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" |
c92fcfc2 | 98 | #include "rw.h" |
9b797a37 | 99 | #include "alloc_cache.h" |
5e2a18d9 | 100 | |
5277deaa | 101 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 102 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 | 103 | |
21b55dbc SG |
104 | #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \ |
105 | IORING_REGISTER_LAST + IORING_OP_LAST) | |
2b188cc1 | 106 | |
68fe256a PB |
107 | #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \ |
108 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) | |
109 | ||
5562a8d7 PB |
110 | #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \ |
111 | IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS) | |
68fe256a | 112 | |
c854357b | 113 | #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ |
9cae36a0 JA |
114 | REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \ |
115 | REQ_F_ASYNC_DATA) | |
b16fed66 | 116 | |
a538be5b PB |
117 | #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\ |
118 | IO_REQ_CLEAN_FLAGS) | |
119 | ||
09899b19 PB |
120 | #define IO_TCTX_REFS_CACHE_NR (1U << 10) |
121 | ||
6dd0be1e | 122 | #define IO_COMPL_BATCH 32 |
bf019da7 | 123 | #define IO_REQ_ALLOC_BATCH 8 |
258b29a9 | 124 | |
10988a0a DY |
125 | enum { |
126 | IO_CHECK_CQ_OVERFLOW_BIT, | |
155bc950 | 127 | IO_CHECK_CQ_DROPPED_BIT, |
10988a0a DY |
128 | }; |
129 | ||
21a091b9 DY |
130 | enum { |
131 | IO_EVENTFD_OP_SIGNAL_BIT, | |
132 | IO_EVENTFD_OP_FREE_BIT, | |
133 | }; | |
134 | ||
27dc8338 PB |
135 | struct io_defer_entry { |
136 | struct list_head list; | |
137 | struct io_kiocb *req; | |
9cf7c104 | 138 | u32 seq; |
2b188cc1 JA |
139 | }; |
140 | ||
0756a869 PB |
141 | /* requests with any of those set should undergo io_disarm_next() */ |
142 | #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) | |
da1a08c5 | 143 | #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK) |
0756a869 | 144 | |
affa87db | 145 | static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
9936c7c2 | 146 | struct task_struct *task, |
3dd0c97a | 147 | bool cancel_all); |
1ffc5422 | 148 | |
c7dae4ba | 149 | static void io_dismantle_req(struct io_kiocb *req); |
68fb8979 | 150 | static void io_clean_op(struct io_kiocb *req); |
cbc2e203 | 151 | static void io_queue_sqe(struct io_kiocb *req); |
c0e0d6ba | 152 | static void io_move_task_work_from_local(struct io_ring_ctx *ctx); |
c450178d | 153 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx); |
77e443ab | 154 | static __cold void io_fallback_tw(struct io_uring_task *tctx); |
de0617e4 | 155 | |
c1755c25 | 156 | struct kmem_cache *req_cachep; |
2b188cc1 | 157 | |
2b188cc1 JA |
158 | struct sock *io_uring_get_socket(struct file *file) |
159 | { | |
160 | #if defined(CONFIG_UNIX) | |
cd40cae2 | 161 | if (io_is_uring_fops(file)) { |
2b188cc1 JA |
162 | struct io_ring_ctx *ctx = file->private_data; |
163 | ||
164 | return ctx->ring_sock->sk; | |
165 | } | |
166 | #endif | |
167 | return NULL; | |
168 | } | |
169 | EXPORT_SYMBOL(io_uring_get_socket); | |
170 | ||
c450178d PB |
171 | static inline void io_submit_flush_completions(struct io_ring_ctx *ctx) |
172 | { | |
931147dd DY |
173 | if (!wq_list_empty(&ctx->submit_state.compl_reqs) || |
174 | ctx->submit_state.cqes_count) | |
c450178d PB |
175 | __io_submit_flush_completions(ctx); |
176 | } | |
177 | ||
faf88dde PB |
178 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
179 | { | |
180 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
181 | } | |
182 | ||
0fc8c2ac DY |
183 | static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx) |
184 | { | |
185 | return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head); | |
186 | } | |
187 | ||
9cae36a0 JA |
188 | static bool io_match_linked(struct io_kiocb *head) |
189 | { | |
190 | struct io_kiocb *req; | |
191 | ||
192 | io_for_each_link(req, head) { | |
193 | if (req->flags & REQ_F_INFLIGHT) | |
194 | return true; | |
195 | } | |
196 | return false; | |
6af3f48b PB |
197 | } |
198 | ||
199 | /* | |
200 | * As io_match_task() but protected against racing with linked timeouts. | |
201 | * User must not hold timeout_lock. | |
202 | */ | |
329061d3 JA |
203 | bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, |
204 | bool cancel_all) | |
6af3f48b | 205 | { |
9cae36a0 JA |
206 | bool matched; |
207 | ||
6af3f48b PB |
208 | if (task && head->task != task) |
209 | return false; | |
9cae36a0 JA |
210 | if (cancel_all) |
211 | return true; | |
212 | ||
213 | if (head->flags & REQ_F_LINK_TIMEOUT) { | |
214 | struct io_ring_ctx *ctx = head->ctx; | |
215 | ||
216 | /* protect against races with linked timeouts */ | |
217 | spin_lock_irq(&ctx->timeout_lock); | |
218 | matched = io_match_linked(head); | |
219 | spin_unlock_irq(&ctx->timeout_lock); | |
220 | } else { | |
221 | matched = io_match_linked(head); | |
222 | } | |
223 | return matched; | |
6af3f48b PB |
224 | } |
225 | ||
a8295b98 HX |
226 | static inline void req_fail_link_node(struct io_kiocb *req, int res) |
227 | { | |
228 | req_set_fail(req); | |
97b388d7 | 229 | io_req_set_res(req, res, 0); |
a8295b98 HX |
230 | } |
231 | ||
fa05457a PB |
232 | static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) |
233 | { | |
234 | wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); | |
c1755c25 | 235 | kasan_poison_object_data(req_cachep, req); |
a8295b98 HX |
236 | } |
237 | ||
c072481d | 238 | static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) |
2b188cc1 JA |
239 | { |
240 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
241 | ||
0f158b4c | 242 | complete(&ctx->ref_comp); |
2b188cc1 JA |
243 | } |
244 | ||
c072481d | 245 | static __cold void io_fallback_req_func(struct work_struct *work) |
f56165e6 PB |
246 | { |
247 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
248 | fallback_work.work); | |
249 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
250 | struct io_kiocb *req, *tmp; | |
a282967c | 251 | struct io_tw_state ts = { .locked = true, }; |
f56165e6 | 252 | |
31f084b7 | 253 | mutex_lock(&ctx->uring_lock); |
3218e5d3 | 254 | llist_for_each_entry_safe(req, tmp, node, io_task_work.node) |
a282967c PB |
255 | req->io_task_work.func(req, &ts); |
256 | if (WARN_ON_ONCE(!ts.locked)) | |
31f084b7 PB |
257 | return; |
258 | io_submit_flush_completions(ctx); | |
259 | mutex_unlock(&ctx->uring_lock); | |
f56165e6 PB |
260 | } |
261 | ||
e6f89be6 PB |
262 | static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) |
263 | { | |
264 | unsigned hash_buckets = 1U << bits; | |
265 | size_t hash_size = hash_buckets * sizeof(table->hbs[0]); | |
266 | ||
267 | table->hbs = kmalloc(hash_size, GFP_KERNEL); | |
268 | if (!table->hbs) | |
269 | return -ENOMEM; | |
270 | ||
271 | table->hash_bits = bits; | |
272 | init_hash_table(table, hash_buckets); | |
273 | return 0; | |
274 | } | |
275 | ||
c072481d | 276 | static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
2b188cc1 JA |
277 | { |
278 | struct io_ring_ctx *ctx; | |
9cfc7e94 | 279 | int hash_bits; |
2b188cc1 JA |
280 | |
281 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
282 | if (!ctx) | |
283 | return NULL; | |
284 | ||
9cfc7e94 JA |
285 | xa_init(&ctx->io_bl_xa); |
286 | ||
78076bb6 JA |
287 | /* |
288 | * Use 5 bits less than the max cq entries, that should give us around | |
4a07723f PB |
289 | * 32 entries per hash list if totally full and uniformly spread, but |
290 | * don't keep too many buckets to not overconsume memory. | |
78076bb6 | 291 | */ |
4a07723f PB |
292 | hash_bits = ilog2(p->cq_entries) - 5; |
293 | hash_bits = clamp(hash_bits, 1, 8); | |
e6f89be6 | 294 | if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) |
78076bb6 | 295 | goto err; |
9ca9fb24 PB |
296 | if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) |
297 | goto err; | |
38513c46 | 298 | |
6224843d PB |
299 | ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL); |
300 | if (!ctx->dummy_ubuf) | |
301 | goto err; | |
302 | /* set invalid range, so io_import_fixed() fails meeting it */ | |
303 | ctx->dummy_ubuf->ubuf = -1UL; | |
304 | ||
21482896 | 305 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
48904229 | 306 | 0, GFP_KERNEL)) |
206aefde | 307 | goto err; |
2b188cc1 JA |
308 | |
309 | ctx->flags = p->flags; | |
90554200 | 310 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 311 | INIT_LIST_HEAD(&ctx->sqd_list); |
1d7bb1d5 | 312 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
cc3cec83 | 313 | INIT_LIST_HEAD(&ctx->io_buffers_cache); |
69bbc6ad PB |
314 | io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX, |
315 | sizeof(struct io_rsrc_node)); | |
316 | io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX, | |
317 | sizeof(struct async_poll)); | |
318 | io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX, | |
319 | sizeof(struct io_async_msghdr)); | |
0f158b4c | 320 | init_completion(&ctx->ref_comp); |
61cf9370 | 321 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 322 | mutex_init(&ctx->uring_lock); |
311997b3 | 323 | init_waitqueue_head(&ctx->cq_wait); |
7b235dd8 | 324 | init_waitqueue_head(&ctx->poll_wq); |
4ea15b56 | 325 | init_waitqueue_head(&ctx->rsrc_quiesce_wq); |
2b188cc1 | 326 | spin_lock_init(&ctx->completion_lock); |
89850fce | 327 | spin_lock_init(&ctx->timeout_lock); |
5eef4e87 | 328 | INIT_WQ_LIST(&ctx->iopoll_list); |
cc3cec83 JA |
329 | INIT_LIST_HEAD(&ctx->io_buffers_pages); |
330 | INIT_LIST_HEAD(&ctx->io_buffers_comp); | |
de0617e4 | 331 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 332 | INIT_LIST_HEAD(&ctx->timeout_list); |
ef9dd637 | 333 | INIT_LIST_HEAD(&ctx->ltimeout_list); |
d67d2263 | 334 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); |
c0e0d6ba | 335 | init_llist_head(&ctx->work_llist); |
13bf43f5 | 336 | INIT_LIST_HEAD(&ctx->tctx_list); |
c2b6c6bc PB |
337 | ctx->submit_state.free_list.next = NULL; |
338 | INIT_WQ_LIST(&ctx->locked_free_list); | |
9011bf9a | 339 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
6f33b0bc | 340 | INIT_WQ_LIST(&ctx->submit_state.compl_reqs); |
2b188cc1 | 341 | return ctx; |
206aefde | 342 | err: |
6224843d | 343 | kfree(ctx->dummy_ubuf); |
e6f89be6 | 344 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 345 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 JA |
346 | kfree(ctx->io_bl); |
347 | xa_destroy(&ctx->io_bl_xa); | |
206aefde JA |
348 | kfree(ctx); |
349 | return NULL; | |
2b188cc1 JA |
350 | } |
351 | ||
8f6ed49a PB |
352 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
353 | { | |
354 | struct io_rings *r = ctx->rings; | |
355 | ||
356 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
357 | ctx->cq_extra--; | |
358 | } | |
359 | ||
9cf7c104 | 360 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 361 | { |
2bc9930e JA |
362 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
363 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 364 | |
8f6ed49a | 365 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 366 | } |
de0617e4 | 367 | |
9d858b21 | 368 | return false; |
de0617e4 JA |
369 | } |
370 | ||
9cae36a0 JA |
371 | static inline void io_req_track_inflight(struct io_kiocb *req) |
372 | { | |
373 | if (!(req->flags & REQ_F_INFLIGHT)) { | |
374 | req->flags |= REQ_F_INFLIGHT; | |
386e4fb6 | 375 | atomic_inc(&req->task->io_uring->inflight_tracked); |
9cae36a0 JA |
376 | } |
377 | } | |
378 | ||
fd08e530 PB |
379 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
380 | { | |
906c6caa PB |
381 | if (WARN_ON_ONCE(!req->link)) |
382 | return NULL; | |
383 | ||
4d13d1a4 PB |
384 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
385 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
386 | |
387 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 388 | io_req_set_refcount(req); |
4d13d1a4 PB |
389 | __io_req_set_refcount(req->link, 2); |
390 | return req->link; | |
fd08e530 PB |
391 | } |
392 | ||
393 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
394 | { | |
4d13d1a4 | 395 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
396 | return NULL; |
397 | return __io_prep_linked_timeout(req); | |
398 | } | |
399 | ||
cb2d344c PB |
400 | static noinline void __io_arm_ltimeout(struct io_kiocb *req) |
401 | { | |
402 | io_queue_linked_timeout(__io_prep_linked_timeout(req)); | |
403 | } | |
404 | ||
405 | static inline void io_arm_ltimeout(struct io_kiocb *req) | |
406 | { | |
407 | if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) | |
408 | __io_arm_ltimeout(req); | |
409 | } | |
410 | ||
1e6fa521 JA |
411 | static void io_prep_async_work(struct io_kiocb *req) |
412 | { | |
a7dd2782 | 413 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
1e6fa521 JA |
414 | struct io_ring_ctx *ctx = req->ctx; |
415 | ||
b8e64b53 PB |
416 | if (!(req->flags & REQ_F_CREDS)) { |
417 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 418 | req->creds = get_current_cred(); |
b8e64b53 | 419 | } |
003e8dcc | 420 | |
e1d675df PB |
421 | req->work.list.next = NULL; |
422 | req->work.flags = 0; | |
8e29da69 | 423 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); |
feaadc4f PB |
424 | if (req->flags & REQ_F_FORCE_ASYNC) |
425 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
426 | ||
f6b543fd JA |
427 | if (req->file && !io_req_ffs_set(req)) |
428 | req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT; | |
429 | ||
8b1df11f | 430 | if (req->file && (req->flags & REQ_F_ISREG)) { |
d4755e15 JA |
431 | bool should_hash = def->hash_reg_file; |
432 | ||
433 | /* don't serialize this request if the fs doesn't need it */ | |
434 | if (should_hash && (req->file->f_flags & O_DIRECT) && | |
435 | (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE)) | |
436 | should_hash = false; | |
437 | if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL)) | |
1e6fa521 | 438 | io_wq_hash_work(&req->work, file_inode(req->file)); |
4b982bd0 | 439 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
440 | if (def->unbound_nonreg_file) |
441 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
442 | } | |
561fb04a | 443 | } |
cccf0ee8 | 444 | |
cbdcb435 | 445 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 446 | { |
cbdcb435 | 447 | struct io_kiocb *cur; |
54a91f3b | 448 | |
44eff40a PB |
449 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
450 | struct io_ring_ctx *ctx = req->ctx; | |
451 | ||
674ee8e1 | 452 | spin_lock_irq(&ctx->timeout_lock); |
44eff40a PB |
453 | io_for_each_link(cur, req) |
454 | io_prep_async_work(cur); | |
674ee8e1 | 455 | spin_unlock_irq(&ctx->timeout_lock); |
44eff40a PB |
456 | } else { |
457 | io_for_each_link(cur, req) | |
458 | io_prep_async_work(cur); | |
459 | } | |
561fb04a JA |
460 | } |
461 | ||
a282967c | 462 | void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use) |
561fb04a | 463 | { |
cbdcb435 | 464 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 465 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 466 | |
3bfe6106 JA |
467 | BUG_ON(!tctx); |
468 | BUG_ON(!tctx->io_wq); | |
561fb04a | 469 | |
cbdcb435 PB |
470 | /* init ->work of the whole link before punting */ |
471 | io_prep_async_link(req); | |
991468dc JA |
472 | |
473 | /* | |
474 | * Not expected to happen, but if we do have a bug where this _can_ | |
475 | * happen, catch it here and ensure the request is marked as | |
476 | * canceled. That will make io-wq go through the usual work cancel | |
477 | * procedure rather than attempt to run this request (or create a new | |
478 | * worker for it). | |
479 | */ | |
480 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
481 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
482 | ||
48863ffd | 483 | trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); |
ebf93667 | 484 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
485 | if (link) |
486 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
487 | } |
488 | ||
c072481d | 489 | static __cold void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 490 | { |
441b8a78 | 491 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
492 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
493 | struct io_defer_entry, list); | |
de0617e4 | 494 | |
9cf7c104 | 495 | if (req_need_defer(de->req, de->seq)) |
04518945 | 496 | break; |
27dc8338 | 497 | list_del_init(&de->list); |
907d1df3 | 498 | io_req_task_queue(de->req); |
27dc8338 | 499 | kfree(de); |
441b8a78 | 500 | } |
04518945 PB |
501 | } |
502 | ||
21a091b9 DY |
503 | |
504 | static void io_eventfd_ops(struct rcu_head *rcu) | |
f2842ab5 | 505 | { |
d8e9214f | 506 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
21a091b9 | 507 | int ops = atomic_xchg(&ev_fd->ops, 0); |
305bef98 | 508 | |
21a091b9 | 509 | if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) |
44648532 | 510 | eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); |
d8e9214f | 511 | |
21a091b9 DY |
512 | /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback |
513 | * ordering in a race but if references are 0 we know we have to free | |
514 | * it regardless. | |
305bef98 | 515 | */ |
21a091b9 DY |
516 | if (atomic_dec_and_test(&ev_fd->refs)) { |
517 | eventfd_ctx_put(ev_fd->cq_ev_fd); | |
518 | kfree(ev_fd); | |
519 | } | |
d8e9214f DY |
520 | } |
521 | ||
77bc59b4 | 522 | static void io_eventfd_signal(struct io_ring_ctx *ctx) |
f2842ab5 | 523 | { |
21a091b9 | 524 | struct io_ev_fd *ev_fd = NULL; |
77bc59b4 | 525 | |
77bc59b4 UA |
526 | rcu_read_lock(); |
527 | /* | |
528 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
529 | * and eventfd_signal | |
530 | */ | |
531 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
532 | ||
533 | /* | |
534 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
535 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
536 | * the function and rcu_read_lock. | |
537 | */ | |
538 | if (unlikely(!ev_fd)) | |
539 | goto out; | |
7e55a19c | 540 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 | 541 | goto out; |
21a091b9 DY |
542 | if (ev_fd->eventfd_async && !io_wq_current_is_worker()) |
543 | goto out; | |
77bc59b4 | 544 | |
21a091b9 | 545 | if (likely(eventfd_signal_allowed())) { |
44648532 | 546 | eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); |
21a091b9 DY |
547 | } else { |
548 | atomic_inc(&ev_fd->refs); | |
549 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) | |
44a84da4 | 550 | call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops); |
21a091b9 DY |
551 | else |
552 | atomic_dec(&ev_fd->refs); | |
553 | } | |
554 | ||
77bc59b4 UA |
555 | out: |
556 | rcu_read_unlock(); | |
f2842ab5 JA |
557 | } |
558 | ||
21a091b9 DY |
559 | static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) |
560 | { | |
561 | bool skip; | |
562 | ||
563 | spin_lock(&ctx->completion_lock); | |
564 | ||
565 | /* | |
566 | * Eventfd should only get triggered when at least one event has been | |
567 | * posted. Some applications rely on the eventfd notification count | |
568 | * only changing IFF a new CQE has been added to the CQ ring. There's | |
569 | * no depedency on 1:1 relationship between how many times this | |
570 | * function is called (and hence the eventfd count) and number of CQEs | |
571 | * posted to the CQ ring. | |
572 | */ | |
573 | skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
574 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
575 | spin_unlock(&ctx->completion_lock); | |
576 | if (skip) | |
577 | return; | |
578 | ||
579 | io_eventfd_signal(ctx); | |
580 | } | |
581 | ||
a830ffd2 PB |
582 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
583 | { | |
bca39f39 PB |
584 | if (ctx->poll_activated) |
585 | io_poll_wq_wake(ctx); | |
e5f30f6f PB |
586 | if (ctx->off_timeout_used) |
587 | io_flush_timeouts(ctx); | |
588 | if (ctx->drain_active) { | |
a830ffd2 | 589 | spin_lock(&ctx->completion_lock); |
e5f30f6f | 590 | io_queue_deferred(ctx); |
a830ffd2 PB |
591 | spin_unlock(&ctx->completion_lock); |
592 | } | |
593 | if (ctx->has_evfd) | |
21a091b9 | 594 | io_eventfd_flush_signal(ctx); |
a830ffd2 PB |
595 | } |
596 | ||
f66f7342 PB |
597 | static inline void __io_cq_lock(struct io_ring_ctx *ctx) |
598 | __acquires(ctx->completion_lock) | |
599 | { | |
600 | if (!ctx->task_complete) | |
601 | spin_lock(&ctx->completion_lock); | |
602 | } | |
603 | ||
604 | static inline void __io_cq_unlock(struct io_ring_ctx *ctx) | |
605 | { | |
606 | if (!ctx->task_complete) | |
607 | spin_unlock(&ctx->completion_lock); | |
608 | } | |
609 | ||
6971253f PB |
610 | static inline void io_cq_lock(struct io_ring_ctx *ctx) |
611 | __acquires(ctx->completion_lock) | |
612 | { | |
613 | spin_lock(&ctx->completion_lock); | |
614 | } | |
615 | ||
616 | static inline void io_cq_unlock(struct io_ring_ctx *ctx) | |
617 | __releases(ctx->completion_lock) | |
618 | { | |
619 | spin_unlock(&ctx->completion_lock); | |
620 | } | |
621 | ||
5d772916 | 622 | /* keep it inlined for io_submit_flush_completions() */ |
f66f7342 | 623 | static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) |
25399321 PB |
624 | __releases(ctx->completion_lock) |
625 | { | |
626 | io_commit_cqring(ctx); | |
f66f7342 | 627 | __io_cq_unlock(ctx); |
6c16fe3c JA |
628 | io_commit_cqring_flush(ctx); |
629 | io_cqring_wake(ctx); | |
25399321 PB |
630 | } |
631 | ||
c66ae3ec | 632 | static void __io_cq_unlock_post_flush(struct io_ring_ctx *ctx) |
3181e22f PB |
633 | __releases(ctx->completion_lock) |
634 | { | |
635 | io_commit_cqring(ctx); | |
3181e22f | 636 | |
c66ae3ec PB |
637 | if (ctx->task_complete) { |
638 | /* | |
639 | * ->task_complete implies that only current might be waiting | |
640 | * for CQEs, and obviously, we currently don't. No one is | |
641 | * waiting, wakeups are futile, skip them. | |
642 | */ | |
643 | io_commit_cqring_flush(ctx); | |
644 | } else { | |
645 | __io_cq_unlock(ctx); | |
646 | io_commit_cqring_flush(ctx); | |
6e7248ad | 647 | io_cqring_wake(ctx); |
3181e22f PB |
648 | } |
649 | } | |
650 | ||
25399321 | 651 | void io_cq_unlock_post(struct io_ring_ctx *ctx) |
5d772916 | 652 | __releases(ctx->completion_lock) |
25399321 | 653 | { |
f66f7342 PB |
654 | io_commit_cqring(ctx); |
655 | spin_unlock(&ctx->completion_lock); | |
656 | io_commit_cqring_flush(ctx); | |
657 | io_cqring_wake(ctx); | |
25399321 PB |
658 | } |
659 | ||
c4a2ed72 | 660 | /* Returns true if there are no backlogged entries after the flush */ |
