]>
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 | |
cbc2e203 | 149 | static void io_queue_sqe(struct io_kiocb *req); |
c0e0d6ba | 150 | static void io_move_task_work_from_local(struct io_ring_ctx *ctx); |
c450178d | 151 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx); |
de0617e4 | 152 | |
c1755c25 | 153 | struct kmem_cache *req_cachep; |
2b188cc1 | 154 | |
2b188cc1 JA |
155 | struct sock *io_uring_get_socket(struct file *file) |
156 | { | |
157 | #if defined(CONFIG_UNIX) | |
cd40cae2 | 158 | if (io_is_uring_fops(file)) { |
2b188cc1 JA |
159 | struct io_ring_ctx *ctx = file->private_data; |
160 | ||
161 | return ctx->ring_sock->sk; | |
162 | } | |
163 | #endif | |
164 | return NULL; | |
165 | } | |
166 | EXPORT_SYMBOL(io_uring_get_socket); | |
167 | ||
c450178d PB |
168 | static inline void io_submit_flush_completions(struct io_ring_ctx *ctx) |
169 | { | |
931147dd DY |
170 | if (!wq_list_empty(&ctx->submit_state.compl_reqs) || |
171 | ctx->submit_state.cqes_count) | |
c450178d PB |
172 | __io_submit_flush_completions(ctx); |
173 | } | |
174 | ||
faf88dde PB |
175 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
176 | { | |
177 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
178 | } | |
179 | ||
0fc8c2ac DY |
180 | static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx) |
181 | { | |
182 | return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head); | |
183 | } | |
184 | ||
9cae36a0 JA |
185 | static bool io_match_linked(struct io_kiocb *head) |
186 | { | |
187 | struct io_kiocb *req; | |
188 | ||
189 | io_for_each_link(req, head) { | |
190 | if (req->flags & REQ_F_INFLIGHT) | |
191 | return true; | |
192 | } | |
193 | return false; | |
6af3f48b PB |
194 | } |
195 | ||
196 | /* | |
197 | * As io_match_task() but protected against racing with linked timeouts. | |
198 | * User must not hold timeout_lock. | |
199 | */ | |
329061d3 JA |
200 | bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, |
201 | bool cancel_all) | |
6af3f48b | 202 | { |
9cae36a0 JA |
203 | bool matched; |
204 | ||
6af3f48b PB |
205 | if (task && head->task != task) |
206 | return false; | |
9cae36a0 JA |
207 | if (cancel_all) |
208 | return true; | |
209 | ||
210 | if (head->flags & REQ_F_LINK_TIMEOUT) { | |
211 | struct io_ring_ctx *ctx = head->ctx; | |
212 | ||
213 | /* protect against races with linked timeouts */ | |
214 | spin_lock_irq(&ctx->timeout_lock); | |
215 | matched = io_match_linked(head); | |
216 | spin_unlock_irq(&ctx->timeout_lock); | |
217 | } else { | |
218 | matched = io_match_linked(head); | |
219 | } | |
220 | return matched; | |
6af3f48b PB |
221 | } |
222 | ||
a8295b98 HX |
223 | static inline void req_fail_link_node(struct io_kiocb *req, int res) |
224 | { | |
225 | req_set_fail(req); | |
97b388d7 | 226 | io_req_set_res(req, res, 0); |
a8295b98 HX |
227 | } |
228 | ||
fa05457a PB |
229 | static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) |
230 | { | |
231 | wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); | |
c1755c25 | 232 | kasan_poison_object_data(req_cachep, req); |
a8295b98 HX |
233 | } |
234 | ||
c072481d | 235 | static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) |
2b188cc1 JA |
236 | { |
237 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
238 | ||
0f158b4c | 239 | complete(&ctx->ref_comp); |
2b188cc1 JA |
240 | } |
241 | ||
c072481d | 242 | static __cold void io_fallback_req_func(struct work_struct *work) |
f56165e6 PB |
243 | { |
244 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
245 | fallback_work.work); | |
246 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
247 | struct io_kiocb *req, *tmp; | |
a282967c | 248 | struct io_tw_state ts = { .locked = true, }; |
f56165e6 | 249 | |
31f084b7 | 250 | mutex_lock(&ctx->uring_lock); |
3218e5d3 | 251 | llist_for_each_entry_safe(req, tmp, node, io_task_work.node) |
a282967c PB |
252 | req->io_task_work.func(req, &ts); |
253 | if (WARN_ON_ONCE(!ts.locked)) | |
31f084b7 PB |
254 | return; |
255 | io_submit_flush_completions(ctx); | |
256 | mutex_unlock(&ctx->uring_lock); | |
f56165e6 PB |
257 | } |
258 | ||
e6f89be6 PB |
259 | static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) |
260 | { | |
261 | unsigned hash_buckets = 1U << bits; | |
262 | size_t hash_size = hash_buckets * sizeof(table->hbs[0]); | |
263 | ||
264 | table->hbs = kmalloc(hash_size, GFP_KERNEL); | |
265 | if (!table->hbs) | |
266 | return -ENOMEM; | |
267 | ||
268 | table->hash_bits = bits; | |
269 | init_hash_table(table, hash_buckets); | |
270 | return 0; | |
271 | } | |
272 | ||
c072481d | 273 | static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
2b188cc1 JA |
274 | { |
275 | struct io_ring_ctx *ctx; | |
9cfc7e94 | 276 | int hash_bits; |
2b188cc1 JA |
277 | |
278 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
279 | if (!ctx) | |
280 | return NULL; | |
281 | ||
9cfc7e94 JA |
282 | xa_init(&ctx->io_bl_xa); |
283 | ||
78076bb6 JA |
284 | /* |
285 | * Use 5 bits less than the max cq entries, that should give us around | |
4a07723f PB |
286 | * 32 entries per hash list if totally full and uniformly spread, but |
287 | * don't keep too many buckets to not overconsume memory. | |
78076bb6 | 288 | */ |
4a07723f PB |
289 | hash_bits = ilog2(p->cq_entries) - 5; |
290 | hash_bits = clamp(hash_bits, 1, 8); | |
e6f89be6 | 291 | if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) |
78076bb6 | 292 | goto err; |
9ca9fb24 PB |
293 | if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) |
294 | goto err; | |
38513c46 | 295 | |
6224843d PB |
296 | ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL); |
297 | if (!ctx->dummy_ubuf) | |
298 | goto err; | |
299 | /* set invalid range, so io_import_fixed() fails meeting it */ | |
300 | ctx->dummy_ubuf->ubuf = -1UL; | |
301 | ||
21482896 | 302 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
48904229 | 303 | 0, GFP_KERNEL)) |
206aefde | 304 | goto err; |
2b188cc1 JA |
305 | |
306 | ctx->flags = p->flags; | |
90554200 | 307 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 308 | INIT_LIST_HEAD(&ctx->sqd_list); |
1d7bb1d5 | 309 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
cc3cec83 | 310 | INIT_LIST_HEAD(&ctx->io_buffers_cache); |
69bbc6ad PB |
311 | io_alloc_cache_init(&ctx->rsrc_node_cache, IO_NODE_ALLOC_CACHE_MAX, |
312 | sizeof(struct io_rsrc_node)); | |
313 | io_alloc_cache_init(&ctx->apoll_cache, IO_ALLOC_CACHE_MAX, | |
314 | sizeof(struct async_poll)); | |
315 | io_alloc_cache_init(&ctx->netmsg_cache, IO_ALLOC_CACHE_MAX, | |
316 | sizeof(struct io_async_msghdr)); | |
0f158b4c | 317 | init_completion(&ctx->ref_comp); |
61cf9370 | 318 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 319 | mutex_init(&ctx->uring_lock); |
311997b3 | 320 | init_waitqueue_head(&ctx->cq_wait); |
7b235dd8 | 321 | init_waitqueue_head(&ctx->poll_wq); |
4ea15b56 | 322 | init_waitqueue_head(&ctx->rsrc_quiesce_wq); |
2b188cc1 | 323 | spin_lock_init(&ctx->completion_lock); |
89850fce | 324 | spin_lock_init(&ctx->timeout_lock); |
5eef4e87 | 325 | INIT_WQ_LIST(&ctx->iopoll_list); |
cc3cec83 JA |
326 | INIT_LIST_HEAD(&ctx->io_buffers_pages); |
327 | INIT_LIST_HEAD(&ctx->io_buffers_comp); | |
de0617e4 | 328 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 329 | INIT_LIST_HEAD(&ctx->timeout_list); |
ef9dd637 | 330 | INIT_LIST_HEAD(&ctx->ltimeout_list); |
d67d2263 | 331 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); |
c0e0d6ba | 332 | init_llist_head(&ctx->work_llist); |
13bf43f5 | 333 | INIT_LIST_HEAD(&ctx->tctx_list); |
c2b6c6bc PB |
334 | ctx->submit_state.free_list.next = NULL; |
335 | INIT_WQ_LIST(&ctx->locked_free_list); | |
9011bf9a | 336 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
6f33b0bc | 337 | INIT_WQ_LIST(&ctx->submit_state.compl_reqs); |
2b188cc1 | 338 | return ctx; |
206aefde | 339 | err: |
6224843d | 340 | kfree(ctx->dummy_ubuf); |
e6f89be6 | 341 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 342 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 JA |
343 | kfree(ctx->io_bl); |
344 | xa_destroy(&ctx->io_bl_xa); | |
206aefde JA |
345 | kfree(ctx); |
346 | return NULL; | |
2b188cc1 JA |
347 | } |
348 | ||
8f6ed49a PB |
349 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
350 | { | |
351 | struct io_rings *r = ctx->rings; | |
352 | ||
353 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
354 | ctx->cq_extra--; | |
355 | } | |
356 | ||
9cf7c104 | 357 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 358 | { |
2bc9930e JA |
359 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
360 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 361 | |
8f6ed49a | 362 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 363 | } |
de0617e4 | 364 | |
9d858b21 | 365 | return false; |
de0617e4 JA |
366 | } |
367 | ||
5a754dea PB |
368 | static void io_clean_op(struct io_kiocb *req) |
369 | { | |
370 | if (req->flags & REQ_F_BUFFER_SELECTED) { | |
371 | spin_lock(&req->ctx->completion_lock); | |
372 | io_put_kbuf_comp(req); | |
373 | spin_unlock(&req->ctx->completion_lock); | |
374 | } | |
375 | ||
376 | if (req->flags & REQ_F_NEED_CLEANUP) { | |
377 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; | |
378 | ||
379 | if (def->cleanup) | |
380 | def->cleanup(req); | |
381 | } | |
382 | if ((req->flags & REQ_F_POLLED) && req->apoll) { | |
383 | kfree(req->apoll->double_poll); | |
384 | kfree(req->apoll); | |
385 | req->apoll = NULL; | |
386 | } | |
387 | if (req->flags & REQ_F_INFLIGHT) { | |
388 | struct io_uring_task *tctx = req->task->io_uring; | |
389 | ||
390 | atomic_dec(&tctx->inflight_tracked); | |
391 | } | |
392 | if (req->flags & REQ_F_CREDS) | |
393 | put_cred(req->creds); | |
394 | if (req->flags & REQ_F_ASYNC_DATA) { | |
395 | kfree(req->async_data); | |
396 | req->async_data = NULL; | |
397 | } | |
398 | req->flags &= ~IO_REQ_CLEAN_FLAGS; | |
399 | } | |
400 | ||
9cae36a0 JA |
401 | static inline void io_req_track_inflight(struct io_kiocb *req) |
402 | { | |
403 | if (!(req->flags & REQ_F_INFLIGHT)) { | |
404 | req->flags |= REQ_F_INFLIGHT; | |
386e4fb6 | 405 | atomic_inc(&req->task->io_uring->inflight_tracked); |
9cae36a0 JA |
406 | } |
407 | } | |
408 | ||
fd08e530 PB |
409 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
410 | { | |
906c6caa PB |
411 | if (WARN_ON_ONCE(!req->link)) |
412 | return NULL; | |
413 | ||
4d13d1a4 PB |
414 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
415 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
416 | |
417 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 418 | io_req_set_refcount(req); |
4d13d1a4 PB |
419 | __io_req_set_refcount(req->link, 2); |
420 | return req->link; | |
fd08e530 PB |
421 | } |
422 | ||
423 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
424 | { | |
4d13d1a4 | 425 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
426 | return NULL; |
427 | return __io_prep_linked_timeout(req); | |
428 | } | |
429 | ||
cb2d344c PB |
430 | static noinline void __io_arm_ltimeout(struct io_kiocb *req) |
431 | { | |
432 | io_queue_linked_timeout(__io_prep_linked_timeout(req)); | |
433 | } | |
434 | ||
435 | static inline void io_arm_ltimeout(struct io_kiocb *req) | |
436 | { | |
437 | if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) | |
438 | __io_arm_ltimeout(req); | |
439 | } | |
440 | ||
1e6fa521 JA |
441 | static void io_prep_async_work(struct io_kiocb *req) |
442 | { | |
a7dd2782 | 443 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
1e6fa521 JA |
444 | struct io_ring_ctx *ctx = req->ctx; |
445 | ||
b8e64b53 PB |
446 | if (!(req->flags & REQ_F_CREDS)) { |
447 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 448 | req->creds = get_current_cred(); |
b8e64b53 | 449 | } |
003e8dcc | 450 | |
e1d675df PB |
451 | req->work.list.next = NULL; |
452 | req->work.flags = 0; | |
8e29da69 | 453 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); |
feaadc4f PB |
454 | if (req->flags & REQ_F_FORCE_ASYNC) |
455 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
456 | ||
3beed235 | 457 | if (req->file && !(req->flags & REQ_F_FIXED_FILE)) |
