | |
67 | #include <linux/nospec.h> | |
aa4c3967 | 68 | #include <linux/highmem.h> |
15b71abe | 69 | #include <linux/fsnotify.h> |
4840e418 | 70 | #include <linux/fadvise.h> |
b41e9852 | 71 | #include <linux/task_work.h> |
0f212204 | 72 | #include <linux/io_uring.h> |
5bd2182d | 73 | #include <linux/audit.h> |
cdc1404a | 74 | #include <linux/security.h> |
d808459b | 75 | #include <asm/shmparam.h> |
2b188cc1 | 76 | |
c826bd7a DD |
77 | #define CREATE_TRACE_POINTS |
78 | #include <trace/events/io_uring.h> | |
79 | ||
2b188cc1 JA |
80 | #include <uapi/linux/io_uring.h> |
81 | ||
561fb04a | 82 | #include "io-wq.h" |
2b188cc1 | 83 | |
de23077e | 84 | #include "io_uring.h" |
329061d3 | 85 | #include "opdef.h" |
e418bbc9 | 86 | #include "refs.h" |
c9f06aa7 | 87 | #include "tctx.h" |
17437f31 | 88 | #include "sqpoll.h" |
a4ad4f74 | 89 | #include "fdinfo.h" |
3b77495a | 90 | #include "kbuf.h" |
73572984 | 91 | #include "rsrc.h" |
38513c46 | 92 | #include "cancel.h" |
43e0bbbd | 93 | #include "net.h" |
eb42cebb | 94 | #include "notif.h" |
e27f928e | 95 | |
59915143 | 96 | #include "timeout.h" |
329061d3 | 97 | #include "poll.h" |
9b797a37 | 98 | #include "alloc_cache.h" |
5e2a18d9 | 99 | |
5277deaa | 100 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 101 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 | 102 | |
21b55dbc SG |
103 | #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \ |
104 | IORING_REGISTER_LAST + IORING_OP_LAST) | |
2b188cc1 | 105 | |
68fe256a PB |
106 | #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \ |
107 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) | |
108 | ||
5562a8d7 PB |
109 | #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \ |
110 | IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS) | |
68fe256a | 111 | |
c854357b | 112 | #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ |
9cae36a0 JA |
113 | REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \ |
114 | REQ_F_ASYNC_DATA) | |
b16fed66 | 115 | |
a538be5b PB |
116 | #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\ |
117 | IO_REQ_CLEAN_FLAGS) | |
118 | ||
09899b19 PB |
119 | #define IO_TCTX_REFS_CACHE_NR (1U << 10) |
120 | ||
6dd0be1e | 121 | #define IO_COMPL_BATCH 32 |
bf019da7 | 122 | #define IO_REQ_ALLOC_BATCH 8 |
258b29a9 | 123 | |
10988a0a DY |
124 | enum { |
125 | IO_CHECK_CQ_OVERFLOW_BIT, | |
155bc950 | 126 | IO_CHECK_CQ_DROPPED_BIT, |
10988a0a DY |
127 | }; |
128 | ||
21a091b9 DY |
129 | enum { |
130 | IO_EVENTFD_OP_SIGNAL_BIT, | |
131 | IO_EVENTFD_OP_FREE_BIT, | |
132 | }; | |
133 | ||
27dc8338 PB |
134 | struct io_defer_entry { |
135 | struct list_head list; | |
136 | struct io_kiocb *req; | |
9cf7c104 | 137 | u32 seq; |
2b188cc1 JA |
138 | }; |
139 | ||
0756a869 PB |
140 | /* requests with any of those set should undergo io_disarm_next() */ |
141 | #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) | |
da1a08c5 | 142 | #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK) |
0756a869 | 143 | |
affa87db | 144 | static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
9936c7c2 | 145 | struct task_struct *task, |
3dd0c97a | 146 | bool cancel_all); |
1ffc5422 | 147 | |
c7dae4ba | 148 | static void io_dismantle_req(struct io_kiocb *req); |
68fb8979 | 149 | static void io_clean_op(struct io_kiocb *req); |
cbc2e203 | 150 | static void io_queue_sqe(struct io_kiocb *req); |
c0e0d6ba | 151 | static void io_move_task_work_from_local(struct io_ring_ctx *ctx); |
c450178d | 152 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx); |
77e443ab | 153 | static __cold void io_fallback_tw(struct io_uring_task *tctx); |
de0617e4 | 154 | |
c1755c25 | 155 | struct kmem_cache *req_cachep; |
2b188cc1 | 156 | |
2b188cc1 JA |
157 | struct sock *io_uring_get_socket(struct file *file) |
158 | { | |
159 | #if defined(CONFIG_UNIX) | |
cd40cae2 | 160 | if (io_is_uring_fops(file)) { |
2b188cc1 JA |
161 | struct io_ring_ctx *ctx = file->private_data; |
162 | ||
163 | return ctx->ring_sock->sk; | |
164 | } | |
165 | #endif | |
166 | return NULL; | |
167 | } | |
168 | EXPORT_SYMBOL(io_uring_get_socket); | |
169 | ||
c450178d PB |
170 | static inline void io_submit_flush_completions(struct io_ring_ctx *ctx) |
171 | { | |
931147dd DY |
172 | if (!wq_list_empty(&ctx->submit_state.compl_reqs) || |
173 | ctx->submit_state.cqes_count) | |
c450178d PB |
174 | __io_submit_flush_completions(ctx); |
175 | } | |
176 | ||
faf88dde PB |
177 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
178 | { | |
179 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
180 | } | |
181 | ||
0fc8c2ac DY |
182 | static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx) |
183 | { | |
184 | return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head); | |
185 | } | |
186 | ||
9cae36a0 JA |
187 | static bool io_match_linked(struct io_kiocb *head) |
188 | { | |
189 | struct io_kiocb *req; | |
190 | ||
191 | io_for_each_link(req, head) { | |
192 | if (req->flags & REQ_F_INFLIGHT) | |
193 | return true; | |
194 | } | |
195 | return false; | |
6af3f48b PB |
196 | } |
197 | ||
198 | /* | |
199 | * As io_match_task() but protected against racing with linked timeouts. | |
200 | * User must not hold timeout_lock. | |
201 | */ | |
329061d3 JA |
202 | bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, |
203 | bool cancel_all) | |
6af3f48b | 204 | { |
9cae36a0 JA |
205 | bool matched; |
206 | ||
6af3f48b PB |
207 | if (task && head->task != task) |
208 | return false; | |
9cae36a0 JA |
209 | if (cancel_all) |
210 | return true; | |
211 | ||
212 | if (head->flags & REQ_F_LINK_TIMEOUT) { | |
213 | struct io_ring_ctx *ctx = head->ctx; | |
214 | ||
215 | /* protect against races with linked timeouts */ | |
216 | spin_lock_irq(&ctx->timeout_lock); | |
217 | matched = io_match_linked(head); | |
218 | spin_unlock_irq(&ctx->timeout_lock); | |
219 | } else { | |
220 | matched = io_match_linked(head); | |
221 | } | |
222 | return matched; | |
6af3f48b PB |
223 | } |
224 | ||
a8295b98 HX |
225 | static inline void req_fail_link_node(struct io_kiocb *req, int res) |
226 | { | |
227 | req_set_fail(req); | |
97b388d7 | 228 | io_req_set_res(req, res, 0); |
a8295b98 HX |
229 | } |
230 | ||
fa05457a PB |
231 | static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) |
232 | { | |
233 | wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); | |
c1755c25 | 234 | kasan_poison_object_data(req_cachep, req); |
a8295b98 HX |
235 | } |
236 | ||
c072481d | 237 | static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) |
2b188cc1 JA |
238 | { |
239 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
240 | ||
0f158b4c | 241 | complete(&ctx->ref_comp); |
2b188cc1 JA |
242 | } |
243 | ||
c072481d | 244 | static __cold void io_fallback_req_func(struct work_struct *work) |
f56165e6 PB |
245 | { |
246 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
247 | fallback_work.work); | |
248 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
249 | struct io_kiocb *req, *tmp; | |
31f084b7 | 250 | bool locked = true; |
f56165e6 | 251 | |
31f084b7 | 252 | mutex_lock(&ctx->uring_lock); |
3218e5d3 | 253 | llist_for_each_entry_safe(req, tmp, node, io_task_work.node) |
f237c30a | 254 | req->io_task_work.func(req, &locked); |
31f084b7 PB |
255 | if (WARN_ON_ONCE(!locked)) |
256 | return; | |
257 | io_submit_flush_completions(ctx); | |
258 | mutex_unlock(&ctx->uring_lock); | |
f56165e6 PB |
259 | } |
260 | ||
e6f89be6 PB |
261 | static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) |
262 | { | |
263 | unsigned hash_buckets = 1U << bits; | |
264 | size_t hash_size = hash_buckets * sizeof(table->hbs[0]); | |
265 | ||
266 | table->hbs = kmalloc(hash_size, GFP_KERNEL); | |
267 | if (!table->hbs) | |
268 | return -ENOMEM; | |
269 | ||
270 | table->hash_bits = bits; | |
271 | init_hash_table(table, hash_buckets); | |
272 | return 0; | |
273 | } | |
274 | ||
c072481d | 275 | static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
2b188cc1 JA |
276 | { |
277 | struct io_ring_ctx *ctx; | |
9cfc7e94 | 278 | int hash_bits; |
2b188cc1 JA |
279 | |
280 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
281 | if (!ctx) | |
282 | return NULL; | |
283 | ||
9cfc7e94 JA |
284 | xa_init(&ctx->io_bl_xa); |
285 | ||
78076bb6 JA |
286 | /* |
287 | * Use 5 bits less than the max cq entries, that should give us around | |
4a07723f PB |
288 | * 32 entries per hash list if totally full and uniformly spread, but |
289 | * don't keep too many buckets to not overconsume memory. | |
78076bb6 | 290 | */ |
4a07723f PB |
291 | hash_bits = ilog2(p->cq_entries) - 5; |
292 | hash_bits = clamp(hash_bits, 1, 8); | |
e6f89be6 | 293 | if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) |
78076bb6 | 294 | goto err; |
9ca9fb24 PB |
295 | if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) |
296 | goto err; | |
38513c46 | 297 | |
6224843d PB |
298 | ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL); |
299 | if (!ctx->dummy_ubuf) | |
300 | goto err; | |
301 | /* set invalid range, so io_import_fixed() fails meeting it */ | |
302 | ctx->dummy_ubuf->ubuf = -1UL; | |
303 | ||
21482896 | 304 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
48904229 | 305 | 0, GFP_KERNEL)) |
206aefde | 306 | goto err; |
2b188cc1 JA |
307 | |
308 | ctx->flags = p->flags; | |
90554200 | 309 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 310 | INIT_LIST_HEAD(&ctx->sqd_list); |
1d7bb1d5 | 311 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
cc3cec83 | 312 | INIT_LIST_HEAD(&ctx->io_buffers_cache); |
9b797a37 | 313 | io_alloc_cache_init(&ctx->apoll_cache); |
43e0bbbd | 314 | io_alloc_cache_init(&ctx->netmsg_cache); |
0f158b4c | 315 | init_completion(&ctx->ref_comp); |
61cf9370 | 316 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 317 | mutex_init(&ctx->uring_lock); |
311997b3 | 318 | init_waitqueue_head(&ctx->cq_wait); |
7b235dd8 | 319 | init_waitqueue_head(&ctx->poll_wq); |
2b188cc1 | 320 | spin_lock_init(&ctx->completion_lock); |
89850fce | 321 | spin_lock_init(&ctx->timeout_lock); |
5eef4e87 | 322 | INIT_WQ_LIST(&ctx->iopoll_list); |
cc3cec83 JA |
323 | INIT_LIST_HEAD(&ctx->io_buffers_pages); |
324 | INIT_LIST_HEAD(&ctx->io_buffers_comp); | |
de0617e4 | 325 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 326 | INIT_LIST_HEAD(&ctx->timeout_list); |
ef9dd637 | 327 | INIT_LIST_HEAD(&ctx->ltimeout_list); |
d67d2263 BM |
328 | spin_lock_init(&ctx->rsrc_ref_lock); |
329 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); | |
269bbe5f | 330 | INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work); |
d34b1b0b | 331 | init_task_work(&ctx->rsrc_put_tw, io_rsrc_put_tw); |
269bbe5f | 332 | init_llist_head(&ctx->rsrc_put_llist); |
c0e0d6ba | 333 | init_llist_head(&ctx->work_llist); |
13bf43f5 | 334 | INIT_LIST_HEAD(&ctx->tctx_list); |
c2b6c6bc PB |
335 | ctx->submit_state.free_list.next = NULL; |
336 | INIT_WQ_LIST(&ctx->locked_free_list); | |
9011bf9a | 337 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
6f33b0bc | 338 | INIT_WQ_LIST(&ctx->submit_state.compl_reqs); |
2b188cc1 | 339 | return ctx; |
206aefde | 340 | err: |
6224843d | 341 | kfree(ctx->dummy_ubuf); |
e6f89be6 | 342 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 343 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 JA |
344 | kfree(ctx->io_bl); |
345 | xa_destroy(&ctx->io_bl_xa); | |
206aefde JA |
346 | kfree(ctx); |
347 | return NULL; | |
2b188cc1 JA |
348 | } |
349 | ||
8f6ed49a PB |
350 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
351 | { | |
352 | struct io_rings *r = ctx->rings; | |
353 | ||
354 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
355 | ctx->cq_extra--; | |
356 | } | |
357 | ||
9cf7c104 | 358 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 359 | { |
2bc9930e JA |
360 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
361 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 362 | |
8f6ed49a | 363 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 364 | } |
de0617e4 | 365 | |
9d858b21 | 366 | return false; |
de0617e4 JA |
367 | } |
368 | ||
9cae36a0 JA |
369 | static inline void io_req_track_inflight(struct io_kiocb *req) |
370 | { | |
371 | if (!(req->flags & REQ_F_INFLIGHT)) { | |
372 | req->flags |= REQ_F_INFLIGHT; | |
386e4fb6 | 373 | atomic_inc(&req->task->io_uring->inflight_tracked); |
9cae36a0 JA |
374 | } |
375 | } | |
376 | ||
fd08e530 PB |
377 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
378 | { | |
906c6caa PB |
379 | if (WARN_ON_ONCE(!req->link)) |
380 | return NULL; | |
381 | ||
4d13d1a4 PB |
382 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
383 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
384 | |
385 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 386 | io_req_set_refcount(req); |
4d13d1a4 PB |
387 | __io_req_set_refcount(req->link, 2); |
388 | return req->link; | |
fd08e530 PB |
389 | } |
390 | ||
391 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
392 | { | |
4d13d1a4 | 393 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
394 | return NULL; |
395 | return __io_prep_linked_timeout(req); | |
396 | } | |
397 | ||
cb2d344c PB |
398 | static noinline void __io_arm_ltimeout(struct io_kiocb *req) |
399 | { | |
400 | io_queue_linked_timeout(__io_prep_linked_timeout(req)); | |
401 | } | |
402 | ||
403 | static inline void io_arm_ltimeout(struct io_kiocb *req) | |
404 | { | |
405 | if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) | |
406 | __io_arm_ltimeout(req); | |
407 | } | |
408 | ||
1e6fa521 JA |
409 | static void io_prep_async_work(struct io_kiocb *req) |
410 | { | |
a7dd2782 | 411 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
1e6fa521 JA |
412 | struct io_ring_ctx *ctx = req->ctx; |
413 | ||
b8e64b53 PB |
414 | if (!(req->flags & REQ_F_CREDS)) { |
415 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 416 | req->creds = get_current_cred(); |
b8e64b53 | 417 | } |
003e8dcc | 418 | |
e1d675df PB |
419 | req->work.list.next = NULL; |
420 | req->work.flags = 0; | |
8e29da69 | 421 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); |
feaadc4f PB |
422 | if (req->flags & REQ_F_FORCE_ASYNC) |
423 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
424 | ||
f6b543fd JA |
425 | if (req->file && !io_req_ffs_set(req)) |
426 | req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT; | |
427 | ||
1e6fa521 | 428 | if (req->flags & REQ_F_ISREG) { |
d4755e15 JA |
429 | bool should_hash = def->hash_reg_file; |
430 | ||
431 | /* don't serialize this request if the fs doesn't need it */ | |
432 | if (should_hash && (req->file->f_flags & O_DIRECT) && | |
433 | (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE)) | |
434 | should_hash = false; | |
435 | if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL)) | |
1e6fa521 | 436 | io_wq_hash_work(&req->work, file_inode(req->file)); |
4b982bd0 | 437 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
438 | if (def->unbound_nonreg_file) |
439 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
440 | } | |
561fb04a | 441 | } |
cccf0ee8 | 442 | |
cbdcb435 | 443 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 444 | { |
cbdcb435 | 445 | struct io_kiocb *cur; |
54a91f3b | 446 | |
44eff40a PB |
447 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
448 | struct io_ring_ctx *ctx = req->ctx; | |
449 | ||
674ee8e1 | 450 | spin_lock_irq(&ctx->timeout_lock); |
44eff40a PB |
451 | io_for_each_link(cur, req) |
452 | io_prep_async_work(cur); | |
674ee8e1 | 453 | spin_unlock_irq(&ctx->timeout_lock); |
44eff40a PB |
454 | } else { |
455 | io_for_each_link(cur, req) | |
456 | io_prep_async_work(cur); | |
457 | } | |
561fb04a JA |
458 | } |
459 | ||
f3b44f92 | 460 | void io_queue_iowq(struct io_kiocb *req, bool *dont_use) |
561fb04a | 461 | { |
cbdcb435 | 462 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 463 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 464 | |
3bfe6106 JA |
465 | BUG_ON(!tctx); |
466 | BUG_ON(!tctx->io_wq); | |
561fb04a | 467 | |
cbdcb435 PB |
468 | /* init ->work of the whole link before punting */ |
469 | io_prep_async_link(req); | |
991468dc JA |
470 | |
471 | /* | |
472 | * Not expected to happen, but if we do have a bug where this _can_ | |
473 | * happen, catch it here and ensure the request is marked as | |
474 | * canceled. That will make io-wq go through the usual work cancel | |
475 | * procedure rather than attempt to run this request (or create a new | |
476 | * worker for it). | |
477 | */ | |
478 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
479 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
480 | ||
48863ffd | 481 | trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); |
ebf93667 | 482 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
483 | if (link) |
484 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
485 | } |
486 | ||
c072481d | 487 | static __cold void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 488 | { |
441b8a78 | 489 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
490 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
491 | struct io_defer_entry, list); | |
de0617e4 | 492 | |
9cf7c104 | 493 | if (req_need_defer(de->req, de->seq)) |
04518945 | 494 | break; |
27dc8338 | 495 | list_del_init(&de->list); |
907d1df3 | 496 | io_req_task_queue(de->req); |
27dc8338 | 497 | kfree(de); |
441b8a78 | 498 | } |
04518945 PB |
499 | } |
500 | ||
21a091b9 DY |
501 | |
502 | static void io_eventfd_ops(struct rcu_head *rcu) | |
f2842ab5 | 503 | { |
d8e9214f | 504 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
21a091b9 | 505 | int ops = atomic_xchg(&ev_fd->ops, 0); |
305bef98 | 506 | |
21a091b9 | 507 | if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) |
44648532 | 508 | eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); |
d8e9214f | 509 | |
21a091b9 DY |
510 | /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback |
511 | * ordering in a race but if references are 0 we know we have to free | |
512 | * it regardless. | |
305bef98 | 513 | */ |
21a091b9 DY |
514 | if (atomic_dec_and_test(&ev_fd->refs)) { |
515 | eventfd_ctx_put(ev_fd->cq_ev_fd); | |
516 | kfree(ev_fd); | |
517 | } | |
d8e9214f DY |
518 | } |
519 | ||
77bc59b4 | 520 | static void io_eventfd_signal(struct io_ring_ctx *ctx) |
f2842ab5 | 521 | { |
21a091b9 | 522 | struct io_ev_fd *ev_fd = NULL; |
77bc59b4 | 523 | |
77bc59b4 UA |
524 | rcu_read_lock(); |
525 | /* | |
526 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
527 | * and eventfd_signal | |
528 | */ | |
529 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
530 | ||
531 | /* | |
532 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
533 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
534 | * the function and rcu_read_lock. | |
535 | */ | |
536 | if (unlikely(!ev_fd)) | |
537 | goto out; | |
7e55a19c | 538 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 | 539 | goto out; |
21a091b9 DY |
540 | if (ev_fd->eventfd_async && !io_wq_current_is_worker()) |
541 | goto out; | |
77bc59b4 | 542 | |
21a091b9 | 543 | if (likely(eventfd_signal_allowed())) { |
44648532 | 544 | eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); |
21a091b9 DY |
545 | } else { |
546 | atomic_inc(&ev_fd->refs); | |
547 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) | |
44a84da4 | 548 | call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops); |
21a091b9 DY |
549 | else |
550 | atomic_dec(&ev_fd->refs); | |
551 | } | |
552 | ||
77bc59b4 UA |
553 | out: |
554 | rcu_read_unlock(); | |
f2842ab5 JA |
555 | } |
556 | ||
21a091b9 DY |
557 | static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) |
558 | { | |
559 | bool skip; | |
560 | ||
561 | spin_lock(&ctx->completion_lock); | |
562 | ||
563 | /* | |
564 | * Eventfd should only get triggered when at least one event has been | |
565 | * posted. Some applications rely on the eventfd notification count | |
566 | * only changing IFF a new CQE has been added to the CQ ring. There's | |
567 | * no depedency on 1:1 relationship between how many times this | |
568 | * function is called (and hence the eventfd count) and number of CQEs | |
569 | * posted to the CQ ring. | |
570 | */ | |
571 | skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
572 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
573 | spin_unlock(&ctx->completion_lock); | |
574 | if (skip) | |
575 | return; | |
576 | ||
577 | io_eventfd_signal(ctx); | |
578 | } | |
579 | ||
a830ffd2 PB |
580 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
581 | { | |
bca39f39 PB |
582 | if (ctx->poll_activated) |
583 | io_poll_wq_wake(ctx); | |
e5f30f6f PB |
584 | if (ctx->off_timeout_used) |
585 | io_flush_timeouts(ctx); | |
586 | if (ctx->drain_active) { | |
a830ffd2 | 587 | spin_lock(&ctx->completion_lock); |
e5f30f6f | 588 | io_queue_deferred(ctx); |
a830ffd2 PB |
589 | spin_unlock(&ctx->completion_lock); |
590 | } | |
591 | if (ctx->has_evfd) | |
21a091b9 | 592 | io_eventfd_flush_signal(ctx); |
a830ffd2 PB |
593 | } |
594 | ||
f66f7342 PB |
595 | static inline void __io_cq_lock(struct io_ring_ctx *ctx) |
596 | __acquires(ctx->completion_lock) | |
597 | { | |
598 | if (!ctx->task_complete) | |
599 | spin_lock(&ctx->completion_lock); | |
600 | } | |
601 | ||
602 | static inline void __io_cq_unlock(struct io_ring_ctx *ctx) | |
603 | { | |
604 | if (!ctx->task_complete) | |
605 | spin_unlock(&ctx->completion_lock); | |
606 | } | |
607 | ||
6971253f PB |
608 | static inline void io_cq_lock(struct io_ring_ctx *ctx) |
609 | __acquires(ctx->completion_lock) | |