a85381d8 PB |
661 | static void io_cqring_overflow_kill(struct io_ring_ctx *ctx) |
662 | { | |
663 | struct io_overflow_cqe *ocqe; | |
664 | LIST_HEAD(list); | |
665 | ||
666 | io_cq_lock(ctx); | |
667 | list_splice_init(&ctx->cq_overflow_list, &list); | |
668 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); | |
669 | io_cq_unlock(ctx); | |
670 | ||
671 | while (!list_empty(&list)) { | |
672 | ocqe = list_first_entry(&list, struct io_overflow_cqe, list); | |
673 | list_del(&ocqe->list); | |
674 | kfree(ocqe); | |
675 | } | |
676 | } | |
677 | ||
1b346e4a | 678 | static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
1d7bb1d5 | 679 | { |
e45a3e05 | 680 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 681 | |
a85381d8 | 682 | if (__io_cqring_events(ctx) == ctx->cq_entries) |
1b346e4a | 683 | return; |
1d7bb1d5 | 684 | |
e45a3e05 SR |
685 | if (ctx->flags & IORING_SETUP_CQE32) |
686 | cqe_size <<= 1; | |
687 | ||
25399321 | 688 | io_cq_lock(ctx); |
6c2450ae | 689 | while (!list_empty(&ctx->cq_overflow_list)) { |
aa1df3a3 | 690 | struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true); |
6c2450ae | 691 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 692 | |
a85381d8 | 693 | if (!cqe) |
1d7bb1d5 | 694 | break; |
6c2450ae PB |
695 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
696 | struct io_overflow_cqe, list); | |
a85381d8 | 697 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae PB |
698 | list_del(&ocqe->list); |
699 | kfree(ocqe); | |
1d7bb1d5 JA |
700 | } |
701 | ||
1b346e4a | 702 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 703 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 704 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 705 | } |
25399321 | 706 | io_cq_unlock_post(ctx); |
1d7bb1d5 JA |
707 | } |
708 | ||
52ea806a JA |
709 | static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx) |
710 | { | |
711 | /* iopoll syncs against uring_lock, not completion_lock */ | |
712 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
713 | mutex_lock(&ctx->uring_lock); | |
714 | __io_cqring_overflow_flush(ctx); | |
715 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
716 | mutex_unlock(&ctx->uring_lock); | |
717 | } | |
718 | ||
1b346e4a | 719 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 720 | { |
52ea806a JA |
721 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) |
722 | io_cqring_do_overflow_flush(ctx); | |
6c503150 PB |
723 | } |
724 | ||
5afa4650 PB |
725 | /* can be called by any task */ |
726 | static void io_put_task_remote(struct task_struct *task, int nr) | |
6a290a14 PB |
727 | { |
728 | struct io_uring_task *tctx = task->io_uring; | |
729 | ||
9d170164 | 730 | percpu_counter_sub(&tctx->inflight, nr); |
8d664282 | 731 | if (unlikely(atomic_read(&tctx->in_cancel))) |
9d170164 PB |
732 | wake_up(&tctx->wait); |
733 | put_task_struct_many(task, nr); | |
734 | } | |
735 | ||
5afa4650 PB |
736 | /* used by a task to put its own references */ |
737 | static void io_put_task_local(struct task_struct *task, int nr) | |
738 | { | |
739 | task->io_uring->cached_refs += nr; | |
740 | } | |
741 | ||
89800a2d PB |
742 | /* must to be called somewhat shortly after putting a request */ |
743 | static inline void io_put_task(struct task_struct *task, int nr) | |
744 | { | |
745 | if (likely(task == current)) | |
5afa4650 | 746 | io_put_task_local(task, nr); |
89800a2d | 747 | else |
5afa4650 | 748 | io_put_task_remote(task, nr); |
89800a2d PB |
749 | } |
750 | ||
63809137 | 751 | void io_task_refs_refill(struct io_uring_task *tctx) |
9a10867a PB |
752 | { |
753 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
754 | ||
755 | percpu_counter_add(&tctx->inflight, refill); | |
756 | refcount_add(refill, ¤t->usage); | |
757 | tctx->cached_refs += refill; | |
758 | } | |
759 | ||
3cc7fdb9 PB |
760 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
761 | { | |
762 | struct io_uring_task *tctx = task->io_uring; | |
763 | unsigned int refs = tctx->cached_refs; | |
764 | ||
765 | if (refs) { | |
766 | tctx->cached_refs = 0; | |
767 | percpu_counter_sub(&tctx->inflight, refs); | |
768 | put_task_struct_many(task, refs); | |
769 | } | |
770 | } | |
771 | ||
68494a65 PB |
772 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
773 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 774 | { |
cce4b8b0 | 775 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
776 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
777 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 778 | |
f26cc959 PB |
779 | lockdep_assert_held(&ctx->completion_lock); |
780 | ||
e45a3e05 SR |
781 | if (is_cqe32) |
782 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 783 | |
e45a3e05 | 784 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 785 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
786 | if (!ocqe) { |
787 | /* | |
788 | * If we're in ring overflow flush mode, or in task cancel mode, | |
789 | * or cannot allocate an overflow entry, then we need to drop it | |
790 | * on the floor. | |
791 | */ | |
8f6ed49a | 792 | io_account_cq_overflow(ctx); |
155bc950 | 793 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 794 | return false; |
2b188cc1 | 795 | } |
cce4b8b0 | 796 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 797 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 798 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 799 | |
cce4b8b0 | 800 | } |
d4d19c19 | 801 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
802 | ocqe->cqe.res = res; |
803 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
804 | if (is_cqe32) { |
805 | ocqe->cqe.big_cqe[0] = extra1; | |
806 | ocqe->cqe.big_cqe[1] = extra2; | |
807 | } | |
cce4b8b0 PB |
808 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
809 | return true; | |
2b188cc1 JA |
810 | } |
811 | ||
68494a65 PB |
812 | bool io_req_cqe_overflow(struct io_kiocb *req) |
813 | { | |
814 | if (!(req->flags & REQ_F_CQE32_INIT)) { | |
815 | req->extra1 = 0; | |
816 | req->extra2 = 0; | |
817 | } | |
818 | return io_cqring_event_overflow(req->ctx, req->cqe.user_data, | |
819 | req->cqe.res, req->cqe.flags, | |
820 | req->extra1, req->extra2); | |
821 | } | |
822 | ||
faf88dde PB |
823 | /* |
824 | * writes to the cq entry need to come after reading head; the | |
825 | * control dependency is enough as we're using WRITE_ONCE to | |
826 | * fill the cq entry | |
827 | */ | |
aa1df3a3 | 828 | struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow) |
faf88dde PB |
829 | { |
830 | struct io_rings *rings = ctx->rings; | |
831 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
832 | unsigned int free, queued, len; |
833 | ||
aa1df3a3 PB |
834 | /* |
835 | * Posting into the CQ when there are pending overflowed CQEs may break | |
836 | * ordering guarantees, which will affect links, F_MORE users and more. | |
837 | * Force overflow the completion. | |
838 | */ | |
839 | if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) | |
840 | return NULL; | |
faf88dde PB |
841 | |
842 | /* userspace may cheat modifying the tail, be safe and do min */ | |
843 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
844 | free = ctx->cq_entries - queued; | |
845 | /* we need a contiguous range, limit based on the current array offset */ | |
846 | len = min(free, ctx->cq_entries - off); | |
847 | if (!len) | |
848 | return NULL; | |
849 | ||
b3659a65 PB |
850 | if (ctx->flags & IORING_SETUP_CQE32) { |
851 | off <<= 1; | |
852 | len <<= 1; | |
853 | } | |
854 | ||
faf88dde PB |
855 | ctx->cqe_cached = &rings->cqes[off]; |
856 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
b3659a65 PB |
857 | |
858 | ctx->cached_cq_tail++; | |
faf88dde | 859 | ctx->cqe_cached++; |
b3659a65 PB |
860 | if (ctx->flags & IORING_SETUP_CQE32) |
861 | ctx->cqe_cached++; | |
862 | return &rings->cqes[off]; | |
faf88dde PB |
863 | } |
864 | ||
f66f7342 PB |
865 | static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, |
866 | u32 cflags) | |
bcda7baa | 867 | { |
cd94903d PB |
868 | struct io_uring_cqe *cqe; |
869 | ||
913a571a | 870 | ctx->cq_extra++; |
cd94903d PB |
871 | |
872 | /* | |
873 | * If we can't get a cq entry, userspace overflowed the | |
874 | * submission (by quite a lot). Increment the overflow count in | |
875 | * the ring. | |
876 | */ | |
877 | cqe = io_get_cqe(ctx); | |
878 | if (likely(cqe)) { | |
e0486f3f DY |
879 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
880 | ||
cd94903d PB |
881 | WRITE_ONCE(cqe->user_data, user_data); |
882 | WRITE_ONCE(cqe->res, res); | |
883 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
884 | |
885 | if (ctx->flags & IORING_SETUP_CQE32) { | |
886 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
887 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
888 | } | |
cd94903d PB |
889 | return true; |
890 | } | |
52120f0f | 891 | return false; |
bcda7baa JA |
892 | } |
893 | ||
931147dd DY |
894 | static void __io_flush_post_cqes(struct io_ring_ctx *ctx) |
895 | __must_hold(&ctx->uring_lock) | |
896 | { | |
897 | struct io_submit_state *state = &ctx->submit_state; | |
898 | unsigned int i; | |
899 | ||
900 | lockdep_assert_held(&ctx->uring_lock); | |
901 | for (i = 0; i < state->cqes_count; i++) { | |
902 | struct io_uring_cqe *cqe = &state->cqes[i]; | |
903 | ||
f66f7342 PB |
904 | if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) { |
905 | if (ctx->task_complete) { | |
906 | spin_lock(&ctx->completion_lock); | |
907 | io_cqring_event_overflow(ctx, cqe->user_data, | |
908 | cqe->res, cqe->flags, 0, 0); | |
909 | spin_unlock(&ctx->completion_lock); | |
910 | } else { | |
911 | io_cqring_event_overflow(ctx, cqe->user_data, | |
912 | cqe->res, cqe->flags, 0, 0); | |
913 | } | |
914 | } | |
931147dd DY |
915 | } |
916 | state->cqes_count = 0; | |
917 | } | |
918 | ||
b529c96a DY |
919 | static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, |
920 | bool allow_overflow) | |
d245bca6 PB |
921 | { |
922 | bool filled; | |
923 | ||
25399321 | 924 | io_cq_lock(ctx); |
f66f7342 PB |
925 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags); |
926 | if (!filled && allow_overflow) | |
927 | filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); | |
928 | ||
25399321 | 929 | io_cq_unlock_post(ctx); |
d245bca6 PB |
930 | return filled; |
931 | } | |
932 | ||
b529c96a DY |
933 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags) |
934 | { | |
935 | return __io_post_aux_cqe(ctx, user_data, res, cflags, true); | |
936 | } | |
937 | ||
d86eaed1 | 938 | bool io_aux_cqe(const struct io_kiocb *req, bool defer, s32 res, u32 cflags, |
9b8c5475 | 939 | bool allow_overflow) |
2b188cc1 | 940 | { |
d86eaed1 JA |
941 | struct io_ring_ctx *ctx = req->ctx; |
942 | u64 user_data = req->cqe.user_data; | |
9b8c5475 | 943 | struct io_uring_cqe *cqe; |
9b8c5475 DY |
944 | |
945 | if (!defer) | |
b529c96a | 946 | return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow); |
9b8c5475 | 947 | |
9b8c5475 DY |
948 | lockdep_assert_held(&ctx->uring_lock); |
949 | ||
003f242b | 950 | if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->submit_state.cqes)) { |
f66f7342 | 951 | __io_cq_lock(ctx); |
9b8c5475 DY |
952 | __io_flush_post_cqes(ctx); |
953 | /* no need to flush - flush is deferred */ | |
f66f7342 | 954 | __io_cq_unlock_post(ctx); |
9b8c5475 DY |
955 | } |
956 | ||
957 | /* For defered completions this is not as strict as it is otherwise, | |
958 | * however it's main job is to prevent unbounded posted completions, | |
959 | * and in that it works just as well. | |
960 | */ | |
961 | if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) | |
962 | return false; | |
963 | ||
964 | cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++]; | |
965 | cqe->user_data = user_data; | |
966 | cqe->res = res; | |
967 | cqe->flags = cflags; | |
968 | return true; | |
969 | } | |
970 | ||
ef8ae64f | 971 | static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
2b188cc1 | 972 | { |
fa18fa22 | 973 | struct io_ring_ctx *ctx = req->ctx; |
2ad4c6d0 | 974 | struct io_rsrc_node *rsrc_node = NULL; |
fa18fa22 PB |
975 | |
976 | io_cq_lock(ctx); | |
977 | if (!(req->flags & REQ_F_CQE_SKIP)) | |
a8cf95f9 | 978 | io_fill_cqe_req(ctx, req); |
fa18fa22 | 979 | |
c7dae4ba JA |
980 | /* |
981 | * If we're the last reference to this request, add to our locked | |
982 | * free_list cache. | |
983 | */ | |
de9b4cca | 984 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 985 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 986 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
987 | io_disarm_next(req); |
988 | if (req->link) { | |
989 | io_req_task_queue(req->link); | |
990 | req->link = NULL; | |
991 | } | |
992 | } | |
68a2cc1b PB |
993 | io_put_kbuf_comp(req); |
994 | io_dismantle_req(req); | |
2ad4c6d0 | 995 | rsrc_node = req->rsrc_node; |
8197b053 PB |
996 | /* |
997 | * Selected buffer deallocation in io_clean_op() assumes that | |
998 | * we don't hold ->completion_lock. Clean them here to avoid | |
999 | * deadlocks. | |
1000 | */ | |
5afa4650 | 1001 | io_put_task_remote(req->task, 1); |
c2b6c6bc | 1002 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 1003 | ctx->locked_free_nr++; |
180f829f | 1004 | } |
25399321 | 1005 | io_cq_unlock_post(ctx); |
2ad4c6d0 | 1006 | |
ef8ae64f PB |
1007 | if (rsrc_node) { |
1008 | io_ring_submit_lock(ctx, issue_flags); | |
1f2c8f61 | 1009 | io_put_rsrc_node(ctx, rsrc_node); |
ef8ae64f PB |
1010 | io_ring_submit_unlock(ctx, issue_flags); |
1011 | } | |
4e3d9ff9 JA |
1012 | } |
1013 | ||
1bec951c | 1014 | void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
bcda7baa | 1015 | { |
860e1c7f | 1016 | if (req->ctx->task_complete && req->ctx->submitter_task != current) { |
e6aeb272 PB |
1017 | req->io_task_work.func = io_req_task_complete; |
1018 | io_req_task_work_add(req); | |
1019 | } else if (!(issue_flags & IO_URING_F_UNLOCKED) || | |
1020 | !(req->ctx->flags & IORING_SETUP_IOPOLL)) { | |
ef8ae64f | 1021 | __io_req_complete_post(req, issue_flags); |
1bec951c PB |
1022 | } else { |
1023 | struct io_ring_ctx *ctx = req->ctx; | |
1024 | ||
1025 | mutex_lock(&ctx->uring_lock); | |
ef8ae64f | 1026 | __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED); |
1bec951c PB |
1027 | mutex_unlock(&ctx->uring_lock); |
1028 | } | |
0ddf92e8 JA |
1029 | } |
1030 | ||
973fc83f | 1031 | void io_req_defer_failed(struct io_kiocb *req, s32 res) |
e276ae34 | 1032 | __must_hold(&ctx->uring_lock) |
f41db273 | 1033 | { |
f30bd4d0 | 1034 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; |
a47b255e | 1035 | |
e276ae34 PB |
1036 | lockdep_assert_held(&req->ctx->uring_lock); |
1037 | ||
93d2bcd2 | 1038 | req_set_fail(req); |
97b388d7 | 1039 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
a47b255e PB |
1040 | if (def->fail) |
1041 | def->fail(req); | |
973fc83f | 1042 | io_req_complete_defer(req); |
f41db273 PB |
1043 | } |
1044 | ||
864ea921 PB |
1045 | /* |
1046 | * Don't initialise the fields below on every allocation, but do that in | |
1047 | * advance and keep them valid across allocations. | |
1048 | */ | |
1049 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
1050 | { | |
1051 | req->ctx = ctx; | |
1052 | req->link = NULL; | |
1053 | req->async_data = NULL; | |
1054 | /* not necessary, but safer to zero */ | |
cef216fc | 1055 | req->cqe.res = 0; |
864ea921 PB |
1056 | } |
1057 | ||
dac7a098 | 1058 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 1059 | struct io_submit_state *state) |
dac7a098 | 1060 | { |
79ebeaee | 1061 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1062 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 1063 | ctx->locked_free_nr = 0; |
79ebeaee | 1064 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
1065 | } |
1066 | ||
5d5901a3 PB |
1067 | /* |
1068 | * A request might get retired back into the request caches even before opcode | |
1069 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
1070 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
1071 | * and with extra caution to not get a request that is still worked on. | |
1072 | */ | |
bd1a3783 | 1073 | __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 1074 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 1075 | { |
864ea921 | 1076 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 1077 | void *reqs[IO_REQ_ALLOC_BATCH]; |
864ea921 | 1078 | int ret, i; |
e5d1bc0a | 1079 | |
23a5c43b PB |
1080 | /* |
1081 | * If we have more than a batch's worth of requests in our IRQ side | |
1082 | * locked cache, grab the lock and move them over to our submission | |
1083 | * side cache. | |
1084 | */ | |
a6d97a8a | 1085 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 1086 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 1087 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
1088 | return true; |
1089 | } | |
e5d1bc0a | 1090 | |
3ab665b7 | 1091 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 1092 | |
864ea921 PB |
1093 | /* |
1094 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1095 | * retry single alloc to be on the safe side. | |
1096 | */ | |
1097 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
1098 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
1099 | if (!reqs[0]) | |
a33ae9ce | 1100 | return false; |
864ea921 | 1101 | ret = 1; |
2b188cc1 | 1102 | } |
864ea921 | 1103 | |
37f0e767 | 1104 | percpu_ref_get_many(&ctx->refs, ret); |
3ab665b7 | 1105 | for (i = 0; i < ret; i++) { |
23a5c43b | 1106 | struct io_kiocb *req = reqs[i]; |
3ab665b7 PB |
1107 | |
1108 | io_preinit_req(req, ctx); | |
fa05457a | 1109 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 1110 | } |
a33ae9ce PB |
1111 | return true; |
1112 | } | |
1113 | ||
6b639522 | 1114 | static inline void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 1115 | { |
094bae49 | 1116 | unsigned int flags = req->flags; |
929a3af9 | 1117 | |
867f8fa5 | 1118 | if (unlikely(flags & IO_REQ_CLEAN_FLAGS)) |
3a0a6902 | 1119 | io_clean_op(req); |
e1d767f0 PB |
1120 | if (!(flags & REQ_F_FIXED_FILE)) |
1121 | io_put_file(req->file); | |
e65ef56d JA |
1122 | } |
1123 | ||
03adabe8 | 1124 | static __cold void io_free_req_tw(struct io_kiocb *req, struct io_tw_state *ts) |
c6ca97b3 | 1125 | { |
51a4cc11 | 1126 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 1127 | |
ef8ae64f PB |
1128 | if (req->rsrc_node) { |
1129 | io_tw_lock(ctx, ts); | |
1f2c8f61 | 1130 | io_put_rsrc_node(ctx, req->rsrc_node); |
ef8ae64f | 1131 | } |
216578e5 | 1132 | io_dismantle_req(req); |
5afa4650 | 1133 | io_put_task_remote(req->task, 1); |
c6ca97b3 | 1134 | |
79ebeaee | 1135 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1136 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
c34b025f | 1137 | ctx->locked_free_nr++; |
79ebeaee | 1138 | spin_unlock(&ctx->completion_lock); |
e65ef56d JA |
1139 | } |
1140 | ||
03adabe8 PB |
1141 | __cold void io_free_req(struct io_kiocb *req) |
1142 | { | |
1143 | req->io_task_work.func = io_free_req_tw; | |
1144 | io_req_task_work_add(req); | |
1145 | } | |
1146 | ||
d81499bf PB |
1147 | static void __io_req_find_next_prep(struct io_kiocb *req) |
1148 | { | |
1149 | struct io_ring_ctx *ctx = req->ctx; | |
d81499bf | 1150 | |
6971253f | 1151 | spin_lock(&ctx->completion_lock); |
305bef98 | 1152 | io_disarm_next(req); |
6971253f | 1153 | spin_unlock(&ctx->completion_lock); |
d81499bf PB |
1154 | } |
1155 | ||
1156 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 1157 | { |
33cc89a9 | 1158 | struct io_kiocb *nxt; |
944e58bf | 1159 | |
9e645e11 JA |
1160 | /* |
1161 | * If LINK is set, we have dependent requests in this chain. If we | |
1162 | * didn't fail this request, queue the first one up, moving any other | |
1163 | * dependencies to the next request. In case of failure, fail the rest | |
1164 | * of the chain. | |
1165 | */ | |
d81499bf PB |
1166 | if (unlikely(req->flags & IO_DISARM_MASK)) |
1167 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
1168 | nxt = req->link; |
1169 | req->link = NULL; | |
1170 | return nxt; | |
4d7dd462 | 1171 | } |
9e645e11 | 1172 | |
a282967c | 1173 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts) |
2c32395d PB |
1174 | { |
1175 | if (!ctx) | |
1176 | return; | |
ef060ea9 JA |
1177 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1178 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
a282967c | 1179 | if (ts->locked) { |
c450178d | 1180 | io_submit_flush_completions(ctx); |
2c32395d | 1181 | mutex_unlock(&ctx->uring_lock); |
a282967c | 1182 | ts->locked = false; |
2c32395d PB |
1183 | } |
1184 | percpu_ref_put(&ctx->refs); | |
1185 | } | |
1186 | ||
c6dd763c | 1187 | static unsigned int handle_tw_list(struct llist_node *node, |
a282967c PB |
1188 | struct io_ring_ctx **ctx, |
1189 | struct io_tw_state *ts, | |
c6dd763c | 1190 | struct llist_node *last) |
9f8d032a | 1191 | { |
c6dd763c DY |
1192 | unsigned int count = 0; |
1193 | ||
cb6bf7f2 | 1194 | while (node && node != last) { |
f88262e6 | 1195 | struct llist_node *next = node->next; |
9f8d032a HX |
1196 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1197 | io_task_work.node); | |
1198 | ||
34d2bfe7 JA |
1199 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); |
1200 | ||
9f8d032a | 1201 | if (req->ctx != *ctx) { |
a282967c | 1202 | ctx_flush_and_put(*ctx, ts); |
9f8d032a HX |
1203 | *ctx = req->ctx; |
1204 | /* if not contended, grab and improve batching */ | |
a282967c | 1205 | ts->locked = mutex_trylock(&(*ctx)->uring_lock); |
9f8d032a | 1206 | percpu_ref_get(&(*ctx)->refs); |
13bfa6f1 | 1207 | } |
c92fcfc2 JA |
1208 | INDIRECT_CALL_2(req->io_task_work.func, |
1209 | io_poll_task_func, io_req_rw_complete, | |
1210 | req, ts); | |
9f8d032a | 1211 | node = next; |
c6dd763c | 1212 | count++; |
f5868008 | 1213 | if (unlikely(need_resched())) { |
a282967c | 1214 | ctx_flush_and_put(*ctx, ts); |
f5868008 JA |
1215 | *ctx = NULL; |
1216 | cond_resched(); | |
1217 | } | |
3a0c037b | 1218 | } |
c6dd763c DY |
1219 | |
1220 | return count; | |
9f8d032a HX |
1221 | } |
1222 | ||
923d1592 DY |
1223 | /** |
1224 | * io_llist_xchg - swap all entries in a lock-less list | |
1225 | * @head: the head of lock-less list to delete all entries | |
1226 | * @new: new entry as the head of the list | |
1227 | * | |