8487f083 | 458 | req->flags |= io_file_get_flags(req->file); |
f6b543fd | 459 | |
8b1df11f | 460 | if (req->file && (req->flags & REQ_F_ISREG)) { |
d4755e15 JA |
461 | bool should_hash = def->hash_reg_file; |
462 | ||
463 | /* don't serialize this request if the fs doesn't need it */ | |
464 | if (should_hash && (req->file->f_flags & O_DIRECT) && | |
465 | (req->file->f_mode & FMODE_DIO_PARALLEL_WRITE)) | |
466 | should_hash = false; | |
467 | if (should_hash || (ctx->flags & IORING_SETUP_IOPOLL)) | |
1e6fa521 | 468 | io_wq_hash_work(&req->work, file_inode(req->file)); |
4b982bd0 | 469 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
470 | if (def->unbound_nonreg_file) |
471 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
472 | } | |
561fb04a | 473 | } |
cccf0ee8 | 474 | |
cbdcb435 | 475 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 476 | { |
cbdcb435 | 477 | struct io_kiocb *cur; |
54a91f3b | 478 | |
44eff40a PB |
479 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
480 | struct io_ring_ctx *ctx = req->ctx; | |
481 | ||
674ee8e1 | 482 | spin_lock_irq(&ctx->timeout_lock); |
44eff40a PB |
483 | io_for_each_link(cur, req) |
484 | io_prep_async_work(cur); | |
674ee8e1 | 485 | spin_unlock_irq(&ctx->timeout_lock); |
44eff40a PB |
486 | } else { |
487 | io_for_each_link(cur, req) | |
488 | io_prep_async_work(cur); | |
489 | } | |
561fb04a JA |
490 | } |
491 | ||
a282967c | 492 | void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use) |
561fb04a | 493 | { |
cbdcb435 | 494 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 495 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 496 | |
3bfe6106 JA |
497 | BUG_ON(!tctx); |
498 | BUG_ON(!tctx->io_wq); | |
561fb04a | 499 | |
cbdcb435 PB |
500 | /* init ->work of the whole link before punting */ |
501 | io_prep_async_link(req); | |
991468dc JA |
502 | |
503 | /* | |
504 | * Not expected to happen, but if we do have a bug where this _can_ | |
505 | * happen, catch it here and ensure the request is marked as | |
506 | * canceled. That will make io-wq go through the usual work cancel | |
507 | * procedure rather than attempt to run this request (or create a new | |
508 | * worker for it). | |
509 | */ | |
510 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
511 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
512 | ||
48863ffd | 513 | trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); |
ebf93667 | 514 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
515 | if (link) |
516 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
517 | } |
518 | ||
c072481d | 519 | static __cold void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 520 | { |
441b8a78 | 521 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
522 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
523 | struct io_defer_entry, list); | |
de0617e4 | 524 | |
9cf7c104 | 525 | if (req_need_defer(de->req, de->seq)) |
04518945 | 526 | break; |
27dc8338 | 527 | list_del_init(&de->list); |
907d1df3 | 528 | io_req_task_queue(de->req); |
27dc8338 | 529 | kfree(de); |
441b8a78 | 530 | } |
04518945 PB |
531 | } |
532 | ||
21a091b9 DY |
533 | |
534 | static void io_eventfd_ops(struct rcu_head *rcu) | |
f2842ab5 | 535 | { |
d8e9214f | 536 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
21a091b9 | 537 | int ops = atomic_xchg(&ev_fd->ops, 0); |
305bef98 | 538 | |
21a091b9 | 539 | if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT)) |
44648532 | 540 | eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); |
d8e9214f | 541 | |
21a091b9 DY |
542 | /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback |
543 | * ordering in a race but if references are 0 we know we have to free | |
544 | * it regardless. | |
305bef98 | 545 | */ |
21a091b9 DY |
546 | if (atomic_dec_and_test(&ev_fd->refs)) { |
547 | eventfd_ctx_put(ev_fd->cq_ev_fd); | |
548 | kfree(ev_fd); | |
549 | } | |
d8e9214f DY |
550 | } |
551 | ||
77bc59b4 | 552 | static void io_eventfd_signal(struct io_ring_ctx *ctx) |
f2842ab5 | 553 | { |
21a091b9 | 554 | struct io_ev_fd *ev_fd = NULL; |
77bc59b4 | 555 | |
77bc59b4 UA |
556 | rcu_read_lock(); |
557 | /* | |
558 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
559 | * and eventfd_signal | |
560 | */ | |
561 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
562 | ||
563 | /* | |
564 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
565 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
566 | * the function and rcu_read_lock. | |
567 | */ | |
568 | if (unlikely(!ev_fd)) | |
569 | goto out; | |
7e55a19c | 570 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 | 571 | goto out; |
21a091b9 DY |
572 | if (ev_fd->eventfd_async && !io_wq_current_is_worker()) |
573 | goto out; | |
77bc59b4 | 574 | |
21a091b9 | 575 | if (likely(eventfd_signal_allowed())) { |
44648532 | 576 | eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE); |
21a091b9 DY |
577 | } else { |
578 | atomic_inc(&ev_fd->refs); | |
579 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops)) | |
44a84da4 | 580 | call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops); |
21a091b9 DY |
581 | else |
582 | atomic_dec(&ev_fd->refs); | |
583 | } | |
584 | ||
77bc59b4 UA |
585 | out: |
586 | rcu_read_unlock(); | |
f2842ab5 JA |
587 | } |
588 | ||
21a091b9 DY |
589 | static void io_eventfd_flush_signal(struct io_ring_ctx *ctx) |
590 | { | |
591 | bool skip; | |
592 | ||
593 | spin_lock(&ctx->completion_lock); | |
594 | ||
595 | /* | |
596 | * Eventfd should only get triggered when at least one event has been | |
597 | * posted. Some applications rely on the eventfd notification count | |
598 | * only changing IFF a new CQE has been added to the CQ ring. There's | |
599 | * no depedency on 1:1 relationship between how many times this | |
600 | * function is called (and hence the eventfd count) and number of CQEs | |
601 | * posted to the CQ ring. | |
602 | */ | |
603 | skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
604 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
605 | spin_unlock(&ctx->completion_lock); | |
606 | if (skip) | |
607 | return; | |
608 | ||
609 | io_eventfd_signal(ctx); | |
610 | } | |
611 | ||
a830ffd2 PB |
612 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
613 | { | |
bca39f39 PB |
614 | if (ctx->poll_activated) |
615 | io_poll_wq_wake(ctx); | |
e5f30f6f PB |
616 | if (ctx->off_timeout_used) |
617 | io_flush_timeouts(ctx); | |
618 | if (ctx->drain_active) { | |
a830ffd2 | 619 | spin_lock(&ctx->completion_lock); |
e5f30f6f | 620 | io_queue_deferred(ctx); |
a830ffd2 PB |
621 | spin_unlock(&ctx->completion_lock); |
622 | } | |
623 | if (ctx->has_evfd) | |
21a091b9 | 624 | io_eventfd_flush_signal(ctx); |
a830ffd2 PB |
625 | } |
626 | ||
f66f7342 | 627 | static inline void __io_cq_lock(struct io_ring_ctx *ctx) |
f66f7342 PB |
628 | { |
629 | if (!ctx->task_complete) | |
630 | spin_lock(&ctx->completion_lock); | |
631 | } | |
632 | ||
6971253f PB |
633 | static inline void io_cq_lock(struct io_ring_ctx *ctx) |
634 | __acquires(ctx->completion_lock) | |
635 | { | |
636 | spin_lock(&ctx->completion_lock); | |
637 | } | |
638 | ||
f66f7342 | 639 | static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) |
3181e22f PB |
640 | { |
641 | io_commit_cqring(ctx); | |
3181e22f | 642 | |
c66ae3ec PB |
643 | if (ctx->task_complete) { |
644 | /* | |
645 | * ->task_complete implies that only current might be waiting | |
646 | * for CQEs, and obviously, we currently don't. No one is | |
647 | * waiting, wakeups are futile, skip them. | |
648 | */ | |
649 | io_commit_cqring_flush(ctx); | |
650 | } else { | |
ff126177 | 651 | spin_unlock(&ctx->completion_lock); |
c66ae3ec | 652 | io_commit_cqring_flush(ctx); |
6e7248ad | 653 | io_cqring_wake(ctx); |
3181e22f PB |
654 | } |
655 | } | |
656 | ||
0fdb9a19 | 657 | static void io_cq_unlock_post(struct io_ring_ctx *ctx) |
5d772916 | 658 | __releases(ctx->completion_lock) |
25399321 | 659 | { |
f66f7342 PB |
660 | io_commit_cqring(ctx); |
661 | spin_unlock(&ctx->completion_lock); | |
662 | io_commit_cqring_flush(ctx); | |
663 | io_cqring_wake(ctx); | |
25399321 PB |
664 | } |
665 | ||
c4a2ed72 | 666 | /* Returns true if there are no backlogged entries after the flush */ |
a85381d8 PB |
667 | static void io_cqring_overflow_kill(struct io_ring_ctx *ctx) |
668 | { | |
669 | struct io_overflow_cqe *ocqe; | |
670 | LIST_HEAD(list); | |
671 | ||
f432b76b | 672 | spin_lock(&ctx->completion_lock); |
a85381d8 PB |
673 | list_splice_init(&ctx->cq_overflow_list, &list); |
674 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); | |
f432b76b | 675 | spin_unlock(&ctx->completion_lock); |
a85381d8 PB |
676 | |
677 | while (!list_empty(&list)) { | |
678 | ocqe = list_first_entry(&list, struct io_overflow_cqe, list); | |
679 | list_del(&ocqe->list); | |
680 | kfree(ocqe); | |
681 | } | |
682 | } | |
683 | ||
1b346e4a | 684 | static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
1d7bb1d5 | 685 | { |
e45a3e05 | 686 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 687 | |
a85381d8 | 688 | if (__io_cqring_events(ctx) == ctx->cq_entries) |
1b346e4a | 689 | return; |
1d7bb1d5 | 690 | |
e45a3e05 SR |
691 | if (ctx->flags & IORING_SETUP_CQE32) |
692 | cqe_size <<= 1; | |
693 | ||
25399321 | 694 | io_cq_lock(ctx); |
6c2450ae | 695 | while (!list_empty(&ctx->cq_overflow_list)) { |
aa1df3a3 | 696 | struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true); |
6c2450ae | 697 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 698 | |
a85381d8 | 699 | if (!cqe) |
1d7bb1d5 | 700 | break; |
6c2450ae PB |
701 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
702 | struct io_overflow_cqe, list); | |
a85381d8 | 703 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae PB |
704 | list_del(&ocqe->list); |
705 | kfree(ocqe); | |
1d7bb1d5 JA |
706 | } |
707 | ||
1b346e4a | 708 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 709 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 710 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 711 | } |
25399321 | 712 | io_cq_unlock_post(ctx); |
1d7bb1d5 JA |
713 | } |
714 | ||
52ea806a JA |
715 | static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx) |
716 | { | |
717 | /* iopoll syncs against uring_lock, not completion_lock */ | |
718 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
719 | mutex_lock(&ctx->uring_lock); | |
720 | __io_cqring_overflow_flush(ctx); | |
721 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
722 | mutex_unlock(&ctx->uring_lock); | |
723 | } | |
724 | ||
1b346e4a | 725 | static void io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 726 | { |
52ea806a JA |
727 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) |
728 | io_cqring_do_overflow_flush(ctx); | |
6c503150 PB |
729 | } |
730 | ||
5afa4650 | 731 | /* can be called by any task */ |
2fdd6fb5 | 732 | static void io_put_task_remote(struct task_struct *task) |
6a290a14 PB |
733 | { |
734 | struct io_uring_task *tctx = task->io_uring; | |
735 | ||
2fdd6fb5 | 736 | percpu_counter_sub(&tctx->inflight, 1); |
8d664282 | 737 | if (unlikely(atomic_read(&tctx->in_cancel))) |
9d170164 | 738 | wake_up(&tctx->wait); |
2fdd6fb5 | 739 | put_task_struct(task); |
9d170164 PB |
740 | } |
741 | ||
5afa4650 | 742 | /* used by a task to put its own references */ |
2fdd6fb5 | 743 | static void io_put_task_local(struct task_struct *task) |
5afa4650 | 744 | { |
2fdd6fb5 | 745 | task->io_uring->cached_refs++; |
5afa4650 PB |
746 | } |
747 | ||
89800a2d | 748 | /* must to be called somewhat shortly after putting a request */ |
2fdd6fb5 | 749 | static inline void io_put_task(struct task_struct *task) |
89800a2d PB |
750 | { |
751 | if (likely(task == current)) | |
2fdd6fb5 | 752 | io_put_task_local(task); |
89800a2d | 753 | else |
2fdd6fb5 | 754 | io_put_task_remote(task); |
89800a2d PB |
755 | } |
756 | ||
63809137 | 757 | void io_task_refs_refill(struct io_uring_task *tctx) |
9a10867a PB |
758 | { |
759 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
760 | ||
761 | percpu_counter_add(&tctx->inflight, refill); | |