610 | { | |
611 | spin_lock(&ctx->completion_lock); | |
612 | } | |
613 | ||
614 | static inline void io_cq_unlock(struct io_ring_ctx *ctx) | |
615 | __releases(ctx->completion_lock) | |
616 | { | |
617 | spin_unlock(&ctx->completion_lock); | |
618 | } | |
619 | ||
5d772916 | 620 | /* keep it inlined for io_submit_flush_completions() */ |
f66f7342 | 621 | static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) |
25399321 PB |
622 | __releases(ctx->completion_lock) |
623 | { | |
624 | io_commit_cqring(ctx); | |
f66f7342 | 625 | __io_cq_unlock(ctx); |
6c16fe3c JA |
626 | io_commit_cqring_flush(ctx); |
627 | io_cqring_wake(ctx); | |
25399321 PB |
628 | } |
629 | ||
3181e22f PB |
630 | static inline void __io_cq_unlock_post_flush(struct io_ring_ctx *ctx) |
631 | __releases(ctx->completion_lock) | |
632 | { | |
633 | io_commit_cqring(ctx); | |
634 | __io_cq_unlock(ctx); | |
635 | io_commit_cqring_flush(ctx); | |
636 | ||
637 | /* | |
638 | * As ->task_complete implies that the ring is single tasked, cq_wait | |
639 | * may only be waited on by the current in io_cqring_wait(), but since | |
640 | * it will re-check the wakeup conditions once we return we can safely | |
641 | * skip waking it up. | |
642 | */ | |
643 | if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) { | |
644 | smp_mb(); | |
645 | __io_cqring_wake(ctx); | |
646 | } | |
647 | } | |
648 | ||
25399321 | 649 | void io_cq_unlock_post(struct io_ring_ctx *ctx) |
5d772916 | 650 | __releases(ctx->completion_lock) |
25399321 | 651 | { |
f66f7342 PB |
652 | io_commit_cqring(ctx); |
653 | spin_unlock(&ctx->completion_lock); | |
654 | io_commit_cqring_flush(ctx); | |
655 | io_cqring_wake(ctx); | |
25399321 PB |
656 | } |
657 | ||
c4a2ed72 | 658 | /* Returns true if there are no backlogged entries after the flush */ |
a85381d8 PB |
659 | static void io_cqring_overflow_kill(struct io_ring_ctx *ctx) |
660 | { | |
661 | struct io_overflow_cqe *ocqe; | |
662 | LIST_HEAD(list); | |
663 | ||
664 | io_cq_lock(ctx); | |
665 | list_splice_init(&ctx->cq_overflow_list, &list); | |
666 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); | |
667 | io_cq_unlock(ctx); | |
668 | ||
669 | while (!list_empty(&list)) { | |
670 | ocqe = list_first_entry(&list, struct io_overflow_cqe, list); | |
671 | list_del(&ocqe->list); | |
672 | kfree(ocqe); | |
673 | } | |
674 | } | |
675 | ||
1b346e4a | 676 | static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
1d7bb1d5 | 677 | { |
e45a3e05 | 678 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 679 | |
a85381d8 | 680 | if (__io_cqring_events(ctx) == ctx->cq_entries) |
1b346e4a | 681 | return; |
1d7bb1d5 | 682 | |
e45a3e05 SR |
683 | if (ctx->flags & IORING_SETUP_CQE32) |
684 | cqe_size <<= 1; | |
685 | ||
25399321 | 686 | io_cq_lock(ctx); |
6c2450ae | 687 | while (!list_empty(&ctx->cq_overflow_list)) { |
aa1df3a3 | 688 | struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true); |
6c2450ae | 689 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 690 | |
a85381d8 | 691 | if (!cqe) |
1d7bb1d5 | 692 | break; |
6c2450ae PB |
693 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
694 | struct io_overflow_cqe, list); | |
a85381d8 | 695 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae PB |
696 | list_del(&ocqe->list); |
697 | kfree(ocqe); | |
1d7bb1d5 JA |
698 | } |
699 | ||
1b346e4a | 700 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 701 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 702 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 703 | } |
25399321 | 704 | io_cq_unlock_post(ctx); |
1d7bb1d5 JA |
705 | } |
706 | ||
52ea806a JA |
707 | static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx) |
708 | { | |
709 | /* iopoll syncs against uring_lock, not completion_lock */ | |
710 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
711 | mutex_lock(&ctx->uring_lock); | |
712 | __io_cqring_overflow_flush(ctx); | |
713 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
714 | mutex_unlock(&ctx->uring_lock); | |
715 | } | |
716 | ||
1b346e4a | 717 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 718 | { |
52ea806a JA |
719 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) |
720 | io_cqring_do_overflow_flush(ctx); | |
6c503150 PB |
721 | } |
722 | ||
5afa4650 PB |
723 | /* can be called by any task */ |
724 | static void io_put_task_remote(struct task_struct *task, int nr) | |
6a290a14 PB |
725 | { |
726 | struct io_uring_task *tctx = task->io_uring; | |
727 | ||
9d170164 | 728 | percpu_counter_sub(&tctx->inflight, nr); |
8d664282 | 729 | if (unlikely(atomic_read(&tctx->in_cancel))) |
9d170164 PB |
730 | wake_up(&tctx->wait); |
731 | put_task_struct_many(task, nr); | |
732 | } | |
733 | ||
5afa4650 PB |
734 | /* used by a task to put its own references */ |
735 | static void io_put_task_local(struct task_struct *task, int nr) | |
736 | { | |
737 | task->io_uring->cached_refs += nr; | |
738 | } | |
739 | ||
89800a2d PB |
740 | /* must to be called somewhat shortly after putting a request */ |
741 | static inline void io_put_task(struct task_struct *task, int nr) | |
742 | { | |
743 | if (likely(task == current)) | |
5afa4650 | 744 | io_put_task_local(task, nr); |
89800a2d | 745 | else |
5afa4650 | 746 | io_put_task_remote(task, nr); |
89800a2d PB |
747 | } |
748 | ||
63809137 | 749 | void io_task_refs_refill(struct io_uring_task *tctx) |
9a10867a PB |
750 | { |
751 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
752 | ||
753 | percpu_counter_add(&tctx->inflight, refill); | |
754 | refcount_add(refill, ¤t->usage); | |
755 | tctx->cached_refs += refill; | |
756 | } | |
757 | ||
3cc7fdb9 PB |
758 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
759 | { | |
760 | struct io_uring_task *tctx = task->io_uring; | |
761 | unsigned int refs = tctx->cached_refs; | |
762 | ||
763 | if (refs) { | |
764 | tctx->cached_refs = 0; | |
765 | percpu_counter_sub(&tctx->inflight, refs); | |
766 | put_task_struct_many(task, refs); | |
767 | } | |
768 | } | |
769 | ||
68494a65 PB |
770 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
771 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 772 | { |
cce4b8b0 | 773 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
774 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
775 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 776 | |
f26cc959 PB |
777 | lockdep_assert_held(&ctx->completion_lock); |
778 | ||
e45a3e05 SR |
779 | if (is_cqe32) |
780 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 781 | |
e45a3e05 | 782 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 783 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
784 | if (!ocqe) { |
785 | /* | |
786 | * If we're in ring overflow flush mode, or in task cancel mode, | |
787 | * or cannot allocate an overflow entry, then we need to drop it | |
788 | * on the floor. | |
789 | */ | |
8f6ed49a | 790 | io_account_cq_overflow(ctx); |
155bc950 | 791 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 792 | return false; |
2b188cc1 | 793 | } |
cce4b8b0 | 794 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 795 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 796 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 797 | |
cce4b8b0 | 798 | } |
d4d19c19 | 799 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
800 | ocqe->cqe.res = res; |
801 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
802 | if (is_cqe32) { |
803 | ocqe->cqe.big_cqe[0] = extra1; | |
804 | ocqe->cqe.big_cqe[1] = extra2; | |
805 | } | |
cce4b8b0 PB |
806 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
807 | return true; | |
2b188cc1 JA |
808 | } |
809 | ||
68494a65 PB |
810 | bool io_req_cqe_overflow(struct io_kiocb *req) |
811 | { | |
812 | if (!(req->flags & REQ_F_CQE32_INIT)) { | |
813 | req->extra1 = 0; | |
814 | req->extra2 = 0; | |
815 | } | |
816 | return io_cqring_event_overflow(req->ctx, req->cqe.user_data, | |
817 | req->cqe.res, req->cqe.flags, | |
818 | req->extra1, req->extra2); | |
819 | } | |
820 | ||
faf88dde PB |
821 | /* |
822 | * writes to the cq entry need to come after reading head; the | |
823 | * control dependency is enough as we're using WRITE_ONCE to | |
824 | * fill the cq entry | |
825 | */ | |
aa1df3a3 | 826 | struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow) |
faf88dde PB |
827 | { |
828 | struct io_rings *rings = ctx->rings; | |
829 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
830 | unsigned int free, queued, len; |
831 | ||
aa1df3a3 PB |
832 | /* |
833 | * Posting into the CQ when there are pending overflowed CQEs may break | |
834 | * ordering guarantees, which will affect links, F_MORE users and more. | |
835 | * Force overflow the completion. | |
836 | */ | |
837 | if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) | |
838 | return NULL; | |
faf88dde PB |
839 | |
840 | /* userspace may cheat modifying the tail, be safe and do min */ | |
841 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
842 | free = ctx->cq_entries - queued; | |
843 | /* we need a contiguous range, limit based on the current array offset */ | |
844 | len = min(free, ctx->cq_entries - off); | |
845 | if (!len) | |
846 | return NULL; | |
847 | ||
b3659a65 PB |
848 | if (ctx->flags & IORING_SETUP_CQE32) { |
849 | off <<= 1; | |
850 | len <<= 1; | |
851 | } | |
852 | ||
faf88dde PB |
853 | ctx->cqe_cached = &rings->cqes[off]; |
854 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
b3659a65 PB |
855 | |
856 | ctx->cached_cq_tail++; | |
faf88dde | 857 | ctx->cqe_cached++; |
b3659a65 PB |
858 | if (ctx->flags & IORING_SETUP_CQE32) |
859 | ctx->cqe_cached++; | |
860 | return &rings->cqes[off]; | |
faf88dde PB |
861 | } |
862 | ||
f66f7342 PB |
863 | static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, |
864 | u32 cflags) | |
bcda7baa | 865 | { |
cd94903d PB |
866 | struct io_uring_cqe *cqe; |
867 | ||
913a571a | 868 | ctx->cq_extra++; |
cd94903d PB |
869 | |
870 | /* | |
871 | * If we can't get a cq entry, userspace overflowed the | |
872 | * submission (by quite a lot). Increment the overflow count in | |
873 | * the ring. | |
874 | */ | |
875 | cqe = io_get_cqe(ctx); | |
876 | if (likely(cqe)) { | |
e0486f3f DY |
877 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
878 | ||
cd94903d PB |
879 | WRITE_ONCE(cqe->user_data, user_data); |
880 | WRITE_ONCE(cqe->res, res); | |
881 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
882 | |
883 | if (ctx->flags & IORING_SETUP_CQE32) { | |
884 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
885 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
886 | } | |
cd94903d PB |
887 | return true; |
888 | } | |
52120f0f | 889 | return false; |
bcda7baa JA |
890 | } |
891 | ||
931147dd DY |
892 | static void __io_flush_post_cqes(struct io_ring_ctx *ctx) |
893 | __must_hold(&ctx->uring_lock) | |
894 | { | |
895 | struct io_submit_state *state = &ctx->submit_state; | |
896 | unsigned int i; | |
897 | ||
898 | lockdep_assert_held(&ctx->uring_lock); | |
899 | for (i = 0; i < state->cqes_count; i++) { | |
900 | struct io_uring_cqe *cqe = &state->cqes[i]; | |
901 | ||
f66f7342 PB |
902 | if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) { |
903 | if (ctx->task_complete) { | |
904 | spin_lock(&ctx->completion_lock); | |
905 | io_cqring_event_overflow(ctx, cqe->user_data, | |
906 | cqe->res, cqe->flags, 0, 0); | |
907 | spin_unlock(&ctx->completion_lock); | |
908 | } else { | |
909 | io_cqring_event_overflow(ctx, cqe->user_data, | |
910 | cqe->res, cqe->flags, 0, 0); | |
911 | } | |
912 | } | |
931147dd DY |
913 | } |
914 | state->cqes_count = 0; | |
915 | } | |
916 | ||
b529c96a DY |
917 | static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, |
918 | bool allow_overflow) | |
d245bca6 PB |
919 | { |
920 | bool filled; | |
921 | ||
25399321 | 922 | io_cq_lock(ctx); |
f66f7342 PB |
923 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags); |
924 | if (!filled && allow_overflow) | |
925 | filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); | |
926 | ||
25399321 | 927 | io_cq_unlock_post(ctx); |
d245bca6 PB |
928 | return filled; |
929 | } | |
930 | ||
b529c96a DY |
931 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags) |
932 | { | |
933 | return __io_post_aux_cqe(ctx, user_data, res, cflags, true); | |
934 | } | |
935 | ||
9b8c5475 DY |
936 | bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags, |
937 | bool allow_overflow) | |
2b188cc1 | 938 | { |
9b8c5475 DY |
939 | struct io_uring_cqe *cqe; |
940 | unsigned int length; | |
941 | ||
942 | if (!defer) | |
b529c96a | 943 | return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow); |
9b8c5475 DY |
944 | |
945 | length = ARRAY_SIZE(ctx->submit_state.cqes); | |
946 | ||
947 | lockdep_assert_held(&ctx->uring_lock); | |
948 | ||
949 | if (ctx->submit_state.cqes_count == length) { | |
f66f7342 | 950 | __io_cq_lock(ctx); |
9b8c5475 DY |
951 | __io_flush_post_cqes(ctx); |
952 | /* no need to flush - flush is deferred */ | |
f66f7342 | 953 | __io_cq_unlock_post(ctx); |
9b8c5475 DY |
954 | } |
955 | ||
956 | /* For defered completions this is not as strict as it is otherwise, | |
957 | * however it's main job is to prevent unbounded posted completions, | |
958 | * and in that it works just as well. | |
959 | */ | |
960 | if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) | |
961 | return false; | |
962 | ||
963 | cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++]; | |
964 | cqe->user_data = user_data; | |
965 | cqe->res = res; | |
966 | cqe->flags = cflags; | |
967 | return true; | |
968 | } | |
969 | ||
1bec951c | 970 | static void __io_req_complete_post(struct io_kiocb *req) |
2b188cc1 | 971 | { |
fa18fa22 PB |
972 | struct io_ring_ctx *ctx = req->ctx; |
973 | ||
974 | io_cq_lock(ctx); | |
975 | if (!(req->flags & REQ_F_CQE_SKIP)) | |
a8cf95f9 | 976 | io_fill_cqe_req(ctx, req); |
fa18fa22 | 977 | |
c7dae4ba JA |
978 | /* |
979 | * If we're the last reference to this request, add to our locked | |
980 | * free_list cache. | |
981 | */ | |
de9b4cca | 982 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 983 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 984 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
985 | io_disarm_next(req); |
986 | if (req->link) { | |
987 | io_req_task_queue(req->link); | |
988 | req->link = NULL; | |
989 | } | |
990 | } | |
68a2cc1b PB |
991 | io_put_kbuf_comp(req); |
992 | io_dismantle_req(req); | |
7ac1edc4 | 993 | io_req_put_rsrc(req); |
8197b053 PB |
994 | /* |
995 | * Selected buffer deallocation in io_clean_op() assumes that | |
996 | * we don't hold ->completion_lock. Clean them here to avoid | |
997 | * deadlocks. | |
998 | */ | |
5afa4650 | 999 | io_put_task_remote(req->task, 1); |
c2b6c6bc | 1000 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 1001 | ctx->locked_free_nr++; |
180f829f | 1002 | } |
25399321 | 1003 | io_cq_unlock_post(ctx); |
4e3d9ff9 JA |
1004 | } |
1005 | ||
1bec951c | 1006 | void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
bcda7baa | 1007 | { |
e6aeb272 PB |
1008 | if (req->ctx->task_complete && (issue_flags & IO_URING_F_IOWQ)) { |
1009 | req->io_task_work.func = io_req_task_complete; | |
1010 | io_req_task_work_add(req); | |
1011 | } else if (!(issue_flags & IO_URING_F_UNLOCKED) || | |
1012 | !(req->ctx->flags & IORING_SETUP_IOPOLL)) { | |
1bec951c PB |
1013 | __io_req_complete_post(req); |
1014 | } else { | |
1015 | struct io_ring_ctx *ctx = req->ctx; | |
1016 | ||
1017 | mutex_lock(&ctx->uring_lock); | |
1018 | __io_req_complete_post(req); | |
1019 | mutex_unlock(&ctx->uring_lock); | |
1020 | } | |
0ddf92e8 JA |
1021 | } |
1022 | ||
973fc83f | 1023 | void io_req_defer_failed(struct io_kiocb *req, s32 res) |
e276ae34 | 1024 | __must_hold(&ctx->uring_lock) |
f41db273 | 1025 | { |
f30bd4d0 | 1026 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; |
a47b255e | 1027 | |
e276ae34 PB |
1028 | lockdep_assert_held(&req->ctx->uring_lock); |
1029 | ||
93d2bcd2 | 1030 | req_set_fail(req); |
97b388d7 | 1031 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
a47b255e PB |
1032 | if (def->fail) |
1033 | def->fail(req); | |
973fc83f | 1034 | io_req_complete_defer(req); |
f41db273 PB |
1035 | } |
1036 | ||
864ea921 PB |
1037 | /* |
1038 | * Don't initialise the fields below on every allocation, but do that in | |
1039 | * advance and keep them valid across allocations. | |
1040 | */ | |
1041 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
1042 | { | |
1043 | req->ctx = ctx; | |
1044 | req->link = NULL; | |
1045 | req->async_data = NULL; | |
1046 | /* not necessary, but safer to zero */ | |
cef216fc | 1047 | req->cqe.res = 0; |
864ea921 PB |
1048 | } |
1049 | ||
dac7a098 | 1050 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 1051 | struct io_submit_state *state) |
dac7a098 | 1052 | { |
79ebeaee | 1053 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1054 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 1055 | ctx->locked_free_nr = 0; |
79ebeaee | 1056 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
1057 | } |
1058 | ||
5d5901a3 PB |
1059 | /* |
1060 | * A request might get retired back into the request caches even before opcode | |
1061 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
1062 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
1063 | * and with extra caution to not get a request that is still worked on. | |
1064 | */ | |
bd1a3783 | 1065 | __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 1066 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 1067 | { |
864ea921 | 1068 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 1069 | void *reqs[IO_REQ_ALLOC_BATCH]; |
864ea921 | 1070 | int ret, i; |
e5d1bc0a | 1071 | |
23a5c43b PB |
1072 | /* |
1073 | * If we have more than a batch's worth of requests in our IRQ side | |
1074 | * locked cache, grab the lock and move them over to our submission | |
1075 | * side cache. | |
1076 | */ | |
a6d97a8a | 1077 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 1078 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 1079 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
1080 | return true; |
1081 | } | |
e5d1bc0a | 1082 | |
3ab665b7 | 1083 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 1084 | |
864ea921 PB |
1085 | /* |
1086 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1087 | * retry single alloc to be on the safe side. | |
1088 | */ | |
1089 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
1090 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
1091 | if (!reqs[0]) | |
a33ae9ce | 1092 | return false; |
864ea921 | 1093 | ret = 1; |
2b188cc1 | 1094 | } |
864ea921 | 1095 | |
37f0e767 | 1096 | percpu_ref_get_many(&ctx->refs, ret); |
3ab665b7 | 1097 | for (i = 0; i < ret; i++) { |
23a5c43b | 1098 | struct io_kiocb *req = reqs[i]; |
3ab665b7 PB |
1099 | |
1100 | io_preinit_req(req, ctx); | |
fa05457a | 1101 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 1102 | } |
a33ae9ce PB |
1103 | return true; |
1104 | } | |
1105 | ||
6b639522 | 1106 | static inline void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 1107 | { |
094bae49 | 1108 | unsigned int flags = req->flags; |
929a3af9 | 1109 | |
867f8fa5 | 1110 | if (unlikely(flags & IO_REQ_CLEAN_FLAGS)) |
3a0a6902 | 1111 | io_clean_op(req); |
e1d767f0 PB |
1112 | if (!(flags & REQ_F_FIXED_FILE)) |
1113 | io_put_file(req->file); | |
e65ef56d JA |
1114 | } |
1115 | ||
59915143 | 1116 | __cold void io_free_req(struct io_kiocb *req) |
c6ca97b3 | 1117 | { |
51a4cc11 | 1118 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 1119 | |
7ac1edc4 | 1120 | io_req_put_rsrc(req); |
216578e5 | 1121 | io_dismantle_req(req); |
5afa4650 | 1122 | io_put_task_remote(req->task, 1); |
c6ca97b3 | 1123 | |
79ebeaee | 1124 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1125 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
c34b025f | 1126 | ctx->locked_free_nr++; |
79ebeaee | 1127 | spin_unlock(&ctx->completion_lock); |
e65ef56d JA |
1128 | } |
1129 | ||
d81499bf PB |
1130 | static void __io_req_find_next_prep(struct io_kiocb *req) |
1131 | { | |
1132 | struct io_ring_ctx *ctx = req->ctx; | |
d81499bf | 1133 | |
6971253f | 1134 | spin_lock(&ctx->completion_lock); |
305bef98 | 1135 | io_disarm_next(req); |
6971253f | 1136 | spin_unlock(&ctx->completion_lock); |
d81499bf PB |
1137 | } |
1138 | ||
1139 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 1140 | { |
33cc89a9 | 1141 | struct io_kiocb *nxt; |
944e58bf | 1142 | |
9e645e11 JA |
1143 | /* |
1144 | * If LINK is set, we have dependent requests in this chain. If we | |
1145 | * didn't fail this request, queue the first one up, moving any other | |
1146 | * dependencies to the next request. In case of failure, fail the rest | |
1147 | * of the chain. | |
1148 | */ | |
d81499bf PB |
1149 | if (unlikely(req->flags & IO_DISARM_MASK)) |
1150 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
1151 | nxt = req->link; |
1152 | req->link = NULL; | |
1153 | return nxt; | |
4d7dd462 | 1154 | } |
9e645e11 | 1155 | |
f237c30a | 1156 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked) |
2c32395d PB |
1157 | { |
1158 | if (!ctx) | |
1159 | return; | |
ef060ea9 JA |
1160 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1161 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
f237c30a | 1162 | if (*locked) { |
c450178d | 1163 | io_submit_flush_completions(ctx); |
2c32395d | 1164 | mutex_unlock(&ctx->uring_lock); |
f237c30a | 1165 | *locked = false; |
2c32395d PB |