1228 | * If list is empty, return NULL, otherwise, return the pointer to the first entry. | |
1229 | * The order of entries returned is from the newest to the oldest added one. | |
1230 | */ | |
1231 | static inline struct llist_node *io_llist_xchg(struct llist_head *head, | |
1232 | struct llist_node *new) | |
1233 | { | |
1234 | return xchg(&head->first, new); | |
1235 | } | |
1236 | ||
1237 | /** | |
1238 | * io_llist_cmpxchg - possibly swap all entries in a lock-less list | |
1239 | * @head: the head of lock-less list to delete all entries | |
1240 | * @old: expected old value of the first entry of the list | |
1241 | * @new: new entry as the head of the list | |
1242 | * | |
1243 | * perform a cmpxchg on the first entry of the list. | |
1244 | */ | |
1245 | ||
1246 | static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head, | |
1247 | struct llist_node *old, | |
1248 | struct llist_node *new) | |
1249 | { | |
1250 | return cmpxchg(&head->first, old, new); | |
1251 | } | |
1252 | ||
c9f06aa7 | 1253 | void tctx_task_work(struct callback_head *cb) |
c40f6379 | 1254 | { |
a282967c | 1255 | struct io_tw_state ts = {}; |
ebd0df2e | 1256 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
1257 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
1258 | task_work); | |
3a0c037b | 1259 | struct llist_node fake = {}; |
77e443ab | 1260 | struct llist_node *node; |
cb6bf7f2 PB |
1261 | unsigned int loops = 0; |
1262 | unsigned int count = 0; | |
3a0c037b | 1263 | |
77e443ab PB |
1264 | if (unlikely(current->flags & PF_EXITING)) { |
1265 | io_fallback_tw(tctx); | |
1266 | return; | |
1267 | } | |
3a0c037b | 1268 | |
cb6bf7f2 | 1269 | do { |
c6dd763c | 1270 | loops++; |
3a0c037b | 1271 | node = io_llist_xchg(&tctx->task_list, &fake); |
a282967c | 1272 | count += handle_tw_list(node, &ctx, &ts, &fake); |
50470fc5 PB |
1273 | |
1274 | /* skip expensive cmpxchg if there are items in the list */ | |
1275 | if (READ_ONCE(tctx->task_list.first) != &fake) | |
1276 | continue; | |
a282967c | 1277 | if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) { |
50470fc5 PB |
1278 | io_submit_flush_completions(ctx); |
1279 | if (READ_ONCE(tctx->task_list.first) != &fake) | |
1280 | continue; | |
1281 | } | |
3a0c037b | 1282 | node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); |
cb6bf7f2 | 1283 | } while (node != &fake); |
ebd0df2e | 1284 | |
a282967c | 1285 | ctx_flush_and_put(ctx, &ts); |
3cc7fdb9 | 1286 | |
8d664282 JA |
1287 | /* relaxed read is enough as only the task itself sets ->in_cancel */ |
1288 | if (unlikely(atomic_read(&tctx->in_cancel))) | |
3cc7fdb9 | 1289 | io_uring_drop_tctx_refs(current); |
c6dd763c DY |
1290 | |
1291 | trace_io_uring_task_work_run(tctx, count, loops); | |
7cbf1722 JA |
1292 | } |
1293 | ||
d7593606 PB |
1294 | static __cold void io_fallback_tw(struct io_uring_task *tctx) |
1295 | { | |
1296 | struct llist_node *node = llist_del_all(&tctx->task_list); | |
1297 | struct io_kiocb *req; | |
1298 | ||
1299 | while (node) { | |
1300 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1301 | node = node->next; | |
1302 | if (llist_add(&req->io_task_work.node, | |
1303 | &req->ctx->fallback_llist)) | |
1304 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1305 | } | |
1306 | } | |
1307 | ||
8751d154 | 1308 | static void io_req_local_work_add(struct io_kiocb *req, unsigned flags) |
c0e0d6ba DY |
1309 | { |
1310 | struct io_ring_ctx *ctx = req->ctx; | |
8751d154 | 1311 | unsigned nr_wait, nr_tw, nr_tw_prev; |
51509400 | 1312 | struct llist_node *first; |
c0e0d6ba | 1313 | |
8751d154 PB |
1314 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) |
1315 | flags &= ~IOU_F_TWQ_LAZY_WAKE; | |
ce8e04f6 | 1316 | |
51509400 PB |
1317 | first = READ_ONCE(ctx->work_llist.first); |
1318 | do { | |
8751d154 PB |
1319 | nr_tw_prev = 0; |
1320 | if (first) { | |
1321 | struct io_kiocb *first_req = container_of(first, | |
1322 | struct io_kiocb, | |
1323 | io_task_work.node); | |
1324 | /* | |
1325 | * Might be executed at any moment, rely on | |
1326 | * SLAB_TYPESAFE_BY_RCU to keep it alive. | |
1327 | */ | |
1328 | nr_tw_prev = READ_ONCE(first_req->nr_tw); | |
1329 | } | |
1330 | nr_tw = nr_tw_prev + 1; | |
1331 | /* Large enough to fail the nr_wait comparison below */ | |
1332 | if (!(flags & IOU_F_TWQ_LAZY_WAKE)) | |
1333 | nr_tw = -1U; | |
1334 | ||
1335 | req->nr_tw = nr_tw; | |
51509400 PB |
1336 | req->io_task_work.node.next = first; |
1337 | } while (!try_cmpxchg(&ctx->work_llist.first, &first, | |
1338 | &req->io_task_work.node)); | |
1339 | ||
8751d154 | 1340 | if (!first) { |
8751d154 PB |
1341 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1342 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1343 | if (ctx->has_evfd) | |
1344 | io_eventfd_signal(ctx); | |
c0e0d6ba DY |
1345 | } |
1346 | ||
8751d154 PB |
1347 | nr_wait = atomic_read(&ctx->cq_wait_nr); |
1348 | /* no one is waiting */ | |
1349 | if (!nr_wait) | |
1350 | return; | |
1351 | /* either not enough or the previous add has already woken it up */ | |
1352 | if (nr_wait > nr_tw || nr_tw_prev >= nr_wait) | |
1353 | return; | |
1354 | /* pairs with set_current_state() in io_cqring_wait() */ | |
1355 | smp_mb__after_atomic(); | |
1356 | wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE); | |
c0e0d6ba DY |
1357 | } |
1358 | ||
8501fe70 | 1359 | void __io_req_task_work_add(struct io_kiocb *req, unsigned flags) |
7cbf1722 | 1360 | { |
c34398a8 | 1361 | struct io_uring_task *tctx = req->task->io_uring; |
9f010507 | 1362 | struct io_ring_ctx *ctx = req->ctx; |
7cbf1722 | 1363 | |
8501fe70 PB |
1364 | if (!(flags & IOU_F_TWQ_FORCE_NORMAL) && |
1365 | (ctx->flags & IORING_SETUP_DEFER_TASKRUN)) { | |
d73a572d | 1366 | rcu_read_lock(); |
8751d154 | 1367 | io_req_local_work_add(req, flags); |
d73a572d | 1368 | rcu_read_unlock(); |
c0e0d6ba DY |
1369 | return; |
1370 | } | |
7cbf1722 JA |
1371 | |
1372 | /* task_work already pending, we're done */ | |
32d91f05 | 1373 | if (!llist_add(&req->io_task_work.node, &tctx->task_list)) |
e09ee510 | 1374 | return; |
7cbf1722 | 1375 | |
ef060ea9 JA |
1376 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1377 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1378 | ||
3fe07bcd | 1379 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1380 | return; |
2215bed9 | 1381 | |
d7593606 | 1382 | io_fallback_tw(tctx); |
c0e0d6ba DY |
1383 | } |
1384 | ||
1385 | static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) | |
1386 | { | |
1387 | struct llist_node *node; | |
1388 | ||
1389 | node = llist_del_all(&ctx->work_llist); | |
1390 | while (node) { | |
1391 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1392 | io_task_work.node); | |
1393 | ||
1394 | node = node->next; | |
8501fe70 | 1395 | __io_req_task_work_add(req, IOU_F_TWQ_FORCE_NORMAL); |
c0e0d6ba DY |
1396 | } |
1397 | } | |
1398 | ||
a282967c | 1399 | static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts) |
c0e0d6ba | 1400 | { |
c0e0d6ba | 1401 | struct llist_node *node; |
c3f4d39e | 1402 | unsigned int loops = 0; |
140102ae | 1403 | int ret = 0; |
c0e0d6ba | 1404 | |
140102ae | 1405 | if (WARN_ON_ONCE(ctx->submitter_task != current)) |
c0e0d6ba | 1406 | return -EEXIST; |
c3f4d39e PB |
1407 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1408 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
c0e0d6ba | 1409 | again: |
3af0356c JA |
1410 | /* |
1411 | * llists are in reverse order, flip it back the right way before | |
1412 | * running the pending items. | |
1413 | */ | |
1414 | node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL)); | |
c3f4d39e | 1415 | while (node) { |
c0e0d6ba DY |
1416 | struct llist_node *next = node->next; |
1417 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1418 | io_task_work.node); | |
1419 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); | |
c92fcfc2 JA |
1420 | INDIRECT_CALL_2(req->io_task_work.func, |
1421 | io_poll_task_func, io_req_rw_complete, | |
1422 | req, ts); | |
c0e0d6ba DY |
1423 | ret++; |
1424 | node = next; | |
1425 | } | |
c3f4d39e | 1426 | loops++; |
c0e0d6ba | 1427 | |
c3f4d39e | 1428 | if (!llist_empty(&ctx->work_llist)) |
c0e0d6ba | 1429 | goto again; |
a282967c | 1430 | if (ts->locked) { |
c0e0d6ba | 1431 | io_submit_flush_completions(ctx); |
b0b7a7d2 PB |
1432 | if (!llist_empty(&ctx->work_llist)) |
1433 | goto again; | |
1434 | } | |
f75d5036 | 1435 | trace_io_uring_local_work_run(ctx, ret, loops); |
c0e0d6ba | 1436 | return ret; |
8ac5d85a JA |
1437 | } |
1438 | ||
360173ab PB |
1439 | static inline int io_run_local_work_locked(struct io_ring_ctx *ctx) |
1440 | { | |
a282967c | 1441 | struct io_tw_state ts = { .locked = true, }; |
360173ab PB |
1442 | int ret; |
1443 | ||
1444 | if (llist_empty(&ctx->work_llist)) | |
1445 | return 0; | |
1446 | ||
a282967c | 1447 | ret = __io_run_local_work(ctx, &ts); |
360173ab | 1448 | /* shouldn't happen! */ |
a282967c | 1449 | if (WARN_ON_ONCE(!ts.locked)) |
360173ab PB |
1450 | mutex_lock(&ctx->uring_lock); |
1451 | return ret; | |
1452 | } | |
1453 | ||
3e565555 | 1454 | static int io_run_local_work(struct io_ring_ctx *ctx) |
8ac5d85a | 1455 | { |
a282967c | 1456 | struct io_tw_state ts = {}; |
8ac5d85a JA |
1457 | int ret; |
1458 | ||
a282967c PB |
1459 | ts.locked = mutex_trylock(&ctx->uring_lock); |
1460 | ret = __io_run_local_work(ctx, &ts); | |
1461 | if (ts.locked) | |
8ac5d85a JA |
1462 | mutex_unlock(&ctx->uring_lock); |
1463 | ||
1464 | return ret; | |
c0e0d6ba DY |
1465 | } |
1466 | ||
a282967c | 1467 | static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts) |
c40f6379 | 1468 | { |
a282967c | 1469 | io_tw_lock(req->ctx, ts); |
973fc83f | 1470 | io_req_defer_failed(req, req->cqe.res); |
c40f6379 JA |
1471 | } |
1472 | ||
a282967c | 1473 | void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts) |
c40f6379 | 1474 | { |
a282967c | 1475 | io_tw_lock(req->ctx, ts); |
316319e8 | 1476 | /* req->task == current here, checking PF_EXITING is safe */ |
6bb30855 | 1477 | if (unlikely(req->task->flags & PF_EXITING)) |
973fc83f | 1478 | io_req_defer_failed(req, -EFAULT); |
6bb30855 | 1479 | else if (req->flags & REQ_F_FORCE_ASYNC) |
a282967c | 1480 | io_queue_iowq(req, ts); |
6bb30855 DY |
1481 | else |
1482 | io_queue_sqe(req); | |
c40f6379 JA |
1483 | } |
1484 | ||
59915143 | 1485 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1486 | { |
97b388d7 | 1487 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1488 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1489 | io_req_task_work_add(req); |
c40f6379 JA |
1490 | } |
1491 | ||
f3b44f92 | 1492 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1493 | { |
5b0a6acc | 1494 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1495 | io_req_task_work_add(req); |
a3df7698 PB |
1496 | } |
1497 | ||
59915143 | 1498 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1499 | { |
9b5f7bd9 | 1500 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1501 | |
1502 | if (nxt) | |
906a8c3f | 1503 | io_req_task_queue(nxt); |
c69f8dbe JL |
1504 | } |
1505 | ||
f3b44f92 | 1506 | void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node) |
3aa83bfb | 1507 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1508 | { |
d4b7a5ef | 1509 | struct task_struct *task = NULL; |
37f0e767 | 1510 | int task_refs = 0; |
5af1d13e | 1511 | |
3aa83bfb PB |
1512 | do { |
1513 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1514 | comp_list); | |
2d6500d4 | 1515 | |
a538be5b PB |
1516 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1517 | if (req->flags & REQ_F_REFCOUNT) { | |
1518 | node = req->comp_list.next; | |
1519 | if (!req_ref_put_and_test(req)) | |
1520 | continue; | |
1521 | } | |
b605a7fa PB |
1522 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1523 | struct async_poll *apoll = req->apoll; | |
1524 | ||
1525 | if (apoll->double_poll) | |
1526 | kfree(apoll->double_poll); | |
9731bc98 JA |
1527 | if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) |
1528 | kfree(apoll); | |
b605a7fa PB |
1529 | req->flags &= ~REQ_F_POLLED; |
1530 | } | |
da1a08c5 | 1531 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1532 | io_queue_next(req); |
a538be5b PB |
1533 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1534 | io_clean_op(req); | |
c1e53a69 | 1535 | } |
a538be5b PB |
1536 | if (!(req->flags & REQ_F_FIXED_FILE)) |
1537 | io_put_file(req->file); | |
2d6500d4 | 1538 | |
ab409402 | 1539 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1540 | |
d4b7a5ef PB |
1541 | if (req->task != task) { |
1542 | if (task) | |
1543 | io_put_task(task, task_refs); | |
1544 | task = req->task; | |
1545 | task_refs = 0; | |
1546 | } | |
1547 | task_refs++; | |
c1e53a69 | 1548 | node = req->comp_list.next; |
fa05457a | 1549 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1550 | } while (node); |
d4b7a5ef | 1551 | |
d4b7a5ef PB |
1552 | if (task) |
1553 | io_put_task(task, task_refs); | |
7a743e22 PB |
1554 | } |
1555 | ||
c450178d | 1556 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1557 | __must_hold(&ctx->uring_lock) |
905c172f | 1558 | { |
cd0ca2e0 | 1559 | struct io_submit_state *state = &ctx->submit_state; |
fa780334 | 1560 | struct io_wq_work_node *node; |
905c172f | 1561 | |
f66f7342 | 1562 | __io_cq_lock(ctx); |
931147dd DY |
1563 | /* must come first to preserve CQE ordering in failure cases */ |
1564 | if (state->cqes_count) | |
1565 | __io_flush_post_cqes(ctx); | |
fa780334 | 1566 | __wq_list_for_each(node, &state->compl_reqs) { |
d9dee430 PB |
1567 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1568 | comp_list); | |
3d4aeb9f | 1569 | |
f66f7342 PB |
1570 | if (!(req->flags & REQ_F_CQE_SKIP) && |
1571 | unlikely(!__io_fill_cqe_req(ctx, req))) { | |
1572 | if (ctx->task_complete) { | |
1573 | spin_lock(&ctx->completion_lock); | |
1574 | io_req_cqe_overflow(req); | |
1575 | spin_unlock(&ctx->completion_lock); | |
1576 | } else { | |
1577 | io_req_cqe_overflow(req); | |
1578 | } | |
1579 | } | |
905c172f | 1580 | } |
3181e22f | 1581 | __io_cq_unlock_post_flush(ctx); |
d9dee430 | 1582 | |
931147dd DY |
1583 | if (!wq_list_empty(&ctx->submit_state.compl_reqs)) { |
1584 | io_free_batch_list(ctx, state->compl_reqs.first); | |
1585 | INIT_WQ_LIST(&state->compl_reqs); | |
1586 | } | |
7a743e22 PB |
1587 | } |
1588 | ||
ba816ad6 JA |
1589 | /* |
1590 | * Drop reference to request, return next in chain (if there is one) if this | |
1591 | * was the last reference to this request. | |
1592 | */ | |
0d85035a | 1593 | static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 1594 | { |
9b5f7bd9 PB |
1595 | struct io_kiocb *nxt = NULL; |
1596 | ||
de9b4cca | 1597 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 1598 | if (unlikely(req->flags & IO_REQ_LINK_FLAGS)) |
7819a1f6 | 1599 | nxt = io_req_find_next(req); |
f5c6cf2a | 1600 | io_free_req(req); |
2a44f467 | 1601 | } |
9b5f7bd9 | 1602 | return nxt; |
2b188cc1 JA |
1603 | } |
1604 | ||
6c503150 | 1605 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1606 | { |
1607 | /* See comment at the top of this file */ | |
1608 | smp_rmb(); | |
e23de15f | 1609 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1610 | } |
1611 | ||
def596e9 JA |
1612 | /* |
1613 | * We can't just wait for polled events to come to us, we have to actively | |
1614 | * find and complete them. | |
1615 | */ | |
c072481d | 1616 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1617 | { |
1618 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1619 | return; | |
1620 | ||
1621 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1622 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1623 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1624 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1625 | break; |
08f5439f JA |
1626 | /* |
1627 | * Ensure we allow local-to-the-cpu processing to take place, | |
1628 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1629 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1630 | */ |
3fcee5a6 PB |
1631 | if (need_resched()) { |
1632 | mutex_unlock(&ctx->uring_lock); | |
1633 | cond_resched(); | |
1634 | mutex_lock(&ctx->uring_lock); | |
1635 | } | |
def596e9 JA |
1636 | } |
1637 | mutex_unlock(&ctx->uring_lock); | |
1638 | } | |
1639 | ||
7668b92a | 1640 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1641 | { |
7668b92a | 1642 | unsigned int nr_events = 0; |
e9979b36 | 1643 | int ret = 0; |
155bc950 | 1644 | unsigned long check_cq; |
500f9fba | 1645 | |
76de6749 PB |
1646 | if (!io_allowed_run_tw(ctx)) |
1647 | return -EEXIST; | |
1648 | ||
3a08576b PB |
1649 | check_cq = READ_ONCE(ctx->check_cq); |
1650 | if (unlikely(check_cq)) { | |
1651 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
a85381d8 | 1652 | __io_cqring_overflow_flush(ctx); |
3a08576b PB |
1653 | /* |
1654 | * Similarly do not spin if we have not informed the user of any | |
1655 | * dropped CQE. | |
1656 | */ | |
1657 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1658 | return -EBADR; | |
1659 | } | |
f39c8a5b PB |
1660 | /* |
1661 | * Don't enter poll loop if we already have events pending. | |
1662 | * If we do, we can potentially be spinning for commands that | |
1663 | * already triggered a CQE (eg in error). | |
1664 | */ | |
f39c8a5b | 1665 | if (io_cqring_events(ctx)) |
d487b43c | 1666 | return 0; |
155bc950 | 1667 | |
def596e9 | 1668 | do { |
500f9fba JA |
1669 | /* |
1670 | * If a submit got punted to a workqueue, we can have the | |
1671 | * application entering polling for a command before it gets | |
1672 | * issued. That app will hold the uring_lock for the duration | |
1673 | * of the poll right here, so we need to take a breather every | |
1674 | * now and then to ensure that the issue has a chance to add | |
1675 | * the poll to the issued list. Otherwise we can spin here | |
1676 | * forever, while the workqueue is stuck trying to acquire the | |
1677 | * very same mutex. | |
1678 | */ | |
dac6a0ea JA |
1679 | if (wq_list_empty(&ctx->iopoll_list) || |
1680 | io_task_work_pending(ctx)) { | |
8f487ef2 PB |
1681 | u32 tail = ctx->cached_cq_tail; |
1682 | ||
8de11cdc | 1683 | (void) io_run_local_work_locked(ctx); |
def596e9 | 1684 | |
dac6a0ea JA |
1685 | if (task_work_pending(current) || |
1686 | wq_list_empty(&ctx->iopoll_list)) { | |
dac6a0ea | 1687 | mutex_unlock(&ctx->uring_lock); |
9d54bd6a | 1688 | io_run_task_work(); |
dac6a0ea | 1689 | mutex_lock(&ctx->uring_lock); |
dac6a0ea | 1690 | } |
8f487ef2 PB |
1691 | /* some requests don't go through iopoll_list */ |
1692 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1693 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1694 | break; |
500f9fba | 1695 | } |
5ba3c874 PB |
1696 | ret = io_do_iopoll(ctx, !min); |
1697 | if (ret < 0) | |
1698 | break; | |
1699 | nr_events += ret; | |
1700 | ret = 0; | |
1701 | } while (nr_events < min && !need_resched()); | |
d487b43c | 1702 | |
def596e9 JA |
1703 | return ret; |
1704 | } | |
7012c815 | 1705 | |
a282967c | 1706 | void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts) |
8ef12efe | 1707 | { |
a282967c | 1708 | if (ts->locked) |
9da070b1 | 1709 | io_req_complete_defer(req); |
7012c815 | 1710 | else |
27f35fe9 | 1711 | io_req_complete_post(req, IO_URING_F_UNLOCKED); |
8ef12efe JA |
1712 | } |
1713 | ||
def596e9 JA |
1714 | /* |
1715 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1716 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1717 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1718 | * accessing the kiocb cookie. |
1719 | */ | |
9882131c | 1720 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1721 | { |
1722 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1723 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1724 | |
1725 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1726 | if (unlikely(needs_lock)) |
cb3d8972 | 1727 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1728 | |
1729 | /* | |
1730 | * Track whether we have multiple files in our lists. This will impact | |
1731 | * how we do polling eventually, not spinning if we're on potentially | |
1732 | * different devices. | |
1733 | */ | |
5eef4e87 | 1734 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1735 | ctx->poll_multi_queue = false; |
1736 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1737 | struct io_kiocb *list_req; |
1738 | ||
5eef4e87 PB |
1739 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1740 | comp_list); | |
30da1b45 | 1741 | if (list_req->file != req->file) |
915b3dde | 1742 | ctx->poll_multi_queue = true; |
def596e9 JA |
1743 | } |
1744 | ||
1745 | /* | |
1746 | * For fast devices, IO may have already completed. If it has, add | |
1747 | * it to the front so we find it first. | |
1748 | */ | |
65a6543d | 1749 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1750 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1751 | else |
5eef4e87 | 1752 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1753 | |
3b44b371 | 1754 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1755 | /* |
1756 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1757 | * in sq thread task context or in io worker task context. If | |
1758 | * current task context is sq thread, we don't need to check | |
1759 | * whether should wake up sq thread. | |
1760 | */ | |
1761 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1762 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1763 | wake_up(&ctx->sq_data->wait); | |
1764 | ||
1765 | mutex_unlock(&ctx->uring_lock); | |
1766 | } | |
def596e9 JA |
1767 | } |
1768 | ||
2b188cc1 JA |
1769 | /* |
1770 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1771 | * any file. For now, just ensure that anything potentially problematic is done | |
1772 | * inline. | |
1773 | */ | |
88459b50 | 1774 | static bool __io_file_supports_nowait(struct file *file, umode_t mode) |
2b188cc1 | 1775 | { |
c5b85625 JA |
1776 | /* any ->read/write should understand O_NONBLOCK */ |
1777 | if (file->f_flags & O_NONBLOCK) | |
1778 | return true; | |
35645ac3 | 1779 | return file->f_mode & FMODE_NOWAIT; |
2b188cc1 | 1780 | } |
c5b85625 | 1781 | |
88459b50 PB |
1782 | /* |
1783 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1784 | * any file. For now, just ensure that anything potentially problematic is done | |