762 | refcount_add(refill, ¤t->usage); | |
763 | tctx->cached_refs += refill; | |
764 | } | |
765 | ||
3cc7fdb9 PB |
766 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
767 | { | |
768 | struct io_uring_task *tctx = task->io_uring; | |
769 | unsigned int refs = tctx->cached_refs; | |
770 | ||
771 | if (refs) { | |
772 | tctx->cached_refs = 0; | |
773 | percpu_counter_sub(&tctx->inflight, refs); | |
774 | put_task_struct_many(task, refs); | |
775 | } | |
776 | } | |
777 | ||
68494a65 PB |
778 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
779 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 780 | { |
cce4b8b0 | 781 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
782 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
783 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 784 | |
f26cc959 PB |
785 | lockdep_assert_held(&ctx->completion_lock); |
786 | ||
e45a3e05 SR |
787 | if (is_cqe32) |
788 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 789 | |
e45a3e05 | 790 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 791 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
792 | if (!ocqe) { |
793 | /* | |
794 | * If we're in ring overflow flush mode, or in task cancel mode, | |
795 | * or cannot allocate an overflow entry, then we need to drop it | |
796 | * on the floor. | |
797 | */ | |
8f6ed49a | 798 | io_account_cq_overflow(ctx); |
155bc950 | 799 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 800 | return false; |
2b188cc1 | 801 | } |
cce4b8b0 | 802 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 803 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 804 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 805 | |
cce4b8b0 | 806 | } |
d4d19c19 | 807 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
808 | ocqe->cqe.res = res; |
809 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
810 | if (is_cqe32) { |
811 | ocqe->cqe.big_cqe[0] = extra1; | |
812 | ocqe->cqe.big_cqe[1] = extra2; | |
813 | } | |
cce4b8b0 PB |
814 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
815 | return true; | |
2b188cc1 JA |
816 | } |
817 | ||
68494a65 PB |
818 | bool io_req_cqe_overflow(struct io_kiocb *req) |
819 | { | |
820 | if (!(req->flags & REQ_F_CQE32_INIT)) { | |
821 | req->extra1 = 0; | |
822 | req->extra2 = 0; | |
823 | } | |
824 | return io_cqring_event_overflow(req->ctx, req->cqe.user_data, | |
825 | req->cqe.res, req->cqe.flags, | |
826 | req->extra1, req->extra2); | |
827 | } | |
828 | ||
faf88dde PB |
829 | /* |
830 | * writes to the cq entry need to come after reading head; the | |
831 | * control dependency is enough as we're using WRITE_ONCE to | |
832 | * fill the cq entry | |
833 | */ | |
aa1df3a3 | 834 | struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow) |
faf88dde PB |
835 | { |
836 | struct io_rings *rings = ctx->rings; | |
837 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
838 | unsigned int free, queued, len; |
839 | ||
aa1df3a3 PB |
840 | /* |
841 | * Posting into the CQ when there are pending overflowed CQEs may break | |
842 | * ordering guarantees, which will affect links, F_MORE users and more. | |
843 | * Force overflow the completion. | |
844 | */ | |
845 | if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))) | |
846 | return NULL; | |
faf88dde PB |
847 | |
848 | /* userspace may cheat modifying the tail, be safe and do min */ | |
849 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
850 | free = ctx->cq_entries - queued; | |
851 | /* we need a contiguous range, limit based on the current array offset */ | |
852 | len = min(free, ctx->cq_entries - off); | |
853 | if (!len) | |
854 | return NULL; | |
855 | ||
b3659a65 PB |
856 | if (ctx->flags & IORING_SETUP_CQE32) { |
857 | off <<= 1; | |
858 | len <<= 1; | |
859 | } | |
860 | ||
faf88dde PB |
861 | ctx->cqe_cached = &rings->cqes[off]; |
862 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
b3659a65 PB |
863 | |
864 | ctx->cached_cq_tail++; | |
faf88dde | 865 | ctx->cqe_cached++; |
b3659a65 PB |
866 | if (ctx->flags & IORING_SETUP_CQE32) |
867 | ctx->cqe_cached++; | |
868 | return &rings->cqes[off]; | |
faf88dde PB |
869 | } |
870 | ||
f66f7342 PB |
871 | static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, |
872 | u32 cflags) | |
bcda7baa | 873 | { |
cd94903d PB |
874 | struct io_uring_cqe *cqe; |
875 | ||
913a571a | 876 | ctx->cq_extra++; |
cd94903d PB |
877 | |
878 | /* | |
879 | * If we can't get a cq entry, userspace overflowed the | |
880 | * submission (by quite a lot). Increment the overflow count in | |
881 | * the ring. | |
882 | */ | |
883 | cqe = io_get_cqe(ctx); | |
884 | if (likely(cqe)) { | |
e0486f3f DY |
885 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
886 | ||
cd94903d PB |
887 | WRITE_ONCE(cqe->user_data, user_data); |
888 | WRITE_ONCE(cqe->res, res); | |
889 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
890 | |
891 | if (ctx->flags & IORING_SETUP_CQE32) { | |
892 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
893 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
894 | } | |
cd94903d PB |
895 | return true; |
896 | } | |
52120f0f | 897 | return false; |
bcda7baa JA |
898 | } |
899 | ||
931147dd DY |
900 | static void __io_flush_post_cqes(struct io_ring_ctx *ctx) |
901 | __must_hold(&ctx->uring_lock) | |
902 | { | |
903 | struct io_submit_state *state = &ctx->submit_state; | |
904 | unsigned int i; | |
905 | ||
906 | lockdep_assert_held(&ctx->uring_lock); | |
907 | for (i = 0; i < state->cqes_count; i++) { | |
908 | struct io_uring_cqe *cqe = &state->cqes[i]; | |
909 | ||
f66f7342 PB |
910 | if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) { |
911 | if (ctx->task_complete) { | |
912 | spin_lock(&ctx->completion_lock); | |
913 | io_cqring_event_overflow(ctx, cqe->user_data, | |
914 | cqe->res, cqe->flags, 0, 0); | |
915 | spin_unlock(&ctx->completion_lock); | |
916 | } else { | |
917 | io_cqring_event_overflow(ctx, cqe->user_data, | |
918 | cqe->res, cqe->flags, 0, 0); | |
919 | } | |
920 | } | |
931147dd DY |
921 | } |
922 | state->cqes_count = 0; | |
923 | } | |
924 | ||
b529c96a DY |
925 | static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, |
926 | bool allow_overflow) | |
d245bca6 PB |
927 | { |
928 | bool filled; | |
929 | ||
25399321 | 930 | io_cq_lock(ctx); |
f66f7342 PB |
931 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags); |
932 | if (!filled && allow_overflow) | |
933 | filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); | |
934 | ||
25399321 | 935 | io_cq_unlock_post(ctx); |
d245bca6 PB |
936 | return filled; |
937 | } | |
938 | ||
b529c96a DY |
939 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags) |
940 | { | |
941 | return __io_post_aux_cqe(ctx, user_data, res, cflags, true); | |
942 | } | |
943 | ||
d86eaed1 | 944 | bool io_aux_cqe(const struct io_kiocb *req, bool defer, s32 res, u32 cflags, |
9b8c5475 | 945 | bool allow_overflow) |
2b188cc1 | 946 | { |
d86eaed1 JA |
947 | struct io_ring_ctx *ctx = req->ctx; |
948 | u64 user_data = req->cqe.user_data; | |
9b8c5475 | 949 | struct io_uring_cqe *cqe; |
9b8c5475 DY |
950 | |
951 | if (!defer) | |
b529c96a | 952 | return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow); |
9b8c5475 | 953 | |
9b8c5475 DY |
954 | lockdep_assert_held(&ctx->uring_lock); |
955 | ||
003f242b | 956 | if (ctx->submit_state.cqes_count == ARRAY_SIZE(ctx->submit_state.cqes)) { |
f66f7342 | 957 | __io_cq_lock(ctx); |
9b8c5475 DY |
958 | __io_flush_post_cqes(ctx); |
959 | /* no need to flush - flush is deferred */ | |
f66f7342 | 960 | __io_cq_unlock_post(ctx); |
9b8c5475 DY |
961 | } |
962 | ||
963 | /* For defered completions this is not as strict as it is otherwise, | |
964 | * however it's main job is to prevent unbounded posted completions, | |
965 | * and in that it works just as well. | |
966 | */ | |
967 | if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) | |
968 | return false; | |
969 | ||
970 | cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++]; | |
971 | cqe->user_data = user_data; | |
972 | cqe->res = res; | |
973 | cqe->flags = cflags; | |
974 | return true; | |
975 | } | |
976 | ||
ef8ae64f | 977 | static void __io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
2b188cc1 | 978 | { |
fa18fa22 | 979 | struct io_ring_ctx *ctx = req->ctx; |
2ad4c6d0 | 980 | struct io_rsrc_node *rsrc_node = NULL; |
fa18fa22 PB |
981 | |
982 | io_cq_lock(ctx); | |
983 | if (!(req->flags & REQ_F_CQE_SKIP)) | |
a8cf95f9 | 984 | io_fill_cqe_req(ctx, req); |
fa18fa22 | 985 | |
c7dae4ba JA |
986 | /* |
987 | * If we're the last reference to this request, add to our locked | |
988 | * free_list cache. | |
989 | */ | |
de9b4cca | 990 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 991 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 992 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
993 | io_disarm_next(req); |
994 | if (req->link) { | |
995 | io_req_task_queue(req->link); | |
996 | req->link = NULL; | |
997 | } | |
998 | } | |
68a2cc1b | 999 | io_put_kbuf_comp(req); |
3b7a612f PB |
1000 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1001 | io_clean_op(req); | |
1002 | if (!(req->flags & REQ_F_FIXED_FILE)) | |
1003 | io_put_file(req->file); | |
1004 | ||
2ad4c6d0 | 1005 | rsrc_node = req->rsrc_node; |
8197b053 PB |
1006 | /* |
1007 | * Selected buffer deallocation in io_clean_op() assumes that | |
1008 | * we don't hold ->completion_lock. Clean them here to avoid | |
1009 | * deadlocks. | |
1010 | */ | |
2fdd6fb5 | 1011 | io_put_task_remote(req->task); |
c2b6c6bc | 1012 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 1013 | ctx->locked_free_nr++; |
180f829f | 1014 | } |
25399321 | 1015 | io_cq_unlock_post(ctx); |
2ad4c6d0 | 1016 | |
ef8ae64f PB |
1017 | if (rsrc_node) { |
1018 | io_ring_submit_lock(ctx, issue_flags); | |
1f2c8f61 | 1019 | io_put_rsrc_node(ctx, rsrc_node); |
ef8ae64f PB |
1020 | io_ring_submit_unlock(ctx, issue_flags); |
1021 | } | |
4e3d9ff9 JA |
1022 | } |
1023 | ||
1bec951c | 1024 | void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags) |
bcda7baa | 1025 | { |
860e1c7f | 1026 | if (req->ctx->task_complete && req->ctx->submitter_task != current) { |
e6aeb272 PB |
1027 | req->io_task_work.func = io_req_task_complete; |
1028 | io_req_task_work_add(req); | |
1029 | } else if (!(issue_flags & IO_URING_F_UNLOCKED) || | |
1030 | !(req->ctx->flags & IORING_SETUP_IOPOLL)) { | |
ef8ae64f | 1031 | __io_req_complete_post(req, issue_flags); |
1bec951c PB |
1032 | } else { |
1033 | struct io_ring_ctx *ctx = req->ctx; | |
1034 | ||
1035 | mutex_lock(&ctx->uring_lock); | |
ef8ae64f | 1036 | __io_req_complete_post(req, issue_flags & ~IO_URING_F_UNLOCKED); |
1bec951c PB |
1037 | mutex_unlock(&ctx->uring_lock); |
1038 | } | |
0ddf92e8 JA |
1039 | } |
1040 | ||
973fc83f | 1041 | void io_req_defer_failed(struct io_kiocb *req, s32 res) |
e276ae34 | 1042 | __must_hold(&ctx->uring_lock) |
f41db273 | 1043 | { |
f30bd4d0 | 1044 | const struct io_cold_def *def = &io_cold_defs[req->opcode]; |
a47b255e | 1045 | |
e276ae34 PB |
1046 | lockdep_assert_held(&req->ctx->uring_lock); |
1047 | ||
93d2bcd2 | 1048 | req_set_fail(req); |
97b388d7 | 1049 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
a47b255e PB |
1050 | if (def->fail) |
1051 | def->fail(req); | |
973fc83f | 1052 | io_req_complete_defer(req); |
f41db273 PB |
1053 | } |
1054 | ||
864ea921 PB |
1055 | /* |
1056 | * Don't initialise the fields below on every allocation, but do that in | |
1057 | * advance and keep them valid across allocations. | |
1058 | */ | |
1059 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
1060 | { | |
1061 | req->ctx = ctx; | |
1062 | req->link = NULL; | |
1063 | req->async_data = NULL; | |
1064 | /* not necessary, but safer to zero */ | |
cef216fc | 1065 | req->cqe.res = 0; |
864ea921 PB |
1066 | } |
1067 | ||
dac7a098 | 1068 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 1069 | struct io_submit_state *state) |
dac7a098 | 1070 | { |