1166 | } |
1167 | percpu_ref_put(&ctx->refs); | |
1168 | } | |
1169 | ||
c6dd763c DY |
1170 | static unsigned int handle_tw_list(struct llist_node *node, |
1171 | struct io_ring_ctx **ctx, bool *locked, | |
1172 | struct llist_node *last) | |
9f8d032a | 1173 | { |
c6dd763c DY |
1174 | unsigned int count = 0; |
1175 | ||
cb6bf7f2 | 1176 | while (node && node != last) { |
f88262e6 | 1177 | struct llist_node *next = node->next; |
9f8d032a HX |
1178 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1179 | io_task_work.node); | |
1180 | ||
34d2bfe7 JA |
1181 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); |
1182 | ||
9f8d032a HX |
1183 | if (req->ctx != *ctx) { |
1184 | ctx_flush_and_put(*ctx, locked); | |
1185 | *ctx = req->ctx; | |
1186 | /* if not contended, grab and improve batching */ | |
1187 | *locked = mutex_trylock(&(*ctx)->uring_lock); | |
1188 | percpu_ref_get(&(*ctx)->refs); | |
f5868008 JA |
1189 | } else if (!*locked) |
1190 | *locked = mutex_trylock(&(*ctx)->uring_lock); | |
9f8d032a HX |
1191 | req->io_task_work.func(req, locked); |
1192 | node = next; | |
c6dd763c | 1193 | count++; |
f5868008 JA |
1194 | if (unlikely(need_resched())) { |
1195 | ctx_flush_and_put(*ctx, locked); | |
1196 | *ctx = NULL; | |
1197 | cond_resched(); | |
1198 | } | |
3a0c037b | 1199 | } |
c6dd763c DY |
1200 | |
1201 | return count; | |
9f8d032a HX |
1202 | } |
1203 | ||
923d1592 DY |
1204 | /** |
1205 | * io_llist_xchg - swap all entries in a lock-less list | |
1206 | * @head: the head of lock-less list to delete all entries | |
1207 | * @new: new entry as the head of the list | |
1208 | * | |
1209 | * If list is empty, return NULL, otherwise, return the pointer to the first entry. | |
1210 | * The order of entries returned is from the newest to the oldest added one. | |
1211 | */ | |
1212 | static inline struct llist_node *io_llist_xchg(struct llist_head *head, | |
1213 | struct llist_node *new) | |
1214 | { | |
1215 | return xchg(&head->first, new); | |
1216 | } | |
1217 | ||
1218 | /** | |
1219 | * io_llist_cmpxchg - possibly swap all entries in a lock-less list | |
1220 | * @head: the head of lock-less list to delete all entries | |
1221 | * @old: expected old value of the first entry of the list | |
1222 | * @new: new entry as the head of the list | |
1223 | * | |
1224 | * perform a cmpxchg on the first entry of the list. | |
1225 | */ | |
1226 | ||
1227 | static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head, | |
1228 | struct llist_node *old, | |
1229 | struct llist_node *new) | |
1230 | { | |
1231 | return cmpxchg(&head->first, old, new); | |
1232 | } | |
1233 | ||
c9f06aa7 | 1234 | void tctx_task_work(struct callback_head *cb) |
c40f6379 | 1235 | { |
f28c240e | 1236 | bool uring_locked = false; |
ebd0df2e | 1237 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
1238 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
1239 | task_work); | |
3a0c037b | 1240 | struct llist_node fake = {}; |
77e443ab | 1241 | struct llist_node *node; |
cb6bf7f2 PB |
1242 | unsigned int loops = 0; |
1243 | unsigned int count = 0; | |
3a0c037b | 1244 | |
77e443ab PB |
1245 | if (unlikely(current->flags & PF_EXITING)) { |
1246 | io_fallback_tw(tctx); | |
1247 | return; | |
1248 | } | |
3a0c037b | 1249 | |
cb6bf7f2 | 1250 | do { |
c6dd763c | 1251 | loops++; |
3a0c037b | 1252 | node = io_llist_xchg(&tctx->task_list, &fake); |
c6dd763c | 1253 | count += handle_tw_list(node, &ctx, &uring_locked, &fake); |
50470fc5 PB |
1254 | |
1255 | /* skip expensive cmpxchg if there are items in the list */ | |
1256 | if (READ_ONCE(tctx->task_list.first) != &fake) | |
1257 | continue; | |
1258 | if (uring_locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) { | |
1259 | io_submit_flush_completions(ctx); | |
1260 | if (READ_ONCE(tctx->task_list.first) != &fake) | |
1261 | continue; | |
1262 | } | |
3a0c037b | 1263 | node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); |
cb6bf7f2 | 1264 | } while (node != &fake); |
ebd0df2e | 1265 | |
f28c240e | 1266 | ctx_flush_and_put(ctx, &uring_locked); |
3cc7fdb9 | 1267 | |
8d664282 JA |
1268 | /* relaxed read is enough as only the task itself sets ->in_cancel */ |
1269 | if (unlikely(atomic_read(&tctx->in_cancel))) | |
3cc7fdb9 | 1270 | io_uring_drop_tctx_refs(current); |
c6dd763c DY |
1271 | |
1272 | trace_io_uring_task_work_run(tctx, count, loops); | |
7cbf1722 JA |
1273 | } |
1274 | ||
d7593606 PB |
1275 | static __cold void io_fallback_tw(struct io_uring_task *tctx) |
1276 | { | |
1277 | struct llist_node *node = llist_del_all(&tctx->task_list); | |
1278 | struct io_kiocb *req; | |
1279 | ||
1280 | while (node) { | |
1281 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1282 | node = node->next; | |
1283 | if (llist_add(&req->io_task_work.node, | |
1284 | &req->ctx->fallback_llist)) | |
1285 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1286 | } | |
1287 | } | |
1288 | ||
c0e0d6ba DY |
1289 | static void io_req_local_work_add(struct io_kiocb *req) |
1290 | { | |
1291 | struct io_ring_ctx *ctx = req->ctx; | |
1292 | ||
9ffa13ff PB |
1293 | percpu_ref_get(&ctx->refs); |
1294 | ||
ce8e04f6 JA |
1295 | if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) |
1296 | goto put_ref; | |
1297 | ||
130bd686 | 1298 | /* needed for the following wake up */ |
fc86f9d3 | 1299 | smp_mb__after_atomic(); |
c0e0d6ba | 1300 | |
8d664282 | 1301 | if (unlikely(atomic_read(&req->task->io_uring->in_cancel))) { |
c0e0d6ba | 1302 | io_move_task_work_from_local(ctx); |
ce8e04f6 | 1303 | goto put_ref; |
c0e0d6ba DY |
1304 | } |
1305 | ||
1306 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) | |
1307 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
21a091b9 DY |
1308 | if (ctx->has_evfd) |
1309 | io_eventfd_signal(ctx); | |
d80c0f00 PB |
1310 | |
1311 | if (READ_ONCE(ctx->cq_waiting)) | |
1312 | wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE); | |
ce8e04f6 JA |
1313 | |
1314 | put_ref: | |
9ffa13ff | 1315 | percpu_ref_put(&ctx->refs); |
c0e0d6ba DY |
1316 | } |
1317 | ||
e52d2e58 | 1318 | void __io_req_task_work_add(struct io_kiocb *req, bool allow_local) |
7cbf1722 | 1319 | { |
c34398a8 | 1320 | struct io_uring_task *tctx = req->task->io_uring; |
9f010507 | 1321 | struct io_ring_ctx *ctx = req->ctx; |
7cbf1722 | 1322 | |
c0e0d6ba DY |
1323 | if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
1324 | io_req_local_work_add(req); | |
1325 | return; | |
1326 | } | |
7cbf1722 JA |
1327 | |
1328 | /* task_work already pending, we're done */ | |
32d91f05 | 1329 | if (!llist_add(&req->io_task_work.node, &tctx->task_list)) |
e09ee510 | 1330 | return; |
7cbf1722 | 1331 | |
ef060ea9 JA |
1332 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1333 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1334 | ||
3fe07bcd | 1335 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1336 | return; |
2215bed9 | 1337 | |
d7593606 | 1338 | io_fallback_tw(tctx); |
c0e0d6ba DY |
1339 | } |
1340 | ||
1341 | static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) | |
1342 | { | |
1343 | struct llist_node *node; | |
1344 | ||
1345 | node = llist_del_all(&ctx->work_llist); | |
1346 | while (node) { | |
1347 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1348 | io_task_work.node); | |
1349 | ||
1350 | node = node->next; | |
1351 | __io_req_task_work_add(req, false); | |
1352 | } | |
1353 | } | |
1354 | ||
360173ab | 1355 | static int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked) |
c0e0d6ba | 1356 | { |
c0e0d6ba | 1357 | struct llist_node *node; |
c3f4d39e | 1358 | unsigned int loops = 0; |
140102ae | 1359 | int ret = 0; |
c0e0d6ba | 1360 | |
140102ae | 1361 | if (WARN_ON_ONCE(ctx->submitter_task != current)) |
c0e0d6ba | 1362 | return -EEXIST; |
c3f4d39e PB |
1363 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1364 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
c0e0d6ba | 1365 | again: |
c3f4d39e PB |
1366 | node = io_llist_xchg(&ctx->work_llist, NULL); |
1367 | while (node) { | |
c0e0d6ba DY |
1368 | struct llist_node *next = node->next; |
1369 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1370 | io_task_work.node); | |
1371 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); | |
b3026767 | 1372 | req->io_task_work.func(req, locked); |
c0e0d6ba DY |
1373 | ret++; |
1374 | node = next; | |
1375 | } | |
c3f4d39e | 1376 | loops++; |
c0e0d6ba | 1377 | |
c3f4d39e | 1378 | if (!llist_empty(&ctx->work_llist)) |
c0e0d6ba | 1379 | goto again; |
b0b7a7d2 | 1380 | if (*locked) { |
c0e0d6ba | 1381 | io_submit_flush_completions(ctx); |
b0b7a7d2 PB |
1382 | if (!llist_empty(&ctx->work_llist)) |
1383 | goto again; | |
1384 | } | |
f75d5036 | 1385 | trace_io_uring_local_work_run(ctx, ret, loops); |
c0e0d6ba | 1386 | return ret; |
8ac5d85a JA |
1387 | } |
1388 | ||
360173ab PB |
1389 | static inline int io_run_local_work_locked(struct io_ring_ctx *ctx) |
1390 | { | |
1391 | bool locked; | |
1392 | int ret; | |
1393 | ||
1394 | if (llist_empty(&ctx->work_llist)) | |
1395 | return 0; | |
1396 | ||
1397 | locked = true; | |
1398 | ret = __io_run_local_work(ctx, &locked); | |
1399 | /* shouldn't happen! */ | |
1400 | if (WARN_ON_ONCE(!locked)) | |
1401 | mutex_lock(&ctx->uring_lock); | |
1402 | return ret; | |
1403 | } | |
1404 | ||
3e565555 | 1405 | static int io_run_local_work(struct io_ring_ctx *ctx) |
8ac5d85a | 1406 | { |
2f413956 | 1407 | bool locked = mutex_trylock(&ctx->uring_lock); |
8ac5d85a JA |
1408 | int ret; |
1409 | ||
b3026767 | 1410 | ret = __io_run_local_work(ctx, &locked); |
8ac5d85a JA |
1411 | if (locked) |
1412 | mutex_unlock(&ctx->uring_lock); | |
1413 | ||
1414 | return ret; | |
c0e0d6ba DY |
1415 | } |
1416 | ||
f237c30a | 1417 | static void io_req_task_cancel(struct io_kiocb *req, bool *locked) |
c40f6379 | 1418 | { |
971cf9c1 | 1419 | io_tw_lock(req->ctx, locked); |
973fc83f | 1420 | io_req_defer_failed(req, req->cqe.res); |
c40f6379 JA |
1421 | } |
1422 | ||
329061d3 | 1423 | void io_req_task_submit(struct io_kiocb *req, bool *locked) |
c40f6379 | 1424 | { |
971cf9c1 | 1425 | io_tw_lock(req->ctx, locked); |
316319e8 | 1426 | /* req->task == current here, checking PF_EXITING is safe */ |
6bb30855 | 1427 | if (unlikely(req->task->flags & PF_EXITING)) |
973fc83f | 1428 | io_req_defer_failed(req, -EFAULT); |
6bb30855 DY |
1429 | else if (req->flags & REQ_F_FORCE_ASYNC) |
1430 | io_queue_iowq(req, locked); | |
1431 | else | |
1432 | io_queue_sqe(req); | |
c40f6379 JA |
1433 | } |
1434 | ||
59915143 | 1435 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1436 | { |
97b388d7 | 1437 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1438 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1439 | io_req_task_work_add(req); |
c40f6379 JA |
1440 | } |
1441 | ||
f3b44f92 | 1442 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1443 | { |
5b0a6acc | 1444 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1445 | io_req_task_work_add(req); |
a3df7698 PB |
1446 | } |
1447 | ||
59915143 | 1448 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1449 | { |
9b5f7bd9 | 1450 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1451 | |
1452 | if (nxt) | |
906a8c3f | 1453 | io_req_task_queue(nxt); |
c69f8dbe JL |
1454 | } |
1455 | ||
f3b44f92 | 1456 | void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node) |
3aa83bfb | 1457 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1458 | { |
d4b7a5ef | 1459 | struct task_struct *task = NULL; |
37f0e767 | 1460 | int task_refs = 0; |
5af1d13e | 1461 | |
3aa83bfb PB |
1462 | do { |
1463 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1464 | comp_list); | |
2d6500d4 | 1465 | |
a538be5b PB |
1466 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1467 | if (req->flags & REQ_F_REFCOUNT) { | |
1468 | node = req->comp_list.next; | |
1469 | if (!req_ref_put_and_test(req)) | |
1470 | continue; | |
1471 | } | |
b605a7fa PB |
1472 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1473 | struct async_poll *apoll = req->apoll; | |
1474 | ||
1475 | if (apoll->double_poll) | |
1476 | kfree(apoll->double_poll); | |
9731bc98 JA |
1477 | if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) |
1478 | kfree(apoll); | |
b605a7fa PB |
1479 | req->flags &= ~REQ_F_POLLED; |
1480 | } | |
da1a08c5 | 1481 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1482 | io_queue_next(req); |
a538be5b PB |
1483 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1484 | io_clean_op(req); | |
c1e53a69 | 1485 | } |
a538be5b PB |
1486 | if (!(req->flags & REQ_F_FIXED_FILE)) |
1487 | io_put_file(req->file); | |
2d6500d4 | 1488 | |
ab409402 | 1489 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1490 | |
d4b7a5ef PB |
1491 | if (req->task != task) { |
1492 | if (task) | |
1493 | io_put_task(task, task_refs); | |
1494 | task = req->task; | |
1495 | task_refs = 0; | |
1496 | } | |
1497 | task_refs++; | |
c1e53a69 | 1498 | node = req->comp_list.next; |
fa05457a | 1499 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1500 | } while (node); |
d4b7a5ef | 1501 | |
d4b7a5ef PB |
1502 | if (task) |
1503 | io_put_task(task, task_refs); | |
7a743e22 PB |
1504 | } |
1505 | ||
c450178d | 1506 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1507 | __must_hold(&ctx->uring_lock) |
905c172f | 1508 | { |
cd0ca2e0 | 1509 | struct io_submit_state *state = &ctx->submit_state; |
fa780334 | 1510 | struct io_wq_work_node *node; |
905c172f | 1511 | |
f66f7342 | 1512 | __io_cq_lock(ctx); |
931147dd DY |
1513 | /* must come first to preserve CQE ordering in failure cases */ |
1514 | if (state->cqes_count) | |
1515 | __io_flush_post_cqes(ctx); | |
fa780334 | 1516 | __wq_list_for_each(node, &state->compl_reqs) { |
d9dee430 PB |
1517 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1518 | comp_list); | |
3d4aeb9f | 1519 | |
f66f7342 PB |
1520 | if (!(req->flags & REQ_F_CQE_SKIP) && |
1521 | unlikely(!__io_fill_cqe_req(ctx, req))) { | |
1522 | if (ctx->task_complete) { | |
1523 | spin_lock(&ctx->completion_lock); | |
1524 | io_req_cqe_overflow(req); | |
1525 | spin_unlock(&ctx->completion_lock); | |
1526 | } else { | |
1527 | io_req_cqe_overflow(req); | |
1528 | } | |
1529 | } | |
905c172f | 1530 | } |
3181e22f | 1531 | __io_cq_unlock_post_flush(ctx); |
d9dee430 | 1532 | |
931147dd DY |
1533 | if (!wq_list_empty(&ctx->submit_state.compl_reqs)) { |
1534 | io_free_batch_list(ctx, state->compl_reqs.first); | |
1535 | INIT_WQ_LIST(&state->compl_reqs); | |
1536 | } | |
7a743e22 PB |
1537 | } |
1538 | ||
ba816ad6 JA |
1539 | /* |
1540 | * Drop reference to request, return next in chain (if there is one) if this | |
1541 | * was the last reference to this request. | |
1542 | */ | |
0d85035a | 1543 | static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 1544 | { |
9b5f7bd9 PB |
1545 | struct io_kiocb *nxt = NULL; |
1546 | ||
de9b4cca | 1547 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 1548 | if (unlikely(req->flags & IO_REQ_LINK_FLAGS)) |
7819a1f6 | 1549 | nxt = io_req_find_next(req); |
f5c6cf2a | 1550 | io_free_req(req); |
2a44f467 | 1551 | } |
9b5f7bd9 | 1552 | return nxt; |
2b188cc1 JA |
1553 | } |
1554 | ||
6c503150 | 1555 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1556 | { |
1557 | /* See comment at the top of this file */ | |
1558 | smp_rmb(); | |
e23de15f | 1559 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1560 | } |
1561 | ||
def596e9 JA |
1562 | /* |
1563 | * We can't just wait for polled events to come to us, we have to actively | |
1564 | * find and complete them. | |
1565 | */ | |
c072481d | 1566 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1567 | { |
1568 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1569 | return; | |
1570 | ||
1571 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1572 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1573 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1574 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1575 | break; |
08f5439f JA |
1576 | /* |
1577 | * Ensure we allow local-to-the-cpu processing to take place, | |
1578 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1579 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1580 | */ |
3fcee5a6 PB |
1581 | if (need_resched()) { |
1582 | mutex_unlock(&ctx->uring_lock); | |
1583 | cond_resched(); | |
1584 | mutex_lock(&ctx->uring_lock); | |
1585 | } | |
def596e9 JA |
1586 | } |
1587 | mutex_unlock(&ctx->uring_lock); | |
1588 | } | |
1589 | ||
7668b92a | 1590 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1591 | { |
7668b92a | 1592 | unsigned int nr_events = 0; |
e9979b36 | 1593 | int ret = 0; |
155bc950 | 1594 | unsigned long check_cq; |
500f9fba | 1595 | |
76de6749 PB |
1596 | if (!io_allowed_run_tw(ctx)) |
1597 | return -EEXIST; | |
1598 | ||
3a08576b PB |
1599 | check_cq = READ_ONCE(ctx->check_cq); |
1600 | if (unlikely(check_cq)) { | |
1601 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
a85381d8 | 1602 | __io_cqring_overflow_flush(ctx); |
3a08576b PB |
1603 | /* |
1604 | * Similarly do not spin if we have not informed the user of any | |
1605 | * dropped CQE. | |
1606 | */ | |
1607 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1608 | return -EBADR; | |
1609 | } | |
f39c8a5b PB |
1610 | /* |
1611 | * Don't enter poll loop if we already have events pending. | |
1612 | * If we do, we can potentially be spinning for commands that | |
1613 | * already triggered a CQE (eg in error). | |
1614 | */ | |
f39c8a5b | 1615 | if (io_cqring_events(ctx)) |
d487b43c | 1616 | return 0; |
155bc950 | 1617 | |
def596e9 | 1618 | do { |
500f9fba JA |
1619 | /* |
1620 | * If a submit got punted to a workqueue, we can have the | |
1621 | * application entering polling for a command before it gets | |
1622 | * issued. That app will hold the uring_lock for the duration | |
1623 | * of the poll right here, so we need to take a breather every | |
1624 | * now and then to ensure that the issue has a chance to add | |
1625 | * the poll to the issued list. Otherwise we can spin here | |
1626 | * forever, while the workqueue is stuck trying to acquire the | |
1627 | * very same mutex. | |
1628 | */ | |
dac6a0ea JA |
1629 | if (wq_list_empty(&ctx->iopoll_list) || |
1630 | io_task_work_pending(ctx)) { | |
8f487ef2 PB |
1631 | u32 tail = ctx->cached_cq_tail; |
1632 | ||
8de11cdc | 1633 | (void) io_run_local_work_locked(ctx); |
def596e9 | 1634 | |
dac6a0ea JA |
1635 | if (task_work_pending(current) || |
1636 | wq_list_empty(&ctx->iopoll_list)) { | |
dac6a0ea | 1637 | mutex_unlock(&ctx->uring_lock); |
9d54bd6a | 1638 | io_run_task_work(); |
dac6a0ea | 1639 | mutex_lock(&ctx->uring_lock); |
dac6a0ea | 1640 | } |
8f487ef2 PB |
1641 | /* some requests don't go through iopoll_list */ |
1642 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1643 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1644 | break; |
500f9fba | 1645 | } |
5ba3c874 PB |
1646 | ret = io_do_iopoll(ctx, !min); |
1647 | if (ret < 0) | |
1648 | break; | |
1649 | nr_events += ret; | |
1650 | ret = 0; | |
1651 | } while (nr_events < min && !need_resched()); | |
d487b43c | 1652 | |
def596e9 JA |
1653 | return ret; |
1654 | } | |
7012c815 PB |
1655 | |
1656 | void io_req_task_complete(struct io_kiocb *req, bool *locked) | |
8ef12efe | 1657 | { |
7012c815 | 1658 | if (*locked) |
9da070b1 | 1659 | io_req_complete_defer(req); |
7012c815 | 1660 | else |
27f35fe9 | 1661 | io_req_complete_post(req, IO_URING_F_UNLOCKED); |
8ef12efe JA |
1662 | } |
1663 | ||
def596e9 JA |
1664 | /* |
1665 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1666 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1667 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1668 | * accessing the kiocb cookie. |
1669 | */ | |
9882131c | 1670 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1671 | { |
1672 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1673 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1674 | |
1675 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1676 | if (unlikely(needs_lock)) |
cb3d8972 | 1677 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1678 | |
1679 | /* | |
1680 | * Track whether we have multiple files in our lists. This will impact | |
1681 | * how we do polling eventually, not spinning if we're on potentially | |
1682 | * different devices. | |
1683 | */ | |
5eef4e87 | 1684 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1685 | ctx->poll_multi_queue = false; |
1686 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1687 | struct io_kiocb *list_req; |
1688 | ||
5eef4e87 PB |
1689 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1690 | comp_list); | |
30da1b45 | 1691 | if (list_req->file != req->file) |
915b3dde | 1692 | ctx->poll_multi_queue = true; |
def596e9 JA |
1693 | } |
1694 | ||
1695 | /* | |
1696 | * For fast devices, IO may have already completed. If it has, add | |
1697 | * it to the front so we find it first. | |
1698 | */ | |
65a6543d | 1699 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1700 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1701 | else |
5eef4e87 | 1702 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1703 | |
3b44b371 | 1704 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1705 | /* |
1706 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1707 | * in sq thread task context or in io worker task context. If | |
1708 | * current task context is sq thread, we don't need to check | |
1709 | * whether should wake up sq thread. | |
1710 | */ | |