1785 | * inline. | |
1786 | */ | |
a4ad4f74 | 1787 | unsigned int io_file_get_flags(struct file *file) |
88459b50 PB |
1788 | { |
1789 | umode_t mode = file_inode(file)->i_mode; | |
1790 | unsigned int res = 0; | |
af197f50 | 1791 | |
88459b50 PB |
1792 | if (S_ISREG(mode)) |
1793 | res |= FFS_ISREG; | |
1794 | if (__io_file_supports_nowait(file, mode)) | |
1795 | res |= FFS_NOWAIT; | |
1796 | return res; | |
2b188cc1 JA |
1797 | } |
1798 | ||
99f15d8d | 1799 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1800 | { |
f30bd4d0 BL |
1801 | WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size); |
1802 | req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL); | |
d886e185 PB |
1803 | if (req->async_data) { |
1804 | req->flags |= REQ_F_ASYNC_DATA; | |
1805 | return false; | |
1806 | } | |
1807 | return true; | |
3d9932a8 XW |
1808 | } |
1809 | ||
f3b44f92 | 1810 | int io_req_prep_async(struct io_kiocb *req) |
f67676d1 | 1811 | { |
f30bd4d0 | 1812 | const struct io_cold_def *cdef = &io_cold_defs[req->opcode]; |
a7dd2782 | 1813 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
0702e536 JA |
1814 | |
1815 | /* assign early for deferred execution for non-fixed file */ | |
54aa7f23 | 1816 | if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file) |
0702e536 | 1817 | req->file = io_file_get_normal(req, req->cqe.fd); |
f30bd4d0 | 1818 | if (!cdef->prep_async) |
0702e536 JA |
1819 | return 0; |
1820 | if (WARN_ON_ONCE(req_has_async_data(req))) | |
1821 | return -EFAULT; | |
f30bd4d0 | 1822 | if (!def->manual_alloc) { |
59169439 PB |
1823 | if (io_alloc_async_data(req)) |
1824 | return -EAGAIN; | |
1825 | } | |
f30bd4d0 | 1826 | return cdef->prep_async(req); |
bfe76559 PB |
1827 | } |
1828 | ||
9cf7c104 PB |
1829 | static u32 io_get_sequence(struct io_kiocb *req) |
1830 | { | |
a3dbdf54 | 1831 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1832 | struct io_kiocb *cur; |
9cf7c104 | 1833 | |
a3dbdf54 | 1834 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1835 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1836 | seq--; |
1837 | return seq; | |
9cf7c104 PB |
1838 | } |
1839 | ||
c072481d | 1840 | static __cold void io_drain_req(struct io_kiocb *req) |
e276ae34 | 1841 | __must_hold(&ctx->uring_lock) |
de0617e4 | 1842 | { |
a197f664 | 1843 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1844 | struct io_defer_entry *de; |
f67676d1 | 1845 | int ret; |
e0eb71dc | 1846 | u32 seq = io_get_sequence(req); |
3c19966d | 1847 | |
9d858b21 | 1848 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1849 | spin_lock(&ctx->completion_lock); |
5e371265 | 1850 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1851 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1852 | queue: |
10c66904 | 1853 | ctx->drain_active = false; |
e0eb71dc PB |
1854 | io_req_task_queue(req); |
1855 | return; | |
10c66904 | 1856 | } |
e302f104 | 1857 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1858 | |
cbdcb435 | 1859 | io_prep_async_link(req); |
27dc8338 | 1860 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1861 | if (!de) { |
1b48773f | 1862 | ret = -ENOMEM; |
ef5c600a DY |
1863 | io_req_defer_failed(req, ret); |
1864 | return; | |
76cc33d7 | 1865 | } |
2d28390a | 1866 | |
79ebeaee | 1867 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1868 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1869 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1870 | kfree(de); |
e0eb71dc | 1871 | goto queue; |
de0617e4 JA |
1872 | } |
1873 | ||
48863ffd | 1874 | trace_io_uring_defer(req); |
27dc8338 | 1875 | de->req = req; |
9cf7c104 | 1876 | de->seq = seq; |
27dc8338 | 1877 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1878 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1879 | } |
1880 | ||
68fb8979 | 1881 | static void io_clean_op(struct io_kiocb *req) |
99bc4c38 | 1882 | { |
8197b053 PB |
1883 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
1884 | spin_lock(&req->ctx->completion_lock); | |
cc3cec83 | 1885 | io_put_kbuf_comp(req); |
8197b053 PB |
1886 | spin_unlock(&req->ctx->completion_lock); |
1887 | } | |
99bc4c38 | 1888 | |
0e1b6fe3 | 1889 | if (req->flags & REQ_F_NEED_CLEANUP) { |
f30bd4d0 | 1890 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; |
bb040a21 | 1891 | |
4d4c9cff JA |
1892 | if (def->cleanup) |
1893 | def->cleanup(req); | |
99bc4c38 | 1894 | } |
75652a30 JA |
1895 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1896 | kfree(req->apoll->double_poll); | |
1897 | kfree(req->apoll); | |
1898 | req->apoll = NULL; | |
1899 | } | |
9cae36a0 JA |
1900 | if (req->flags & REQ_F_INFLIGHT) { |
1901 | struct io_uring_task *tctx = req->task->io_uring; | |
1902 | ||
1903 | atomic_dec(&tctx->inflight_tracked); | |
1904 | } | |
c854357b | 1905 | if (req->flags & REQ_F_CREDS) |
b8e64b53 | 1906 | put_cred(req->creds); |
d886e185 PB |
1907 | if (req->flags & REQ_F_ASYNC_DATA) { |
1908 | kfree(req->async_data); | |
1909 | req->async_data = NULL; | |
1910 | } | |
c854357b | 1911 | req->flags &= ~IO_REQ_CLEAN_FLAGS; |
99bc4c38 PB |
1912 | } |
1913 | ||
f4992544 JA |
1914 | static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def, |
1915 | unsigned int issue_flags) | |
6bf9c47a | 1916 | { |
f4992544 | 1917 | if (req->file || !def->needs_file) |
6bf9c47a JA |
1918 | return true; |
1919 | ||
1920 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1921 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1922 | else |
cef216fc | 1923 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1924 | |
772f5e00 | 1925 | return !!req->file; |
6bf9c47a JA |
1926 | } |
1927 | ||
889fca73 | 1928 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1929 | { |
a7dd2782 | 1930 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
5730b27e | 1931 | const struct cred *creds = NULL; |
d625c6ee | 1932 | int ret; |
2b188cc1 | 1933 | |
f4992544 | 1934 | if (unlikely(!io_assign_file(req, def, issue_flags))) |
70152140 JA |
1935 | return -EBADF; |
1936 | ||
6878b40e | 1937 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1938 | creds = override_creds(req->creds); |
5730b27e | 1939 | |
fcde59fe | 1940 | if (!def->audit_skip) |
5bd2182d PM |
1941 | audit_uring_entry(req->opcode); |
1942 | ||
0702e536 | 1943 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1944 | |
fcde59fe | 1945 | if (!def->audit_skip) |
5bd2182d PM |
1946 | audit_uring_exit(!ret, ret); |
1947 | ||
5730b27e JA |
1948 | if (creds) |
1949 | revert_creds(creds); | |
97b388d7 | 1950 | |
75d7b3ae PB |
1951 | if (ret == IOU_OK) { |
1952 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
9da070b1 | 1953 | io_req_complete_defer(req); |
75d7b3ae | 1954 | else |
1bec951c | 1955 | io_req_complete_post(req, issue_flags); |
75d7b3ae | 1956 | } else if (ret != IOU_ISSUE_SKIP_COMPLETE) |
def596e9 | 1957 | return ret; |
97b388d7 | 1958 | |
b532576e | 1959 | /* If the op doesn't have a file, we're not polling for it */ |
ef0ec1ad | 1960 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue) |
9882131c | 1961 | io_iopoll_req_issued(req, issue_flags); |
def596e9 JA |
1962 | |
1963 | return 0; | |
2b188cc1 JA |
1964 | } |
1965 | ||
a282967c | 1966 | int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts) |
329061d3 | 1967 | { |
a282967c | 1968 | io_tw_lock(req->ctx, ts); |
9a692451 DY |
1969 | return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT| |
1970 | IO_URING_F_COMPLETE_DEFER); | |
329061d3 JA |
1971 | } |
1972 | ||
c9f06aa7 | 1973 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1974 | { |
1975 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1976 | ||
1977 | req = io_put_req_find_next(req); | |
1978 | return req ? &req->work : NULL; | |
1979 | } | |
1980 | ||
c9f06aa7 | 1981 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1982 | { |
1983 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
a7dd2782 | 1984 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
e6aeb272 | 1985 | unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ; |
d01905db | 1986 | bool needs_poll = false; |
6bf9c47a | 1987 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1988 | |
23a6c9ac | 1989 | /* one will be dropped by ->io_wq_free_work() after returning to io-wq */ |
48dcd38d PB |
1990 | if (!(req->flags & REQ_F_REFCOUNT)) |
1991 | __io_req_set_refcount(req, 2); | |
1992 | else | |
1993 | req_ref_get(req); | |
5d5901a3 | 1994 | |
cb2d344c | 1995 | io_arm_ltimeout(req); |
6bf9c47a | 1996 | |
dadebc35 | 1997 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1998 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1999 | fail: |
6bf9c47a | 2000 | io_req_task_queue_fail(req, err); |
d01905db PB |
2001 | return; |
2002 | } | |
f4992544 | 2003 | if (!io_assign_file(req, def, issue_flags)) { |
0f8da75b PB |
2004 | err = -EBADF; |
2005 | work->flags |= IO_WQ_WORK_CANCEL; | |
2006 | goto fail; | |
2007 | } | |
31b51510 | 2008 | |
d01905db | 2009 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
2010 | bool opcode_poll = def->pollin || def->pollout; |
2011 | ||
2012 | if (opcode_poll && file_can_poll(req->file)) { | |
2013 | needs_poll = true; | |
d01905db | 2014 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 2015 | } |
561fb04a | 2016 | } |
31b51510 | 2017 | |
d01905db PB |
2018 | do { |
2019 | ret = io_issue_sqe(req, issue_flags); | |
2020 | if (ret != -EAGAIN) | |
2021 | break; | |
2022 | /* | |
2023 | * We can get EAGAIN for iopolled IO even though we're | |
2024 | * forcing a sync submission from here, since we can't | |
2025 | * wait for request slots on the block side. | |
2026 | */ | |
2027 | if (!needs_poll) { | |
e0deb6a0 PB |
2028 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
2029 | break; | |
d01905db PB |
2030 | cond_resched(); |
2031 | continue; | |
90fa0288 HX |
2032 | } |
2033 | ||
4d9237e3 | 2034 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
2035 | return; |
2036 | /* aborted or ready, in either case retry blocking */ | |
2037 | needs_poll = false; | |
2038 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
2039 | } while (1); | |
31b51510 | 2040 | |
a3df7698 | 2041 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 2042 | if (ret < 0) |
a3df7698 | 2043 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
2044 | } |
2045 | ||
531113bb JA |
2046 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
2047 | unsigned int issue_flags) | |
09bb8394 | 2048 | { |
5106dd6e JA |
2049 | struct io_ring_ctx *ctx = req->ctx; |
2050 | struct file *file = NULL; | |
ac177053 | 2051 | unsigned long file_ptr; |
09bb8394 | 2052 | |
93f052cb | 2053 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 2054 | |
ac177053 | 2055 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 2056 | goto out; |
ac177053 PB |
2057 | fd = array_index_nospec(fd, ctx->nr_user_files); |
2058 | file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; | |
2059 | file = (struct file *) (file_ptr & FFS_MASK); | |
2060 | file_ptr &= ~FFS_MASK; | |
2061 | /* mask in overlapping REQ_F and FFS bits */ | |
35645ac3 | 2062 | req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT); |
5106dd6e JA |
2063 | io_req_set_rsrc_node(req, ctx, 0); |
2064 | out: | |
93f052cb | 2065 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
2066 | return file; |
2067 | } | |
d44f554e | 2068 | |
531113bb | 2069 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 2070 | { |
62906e89 | 2071 | struct file *file = fget(fd); |
ac177053 | 2072 | |
48863ffd | 2073 | trace_io_uring_file_get(req, fd); |
09bb8394 | 2074 | |
ac177053 | 2075 | /* we don't allow fixed io_uring files */ |
e5550a14 | 2076 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 2077 | io_req_track_inflight(req); |
8371adf5 | 2078 | return file; |
09bb8394 JA |
2079 | } |
2080 | ||
7bfa9bad | 2081 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
2082 | __must_hold(&req->ctx->uring_lock) |
2083 | { | |
7bfa9bad PB |
2084 | struct io_kiocb *linked_timeout; |
2085 | ||
2086 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
973fc83f | 2087 | io_req_defer_failed(req, ret); |
7bfa9bad PB |
2088 | return; |
2089 | } | |
2090 | ||
2091 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 2092 | |
4d9237e3 | 2093 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 2094 | case IO_APOLL_READY: |
336d28a8 | 2095 | io_kbuf_recycle(req, 0); |
d475a9a6 PB |
2096 | io_req_task_queue(req); |
2097 | break; | |
2098 | case IO_APOLL_ABORTED: | |
6436c770 | 2099 | io_kbuf_recycle(req, 0); |
77955efb | 2100 | io_queue_iowq(req, NULL); |
d475a9a6 | 2101 | break; |
b1c62645 | 2102 | case IO_APOLL_OK: |
b1c62645 | 2103 | break; |
d475a9a6 PB |
2104 | } |
2105 | ||
2106 | if (linked_timeout) | |
2107 | io_queue_linked_timeout(linked_timeout); | |
2108 | } | |
2109 | ||
cbc2e203 | 2110 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 2111 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 2112 | { |
e0c5c576 | 2113 | int ret; |
2b188cc1 | 2114 | |
c5eef2b9 | 2115 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 2116 | |
491381ce JA |
2117 | /* |
2118 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2119 | * doesn't support non-blocking read/write attempts | |
2120 | */ | |
7bfa9bad | 2121 | if (likely(!ret)) |
cb2d344c | 2122 | io_arm_ltimeout(req); |
7bfa9bad PB |
2123 | else |
2124 | io_queue_async(req, ret); | |
2b188cc1 JA |
2125 | } |
2126 | ||
4652fe3f | 2127 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 2128 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 2129 | { |
17b147f6 PB |
2130 | if (unlikely(req->flags & REQ_F_FAIL)) { |
2131 | /* | |
2132 | * We don't submit, fail them all, for that replace hardlinks | |
2133 | * with normal links. Extra REQ_F_LINK is tolerated. | |
2134 | */ | |
2135 | req->flags &= ~REQ_F_HARDLINK; | |
2136 | req->flags |= REQ_F_LINK; | |
973fc83f | 2137 | io_req_defer_failed(req, req->cqe.res); |
76cc33d7 PB |
2138 | } else { |
2139 | int ret = io_req_prep_async(req); | |
2140 | ||
ef5c600a | 2141 | if (unlikely(ret)) { |
973fc83f | 2142 | io_req_defer_failed(req, ret); |
ef5c600a DY |
2143 | return; |
2144 | } | |
2145 | ||
2146 | if (unlikely(req->ctx->drain_active)) | |
2147 | io_drain_req(req); | |
76cc33d7 | 2148 | else |
77955efb | 2149 | io_queue_iowq(req, NULL); |
ce35a47a | 2150 | } |
4fe2c963 JL |
2151 | } |
2152 | ||
b16fed66 PB |
2153 | /* |
2154 | * Check SQE restrictions (opcode and flags). | |
2155 | * | |
2156 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
2157 | */ | |
2158 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
2159 | struct io_kiocb *req, | |
2160 | unsigned int sqe_flags) | |
4fe2c963 | 2161 | { |
b16fed66 PB |
2162 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
2163 | return false; | |
2164 | ||
2165 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
2166 | ctx->restrictions.sqe_flags_required) | |
2167 | return false; | |
2168 | ||
2169 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
2170 | ctx->restrictions.sqe_flags_required)) | |
2171 | return false; | |
2172 | ||
2173 | return true; | |
4fe2c963 JL |
2174 | } |
2175 | ||
22b2ca31 PB |
2176 | static void io_init_req_drain(struct io_kiocb *req) |
2177 | { | |
2178 | struct io_ring_ctx *ctx = req->ctx; | |
2179 | struct io_kiocb *head = ctx->submit_state.link.head; | |
2180 | ||
2181 | ctx->drain_active = true; | |
2182 | if (head) { | |
2183 | /* | |
2184 | * If we need to drain a request in the middle of a link, drain | |
2185 | * the head request and the next request/link after the current | |
2186 | * link. Considering sequential execution of links, | |
b6c7db32 | 2187 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
2188 | * link. |
2189 | */ | |
b6c7db32 | 2190 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
2191 | ctx->drain_next = true; |
2192 | } | |
2193 | } | |
2194 | ||
b16fed66 PB |
2195 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2196 | const struct io_uring_sqe *sqe) | |
282cdc86 | 2197 | __must_hold(&ctx->uring_lock) |
b16fed66 | 2198 | { |
a7dd2782 | 2199 | const struct io_issue_def *def; |
b16fed66 | 2200 | unsigned int sqe_flags; |
fc0ae024 | 2201 | int personality; |
4a04d1d1 | 2202 | u8 opcode; |
b16fed66 | 2203 | |
864ea921 | 2204 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 2205 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 PB |
2206 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
2207 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
cef216fc | 2208 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 2209 | req->file = NULL; |
c1bdf8ed | 2210 | req->rsrc_node = NULL; |
b16fed66 | 2211 | req->task = current; |
b16fed66 | 2212 | |
4a04d1d1 PB |
2213 | if (unlikely(opcode >= IORING_OP_LAST)) { |
2214 | req->opcode = 0; | |
b16fed66 | 2215 | return -EINVAL; |
4a04d1d1 | 2216 | } |
a7dd2782 | 2217 | def = &io_issue_defs[opcode]; |
68fe256a PB |
2218 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
2219 | /* enforce forwards compatibility on users */ | |
2220 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
2221 | return -EINVAL; | |
4e906702 | 2222 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 2223 | if (!def->buffer_select) |
4e906702 JA |
2224 | return -EOPNOTSUPP; |
2225 | req->buf_index = READ_ONCE(sqe->buf_group); | |
2226 | } | |
5562a8d7 PB |
2227 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
2228 | ctx->drain_disabled = true; | |
2229 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
2230 | if (ctx->drain_disabled) | |
2231 | return -EOPNOTSUPP; | |
22b2ca31 | 2232 | io_init_req_drain(req); |
5562a8d7 | 2233 | } |
2a56a9bd PB |
2234 | } |
2235 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
2236 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
2237 | return -EACCES; | |
2238 | /* knock it to the slow queue path, will be drained there */ | |
2239 | if (ctx->drain_active) | |
2240 | req->flags |= REQ_F_FORCE_ASYNC; | |
2241 | /* if there is no link, we're at "next" request and need to drain */ | |
2242 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
2243 | ctx->drain_next = false; | |
2244 | ctx->drain_active = true; | |
b6c7db32 | 2245 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 2246 | } |
68fe256a | 2247 | } |
b16fed66 | 2248 | |
fcde59fe | 2249 | if (!def->ioprio && sqe->ioprio) |
73911426 | 2250 | return -EINVAL; |
fcde59fe | 2251 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
73911426 JA |
2252 | return -EINVAL; |
2253 | ||
fcde59fe | 2254 | if (def->needs_file) { |
6d63416d PB |
2255 | struct io_submit_state *state = &ctx->submit_state; |
2256 | ||
cef216fc | 2257 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 2258 | |
6d63416d PB |
2259 | /* |
2260 | * Plug now if we have more than 2 IO left after this, and the | |
2261 | * target is potentially a read/write to block based storage. | |
2262 | */ | |
fcde59fe | 2263 | if (state->need_plug && def->plug) { |
6d63416d PB |
2264 | state->plug_started = true; |
2265 | state->need_plug = false; | |
5ca7a8b3 | 2266 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 2267 | } |
b16fed66 | 2268 | } |
863e0560 | 2269 | |
003e8dcc JA |
2270 | personality = READ_ONCE(sqe->personality); |
2271 | if (personality) { | |
cdab10bf LT |
2272 | int ret; |
2273 | ||
c10d1f98 PB |
2274 | req->creds = xa_load(&ctx->personalities, personality); |
2275 | if (!req->creds) | |
003e8dcc | 2276 | return -EINVAL; |
c10d1f98 | 2277 | get_cred(req->creds); |
cdc1404a PM |
2278 | ret = security_uring_override_creds(req->creds); |
2279 | if (ret) { | |
2280 | put_cred(req->creds); | |
2281 | return ret; | |
2282 | } | |
b8e64b53 | 2283 | req->flags |= REQ_F_CREDS; |
003e8dcc | 2284 | } |
b16fed66 | 2285 | |
0702e536 | 2286 | return def->prep(req, sqe); |
b16fed66 PB |
2287 | } |
2288 | ||
df3becde PB |
2289 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
2290 | struct io_kiocb *req, int ret) | |
2291 | { | |
2292 | struct io_ring_ctx *ctx = req->ctx; | |
2293 | struct io_submit_link *link = &ctx->submit_state.link; | |
2294 | struct io_kiocb *head = link->head; | |
2295 | ||
48863ffd | 2296 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
2297 | |
2298 | /* | |
2299 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
2300 | * unusable. Instead of failing eagerly, continue assembling the link if | |
2301 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
2302 | * should find the flag and handle the rest. | |
2303 | */ | |
2304 | req_fail_link_node(req, ret); | |
2305 | if (head && !(head->flags & REQ_F_FAIL)) | |
2306 | req_fail_link_node(head, -ECANCELED); | |
2307 | ||
2308 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
2309 | if (head) { | |
2310 | link->last->link = req; | |
2311 | link->head = NULL; | |
2312 | req = head; | |
2313 | } | |
2314 | io_queue_sqe_fallback(req); | |
2315 | return ret; | |
2316 | } | |
2317 | ||
2318 | if (head) | |
2319 | link->last->link = req; | |
2320 | else | |
2321 | link->head = req; | |
2322 | link->last = req; | |
2323 | return 0; | |
2324 | } | |
2325 | ||
2326 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 2327 | const struct io_uring_sqe *sqe) |
282cdc86 | 2328 | __must_hold(&ctx->uring_lock) |
9e645e11 | 2329 | { |
a1ab7b35 | 2330 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 2331 | int ret; |
9e645e11 | 2332 | |
a6b8cadc | 2333 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
2334 | if (unlikely(ret)) |
2335 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 2336 | |
2ad57931 | 2337 | trace_io_uring_submit_req(req); |
a6b8cadc | 2338 | |
9e645e11 JA |
2339 | /* |
2340 | * If we already have a head request, queue this one for async | |
2341 | * submittal once the head completes. If we don't have a head but | |
2342 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2343 | * submitted sync once the chain is complete. If none of those | |
2344 | * conditions are true (normal request), then just queue it. | |
2345 | */ | |
924a07e4 | 2346 | if (unlikely(link->head)) { |
df3becde PB |
2347 | ret = io_req_prep_async(req); |
2348 | if (unlikely(ret)) | |
2349 | return io_submit_fail_init(sqe, req, ret); | |
2350 | ||
48863ffd | 2351 | trace_io_uring_link(req, link->head); |
f2f87370 | 2352 | link->last->link = req; |
863e0560 | 2353 | link->last = req; |
32fe525b | 2354 | |