79ebeaee | 1071 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 1072 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 1073 | ctx->locked_free_nr = 0; |
79ebeaee | 1074 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
1075 | } |
1076 | ||
5d5901a3 PB |
1077 | /* |
1078 | * A request might get retired back into the request caches even before opcode | |
1079 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
1080 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
1081 | * and with extra caution to not get a request that is still worked on. | |
1082 | */ | |
bd1a3783 | 1083 | __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 1084 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 1085 | { |
864ea921 | 1086 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 1087 | void *reqs[IO_REQ_ALLOC_BATCH]; |
864ea921 | 1088 | int ret, i; |
e5d1bc0a | 1089 | |
23a5c43b PB |
1090 | /* |
1091 | * If we have more than a batch's worth of requests in our IRQ side | |
1092 | * locked cache, grab the lock and move them over to our submission | |
1093 | * side cache. | |
1094 | */ | |
a6d97a8a | 1095 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 1096 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 1097 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
1098 | return true; |
1099 | } | |
e5d1bc0a | 1100 | |
3ab665b7 | 1101 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 1102 | |
864ea921 PB |
1103 | /* |
1104 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1105 | * retry single alloc to be on the safe side. | |
1106 | */ | |
1107 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
1108 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
1109 | if (!reqs[0]) | |
a33ae9ce | 1110 | return false; |
864ea921 | 1111 | ret = 1; |
2b188cc1 | 1112 | } |
864ea921 | 1113 | |
37f0e767 | 1114 | percpu_ref_get_many(&ctx->refs, ret); |
3ab665b7 | 1115 | for (i = 0; i < ret; i++) { |
23a5c43b | 1116 | struct io_kiocb *req = reqs[i]; |
3ab665b7 PB |
1117 | |
1118 | io_preinit_req(req, ctx); | |
fa05457a | 1119 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 1120 | } |
a33ae9ce PB |
1121 | return true; |
1122 | } | |
1123 | ||
03adabe8 PB |
1124 | __cold void io_free_req(struct io_kiocb *req) |
1125 | { | |
6ec9afc7 PB |
1126 | /* refs were already put, restore them for io_req_task_complete() */ |
1127 | req->flags &= ~REQ_F_REFCOUNT; | |
1128 | /* we only want to free it, don't post CQEs */ | |
1129 | req->flags |= REQ_F_CQE_SKIP; | |
1130 | req->io_task_work.func = io_req_task_complete; | |
03adabe8 PB |
1131 | io_req_task_work_add(req); |
1132 | } | |
1133 | ||
d81499bf PB |
1134 | static void __io_req_find_next_prep(struct io_kiocb *req) |
1135 | { | |
1136 | struct io_ring_ctx *ctx = req->ctx; | |
d81499bf | 1137 | |
6971253f | 1138 | spin_lock(&ctx->completion_lock); |
305bef98 | 1139 | io_disarm_next(req); |
6971253f | 1140 | spin_unlock(&ctx->completion_lock); |
d81499bf PB |
1141 | } |
1142 | ||
1143 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 1144 | { |
33cc89a9 | 1145 | struct io_kiocb *nxt; |
944e58bf | 1146 | |
9e645e11 JA |
1147 | /* |
1148 | * If LINK is set, we have dependent requests in this chain. If we | |
1149 | * didn't fail this request, queue the first one up, moving any other | |
1150 | * dependencies to the next request. In case of failure, fail the rest | |
1151 | * of the chain. | |
1152 | */ | |
d81499bf PB |
1153 | if (unlikely(req->flags & IO_DISARM_MASK)) |
1154 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
1155 | nxt = req->link; |
1156 | req->link = NULL; | |
1157 | return nxt; | |
4d7dd462 | 1158 | } |
9e645e11 | 1159 | |
a282967c | 1160 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, struct io_tw_state *ts) |
2c32395d PB |
1161 | { |
1162 | if (!ctx) | |
1163 | return; | |
ef060ea9 JA |
1164 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1165 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
a282967c | 1166 | if (ts->locked) { |
c450178d | 1167 | io_submit_flush_completions(ctx); |
2c32395d | 1168 | mutex_unlock(&ctx->uring_lock); |
a282967c | 1169 | ts->locked = false; |
2c32395d PB |
1170 | } |
1171 | percpu_ref_put(&ctx->refs); | |
1172 | } | |
1173 | ||
c6dd763c | 1174 | static unsigned int handle_tw_list(struct llist_node *node, |
a282967c PB |
1175 | struct io_ring_ctx **ctx, |
1176 | struct io_tw_state *ts, | |
c6dd763c | 1177 | struct llist_node *last) |
9f8d032a | 1178 | { |
c6dd763c DY |
1179 | unsigned int count = 0; |
1180 | ||
cb6bf7f2 | 1181 | while (node && node != last) { |
f88262e6 | 1182 | struct llist_node *next = node->next; |
9f8d032a HX |
1183 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1184 | io_task_work.node); | |
1185 | ||
34d2bfe7 JA |
1186 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); |
1187 | ||
9f8d032a | 1188 | if (req->ctx != *ctx) { |
a282967c | 1189 | ctx_flush_and_put(*ctx, ts); |
9f8d032a HX |
1190 | *ctx = req->ctx; |
1191 | /* if not contended, grab and improve batching */ | |
a282967c | 1192 | ts->locked = mutex_trylock(&(*ctx)->uring_lock); |
9f8d032a | 1193 | percpu_ref_get(&(*ctx)->refs); |
13bfa6f1 | 1194 | } |
c92fcfc2 JA |
1195 | INDIRECT_CALL_2(req->io_task_work.func, |
1196 | io_poll_task_func, io_req_rw_complete, | |
1197 | req, ts); | |
9f8d032a | 1198 | node = next; |
c6dd763c | 1199 | count++; |
f5868008 | 1200 | if (unlikely(need_resched())) { |
a282967c | 1201 | ctx_flush_and_put(*ctx, ts); |
f5868008 JA |
1202 | *ctx = NULL; |
1203 | cond_resched(); | |
1204 | } | |
3a0c037b | 1205 | } |
c6dd763c DY |
1206 | |
1207 | return count; | |
9f8d032a HX |
1208 | } |
1209 | ||
923d1592 DY |
1210 | /** |
1211 | * io_llist_xchg - swap all entries in a lock-less list | |
1212 | * @head: the head of lock-less list to delete all entries | |
1213 | * @new: new entry as the head of the list | |
1214 | * | |
1215 | * If list is empty, return NULL, otherwise, return the pointer to the first entry. | |
1216 | * The order of entries returned is from the newest to the oldest added one. | |
1217 | */ | |
1218 | static inline struct llist_node *io_llist_xchg(struct llist_head *head, | |
1219 | struct llist_node *new) | |
1220 | { | |
1221 | return xchg(&head->first, new); | |
1222 | } | |
1223 | ||
1224 | /** | |
1225 | * io_llist_cmpxchg - possibly swap all entries in a lock-less list | |
1226 | * @head: the head of lock-less list to delete all entries | |
1227 | * @old: expected old value of the first entry of the list | |
1228 | * @new: new entry as the head of the list | |
1229 | * | |
1230 | * perform a cmpxchg on the first entry of the list. | |
1231 | */ | |
1232 | ||
1233 | static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head, | |
1234 | struct llist_node *old, | |
1235 | struct llist_node *new) | |
1236 | { | |
1237 | return cmpxchg(&head->first, old, new); | |
1238 | } | |
1239 | ||
dfbe5561 | 1240 | static __cold void io_fallback_tw(struct io_uring_task *tctx, bool sync) |
10e1c0d5 JA |
1241 | { |
1242 | struct llist_node *node = llist_del_all(&tctx->task_list); | |
dfbe5561 | 1243 | struct io_ring_ctx *last_ctx = NULL; |
10e1c0d5 JA |
1244 | struct io_kiocb *req; |
1245 | ||
1246 | while (node) { | |
1247 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1248 | node = node->next; | |
dfbe5561 JA |
1249 | if (sync && last_ctx != req->ctx) { |
1250 | if (last_ctx) { | |
1251 | flush_delayed_work(&last_ctx->fallback_work); | |
1252 | percpu_ref_put(&last_ctx->refs); | |
1253 | } | |
1254 | last_ctx = req->ctx; | |
1255 | percpu_ref_get(&last_ctx->refs); | |
1256 | } | |
10e1c0d5 JA |
1257 | if (llist_add(&req->io_task_work.node, |
1258 | &req->ctx->fallback_llist)) | |
1259 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1260 | } | |
dfbe5561 JA |
1261 | |
1262 | if (last_ctx) { | |
1263 | flush_delayed_work(&last_ctx->fallback_work); | |
1264 | percpu_ref_put(&last_ctx->refs); | |
1265 | } | |
10e1c0d5 JA |
1266 | } |
1267 | ||
c9f06aa7 | 1268 | void tctx_task_work(struct callback_head *cb) |
c40f6379 | 1269 | { |
a282967c | 1270 | struct io_tw_state ts = {}; |
ebd0df2e | 1271 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
1272 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
1273 | task_work); | |
3a0c037b | 1274 | struct llist_node fake = {}; |
77e443ab | 1275 | struct llist_node *node; |
cb6bf7f2 PB |
1276 | unsigned int loops = 0; |
1277 | unsigned int count = 0; | |
3a0c037b | 1278 | |
77e443ab | 1279 | if (unlikely(current->flags & PF_EXITING)) { |
dfbe5561 | 1280 | io_fallback_tw(tctx, true); |
77e443ab PB |
1281 | return; |
1282 | } | |
3a0c037b | 1283 | |
cb6bf7f2 | 1284 | do { |
c6dd763c | 1285 | loops++; |
3a0c037b | 1286 | node = io_llist_xchg(&tctx->task_list, &fake); |
a282967c | 1287 | count += handle_tw_list(node, &ctx, &ts, &fake); |
50470fc5 PB |
1288 | |
1289 | /* skip expensive cmpxchg if there are items in the list */ | |
1290 | if (READ_ONCE(tctx->task_list.first) != &fake) | |
1291 | continue; | |
a282967c | 1292 | if (ts.locked && !wq_list_empty(&ctx->submit_state.compl_reqs)) { |
50470fc5 PB |
1293 | io_submit_flush_completions(ctx); |
1294 | if (READ_ONCE(tctx->task_list.first) != &fake) | |
1295 | continue; | |
1296 | } | |
3a0c037b | 1297 | node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); |
cb6bf7f2 | 1298 | } while (node != &fake); |
ebd0df2e | 1299 | |
a282967c | 1300 | ctx_flush_and_put(ctx, &ts); |
3cc7fdb9 | 1301 | |
8d664282 JA |
1302 | /* relaxed read is enough as only the task itself sets ->in_cancel */ |
1303 | if (unlikely(atomic_read(&tctx->in_cancel))) | |
3cc7fdb9 | 1304 | io_uring_drop_tctx_refs(current); |
c6dd763c DY |
1305 | |
1306 | trace_io_uring_task_work_run(tctx, count, loops); | |
7cbf1722 JA |
1307 | } |
1308 | ||
91c7884a | 1309 | static inline void io_req_local_work_add(struct io_kiocb *req, unsigned flags) |
c0e0d6ba DY |
1310 | { |
1311 | struct io_ring_ctx *ctx = req->ctx; | |
8751d154 | 1312 | unsigned nr_wait, nr_tw, nr_tw_prev; |
51509400 | 1313 | struct llist_node *first; |
c0e0d6ba | 1314 | |
8751d154 PB |
1315 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) |
1316 | flags &= ~IOU_F_TWQ_LAZY_WAKE; | |
ce8e04f6 | 1317 | |
51509400 PB |
1318 | first = READ_ONCE(ctx->work_llist.first); |
1319 | do { | |
8751d154 PB |
1320 | nr_tw_prev = 0; |
1321 | if (first) { | |
1322 | struct io_kiocb *first_req = container_of(first, | |
1323 | struct io_kiocb, | |
1324 | io_task_work.node); | |
1325 | /* | |
1326 | * Might be executed at any moment, rely on | |
1327 | * SLAB_TYPESAFE_BY_RCU to keep it alive. | |
1328 | */ | |
1329 | nr_tw_prev = READ_ONCE(first_req->nr_tw); | |
1330 | } | |
1331 | nr_tw = nr_tw_prev + 1; | |
1332 | /* Large enough to fail the nr_wait comparison below */ | |
1333 | if (!(flags & IOU_F_TWQ_LAZY_WAKE)) | |
1334 | nr_tw = -1U; | |
1335 | ||
1336 | req->nr_tw = nr_tw; | |
51509400 PB |
1337 | req->io_task_work.node.next = first; |
1338 | } while (!try_cmpxchg(&ctx->work_llist.first, &first, | |
1339 | &req->io_task_work.node)); | |
1340 | ||
8751d154 | 1341 | if (!first) { |
8751d154 PB |
1342 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1343 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1344 | if (ctx->has_evfd) | |
1345 | io_eventfd_signal(ctx); | |
c0e0d6ba DY |
1346 | } |
1347 | ||
8751d154 PB |
1348 | nr_wait = atomic_read(&ctx->cq_wait_nr); |
1349 | /* no one is waiting */ | |
1350 | if (!nr_wait) | |
1351 | return; | |
1352 | /* either not enough or the previous add has already woken it up */ | |
1353 | if (nr_wait > nr_tw || nr_tw_prev >= nr_wait) | |
1354 | return; | |
1355 | /* pairs with set_current_state() in io_cqring_wait() */ | |
1356 | smp_mb__after_atomic(); | |
1357 | wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE); | |
c0e0d6ba DY |
1358 | } |
1359 | ||
91c7884a | 1360 | static void io_req_normal_work_add(struct io_kiocb *req) |
7cbf1722 | 1361 | { |
c34398a8 | 1362 | struct io_uring_task *tctx = req->task->io_uring; |
9f010507 | 1363 | struct io_ring_ctx *ctx = req->ctx; |
7cbf1722 | 1364 | |
7cbf1722 | 1365 | /* task_work already pending, we're done */ |