1711 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1712 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1713 | wake_up(&ctx->sq_data->wait); | |
1714 | ||
1715 | mutex_unlock(&ctx->uring_lock); | |
1716 | } | |
def596e9 JA |
1717 | } |
1718 | ||
4503b767 JA |
1719 | static bool io_bdev_nowait(struct block_device *bdev) |
1720 | { | |
568ec936 | 1721 | return !bdev || bdev_nowait(bdev); |
4503b767 JA |
1722 | } |
1723 | ||
2b188cc1 JA |
1724 | /* |
1725 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1726 | * any file. For now, just ensure that anything potentially problematic is done | |
1727 | * inline. | |
1728 | */ | |
88459b50 | 1729 | static bool __io_file_supports_nowait(struct file *file, umode_t mode) |
2b188cc1 | 1730 | { |
4503b767 | 1731 | if (S_ISBLK(mode)) { |
4e7b5671 CH |
1732 | if (IS_ENABLED(CONFIG_BLOCK) && |
1733 | io_bdev_nowait(I_BDEV(file->f_mapping->host))) | |
4503b767 JA |
1734 | return true; |
1735 | return false; | |
1736 | } | |
976517f1 | 1737 | if (S_ISSOCK(mode)) |
2b188cc1 | 1738 | return true; |
4503b767 | 1739 | if (S_ISREG(mode)) { |
4e7b5671 CH |
1740 | if (IS_ENABLED(CONFIG_BLOCK) && |
1741 | io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
e5550a14 | 1742 | !io_is_uring_fops(file)) |
4503b767 JA |
1743 | return true; |
1744 | return false; | |
1745 | } | |
2b188cc1 | 1746 | |
c5b85625 JA |
1747 | /* any ->read/write should understand O_NONBLOCK */ |
1748 | if (file->f_flags & O_NONBLOCK) | |
1749 | return true; | |
35645ac3 | 1750 | return file->f_mode & FMODE_NOWAIT; |
2b188cc1 | 1751 | } |
c5b85625 | 1752 | |
88459b50 PB |
1753 | /* |
1754 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1755 | * any file. For now, just ensure that anything potentially problematic is done | |
1756 | * inline. | |
1757 | */ | |
a4ad4f74 | 1758 | unsigned int io_file_get_flags(struct file *file) |
88459b50 PB |
1759 | { |
1760 | umode_t mode = file_inode(file)->i_mode; | |
1761 | unsigned int res = 0; | |
af197f50 | 1762 | |
88459b50 PB |
1763 | if (S_ISREG(mode)) |
1764 | res |= FFS_ISREG; | |
1765 | if (__io_file_supports_nowait(file, mode)) | |
1766 | res |= FFS_NOWAIT; | |
1767 | return res; | |
2b188cc1 JA |
1768 | } |
1769 | ||
99f15d8d | 1770 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1771 | { |
f30bd4d0 BL |
1772 | WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size); |
1773 | req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL); | |
d886e185 PB |
1774 | if (req->async_data) { |
1775 | req->flags |= REQ_F_ASYNC_DATA; | |
1776 | return false; | |
1777 | } | |
1778 | return true; | |
3d9932a8 XW |
1779 | } |
1780 | ||
f3b44f92 | 1781 | int io_req_prep_async(struct io_kiocb *req) |
f67676d1 | 1782 | { |
f30bd4d0 | 1783 | const struct io_cold_def *cdef = &io_cold_defs[req->opcode]; |
a7dd2782 | 1784 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
0702e536 JA |
1785 | |
1786 | /* assign early for deferred execution for non-fixed file */ | |
54aa7f23 | 1787 | if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file) |
0702e536 | 1788 | req->file = io_file_get_normal(req, req->cqe.fd); |
f30bd4d0 | 1789 | if (!cdef->prep_async) |
0702e536 JA |
1790 | return 0; |
1791 | if (WARN_ON_ONCE(req_has_async_data(req))) | |
1792 | return -EFAULT; | |
f30bd4d0 | 1793 | if (!def->manual_alloc) { |
59169439 PB |
1794 | if (io_alloc_async_data(req)) |
1795 | return -EAGAIN; | |
1796 | } | |
f30bd4d0 | 1797 | return cdef->prep_async(req); |
bfe76559 PB |
1798 | } |
1799 | ||
9cf7c104 PB |
1800 | static u32 io_get_sequence(struct io_kiocb *req) |
1801 | { | |
a3dbdf54 | 1802 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1803 | struct io_kiocb *cur; |
9cf7c104 | 1804 | |
a3dbdf54 | 1805 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1806 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1807 | seq--; |
1808 | return seq; | |
9cf7c104 PB |
1809 | } |
1810 | ||
c072481d | 1811 | static __cold void io_drain_req(struct io_kiocb *req) |
e276ae34 | 1812 | __must_hold(&ctx->uring_lock) |
de0617e4 | 1813 | { |
a197f664 | 1814 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1815 | struct io_defer_entry *de; |
f67676d1 | 1816 | int ret; |
e0eb71dc | 1817 | u32 seq = io_get_sequence(req); |
3c19966d | 1818 | |
9d858b21 | 1819 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1820 | spin_lock(&ctx->completion_lock); |
5e371265 | 1821 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1822 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1823 | queue: |
10c66904 | 1824 | ctx->drain_active = false; |
e0eb71dc PB |
1825 | io_req_task_queue(req); |
1826 | return; | |
10c66904 | 1827 | } |
e302f104 | 1828 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1829 | |
cbdcb435 | 1830 | io_prep_async_link(req); |
27dc8338 | 1831 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1832 | if (!de) { |
1b48773f | 1833 | ret = -ENOMEM; |
ef5c600a DY |
1834 | io_req_defer_failed(req, ret); |
1835 | return; | |
76cc33d7 | 1836 | } |
2d28390a | 1837 | |
79ebeaee | 1838 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1839 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1840 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1841 | kfree(de); |
e0eb71dc | 1842 | goto queue; |
de0617e4 JA |
1843 | } |
1844 | ||
48863ffd | 1845 | trace_io_uring_defer(req); |
27dc8338 | 1846 | de->req = req; |
9cf7c104 | 1847 | de->seq = seq; |
27dc8338 | 1848 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1849 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1850 | } |
1851 | ||
68fb8979 | 1852 | static void io_clean_op(struct io_kiocb *req) |
99bc4c38 | 1853 | { |
8197b053 PB |
1854 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
1855 | spin_lock(&req->ctx->completion_lock); | |
cc3cec83 | 1856 | io_put_kbuf_comp(req); |
8197b053 PB |
1857 | spin_unlock(&req->ctx->completion_lock); |
1858 | } | |
99bc4c38 | 1859 | |
0e1b6fe3 | 1860 | if (req->flags & REQ_F_NEED_CLEANUP) { |
f30bd4d0 | 1861 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; |
bb040a21 | 1862 | |
4d4c9cff JA |
1863 | if (def->cleanup) |
1864 | def->cleanup(req); | |
99bc4c38 | 1865 | } |
75652a30 JA |
1866 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1867 | kfree(req->apoll->double_poll); | |
1868 | kfree(req->apoll); | |
1869 | req->apoll = NULL; | |
1870 | } | |
9cae36a0 JA |
1871 | if (req->flags & REQ_F_INFLIGHT) { |
1872 | struct io_uring_task *tctx = req->task->io_uring; | |
1873 | ||
1874 | atomic_dec(&tctx->inflight_tracked); | |
1875 | } | |
c854357b | 1876 | if (req->flags & REQ_F_CREDS) |
b8e64b53 | 1877 | put_cred(req->creds); |
d886e185 PB |
1878 | if (req->flags & REQ_F_ASYNC_DATA) { |
1879 | kfree(req->async_data); | |
1880 | req->async_data = NULL; | |
1881 | } | |
c854357b | 1882 | req->flags &= ~IO_REQ_CLEAN_FLAGS; |
99bc4c38 PB |
1883 | } |
1884 | ||
f4992544 JA |
1885 | static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def, |
1886 | unsigned int issue_flags) | |
6bf9c47a | 1887 | { |
f4992544 | 1888 | if (req->file || !def->needs_file) |
6bf9c47a JA |
1889 | return true; |
1890 | ||
1891 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1892 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1893 | else |
cef216fc | 1894 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1895 | |
772f5e00 | 1896 | return !!req->file; |
6bf9c47a JA |
1897 | } |
1898 | ||
889fca73 | 1899 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1900 | { |
a7dd2782 | 1901 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
5730b27e | 1902 | const struct cred *creds = NULL; |
d625c6ee | 1903 | int ret; |
2b188cc1 | 1904 | |
f4992544 | 1905 | if (unlikely(!io_assign_file(req, def, issue_flags))) |
70152140 JA |
1906 | return -EBADF; |
1907 | ||
6878b40e | 1908 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1909 | creds = override_creds(req->creds); |
5730b27e | 1910 | |
fcde59fe | 1911 | if (!def->audit_skip) |
5bd2182d PM |
1912 | audit_uring_entry(req->opcode); |
1913 | ||
0702e536 | 1914 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1915 | |
fcde59fe | 1916 | if (!def->audit_skip) |
5bd2182d PM |
1917 | audit_uring_exit(!ret, ret); |
1918 | ||
5730b27e JA |
1919 | if (creds) |
1920 | revert_creds(creds); | |
97b388d7 | 1921 | |
75d7b3ae PB |
1922 | if (ret == IOU_OK) { |
1923 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
9da070b1 | 1924 | io_req_complete_defer(req); |
75d7b3ae | 1925 | else |
1bec951c | 1926 | io_req_complete_post(req, issue_flags); |
75d7b3ae | 1927 | } else if (ret != IOU_ISSUE_SKIP_COMPLETE) |
def596e9 | 1928 | return ret; |
97b388d7 | 1929 | |
b532576e | 1930 | /* If the op doesn't have a file, we're not polling for it */ |
ef0ec1ad | 1931 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue) |
9882131c | 1932 | io_iopoll_req_issued(req, issue_flags); |
def596e9 JA |
1933 | |
1934 | return 0; | |
2b188cc1 JA |
1935 | } |
1936 | ||
329061d3 JA |
1937 | int io_poll_issue(struct io_kiocb *req, bool *locked) |
1938 | { | |
1939 | io_tw_lock(req->ctx, locked); | |
9a692451 DY |
1940 | return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT| |
1941 | IO_URING_F_COMPLETE_DEFER); | |
329061d3 JA |
1942 | } |
1943 | ||
c9f06aa7 | 1944 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1945 | { |
1946 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1947 | ||
1948 | req = io_put_req_find_next(req); | |
1949 | return req ? &req->work : NULL; | |
1950 | } | |
1951 | ||
c9f06aa7 | 1952 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1953 | { |
1954 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
a7dd2782 | 1955 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
e6aeb272 | 1956 | unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ; |
d01905db | 1957 | bool needs_poll = false; |
6bf9c47a | 1958 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1959 | |
23a6c9ac | 1960 | /* one will be dropped by ->io_wq_free_work() after returning to io-wq */ |
48dcd38d PB |
1961 | if (!(req->flags & REQ_F_REFCOUNT)) |
1962 | __io_req_set_refcount(req, 2); | |
1963 | else | |
1964 | req_ref_get(req); | |
5d5901a3 | 1965 | |
cb2d344c | 1966 | io_arm_ltimeout(req); |
6bf9c47a | 1967 | |
dadebc35 | 1968 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1969 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1970 | fail: |
6bf9c47a | 1971 | io_req_task_queue_fail(req, err); |
d01905db PB |
1972 | return; |
1973 | } | |
f4992544 | 1974 | if (!io_assign_file(req, def, issue_flags)) { |
0f8da75b PB |
1975 | err = -EBADF; |
1976 | work->flags |= IO_WQ_WORK_CANCEL; | |
1977 | goto fail; | |
1978 | } | |
31b51510 | 1979 | |
d01905db | 1980 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
1981 | bool opcode_poll = def->pollin || def->pollout; |
1982 | ||
1983 | if (opcode_poll && file_can_poll(req->file)) { | |
1984 | needs_poll = true; | |
d01905db | 1985 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 1986 | } |
561fb04a | 1987 | } |
31b51510 | 1988 | |
d01905db PB |
1989 | do { |
1990 | ret = io_issue_sqe(req, issue_flags); | |
1991 | if (ret != -EAGAIN) | |
1992 | break; | |
1993 | /* | |
1994 | * We can get EAGAIN for iopolled IO even though we're | |
1995 | * forcing a sync submission from here, since we can't | |
1996 | * wait for request slots on the block side. | |
1997 | */ | |
1998 | if (!needs_poll) { | |
e0deb6a0 PB |
1999 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
2000 | break; | |
d01905db PB |
2001 | cond_resched(); |
2002 | continue; | |
90fa0288 HX |
2003 | } |
2004 | ||
4d9237e3 | 2005 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
2006 | return; |
2007 | /* aborted or ready, in either case retry blocking */ | |
2008 | needs_poll = false; | |
2009 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
2010 | } while (1); | |
31b51510 | 2011 | |
a3df7698 | 2012 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 2013 | if (ret < 0) |
a3df7698 | 2014 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
2015 | } |
2016 | ||
531113bb JA |
2017 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
2018 | unsigned int issue_flags) | |
09bb8394 | 2019 | { |
5106dd6e JA |
2020 | struct io_ring_ctx *ctx = req->ctx; |
2021 | struct file *file = NULL; | |
ac177053 | 2022 | unsigned long file_ptr; |
09bb8394 | 2023 | |
93f052cb | 2024 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 2025 | |
ac177053 | 2026 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 2027 | goto out; |
ac177053 PB |
2028 | fd = array_index_nospec(fd, ctx->nr_user_files); |
2029 | file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; | |
2030 | file = (struct file *) (file_ptr & FFS_MASK); | |
2031 | file_ptr &= ~FFS_MASK; | |
2032 | /* mask in overlapping REQ_F and FFS bits */ | |
35645ac3 | 2033 | req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT); |
5106dd6e JA |
2034 | io_req_set_rsrc_node(req, ctx, 0); |
2035 | out: | |
93f052cb | 2036 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
2037 | return file; |
2038 | } | |
d44f554e | 2039 | |
531113bb | 2040 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 2041 | { |
62906e89 | 2042 | struct file *file = fget(fd); |
ac177053 | 2043 | |
48863ffd | 2044 | trace_io_uring_file_get(req, fd); |
09bb8394 | 2045 | |
ac177053 | 2046 | /* we don't allow fixed io_uring files */ |
e5550a14 | 2047 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 2048 | io_req_track_inflight(req); |
8371adf5 | 2049 | return file; |
09bb8394 JA |
2050 | } |
2051 | ||
7bfa9bad | 2052 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
2053 | __must_hold(&req->ctx->uring_lock) |
2054 | { | |
7bfa9bad PB |
2055 | struct io_kiocb *linked_timeout; |
2056 | ||
2057 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
973fc83f | 2058 | io_req_defer_failed(req, ret); |
7bfa9bad PB |
2059 | return; |
2060 | } | |
2061 | ||
2062 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 2063 | |
4d9237e3 | 2064 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 2065 | case IO_APOLL_READY: |
336d28a8 | 2066 | io_kbuf_recycle(req, 0); |
d475a9a6 PB |
2067 | io_req_task_queue(req); |
2068 | break; | |
2069 | case IO_APOLL_ABORTED: | |
6436c770 | 2070 | io_kbuf_recycle(req, 0); |
77955efb | 2071 | io_queue_iowq(req, NULL); |
d475a9a6 | 2072 | break; |
b1c62645 | 2073 | case IO_APOLL_OK: |
b1c62645 | 2074 | break; |
d475a9a6 PB |
2075 | } |
2076 | ||
2077 | if (linked_timeout) | |
2078 | io_queue_linked_timeout(linked_timeout); | |
2079 | } | |
2080 | ||
cbc2e203 | 2081 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 2082 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 2083 | { |
e0c5c576 | 2084 | int ret; |
2b188cc1 | 2085 | |
c5eef2b9 | 2086 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 2087 | |
491381ce JA |
2088 | /* |
2089 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2090 | * doesn't support non-blocking read/write attempts | |
2091 | */ | |
7bfa9bad | 2092 | if (likely(!ret)) |
cb2d344c | 2093 | io_arm_ltimeout(req); |
7bfa9bad PB |
2094 | else |
2095 | io_queue_async(req, ret); | |
2b188cc1 JA |
2096 | } |
2097 | ||
4652fe3f | 2098 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 2099 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 2100 | { |
17b147f6 PB |
2101 | if (unlikely(req->flags & REQ_F_FAIL)) { |
2102 | /* | |
2103 | * We don't submit, fail them all, for that replace hardlinks | |
2104 | * with normal links. Extra REQ_F_LINK is tolerated. | |
2105 | */ | |
2106 | req->flags &= ~REQ_F_HARDLINK; | |
2107 | req->flags |= REQ_F_LINK; | |
973fc83f | 2108 | io_req_defer_failed(req, req->cqe.res); |
76cc33d7 PB |
2109 | } else { |
2110 | int ret = io_req_prep_async(req); | |
2111 | ||
ef5c600a | 2112 | if (unlikely(ret)) { |
973fc83f | 2113 | io_req_defer_failed(req, ret); |
ef5c600a DY |
2114 | return; |
2115 | } | |
2116 | ||
2117 | if (unlikely(req->ctx->drain_active)) | |
2118 | io_drain_req(req); | |
76cc33d7 | 2119 | else |
77955efb | 2120 | io_queue_iowq(req, NULL); |
ce35a47a | 2121 | } |
4fe2c963 JL |
2122 | } |
2123 | ||
b16fed66 PB |
2124 | /* |
2125 | * Check SQE restrictions (opcode and flags). | |
2126 | * | |
2127 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
2128 | */ | |
2129 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
2130 | struct io_kiocb *req, | |
2131 | unsigned int sqe_flags) | |
4fe2c963 | 2132 | { |
b16fed66 PB |
2133 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
2134 | return false; | |
2135 | ||
2136 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
2137 | ctx->restrictions.sqe_flags_required) | |
2138 | return false; | |
2139 | ||
2140 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
2141 | ctx->restrictions.sqe_flags_required)) | |
2142 | return false; | |
2143 | ||
2144 | return true; | |
4fe2c963 JL |
2145 | } |
2146 | ||
22b2ca31 PB |
2147 | static void io_init_req_drain(struct io_kiocb *req) |
2148 | { | |
2149 | struct io_ring_ctx *ctx = req->ctx; | |
2150 | struct io_kiocb *head = ctx->submit_state.link.head; | |
2151 | ||
2152 | ctx->drain_active = true; | |
2153 | if (head) { | |
2154 | /* | |
2155 | * If we need to drain a request in the middle of a link, drain | |
2156 | * the head request and the next request/link after the current | |
2157 | * link. Considering sequential execution of links, | |
b6c7db32 | 2158 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
2159 | * link. |
2160 | */ | |
b6c7db32 | 2161 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
2162 | ctx->drain_next = true; |
2163 | } | |
2164 | } | |
2165 | ||
b16fed66 PB |
2166 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2167 | const struct io_uring_sqe *sqe) | |
282cdc86 | 2168 | __must_hold(&ctx->uring_lock) |
b16fed66 | 2169 | { |
a7dd2782 | 2170 | const struct io_issue_def *def; |
b16fed66 | 2171 | unsigned int sqe_flags; |
fc0ae024 | 2172 | int personality; |
4a04d1d1 | 2173 | u8 opcode; |
b16fed66 | 2174 | |
864ea921 | 2175 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 2176 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 PB |
2177 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
2178 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
cef216fc | 2179 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 2180 | req->file = NULL; |
c1bdf8ed | 2181 | req->rsrc_node = NULL; |
b16fed66 | 2182 | req->task = current; |
b16fed66 | 2183 | |
4a04d1d1 PB |
2184 | if (unlikely(opcode >= IORING_OP_LAST)) { |
2185 | req->opcode = 0; | |
b16fed66 | 2186 | return -EINVAL; |
4a04d1d1 | 2187 | } |
a7dd2782 | 2188 | def = &io_issue_defs[opcode]; |
68fe256a PB |
2189 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
2190 | /* enforce forwards compatibility on users */ | |
2191 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
2192 | return -EINVAL; | |
4e906702 | 2193 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 2194 | if (!def->buffer_select) |
4e906702 JA |
2195 | return -EOPNOTSUPP; |
2196 | req->buf_index = READ_ONCE(sqe->buf_group); | |
2197 | } | |
5562a8d7 PB |
2198 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
2199 | ctx->drain_disabled = true; | |
2200 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
2201 | if (ctx->drain_disabled) | |
2202 | return -EOPNOTSUPP; | |
22b2ca31 | 2203 | io_init_req_drain(req); |
5562a8d7 | 2204 | } |
2a56a9bd PB |
2205 | } |
2206 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
2207 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
2208 | return -EACCES; | |
2209 | /* knock it to the slow queue path, will be drained there */ | |
2210 | if (ctx->drain_active) | |
2211 | req->flags |= REQ_F_FORCE_ASYNC; | |
2212 | /* if there is no link, we're at "next" request and need to drain */ | |
2213 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
2214 | ctx->drain_next = false; | |
2215 | ctx->drain_active = true; | |
b6c7db32 | 2216 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 2217 | } |
68fe256a | 2218 | } |
b16fed66 | 2219 | |
fcde59fe | 2220 | if (!def->ioprio && sqe->ioprio) |
73911426 | 2221 | return -EINVAL; |
fcde59fe | 2222 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
73911426 JA |
2223 | return -EINVAL; |
2224 | ||
fcde59fe | 2225 | if (def->needs_file) { |
6d63416d PB |
2226 | struct io_submit_state *state = &ctx->submit_state; |
2227 | ||
cef216fc | 2228 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 2229 | |
6d63416d PB |
2230 | /* |
2231 | * Plug now if we have more than 2 IO left after this, and the | |
2232 | * target is potentially a read/write to block based storage. | |
2233 | */ | |
fcde59fe | 2234 | if (state->need_plug && def->plug) { |
6d63416d PB |
2235 | state->plug_started = true; |
2236 | state->need_plug = false; | |
5ca7a8b3 | 2237 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 2238 | } |
b16fed66 | 2239 | } |
863e0560 | 2240 | |
003e8dcc JA |
2241 | personality = READ_ONCE(sqe->personality); |
2242 | if (personality) { | |
cdab10bf LT |
2243 | int ret; |
2244 | ||
c10d1f98 PB |
2245 | req->creds = xa_load(&ctx->personalities, personality); |
2246 | if (!req->creds) | |
003e8dcc | 2247 | return -EINVAL; |
c10d1f98 | 2248 | get_cred(req->creds); |
cdc1404a PM |
2249 | ret = security_uring_override_creds(req->creds); |
2250 | if (ret) { | |
2251 | put_cred(req->creds); | |
2252 | return ret; | |
2253 | } | |
b8e64b53 | 2254 | req->flags |= REQ_F_CREDS; |
003e8dcc | 2255 | } |
b16fed66 | 2256 | |
0702e536 | 2257 | return def->prep(req, sqe); |
b16fed66 PB |
2258 | } |
2259 | ||
df3becde PB |