da1a08c5 | 2355 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 2356 | return 0; |
df3becde PB |
2357 | /* last request of the link, flush it */ |
2358 | req = link->head; | |
f15a3431 | 2359 | link->head = NULL; |
924a07e4 PB |
2360 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
2361 | goto fallback; | |
2362 | ||
2363 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2364 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2365 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2366 | link->head = req; | |
2367 | link->last = req; | |
2368 | } else { | |
2369 | fallback: | |
2370 | io_queue_sqe_fallback(req); | |
2371 | } | |
f15a3431 | 2372 | return 0; |
9e645e11 | 2373 | } |
2e6e1fde | 2374 | |
f15a3431 | 2375 | io_queue_sqe(req); |
1d4240cc | 2376 | return 0; |
9e645e11 JA |
2377 | } |
2378 | ||
9a56a232 JA |
2379 | /* |
2380 | * Batched submission is done, ensure local IO is flushed out. | |
2381 | */ | |
553deffd | 2382 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2383 | { |
553deffd PB |
2384 | struct io_submit_state *state = &ctx->submit_state; |
2385 | ||
e126391c PB |
2386 | if (unlikely(state->link.head)) |
2387 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2388 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2389 | io_submit_flush_completions(ctx); |
27926b68 JA |
2390 | if (state->plug_started) |
2391 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2392 | } |
2393 | ||
2394 | /* | |
2395 | * Start submission side cache. | |
2396 | */ | |
2397 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2398 | unsigned int max_ios) |
9a56a232 | 2399 | { |
27926b68 | 2400 | state->plug_started = false; |
4b628aeb | 2401 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2402 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2403 | /* set only head, no need to init link_last in advance */ |
2404 | state->link.head = NULL; | |
9a56a232 JA |
2405 | } |
2406 | ||
2b188cc1 JA |
2407 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2408 | { | |
75b28aff | 2409 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2410 | |
caf582c6 PB |
2411 | /* |
2412 | * Ensure any loads from the SQEs are done at this point, | |
2413 | * since once we write the new head, the application could | |
2414 | * write new data to them. | |
2415 | */ | |
2416 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2417 | } |
2418 | ||
2b188cc1 | 2419 | /* |
dd9ae8a0 | 2420 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2421 | * that is mapped by userspace. This means that care needs to be taken to |
2422 | * ensure that reads are stable, as we cannot rely on userspace always | |
2423 | * being a good citizen. If members of the sqe are validated and then later | |
2424 | * used, it's important that those reads are done through READ_ONCE() to | |
2425 | * prevent a re-load down the line. | |
2426 | */ | |
b5083dfa | 2427 | static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe) |
2b188cc1 | 2428 | { |
ea5ab3b5 | 2429 | unsigned head, mask = ctx->sq_entries - 1; |
17d3aeb3 | 2430 | unsigned sq_idx = ctx->cached_sq_head++ & mask; |
2b188cc1 JA |
2431 | |
2432 | /* | |
2433 | * The cached sq head (or cq tail) serves two purposes: | |
2434 | * | |
2435 | * 1) allows us to batch the cost of updating the user visible | |
2436 | * head updates. | |
2437 | * 2) allows the kernel side to track the head on its own, even | |
2438 | * though the application is the one updating it. | |
2439 | */ | |
17d3aeb3 | 2440 | head = READ_ONCE(ctx->sq_array[sq_idx]); |
ebdeb7c0 JA |
2441 | if (likely(head < ctx->sq_entries)) { |
2442 | /* double index for 128-byte SQEs, twice as long */ | |
2443 | if (ctx->flags & IORING_SETUP_SQE128) | |
2444 | head <<= 1; | |
b5083dfa PB |
2445 | *sqe = &ctx->sq_sqes[head]; |
2446 | return true; | |
ebdeb7c0 | 2447 | } |
2b188cc1 JA |
2448 | |
2449 | /* drop invalid entries */ | |
15641e42 PB |
2450 | ctx->cq_extra--; |
2451 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2452 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
b5083dfa | 2453 | return false; |
709b302f PB |
2454 | } |
2455 | ||
17437f31 | 2456 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2457 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2458 | { |
69629809 | 2459 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2460 | unsigned int left; |
2461 | int ret; | |
6c271ce2 | 2462 | |
51d48dab | 2463 | if (unlikely(!entries)) |
69629809 | 2464 | return 0; |
ee7d46d9 | 2465 | /* make sure SQ entry isn't read before tail */ |
e3ef728f | 2466 | ret = left = min(nr, entries); |
8e6971a8 PB |
2467 | io_get_task_refs(left); |
2468 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2469 | |
69629809 | 2470 | do { |
3529d8c2 | 2471 | const struct io_uring_sqe *sqe; |
196be95c | 2472 | struct io_kiocb *req; |
fb5ccc98 | 2473 | |
c8576f3e | 2474 | if (unlikely(!io_alloc_req(ctx, &req))) |
fb5ccc98 | 2475 | break; |
b5083dfa | 2476 | if (unlikely(!io_get_sqe(ctx, &sqe))) { |
fa05457a | 2477 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2478 | break; |
2479 | } | |
6c271ce2 | 2480 | |
1cd15904 PB |
2481 | /* |
2482 | * Continue submitting even for sqe failure if the | |
2483 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2484 | */ | |
2485 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2486 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2487 | left--; | |
2488 | break; | |
bcbb7bf6 | 2489 | } |
1cd15904 | 2490 | } while (--left); |
9466f437 | 2491 | |
8e6971a8 PB |
2492 | if (unlikely(left)) { |
2493 | ret -= left; | |
2494 | /* try again if it submitted nothing and can't allocate a req */ | |
2495 | if (!ret && io_req_cache_empty(ctx)) | |
2496 | ret = -EAGAIN; | |
2497 | current->io_uring->cached_refs += left; | |
9466f437 | 2498 | } |
6c271ce2 | 2499 | |
553deffd | 2500 | io_submit_state_end(ctx); |
ae9428ca PB |
2501 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2502 | io_commit_sqring(ctx); | |
8e6971a8 | 2503 | return ret; |
6c271ce2 JA |
2504 | } |
2505 | ||
bda52162 JA |
2506 | struct io_wait_queue { |
2507 | struct wait_queue_entry wq; | |
2508 | struct io_ring_ctx *ctx; | |
5fd46178 | 2509 | unsigned cq_tail; |
bda52162 | 2510 | unsigned nr_timeouts; |
d33a39e5 | 2511 | ktime_t timeout; |
bda52162 JA |
2512 | }; |
2513 | ||
b4c98d59 DY |
2514 | static inline bool io_has_work(struct io_ring_ctx *ctx) |
2515 | { | |
c0e0d6ba | 2516 | return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) || |
490c00eb | 2517 | !llist_empty(&ctx->work_llist); |
b4c98d59 DY |
2518 | } |
2519 | ||
6c503150 | 2520 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
2521 | { |
2522 | struct io_ring_ctx *ctx = iowq->ctx; | |
0fc8c2ac | 2523 | int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail; |
bda52162 JA |
2524 | |
2525 | /* | |
d195a66e | 2526 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
2527 | * started waiting. For timeouts, we always want to return to userspace, |
2528 | * regardless of event count. | |
2529 | */ | |
5fd46178 | 2530 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
2531 | } |
2532 | ||
2533 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2534 | int wake_flags, void *key) | |
2535 | { | |
bd550173 | 2536 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq); |
bda52162 | 2537 | |
6c503150 PB |
2538 | /* |
2539 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2540 | * the task, and the next invocation will do it. | |
2541 | */ | |
bd550173 | 2542 | if (io_should_wake(iowq) || io_has_work(iowq->ctx)) |
6c503150 PB |
2543 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2544 | return -1; | |
bda52162 JA |
2545 | } |
2546 | ||
c0e0d6ba | 2547 | int io_run_task_work_sig(struct io_ring_ctx *ctx) |
af9c1a44 | 2548 | { |
1414d629 | 2549 | if (!llist_empty(&ctx->work_llist)) { |
2f413956 | 2550 | __set_current_state(TASK_RUNNING); |
1414d629 PB |
2551 | if (io_run_local_work(ctx) > 0) |
2552 | return 1; | |
2553 | } | |
2554 | if (io_run_task_work() > 0) | |
af9c1a44 | 2555 | return 1; |
c5020bc8 OL |
2556 | if (task_sigpending(current)) |
2557 | return -EINTR; | |
2558 | return 0; | |
af9c1a44 JA |
2559 | } |
2560 | ||
eeb60b9a PB |
2561 | /* when returns >0, the caller should retry */ |
2562 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
d33a39e5 | 2563 | struct io_wait_queue *iowq) |
eeb60b9a | 2564 | { |
3fcf19d5 PB |
2565 | if (unlikely(READ_ONCE(ctx->check_cq))) |
2566 | return 1; | |
846072f1 PB |
2567 | if (unlikely(!llist_empty(&ctx->work_llist))) |
2568 | return 1; | |
2569 | if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) | |
2570 | return 1; | |
2571 | if (unlikely(task_sigpending(current))) | |
2572 | return -EINTR; | |
2573 | if (unlikely(io_should_wake(iowq))) | |
2574 | return 0; | |
d33a39e5 | 2575 | if (iowq->timeout == KTIME_MAX) |
46ae7eef | 2576 | schedule(); |
d33a39e5 | 2577 | else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS)) |
22833966 | 2578 | return -ETIME; |
846072f1 | 2579 | return 0; |
eeb60b9a PB |
2580 | } |
2581 | ||
2b188cc1 JA |
2582 | /* |
2583 | * Wait until events become available, if we don't already have some. The | |
2584 | * application must reap them itself, as they reside on the shared cq ring. | |
2585 | */ | |
2586 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2587 | const sigset_t __user *sig, size_t sigsz, |
2588 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2589 | { |
90291099 | 2590 | struct io_wait_queue iowq; |
75b28aff | 2591 | struct io_rings *rings = ctx->rings; |
c1d5a224 | 2592 | int ret; |
2b188cc1 | 2593 | |
76de6749 PB |
2594 | if (!io_allowed_run_tw(ctx)) |
2595 | return -EEXIST; | |
140102ae PB |
2596 | if (!llist_empty(&ctx->work_llist)) |
2597 | io_run_local_work(ctx); | |
f36ba6cf PB |
2598 | io_run_task_work(); |
2599 | io_cqring_overflow_flush(ctx); | |
2600 | /* if user messes with these they will just get an early return */ | |
2601 | if (__io_cqring_events_user(ctx) >= min_events) | |
2602 | return 0; | |
2b188cc1 JA |
2603 | |
2604 | if (sig) { | |
9e75ad5d AB |
2605 | #ifdef CONFIG_COMPAT |
2606 | if (in_compat_syscall()) | |
2607 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2608 | sigsz); |
9e75ad5d AB |
2609 | else |
2610 | #endif | |
b772434b | 2611 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2612 | |
2b188cc1 JA |
2613 | if (ret) |
2614 | return ret; | |
2615 | } | |
2616 | ||
90291099 PB |
2617 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
2618 | iowq.wq.private = current; | |
2619 | INIT_LIST_HEAD(&iowq.wq.entry); | |
2620 | iowq.ctx = ctx; | |
bda52162 | 2621 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 2622 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
d33a39e5 PB |
2623 | iowq.timeout = KTIME_MAX; |
2624 | ||
2625 | if (uts) { | |
2626 | struct timespec64 ts; | |
2627 | ||
2628 | if (get_timespec64(&ts, uts)) | |
2629 | return -EFAULT; | |
2630 | iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); | |
2631 | } | |
90291099 | 2632 | |
c826bd7a | 2633 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2634 | do { |
3fcf19d5 PB |
2635 | unsigned long check_cq; |
2636 | ||
130bd686 | 2637 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
8751d154 PB |
2638 | int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail); |
2639 | ||
2640 | atomic_set(&ctx->cq_wait_nr, nr_wait); | |
130bd686 PB |
2641 | set_current_state(TASK_INTERRUPTIBLE); |
2642 | } else { | |
2643 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, | |
2644 | TASK_INTERRUPTIBLE); | |
2645 | } | |
2646 | ||
d33a39e5 | 2647 | ret = io_cqring_wait_schedule(ctx, &iowq); |
130bd686 | 2648 | __set_current_state(TASK_RUNNING); |
8751d154 | 2649 | atomic_set(&ctx->cq_wait_nr, 0); |
d80c0f00 | 2650 | |
846072f1 PB |
2651 | if (ret < 0) |
2652 | break; | |
2653 | /* | |
2654 | * Run task_work after scheduling and before io_should_wake(). | |
2655 | * If we got woken because of task_work being processed, run it | |
2656 | * now rather than let the caller do another wait loop. | |
2657 | */ | |
2658 | io_run_task_work(); | |
2659 | if (!llist_empty(&ctx->work_llist)) | |
2660 | io_run_local_work(ctx); | |
3fcf19d5 PB |
2661 | |
2662 | check_cq = READ_ONCE(ctx->check_cq); | |
2663 | if (unlikely(check_cq)) { | |
2664 | /* let the caller flush overflows, retry */ | |
326a9e48 | 2665 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) |
3fcf19d5 | 2666 | io_cqring_do_overflow_flush(ctx); |
3fcf19d5 PB |
2667 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) { |
2668 | ret = -EBADR; | |
2669 | break; | |
2670 | } | |
2671 | } | |
2672 | ||
846072f1 PB |
2673 | if (io_should_wake(&iowq)) { |
2674 | ret = 0; | |
35d90f95 | 2675 | break; |
846072f1 | 2676 | } |
ca0a2651 | 2677 | cond_resched(); |
846072f1 | 2678 | } while (1); |
bda52162 | 2679 | |
130bd686 PB |
2680 | if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) |
2681 | finish_wait(&ctx->cq_wait, &iowq.wq); | |
b7db41c9 | 2682 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2683 | |
75b28aff | 2684 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2685 | } |
2686 | ||
73572984 | 2687 | static void io_mem_free(void *ptr) |
b9bd2bea | 2688 | { |
73572984 | 2689 | struct page *page; |
b36a2050 | 2690 | |
73572984 JA |
2691 | if (!ptr) |
2692 | return; | |
b9bd2bea | 2693 | |
73572984 JA |
2694 | page = virt_to_head_page(ptr); |
2695 | if (put_page_testzero(page)) | |
2696 | free_compound_page(page); | |
b9bd2bea PB |
2697 | } |
2698 | ||
03d89a2d JA |
2699 | static void io_pages_free(struct page ***pages, int npages) |
2700 | { | |
2701 | struct page **page_array; | |
2702 | int i; | |
2703 | ||
2704 | if (!pages) | |
2705 | return; | |
2706 | page_array = *pages; | |
2707 | for (i = 0; i < npages; i++) | |
2708 | unpin_user_page(page_array[i]); | |
2709 | kvfree(page_array); | |
2710 | *pages = NULL; | |
2711 | } | |
2712 | ||
2713 | static void *__io_uaddr_map(struct page ***pages, unsigned short *npages, | |
2714 | unsigned long uaddr, size_t size) | |
2715 | { | |
2716 | struct page **page_array; | |
2717 | unsigned int nr_pages; | |
2718 | int ret; | |
2719 | ||
2720 | *npages = 0; | |
2721 | ||
2722 | if (uaddr & (PAGE_SIZE - 1) || !size) | |
2723 | return ERR_PTR(-EINVAL); | |
2724 | ||
2725 | nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2726 | if (nr_pages > USHRT_MAX) | |
2727 | return ERR_PTR(-EINVAL); | |
2728 | page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); | |
2729 | if (!page_array) | |
2730 | return ERR_PTR(-ENOMEM); | |
2731 | ||
2732 | ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM, | |
2733 | page_array); | |
2734 | if (ret != nr_pages) { | |
2735 | err: | |
2736 | io_pages_free(&page_array, ret > 0 ? ret : 0); | |
2737 | return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT); | |
2738 | } | |
2739 | /* | |
2740 | * Should be a single page. If the ring is small enough that we can | |
2741 | * use a normal page, that is fine. If we need multiple pages, then | |
2742 | * userspace should use a huge page. That's the only way to guarantee | |
2743 | * that we get contigious memory, outside of just being lucky or | |
2744 | * (currently) having low memory fragmentation. | |
2745 | */ | |
2746 | if (page_array[0] != page_array[ret - 1]) | |
2747 | goto err; | |
2748 | *pages = page_array; | |
2749 | *npages = nr_pages; | |
2750 | return page_to_virt(page_array[0]); | |
2751 | } | |
2752 | ||
2753 | static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr, | |
2754 | size_t size) | |
2755 | { | |
2756 | return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr, | |
2757 | size); | |
2758 | } | |
2759 | ||
2760 | static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr, | |
2761 | size_t size) | |
2762 | { | |
2763 | return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr, | |
2764 | size); | |
2765 | } | |
2766 | ||
9c189eee JA |
2767 | static void io_rings_free(struct io_ring_ctx *ctx) |
2768 | { | |
03d89a2d JA |
2769 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) { |
2770 | io_mem_free(ctx->rings); | |
2771 | io_mem_free(ctx->sq_sqes); | |
2772 | ctx->rings = NULL; | |
2773 | ctx->sq_sqes = NULL; | |
2774 | } else { | |
2775 | io_pages_free(&ctx->ring_pages, ctx->n_ring_pages); | |
2776 | io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages); | |
2777 | } | |
9c189eee JA |
2778 | } |
2779 | ||
73572984 | 2780 | static void *io_mem_alloc(size_t size) |
b9bd2bea | 2781 | { |
73572984 | 2782 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; |
e27cef86 | 2783 | void *ret; |
b9bd2bea | 2784 | |
e27cef86 JA |
2785 | ret = (void *) __get_free_pages(gfp, get_order(size)); |
2786 | if (ret) | |
2787 | return ret; | |
2788 | return ERR_PTR(-ENOMEM); | |
b9bd2bea PB |
2789 | } |
2790 | ||
73572984 JA |
2791 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2792 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2793 | { |
73572984 JA |
2794 | struct io_rings *rings; |
2795 | size_t off, sq_array_size; | |
6b06314c | 2796 | |
73572984 JA |
2797 | off = struct_size(rings, cqes, cq_entries); |
2798 | if (off == SIZE_MAX) | |
2799 | return SIZE_MAX; | |
2800 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2801 | if (check_shl_overflow(off, 1, &off)) | |
2802 | return SIZE_MAX; | |
2803 | } | |
ab409402 | 2804 | |
73572984 JA |
2805 | #ifdef CONFIG_SMP |
2806 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2807 | if (off == 0) | |
2808 | return SIZE_MAX; | |
2809 | #endif | |
82fbcfa9 | 2810 | |
73572984 JA |
2811 | if (sq_offset) |
2812 | *sq_offset = off; | |
82fbcfa9 | 2813 | |
73572984 JA |
2814 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2815 | if (sq_array_size == SIZE_MAX) | |
2816 | return SIZE_MAX; | |
6b06314c | 2817 | |
73572984 JA |
2818 | if (check_add_overflow(off, sq_array_size, &off)) |
2819 | return SIZE_MAX; | |
8bad28d8 | 2820 | |
73572984 | 2821 | return off; |
8bad28d8 HX |
2822 | } |
2823 | ||
73572984 JA |
2824 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, |
2825 | unsigned int eventfd_async) | |
8bad28d8 | 2826 | { |
73572984 JA |
2827 | struct io_ev_fd *ev_fd; |
2828 | __s32 __user *fds = arg; | |
2829 | int fd; | |
f2303b1f | 2830 | |
73572984 JA |
2831 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, |
2832 | lockdep_is_held(&ctx->uring_lock)); | |
2833 | if (ev_fd) | |
2834 | return -EBUSY; | |
8bad28d8 | 2835 | |
73572984 JA |
2836 | if (copy_from_user(&fd, fds, sizeof(*fds))) |
2837 | return -EFAULT; | |
8dd03afe | 2838 | |
73572984 JA |
2839 | ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); |
2840 | if (!ev_fd) | |
2841 | return -ENOMEM; | |
05f3fb3c | 2842 | |
73572984 JA |
2843 | ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); |
2844 | if (IS_ERR(ev_fd->cq_ev_fd)) { | |
2845 | int ret = PTR_ERR(ev_fd->cq_ev_fd); | |
2846 | kfree(ev_fd); | |
2847 | return ret; | |
2848 | } | |
305bef98 PB |
2849 | |
2850 | spin_lock(&ctx->completion_lock); | |
2851 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
2852 | spin_unlock(&ctx->completion_lock); | |
2853 | ||
73572984 JA |
2854 | ev_fd->eventfd_async = eventfd_async; |
2855 | ctx->has_evfd = true; | |
2856 | rcu_assign_pointer(ctx->io_ev_fd, ev_fd); | |
21a091b9 DY |
2857 | atomic_set(&ev_fd->refs, 1); |
2858 | atomic_set(&ev_fd->ops, 0); | |
73572984 | 2859 | return 0; |
d7954b2b BM |
2860 | } |
2861 | ||
73572984 | 2862 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 2863 | { |
73572984 JA |
2864 | struct io_ev_fd *ev_fd; |
2865 | ||
2866 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, | |
2867 | lockdep_is_held(&ctx->uring_lock)); | |
2868 | if (ev_fd) { | |
2869 | ctx->has_evfd = false; | |
2870 | rcu_assign_pointer(ctx->io_ev_fd, NULL); | |
21a091b9 DY |
2871 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops)) |
2872 | call_rcu(&ev_fd->rcu, io_eventfd_ops); | |
73572984 JA |
2873 | return 0; |
2874 | } | |
2d091d62 | 2875 | |
73572984 | 2876 | return -ENXIO; |
44b31f2f PB |
2877 | } |
2878 | ||
73572984 | 2879 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2880 | { |
c8576f3e | 2881 | struct io_kiocb *req; |
37f0e767 | 2882 | int nr = 0; |
bf019da7 | 2883 | |
9a4fdbd8 | 2884 | mutex_lock(&ctx->uring_lock); |
34f0bc42 | 2885 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
9a4fdbd8 | 2886 | |
88ab95be | 2887 | while (!io_req_cache_empty(ctx)) { |
c8576f3e | 2888 | req = io_extract_req(ctx); |
c2b6c6bc | 2889 | kmem_cache_free(req_cachep, req); |
37f0e767 | 2890 | nr++; |
c2b6c6bc | 2891 | } |
37f0e767 PB |
2892 | if (nr) |
2893 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2894 | mutex_unlock(&ctx->uring_lock); |
2895 | } | |
2896 | ||
9eae8655 PB |
2897 | static void io_rsrc_node_cache_free(struct io_cache_entry *entry) |
2898 | { | |
2899 | kfree(container_of(entry, struct io_rsrc_node, cache)); | |
2900 | } | |
2901 | ||
c072481d | 2902 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2903 | { |
37d1e2e3 | 2904 | io_sq_thread_finish(ctx); |
43597aac | 2905 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
0b222eeb PB |
2906 | if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list))) |
2907 | return; | |
43597aac | 2908 | |
8bad28d8 | 2909 | mutex_lock(&ctx->uring_lock); |
43597aac | 2910 | if (ctx->buf_data) |
bd54b6fe | 2911 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2912 | if (ctx->file_data) |
08480400 | 2913 | __io_sqe_files_unregister(ctx); |
a85381d8 | 2914 | io_cqring_overflow_kill(ctx); |
9b402849 | 2915 | io_eventfd_unregister(ctx); |
9b797a37 | 2916 | io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); |
43e0bbbd | 2917 | io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); |
5a2e745d | 2918 | io_destroy_buffers(ctx); |
b4a72c05 | 2919 | mutex_unlock(&ctx->uring_lock); |
07db298a PB |
2920 | if (ctx->sq_creds) |
2921 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2922 | if (ctx->submitter_task) |
2923 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2924 | |
a7f0ed5a PB |
2925 | /* there are no registered resources left, nobody uses it */ |
2926 | if (ctx->rsrc_node) | |
9eae8655 | 2927 | io_rsrc_node_destroy(ctx, ctx->rsrc_node); |
a7f0ed5a PB |
2928 | |
2929 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
def596e9 | 2930 | |