32d91f05 | 1366 | if (!llist_add(&req->io_task_work.node, &tctx->task_list)) |
e09ee510 | 1367 | return; |
7cbf1722 | 1368 | |
ef060ea9 JA |
1369 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1370 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1371 | ||
3fe07bcd | 1372 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1373 | return; |
2215bed9 | 1374 | |
dfbe5561 | 1375 | io_fallback_tw(tctx, false); |
c0e0d6ba DY |
1376 | } |
1377 | ||
91c7884a PB |
1378 | void __io_req_task_work_add(struct io_kiocb *req, unsigned flags) |
1379 | { | |
1380 | if (req->ctx->flags & IORING_SETUP_DEFER_TASKRUN) { | |
1381 | rcu_read_lock(); | |
1382 | io_req_local_work_add(req, flags); | |
1383 | rcu_read_unlock(); | |
1384 | } else { | |
1385 | io_req_normal_work_add(req); | |
1386 | } | |
1387 | } | |
1388 | ||
c0e0d6ba DY |
1389 | static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx) |
1390 | { | |
1391 | struct llist_node *node; | |
1392 | ||
1393 | node = llist_del_all(&ctx->work_llist); | |
1394 | while (node) { | |
1395 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1396 | io_task_work.node); | |
1397 | ||
1398 | node = node->next; | |
91c7884a | 1399 | io_req_normal_work_add(req); |
c0e0d6ba DY |
1400 | } |
1401 | } | |
1402 | ||
a282967c | 1403 | static int __io_run_local_work(struct io_ring_ctx *ctx, struct io_tw_state *ts) |
c0e0d6ba | 1404 | { |
c0e0d6ba | 1405 | struct llist_node *node; |
c3f4d39e | 1406 | unsigned int loops = 0; |
140102ae | 1407 | int ret = 0; |
c0e0d6ba | 1408 | |
140102ae | 1409 | if (WARN_ON_ONCE(ctx->submitter_task != current)) |
c0e0d6ba | 1410 | return -EEXIST; |
c3f4d39e PB |
1411 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1412 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
c0e0d6ba | 1413 | again: |
3af0356c JA |
1414 | /* |
1415 | * llists are in reverse order, flip it back the right way before | |
1416 | * running the pending items. | |
1417 | */ | |
1418 | node = llist_reverse_order(io_llist_xchg(&ctx->work_llist, NULL)); | |
c3f4d39e | 1419 | while (node) { |
c0e0d6ba DY |
1420 | struct llist_node *next = node->next; |
1421 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1422 | io_task_work.node); | |
1423 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); | |
c92fcfc2 JA |
1424 | INDIRECT_CALL_2(req->io_task_work.func, |
1425 | io_poll_task_func, io_req_rw_complete, | |
1426 | req, ts); | |
c0e0d6ba DY |
1427 | ret++; |
1428 | node = next; | |
1429 | } | |
c3f4d39e | 1430 | loops++; |
c0e0d6ba | 1431 | |
c3f4d39e | 1432 | if (!llist_empty(&ctx->work_llist)) |
c0e0d6ba | 1433 | goto again; |
a282967c | 1434 | if (ts->locked) { |
c0e0d6ba | 1435 | io_submit_flush_completions(ctx); |
b0b7a7d2 PB |
1436 | if (!llist_empty(&ctx->work_llist)) |
1437 | goto again; | |
1438 | } | |
f75d5036 | 1439 | trace_io_uring_local_work_run(ctx, ret, loops); |
c0e0d6ba | 1440 | return ret; |
8ac5d85a JA |
1441 | } |
1442 | ||
360173ab PB |
1443 | static inline int io_run_local_work_locked(struct io_ring_ctx *ctx) |
1444 | { | |
a282967c | 1445 | struct io_tw_state ts = { .locked = true, }; |
360173ab PB |
1446 | int ret; |
1447 | ||
1448 | if (llist_empty(&ctx->work_llist)) | |
1449 | return 0; | |
1450 | ||
a282967c | 1451 | ret = __io_run_local_work(ctx, &ts); |
360173ab | 1452 | /* shouldn't happen! */ |
a282967c | 1453 | if (WARN_ON_ONCE(!ts.locked)) |
360173ab PB |
1454 | mutex_lock(&ctx->uring_lock); |
1455 | return ret; | |
1456 | } | |
1457 | ||
3e565555 | 1458 | static int io_run_local_work(struct io_ring_ctx *ctx) |
8ac5d85a | 1459 | { |
a282967c | 1460 | struct io_tw_state ts = {}; |
8ac5d85a JA |
1461 | int ret; |
1462 | ||
a282967c PB |
1463 | ts.locked = mutex_trylock(&ctx->uring_lock); |
1464 | ret = __io_run_local_work(ctx, &ts); | |
1465 | if (ts.locked) | |
8ac5d85a JA |
1466 | mutex_unlock(&ctx->uring_lock); |
1467 | ||
1468 | return ret; | |
c0e0d6ba DY |
1469 | } |
1470 | ||
a282967c | 1471 | static void io_req_task_cancel(struct io_kiocb *req, struct io_tw_state *ts) |
c40f6379 | 1472 | { |
a282967c | 1473 | io_tw_lock(req->ctx, ts); |
973fc83f | 1474 | io_req_defer_failed(req, req->cqe.res); |
c40f6379 JA |
1475 | } |
1476 | ||
a282967c | 1477 | void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts) |
c40f6379 | 1478 | { |
a282967c | 1479 | io_tw_lock(req->ctx, ts); |
316319e8 | 1480 | /* req->task == current here, checking PF_EXITING is safe */ |
6bb30855 | 1481 | if (unlikely(req->task->flags & PF_EXITING)) |
973fc83f | 1482 | io_req_defer_failed(req, -EFAULT); |
6bb30855 | 1483 | else if (req->flags & REQ_F_FORCE_ASYNC) |
a282967c | 1484 | io_queue_iowq(req, ts); |
6bb30855 DY |
1485 | else |
1486 | io_queue_sqe(req); | |
c40f6379 JA |
1487 | } |
1488 | ||
59915143 | 1489 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1490 | { |
97b388d7 | 1491 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1492 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1493 | io_req_task_work_add(req); |
c40f6379 JA |
1494 | } |
1495 | ||
f3b44f92 | 1496 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1497 | { |
5b0a6acc | 1498 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1499 | io_req_task_work_add(req); |
a3df7698 PB |
1500 | } |
1501 | ||
59915143 | 1502 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1503 | { |
9b5f7bd9 | 1504 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1505 | |
1506 | if (nxt) | |
906a8c3f | 1507 | io_req_task_queue(nxt); |
c69f8dbe JL |
1508 | } |
1509 | ||
f3b44f92 | 1510 | void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node) |
3aa83bfb | 1511 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1512 | { |
3aa83bfb PB |
1513 | do { |
1514 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1515 | comp_list); | |
2d6500d4 | 1516 | |
a538be5b PB |
1517 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1518 | if (req->flags & REQ_F_REFCOUNT) { | |
1519 | node = req->comp_list.next; | |
1520 | if (!req_ref_put_and_test(req)) | |
1521 | continue; | |
1522 | } | |
b605a7fa PB |
1523 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1524 | struct async_poll *apoll = req->apoll; | |
1525 | ||
1526 | if (apoll->double_poll) | |
1527 | kfree(apoll->double_poll); | |
9731bc98 JA |
1528 | if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) |
1529 | kfree(apoll); | |
b605a7fa PB |
1530 | req->flags &= ~REQ_F_POLLED; |
1531 | } | |
da1a08c5 | 1532 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1533 | io_queue_next(req); |
a538be5b PB |
1534 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1535 | io_clean_op(req); | |
c1e53a69 | 1536 | } |
a538be5b PB |
1537 | if (!(req->flags & REQ_F_FIXED_FILE)) |
1538 | io_put_file(req->file); | |
2d6500d4 | 1539 | |
ab409402 | 1540 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1541 | |
2fdd6fb5 | 1542 | io_put_task(req->task); |
c1e53a69 | 1543 | node = req->comp_list.next; |
fa05457a | 1544 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1545 | } while (node); |
7a743e22 PB |
1546 | } |
1547 | ||
c450178d | 1548 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1549 | __must_hold(&ctx->uring_lock) |
905c172f | 1550 | { |
cd0ca2e0 | 1551 | struct io_submit_state *state = &ctx->submit_state; |
fa780334 | 1552 | struct io_wq_work_node *node; |
905c172f | 1553 | |
f66f7342 | 1554 | __io_cq_lock(ctx); |
931147dd DY |
1555 | /* must come first to preserve CQE ordering in failure cases */ |
1556 | if (state->cqes_count) | |
1557 | __io_flush_post_cqes(ctx); | |
fa780334 | 1558 | __wq_list_for_each(node, &state->compl_reqs) { |
d9dee430 PB |
1559 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
1560 | comp_list); | |
3d4aeb9f | 1561 | |
f66f7342 PB |
1562 | if (!(req->flags & REQ_F_CQE_SKIP) && |
1563 | unlikely(!__io_fill_cqe_req(ctx, req))) { | |
1564 | if (ctx->task_complete) { | |
1565 | spin_lock(&ctx->completion_lock); | |
1566 | io_req_cqe_overflow(req); | |
1567 | spin_unlock(&ctx->completion_lock); | |
1568 | } else { | |
1569 | io_req_cqe_overflow(req); | |
1570 | } | |
1571 | } | |
905c172f | 1572 | } |
c98c81a4 | 1573 | __io_cq_unlock_post(ctx); |
d9dee430 | 1574 | |
931147dd DY |
1575 | if (!wq_list_empty(&ctx->submit_state.compl_reqs)) { |
1576 | io_free_batch_list(ctx, state->compl_reqs.first); | |
1577 | INIT_WQ_LIST(&state->compl_reqs); | |
1578 | } | |
7a743e22 PB |
1579 | } |
1580 | ||
6c503150 | 1581 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1582 | { |
1583 | /* See comment at the top of this file */ | |
1584 | smp_rmb(); | |
e23de15f | 1585 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1586 | } |
1587 | ||
def596e9 JA |
1588 | /* |
1589 | * We can't just wait for polled events to come to us, we have to actively | |
1590 | * find and complete them. | |
1591 | */ | |
c072481d | 1592 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1593 | { |
1594 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1595 | return; | |
1596 | ||
1597 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1598 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1599 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1600 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1601 | break; |
08f5439f JA |
1602 | /* |
1603 | * Ensure we allow local-to-the-cpu processing to take place, | |
1604 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1605 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1606 | */ |
3fcee5a6 PB |
1607 | if (need_resched()) { |
1608 | mutex_unlock(&ctx->uring_lock); | |
1609 | cond_resched(); | |
1610 | mutex_lock(&ctx->uring_lock); | |
1611 | } | |
def596e9 JA |
1612 | } |
1613 | mutex_unlock(&ctx->uring_lock); | |
1614 | } | |
1615 | ||
7668b92a | 1616 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1617 | { |
7668b92a | 1618 | unsigned int nr_events = 0; |
e9979b36 | 1619 | int ret = 0; |
155bc950 | 1620 | unsigned long check_cq; |
500f9fba | 1621 | |
76de6749 PB |
1622 | if (!io_allowed_run_tw(ctx)) |
1623 | return -EEXIST; | |
1624 | ||
3a08576b PB |
1625 | check_cq = READ_ONCE(ctx->check_cq); |
1626 | if (unlikely(check_cq)) { | |
1627 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
a85381d8 | 1628 | __io_cqring_overflow_flush(ctx); |
3a08576b PB |
1629 | /* |
1630 | * Similarly do not spin if we have not informed the user of any | |
1631 | * dropped CQE. | |
1632 | */ | |
1633 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1634 | return -EBADR; | |
1635 | } | |
f39c8a5b PB |
1636 | /* |
1637 | * Don't enter poll loop if we already have events pending. | |
1638 | * If we do, we can potentially be spinning for commands that | |
1639 | * already triggered a CQE (eg in error). | |
1640 | */ | |
f39c8a5b | 1641 | if (io_cqring_events(ctx)) |
d487b43c | 1642 | return 0; |
155bc950 | 1643 | |
def596e9 | 1644 | do { |
500f9fba JA |
1645 | /* |
1646 | * If a submit got punted to a workqueue, we can have the | |
1647 | * application entering polling for a command before it gets | |
1648 | * issued. That app will hold the uring_lock for the duration | |
1649 | * of the poll right here, so we need to take a breather every | |
1650 | * now and then to ensure that the issue has a chance to add | |
1651 | * the poll to the issued list. Otherwise we can spin here | |
1652 | * forever, while the workqueue is stuck trying to acquire the | |
1653 | * very same mutex. | |
1654 | */ | |
dac6a0ea JA |
1655 | if (wq_list_empty(&ctx->iopoll_list) || |
1656 | io_task_work_pending(ctx)) { | |
8f487ef2 PB |
1657 | u32 tail = ctx->cached_cq_tail; |
1658 | ||
8de11cdc | 1659 | (void) io_run_local_work_locked(ctx); |
def596e9 | 1660 | |
dac6a0ea JA |
1661 | if (task_work_pending(current) || |
1662 | wq_list_empty(&ctx->iopoll_list)) { | |
dac6a0ea | 1663 | mutex_unlock(&ctx->uring_lock); |
9d54bd6a | 1664 | io_run_task_work(); |
dac6a0ea | 1665 | mutex_lock(&ctx->uring_lock); |
dac6a0ea | 1666 | } |
8f487ef2 PB |
1667 | /* some requests don't go through iopoll_list */ |
1668 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1669 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1670 | break; |