2260 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
2261 | struct io_kiocb *req, int ret) | |
2262 | { | |
2263 | struct io_ring_ctx *ctx = req->ctx; | |
2264 | struct io_submit_link *link = &ctx->submit_state.link; | |
2265 | struct io_kiocb *head = link->head; | |
2266 | ||
48863ffd | 2267 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
2268 | |
2269 | /* | |
2270 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
2271 | * unusable. Instead of failing eagerly, continue assembling the link if | |
2272 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
2273 | * should find the flag and handle the rest. | |
2274 | */ | |
2275 | req_fail_link_node(req, ret); | |
2276 | if (head && !(head->flags & REQ_F_FAIL)) | |
2277 | req_fail_link_node(head, -ECANCELED); | |
2278 | ||
2279 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
2280 | if (head) { | |
2281 | link->last->link = req; | |
2282 | link->head = NULL; | |
2283 | req = head; | |
2284 | } | |
2285 | io_queue_sqe_fallback(req); | |
2286 | return ret; | |
2287 | } | |
2288 | ||
2289 | if (head) | |
2290 | link->last->link = req; | |
2291 | else | |
2292 | link->head = req; | |
2293 | link->last = req; | |
2294 | return 0; | |
2295 | } | |
2296 | ||
2297 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 2298 | const struct io_uring_sqe *sqe) |
282cdc86 | 2299 | __must_hold(&ctx->uring_lock) |
9e645e11 | 2300 | { |
a1ab7b35 | 2301 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 2302 | int ret; |
9e645e11 | 2303 | |
a6b8cadc | 2304 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
2305 | if (unlikely(ret)) |
2306 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 2307 | |
be7053b7 | 2308 | /* don't need @sqe from now on */ |
48863ffd | 2309 | trace_io_uring_submit_sqe(req, true); |
a6b8cadc | 2310 | |
9e645e11 JA |
2311 | /* |
2312 | * If we already have a head request, queue this one for async | |
2313 | * submittal once the head completes. If we don't have a head but | |
2314 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2315 | * submitted sync once the chain is complete. If none of those | |
2316 | * conditions are true (normal request), then just queue it. | |
2317 | */ | |
924a07e4 | 2318 | if (unlikely(link->head)) { |
df3becde PB |
2319 | ret = io_req_prep_async(req); |
2320 | if (unlikely(ret)) | |
2321 | return io_submit_fail_init(sqe, req, ret); | |
2322 | ||
48863ffd | 2323 | trace_io_uring_link(req, link->head); |
f2f87370 | 2324 | link->last->link = req; |
863e0560 | 2325 | link->last = req; |
32fe525b | 2326 | |
da1a08c5 | 2327 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 2328 | return 0; |
df3becde PB |
2329 | /* last request of the link, flush it */ |
2330 | req = link->head; | |
f15a3431 | 2331 | link->head = NULL; |
924a07e4 PB |
2332 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
2333 | goto fallback; | |
2334 | ||
2335 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2336 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2337 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2338 | link->head = req; | |
2339 | link->last = req; | |
2340 | } else { | |
2341 | fallback: | |
2342 | io_queue_sqe_fallback(req); | |
2343 | } | |
f15a3431 | 2344 | return 0; |
9e645e11 | 2345 | } |
2e6e1fde | 2346 | |
f15a3431 | 2347 | io_queue_sqe(req); |
1d4240cc | 2348 | return 0; |
9e645e11 JA |
2349 | } |
2350 | ||
9a56a232 JA |
2351 | /* |
2352 | * Batched submission is done, ensure local IO is flushed out. | |
2353 | */ | |
553deffd | 2354 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2355 | { |
553deffd PB |
2356 | struct io_submit_state *state = &ctx->submit_state; |
2357 | ||
e126391c PB |
2358 | if (unlikely(state->link.head)) |
2359 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2360 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2361 | io_submit_flush_completions(ctx); |
27926b68 JA |
2362 | if (state->plug_started) |
2363 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2364 | } |
2365 | ||
2366 | /* | |
2367 | * Start submission side cache. | |
2368 | */ | |
2369 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2370 | unsigned int max_ios) |
9a56a232 | 2371 | { |
27926b68 | 2372 | state->plug_started = false; |
4b628aeb | 2373 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2374 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2375 | /* set only head, no need to init link_last in advance */ |
2376 | state->link.head = NULL; | |
9a56a232 JA |
2377 | } |
2378 | ||
2b188cc1 JA |
2379 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2380 | { | |
75b28aff | 2381 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2382 | |
caf582c6 PB |
2383 | /* |
2384 | * Ensure any loads from the SQEs are done at this point, | |
2385 | * since once we write the new head, the application could | |
2386 | * write new data to them. | |
2387 | */ | |
2388 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2389 | } |
2390 | ||
2b188cc1 | 2391 | /* |
dd9ae8a0 | 2392 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2393 | * that is mapped by userspace. This means that care needs to be taken to |
2394 | * ensure that reads are stable, as we cannot rely on userspace always | |
2395 | * being a good citizen. If members of the sqe are validated and then later | |
2396 | * used, it's important that those reads are done through READ_ONCE() to | |
2397 | * prevent a re-load down the line. | |
2398 | */ | |
b5083dfa | 2399 | static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe) |
2b188cc1 | 2400 | { |
ea5ab3b5 | 2401 | unsigned head, mask = ctx->sq_entries - 1; |
17d3aeb3 | 2402 | unsigned sq_idx = ctx->cached_sq_head++ & mask; |
2b188cc1 JA |
2403 | |
2404 | /* | |
2405 | * The cached sq head (or cq tail) serves two purposes: | |
2406 | * | |
2407 | * 1) allows us to batch the cost of updating the user visible | |
2408 | * head updates. | |
2409 | * 2) allows the kernel side to track the head on its own, even | |
2410 | * though the application is the one updating it. | |
2411 | */ | |
17d3aeb3 | 2412 | head = READ_ONCE(ctx->sq_array[sq_idx]); |
ebdeb7c0 JA |
2413 | if (likely(head < ctx->sq_entries)) { |
2414 | /* double index for 128-byte SQEs, twice as long */ | |
2415 | if (ctx->flags & IORING_SETUP_SQE128) | |
2416 | head <<= 1; | |
b5083dfa PB |
2417 | *sqe = &ctx->sq_sqes[head]; |
2418 | return true; | |
ebdeb7c0 | 2419 | } |
2b188cc1 JA |
2420 | |
2421 | /* drop invalid entries */ | |
15641e42 PB |
2422 | ctx->cq_extra--; |
2423 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2424 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
b5083dfa | 2425 | return false; |
709b302f PB |
2426 | } |
2427 | ||
17437f31 | 2428 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2429 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2430 | { |
69629809 | 2431 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2432 | unsigned int left; |
2433 | int ret; | |
6c271ce2 | 2434 | |
51d48dab | 2435 | if (unlikely(!entries)) |
69629809 | 2436 | return 0; |
ee7d46d9 | 2437 | /* make sure SQ entry isn't read before tail */ |
8e6971a8 PB |
2438 | ret = left = min3(nr, ctx->sq_entries, entries); |
2439 | io_get_task_refs(left); | |
2440 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2441 | |
69629809 | 2442 | do { |
3529d8c2 | 2443 | const struct io_uring_sqe *sqe; |
196be95c | 2444 | struct io_kiocb *req; |
fb5ccc98 | 2445 | |
c8576f3e | 2446 | if (unlikely(!io_alloc_req(ctx, &req))) |
fb5ccc98 | 2447 | break; |
b5083dfa | 2448 | if (unlikely(!io_get_sqe(ctx, &sqe))) { |
fa05457a | 2449 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2450 | break; |
2451 | } | |
6c271ce2 | 2452 | |
1cd15904 PB |
2453 | /* |
2454 | * Continue submitting even for sqe failure if the | |
2455 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2456 | */ | |
2457 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2458 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2459 | left--; | |
2460 | break; | |
bcbb7bf6 | 2461 | } |
1cd15904 | 2462 | } while (--left); |
9466f437 | 2463 | |
8e6971a8 PB |
2464 | if (unlikely(left)) { |
2465 | ret -= left; | |
2466 | /* try again if it submitted nothing and can't allocate a req */ | |
2467 | if (!ret && io_req_cache_empty(ctx)) | |
2468 | ret = -EAGAIN; | |
2469 | current->io_uring->cached_refs += left; | |
9466f437 | 2470 | } |
6c271ce2 | 2471 | |
553deffd | 2472 | io_submit_state_end(ctx); |
ae9428ca PB |
2473 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2474 | io_commit_sqring(ctx); | |
8e6971a8 | 2475 | return ret; |
6c271ce2 JA |
2476 | } |
2477 | ||
bda52162 JA |
2478 | struct io_wait_queue { |
2479 | struct wait_queue_entry wq; | |
2480 | struct io_ring_ctx *ctx; | |
5fd46178 | 2481 | unsigned cq_tail; |
bda52162 | 2482 | unsigned nr_timeouts; |
d33a39e5 | 2483 | ktime_t timeout; |
bda52162 JA |
2484 | }; |
2485 | ||
b4c98d59 DY |
2486 | static inline bool io_has_work(struct io_ring_ctx *ctx) |
2487 | { | |
c0e0d6ba | 2488 | return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) || |
490c00eb | 2489 | !llist_empty(&ctx->work_llist); |
b4c98d59 DY |
2490 | } |
2491 | ||
6c503150 | 2492 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
2493 | { |
2494 | struct io_ring_ctx *ctx = iowq->ctx; | |
0fc8c2ac | 2495 | int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail; |
bda52162 JA |
2496 | |
2497 | /* | |
d195a66e | 2498 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
2499 | * started waiting. For timeouts, we always want to return to userspace, |
2500 | * regardless of event count. | |
2501 | */ | |
5fd46178 | 2502 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
2503 | } |
2504 | ||
2505 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2506 | int wake_flags, void *key) | |
2507 | { | |
bd550173 | 2508 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq); |
bda52162 | 2509 | |
6c503150 PB |
2510 | /* |
2511 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2512 | * the task, and the next invocation will do it. | |
2513 | */ | |
bd550173 | 2514 | if (io_should_wake(iowq) || io_has_work(iowq->ctx)) |
6c503150 PB |
2515 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2516 | return -1; | |
bda52162 JA |
2517 | } |
2518 | ||
c0e0d6ba | 2519 | int io_run_task_work_sig(struct io_ring_ctx *ctx) |
af9c1a44 | 2520 | { |
1414d629 | 2521 | if (!llist_empty(&ctx->work_llist)) { |
2f413956 | 2522 | __set_current_state(TASK_RUNNING); |
1414d629 PB |
2523 | if (io_run_local_work(ctx) > 0) |
2524 | return 1; | |
2525 | } | |
2526 | if (io_run_task_work() > 0) | |
af9c1a44 | 2527 | return 1; |
c5020bc8 OL |
2528 | if (task_sigpending(current)) |
2529 | return -EINTR; | |
2530 | return 0; | |
af9c1a44 JA |
2531 | } |
2532 | ||
eeb60b9a PB |
2533 | /* when returns >0, the caller should retry */ |
2534 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
d33a39e5 | 2535 | struct io_wait_queue *iowq) |
eeb60b9a | 2536 | { |
3fcf19d5 PB |
2537 | if (unlikely(READ_ONCE(ctx->check_cq))) |
2538 | return 1; | |
846072f1 PB |
2539 | if (unlikely(!llist_empty(&ctx->work_llist))) |
2540 | return 1; | |
2541 | if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) | |
2542 | return 1; | |
2543 | if (unlikely(task_sigpending(current))) | |
2544 | return -EINTR; | |
2545 | if (unlikely(io_should_wake(iowq))) | |
2546 | return 0; | |
d33a39e5 | 2547 | if (iowq->timeout == KTIME_MAX) |
46ae7eef | 2548 | schedule(); |
d33a39e5 | 2549 | else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS)) |
22833966 | 2550 | return -ETIME; |
846072f1 | 2551 | return 0; |
eeb60b9a PB |
2552 | } |
2553 | ||
2b188cc1 JA |
2554 | /* |
2555 | * Wait until events become available, if we don't already have some. The | |
2556 | * application must reap them itself, as they reside on the shared cq ring. | |
2557 | */ | |
2558 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2559 | const sigset_t __user *sig, size_t sigsz, |
2560 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2561 | { |
90291099 | 2562 | struct io_wait_queue iowq; |
75b28aff | 2563 | struct io_rings *rings = ctx->rings; |
c1d5a224 | 2564 | int ret; |
2b188cc1 | 2565 | |
76de6749 PB |
2566 | if (!io_allowed_run_tw(ctx)) |
2567 | return -EEXIST; | |
140102ae PB |
2568 | if (!llist_empty(&ctx->work_llist)) |
2569 | io_run_local_work(ctx); | |
f36ba6cf PB |
2570 | io_run_task_work(); |
2571 | io_cqring_overflow_flush(ctx); | |
2572 | /* if user messes with these they will just get an early return */ | |
2573 | if (__io_cqring_events_user(ctx) >= min_events) | |
2574 | return 0; | |
2b188cc1 JA |
2575 | |
2576 | if (sig) { | |
9e75ad5d AB |
2577 | #ifdef CONFIG_COMPAT |
2578 | if (in_compat_syscall()) | |
2579 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2580 | sigsz); |
9e75ad5d AB |
2581 | else |
2582 | #endif | |
b772434b | 2583 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2584 | |
2b188cc1 JA |
2585 | if (ret) |
2586 | return ret; | |
2587 | } | |
2588 | ||
90291099 PB |
2589 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
2590 | iowq.wq.private = current; | |
2591 | INIT_LIST_HEAD(&iowq.wq.entry); | |
2592 | iowq.ctx = ctx; | |
bda52162 | 2593 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 2594 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
d33a39e5 PB |
2595 | iowq.timeout = KTIME_MAX; |
2596 | ||
2597 | if (uts) { | |
2598 | struct timespec64 ts; | |
2599 | ||
2600 | if (get_timespec64(&ts, uts)) | |
2601 | return -EFAULT; | |
2602 | iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); | |
2603 | } | |
90291099 | 2604 | |
c826bd7a | 2605 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2606 | do { |
3fcf19d5 PB |
2607 | unsigned long check_cq; |
2608 | ||
130bd686 | 2609 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
d80c0f00 | 2610 | WRITE_ONCE(ctx->cq_waiting, 1); |
130bd686 PB |
2611 | set_current_state(TASK_INTERRUPTIBLE); |
2612 | } else { | |
2613 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, | |
2614 | TASK_INTERRUPTIBLE); | |
2615 | } | |
2616 | ||
d33a39e5 | 2617 | ret = io_cqring_wait_schedule(ctx, &iowq); |
130bd686 | 2618 | __set_current_state(TASK_RUNNING); |
d80c0f00 PB |
2619 | WRITE_ONCE(ctx->cq_waiting, 0); |
2620 | ||
846072f1 PB |
2621 | if (ret < 0) |
2622 | break; | |
2623 | /* | |
2624 | * Run task_work after scheduling and before io_should_wake(). | |
2625 | * If we got woken because of task_work being processed, run it | |
2626 | * now rather than let the caller do another wait loop. | |
2627 | */ | |
2628 | io_run_task_work(); | |
2629 | if (!llist_empty(&ctx->work_llist)) | |
2630 | io_run_local_work(ctx); | |
3fcf19d5 PB |
2631 | |
2632 | check_cq = READ_ONCE(ctx->check_cq); | |
2633 | if (unlikely(check_cq)) { | |
2634 | /* let the caller flush overflows, retry */ | |
326a9e48 | 2635 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) |
3fcf19d5 | 2636 | io_cqring_do_overflow_flush(ctx); |
3fcf19d5 PB |
2637 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) { |
2638 | ret = -EBADR; | |
2639 | break; | |
2640 | } | |
2641 | } | |
2642 | ||
846072f1 PB |
2643 | if (io_should_wake(&iowq)) { |
2644 | ret = 0; | |
35d90f95 | 2645 | break; |
846072f1 | 2646 | } |
ca0a2651 | 2647 | cond_resched(); |
846072f1 | 2648 | } while (1); |
bda52162 | 2649 | |
130bd686 PB |
2650 | if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) |
2651 | finish_wait(&ctx->cq_wait, &iowq.wq); | |
b7db41c9 | 2652 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2653 | |
75b28aff | 2654 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2655 | } |
2656 | ||
73572984 | 2657 | static void io_mem_free(void *ptr) |
b9bd2bea | 2658 | { |
73572984 | 2659 | struct page *page; |
b36a2050 | 2660 | |
73572984 JA |
2661 | if (!ptr) |
2662 | return; | |
b9bd2bea | 2663 | |
73572984 JA |
2664 | page = virt_to_head_page(ptr); |
2665 | if (put_page_testzero(page)) | |
2666 | free_compound_page(page); | |
b9bd2bea PB |
2667 | } |
2668 | ||
73572984 | 2669 | static void *io_mem_alloc(size_t size) |
b9bd2bea | 2670 | { |
73572984 | 2671 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; |
b9bd2bea | 2672 | |
73572984 | 2673 | return (void *) __get_free_pages(gfp, get_order(size)); |
b9bd2bea PB |
2674 | } |
2675 | ||
73572984 JA |
2676 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2677 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2678 | { |
73572984 JA |
2679 | struct io_rings *rings; |
2680 | size_t off, sq_array_size; | |
6b06314c | 2681 | |
73572984 JA |
2682 | off = struct_size(rings, cqes, cq_entries); |
2683 | if (off == SIZE_MAX) | |
2684 | return SIZE_MAX; | |
2685 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2686 | if (check_shl_overflow(off, 1, &off)) | |
2687 | return SIZE_MAX; | |
2688 | } | |
ab409402 | 2689 | |
73572984 JA |
2690 | #ifdef CONFIG_SMP |
2691 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2692 | if (off == 0) | |
2693 | return SIZE_MAX; | |
2694 | #endif | |
82fbcfa9 | 2695 | |
73572984 JA |
2696 | if (sq_offset) |
2697 | *sq_offset = off; | |
82fbcfa9 | 2698 | |
73572984 JA |
2699 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2700 | if (sq_array_size == SIZE_MAX) | |
2701 | return SIZE_MAX; | |
6b06314c | 2702 | |
73572984 JA |
2703 | if (check_add_overflow(off, sq_array_size, &off)) |
2704 | return SIZE_MAX; | |
8bad28d8 | 2705 | |
73572984 | 2706 | return off; |
8bad28d8 HX |
2707 | } |
2708 | ||
73572984 JA |
2709 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, |
2710 | unsigned int eventfd_async) | |
8bad28d8 | 2711 | { |
73572984 JA |
2712 | struct io_ev_fd *ev_fd; |
2713 | __s32 __user *fds = arg; | |
2714 | int fd; | |
f2303b1f | 2715 | |
73572984 JA |
2716 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, |
2717 | lockdep_is_held(&ctx->uring_lock)); | |
2718 | if (ev_fd) | |
2719 | return -EBUSY; | |
8bad28d8 | 2720 | |
73572984 JA |
2721 | if (copy_from_user(&fd, fds, sizeof(*fds))) |
2722 | return -EFAULT; | |
8dd03afe | 2723 | |
73572984 JA |
2724 | ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); |
2725 | if (!ev_fd) | |
2726 | return -ENOMEM; | |
05f3fb3c | 2727 | |
73572984 JA |
2728 | ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); |
2729 | if (IS_ERR(ev_fd->cq_ev_fd)) { | |
2730 | int ret = PTR_ERR(ev_fd->cq_ev_fd); | |
2731 | kfree(ev_fd); | |
2732 | return ret; | |
2733 | } | |
305bef98 PB |
2734 | |
2735 | spin_lock(&ctx->completion_lock); | |
2736 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
2737 | spin_unlock(&ctx->completion_lock); | |
2738 | ||
73572984 JA |
2739 | ev_fd->eventfd_async = eventfd_async; |
2740 | ctx->has_evfd = true; | |
2741 | rcu_assign_pointer(ctx->io_ev_fd, ev_fd); | |
21a091b9 DY |
2742 | atomic_set(&ev_fd->refs, 1); |
2743 | atomic_set(&ev_fd->ops, 0); | |
73572984 | 2744 | return 0; |
d7954b2b BM |
2745 | } |
2746 | ||
73572984 | 2747 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 2748 | { |
73572984 JA |
2749 | struct io_ev_fd *ev_fd; |
2750 | ||
2751 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, | |
2752 | lockdep_is_held(&ctx->uring_lock)); | |
2753 | if (ev_fd) { | |
2754 | ctx->has_evfd = false; | |
2755 | rcu_assign_pointer(ctx->io_ev_fd, NULL); | |
21a091b9 DY |
2756 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops)) |
2757 | call_rcu(&ev_fd->rcu, io_eventfd_ops); | |
73572984 JA |
2758 | return 0; |
2759 | } | |
2d091d62 | 2760 | |
73572984 | 2761 | return -ENXIO; |
44b31f2f PB |
2762 | } |
2763 | ||
73572984 | 2764 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2765 | { |
c8576f3e | 2766 | struct io_kiocb *req; |
37f0e767 | 2767 | int nr = 0; |
bf019da7 | 2768 | |
9a4fdbd8 | 2769 | mutex_lock(&ctx->uring_lock); |
34f0bc42 | 2770 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
9a4fdbd8 | 2771 | |
88ab95be | 2772 | while (!io_req_cache_empty(ctx)) { |
c8576f3e | 2773 | req = io_extract_req(ctx); |
c2b6c6bc | 2774 | kmem_cache_free(req_cachep, req); |
37f0e767 | 2775 | nr++; |
c2b6c6bc | 2776 | } |
37f0e767 PB |
2777 | if (nr) |
2778 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2779 | mutex_unlock(&ctx->uring_lock); |
2780 | } | |
2781 | ||
c072481d | 2782 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2783 | { |
37d1e2e3 | 2784 | io_sq_thread_finish(ctx); |
ab409402 | 2785 | io_rsrc_refs_drop(ctx); |
43597aac PB |
2786 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
2787 | io_wait_rsrc_data(ctx->buf_data); | |
2788 | io_wait_rsrc_data(ctx->file_data); | |
2789 | ||
8bad28d8 | 2790 | mutex_lock(&ctx->uring_lock); |
43597aac | 2791 | if (ctx->buf_data) |
bd54b6fe | 2792 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2793 | if (ctx->file_data) |
08480400 | 2794 | __io_sqe_files_unregister(ctx); |
a85381d8 | 2795 | io_cqring_overflow_kill(ctx); |
9b402849 | 2796 | io_eventfd_unregister(ctx); |
9b797a37 | 2797 | io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); |
43e0bbbd | 2798 | io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); |
77bc59b4 | 2799 | mutex_unlock(&ctx->uring_lock); |
5a2e745d | 2800 | io_destroy_buffers(ctx); |
07db298a PB |
2801 | if (ctx->sq_creds) |
2802 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2803 | if (ctx->submitter_task) |
2804 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2805 | |
a7f0ed5a PB |
2806 | /* there are no registered resources left, nobody uses it */ |
2807 | if (ctx->rsrc_node) | |
2808 | io_rsrc_node_destroy(ctx->rsrc_node); | |
8dd03afe | 2809 | if (ctx->rsrc_backup_node) |