2b188cc1 | 2931 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
2932 | if (ctx->ring_sock) { |
2933 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 2934 | sock_release(ctx->ring_sock); |
355e8d26 | 2935 | } |
2b188cc1 | 2936 | #endif |
ef9dd637 | 2937 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
2b188cc1 | 2938 | |
9eae8655 | 2939 | io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free); |
42b6419d PB |
2940 | if (ctx->mm_account) { |
2941 | mmdrop(ctx->mm_account); | |
2942 | ctx->mm_account = NULL; | |
2943 | } | |
9c189eee | 2944 | io_rings_free(ctx); |
2b188cc1 JA |
2945 | |
2946 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2947 | free_uid(ctx->user); |
4010fec4 | 2948 | io_req_caches_free(ctx); |
e941894e JA |
2949 | if (ctx->hash_map) |
2950 | io_wq_put_hash(ctx->hash_map); | |
e6f89be6 | 2951 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2952 | kfree(ctx->cancel_table_locked.hbs); |
6224843d | 2953 | kfree(ctx->dummy_ubuf); |
9cfc7e94 JA |
2954 | kfree(ctx->io_bl); |
2955 | xa_destroy(&ctx->io_bl_xa); | |
2b188cc1 JA |
2956 | kfree(ctx); |
2957 | } | |
2958 | ||
bca39f39 PB |
2959 | static __cold void io_activate_pollwq_cb(struct callback_head *cb) |
2960 | { | |
2961 | struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx, | |
2962 | poll_wq_task_work); | |
2963 | ||
2964 | mutex_lock(&ctx->uring_lock); | |
2965 | ctx->poll_activated = true; | |
2966 | mutex_unlock(&ctx->uring_lock); | |
2967 | ||
2968 | /* | |
2969 | * Wake ups for some events between start of polling and activation | |
2970 | * might've been lost due to loose synchronisation. | |
2971 | */ | |
2972 | wake_up_all(&ctx->poll_wq); | |
2973 | percpu_ref_put(&ctx->refs); | |
2974 | } | |
2975 | ||
2976 | static __cold void io_activate_pollwq(struct io_ring_ctx *ctx) | |
2977 | { | |
2978 | spin_lock(&ctx->completion_lock); | |
2979 | /* already activated or in progress */ | |
2980 | if (ctx->poll_activated || ctx->poll_wq_task_work.func) | |
2981 | goto out; | |
2982 | if (WARN_ON_ONCE(!ctx->task_complete)) | |
2983 | goto out; | |
2984 | if (!ctx->submitter_task) | |
2985 | goto out; | |
2986 | /* | |
2987 | * with ->submitter_task only the submitter task completes requests, we | |
2988 | * only need to sync with it, which is done by injecting a tw | |
2989 | */ | |
2990 | init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb); | |
2991 | percpu_ref_get(&ctx->refs); | |
2992 | if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL)) | |
2993 | percpu_ref_put(&ctx->refs); | |
2994 | out: | |
2995 | spin_unlock(&ctx->completion_lock); | |
2996 | } | |
2997 | ||
2b188cc1 JA |
2998 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) |
2999 | { | |
3000 | struct io_ring_ctx *ctx = file->private_data; | |
3001 | __poll_t mask = 0; | |
3002 | ||
bca39f39 PB |
3003 | if (unlikely(!ctx->poll_activated)) |
3004 | io_activate_pollwq(ctx); | |
3005 | ||
7b235dd8 | 3006 | poll_wait(file, &ctx->poll_wq, wait); |
4f7067c3 SB |
3007 | /* |
3008 | * synchronizes with barrier from wq_has_sleeper call in | |
3009 | * io_commit_cqring | |
3010 | */ | |
2b188cc1 | 3011 | smp_rmb(); |
90554200 | 3012 | if (!io_sqring_full(ctx)) |
2b188cc1 | 3013 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
3014 | |
3015 | /* | |
3016 | * Don't flush cqring overflow list here, just do a simple check. | |
3017 | * Otherwise there could possible be ABBA deadlock: | |
3018 | * CPU0 CPU1 | |
3019 | * ---- ---- | |
3020 | * lock(&ctx->uring_lock); | |
3021 | * lock(&ep->mtx); | |
3022 | * lock(&ctx->uring_lock); | |
3023 | * lock(&ep->mtx); | |
3024 | * | |
3025 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
10d8bc35 | 3026 | * pushes them to do the flush. |
ed670c3f | 3027 | */ |
b4c98d59 | 3028 | |
c10bb646 | 3029 | if (__io_cqring_events_user(ctx) || io_has_work(ctx)) |
2b188cc1 JA |
3030 | mask |= EPOLLIN | EPOLLRDNORM; |
3031 | ||
3032 | return mask; | |
3033 | } | |
3034 | ||
0bead8cd | 3035 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 3036 | { |
4379bf8b | 3037 | const struct cred *creds; |
071698e1 | 3038 | |
61cf9370 | 3039 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
3040 | if (creds) { |
3041 | put_cred(creds); | |
0bead8cd | 3042 | return 0; |
1e6fa521 | 3043 | } |
0bead8cd YD |
3044 | |
3045 | return -EINVAL; | |
3046 | } | |
3047 | ||
d56d938b PB |
3048 | struct io_tctx_exit { |
3049 | struct callback_head task_work; | |
3050 | struct completion completion; | |
baf186c4 | 3051 | struct io_ring_ctx *ctx; |
d56d938b PB |
3052 | }; |
3053 | ||
c072481d | 3054 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
3055 | { |
3056 | struct io_uring_task *tctx = current->io_uring; | |
3057 | struct io_tctx_exit *work; | |
3058 | ||
3059 | work = container_of(cb, struct io_tctx_exit, task_work); | |
3060 | /* | |
8d664282 | 3061 | * When @in_cancel, we're in cancellation and it's racy to remove the |
d56d938b | 3062 | * node. It'll be removed by the end of cancellation, just ignore it. |
998b30c3 HM |
3063 | * tctx can be NULL if the queueing of this task_work raced with |
3064 | * work cancelation off the exec path. | |
d56d938b | 3065 | */ |
8d664282 | 3066 | if (tctx && !atomic_read(&tctx->in_cancel)) |
eef51daa | 3067 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
3068 | complete(&work->completion); |
3069 | } | |
3070 | ||
c072481d | 3071 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
3072 | { |
3073 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
3074 | ||
3075 | return req->ctx == data; | |
3076 | } | |
3077 | ||
c072481d | 3078 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 3079 | { |
d56d938b | 3080 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 3081 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 3082 | unsigned long interval = HZ / 20; |
d56d938b PB |
3083 | struct io_tctx_exit exit; |
3084 | struct io_tctx_node *node; | |
3085 | int ret; | |
85faa7b8 | 3086 | |
56952e91 JA |
3087 | /* |
3088 | * If we're doing polled IO and end up having requests being | |
3089 | * submitted async (out-of-line), then completions can come in while | |
3090 | * we're waiting for refs to drop. We need to reap these manually, | |
3091 | * as nobody else will be looking for them. | |
3092 | */ | |
b2edc0a7 | 3093 | do { |
a85381d8 PB |
3094 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
3095 | mutex_lock(&ctx->uring_lock); | |
3096 | io_cqring_overflow_kill(ctx); | |
3097 | mutex_unlock(&ctx->uring_lock); | |
3098 | } | |
3099 | ||
c0e0d6ba DY |
3100 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) |
3101 | io_move_task_work_from_local(ctx); | |
3102 | ||
affa87db PB |
3103 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
3104 | cond_resched(); | |
3105 | ||
28090c13 PB |
3106 | if (ctx->sq_data) { |
3107 | struct io_sq_data *sqd = ctx->sq_data; | |
3108 | struct task_struct *tsk; | |
3109 | ||
3110 | io_sq_thread_park(sqd); | |
3111 | tsk = sqd->thread; | |
3112 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
3113 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
3114 | io_cancel_ctx_cb, ctx, true); | |
3115 | io_sq_thread_unpark(sqd); | |
3116 | } | |
b5bb3a24 | 3117 | |
37f0e767 PB |
3118 | io_req_caches_free(ctx); |
3119 | ||
58d3be2c PB |
3120 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
3121 | /* there is little hope left, don't run it too often */ | |
3122 | interval = HZ * 60; | |
3123 | } | |
4826c594 JA |
3124 | /* |
3125 | * This is really an uninterruptible wait, as it has to be | |
3126 | * complete. But it's also run from a kworker, which doesn't | |
3127 | * take signals, so it's fine to make it interruptible. This | |
3128 | * avoids scenarios where we knowingly can wait much longer | |
3129 | * on completions, for example if someone does a SIGSTOP on | |
3130 | * a task that needs to finish task_work to make this loop | |
3131 | * complete. That's a synthetic situation that should not | |
3132 | * cause a stuck task backtrace, and hence a potential panic | |
3133 | * on stuck tasks if that is enabled. | |
3134 | */ | |
3135 | } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 3136 | |
7f00651a PB |
3137 | init_completion(&exit.completion); |
3138 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
3139 | exit.ctx = ctx; | |
89b5066e PB |
3140 | /* |
3141 | * Some may use context even when all refs and requests have been put, | |
3142 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 3143 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
3144 | * this lock/unlock section also waits them to finish. |
3145 | */ | |
d56d938b PB |
3146 | mutex_lock(&ctx->uring_lock); |
3147 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
3148 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
3149 | ||
d56d938b PB |
3150 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
3151 | ctx_node); | |
7f00651a PB |
3152 | /* don't spin on a single task if cancellation failed */ |
3153 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
3154 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
3155 | if (WARN_ON_ONCE(ret)) | |
3156 | continue; | |
d56d938b PB |
3157 | |
3158 | mutex_unlock(&ctx->uring_lock); | |
4826c594 JA |
3159 | /* |
3160 | * See comment above for | |
3161 | * wait_for_completion_interruptible_timeout() on why this | |
3162 | * wait is marked as interruptible. | |
3163 | */ | |
3164 | wait_for_completion_interruptible(&exit.completion); | |
d56d938b PB |
3165 | mutex_lock(&ctx->uring_lock); |
3166 | } | |
3167 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
3168 | spin_lock(&ctx->completion_lock); |
3169 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 3170 | |
d73a572d PB |
3171 | /* pairs with RCU read section in io_req_local_work_add() */ |
3172 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3173 | synchronize_rcu(); | |
3174 | ||
85faa7b8 JA |
3175 | io_ring_ctx_free(ctx); |
3176 | } | |
3177 | ||
c072481d | 3178 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 3179 | { |
61cf9370 MWO |
3180 | unsigned long index; |
3181 | struct creds *creds; | |
3182 | ||
2b188cc1 JA |
3183 | mutex_lock(&ctx->uring_lock); |
3184 | percpu_ref_kill(&ctx->refs); | |
61cf9370 MWO |
3185 | xa_for_each(&ctx->personalities, index, creds) |
3186 | io_unregister_personality(ctx, index); | |
9ca9fb24 PB |
3187 | if (ctx->rings) |
3188 | io_poll_remove_all(ctx, NULL, true); | |
2b188cc1 JA |
3189 | mutex_unlock(&ctx->uring_lock); |
3190 | ||
02bac94b PB |
3191 | /* |
3192 | * If we failed setting up the ctx, we might not have any rings | |
3193 | * and therefore did not submit any requests | |
3194 | */ | |
3195 | if (ctx->rings) | |
60053be8 | 3196 | io_kill_timeouts(ctx, NULL, true); |
309fc03a | 3197 | |
85faa7b8 | 3198 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
3199 | /* |
3200 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
3201 | * if we're exiting a ton of rings at the same time. It just adds | |
3202 | * noise and overhead, there's no discernable change in runtime | |
3203 | * over using system_wq. | |
3204 | */ | |
3205 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
3206 | } |
3207 | ||
3208 | static int io_uring_release(struct inode *inode, struct file *file) | |
3209 | { | |
3210 | struct io_ring_ctx *ctx = file->private_data; | |
3211 | ||
3212 | file->private_data = NULL; | |
3213 | io_ring_ctx_wait_and_kill(ctx); | |
3214 | return 0; | |
3215 | } | |
3216 | ||
f6edbabb PB |
3217 | struct io_task_cancel { |
3218 | struct task_struct *task; | |
3dd0c97a | 3219 | bool all; |
f6edbabb | 3220 | }; |
f254ac04 | 3221 | |
f6edbabb | 3222 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 3223 | { |
9a472ef7 | 3224 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 3225 | struct io_task_cancel *cancel = data; |
9a472ef7 | 3226 | |
6af3f48b | 3227 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
3228 | } |
3229 | ||
c072481d PB |
3230 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
3231 | struct task_struct *task, | |
3232 | bool cancel_all) | |
b7ddce3c | 3233 | { |
e1915f76 | 3234 | struct io_defer_entry *de; |
b7ddce3c PB |
3235 | LIST_HEAD(list); |
3236 | ||
79ebeaee | 3237 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 3238 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 3239 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
3240 | list_cut_position(&list, &ctx->defer_list, &de->list); |
3241 | break; | |
3242 | } | |
3243 | } | |
79ebeaee | 3244 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
3245 | if (list_empty(&list)) |
3246 | return false; | |
b7ddce3c PB |
3247 | |
3248 | while (!list_empty(&list)) { | |
3249 | de = list_first_entry(&list, struct io_defer_entry, list); | |
3250 | list_del_init(&de->list); | |
e276ae34 | 3251 | io_req_task_queue_fail(de->req, -ECANCELED); |
b7ddce3c PB |
3252 | kfree(de); |
3253 | } | |
e1915f76 | 3254 | return true; |
b7ddce3c PB |
3255 | } |
3256 | ||
c072481d | 3257 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
3258 | { |
3259 | struct io_tctx_node *node; | |
3260 | enum io_wq_cancel cret; | |
3261 | bool ret = false; | |
3262 | ||
3263 | mutex_lock(&ctx->uring_lock); | |
3264 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3265 | struct io_uring_task *tctx = node->task->io_uring; | |
3266 | ||
3267 | /* | |
3268 | * io_wq will stay alive while we hold uring_lock, because it's | |
3269 | * killed after ctx nodes, which requires to take the lock. | |
3270 | */ | |
3271 | if (!tctx || !tctx->io_wq) | |
3272 | continue; | |
3273 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
3274 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3275 | } | |
3276 | mutex_unlock(&ctx->uring_lock); | |
3277 | ||
3278 | return ret; | |
3279 | } | |
3280 | ||
affa87db | 3281 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
3282 | struct task_struct *task, |
3283 | bool cancel_all) | |
9936c7c2 | 3284 | { |
3dd0c97a | 3285 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 3286 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
3287 | enum io_wq_cancel cret; |
3288 | bool ret = false; | |
9936c7c2 | 3289 | |
360cd42c PB |
3290 | /* set it so io_req_local_work_add() would wake us up */ |
3291 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { | |
3292 | atomic_set(&ctx->cq_wait_nr, 1); | |
3293 | smp_mb(); | |
3294 | } | |
3295 | ||
60053be8 PB |
3296 | /* failed during ring init, it couldn't have issued any requests */ |
3297 | if (!ctx->rings) | |
affa87db | 3298 | return false; |
60053be8 | 3299 | |
affa87db PB |
3300 | if (!task) { |
3301 | ret |= io_uring_try_cancel_iowq(ctx); | |
3302 | } else if (tctx && tctx->io_wq) { | |
3303 | /* | |
3304 | * Cancels requests of all rings, not only @ctx, but | |
3305 | * it's fine as the task is in exit/exec. | |
3306 | */ | |
3307 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
3308 | &cancel, true); | |
3309 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3310 | } | |
9936c7c2 | 3311 | |
affa87db PB |
3312 | /* SQPOLL thread does its own polling */ |
3313 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
3314 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
3315 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
3316 | io_iopoll_try_reap_events(ctx); | |
3317 | ret = true; | |
fcc926bb | 3318 | cond_resched(); |
9936c7c2 | 3319 | } |
9936c7c2 | 3320 | } |
affa87db | 3321 | |
140102ae PB |
3322 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3323 | io_allowed_defer_tw_run(ctx)) | |
c0e0d6ba | 3324 | ret |= io_run_local_work(ctx) > 0; |
affa87db PB |
3325 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
3326 | mutex_lock(&ctx->uring_lock); | |
3327 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
3328 | mutex_unlock(&ctx->uring_lock); | |
3329 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
3330 | if (task) | |
c0e0d6ba | 3331 | ret |= io_run_task_work() > 0; |
affa87db | 3332 | return ret; |
9936c7c2 PB |
3333 | } |
3334 | ||
3f48cf18 | 3335 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 3336 | { |
3f48cf18 | 3337 | if (tracked) |
9cae36a0 | 3338 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
3339 | return percpu_counter_sum(&tctx->inflight); |
3340 | } | |
3341 | ||
78cc687b PB |
3342 | /* |
3343 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 3344 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 3345 | */ |
17437f31 | 3346 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 3347 | { |
521d6a73 | 3348 | struct io_uring_task *tctx = current->io_uring; |
734551df | 3349 | struct io_ring_ctx *ctx; |
360cd42c PB |
3350 | struct io_tctx_node *node; |
3351 | unsigned long index; | |
0e9ddb39 PB |
3352 | s64 inflight; |
3353 | DEFINE_WAIT(wait); | |
fdaf083c | 3354 | |
78cc687b PB |
3355 | WARN_ON_ONCE(sqd && sqd->thread != current); |
3356 | ||
6d042ffb PO |
3357 | if (!current->io_uring) |
3358 | return; | |
17a91051 PB |
3359 | if (tctx->io_wq) |
3360 | io_wq_exit_start(tctx->io_wq); | |
3361 | ||
8d664282 | 3362 | atomic_inc(&tctx->in_cancel); |
0e9ddb39 | 3363 | do { |
affa87db PB |
3364 | bool loop = false; |
3365 | ||
e9dbe221 | 3366 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 3367 | /* read completions before cancelations */ |
78cc687b | 3368 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
3369 | if (!inflight) |
3370 | break; | |
fdaf083c | 3371 | |
78cc687b | 3372 | if (!sqd) { |
78cc687b PB |
3373 | xa_for_each(&tctx->xa, index, node) { |
3374 | /* sqpoll task will cancel all its requests */ | |
3375 | if (node->ctx->sq_data) | |
3376 | continue; | |
affa87db PB |
3377 | loop |= io_uring_try_cancel_requests(node->ctx, |
3378 | current, cancel_all); | |
78cc687b PB |
3379 | } |
3380 | } else { | |
3381 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
3382 | loop |= io_uring_try_cancel_requests(ctx, |
3383 | current, | |
3384 | cancel_all); | |
3385 | } | |
3386 | ||
3387 | if (loop) { | |
3388 | cond_resched(); | |
3389 | continue; | |
78cc687b | 3390 | } |
17a91051 | 3391 | |
78a78060 JA |
3392 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
3393 | io_run_task_work(); | |
e9dbe221 | 3394 | io_uring_drop_tctx_refs(current); |
360cd42c PB |
3395 | xa_for_each(&tctx->xa, index, node) { |
3396 | if (!llist_empty(&node->ctx->work_llist)) { | |
3397 | WARN_ON_ONCE(node->ctx->submitter_task && | |
3398 | node->ctx->submitter_task != current); | |
3399 | goto end_wait; | |
3400 | } | |
3401 | } | |
0f212204 | 3402 | /* |
a1bb3cd5 PB |
3403 | * If we've seen completions, retry without waiting. This |
3404 | * avoids a race where a completion comes in before we did | |
3405 | * prepare_to_wait(). | |
0f212204 | 3406 | */ |
3dd0c97a | 3407 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 3408 | schedule(); |
360cd42c | 3409 | end_wait: |
f57555ed | 3410 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 3411 | } while (1); |
de7f1d9e | 3412 | |
8452d4a6 | 3413 | io_uring_clean_tctx(tctx); |
3dd0c97a | 3414 | if (cancel_all) { |
3cc7fdb9 PB |
3415 | /* |
3416 | * We shouldn't run task_works after cancel, so just leave | |
8d664282 | 3417 | * ->in_cancel set for normal exit. |
3cc7fdb9 | 3418 | */ |
8d664282 | 3419 | atomic_dec(&tctx->in_cancel); |
3f48cf18 PB |
3420 | /* for exec all current's requests should be gone, kill tctx */ |
3421 | __io_uring_free(current); | |
3422 | } | |
44e728b8 PB |
3423 | } |
3424 | ||
f552a27a | 3425 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 3426 | { |
f552a27a | 3427 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
3428 | } |
3429 | ||
6c5c240e RP |
3430 | static void *io_uring_validate_mmap_request(struct file *file, |
3431 | loff_t pgoff, size_t sz) | |
2b188cc1 | 3432 | { |
2b188cc1 | 3433 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 3434 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
3435 | struct page *page; |
3436 | void *ptr; | |
3437 | ||
03d89a2d JA |
3438 | /* Don't allow mmap if the ring was setup without it */ |
3439 | if (ctx->flags & IORING_SETUP_NO_MMAP) | |
3440 | return ERR_PTR(-EINVAL); | |
3441 | ||
c56e022c | 3442 | switch (offset & IORING_OFF_MMAP_MASK) { |
2b188cc1 | 3443 | case IORING_OFF_SQ_RING: |
75b28aff HV |
3444 | case IORING_OFF_CQ_RING: |
3445 | ptr = ctx->rings; | |
2b188cc1 JA |
3446 | break; |
3447 | case IORING_OFF_SQES: | |
3448 | ptr = ctx->sq_sqes; | |
3449 | break; | |
c56e022c JA |
3450 | case IORING_OFF_PBUF_RING: { |
3451 | unsigned int bgid; | |
3452 | ||
3453 | bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT; | |
3454 | mutex_lock(&ctx->uring_lock); | |
3455 | ptr = io_pbuf_get_address(ctx, bgid); | |
3456 | mutex_unlock(&ctx->uring_lock); | |
3457 | if (!ptr) | |
3458 | return ERR_PTR(-EINVAL); | |
3459 | break; | |
3460 | } | |
2b188cc1 | 3461 | default: |
6c5c240e | 3462 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
3463 | } |
3464 | ||
3465 | page = virt_to_head_page(ptr); | |
a50b854e | 3466 | if (sz > page_size(page)) |
6c5c240e RP |
3467 | return ERR_PTR(-EINVAL); |
3468 | ||
3469 | return ptr; | |
3470 | } | |
3471 | ||
3472 | #ifdef CONFIG_MMU | |
3473 | ||
c072481d | 3474 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e RP |
3475 | { |
3476 | size_t sz = vma->vm_end - vma->vm_start; | |
3477 | unsigned long pfn; | |
3478 | void *ptr; | |
3479 | ||
3480 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
3481 | if (IS_ERR(ptr)) | |
3482 | return PTR_ERR(ptr); | |
2b188cc1 JA |
3483 | |
3484 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3485 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3486 | } | |
3487 | ||
d808459b HD |
3488 | static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp, |
3489 | unsigned long addr, unsigned long len, | |
3490 | unsigned long pgoff, unsigned long flags) | |
3491 | { | |
3492 | const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags); | |
3493 | struct vm_unmapped_area_info info; | |
3494 | void *ptr; | |
3495 | ||
3496 | /* | |
3497 | * Do not allow to map to user-provided address to avoid breaking the | |
3498 | * aliasing rules. Userspace is not able to guess the offset address of | |
3499 | * kernel kmalloc()ed memory area. | |
3500 | */ | |
3501 | if (addr) | |
3502 | return -EINVAL; | |
3503 | ||
3504 | ptr = io_uring_validate_mmap_request(filp, pgoff, len); | |