500f9fba | 1671 | } |
5ba3c874 PB |
1672 | ret = io_do_iopoll(ctx, !min); |
1673 | if (ret < 0) | |
1674 | break; | |
1675 | nr_events += ret; | |
1676 | ret = 0; | |
1677 | } while (nr_events < min && !need_resched()); | |
d487b43c | 1678 | |
def596e9 JA |
1679 | return ret; |
1680 | } | |
7012c815 | 1681 | |
a282967c | 1682 | void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts) |
8ef12efe | 1683 | { |
a282967c | 1684 | if (ts->locked) |
9da070b1 | 1685 | io_req_complete_defer(req); |
7012c815 | 1686 | else |
27f35fe9 | 1687 | io_req_complete_post(req, IO_URING_F_UNLOCKED); |
8ef12efe JA |
1688 | } |
1689 | ||
def596e9 JA |
1690 | /* |
1691 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1692 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1693 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1694 | * accessing the kiocb cookie. |
1695 | */ | |
9882131c | 1696 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1697 | { |
1698 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1699 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1700 | |
1701 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1702 | if (unlikely(needs_lock)) |
cb3d8972 | 1703 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1704 | |
1705 | /* | |
1706 | * Track whether we have multiple files in our lists. This will impact | |
1707 | * how we do polling eventually, not spinning if we're on potentially | |
1708 | * different devices. | |
1709 | */ | |
5eef4e87 | 1710 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1711 | ctx->poll_multi_queue = false; |
1712 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1713 | struct io_kiocb *list_req; |
1714 | ||
5eef4e87 PB |
1715 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1716 | comp_list); | |
30da1b45 | 1717 | if (list_req->file != req->file) |
915b3dde | 1718 | ctx->poll_multi_queue = true; |
def596e9 JA |
1719 | } |
1720 | ||
1721 | /* | |
1722 | * For fast devices, IO may have already completed. If it has, add | |
1723 | * it to the front so we find it first. | |
1724 | */ | |
65a6543d | 1725 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1726 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1727 | else |
5eef4e87 | 1728 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1729 | |
3b44b371 | 1730 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1731 | /* |
1732 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1733 | * in sq thread task context or in io worker task context. If | |
1734 | * current task context is sq thread, we don't need to check | |
1735 | * whether should wake up sq thread. | |
1736 | */ | |
1737 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1738 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1739 | wake_up(&ctx->sq_data->wait); | |
1740 | ||
1741 | mutex_unlock(&ctx->uring_lock); | |
1742 | } | |
def596e9 JA |
1743 | } |
1744 | ||
a4ad4f74 | 1745 | unsigned int io_file_get_flags(struct file *file) |
88459b50 | 1746 | { |
88459b50 | 1747 | unsigned int res = 0; |
af197f50 | 1748 | |
53cfd5ce | 1749 | if (S_ISREG(file_inode(file)->i_mode)) |
8487f083 | 1750 | res |= REQ_F_ISREG; |
b9a6c945 | 1751 | if ((file->f_flags & O_NONBLOCK) || (file->f_mode & FMODE_NOWAIT)) |
8487f083 | 1752 | res |= REQ_F_SUPPORT_NOWAIT; |
88459b50 | 1753 | return res; |
2b188cc1 JA |
1754 | } |
1755 | ||
99f15d8d | 1756 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1757 | { |
f30bd4d0 BL |
1758 | WARN_ON_ONCE(!io_cold_defs[req->opcode].async_size); |
1759 | req->async_data = kmalloc(io_cold_defs[req->opcode].async_size, GFP_KERNEL); | |
d886e185 PB |
1760 | if (req->async_data) { |
1761 | req->flags |= REQ_F_ASYNC_DATA; | |
1762 | return false; | |
1763 | } | |
1764 | return true; | |
3d9932a8 XW |
1765 | } |
1766 | ||
f3b44f92 | 1767 | int io_req_prep_async(struct io_kiocb *req) |
f67676d1 | 1768 | { |
f30bd4d0 | 1769 | const struct io_cold_def *cdef = &io_cold_defs[req->opcode]; |
a7dd2782 | 1770 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
0702e536 JA |
1771 | |
1772 | /* assign early for deferred execution for non-fixed file */ | |
54aa7f23 | 1773 | if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE) && !req->file) |
0702e536 | 1774 | req->file = io_file_get_normal(req, req->cqe.fd); |
f30bd4d0 | 1775 | if (!cdef->prep_async) |
0702e536 JA |
1776 | return 0; |
1777 | if (WARN_ON_ONCE(req_has_async_data(req))) | |
1778 | return -EFAULT; | |
f30bd4d0 | 1779 | if (!def->manual_alloc) { |
59169439 PB |
1780 | if (io_alloc_async_data(req)) |
1781 | return -EAGAIN; | |
1782 | } | |
f30bd4d0 | 1783 | return cdef->prep_async(req); |
bfe76559 PB |
1784 | } |
1785 | ||
9cf7c104 PB |
1786 | static u32 io_get_sequence(struct io_kiocb *req) |
1787 | { | |
a3dbdf54 | 1788 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1789 | struct io_kiocb *cur; |
9cf7c104 | 1790 | |
a3dbdf54 | 1791 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1792 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1793 | seq--; |
1794 | return seq; | |
9cf7c104 PB |
1795 | } |
1796 | ||
c072481d | 1797 | static __cold void io_drain_req(struct io_kiocb *req) |
e276ae34 | 1798 | __must_hold(&ctx->uring_lock) |
de0617e4 | 1799 | { |
a197f664 | 1800 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1801 | struct io_defer_entry *de; |
f67676d1 | 1802 | int ret; |
e0eb71dc | 1803 | u32 seq = io_get_sequence(req); |
3c19966d | 1804 | |
9d858b21 | 1805 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1806 | spin_lock(&ctx->completion_lock); |
5e371265 | 1807 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1808 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1809 | queue: |
10c66904 | 1810 | ctx->drain_active = false; |
e0eb71dc PB |
1811 | io_req_task_queue(req); |
1812 | return; | |
10c66904 | 1813 | } |
e302f104 | 1814 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1815 | |
cbdcb435 | 1816 | io_prep_async_link(req); |
27dc8338 | 1817 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1818 | if (!de) { |
1b48773f | 1819 | ret = -ENOMEM; |
ef5c600a DY |
1820 | io_req_defer_failed(req, ret); |
1821 | return; | |
76cc33d7 | 1822 | } |
2d28390a | 1823 | |
79ebeaee | 1824 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1825 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1826 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1827 | kfree(de); |
e0eb71dc | 1828 | goto queue; |
de0617e4 JA |
1829 | } |
1830 | ||
48863ffd | 1831 | trace_io_uring_defer(req); |
27dc8338 | 1832 | de->req = req; |
9cf7c104 | 1833 | de->seq = seq; |
27dc8338 | 1834 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1835 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1836 | } |
1837 | ||
f4992544 JA |
1838 | static bool io_assign_file(struct io_kiocb *req, const struct io_issue_def *def, |
1839 | unsigned int issue_flags) | |
6bf9c47a | 1840 | { |
f4992544 | 1841 | if (req->file || !def->needs_file) |
6bf9c47a JA |
1842 | return true; |
1843 | ||
1844 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1845 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1846 | else |
cef216fc | 1847 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1848 | |
772f5e00 | 1849 | return !!req->file; |
6bf9c47a JA |
1850 | } |
1851 | ||
889fca73 | 1852 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1853 | { |
a7dd2782 | 1854 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
5730b27e | 1855 | const struct cred *creds = NULL; |
d625c6ee | 1856 | int ret; |
2b188cc1 | 1857 | |
f4992544 | 1858 | if (unlikely(!io_assign_file(req, def, issue_flags))) |
70152140 JA |
1859 | return -EBADF; |
1860 | ||
6878b40e | 1861 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1862 | creds = override_creds(req->creds); |
5730b27e | 1863 | |
fcde59fe | 1864 | if (!def->audit_skip) |
5bd2182d PM |
1865 | audit_uring_entry(req->opcode); |
1866 | ||
0702e536 | 1867 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1868 | |
fcde59fe | 1869 | if (!def->audit_skip) |
5bd2182d PM |
1870 | audit_uring_exit(!ret, ret); |
1871 | ||
5730b27e JA |
1872 | if (creds) |
1873 | revert_creds(creds); | |
97b388d7 | 1874 | |
75d7b3ae PB |
1875 | if (ret == IOU_OK) { |
1876 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
9da070b1 | 1877 | io_req_complete_defer(req); |
75d7b3ae | 1878 | else |
1bec951c | 1879 | io_req_complete_post(req, issue_flags); |
75d7b3ae | 1880 | } else if (ret != IOU_ISSUE_SKIP_COMPLETE) |
def596e9 | 1881 | return ret; |
97b388d7 | 1882 | |
b532576e | 1883 | /* If the op doesn't have a file, we're not polling for it */ |
ef0ec1ad | 1884 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue) |
9882131c | 1885 | io_iopoll_req_issued(req, issue_flags); |
def596e9 JA |
1886 | |
1887 | return 0; | |
2b188cc1 JA |
1888 | } |
1889 | ||
a282967c | 1890 | int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts) |
329061d3 | 1891 | { |
a282967c | 1892 | io_tw_lock(req->ctx, ts); |
9a692451 DY |
1893 | return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT| |
1894 | IO_URING_F_COMPLETE_DEFER); | |
329061d3 JA |
1895 | } |
1896 | ||
c9f06aa7 | 1897 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1898 | { |
1899 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
247f97a5 | 1900 | struct io_kiocb *nxt = NULL; |
ebc11b6c | 1901 | |
247f97a5 PB |
1902 | if (req_ref_put_and_test(req)) { |
1903 | if (req->flags & IO_REQ_LINK_FLAGS) | |
1904 | nxt = io_req_find_next(req); | |
1905 | io_free_req(req); | |
1906 | } | |
1907 | return nxt ? &nxt->work : NULL; | |
ebc11b6c PB |
1908 | } |
1909 | ||
c9f06aa7 | 1910 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1911 | { |
1912 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
a7dd2782 | 1913 | const struct io_issue_def *def = &io_issue_defs[req->opcode]; |
e6aeb272 | 1914 | unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ; |
d01905db | 1915 | bool needs_poll = false; |
6bf9c47a | 1916 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1917 | |
23a6c9ac | 1918 | /* one will be dropped by ->io_wq_free_work() after returning to io-wq */ |
48dcd38d PB |
1919 | if (!(req->flags & REQ_F_REFCOUNT)) |
1920 | __io_req_set_refcount(req, 2); | |
1921 | else | |
1922 | req_ref_get(req); | |
5d5901a3 | 1923 | |
cb2d344c | 1924 | io_arm_ltimeout(req); |
6bf9c47a | 1925 | |
dadebc35 | 1926 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1927 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1928 | fail: |
6bf9c47a | 1929 | io_req_task_queue_fail(req, err); |
d01905db PB |
1930 | return; |
1931 | } | |
f4992544 | 1932 | if (!io_assign_file(req, def, issue_flags)) { |
0f8da75b PB |
1933 | err = -EBADF; |
1934 | work->flags |= IO_WQ_WORK_CANCEL; | |
1935 | goto fail; | |
1936 | } | |
31b51510 | 1937 | |
d01905db | 1938 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
1939 | bool opcode_poll = def->pollin || def->pollout; |
1940 | ||
1941 | if (opcode_poll && file_can_poll(req->file)) { | |
1942 | needs_poll = true; | |
d01905db | 1943 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 1944 | } |
561fb04a | 1945 | } |
31b51510 | 1946 | |
d01905db PB |
1947 | do { |
1948 | ret = io_issue_sqe(req, issue_flags); | |
1949 | if (ret != -EAGAIN) | |
1950 | break; | |
1951 | /* | |
1952 | * We can get EAGAIN for iopolled IO even though we're | |
1953 | * forcing a sync submission from here, since we can't | |
1954 | * wait for request slots on the block side. | |
1955 | */ | |
1956 | if (!needs_poll) { | |
e0deb6a0 PB |
1957 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
1958 | break; | |
d01905db PB |
1959 | cond_resched(); |
1960 | continue; | |
90fa0288 HX |
1961 | } |
1962 | ||
4d9237e3 | 1963 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
1964 | return; |
1965 | /* aborted or ready, in either case retry blocking */ | |
1966 | needs_poll = false; | |
1967 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
1968 | } while (1); | |
31b51510 | 1969 | |
a3df7698 | 1970 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 1971 | if (ret < 0) |
a3df7698 | 1972 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