b895c9a6 | 2810 | io_rsrc_node_destroy(ctx->rsrc_backup_node); |
a7f0ed5a | 2811 | flush_delayed_work(&ctx->rsrc_put_work); |
756ab7c0 | 2812 | flush_delayed_work(&ctx->fallback_work); |
a7f0ed5a PB |
2813 | |
2814 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
2815 | WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist)); | |
def596e9 | 2816 | |
2b188cc1 | 2817 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
2818 | if (ctx->ring_sock) { |
2819 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 2820 | sock_release(ctx->ring_sock); |
355e8d26 | 2821 | } |
2b188cc1 | 2822 | #endif |
ef9dd637 | 2823 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
2b188cc1 | 2824 | |
42b6419d PB |
2825 | if (ctx->mm_account) { |
2826 | mmdrop(ctx->mm_account); | |
2827 | ctx->mm_account = NULL; | |
2828 | } | |
75b28aff | 2829 | io_mem_free(ctx->rings); |
2b188cc1 | 2830 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
2831 | |
2832 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2833 | free_uid(ctx->user); |
4010fec4 | 2834 | io_req_caches_free(ctx); |
e941894e JA |
2835 | if (ctx->hash_map) |
2836 | io_wq_put_hash(ctx->hash_map); | |
e6f89be6 | 2837 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2838 | kfree(ctx->cancel_table_locked.hbs); |
6224843d | 2839 | kfree(ctx->dummy_ubuf); |
9cfc7e94 JA |
2840 | kfree(ctx->io_bl); |
2841 | xa_destroy(&ctx->io_bl_xa); | |
2b188cc1 JA |
2842 | kfree(ctx); |
2843 | } | |
2844 | ||
bca39f39 PB |
2845 | static __cold void io_activate_pollwq_cb(struct callback_head *cb) |
2846 | { | |
2847 | struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx, | |
2848 | poll_wq_task_work); | |
2849 | ||
2850 | mutex_lock(&ctx->uring_lock); | |
2851 | ctx->poll_activated = true; | |
2852 | mutex_unlock(&ctx->uring_lock); | |
2853 | ||
2854 | /* | |
2855 | * Wake ups for some events between start of polling and activation | |
2856 | * might've been lost due to loose synchronisation. | |
2857 | */ | |
2858 | wake_up_all(&ctx->poll_wq); | |
2859 | percpu_ref_put(&ctx->refs); | |
2860 | } | |
2861 | ||
2862 | static __cold void io_activate_pollwq(struct io_ring_ctx *ctx) | |
2863 | { | |
2864 | spin_lock(&ctx->completion_lock); | |
2865 | /* already activated or in progress */ | |
2866 | if (ctx->poll_activated || ctx->poll_wq_task_work.func) | |
2867 | goto out; | |
2868 | if (WARN_ON_ONCE(!ctx->task_complete)) | |
2869 | goto out; | |
2870 | if (!ctx->submitter_task) | |
2871 | goto out; | |
2872 | /* | |
2873 | * with ->submitter_task only the submitter task completes requests, we | |
2874 | * only need to sync with it, which is done by injecting a tw | |
2875 | */ | |
2876 | init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb); | |
2877 | percpu_ref_get(&ctx->refs); | |
2878 | if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL)) | |
2879 | percpu_ref_put(&ctx->refs); | |
2880 | out: | |
2881 | spin_unlock(&ctx->completion_lock); | |
2882 | } | |
2883 | ||
2b188cc1 JA |
2884 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) |
2885 | { | |
2886 | struct io_ring_ctx *ctx = file->private_data; | |
2887 | __poll_t mask = 0; | |
2888 | ||
bca39f39 PB |
2889 | if (unlikely(!ctx->poll_activated)) |
2890 | io_activate_pollwq(ctx); | |
2891 | ||
7b235dd8 | 2892 | poll_wait(file, &ctx->poll_wq, wait); |
4f7067c3 SB |
2893 | /* |
2894 | * synchronizes with barrier from wq_has_sleeper call in | |
2895 | * io_commit_cqring | |
2896 | */ | |
2b188cc1 | 2897 | smp_rmb(); |
90554200 | 2898 | if (!io_sqring_full(ctx)) |
2b188cc1 | 2899 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
2900 | |
2901 | /* | |
2902 | * Don't flush cqring overflow list here, just do a simple check. | |
2903 | * Otherwise there could possible be ABBA deadlock: | |
2904 | * CPU0 CPU1 | |
2905 | * ---- ---- | |
2906 | * lock(&ctx->uring_lock); | |
2907 | * lock(&ep->mtx); | |
2908 | * lock(&ctx->uring_lock); | |
2909 | * lock(&ep->mtx); | |
2910 | * | |
2911 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
10d8bc35 | 2912 | * pushes them to do the flush. |
ed670c3f | 2913 | */ |
b4c98d59 | 2914 | |
c10bb646 | 2915 | if (__io_cqring_events_user(ctx) || io_has_work(ctx)) |
2b188cc1 JA |
2916 | mask |= EPOLLIN | EPOLLRDNORM; |
2917 | ||
2918 | return mask; | |
2919 | } | |
2920 | ||
0bead8cd | 2921 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 2922 | { |
4379bf8b | 2923 | const struct cred *creds; |
071698e1 | 2924 | |
61cf9370 | 2925 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
2926 | if (creds) { |
2927 | put_cred(creds); | |
0bead8cd | 2928 | return 0; |
1e6fa521 | 2929 | } |
0bead8cd YD |
2930 | |
2931 | return -EINVAL; | |
2932 | } | |
2933 | ||
d56d938b PB |
2934 | struct io_tctx_exit { |
2935 | struct callback_head task_work; | |
2936 | struct completion completion; | |
baf186c4 | 2937 | struct io_ring_ctx *ctx; |
d56d938b PB |
2938 | }; |
2939 | ||
c072481d | 2940 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
2941 | { |
2942 | struct io_uring_task *tctx = current->io_uring; | |
2943 | struct io_tctx_exit *work; | |
2944 | ||
2945 | work = container_of(cb, struct io_tctx_exit, task_work); | |
2946 | /* | |
8d664282 | 2947 | * When @in_cancel, we're in cancellation and it's racy to remove the |
d56d938b | 2948 | * node. It'll be removed by the end of cancellation, just ignore it. |
998b30c3 HM |
2949 | * tctx can be NULL if the queueing of this task_work raced with |
2950 | * work cancelation off the exec path. | |
d56d938b | 2951 | */ |
8d664282 | 2952 | if (tctx && !atomic_read(&tctx->in_cancel)) |
eef51daa | 2953 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
2954 | complete(&work->completion); |
2955 | } | |
2956 | ||
c072481d | 2957 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
2958 | { |
2959 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2960 | ||
2961 | return req->ctx == data; | |
2962 | } | |
2963 | ||
c072481d | 2964 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 2965 | { |
d56d938b | 2966 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 2967 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 2968 | unsigned long interval = HZ / 20; |
d56d938b PB |
2969 | struct io_tctx_exit exit; |
2970 | struct io_tctx_node *node; | |
2971 | int ret; | |
85faa7b8 | 2972 | |
56952e91 JA |
2973 | /* |
2974 | * If we're doing polled IO and end up having requests being | |
2975 | * submitted async (out-of-line), then completions can come in while | |
2976 | * we're waiting for refs to drop. We need to reap these manually, | |
2977 | * as nobody else will be looking for them. | |
2978 | */ | |
b2edc0a7 | 2979 | do { |
a85381d8 PB |
2980 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
2981 | mutex_lock(&ctx->uring_lock); | |
2982 | io_cqring_overflow_kill(ctx); | |
2983 | mutex_unlock(&ctx->uring_lock); | |
2984 | } | |
2985 | ||
c0e0d6ba DY |
2986 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) |
2987 | io_move_task_work_from_local(ctx); | |
2988 | ||
affa87db PB |
2989 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
2990 | cond_resched(); | |
2991 | ||
28090c13 PB |
2992 | if (ctx->sq_data) { |
2993 | struct io_sq_data *sqd = ctx->sq_data; | |
2994 | struct task_struct *tsk; | |
2995 | ||
2996 | io_sq_thread_park(sqd); | |
2997 | tsk = sqd->thread; | |
2998 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
2999 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
3000 | io_cancel_ctx_cb, ctx, true); | |
3001 | io_sq_thread_unpark(sqd); | |
3002 | } | |
b5bb3a24 | 3003 | |
37f0e767 PB |
3004 | io_req_caches_free(ctx); |
3005 | ||
58d3be2c PB |
3006 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
3007 | /* there is little hope left, don't run it too often */ | |
3008 | interval = HZ * 60; | |
3009 | } | |
3010 | } while (!wait_for_completion_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 3011 | |
7f00651a PB |
3012 | init_completion(&exit.completion); |
3013 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
3014 | exit.ctx = ctx; | |
89b5066e PB |
3015 | /* |
3016 | * Some may use context even when all refs and requests have been put, | |
3017 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 3018 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
3019 | * this lock/unlock section also waits them to finish. |
3020 | */ | |
d56d938b PB |
3021 | mutex_lock(&ctx->uring_lock); |
3022 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
3023 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
3024 | ||
d56d938b PB |
3025 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
3026 | ctx_node); | |
7f00651a PB |
3027 | /* don't spin on a single task if cancellation failed */ |
3028 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
3029 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
3030 | if (WARN_ON_ONCE(ret)) | |
3031 | continue; | |
d56d938b PB |
3032 | |
3033 | mutex_unlock(&ctx->uring_lock); | |
3034 | wait_for_completion(&exit.completion); | |
d56d938b PB |
3035 | mutex_lock(&ctx->uring_lock); |
3036 | } | |
3037 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
3038 | spin_lock(&ctx->completion_lock); |
3039 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 3040 | |
85faa7b8 JA |
3041 | io_ring_ctx_free(ctx); |
3042 | } | |
3043 | ||
c072481d | 3044 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 3045 | { |
61cf9370 MWO |
3046 | unsigned long index; |
3047 | struct creds *creds; | |
3048 | ||
2b188cc1 JA |
3049 | mutex_lock(&ctx->uring_lock); |
3050 | percpu_ref_kill(&ctx->refs); | |
61cf9370 MWO |
3051 | xa_for_each(&ctx->personalities, index, creds) |
3052 | io_unregister_personality(ctx, index); | |
9ca9fb24 PB |
3053 | if (ctx->rings) |
3054 | io_poll_remove_all(ctx, NULL, true); | |
2b188cc1 JA |
3055 | mutex_unlock(&ctx->uring_lock); |
3056 | ||
02bac94b PB |
3057 | /* |
3058 | * If we failed setting up the ctx, we might not have any rings | |
3059 | * and therefore did not submit any requests | |
3060 | */ | |
3061 | if (ctx->rings) | |
60053be8 | 3062 | io_kill_timeouts(ctx, NULL, true); |
309fc03a | 3063 | |
85faa7b8 | 3064 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
3065 | /* |
3066 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
3067 | * if we're exiting a ton of rings at the same time. It just adds | |
3068 | * noise and overhead, there's no discernable change in runtime | |
3069 | * over using system_wq. | |
3070 | */ | |
3071 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
3072 | } |
3073 | ||
3074 | static int io_uring_release(struct inode *inode, struct file *file) | |
3075 | { | |
3076 | struct io_ring_ctx *ctx = file->private_data; | |
3077 | ||
3078 | file->private_data = NULL; | |
3079 | io_ring_ctx_wait_and_kill(ctx); | |
3080 | return 0; | |
3081 | } | |
3082 | ||
f6edbabb PB |
3083 | struct io_task_cancel { |
3084 | struct task_struct *task; | |
3dd0c97a | 3085 | bool all; |
f6edbabb | 3086 | }; |
f254ac04 | 3087 | |
f6edbabb | 3088 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 3089 | { |
9a472ef7 | 3090 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 3091 | struct io_task_cancel *cancel = data; |
9a472ef7 | 3092 | |
6af3f48b | 3093 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
3094 | } |
3095 | ||
c072481d PB |
3096 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
3097 | struct task_struct *task, | |
3098 | bool cancel_all) | |
b7ddce3c | 3099 | { |
e1915f76 | 3100 | struct io_defer_entry *de; |
b7ddce3c PB |
3101 | LIST_HEAD(list); |
3102 | ||
79ebeaee | 3103 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 3104 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 3105 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
3106 | list_cut_position(&list, &ctx->defer_list, &de->list); |
3107 | break; | |
3108 | } | |
3109 | } | |
79ebeaee | 3110 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
3111 | if (list_empty(&list)) |
3112 | return false; | |
b7ddce3c PB |
3113 | |
3114 | while (!list_empty(&list)) { | |
3115 | de = list_first_entry(&list, struct io_defer_entry, list); | |
3116 | list_del_init(&de->list); | |
e276ae34 | 3117 | io_req_task_queue_fail(de->req, -ECANCELED); |
b7ddce3c PB |
3118 | kfree(de); |
3119 | } | |
e1915f76 | 3120 | return true; |
b7ddce3c PB |
3121 | } |
3122 | ||
c072481d | 3123 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
3124 | { |
3125 | struct io_tctx_node *node; | |
3126 | enum io_wq_cancel cret; | |
3127 | bool ret = false; | |
3128 | ||
3129 | mutex_lock(&ctx->uring_lock); | |
3130 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3131 | struct io_uring_task *tctx = node->task->io_uring; | |
3132 | ||
3133 | /* | |
3134 | * io_wq will stay alive while we hold uring_lock, because it's | |
3135 | * killed after ctx nodes, which requires to take the lock. | |
3136 | */ | |
3137 | if (!tctx || !tctx->io_wq) | |
3138 | continue; | |
3139 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
3140 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3141 | } | |
3142 | mutex_unlock(&ctx->uring_lock); | |
3143 | ||
3144 | return ret; | |
3145 | } | |
3146 | ||
affa87db | 3147 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
3148 | struct task_struct *task, |
3149 | bool cancel_all) | |
9936c7c2 | 3150 | { |
3dd0c97a | 3151 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 3152 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
3153 | enum io_wq_cancel cret; |
3154 | bool ret = false; | |
9936c7c2 | 3155 | |
60053be8 PB |
3156 | /* failed during ring init, it couldn't have issued any requests */ |
3157 | if (!ctx->rings) | |
affa87db | 3158 | return false; |
60053be8 | 3159 | |
affa87db PB |
3160 | if (!task) { |
3161 | ret |= io_uring_try_cancel_iowq(ctx); | |
3162 | } else if (tctx && tctx->io_wq) { | |
3163 | /* | |
3164 | * Cancels requests of all rings, not only @ctx, but | |
3165 | * it's fine as the task is in exit/exec. | |
3166 | */ | |
3167 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
3168 | &cancel, true); | |
3169 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3170 | } | |
9936c7c2 | 3171 | |
affa87db PB |
3172 | /* SQPOLL thread does its own polling */ |
3173 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
3174 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
3175 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
3176 | io_iopoll_try_reap_events(ctx); | |
3177 | ret = true; | |
fcc926bb | 3178 | cond_resched(); |
9936c7c2 | 3179 | } |
9936c7c2 | 3180 | } |
affa87db | 3181 | |
140102ae PB |
3182 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3183 | io_allowed_defer_tw_run(ctx)) | |
c0e0d6ba | 3184 | ret |= io_run_local_work(ctx) > 0; |
affa87db PB |
3185 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
3186 | mutex_lock(&ctx->uring_lock); | |
3187 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
3188 | mutex_unlock(&ctx->uring_lock); | |
3189 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
3190 | if (task) | |
c0e0d6ba | 3191 | ret |= io_run_task_work() > 0; |
affa87db | 3192 | return ret; |
9936c7c2 PB |
3193 | } |
3194 | ||
3f48cf18 | 3195 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 3196 | { |
3f48cf18 | 3197 | if (tracked) |
9cae36a0 | 3198 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
3199 | return percpu_counter_sum(&tctx->inflight); |
3200 | } | |
3201 | ||
78cc687b PB |
3202 | /* |
3203 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 3204 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 3205 | */ |
17437f31 | 3206 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 3207 | { |
521d6a73 | 3208 | struct io_uring_task *tctx = current->io_uring; |
734551df | 3209 | struct io_ring_ctx *ctx; |
0e9ddb39 PB |
3210 | s64 inflight; |
3211 | DEFINE_WAIT(wait); | |
fdaf083c | 3212 | |
78cc687b PB |
3213 | WARN_ON_ONCE(sqd && sqd->thread != current); |
3214 | ||
6d042ffb PO |
3215 | if (!current->io_uring) |
3216 | return; | |
17a91051 PB |
3217 | if (tctx->io_wq) |
3218 | io_wq_exit_start(tctx->io_wq); | |
3219 | ||
8d664282 | 3220 | atomic_inc(&tctx->in_cancel); |
0e9ddb39 | 3221 | do { |
affa87db PB |
3222 | bool loop = false; |
3223 | ||
e9dbe221 | 3224 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 3225 | /* read completions before cancelations */ |
78cc687b | 3226 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
3227 | if (!inflight) |
3228 | break; | |
fdaf083c | 3229 | |
78cc687b PB |
3230 | if (!sqd) { |
3231 | struct io_tctx_node *node; | |
3232 | unsigned long index; | |
0f212204 | 3233 | |
78cc687b PB |
3234 | xa_for_each(&tctx->xa, index, node) { |
3235 | /* sqpoll task will cancel all its requests */ | |
3236 | if (node->ctx->sq_data) | |
3237 | continue; | |
affa87db PB |
3238 | loop |= io_uring_try_cancel_requests(node->ctx, |
3239 | current, cancel_all); | |
78cc687b PB |
3240 | } |
3241 | } else { | |
3242 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
3243 | loop |= io_uring_try_cancel_requests(ctx, |
3244 | current, | |
3245 | cancel_all); | |
3246 | } | |
3247 | ||
3248 | if (loop) { | |
3249 | cond_resched(); | |
3250 | continue; | |
78cc687b | 3251 | } |
17a91051 | 3252 | |
78a78060 JA |
3253 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
3254 | io_run_task_work(); | |
e9dbe221 | 3255 | io_uring_drop_tctx_refs(current); |
78a78060 | 3256 | |
0f212204 | 3257 | /* |
a1bb3cd5 PB |
3258 | * If we've seen completions, retry without waiting. This |
3259 | * avoids a race where a completion comes in before we did | |
3260 | * prepare_to_wait(). | |
0f212204 | 3261 | */ |
3dd0c97a | 3262 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 3263 | schedule(); |
f57555ed | 3264 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 3265 | } while (1); |
de7f1d9e | 3266 | |
8452d4a6 | 3267 | io_uring_clean_tctx(tctx); |
3dd0c97a | 3268 | if (cancel_all) { |
3cc7fdb9 PB |
3269 | /* |
3270 | * We shouldn't run task_works after cancel, so just leave | |
8d664282 | 3271 | * ->in_cancel set for normal exit. |
3cc7fdb9 | 3272 | */ |
8d664282 | 3273 | atomic_dec(&tctx->in_cancel); |
3f48cf18 PB |
3274 | /* for exec all current's requests should be gone, kill tctx */ |
3275 | __io_uring_free(current); | |
3276 | } | |
44e728b8 PB |
3277 | } |
3278 | ||
f552a27a | 3279 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 3280 | { |
f552a27a | 3281 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
3282 | } |
3283 | ||
6c5c240e RP |
3284 | static void *io_uring_validate_mmap_request(struct file *file, |
3285 | loff_t pgoff, size_t sz) | |
2b188cc1 | 3286 | { |
2b188cc1 | 3287 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 3288 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
3289 | struct page *page; |
3290 | void *ptr; | |
3291 | ||
c56e022c | 3292 | switch (offset & IORING_OFF_MMAP_MASK) { |
2b188cc1 | 3293 | case IORING_OFF_SQ_RING: |
75b28aff HV |
3294 | case IORING_OFF_CQ_RING: |
3295 | ptr = ctx->rings; | |
2b188cc1 JA |
3296 | break; |
3297 | case IORING_OFF_SQES: | |
3298 | ptr = ctx->sq_sqes; | |
3299 | break; | |
c56e022c JA |
3300 | case IORING_OFF_PBUF_RING: { |
3301 | unsigned int bgid; | |
3302 | ||
3303 | bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT; | |
3304 | mutex_lock(&ctx->uring_lock); | |
3305 | ptr = io_pbuf_get_address(ctx, bgid); | |
3306 | mutex_unlock(&ctx->uring_lock); | |
3307 | if (!ptr) | |
3308 | return ERR_PTR(-EINVAL); | |
3309 | break; | |
3310 | } | |
2b188cc1 | 3311 | default: |
6c5c240e | 3312 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
3313 | } |
3314 | ||
3315 | page = virt_to_head_page(ptr); | |
a50b854e | 3316 | if (sz > page_size(page)) |
6c5c240e RP |
3317 | return ERR_PTR(-EINVAL); |
3318 | ||
3319 | return ptr; | |
3320 | } | |
3321 | ||
3322 | #ifdef CONFIG_MMU | |
3323 | ||
c072481d | 3324 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e RP |
3325 | { |
3326 | size_t sz = vma->vm_end - vma->vm_start; | |
3327 | unsigned long pfn; | |
3328 | void *ptr; | |
3329 | ||
3330 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
3331 | if (IS_ERR(ptr)) | |
3332 | return PTR_ERR(ptr); | |
2b188cc1 JA |
3333 | |
3334 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3335 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3336 | } | |
3337 | ||
d808459b HD |
3338 | static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp, |
3339 | unsigned long addr, unsigned long len, | |
3340 | unsigned long pgoff, unsigned long flags) | |
3341 | { | |
3342 | const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags); | |
3343 | struct vm_unmapped_area_info info; | |
3344 | void *ptr; | |
3345 | ||
3346 | /* | |
3347 | * Do not allow to map to user-provided address to avoid breaking the | |
3348 | * aliasing rules. Userspace is not able to guess the offset address of | |
3349 | * kernel kmalloc()ed memory area. | |
3350 | */ | |
3351 | if (addr) | |
3352 | return -EINVAL; | |
3353 | ||
3354 | ptr = io_uring_validate_mmap_request(filp, pgoff, len); | |
3355 | if (IS_ERR(ptr)) | |
3356 | return -ENOMEM; | |
3357 | ||
3358 | info.flags = VM_UNMAPPED_AREA_TOPDOWN; | |
3359 | info.length = len; | |
3360 | info.low_limit = max(PAGE_SIZE, mmap_min_addr); | |
3361 | info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base); | |
3362 | #ifdef SHM_COLOUR | |
3363 | info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL); | |
3364 | #else | |
3365 | info.align_mask = PAGE_MASK & (SHMLBA - 1UL); | |
3366 | #endif | |
3367 | info.align_offset = (unsigned long) ptr; | |
3368 | ||
3369 | /* | |