3505 | if (IS_ERR(ptr)) | |
3506 | return -ENOMEM; | |
3507 | ||
3508 | info.flags = VM_UNMAPPED_AREA_TOPDOWN; | |
3509 | info.length = len; | |
3510 | info.low_limit = max(PAGE_SIZE, mmap_min_addr); | |
3511 | info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base); | |
3512 | #ifdef SHM_COLOUR | |
3513 | info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL); | |
3514 | #else | |
3515 | info.align_mask = PAGE_MASK & (SHMLBA - 1UL); | |
3516 | #endif | |
3517 | info.align_offset = (unsigned long) ptr; | |
3518 | ||
3519 | /* | |
3520 | * A failed mmap() very likely causes application failure, | |
3521 | * so fall back to the bottom-up function here. This scenario | |
3522 | * can happen with large stack limits and large mmap() | |
3523 | * allocations. | |
3524 | */ | |
3525 | addr = vm_unmapped_area(&info); | |
3526 | if (offset_in_page(addr)) { | |
3527 | info.flags = 0; | |
3528 | info.low_limit = TASK_UNMAPPED_BASE; | |
3529 | info.high_limit = mmap_end; | |
3530 | addr = vm_unmapped_area(&info); | |
3531 | } | |
3532 | ||
3533 | return addr; | |
3534 | } | |
3535 | ||
6c5c240e RP |
3536 | #else /* !CONFIG_MMU */ |
3537 | ||
3538 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3539 | { | |
fc4f4be9 | 3540 | return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL; |
6c5c240e RP |
3541 | } |
3542 | ||
3543 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
3544 | { | |
3545 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
3546 | } | |
3547 | ||
3548 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
3549 | unsigned long addr, unsigned long len, | |
3550 | unsigned long pgoff, unsigned long flags) | |
3551 | { | |
3552 | void *ptr; | |
3553 | ||
3554 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
3555 | if (IS_ERR(ptr)) | |
3556 | return PTR_ERR(ptr); | |
3557 | ||
3558 | return (unsigned long) ptr; | |
3559 | } | |
3560 | ||
3561 | #endif /* !CONFIG_MMU */ | |
3562 | ||
f81440d3 PB |
3563 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
3564 | { | |
3565 | if (flags & IORING_ENTER_EXT_ARG) { | |
3566 | struct io_uring_getevents_arg arg; | |
3567 | ||
3568 | if (argsz != sizeof(arg)) | |
3569 | return -EINVAL; | |
3570 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3571 | return -EFAULT; | |
3572 | } | |
3573 | return 0; | |
3574 | } | |
3575 | ||
c73ebb68 HX |
3576 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
3577 | struct __kernel_timespec __user **ts, | |
3578 | const sigset_t __user **sig) | |
3579 | { | |
3580 | struct io_uring_getevents_arg arg; | |
3581 | ||
3582 | /* | |
3583 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
3584 | * is just a pointer to the sigset_t. | |
3585 | */ | |
3586 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
3587 | *sig = (const sigset_t __user *) argp; | |
3588 | *ts = NULL; | |
3589 | return 0; | |
3590 | } | |
3591 | ||
3592 | /* | |
3593 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
3594 | * timespec and sigset_t pointers if good. | |
3595 | */ | |
3596 | if (*argsz != sizeof(arg)) | |
3597 | return -EINVAL; | |
3598 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3599 | return -EFAULT; | |
d2347b96 DY |
3600 | if (arg.pad) |
3601 | return -EINVAL; | |
c73ebb68 HX |
3602 | *sig = u64_to_user_ptr(arg.sigmask); |
3603 | *argsz = arg.sigmask_sz; | |
3604 | *ts = u64_to_user_ptr(arg.ts); | |
3605 | return 0; | |
3606 | } | |
3607 | ||
2b188cc1 | 3608 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
3609 | u32, min_complete, u32, flags, const void __user *, argp, |
3610 | size_t, argsz) | |
2b188cc1 JA |
3611 | { |
3612 | struct io_ring_ctx *ctx; | |
2b188cc1 | 3613 | struct fd f; |
33f993da | 3614 | long ret; |
2b188cc1 | 3615 | |
33f993da | 3616 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
3617 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
3618 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
3619 | return -EINVAL; |
3620 | ||
e7a6c00d JA |
3621 | /* |
3622 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3623 | * need only dereference our task private array to find it. | |
3624 | */ | |
3625 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3626 | struct io_uring_task *tctx = current->io_uring; | |
3627 | ||
3273c440 | 3628 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
e7a6c00d JA |
3629 | return -EINVAL; |
3630 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
3631 | f.file = tctx->registered_rings[fd]; | |
4329490a | 3632 | f.flags = 0; |
3273c440 PB |
3633 | if (unlikely(!f.file)) |
3634 | return -EBADF; | |
e7a6c00d JA |
3635 | } else { |
3636 | f = fdget(fd); | |
3273c440 PB |
3637 | if (unlikely(!f.file)) |
3638 | return -EBADF; | |
3639 | ret = -EOPNOTSUPP; | |
3640 | if (unlikely(!io_is_uring_fops(f.file))) | |
fbb8bb02 | 3641 | goto out; |
e7a6c00d | 3642 | } |
2b188cc1 | 3643 | |
2b188cc1 | 3644 | ctx = f.file->private_data; |
7e84e1c7 | 3645 | ret = -EBADFD; |
33f993da | 3646 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3647 | goto out; |
3648 | ||
6c271ce2 JA |
3649 | /* |
3650 | * For SQ polling, the thread will do all submissions and completions. | |
3651 | * Just return the requested submit count, and wake the thread if | |
3652 | * we were asked to. | |
3653 | */ | |
b2a9eada | 3654 | ret = 0; |
6c271ce2 | 3655 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3656 | io_cqring_overflow_flush(ctx); |
89448c47 | 3657 | |
21f96522 JA |
3658 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3659 | ret = -EOWNERDEAD; | |
04147488 | 3660 | goto out; |
21f96522 | 3661 | } |
6c271ce2 | 3662 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3663 | wake_up(&ctx->sq_data->wait); |
88b80534 QF |
3664 | if (flags & IORING_ENTER_SQ_WAIT) |
3665 | io_sqpoll_wait_sq(ctx); | |
3666 | ||
3e813c90 | 3667 | ret = to_submit; |
b2a9eada | 3668 | } else if (to_submit) { |
eef51daa | 3669 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3670 | if (unlikely(ret)) |
3671 | goto out; | |
7c504e65 | 3672 | |
2b188cc1 | 3673 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3674 | ret = io_submit_sqes(ctx, to_submit); |
3675 | if (ret != to_submit) { | |
d487b43c | 3676 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3677 | goto out; |
d487b43c | 3678 | } |
44f87745 PB |
3679 | if (flags & IORING_ENTER_GETEVENTS) { |
3680 | if (ctx->syscall_iopoll) | |
3681 | goto iopoll_locked; | |
3682 | /* | |
3683 | * Ignore errors, we'll soon call io_cqring_wait() and | |
3684 | * it should handle ownership problems if any. | |
3685 | */ | |
3686 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3687 | (void)io_run_local_work_locked(ctx); | |
3688 | } | |
d487b43c | 3689 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 | 3690 | } |
c0e0d6ba | 3691 | |
2b188cc1 | 3692 | if (flags & IORING_ENTER_GETEVENTS) { |
3e813c90 | 3693 | int ret2; |
c0e0d6ba | 3694 | |
773697b6 | 3695 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3696 | /* |
3697 | * We disallow the app entering submit/complete with | |
3698 | * polling, but we still need to lock the ring to | |
3699 | * prevent racing with polled issue that got punted to | |
3700 | * a workqueue. | |
3701 | */ | |
3702 | mutex_lock(&ctx->uring_lock); | |
3703 | iopoll_locked: | |
3e813c90 DY |
3704 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3705 | if (likely(!ret2)) { | |
3706 | min_complete = min(min_complete, | |
3707 | ctx->cq_entries); | |
3708 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3709 | } |
3710 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3711 | } else { |
f81440d3 PB |
3712 | const sigset_t __user *sig; |
3713 | struct __kernel_timespec __user *ts; | |
3714 | ||
3e813c90 DY |
3715 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3716 | if (likely(!ret2)) { | |
3717 | min_complete = min(min_complete, | |
3718 | ctx->cq_entries); | |
3719 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3720 | argsz, ts); | |
3721 | } | |
def596e9 | 3722 | } |
c73ebb68 | 3723 | |
155bc950 | 3724 | if (!ret) { |
3e813c90 | 3725 | ret = ret2; |
2b188cc1 | 3726 | |
155bc950 DY |
3727 | /* |
3728 | * EBADR indicates that one or more CQE were dropped. | |
3729 | * Once the user has been informed we can clear the bit | |
3730 | * as they are obviously ok with those drops. | |
3731 | */ | |
3732 | if (unlikely(ret2 == -EBADR)) | |
3733 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3734 | &ctx->check_cq); | |
def596e9 | 3735 | } |
2b188cc1 | 3736 | } |
7c504e65 | 3737 | out: |
4329490a | 3738 | fdput(f); |
3e813c90 | 3739 | return ret; |
2b188cc1 JA |
3740 | } |
3741 | ||
3742 | static const struct file_operations io_uring_fops = { | |
3743 | .release = io_uring_release, | |
3744 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3745 | #ifndef CONFIG_MMU |
3746 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3747 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
d808459b HD |
3748 | #else |
3749 | .get_unmapped_area = io_uring_mmu_get_unmapped_area, | |
6c5c240e | 3750 | #endif |
2b188cc1 | 3751 | .poll = io_uring_poll, |
bebdb65e | 3752 | #ifdef CONFIG_PROC_FS |
87ce955b | 3753 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3754 | #endif |
2b188cc1 JA |
3755 | }; |
3756 | ||
92ac8bea JA |
3757 | bool io_is_uring_fops(struct file *file) |
3758 | { | |
3759 | return file->f_op == &io_uring_fops; | |
3760 | } | |
3761 | ||
c072481d PB |
3762 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3763 | struct io_uring_params *p) | |
2b188cc1 | 3764 | { |
75b28aff HV |
3765 | struct io_rings *rings; |
3766 | size_t size, sq_array_offset; | |
e27cef86 | 3767 | void *ptr; |
2b188cc1 | 3768 | |
bd740481 JA |
3769 | /* make sure these are sane, as we already accounted them */ |
3770 | ctx->sq_entries = p->sq_entries; | |
3771 | ctx->cq_entries = p->cq_entries; | |
3772 | ||
baf9cb64 | 3773 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3774 | if (size == SIZE_MAX) |
3775 | return -EOVERFLOW; | |
3776 | ||
03d89a2d JA |
3777 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3778 | rings = io_mem_alloc(size); | |
3779 | else | |
3780 | rings = io_rings_map(ctx, p->cq_off.user_addr, size); | |
3781 | ||
e27cef86 JA |
3782 | if (IS_ERR(rings)) |
3783 | return PTR_ERR(rings); | |
2b188cc1 | 3784 | |
75b28aff HV |
3785 | ctx->rings = rings; |
3786 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
3787 | rings->sq_ring_mask = p->sq_entries - 1; | |
3788 | rings->cq_ring_mask = p->cq_entries - 1; | |
3789 | rings->sq_ring_entries = p->sq_entries; | |
3790 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3791 | |
ebdeb7c0 JA |
3792 | if (p->flags & IORING_SETUP_SQE128) |
3793 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3794 | else | |
3795 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 | 3796 | if (size == SIZE_MAX) { |
9c189eee | 3797 | io_rings_free(ctx); |
2b188cc1 | 3798 | return -EOVERFLOW; |
eb065d30 | 3799 | } |
2b188cc1 | 3800 | |
03d89a2d JA |
3801 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3802 | ptr = io_mem_alloc(size); | |
3803 | else | |
3804 | ptr = io_sqes_map(ctx, p->sq_off.user_addr, size); | |
3805 | ||
e27cef86 | 3806 | if (IS_ERR(ptr)) { |
9c189eee | 3807 | io_rings_free(ctx); |
e27cef86 | 3808 | return PTR_ERR(ptr); |
eb065d30 | 3809 | } |
2b188cc1 | 3810 | |
e27cef86 | 3811 | ctx->sq_sqes = ptr; |
2b188cc1 JA |
3812 | return 0; |
3813 | } | |
3814 | ||
6e76ac59 | 3815 | static int io_uring_install_fd(struct file *file) |
9faadcc8 | 3816 | { |
6e76ac59 | 3817 | int fd; |
9faadcc8 PB |
3818 | |
3819 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3820 | if (fd < 0) | |
3821 | return fd; | |
9faadcc8 PB |
3822 | fd_install(fd, file); |
3823 | return fd; | |
3824 | } | |
3825 | ||
2b188cc1 JA |
3826 | /* |
3827 | * Allocate an anonymous fd, this is what constitutes the application | |
3828 | * visible backing of an io_uring instance. The application mmaps this | |
3829 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3830 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3831 | */ | |
9faadcc8 | 3832 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
3833 | { |
3834 | struct file *file; | |
9faadcc8 | 3835 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
3836 | int ret; |
3837 | ||
2b188cc1 JA |
3838 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
3839 | &ctx->ring_sock); | |
3840 | if (ret) | |
9faadcc8 | 3841 | return ERR_PTR(ret); |
2b188cc1 JA |
3842 | #endif |
3843 | ||
91a9ab7c PM |
3844 | file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, |
3845 | O_RDWR | O_CLOEXEC, NULL); | |
2b188cc1 | 3846 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
3847 | if (IS_ERR(file)) { |
3848 | sock_release(ctx->ring_sock); | |
3849 | ctx->ring_sock = NULL; | |
3850 | } else { | |
3851 | ctx->ring_sock->file = file; | |
0f212204 | 3852 | } |
2b188cc1 | 3853 | #endif |
9faadcc8 | 3854 | return file; |
2b188cc1 JA |
3855 | } |
3856 | ||
c072481d PB |
3857 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3858 | struct io_uring_params __user *params) | |
2b188cc1 | 3859 | { |
2b188cc1 | 3860 | struct io_ring_ctx *ctx; |
6e76ac59 | 3861 | struct io_uring_task *tctx; |
9faadcc8 | 3862 | struct file *file; |
2b188cc1 JA |
3863 | int ret; |
3864 | ||
8110c1a6 | 3865 | if (!entries) |
2b188cc1 | 3866 | return -EINVAL; |
8110c1a6 JA |
3867 | if (entries > IORING_MAX_ENTRIES) { |
3868 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3869 | return -EINVAL; | |
3870 | entries = IORING_MAX_ENTRIES; | |
3871 | } | |
2b188cc1 | 3872 | |
6e76ac59 JT |
3873 | if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY) |
3874 | && !(p->flags & IORING_SETUP_NO_MMAP)) | |
3875 | return -EINVAL; | |
3876 | ||
2b188cc1 JA |
3877 | /* |
3878 | * Use twice as many entries for the CQ ring. It's possible for the | |
3879 | * application to drive a higher depth than the size of the SQ ring, | |
3880 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3881 | * some flexibility in overcommitting a bit. If the application has |
3882 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3883 | * of CQ ring entries manually. | |
2b188cc1 JA |
3884 | */ |
3885 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3886 | if (p->flags & IORING_SETUP_CQSIZE) { |
3887 | /* | |
3888 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3889 | * to a power-of-two, if it isn't already. We do NOT impose | |
3890 | * any cq vs sq ring sizing. | |
3891 | */ | |
eb2667b3 | 3892 | if (!p->cq_entries) |
33a107f0 | 3893 | return -EINVAL; |
8110c1a6 JA |
3894 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3895 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3896 | return -EINVAL; | |
3897 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3898 | } | |
eb2667b3 JQ |
3899 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3900 | if (p->cq_entries < p->sq_entries) | |
3901 | return -EINVAL; | |
33a107f0 JA |
3902 | } else { |
3903 | p->cq_entries = 2 * p->sq_entries; | |
3904 | } | |
2b188cc1 | 3905 | |
2b188cc1 | 3906 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3907 | if (!ctx) |
2b188cc1 | 3908 | return -ENOMEM; |
773697b6 | 3909 | |
e6aeb272 PB |
3910 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3911 | !(ctx->flags & IORING_SETUP_IOPOLL) && | |
3912 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3913 | ctx->task_complete = true; | |
3914 | ||
bca39f39 PB |
3915 | /* |
3916 | * lazy poll_wq activation relies on ->task_complete for synchronisation | |
3917 | * purposes, see io_activate_pollwq() | |
3918 | */ | |
3919 | if (!ctx->task_complete) | |
3920 | ctx->poll_activated = true; | |
3921 | ||
773697b6 PB |
3922 | /* |
3923 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3924 | * space applications don't need to do io completion events | |
3925 | * polling again, they can rely on io_sq_thread to do polling | |
3926 | * work, which can reduce cpu usage and uring_lock contention. | |
3927 | */ | |
3928 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3929 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3930 | ctx->syscall_iopoll = 1; | |
3931 | ||
2b188cc1 | 3932 | ctx->compat = in_compat_syscall(); |
62e398be JA |
3933 | if (!capable(CAP_IPC_LOCK)) |
3934 | ctx->user = get_uid(current_user()); | |
2aede0e4 | 3935 | |
9f010507 | 3936 | /* |
e1169f06 JA |
3937 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3938 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3939 | */ |
e1169f06 JA |
3940 | ret = -EINVAL; |
3941 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3942 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 | 3943 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
c0e0d6ba DY |
3944 | IORING_SETUP_TASKRUN_FLAG | |
3945 | IORING_SETUP_DEFER_TASKRUN)) | |
e1169f06 | 3946 | goto err; |
9f010507 | 3947 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3948 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3949 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3950 | } else { | |
c0e0d6ba DY |
3951 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && |
3952 | !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) | |
ef060ea9 | 3953 | goto err; |
9f010507 | 3954 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3955 | } |
9f010507 | 3956 | |
c0e0d6ba DY |
3957 | /* |
3958 | * For DEFER_TASKRUN we require the completion task to be the same as the | |
3959 | * submission task. This implies that there is only one submitter, so enforce | |
3960 | * that. | |
3961 | */ | |
3962 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && | |
3963 | !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { | |
3964 | goto err; | |
3965 | } | |
3966 | ||
2aede0e4 JA |
3967 | /* |
3968 | * This is just grabbed for accounting purposes. When a process exits, | |
3969 | * the mm is exited and dropped before the files, hence we need to hang | |
3970 | * on to this mm purely for the purposes of being able to unaccount | |
3971 | * memory (locked/pinned vm). It's not used for anything else. | |
3972 | */ | |
6b7898eb | 3973 | mmgrab(current->mm); |
2aede0e4 | 3974 | ctx->mm_account = current->mm; |
6b7898eb | 3975 | |
2b188cc1 JA |
3976 | ret = io_allocate_scq_urings(ctx, p); |
3977 | if (ret) | |
3978 | goto err; | |
3979 | ||
7e84e1c7 | 3980 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
3981 | if (ret) |
3982 | goto err; | |
2933ae6e PB |
3983 | |
3984 | ret = io_rsrc_init(ctx); | |
47b228ce PB |
3985 | if (ret) |
3986 | goto err; | |
2b188cc1 | 3987 | |
75b28aff HV |
3988 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3989 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3990 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3991 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3992 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3993 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
3994 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
9b1b58ca | 3995 | p->sq_off.resv1 = 0; |
03d89a2d JA |
3996 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3997 | p->sq_off.user_addr = 0; | |
2b188cc1 | 3998 | |
75b28aff HV |
3999 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
4000 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
4001 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
4002 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
4003 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
4004 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 4005 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
9b1b58ca | 4006 | p->cq_off.resv1 = 0; |
03d89a2d JA |
4007 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
4008 | p->cq_off.user_addr = 0; | |
ac90f249 | 4009 | |
7f13657d XW |
4010 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
4011 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 4012 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 4013 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 4014 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e | 4015 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
7d3fd88d | 4016 | IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING; |
7f13657d XW |
4017 | |
4018 | if (copy_to_user(params, p, sizeof(*p))) { | |
4019 | ret = -EFAULT; | |
4020 | goto err; | |
4021 | } | |
d1719f70 | 4022 | |
7cae596b DY |
4023 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER |
4024 | && !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
8579538c | 4025 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
7cae596b | 4026 | |
9faadcc8 PB |
4027 | file = io_uring_get_file(ctx); |
4028 | if (IS_ERR(file)) { | |
4029 | ret = PTR_ERR(file); | |
4030 | goto err; | |
4031 | } | |
4032 | ||
6e76ac59 JT |
4033 | ret = __io_uring_add_tctx_node(ctx); |
4034 | if (ret) | |
4035 | goto err_fput; | |
4036 | tctx = current->io_uring; | |
4037 | ||
044c1ab3 JA |
4038 | /* |
4039 | * Install ring fd as the very last thing, so we don't risk someone | |
4040 | * having closed it before we finish setup | |
4041 | */ | |
6e76ac59 JT |
4042 | if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY) |
4043 | ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX); | |
4044 | else | |
4045 | ret = io_uring_install_fd(file); | |
4046 | if (ret < 0) | |
4047 | goto err_fput; | |
044c1ab3 | 4048 | |
c826bd7a | 4049 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
4050 | return ret; |
4051 | err: | |
4052 | io_ring_ctx_wait_and_kill(ctx); | |
4053 | return ret; | |
6e76ac59 JT |
4054 | err_fput: |
4055 | fput(file); | |
4056 | return ret; | |
2b188cc1 JA |
4057 | } |
4058 | ||
4059 | /* | |
4060 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
4061 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
4062 | * params structure passed in. | |
4063 | */ | |
4064 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
4065 | { | |
4066 | struct io_uring_params p; | |
2b188cc1 JA |
4067 | int i; |
4068 | ||
4069 | if (copy_from_user(&p, params, sizeof(p))) | |
4070 | return -EFAULT; | |
4071 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
4072 | if (p.resv[i]) | |
4073 | return -EINVAL; | |
4074 | } | |
4075 | ||