1973 | } |
1974 | ||
531113bb JA |
1975 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
1976 | unsigned int issue_flags) | |
09bb8394 | 1977 | { |
5106dd6e | 1978 | struct io_ring_ctx *ctx = req->ctx; |
4bfb0c9a | 1979 | struct io_fixed_file *slot; |
5106dd6e | 1980 | struct file *file = NULL; |
09bb8394 | 1981 | |
93f052cb | 1982 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 1983 | |
ac177053 | 1984 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 1985 | goto out; |
ac177053 | 1986 | fd = array_index_nospec(fd, ctx->nr_user_files); |
4bfb0c9a CH |
1987 | slot = io_fixed_file_slot(&ctx->file_table, fd); |
1988 | file = io_slot_file(slot); | |
1989 | req->flags |= io_slot_flags(slot); | |
5106dd6e JA |
1990 | io_req_set_rsrc_node(req, ctx, 0); |
1991 | out: | |
93f052cb | 1992 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
1993 | return file; |
1994 | } | |
d44f554e | 1995 | |
531113bb | 1996 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 1997 | { |
62906e89 | 1998 | struct file *file = fget(fd); |
ac177053 | 1999 | |
48863ffd | 2000 | trace_io_uring_file_get(req, fd); |
09bb8394 | 2001 | |
ac177053 | 2002 | /* we don't allow fixed io_uring files */ |
e5550a14 | 2003 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 2004 | io_req_track_inflight(req); |
8371adf5 | 2005 | return file; |
09bb8394 JA |
2006 | } |
2007 | ||
7bfa9bad | 2008 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
2009 | __must_hold(&req->ctx->uring_lock) |
2010 | { | |
7bfa9bad PB |
2011 | struct io_kiocb *linked_timeout; |
2012 | ||
2013 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
973fc83f | 2014 | io_req_defer_failed(req, ret); |
7bfa9bad PB |
2015 | return; |
2016 | } | |
2017 | ||
2018 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 2019 | |
4d9237e3 | 2020 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 2021 | case IO_APOLL_READY: |
336d28a8 | 2022 | io_kbuf_recycle(req, 0); |
d475a9a6 PB |
2023 | io_req_task_queue(req); |
2024 | break; | |
2025 | case IO_APOLL_ABORTED: | |
6436c770 | 2026 | io_kbuf_recycle(req, 0); |
77955efb | 2027 | io_queue_iowq(req, NULL); |
d475a9a6 | 2028 | break; |
b1c62645 | 2029 | case IO_APOLL_OK: |
b1c62645 | 2030 | break; |
d475a9a6 PB |
2031 | } |
2032 | ||
2033 | if (linked_timeout) | |
2034 | io_queue_linked_timeout(linked_timeout); | |
2035 | } | |
2036 | ||
cbc2e203 | 2037 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 2038 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 2039 | { |
e0c5c576 | 2040 | int ret; |
2b188cc1 | 2041 | |
c5eef2b9 | 2042 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 2043 | |
491381ce JA |
2044 | /* |
2045 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
2046 | * doesn't support non-blocking read/write attempts | |
2047 | */ | |
7bfa9bad | 2048 | if (likely(!ret)) |
cb2d344c | 2049 | io_arm_ltimeout(req); |
7bfa9bad PB |
2050 | else |
2051 | io_queue_async(req, ret); | |
2b188cc1 JA |
2052 | } |
2053 | ||
4652fe3f | 2054 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 2055 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 2056 | { |
17b147f6 PB |
2057 | if (unlikely(req->flags & REQ_F_FAIL)) { |
2058 | /* | |
2059 | * We don't submit, fail them all, for that replace hardlinks | |
2060 | * with normal links. Extra REQ_F_LINK is tolerated. | |
2061 | */ | |
2062 | req->flags &= ~REQ_F_HARDLINK; | |
2063 | req->flags |= REQ_F_LINK; | |
973fc83f | 2064 | io_req_defer_failed(req, req->cqe.res); |
76cc33d7 PB |
2065 | } else { |
2066 | int ret = io_req_prep_async(req); | |
2067 | ||
ef5c600a | 2068 | if (unlikely(ret)) { |
973fc83f | 2069 | io_req_defer_failed(req, ret); |
ef5c600a DY |
2070 | return; |
2071 | } | |
2072 | ||
2073 | if (unlikely(req->ctx->drain_active)) | |
2074 | io_drain_req(req); | |
76cc33d7 | 2075 | else |
77955efb | 2076 | io_queue_iowq(req, NULL); |
ce35a47a | 2077 | } |
4fe2c963 JL |
2078 | } |
2079 | ||
b16fed66 PB |
2080 | /* |
2081 | * Check SQE restrictions (opcode and flags). | |
2082 | * | |
2083 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
2084 | */ | |
2085 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
2086 | struct io_kiocb *req, | |
2087 | unsigned int sqe_flags) | |
4fe2c963 | 2088 | { |
b16fed66 PB |
2089 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
2090 | return false; | |
2091 | ||
2092 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
2093 | ctx->restrictions.sqe_flags_required) | |
2094 | return false; | |
2095 | ||
2096 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
2097 | ctx->restrictions.sqe_flags_required)) | |
2098 | return false; | |
2099 | ||
2100 | return true; | |
4fe2c963 JL |
2101 | } |
2102 | ||
22b2ca31 PB |
2103 | static void io_init_req_drain(struct io_kiocb *req) |
2104 | { | |
2105 | struct io_ring_ctx *ctx = req->ctx; | |
2106 | struct io_kiocb *head = ctx->submit_state.link.head; | |
2107 | ||
2108 | ctx->drain_active = true; | |
2109 | if (head) { | |
2110 | /* | |
2111 | * If we need to drain a request in the middle of a link, drain | |
2112 | * the head request and the next request/link after the current | |
2113 | * link. Considering sequential execution of links, | |
b6c7db32 | 2114 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
2115 | * link. |
2116 | */ | |
b6c7db32 | 2117 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
2118 | ctx->drain_next = true; |
2119 | } | |
2120 | } | |
2121 | ||
b16fed66 PB |
2122 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
2123 | const struct io_uring_sqe *sqe) | |
282cdc86 | 2124 | __must_hold(&ctx->uring_lock) |
b16fed66 | 2125 | { |
a7dd2782 | 2126 | const struct io_issue_def *def; |
b16fed66 | 2127 | unsigned int sqe_flags; |
fc0ae024 | 2128 | int personality; |
4a04d1d1 | 2129 | u8 opcode; |
b16fed66 | 2130 | |
864ea921 | 2131 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 2132 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 PB |
2133 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
2134 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
cef216fc | 2135 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 2136 | req->file = NULL; |
c1bdf8ed | 2137 | req->rsrc_node = NULL; |
b16fed66 | 2138 | req->task = current; |
b16fed66 | 2139 | |
4a04d1d1 PB |
2140 | if (unlikely(opcode >= IORING_OP_LAST)) { |
2141 | req->opcode = 0; | |
b16fed66 | 2142 | return -EINVAL; |
4a04d1d1 | 2143 | } |
a7dd2782 | 2144 | def = &io_issue_defs[opcode]; |
68fe256a PB |
2145 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
2146 | /* enforce forwards compatibility on users */ | |
2147 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
2148 | return -EINVAL; | |
4e906702 | 2149 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 2150 | if (!def->buffer_select) |
4e906702 JA |
2151 | return -EOPNOTSUPP; |
2152 | req->buf_index = READ_ONCE(sqe->buf_group); | |
2153 | } | |
5562a8d7 PB |
2154 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
2155 | ctx->drain_disabled = true; | |
2156 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
2157 | if (ctx->drain_disabled) | |
2158 | return -EOPNOTSUPP; | |
22b2ca31 | 2159 | io_init_req_drain(req); |
5562a8d7 | 2160 | } |
2a56a9bd PB |
2161 | } |
2162 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
2163 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
2164 | return -EACCES; | |
2165 | /* knock it to the slow queue path, will be drained there */ | |
2166 | if (ctx->drain_active) | |
2167 | req->flags |= REQ_F_FORCE_ASYNC; | |
2168 | /* if there is no link, we're at "next" request and need to drain */ | |
2169 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
2170 | ctx->drain_next = false; | |
2171 | ctx->drain_active = true; | |
b6c7db32 | 2172 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 2173 | } |
68fe256a | 2174 | } |
b16fed66 | 2175 | |
fcde59fe | 2176 | if (!def->ioprio && sqe->ioprio) |
73911426 | 2177 | return -EINVAL; |
fcde59fe | 2178 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
73911426 JA |
2179 | return -EINVAL; |
2180 | ||
fcde59fe | 2181 | if (def->needs_file) { |
6d63416d PB |
2182 | struct io_submit_state *state = &ctx->submit_state; |
2183 | ||
cef216fc | 2184 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 2185 | |
6d63416d PB |
2186 | /* |
2187 | * Plug now if we have more than 2 IO left after this, and the | |
2188 | * target is potentially a read/write to block based storage. | |
2189 | */ | |
fcde59fe | 2190 | if (state->need_plug && def->plug) { |
6d63416d PB |
2191 | state->plug_started = true; |
2192 | state->need_plug = false; | |
5ca7a8b3 | 2193 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 2194 | } |
b16fed66 | 2195 | } |
863e0560 | 2196 | |
003e8dcc JA |
2197 | personality = READ_ONCE(sqe->personality); |
2198 | if (personality) { | |
cdab10bf LT |
2199 | int ret; |
2200 | ||
c10d1f98 PB |
2201 | req->creds = xa_load(&ctx->personalities, personality); |
2202 | if (!req->creds) | |
003e8dcc | 2203 | return -EINVAL; |
c10d1f98 | 2204 | get_cred(req->creds); |
cdc1404a PM |
2205 | ret = security_uring_override_creds(req->creds); |
2206 | if (ret) { | |
2207 | put_cred(req->creds); | |
2208 | return ret; | |
2209 | } | |
b8e64b53 | 2210 | req->flags |= REQ_F_CREDS; |
003e8dcc | 2211 | } |
b16fed66 | 2212 | |
0702e536 | 2213 | return def->prep(req, sqe); |
b16fed66 PB |
2214 | } |
2215 | ||
df3becde PB |
2216 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
2217 | struct io_kiocb *req, int ret) | |
2218 | { | |
2219 | struct io_ring_ctx *ctx = req->ctx; | |
2220 | struct io_submit_link *link = &ctx->submit_state.link; | |
2221 | struct io_kiocb *head = link->head; | |
2222 | ||
48863ffd | 2223 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
2224 | |
2225 | /* | |
2226 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
2227 | * unusable. Instead of failing eagerly, continue assembling the link if | |
2228 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
2229 | * should find the flag and handle the rest. | |
2230 | */ | |
2231 | req_fail_link_node(req, ret); | |
2232 | if (head && !(head->flags & REQ_F_FAIL)) | |
2233 | req_fail_link_node(head, -ECANCELED); | |
2234 | ||
2235 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
2236 | if (head) { | |
2237 | link->last->link = req; | |
2238 | link->head = NULL; | |
2239 | req = head; | |
2240 | } | |
2241 | io_queue_sqe_fallback(req); | |
2242 | return ret; | |
2243 | } | |
2244 | ||
2245 | if (head) | |
2246 | link->last->link = req; | |
2247 | else | |
2248 | link->head = req; | |
2249 | link->last = req; | |
2250 | return 0; | |
2251 | } | |
2252 | ||
2253 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 2254 | const struct io_uring_sqe *sqe) |
282cdc86 | 2255 | __must_hold(&ctx->uring_lock) |
9e645e11 | 2256 | { |
a1ab7b35 | 2257 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 2258 | int ret; |
9e645e11 | 2259 | |
a6b8cadc | 2260 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
2261 | if (unlikely(ret)) |
2262 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 2263 | |
2ad57931 | 2264 | trace_io_uring_submit_req(req); |
a6b8cadc | 2265 | |
9e645e11 JA |
2266 | /* |
2267 | * If we already have a head request, queue this one for async | |
2268 | * submittal once the head completes. If we don't have a head but | |
2269 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2270 | * submitted sync once the chain is complete. If none of those | |
2271 | * conditions are true (normal request), then just queue it. | |
2272 | */ | |
924a07e4 | 2273 | if (unlikely(link->head)) { |
df3becde PB |
2274 | ret = io_req_prep_async(req); |
2275 | if (unlikely(ret)) | |
2276 | return io_submit_fail_init(sqe, req, ret); | |
2277 | ||
48863ffd | 2278 | trace_io_uring_link(req, link->head); |
f2f87370 | 2279 | link->last->link = req; |
863e0560 | 2280 | link->last = req; |
32fe525b | 2281 | |
da1a08c5 | 2282 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 2283 | return 0; |
df3becde PB |
2284 | /* last request of the link, flush it */ |