3370 | * A failed mmap() very likely causes application failure, | |
3371 | * so fall back to the bottom-up function here. This scenario | |
3372 | * can happen with large stack limits and large mmap() | |
3373 | * allocations. | |
3374 | */ | |
3375 | addr = vm_unmapped_area(&info); | |
3376 | if (offset_in_page(addr)) { | |
3377 | info.flags = 0; | |
3378 | info.low_limit = TASK_UNMAPPED_BASE; | |
3379 | info.high_limit = mmap_end; | |
3380 | addr = vm_unmapped_area(&info); | |
3381 | } | |
3382 | ||
3383 | return addr; | |
3384 | } | |
3385 | ||
6c5c240e RP |
3386 | #else /* !CONFIG_MMU */ |
3387 | ||
3388 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3389 | { | |
fc4f4be9 | 3390 | return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL; |
6c5c240e RP |
3391 | } |
3392 | ||
3393 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
3394 | { | |
3395 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
3396 | } | |
3397 | ||
3398 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
3399 | unsigned long addr, unsigned long len, | |
3400 | unsigned long pgoff, unsigned long flags) | |
3401 | { | |
3402 | void *ptr; | |
3403 | ||
3404 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
3405 | if (IS_ERR(ptr)) | |
3406 | return PTR_ERR(ptr); | |
3407 | ||
3408 | return (unsigned long) ptr; | |
3409 | } | |
3410 | ||
3411 | #endif /* !CONFIG_MMU */ | |
3412 | ||
f81440d3 PB |
3413 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
3414 | { | |
3415 | if (flags & IORING_ENTER_EXT_ARG) { | |
3416 | struct io_uring_getevents_arg arg; | |
3417 | ||
3418 | if (argsz != sizeof(arg)) | |
3419 | return -EINVAL; | |
3420 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3421 | return -EFAULT; | |
3422 | } | |
3423 | return 0; | |
3424 | } | |
3425 | ||
c73ebb68 HX |
3426 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
3427 | struct __kernel_timespec __user **ts, | |
3428 | const sigset_t __user **sig) | |
3429 | { | |
3430 | struct io_uring_getevents_arg arg; | |
3431 | ||
3432 | /* | |
3433 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
3434 | * is just a pointer to the sigset_t. | |
3435 | */ | |
3436 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
3437 | *sig = (const sigset_t __user *) argp; | |
3438 | *ts = NULL; | |
3439 | return 0; | |
3440 | } | |
3441 | ||
3442 | /* | |
3443 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
3444 | * timespec and sigset_t pointers if good. | |
3445 | */ | |
3446 | if (*argsz != sizeof(arg)) | |
3447 | return -EINVAL; | |
3448 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3449 | return -EFAULT; | |
d2347b96 DY |
3450 | if (arg.pad) |
3451 | return -EINVAL; | |
c73ebb68 HX |
3452 | *sig = u64_to_user_ptr(arg.sigmask); |
3453 | *argsz = arg.sigmask_sz; | |
3454 | *ts = u64_to_user_ptr(arg.ts); | |
3455 | return 0; | |
3456 | } | |
3457 | ||
2b188cc1 | 3458 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
3459 | u32, min_complete, u32, flags, const void __user *, argp, |
3460 | size_t, argsz) | |
2b188cc1 JA |
3461 | { |
3462 | struct io_ring_ctx *ctx; | |
2b188cc1 | 3463 | struct fd f; |
33f993da | 3464 | long ret; |
2b188cc1 | 3465 | |
33f993da | 3466 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
3467 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
3468 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
3469 | return -EINVAL; |
3470 | ||
e7a6c00d JA |
3471 | /* |
3472 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3473 | * need only dereference our task private array to find it. | |
3474 | */ | |
3475 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3476 | struct io_uring_task *tctx = current->io_uring; | |
3477 | ||
3273c440 | 3478 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
e7a6c00d JA |
3479 | return -EINVAL; |
3480 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
3481 | f.file = tctx->registered_rings[fd]; | |
4329490a | 3482 | f.flags = 0; |
3273c440 PB |
3483 | if (unlikely(!f.file)) |
3484 | return -EBADF; | |
e7a6c00d JA |
3485 | } else { |
3486 | f = fdget(fd); | |
3273c440 PB |
3487 | if (unlikely(!f.file)) |
3488 | return -EBADF; | |
3489 | ret = -EOPNOTSUPP; | |
3490 | if (unlikely(!io_is_uring_fops(f.file))) | |
fbb8bb02 | 3491 | goto out; |
e7a6c00d | 3492 | } |
2b188cc1 | 3493 | |
2b188cc1 | 3494 | ctx = f.file->private_data; |
7e84e1c7 | 3495 | ret = -EBADFD; |
33f993da | 3496 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3497 | goto out; |
3498 | ||
6c271ce2 JA |
3499 | /* |
3500 | * For SQ polling, the thread will do all submissions and completions. | |
3501 | * Just return the requested submit count, and wake the thread if | |
3502 | * we were asked to. | |
3503 | */ | |
b2a9eada | 3504 | ret = 0; |
6c271ce2 | 3505 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3506 | io_cqring_overflow_flush(ctx); |
89448c47 | 3507 | |
21f96522 JA |
3508 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3509 | ret = -EOWNERDEAD; | |
04147488 | 3510 | goto out; |
21f96522 | 3511 | } |
6c271ce2 | 3512 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3513 | wake_up(&ctx->sq_data->wait); |
88b80534 QF |
3514 | if (flags & IORING_ENTER_SQ_WAIT) |
3515 | io_sqpoll_wait_sq(ctx); | |
3516 | ||
3e813c90 | 3517 | ret = to_submit; |
b2a9eada | 3518 | } else if (to_submit) { |
eef51daa | 3519 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3520 | if (unlikely(ret)) |
3521 | goto out; | |
7c504e65 | 3522 | |
2b188cc1 | 3523 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3524 | ret = io_submit_sqes(ctx, to_submit); |
3525 | if (ret != to_submit) { | |
d487b43c | 3526 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3527 | goto out; |
d487b43c | 3528 | } |
44f87745 PB |
3529 | if (flags & IORING_ENTER_GETEVENTS) { |
3530 | if (ctx->syscall_iopoll) | |
3531 | goto iopoll_locked; | |
3532 | /* | |
3533 | * Ignore errors, we'll soon call io_cqring_wait() and | |
3534 | * it should handle ownership problems if any. | |
3535 | */ | |
3536 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3537 | (void)io_run_local_work_locked(ctx); | |
3538 | } | |
d487b43c | 3539 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 | 3540 | } |
c0e0d6ba | 3541 | |
2b188cc1 | 3542 | if (flags & IORING_ENTER_GETEVENTS) { |
3e813c90 | 3543 | int ret2; |
c0e0d6ba | 3544 | |
773697b6 | 3545 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3546 | /* |
3547 | * We disallow the app entering submit/complete with | |
3548 | * polling, but we still need to lock the ring to | |
3549 | * prevent racing with polled issue that got punted to | |
3550 | * a workqueue. | |
3551 | */ | |
3552 | mutex_lock(&ctx->uring_lock); | |
3553 | iopoll_locked: | |
3e813c90 DY |
3554 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3555 | if (likely(!ret2)) { | |
3556 | min_complete = min(min_complete, | |
3557 | ctx->cq_entries); | |
3558 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3559 | } |
3560 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3561 | } else { |
f81440d3 PB |
3562 | const sigset_t __user *sig; |
3563 | struct __kernel_timespec __user *ts; | |
3564 | ||
3e813c90 DY |
3565 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3566 | if (likely(!ret2)) { | |
3567 | min_complete = min(min_complete, | |
3568 | ctx->cq_entries); | |
3569 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3570 | argsz, ts); | |
3571 | } | |
def596e9 | 3572 | } |
c73ebb68 | 3573 | |
155bc950 | 3574 | if (!ret) { |
3e813c90 | 3575 | ret = ret2; |
2b188cc1 | 3576 | |
155bc950 DY |
3577 | /* |
3578 | * EBADR indicates that one or more CQE were dropped. | |
3579 | * Once the user has been informed we can clear the bit | |
3580 | * as they are obviously ok with those drops. | |
3581 | */ | |
3582 | if (unlikely(ret2 == -EBADR)) | |
3583 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3584 | &ctx->check_cq); | |
def596e9 | 3585 | } |
2b188cc1 | 3586 | } |
7c504e65 | 3587 | out: |
4329490a | 3588 | fdput(f); |
3e813c90 | 3589 | return ret; |
2b188cc1 JA |
3590 | } |
3591 | ||
3592 | static const struct file_operations io_uring_fops = { | |
3593 | .release = io_uring_release, | |
3594 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3595 | #ifndef CONFIG_MMU |
3596 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3597 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
d808459b HD |
3598 | #else |
3599 | .get_unmapped_area = io_uring_mmu_get_unmapped_area, | |
6c5c240e | 3600 | #endif |
2b188cc1 | 3601 | .poll = io_uring_poll, |
bebdb65e | 3602 | #ifdef CONFIG_PROC_FS |
87ce955b | 3603 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3604 | #endif |
2b188cc1 JA |
3605 | }; |
3606 | ||
92ac8bea JA |
3607 | bool io_is_uring_fops(struct file *file) |
3608 | { | |
3609 | return file->f_op == &io_uring_fops; | |
3610 | } | |
3611 | ||
c072481d PB |
3612 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3613 | struct io_uring_params *p) | |
2b188cc1 | 3614 | { |
75b28aff HV |
3615 | struct io_rings *rings; |
3616 | size_t size, sq_array_offset; | |
2b188cc1 | 3617 | |
bd740481 JA |
3618 | /* make sure these are sane, as we already accounted them */ |
3619 | ctx->sq_entries = p->sq_entries; | |
3620 | ctx->cq_entries = p->cq_entries; | |
3621 | ||
baf9cb64 | 3622 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3623 | if (size == SIZE_MAX) |
3624 | return -EOVERFLOW; | |
3625 | ||
3626 | rings = io_mem_alloc(size); | |
3627 | if (!rings) | |
2b188cc1 JA |
3628 | return -ENOMEM; |
3629 | ||
75b28aff HV |
3630 | ctx->rings = rings; |
3631 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
3632 | rings->sq_ring_mask = p->sq_entries - 1; | |
3633 | rings->cq_ring_mask = p->cq_entries - 1; | |
3634 | rings->sq_ring_entries = p->sq_entries; | |
3635 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3636 | |
ebdeb7c0 JA |
3637 | if (p->flags & IORING_SETUP_SQE128) |
3638 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3639 | else | |
3640 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
3641 | if (size == SIZE_MAX) { |
3642 | io_mem_free(ctx->rings); | |
3643 | ctx->rings = NULL; | |
2b188cc1 | 3644 | return -EOVERFLOW; |
eb065d30 | 3645 | } |
2b188cc1 JA |
3646 | |
3647 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
3648 | if (!ctx->sq_sqes) { |
3649 | io_mem_free(ctx->rings); | |
3650 | ctx->rings = NULL; | |
2b188cc1 | 3651 | return -ENOMEM; |
eb065d30 | 3652 | } |
2b188cc1 | 3653 | |
2b188cc1 JA |
3654 | return 0; |
3655 | } | |
3656 | ||
9faadcc8 PB |
3657 | static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) |
3658 | { | |
3659 | int ret, fd; | |
3660 | ||
3661 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3662 | if (fd < 0) | |
3663 | return fd; | |
3664 | ||
97c96e9f | 3665 | ret = __io_uring_add_tctx_node(ctx); |
9faadcc8 PB |
3666 | if (ret) { |
3667 | put_unused_fd(fd); | |
3668 | return ret; | |
3669 | } | |
3670 | fd_install(fd, file); | |
3671 | return fd; | |
3672 | } | |
3673 | ||
2b188cc1 JA |
3674 | /* |
3675 | * Allocate an anonymous fd, this is what constitutes the application | |
3676 | * visible backing of an io_uring instance. The application mmaps this | |
3677 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3678 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3679 | */ | |
9faadcc8 | 3680 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
3681 | { |
3682 | struct file *file; | |
9faadcc8 | 3683 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
3684 | int ret; |
3685 | ||
2b188cc1 JA |
3686 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
3687 | &ctx->ring_sock); | |
3688 | if (ret) | |
9faadcc8 | 3689 | return ERR_PTR(ret); |
2b188cc1 JA |
3690 | #endif |
3691 | ||
91a9ab7c PM |
3692 | file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, |
3693 | O_RDWR | O_CLOEXEC, NULL); | |
2b188cc1 | 3694 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
3695 | if (IS_ERR(file)) { |
3696 | sock_release(ctx->ring_sock); | |
3697 | ctx->ring_sock = NULL; | |
3698 | } else { | |
3699 | ctx->ring_sock->file = file; | |
0f212204 | 3700 | } |
2b188cc1 | 3701 | #endif |
9faadcc8 | 3702 | return file; |
2b188cc1 JA |
3703 | } |
3704 | ||
c072481d PB |
3705 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3706 | struct io_uring_params __user *params) | |
2b188cc1 | 3707 | { |
2b188cc1 | 3708 | struct io_ring_ctx *ctx; |
9faadcc8 | 3709 | struct file *file; |
2b188cc1 JA |
3710 | int ret; |
3711 | ||
8110c1a6 | 3712 | if (!entries) |
2b188cc1 | 3713 | return -EINVAL; |
8110c1a6 JA |
3714 | if (entries > IORING_MAX_ENTRIES) { |
3715 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3716 | return -EINVAL; | |
3717 | entries = IORING_MAX_ENTRIES; | |
3718 | } | |
2b188cc1 JA |
3719 | |
3720 | /* | |
3721 | * Use twice as many entries for the CQ ring. It's possible for the | |
3722 | * application to drive a higher depth than the size of the SQ ring, | |
3723 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3724 | * some flexibility in overcommitting a bit. If the application has |
3725 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3726 | * of CQ ring entries manually. | |
2b188cc1 JA |
3727 | */ |
3728 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3729 | if (p->flags & IORING_SETUP_CQSIZE) { |
3730 | /* | |
3731 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3732 | * to a power-of-two, if it isn't already. We do NOT impose | |
3733 | * any cq vs sq ring sizing. | |
3734 | */ | |
eb2667b3 | 3735 | if (!p->cq_entries) |
33a107f0 | 3736 | return -EINVAL; |
8110c1a6 JA |
3737 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3738 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3739 | return -EINVAL; | |
3740 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3741 | } | |
eb2667b3 JQ |
3742 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3743 | if (p->cq_entries < p->sq_entries) | |
3744 | return -EINVAL; | |
33a107f0 JA |
3745 | } else { |
3746 | p->cq_entries = 2 * p->sq_entries; | |
3747 | } | |
2b188cc1 | 3748 | |
2b188cc1 | 3749 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3750 | if (!ctx) |
2b188cc1 | 3751 | return -ENOMEM; |
773697b6 | 3752 | |
e6aeb272 PB |
3753 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3754 | !(ctx->flags & IORING_SETUP_IOPOLL) && | |
3755 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3756 | ctx->task_complete = true; | |
3757 | ||
bca39f39 PB |
3758 | /* |
3759 | * lazy poll_wq activation relies on ->task_complete for synchronisation | |
3760 | * purposes, see io_activate_pollwq() | |
3761 | */ | |
3762 | if (!ctx->task_complete) | |
3763 | ctx->poll_activated = true; | |
3764 | ||
773697b6 PB |
3765 | /* |
3766 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3767 | * space applications don't need to do io completion events | |
3768 | * polling again, they can rely on io_sq_thread to do polling | |
3769 | * work, which can reduce cpu usage and uring_lock contention. | |
3770 | */ | |
3771 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3772 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3773 | ctx->syscall_iopoll = 1; | |
3774 | ||
2b188cc1 | 3775 | ctx->compat = in_compat_syscall(); |
62e398be JA |
3776 | if (!capable(CAP_IPC_LOCK)) |
3777 | ctx->user = get_uid(current_user()); | |
2aede0e4 | 3778 | |
9f010507 | 3779 | /* |
e1169f06 JA |
3780 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3781 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3782 | */ |
e1169f06 JA |
3783 | ret = -EINVAL; |
3784 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3785 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 | 3786 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
c0e0d6ba DY |
3787 | IORING_SETUP_TASKRUN_FLAG | |
3788 | IORING_SETUP_DEFER_TASKRUN)) | |
e1169f06 | 3789 | goto err; |
9f010507 | 3790 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3791 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3792 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3793 | } else { | |
c0e0d6ba DY |
3794 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && |
3795 | !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) | |
ef060ea9 | 3796 | goto err; |
9f010507 | 3797 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3798 | } |
9f010507 | 3799 | |
c0e0d6ba DY |
3800 | /* |
3801 | * For DEFER_TASKRUN we require the completion task to be the same as the | |
3802 | * submission task. This implies that there is only one submitter, so enforce | |
3803 | * that. | |
3804 | */ | |
3805 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && | |
3806 | !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { | |
3807 | goto err; | |
3808 | } | |
3809 | ||
2aede0e4 JA |
3810 | /* |
3811 | * This is just grabbed for accounting purposes. When a process exits, | |
3812 | * the mm is exited and dropped before the files, hence we need to hang | |
3813 | * on to this mm purely for the purposes of being able to unaccount | |
3814 | * memory (locked/pinned vm). It's not used for anything else. | |
3815 | */ | |
6b7898eb | 3816 | mmgrab(current->mm); |
2aede0e4 | 3817 | ctx->mm_account = current->mm; |
6b7898eb | 3818 | |
2b188cc1 JA |
3819 | ret = io_allocate_scq_urings(ctx, p); |
3820 | if (ret) | |
3821 | goto err; | |
3822 | ||
7e84e1c7 | 3823 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
3824 | if (ret) |
3825 | goto err; | |
eae071c9 | 3826 | /* always set a rsrc node */ |
47b228ce PB |
3827 | ret = io_rsrc_node_switch_start(ctx); |
3828 | if (ret) | |
3829 | goto err; | |
eae071c9 | 3830 | io_rsrc_node_switch(ctx, NULL); |
2b188cc1 | 3831 | |
2b188cc1 | 3832 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
3833 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3834 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3835 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3836 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3837 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3838 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
3839 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
3840 | |
3841 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
3842 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3843 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3844 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3845 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3846 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3847 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 3848 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 3849 | |
7f13657d XW |
3850 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
3851 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 3852 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 3853 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 3854 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e | 3855 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
7d3fd88d | 3856 | IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING; |
7f13657d XW |
3857 | |
3858 | if (copy_to_user(params, p, sizeof(*p))) { | |
3859 | ret = -EFAULT; | |
3860 | goto err; | |
3861 | } | |
d1719f70 | 3862 | |
7cae596b DY |
3863 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER |
3864 | && !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
8579538c | 3865 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
7cae596b | 3866 | |
9faadcc8 PB |
3867 | file = io_uring_get_file(ctx); |
3868 | if (IS_ERR(file)) { | |
3869 | ret = PTR_ERR(file); | |
3870 | goto err; | |
3871 | } | |
3872 | ||
044c1ab3 JA |
3873 | /* |
3874 | * Install ring fd as the very last thing, so we don't risk someone | |
3875 | * having closed it before we finish setup | |
3876 | */ | |
9faadcc8 PB |
3877 | ret = io_uring_install_fd(ctx, file); |
3878 | if (ret < 0) { | |
3879 | /* fput will clean it up */ | |
3880 | fput(file); | |
3881 | return ret; | |
3882 | } | |
044c1ab3 | 3883 | |
c826bd7a | 3884 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
3885 | return ret; |
3886 | err: | |
3887 | io_ring_ctx_wait_and_kill(ctx); | |
3888 | return ret; | |
3889 | } | |
3890 | ||
3891 | /* | |
3892 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3893 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3894 | * params structure passed in. | |
3895 | */ | |
3896 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3897 | { | |
3898 | struct io_uring_params p; | |
2b188cc1 JA |
3899 | int i; |
3900 | ||
3901 | if (copy_from_user(&p, params, sizeof(p))) | |
3902 | return -EFAULT; | |
3903 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3904 | if (p.resv[i]) | |
3905 | return -EINVAL; | |
3906 | } | |
3907 | ||
6c271ce2 | 3908 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 3909 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 3910 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 3911 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 3912 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 | 3913 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
c0e0d6ba | 3914 | IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN)) |
2b188cc1 JA |
3915 | return -EINVAL; |
3916 | ||
ef060ea9 | 3917 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
3918 | } |
3919 | ||
3920 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3921 | struct io_uring_params __user *, params) | |
3922 | { | |
3923 | return io_uring_setup(entries, params); | |
3924 | } | |
3925 | ||
c072481d PB |