6c271ce2 | 4076 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 4077 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 4078 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 4079 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 4080 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 | 4081 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
03d89a2d | 4082 | IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN | |
6e76ac59 | 4083 | IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY)) |
2b188cc1 JA |
4084 | return -EINVAL; |
4085 | ||
ef060ea9 | 4086 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
4087 | } |
4088 | ||
4089 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
4090 | struct io_uring_params __user *, params) | |
4091 | { | |
4092 | return io_uring_setup(entries, params); | |
4093 | } | |
4094 | ||
c072481d PB |
4095 | static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, |
4096 | unsigned nr_args) | |
66f4af93 JA |
4097 | { |
4098 | struct io_uring_probe *p; | |
4099 | size_t size; | |
4100 | int i, ret; | |
4101 | ||
4102 | size = struct_size(p, ops, nr_args); | |
4103 | if (size == SIZE_MAX) | |
4104 | return -EOVERFLOW; | |
4105 | p = kzalloc(size, GFP_KERNEL); | |
4106 | if (!p) | |
4107 | return -ENOMEM; | |
4108 | ||
4109 | ret = -EFAULT; | |
4110 | if (copy_from_user(p, arg, size)) | |
4111 | goto out; | |
4112 | ret = -EINVAL; | |
4113 | if (memchr_inv(p, 0, size)) | |
4114 | goto out; | |
4115 | ||
4116 | p->last_op = IORING_OP_LAST - 1; | |
4117 | if (nr_args > IORING_OP_LAST) | |
4118 | nr_args = IORING_OP_LAST; | |
4119 | ||
4120 | for (i = 0; i < nr_args; i++) { | |
4121 | p->ops[i].op = i; | |
a7dd2782 | 4122 | if (!io_issue_defs[i].not_supported) |
66f4af93 JA |
4123 | p->ops[i].flags = IO_URING_OP_SUPPORTED; |
4124 | } | |
4125 | p->ops_len = i; | |
4126 | ||
4127 | ret = 0; | |
4128 | if (copy_to_user(arg, p, size)) | |
4129 | ret = -EFAULT; | |
4130 | out: | |
4131 | kfree(p); | |
4132 | return ret; | |
4133 | } | |
4134 | ||
071698e1 JA |
4135 | static int io_register_personality(struct io_ring_ctx *ctx) |
4136 | { | |
4379bf8b | 4137 | const struct cred *creds; |
61cf9370 | 4138 | u32 id; |
1e6fa521 | 4139 | int ret; |
071698e1 | 4140 | |
4379bf8b | 4141 | creds = get_current_cred(); |
1e6fa521 | 4142 | |
61cf9370 MWO |
4143 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
4144 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
4145 | if (ret < 0) { |
4146 | put_cred(creds); | |
4147 | return ret; | |
4148 | } | |
4149 | return id; | |
071698e1 JA |
4150 | } |
4151 | ||
c072481d PB |
4152 | static __cold int io_register_restrictions(struct io_ring_ctx *ctx, |
4153 | void __user *arg, unsigned int nr_args) | |
21b55dbc SG |
4154 | { |
4155 | struct io_uring_restriction *res; | |
4156 | size_t size; | |
4157 | int i, ret; | |
4158 | ||
7e84e1c7 SG |
4159 | /* Restrictions allowed only if rings started disabled */ |
4160 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
4161 | return -EBADFD; | |
4162 | ||
21b55dbc | 4163 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 4164 | if (ctx->restrictions.registered) |
21b55dbc SG |
4165 | return -EBUSY; |
4166 | ||
4167 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
4168 | return -EINVAL; | |
4169 | ||
4170 | size = array_size(nr_args, sizeof(*res)); | |
4171 | if (size == SIZE_MAX) | |
4172 | return -EOVERFLOW; | |
4173 | ||
4174 | res = memdup_user(arg, size); | |
4175 | if (IS_ERR(res)) | |
4176 | return PTR_ERR(res); | |
4177 | ||
4178 | ret = 0; | |
4179 | ||
4180 | for (i = 0; i < nr_args; i++) { | |
4181 | switch (res[i].opcode) { | |
4182 | case IORING_RESTRICTION_REGISTER_OP: | |
4183 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
4184 | ret = -EINVAL; | |
4185 | goto out; | |
4186 | } | |
4187 | ||
4188 | __set_bit(res[i].register_op, | |
4189 | ctx->restrictions.register_op); | |
4190 | break; | |
4191 | case IORING_RESTRICTION_SQE_OP: | |
4192 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
4193 | ret = -EINVAL; | |
4194 | goto out; | |
4195 | } | |
4196 | ||
4197 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
4198 | break; | |
4199 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
4200 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
4201 | break; | |
4202 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
4203 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
4204 | break; | |
4205 | default: | |
4206 | ret = -EINVAL; | |
4207 | goto out; | |
4208 | } | |
4209 | } | |
4210 | ||
4211 | out: | |
4212 | /* Reset all restrictions if an error happened */ | |
4213 | if (ret != 0) | |
4214 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
4215 | else | |
7e84e1c7 | 4216 | ctx->restrictions.registered = true; |
21b55dbc SG |
4217 | |
4218 | kfree(res); | |
4219 | return ret; | |
4220 | } | |
4221 | ||
7e84e1c7 SG |
4222 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
4223 | { | |
4224 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
4225 | return -EBADFD; | |
4226 | ||
bca39f39 | 4227 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) { |
8579538c | 4228 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
bca39f39 PB |
4229 | /* |
4230 | * Lazy activation attempts would fail if it was polled before | |
4231 | * submitter_task is set. | |
4232 | */ | |
4233 | if (wq_has_sleeper(&ctx->poll_wq)) | |
4234 | io_activate_pollwq(ctx); | |
4235 | } | |
7cae596b | 4236 | |
7e84e1c7 SG |
4237 | if (ctx->restrictions.registered) |
4238 | ctx->restricted = 1; | |
4239 | ||
0298ef96 PB |
4240 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
4241 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
4242 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
4243 | return 0; |
4244 | } | |
4245 | ||
c072481d PB |
4246 | static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, |
4247 | void __user *arg, unsigned len) | |
fe76421d JA |
4248 | { |
4249 | struct io_uring_task *tctx = current->io_uring; | |
4250 | cpumask_var_t new_mask; | |
4251 | int ret; | |
4252 | ||
4253 | if (!tctx || !tctx->io_wq) | |
4254 | return -EINVAL; | |
4255 | ||
4256 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
4257 | return -ENOMEM; | |
4258 | ||
4259 | cpumask_clear(new_mask); | |
4260 | if (len > cpumask_size()) | |
4261 | len = cpumask_size(); | |
4262 | ||
0f5e4b83 ES |
4263 | if (in_compat_syscall()) { |
4264 | ret = compat_get_bitmap(cpumask_bits(new_mask), | |
4265 | (const compat_ulong_t __user *)arg, | |
4266 | len * 8 /* CHAR_BIT */); | |
4267 | } else { | |
4268 | ret = copy_from_user(new_mask, arg, len); | |
4269 | } | |
4270 | ||
4271 | if (ret) { | |
fe76421d JA |
4272 | free_cpumask_var(new_mask); |
4273 | return -EFAULT; | |
4274 | } | |
4275 | ||
4276 | ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); | |
4277 | free_cpumask_var(new_mask); | |
4278 | return ret; | |
4279 | } | |
4280 | ||
c072481d | 4281 | static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) |
fe76421d JA |
4282 | { |
4283 | struct io_uring_task *tctx = current->io_uring; | |
4284 | ||
4285 | if (!tctx || !tctx->io_wq) | |
4286 | return -EINVAL; | |
4287 | ||
4288 | return io_wq_cpu_affinity(tctx->io_wq, NULL); | |
4289 | } | |
4290 | ||
c072481d PB |
4291 | static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, |
4292 | void __user *arg) | |
b22fa62a | 4293 | __must_hold(&ctx->uring_lock) |
2e480058 | 4294 | { |
b22fa62a | 4295 | struct io_tctx_node *node; |
fa84693b JA |
4296 | struct io_uring_task *tctx = NULL; |
4297 | struct io_sq_data *sqd = NULL; | |
2e480058 JA |
4298 | __u32 new_count[2]; |
4299 | int i, ret; | |
4300 | ||
2e480058 JA |
4301 | if (copy_from_user(new_count, arg, sizeof(new_count))) |
4302 | return -EFAULT; | |
4303 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
4304 | if (new_count[i] > INT_MAX) | |
4305 | return -EINVAL; | |
4306 | ||
fa84693b JA |
4307 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
4308 | sqd = ctx->sq_data; | |
4309 | if (sqd) { | |
009ad9f0 JA |
4310 | /* |
4311 | * Observe the correct sqd->lock -> ctx->uring_lock | |
4312 | * ordering. Fine to drop uring_lock here, we hold | |
4313 | * a ref to the ctx. | |
4314 | */ | |
41d3a6bd | 4315 | refcount_inc(&sqd->refs); |
009ad9f0 | 4316 | mutex_unlock(&ctx->uring_lock); |
fa84693b | 4317 | mutex_lock(&sqd->lock); |
009ad9f0 | 4318 | mutex_lock(&ctx->uring_lock); |
41d3a6bd JA |
4319 | if (sqd->thread) |
4320 | tctx = sqd->thread->io_uring; | |
fa84693b JA |
4321 | } |
4322 | } else { | |
4323 | tctx = current->io_uring; | |
4324 | } | |
4325 | ||
e139a1ec | 4326 | BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); |
fa84693b | 4327 | |
bad119b9 PB |
4328 | for (i = 0; i < ARRAY_SIZE(new_count); i++) |
4329 | if (new_count[i]) | |
4330 | ctx->iowq_limits[i] = new_count[i]; | |
e139a1ec PB |
4331 | ctx->iowq_limits_set = true; |
4332 | ||
e139a1ec PB |
4333 | if (tctx && tctx->io_wq) { |
4334 | ret = io_wq_max_workers(tctx->io_wq, new_count); | |
4335 | if (ret) | |
4336 | goto err; | |
4337 | } else { | |
4338 | memset(new_count, 0, sizeof(new_count)); | |
4339 | } | |
fa84693b | 4340 | |
41d3a6bd | 4341 | if (sqd) { |
fa84693b | 4342 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
4343 | io_put_sq_data(sqd); |
4344 | } | |
2e480058 JA |
4345 | |
4346 | if (copy_to_user(arg, new_count, sizeof(new_count))) | |
4347 | return -EFAULT; | |
4348 | ||
b22fa62a PB |
4349 | /* that's it for SQPOLL, only the SQPOLL task creates requests */ |
4350 | if (sqd) | |
4351 | return 0; | |
4352 | ||
4353 | /* now propagate the restriction to all registered users */ | |
4354 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
4355 | struct io_uring_task *tctx = node->task->io_uring; | |
4356 | ||
4357 | if (WARN_ON_ONCE(!tctx->io_wq)) | |
4358 | continue; | |
4359 | ||
4360 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
4361 | new_count[i] = ctx->iowq_limits[i]; | |
4362 | /* ignore errors, it always returns zero anyway */ | |
4363 | (void)io_wq_max_workers(tctx->io_wq, new_count); | |
4364 | } | |
2e480058 | 4365 | return 0; |
fa84693b | 4366 | err: |
41d3a6bd | 4367 | if (sqd) { |
fa84693b | 4368 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
4369 | io_put_sq_data(sqd); |
4370 | } | |
fa84693b | 4371 | return ret; |
2e480058 JA |
4372 | } |
4373 | ||
edafccee JA |
4374 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
4375 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
4376 | __releases(ctx->uring_lock) |
4377 | __acquires(ctx->uring_lock) | |
edafccee JA |
4378 | { |
4379 | int ret; | |
4380 | ||
35fa71a0 | 4381 | /* |
fbb8bb02 PB |
4382 | * We don't quiesce the refs for register anymore and so it can't be |
4383 | * dying as we're holding a file ref here. | |
35fa71a0 | 4384 | */ |
fbb8bb02 | 4385 | if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs))) |
35fa71a0 JA |
4386 | return -ENXIO; |
4387 | ||
d7cce96c PB |
4388 | if (ctx->submitter_task && ctx->submitter_task != current) |
4389 | return -EEXIST; | |
4390 | ||
75c4021a | 4391 | if (ctx->restricted) { |
75c4021a PB |
4392 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); |
4393 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
4394 | return -EACCES; | |
4395 | } | |
4396 | ||
edafccee JA |
4397 | switch (opcode) { |
4398 | case IORING_REGISTER_BUFFERS: | |
0184f08e PB |
4399 | ret = -EFAULT; |
4400 | if (!arg) | |
4401 | break; | |
634d00df | 4402 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
4403 | break; |
4404 | case IORING_UNREGISTER_BUFFERS: | |
4405 | ret = -EINVAL; | |
4406 | if (arg || nr_args) | |
4407 | break; | |
0a96bbe4 | 4408 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 4409 | break; |
6b06314c | 4410 | case IORING_REGISTER_FILES: |
a8da73a3 JA |
4411 | ret = -EFAULT; |
4412 | if (!arg) | |
4413 | break; | |
792e3582 | 4414 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
4415 | break; |
4416 | case IORING_UNREGISTER_FILES: | |
4417 | ret = -EINVAL; | |
4418 | if (arg || nr_args) | |
4419 | break; | |
4420 | ret = io_sqe_files_unregister(ctx); | |
4421 | break; | |
c3a31e60 | 4422 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 4423 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 4424 | break; |
9b402849 JA |
4425 | case IORING_REGISTER_EVENTFD: |
4426 | ret = -EINVAL; | |
4427 | if (nr_args != 1) | |
4428 | break; | |
c75312dd UA |
4429 | ret = io_eventfd_register(ctx, arg, 0); |
4430 | break; | |
4431 | case IORING_REGISTER_EVENTFD_ASYNC: | |
4432 | ret = -EINVAL; | |
4433 | if (nr_args != 1) | |
f2842ab5 | 4434 | break; |
c75312dd | 4435 | ret = io_eventfd_register(ctx, arg, 1); |
9b402849 JA |
4436 | break; |
4437 | case IORING_UNREGISTER_EVENTFD: | |
4438 | ret = -EINVAL; | |
4439 | if (arg || nr_args) | |
4440 | break; | |
4441 | ret = io_eventfd_unregister(ctx); | |
4442 | break; | |
66f4af93 JA |
4443 | case IORING_REGISTER_PROBE: |
4444 | ret = -EINVAL; | |
4445 | if (!arg || nr_args > 256) | |
4446 | break; | |
4447 | ret = io_probe(ctx, arg, nr_args); | |
4448 | break; | |
071698e1 JA |
4449 | case IORING_REGISTER_PERSONALITY: |
4450 | ret = -EINVAL; | |
4451 | if (arg || nr_args) | |
4452 | break; | |
4453 | ret = io_register_personality(ctx); | |
4454 | break; | |
4455 | case IORING_UNREGISTER_PERSONALITY: | |
4456 | ret = -EINVAL; | |
4457 | if (arg) | |
4458 | break; | |
4459 | ret = io_unregister_personality(ctx, nr_args); | |
4460 | break; | |
7e84e1c7 SG |
4461 | case IORING_REGISTER_ENABLE_RINGS: |
4462 | ret = -EINVAL; | |
4463 | if (arg || nr_args) | |
4464 | break; | |
4465 | ret = io_register_enable_rings(ctx); | |
4466 | break; | |
21b55dbc SG |
4467 | case IORING_REGISTER_RESTRICTIONS: |
4468 | ret = io_register_restrictions(ctx, arg, nr_args); | |
4469 | break; | |
992da01a PB |
4470 | case IORING_REGISTER_FILES2: |
4471 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
4472 | break; | |
4473 | case IORING_REGISTER_FILES_UPDATE2: | |
4474 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
4475 | IORING_RSRC_FILE); | |
4476 | break; | |
4477 | case IORING_REGISTER_BUFFERS2: | |
4478 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 4479 | break; |
992da01a PB |
4480 | case IORING_REGISTER_BUFFERS_UPDATE: |
4481 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
4482 | IORING_RSRC_BUFFER); | |
c3bdad02 | 4483 | break; |
fe76421d JA |
4484 | case IORING_REGISTER_IOWQ_AFF: |
4485 | ret = -EINVAL; | |
4486 | if (!arg || !nr_args) | |
4487 | break; | |
4488 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
4489 | break; | |
4490 | case IORING_UNREGISTER_IOWQ_AFF: | |
4491 | ret = -EINVAL; | |
4492 | if (arg || nr_args) | |
4493 | break; | |
4494 | ret = io_unregister_iowq_aff(ctx); | |
4495 | break; | |
2e480058 JA |
4496 | case IORING_REGISTER_IOWQ_MAX_WORKERS: |
4497 | ret = -EINVAL; | |
4498 | if (!arg || nr_args != 2) | |
4499 | break; | |
4500 | ret = io_register_iowq_max_workers(ctx, arg); | |
4501 | break; | |
e7a6c00d JA |
4502 | case IORING_REGISTER_RING_FDS: |
4503 | ret = io_ringfd_register(ctx, arg, nr_args); | |
4504 | break; | |
4505 | case IORING_UNREGISTER_RING_FDS: | |
4506 | ret = io_ringfd_unregister(ctx, arg, nr_args); | |
4507 | break; | |
c7fb1942 JA |
4508 | case IORING_REGISTER_PBUF_RING: |
4509 | ret = -EINVAL; | |
4510 | if (!arg || nr_args != 1) | |
4511 | break; | |
4512 | ret = io_register_pbuf_ring(ctx, arg); | |
4513 | break; | |
4514 | case IORING_UNREGISTER_PBUF_RING: | |
4515 | ret = -EINVAL; | |
4516 | if (!arg || nr_args != 1) | |
4517 | break; | |
4518 | ret = io_unregister_pbuf_ring(ctx, arg); | |
4519 | break; | |
78a861b9 JA |
4520 | case IORING_REGISTER_SYNC_CANCEL: |
4521 | ret = -EINVAL; | |
4522 | if (!arg || nr_args != 1) | |
4523 | break; | |
4524 | ret = io_sync_cancel(ctx, arg); | |
4525 | break; | |
6e73dffb PB |
4526 | case IORING_REGISTER_FILE_ALLOC_RANGE: |
4527 | ret = -EINVAL; | |
4528 | if (!arg || nr_args) | |
4529 | break; | |
4530 | ret = io_register_file_alloc_range(ctx, arg); | |
4531 | break; | |
edafccee JA |
4532 | default: |
4533 | ret = -EINVAL; | |
4534 | break; | |
4535 | } | |
4536 | ||
edafccee JA |
4537 | return ret; |
4538 | } | |
4539 | ||
4540 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
4541 | void __user *, arg, unsigned int, nr_args) | |
4542 | { | |
4543 | struct io_ring_ctx *ctx; | |
4544 | long ret = -EBADF; | |
4545 | struct fd f; | |
7d3fd88d JT |
4546 | bool use_registered_ring; |
4547 | ||
4548 | use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING); | |
4549 | opcode &= ~IORING_REGISTER_USE_REGISTERED_RING; | |
edafccee | 4550 | |
34319084 JA |
4551 | if (opcode >= IORING_REGISTER_LAST) |
4552 | return -EINVAL; | |
4553 | ||
7d3fd88d JT |
4554 | if (use_registered_ring) { |
4555 | /* | |
4556 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
4557 | * need only dereference our task private array to find it. | |
4558 | */ | |
4559 | struct io_uring_task *tctx = current->io_uring; | |
edafccee | 4560 | |
7d3fd88d JT |
4561 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
4562 | return -EINVAL; | |
4563 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
4564 | f.file = tctx->registered_rings[fd]; | |
4565 | f.flags = 0; | |
4566 | if (unlikely(!f.file)) | |
4567 | return -EBADF; | |
4568 | } else { | |
4569 | f = fdget(fd); | |
4570 | if (unlikely(!f.file)) | |
4571 | return -EBADF; | |
4572 | ret = -EOPNOTSUPP; | |
4573 | if (!io_is_uring_fops(f.file)) | |
4574 | goto out_fput; | |
4575 | } | |
edafccee JA |
4576 | |
4577 | ctx = f.file->private_data; | |
4578 | ||
4579 | mutex_lock(&ctx->uring_lock); | |
4580 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
4581 | mutex_unlock(&ctx->uring_lock); | |
2757be22 | 4582 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); |
edafccee JA |
4583 | out_fput: |
4584 | fdput(f); | |
4585 | return ret; | |
4586 | } | |
4587 | ||
2b188cc1 JA |
4588 | static int __init io_uring_init(void) |
4589 | { | |
9c71d39a | 4590 | #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ |
d7f62e82 | 4591 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ |
9c71d39a | 4592 | BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ |
d7f62e82 SM |
4593 | } while (0) |
4594 | ||
4595 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9c71d39a SM |
4596 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) |
4597 | #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ | |
4598 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) | |
d7f62e82 SM |
4599 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); |
4600 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
4601 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
4602 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
4603 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
4604 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
4605 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9c71d39a SM |
4606 | BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); |
4607 | BUILD_BUG_SQE_ELEM(12, __u32, __pad1); | |
d7f62e82 | 4608 | BUILD_BUG_SQE_ELEM(16, __u64, addr); |
7d67af2c | 4609 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
4610 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
4611 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
4612 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
4613 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
4614 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
4615 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
4616 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
4617 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
4618 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
4619 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
4620 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
4621 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
4622 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
4623 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
4624 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 4625 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
9c71d39a SM |
4626 | BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); |
4627 | BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); | |
4628 | BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); | |
4629 | BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); | |
4630 | BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); | |
d7f62e82 SM |
4631 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
4632 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 4633 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 4634 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 4635 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 4636 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
b48c312b PB |
4637 | BUILD_BUG_SQE_ELEM(44, __u16, addr_len); |
4638 | BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); | |
e9621e2b | 4639 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
9c71d39a SM |
4640 | BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); |
4641 | BUILD_BUG_SQE_ELEM(56, __u64, __pad2); | |
d7f62e82 | 4642 | |
b0d658ec PB |
4643 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
4644 | sizeof(struct io_uring_rsrc_update)); | |
4645 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
4646 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
4647 | |
4648 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
4649 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
4650 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
4651 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 4652 | |
b0d658ec PB |
4653 | /* should fit into one byte */ |
4654 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
4655 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
4656 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 4657 | |
32c2d33e | 4658 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); |
16340eab | 4659 | |
3a4b89a2 JA |
4660 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
4661 | ||
d9b57aa3 | 4662 | io_uring_optable_init(); |
0702e536 | 4663 | |
91f245d5 | 4664 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | |
8751d154 | 4665 | SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU); |
2b188cc1 JA |
4666 | return 0; |
4667 | }; | |
4668 | __initcall(io_uring_init); |