3926 | static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, |
3927 | unsigned nr_args) | |
66f4af93 JA |
3928 | { |
3929 | struct io_uring_probe *p; | |
3930 | size_t size; | |
3931 | int i, ret; | |
3932 | ||
3933 | size = struct_size(p, ops, nr_args); | |
3934 | if (size == SIZE_MAX) | |
3935 | return -EOVERFLOW; | |
3936 | p = kzalloc(size, GFP_KERNEL); | |
3937 | if (!p) | |
3938 | return -ENOMEM; | |
3939 | ||
3940 | ret = -EFAULT; | |
3941 | if (copy_from_user(p, arg, size)) | |
3942 | goto out; | |
3943 | ret = -EINVAL; | |
3944 | if (memchr_inv(p, 0, size)) | |
3945 | goto out; | |
3946 | ||
3947 | p->last_op = IORING_OP_LAST - 1; | |
3948 | if (nr_args > IORING_OP_LAST) | |
3949 | nr_args = IORING_OP_LAST; | |
3950 | ||
3951 | for (i = 0; i < nr_args; i++) { | |
3952 | p->ops[i].op = i; | |
a7dd2782 | 3953 | if (!io_issue_defs[i].not_supported) |
66f4af93 JA |
3954 | p->ops[i].flags = IO_URING_OP_SUPPORTED; |
3955 | } | |
3956 | p->ops_len = i; | |
3957 | ||
3958 | ret = 0; | |
3959 | if (copy_to_user(arg, p, size)) | |
3960 | ret = -EFAULT; | |
3961 | out: | |
3962 | kfree(p); | |
3963 | return ret; | |
3964 | } | |
3965 | ||
071698e1 JA |
3966 | static int io_register_personality(struct io_ring_ctx *ctx) |
3967 | { | |
4379bf8b | 3968 | const struct cred *creds; |
61cf9370 | 3969 | u32 id; |
1e6fa521 | 3970 | int ret; |
071698e1 | 3971 | |
4379bf8b | 3972 | creds = get_current_cred(); |
1e6fa521 | 3973 | |
61cf9370 MWO |
3974 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
3975 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
3976 | if (ret < 0) { |
3977 | put_cred(creds); | |
3978 | return ret; | |
3979 | } | |
3980 | return id; | |
071698e1 JA |
3981 | } |
3982 | ||
c072481d PB |
3983 | static __cold int io_register_restrictions(struct io_ring_ctx *ctx, |
3984 | void __user *arg, unsigned int nr_args) | |
21b55dbc SG |
3985 | { |
3986 | struct io_uring_restriction *res; | |
3987 | size_t size; | |
3988 | int i, ret; | |
3989 | ||
7e84e1c7 SG |
3990 | /* Restrictions allowed only if rings started disabled */ |
3991 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3992 | return -EBADFD; | |
3993 | ||
21b55dbc | 3994 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 3995 | if (ctx->restrictions.registered) |
21b55dbc SG |
3996 | return -EBUSY; |
3997 | ||
3998 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
3999 | return -EINVAL; | |
4000 | ||
4001 | size = array_size(nr_args, sizeof(*res)); | |
4002 | if (size == SIZE_MAX) | |
4003 | return -EOVERFLOW; | |
4004 | ||
4005 | res = memdup_user(arg, size); | |
4006 | if (IS_ERR(res)) | |
4007 | return PTR_ERR(res); | |
4008 | ||
4009 | ret = 0; | |
4010 | ||
4011 | for (i = 0; i < nr_args; i++) { | |
4012 | switch (res[i].opcode) { | |
4013 | case IORING_RESTRICTION_REGISTER_OP: | |
4014 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
4015 | ret = -EINVAL; | |
4016 | goto out; | |
4017 | } | |
4018 | ||
4019 | __set_bit(res[i].register_op, | |
4020 | ctx->restrictions.register_op); | |
4021 | break; | |
4022 | case IORING_RESTRICTION_SQE_OP: | |
4023 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
4024 | ret = -EINVAL; | |
4025 | goto out; | |
4026 | } | |
4027 | ||
4028 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
4029 | break; | |
4030 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
4031 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
4032 | break; | |
4033 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
4034 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
4035 | break; | |
4036 | default: | |
4037 | ret = -EINVAL; | |
4038 | goto out; | |
4039 | } | |
4040 | } | |
4041 | ||
4042 | out: | |
4043 | /* Reset all restrictions if an error happened */ | |
4044 | if (ret != 0) | |
4045 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
4046 | else | |
7e84e1c7 | 4047 | ctx->restrictions.registered = true; |
21b55dbc SG |
4048 | |
4049 | kfree(res); | |
4050 | return ret; | |
4051 | } | |
4052 | ||
7e84e1c7 SG |
4053 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
4054 | { | |
4055 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
4056 | return -EBADFD; | |
4057 | ||
bca39f39 | 4058 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) { |
8579538c | 4059 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
bca39f39 PB |
4060 | /* |
4061 | * Lazy activation attempts would fail if it was polled before | |
4062 | * submitter_task is set. | |
4063 | */ | |
4064 | if (wq_has_sleeper(&ctx->poll_wq)) | |
4065 | io_activate_pollwq(ctx); | |
4066 | } | |
7cae596b | 4067 | |
7e84e1c7 SG |
4068 | if (ctx->restrictions.registered) |
4069 | ctx->restricted = 1; | |
4070 | ||
0298ef96 PB |
4071 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
4072 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
4073 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
4074 | return 0; |
4075 | } | |
4076 | ||
c072481d PB |
4077 | static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, |
4078 | void __user *arg, unsigned len) | |
fe76421d JA |
4079 | { |
4080 | struct io_uring_task *tctx = current->io_uring; | |
4081 | cpumask_var_t new_mask; | |
4082 | int ret; | |
4083 | ||
4084 | if (!tctx || !tctx->io_wq) | |
4085 | return -EINVAL; | |
4086 | ||
4087 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
4088 | return -ENOMEM; | |
4089 | ||
4090 | cpumask_clear(new_mask); | |
4091 | if (len > cpumask_size()) | |
4092 | len = cpumask_size(); | |
4093 | ||
0f5e4b83 ES |
4094 | if (in_compat_syscall()) { |
4095 | ret = compat_get_bitmap(cpumask_bits(new_mask), | |
4096 | (const compat_ulong_t __user *)arg, | |
4097 | len * 8 /* CHAR_BIT */); | |
4098 | } else { | |
4099 | ret = copy_from_user(new_mask, arg, len); | |
4100 | } | |
4101 | ||
4102 | if (ret) { | |
fe76421d JA |
4103 | free_cpumask_var(new_mask); |
4104 | return -EFAULT; | |
4105 | } | |
4106 | ||
4107 | ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); | |
4108 | free_cpumask_var(new_mask); | |
4109 | return ret; | |
4110 | } | |
4111 | ||
c072481d | 4112 | static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) |
fe76421d JA |
4113 | { |
4114 | struct io_uring_task *tctx = current->io_uring; | |
4115 | ||
4116 | if (!tctx || !tctx->io_wq) | |
4117 | return -EINVAL; | |
4118 | ||
4119 | return io_wq_cpu_affinity(tctx->io_wq, NULL); | |
4120 | } | |
4121 | ||
c072481d PB |
4122 | static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, |
4123 | void __user *arg) | |
b22fa62a | 4124 | __must_hold(&ctx->uring_lock) |
2e480058 | 4125 | { |
b22fa62a | 4126 | struct io_tctx_node *node; |
fa84693b JA |
4127 | struct io_uring_task *tctx = NULL; |
4128 | struct io_sq_data *sqd = NULL; | |
2e480058 JA |
4129 | __u32 new_count[2]; |
4130 | int i, ret; | |
4131 | ||
2e480058 JA |
4132 | if (copy_from_user(new_count, arg, sizeof(new_count))) |
4133 | return -EFAULT; | |
4134 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
4135 | if (new_count[i] > INT_MAX) | |
4136 | return -EINVAL; | |
4137 | ||
fa84693b JA |
4138 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
4139 | sqd = ctx->sq_data; | |
4140 | if (sqd) { | |
009ad9f0 JA |
4141 | /* |
4142 | * Observe the correct sqd->lock -> ctx->uring_lock | |
4143 | * ordering. Fine to drop uring_lock here, we hold | |
4144 | * a ref to the ctx. | |
4145 | */ | |
41d3a6bd | 4146 | refcount_inc(&sqd->refs); |
009ad9f0 | 4147 | mutex_unlock(&ctx->uring_lock); |
fa84693b | 4148 | mutex_lock(&sqd->lock); |
009ad9f0 | 4149 | mutex_lock(&ctx->uring_lock); |
41d3a6bd JA |
4150 | if (sqd->thread) |
4151 | tctx = sqd->thread->io_uring; | |
fa84693b JA |
4152 | } |
4153 | } else { | |
4154 | tctx = current->io_uring; | |
4155 | } | |
4156 | ||
e139a1ec | 4157 | BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); |
fa84693b | 4158 | |
bad119b9 PB |
4159 | for (i = 0; i < ARRAY_SIZE(new_count); i++) |
4160 | if (new_count[i]) | |
4161 | ctx->iowq_limits[i] = new_count[i]; | |
e139a1ec PB |
4162 | ctx->iowq_limits_set = true; |
4163 | ||
e139a1ec PB |
4164 | if (tctx && tctx->io_wq) { |
4165 | ret = io_wq_max_workers(tctx->io_wq, new_count); | |
4166 | if (ret) | |
4167 | goto err; | |
4168 | } else { | |
4169 | memset(new_count, 0, sizeof(new_count)); | |
4170 | } | |
fa84693b | 4171 | |
41d3a6bd | 4172 | if (sqd) { |
fa84693b | 4173 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
4174 | io_put_sq_data(sqd); |
4175 | } | |
2e480058 JA |
4176 | |
4177 | if (copy_to_user(arg, new_count, sizeof(new_count))) | |
4178 | return -EFAULT; | |
4179 | ||
b22fa62a PB |
4180 | /* that's it for SQPOLL, only the SQPOLL task creates requests */ |
4181 | if (sqd) | |
4182 | return 0; | |
4183 | ||
4184 | /* now propagate the restriction to all registered users */ | |
4185 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
4186 | struct io_uring_task *tctx = node->task->io_uring; | |
4187 | ||
4188 | if (WARN_ON_ONCE(!tctx->io_wq)) | |
4189 | continue; | |
4190 | ||
4191 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
4192 | new_count[i] = ctx->iowq_limits[i]; | |
4193 | /* ignore errors, it always returns zero anyway */ | |
4194 | (void)io_wq_max_workers(tctx->io_wq, new_count); | |
4195 | } | |
2e480058 | 4196 | return 0; |
fa84693b | 4197 | err: |
41d3a6bd | 4198 | if (sqd) { |
fa84693b | 4199 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
4200 | io_put_sq_data(sqd); |
4201 | } | |
fa84693b | 4202 | return ret; |
2e480058 JA |
4203 | } |
4204 | ||
edafccee JA |
4205 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
4206 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
4207 | __releases(ctx->uring_lock) |
4208 | __acquires(ctx->uring_lock) | |
edafccee JA |
4209 | { |
4210 | int ret; | |
4211 | ||
35fa71a0 | 4212 | /* |
fbb8bb02 PB |
4213 | * We don't quiesce the refs for register anymore and so it can't be |
4214 | * dying as we're holding a file ref here. | |
35fa71a0 | 4215 | */ |
fbb8bb02 | 4216 | if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs))) |
35fa71a0 JA |
4217 | return -ENXIO; |
4218 | ||
d7cce96c PB |
4219 | if (ctx->submitter_task && ctx->submitter_task != current) |
4220 | return -EEXIST; | |
4221 | ||
75c4021a | 4222 | if (ctx->restricted) { |
75c4021a PB |
4223 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); |
4224 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
4225 | return -EACCES; | |
4226 | } | |
4227 | ||
edafccee JA |
4228 | switch (opcode) { |
4229 | case IORING_REGISTER_BUFFERS: | |
0184f08e PB |
4230 | ret = -EFAULT; |
4231 | if (!arg) | |
4232 | break; | |
634d00df | 4233 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
4234 | break; |
4235 | case IORING_UNREGISTER_BUFFERS: | |
4236 | ret = -EINVAL; | |
4237 | if (arg || nr_args) | |
4238 | break; | |
0a96bbe4 | 4239 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 4240 | break; |
6b06314c | 4241 | case IORING_REGISTER_FILES: |
a8da73a3 JA |
4242 | ret = -EFAULT; |
4243 | if (!arg) | |
4244 | break; | |
792e3582 | 4245 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
4246 | break; |
4247 | case IORING_UNREGISTER_FILES: | |
4248 | ret = -EINVAL; | |
4249 | if (arg || nr_args) | |
4250 | break; | |
4251 | ret = io_sqe_files_unregister(ctx); | |
4252 | break; | |
c3a31e60 | 4253 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 4254 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 4255 | break; |
9b402849 JA |
4256 | case IORING_REGISTER_EVENTFD: |
4257 | ret = -EINVAL; | |
4258 | if (nr_args != 1) | |
4259 | break; | |
c75312dd UA |
4260 | ret = io_eventfd_register(ctx, arg, 0); |
4261 | break; | |
4262 | case IORING_REGISTER_EVENTFD_ASYNC: | |
4263 | ret = -EINVAL; | |
4264 | if (nr_args != 1) | |
f2842ab5 | 4265 | break; |
c75312dd | 4266 | ret = io_eventfd_register(ctx, arg, 1); |
9b402849 JA |
4267 | break; |
4268 | case IORING_UNREGISTER_EVENTFD: | |
4269 | ret = -EINVAL; | |
4270 | if (arg || nr_args) | |
4271 | break; | |
4272 | ret = io_eventfd_unregister(ctx); | |
4273 | break; | |
66f4af93 JA |
4274 | case IORING_REGISTER_PROBE: |
4275 | ret = -EINVAL; | |
4276 | if (!arg || nr_args > 256) | |
4277 | break; | |
4278 | ret = io_probe(ctx, arg, nr_args); | |
4279 | break; | |
071698e1 JA |
4280 | case IORING_REGISTER_PERSONALITY: |
4281 | ret = -EINVAL; | |
4282 | if (arg || nr_args) | |
4283 | break; | |
4284 | ret = io_register_personality(ctx); | |
4285 | break; | |
4286 | case IORING_UNREGISTER_PERSONALITY: | |
4287 | ret = -EINVAL; | |
4288 | if (arg) | |
4289 | break; | |
4290 | ret = io_unregister_personality(ctx, nr_args); | |
4291 | break; | |
7e84e1c7 SG |
4292 | case IORING_REGISTER_ENABLE_RINGS: |
4293 | ret = -EINVAL; | |
4294 | if (arg || nr_args) | |
4295 | break; | |
4296 | ret = io_register_enable_rings(ctx); | |
4297 | break; | |
21b55dbc SG |
4298 | case IORING_REGISTER_RESTRICTIONS: |
4299 | ret = io_register_restrictions(ctx, arg, nr_args); | |
4300 | break; | |
992da01a PB |
4301 | case IORING_REGISTER_FILES2: |
4302 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
4303 | break; | |
4304 | case IORING_REGISTER_FILES_UPDATE2: | |
4305 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
4306 | IORING_RSRC_FILE); | |
4307 | break; | |
4308 | case IORING_REGISTER_BUFFERS2: | |
4309 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 4310 | break; |
992da01a PB |
4311 | case IORING_REGISTER_BUFFERS_UPDATE: |
4312 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
4313 | IORING_RSRC_BUFFER); | |
c3bdad02 | 4314 | break; |
fe76421d JA |
4315 | case IORING_REGISTER_IOWQ_AFF: |
4316 | ret = -EINVAL; | |
4317 | if (!arg || !nr_args) | |
4318 | break; | |
4319 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
4320 | break; | |
4321 | case IORING_UNREGISTER_IOWQ_AFF: | |
4322 | ret = -EINVAL; | |
4323 | if (arg || nr_args) | |
4324 | break; | |
4325 | ret = io_unregister_iowq_aff(ctx); | |
4326 | break; | |
2e480058 JA |
4327 | case IORING_REGISTER_IOWQ_MAX_WORKERS: |
4328 | ret = -EINVAL; | |
4329 | if (!arg || nr_args != 2) | |
4330 | break; | |
4331 | ret = io_register_iowq_max_workers(ctx, arg); | |
4332 | break; | |
e7a6c00d JA |
4333 | case IORING_REGISTER_RING_FDS: |
4334 | ret = io_ringfd_register(ctx, arg, nr_args); | |
4335 | break; | |
4336 | case IORING_UNREGISTER_RING_FDS: | |
4337 | ret = io_ringfd_unregister(ctx, arg, nr_args); | |
4338 | break; | |
c7fb1942 JA |
4339 | case IORING_REGISTER_PBUF_RING: |
4340 | ret = -EINVAL; | |
4341 | if (!arg || nr_args != 1) | |
4342 | break; | |
4343 | ret = io_register_pbuf_ring(ctx, arg); | |
4344 | break; | |
4345 | case IORING_UNREGISTER_PBUF_RING: | |
4346 | ret = -EINVAL; | |
4347 | if (!arg || nr_args != 1) | |
4348 | break; | |
4349 | ret = io_unregister_pbuf_ring(ctx, arg); | |
4350 | break; | |
78a861b9 JA |
4351 | case IORING_REGISTER_SYNC_CANCEL: |
4352 | ret = -EINVAL; | |
4353 | if (!arg || nr_args != 1) | |
4354 | break; | |
4355 | ret = io_sync_cancel(ctx, arg); | |
4356 | break; | |
6e73dffb PB |
4357 | case IORING_REGISTER_FILE_ALLOC_RANGE: |
4358 | ret = -EINVAL; | |
4359 | if (!arg || nr_args) | |
4360 | break; | |
4361 | ret = io_register_file_alloc_range(ctx, arg); | |
4362 | break; | |
edafccee JA |
4363 | default: |
4364 | ret = -EINVAL; | |
4365 | break; | |
4366 | } | |
4367 | ||
edafccee JA |
4368 | return ret; |
4369 | } | |
4370 | ||
4371 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
4372 | void __user *, arg, unsigned int, nr_args) | |
4373 | { | |
4374 | struct io_ring_ctx *ctx; | |
4375 | long ret = -EBADF; | |
4376 | struct fd f; | |
7d3fd88d JT |
4377 | bool use_registered_ring; |
4378 | ||
4379 | use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING); | |
4380 | opcode &= ~IORING_REGISTER_USE_REGISTERED_RING; | |
edafccee | 4381 | |
34319084 JA |
4382 | if (opcode >= IORING_REGISTER_LAST) |
4383 | return -EINVAL; | |
4384 | ||
7d3fd88d JT |
4385 | if (use_registered_ring) { |
4386 | /* | |
4387 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
4388 | * need only dereference our task private array to find it. | |
4389 | */ | |
4390 | struct io_uring_task *tctx = current->io_uring; | |
edafccee | 4391 | |
7d3fd88d JT |
4392 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
4393 | return -EINVAL; | |
4394 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
4395 | f.file = tctx->registered_rings[fd]; | |
4396 | f.flags = 0; | |
4397 | if (unlikely(!f.file)) | |
4398 | return -EBADF; | |
4399 | } else { | |
4400 | f = fdget(fd); | |
4401 | if (unlikely(!f.file)) | |
4402 | return -EBADF; | |
4403 | ret = -EOPNOTSUPP; | |
4404 | if (!io_is_uring_fops(f.file)) | |
4405 | goto out_fput; | |
4406 | } | |
edafccee JA |
4407 | |
4408 | ctx = f.file->private_data; | |
4409 | ||
4410 | mutex_lock(&ctx->uring_lock); | |
4411 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
4412 | mutex_unlock(&ctx->uring_lock); | |
2757be22 | 4413 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); |
edafccee JA |
4414 | out_fput: |
4415 | fdput(f); | |
4416 | return ret; | |
4417 | } | |
4418 | ||
2b188cc1 JA |
4419 | static int __init io_uring_init(void) |
4420 | { | |
9c71d39a | 4421 | #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ |
d7f62e82 | 4422 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ |
9c71d39a | 4423 | BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ |
d7f62e82 SM |
4424 | } while (0) |
4425 | ||
4426 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9c71d39a SM |
4427 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) |
4428 | #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ | |
4429 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) | |
d7f62e82 SM |
4430 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); |
4431 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
4432 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
4433 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
4434 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
4435 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
4436 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9c71d39a SM |
4437 | BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); |
4438 | BUILD_BUG_SQE_ELEM(12, __u32, __pad1); | |
d7f62e82 | 4439 | BUILD_BUG_SQE_ELEM(16, __u64, addr); |
7d67af2c | 4440 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
4441 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
4442 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
4443 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
4444 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
4445 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
4446 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
4447 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
4448 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
4449 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
4450 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
4451 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
4452 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
4453 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
4454 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
4455 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 4456 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
9c71d39a SM |
4457 | BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); |
4458 | BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); | |
4459 | BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); | |
4460 | BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); | |
4461 | BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); | |
d7f62e82 SM |
4462 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
4463 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 4464 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 4465 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 4466 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 4467 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
b48c312b PB |
4468 | BUILD_BUG_SQE_ELEM(44, __u16, addr_len); |
4469 | BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); | |
e9621e2b | 4470 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
9c71d39a SM |
4471 | BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); |
4472 | BUILD_BUG_SQE_ELEM(56, __u64, __pad2); | |
d7f62e82 | 4473 | |
b0d658ec PB |
4474 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
4475 | sizeof(struct io_uring_rsrc_update)); | |
4476 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
4477 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
4478 | |
4479 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
4480 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
4481 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
4482 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 4483 | |
b0d658ec PB |
4484 | /* should fit into one byte */ |
4485 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
4486 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
4487 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 4488 | |
32c2d33e | 4489 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); |
16340eab | 4490 | |
3a4b89a2 JA |
4491 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
4492 | ||
d9b57aa3 | 4493 | io_uring_optable_init(); |
0702e536 | 4494 | |
91f245d5 JA |
4495 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | |
4496 | SLAB_ACCOUNT); | |
2b188cc1 JA |
4497 | return 0; |
4498 | }; | |
4499 | __initcall(io_uring_init); |