2285 | req = link->head; | |
f15a3431 | 2286 | link->head = NULL; |
924a07e4 PB |
2287 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
2288 | goto fallback; | |
2289 | ||
2290 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2291 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2292 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2293 | link->head = req; | |
2294 | link->last = req; | |
2295 | } else { | |
2296 | fallback: | |
2297 | io_queue_sqe_fallback(req); | |
2298 | } | |
f15a3431 | 2299 | return 0; |
9e645e11 | 2300 | } |
2e6e1fde | 2301 | |
f15a3431 | 2302 | io_queue_sqe(req); |
1d4240cc | 2303 | return 0; |
9e645e11 JA |
2304 | } |
2305 | ||
9a56a232 JA |
2306 | /* |
2307 | * Batched submission is done, ensure local IO is flushed out. | |
2308 | */ | |
553deffd | 2309 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2310 | { |
553deffd PB |
2311 | struct io_submit_state *state = &ctx->submit_state; |
2312 | ||
e126391c PB |
2313 | if (unlikely(state->link.head)) |
2314 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2315 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2316 | io_submit_flush_completions(ctx); |
27926b68 JA |
2317 | if (state->plug_started) |
2318 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2319 | } |
2320 | ||
2321 | /* | |
2322 | * Start submission side cache. | |
2323 | */ | |
2324 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2325 | unsigned int max_ios) |
9a56a232 | 2326 | { |
27926b68 | 2327 | state->plug_started = false; |
4b628aeb | 2328 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2329 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2330 | /* set only head, no need to init link_last in advance */ |
2331 | state->link.head = NULL; | |
9a56a232 JA |
2332 | } |
2333 | ||
2b188cc1 JA |
2334 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2335 | { | |
75b28aff | 2336 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2337 | |
caf582c6 PB |
2338 | /* |
2339 | * Ensure any loads from the SQEs are done at this point, | |
2340 | * since once we write the new head, the application could | |
2341 | * write new data to them. | |
2342 | */ | |
2343 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2344 | } |
2345 | ||
2b188cc1 | 2346 | /* |
dd9ae8a0 | 2347 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2348 | * that is mapped by userspace. This means that care needs to be taken to |
2349 | * ensure that reads are stable, as we cannot rely on userspace always | |
2350 | * being a good citizen. If members of the sqe are validated and then later | |
2351 | * used, it's important that those reads are done through READ_ONCE() to | |
2352 | * prevent a re-load down the line. | |
2353 | */ | |
b5083dfa | 2354 | static bool io_get_sqe(struct io_ring_ctx *ctx, const struct io_uring_sqe **sqe) |
2b188cc1 | 2355 | { |
ea5ab3b5 | 2356 | unsigned head, mask = ctx->sq_entries - 1; |
17d3aeb3 | 2357 | unsigned sq_idx = ctx->cached_sq_head++ & mask; |
2b188cc1 JA |
2358 | |
2359 | /* | |
2360 | * The cached sq head (or cq tail) serves two purposes: | |
2361 | * | |
2362 | * 1) allows us to batch the cost of updating the user visible | |
2363 | * head updates. | |
2364 | * 2) allows the kernel side to track the head on its own, even | |
2365 | * though the application is the one updating it. | |
2366 | */ | |
17d3aeb3 | 2367 | head = READ_ONCE(ctx->sq_array[sq_idx]); |
ebdeb7c0 JA |
2368 | if (likely(head < ctx->sq_entries)) { |
2369 | /* double index for 128-byte SQEs, twice as long */ | |
2370 | if (ctx->flags & IORING_SETUP_SQE128) | |
2371 | head <<= 1; | |
b5083dfa PB |
2372 | *sqe = &ctx->sq_sqes[head]; |
2373 | return true; | |
ebdeb7c0 | 2374 | } |
2b188cc1 JA |
2375 | |
2376 | /* drop invalid entries */ | |
15641e42 PB |
2377 | ctx->cq_extra--; |
2378 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2379 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
b5083dfa | 2380 | return false; |
709b302f PB |
2381 | } |
2382 | ||
17437f31 | 2383 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2384 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2385 | { |
69629809 | 2386 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2387 | unsigned int left; |
2388 | int ret; | |
6c271ce2 | 2389 | |
51d48dab | 2390 | if (unlikely(!entries)) |
69629809 | 2391 | return 0; |
ee7d46d9 | 2392 | /* make sure SQ entry isn't read before tail */ |
e3ef728f | 2393 | ret = left = min(nr, entries); |
8e6971a8 PB |
2394 | io_get_task_refs(left); |
2395 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2396 | |
69629809 | 2397 | do { |
3529d8c2 | 2398 | const struct io_uring_sqe *sqe; |
196be95c | 2399 | struct io_kiocb *req; |
fb5ccc98 | 2400 | |
c8576f3e | 2401 | if (unlikely(!io_alloc_req(ctx, &req))) |
fb5ccc98 | 2402 | break; |
b5083dfa | 2403 | if (unlikely(!io_get_sqe(ctx, &sqe))) { |
fa05457a | 2404 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2405 | break; |
2406 | } | |
6c271ce2 | 2407 | |
1cd15904 PB |
2408 | /* |
2409 | * Continue submitting even for sqe failure if the | |
2410 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2411 | */ | |
2412 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2413 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2414 | left--; | |
2415 | break; | |
bcbb7bf6 | 2416 | } |
1cd15904 | 2417 | } while (--left); |
9466f437 | 2418 | |
8e6971a8 PB |
2419 | if (unlikely(left)) { |
2420 | ret -= left; | |
2421 | /* try again if it submitted nothing and can't allocate a req */ | |
2422 | if (!ret && io_req_cache_empty(ctx)) | |
2423 | ret = -EAGAIN; | |
2424 | current->io_uring->cached_refs += left; | |
9466f437 | 2425 | } |
6c271ce2 | 2426 | |
553deffd | 2427 | io_submit_state_end(ctx); |
ae9428ca PB |
2428 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2429 | io_commit_sqring(ctx); | |
8e6971a8 | 2430 | return ret; |
6c271ce2 JA |
2431 | } |
2432 | ||
bda52162 JA |
2433 | struct io_wait_queue { |
2434 | struct wait_queue_entry wq; | |
2435 | struct io_ring_ctx *ctx; | |
5fd46178 | 2436 | unsigned cq_tail; |
bda52162 | 2437 | unsigned nr_timeouts; |
d33a39e5 | 2438 | ktime_t timeout; |
bda52162 JA |
2439 | }; |
2440 | ||
b4c98d59 DY |
2441 | static inline bool io_has_work(struct io_ring_ctx *ctx) |
2442 | { | |
c0e0d6ba | 2443 | return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) || |
490c00eb | 2444 | !llist_empty(&ctx->work_llist); |
b4c98d59 DY |
2445 | } |
2446 | ||
6c503150 | 2447 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
2448 | { |
2449 | struct io_ring_ctx *ctx = iowq->ctx; | |
0fc8c2ac | 2450 | int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail; |
bda52162 JA |
2451 | |
2452 | /* | |
d195a66e | 2453 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
2454 | * started waiting. For timeouts, we always want to return to userspace, |
2455 | * regardless of event count. | |
2456 | */ | |
5fd46178 | 2457 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
2458 | } |
2459 | ||
2460 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2461 | int wake_flags, void *key) | |
2462 | { | |
bd550173 | 2463 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, wq); |
bda52162 | 2464 | |
6c503150 PB |
2465 | /* |
2466 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2467 | * the task, and the next invocation will do it. | |
2468 | */ | |
bd550173 | 2469 | if (io_should_wake(iowq) || io_has_work(iowq->ctx)) |
6c503150 PB |
2470 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2471 | return -1; | |
bda52162 JA |
2472 | } |
2473 | ||
c0e0d6ba | 2474 | int io_run_task_work_sig(struct io_ring_ctx *ctx) |
af9c1a44 | 2475 | { |
1414d629 | 2476 | if (!llist_empty(&ctx->work_llist)) { |
2f413956 | 2477 | __set_current_state(TASK_RUNNING); |
1414d629 PB |
2478 | if (io_run_local_work(ctx) > 0) |
2479 | return 1; | |
2480 | } | |
2481 | if (io_run_task_work() > 0) | |
af9c1a44 | 2482 | return 1; |
c5020bc8 OL |
2483 | if (task_sigpending(current)) |
2484 | return -EINTR; | |
2485 | return 0; | |
af9c1a44 JA |
2486 | } |
2487 | ||
eeb60b9a PB |
2488 | /* when returns >0, the caller should retry */ |
2489 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
d33a39e5 | 2490 | struct io_wait_queue *iowq) |
eeb60b9a | 2491 | { |
3fcf19d5 PB |
2492 | if (unlikely(READ_ONCE(ctx->check_cq))) |
2493 | return 1; | |
846072f1 PB |
2494 | if (unlikely(!llist_empty(&ctx->work_llist))) |
2495 | return 1; | |
2496 | if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) | |
2497 | return 1; | |
2498 | if (unlikely(task_sigpending(current))) | |
2499 | return -EINTR; | |
2500 | if (unlikely(io_should_wake(iowq))) | |
2501 | return 0; | |
d33a39e5 | 2502 | if (iowq->timeout == KTIME_MAX) |
46ae7eef | 2503 | schedule(); |
d33a39e5 | 2504 | else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS)) |
22833966 | 2505 | return -ETIME; |
846072f1 | 2506 | return 0; |
eeb60b9a PB |
2507 | } |
2508 | ||
2b188cc1 JA |
2509 | /* |
2510 | * Wait until events become available, if we don't already have some. The | |
2511 | * application must reap them itself, as they reside on the shared cq ring. | |
2512 | */ | |
2513 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2514 | const sigset_t __user *sig, size_t sigsz, |
2515 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2516 | { |
90291099 | 2517 | struct io_wait_queue iowq; |
75b28aff | 2518 | struct io_rings *rings = ctx->rings; |
c1d5a224 | 2519 | int ret; |
2b188cc1 | 2520 | |
76de6749 PB |
2521 | if (!io_allowed_run_tw(ctx)) |
2522 | return -EEXIST; | |
140102ae PB |
2523 | if (!llist_empty(&ctx->work_llist)) |
2524 | io_run_local_work(ctx); | |
f36ba6cf PB |
2525 | io_run_task_work(); |
2526 | io_cqring_overflow_flush(ctx); | |
2527 | /* if user messes with these they will just get an early return */ | |
2528 | if (__io_cqring_events_user(ctx) >= min_events) | |
2529 | return 0; | |
2b188cc1 JA |
2530 | |
2531 | if (sig) { | |
9e75ad5d AB |
2532 | #ifdef CONFIG_COMPAT |
2533 | if (in_compat_syscall()) | |
2534 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2535 | sigsz); |
9e75ad5d AB |
2536 | else |
2537 | #endif | |
b772434b | 2538 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2539 | |
2b188cc1 JA |
2540 | if (ret) |
2541 | return ret; | |
2542 | } | |
2543 | ||
90291099 PB |
2544 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
2545 | iowq.wq.private = current; | |
2546 | INIT_LIST_HEAD(&iowq.wq.entry); | |
2547 | iowq.ctx = ctx; | |
bda52162 | 2548 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 2549 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
d33a39e5 PB |
2550 | iowq.timeout = KTIME_MAX; |
2551 | ||
2552 | if (uts) { | |
2553 | struct timespec64 ts; | |
2554 | ||
2555 | if (get_timespec64(&ts, uts)) | |
2556 | return -EFAULT; | |
2557 | iowq.timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); | |
2558 | } | |
90291099 | 2559 | |
c826bd7a | 2560 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2561 | do { |
3fcf19d5 PB |
2562 | unsigned long check_cq; |
2563 | ||
130bd686 | 2564 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { |
8751d154 PB |
2565 | int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail); |
2566 | ||
2567 | atomic_set(&ctx->cq_wait_nr, nr_wait); | |
130bd686 PB |
2568 | set_current_state(TASK_INTERRUPTIBLE); |
2569 | } else { | |
2570 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, | |
2571 | TASK_INTERRUPTIBLE); | |
2572 | } | |
2573 | ||
d33a39e5 | 2574 | ret = io_cqring_wait_schedule(ctx, &iowq); |
130bd686 | 2575 | __set_current_state(TASK_RUNNING); |
8751d154 | 2576 | atomic_set(&ctx->cq_wait_nr, 0); |
d80c0f00 | 2577 | |
846072f1 PB |
2578 | if (ret < 0) |
2579 | break; | |
2580 | /* | |
2581 | * Run task_work after scheduling and before io_should_wake(). | |
2582 | * If we got woken because of task_work being processed, run it | |
2583 | * now rather than let the caller do another wait loop. | |
2584 | */ | |
2585 | io_run_task_work(); | |
2586 | if (!llist_empty(&ctx->work_llist)) | |
2587 | io_run_local_work(ctx); | |
3fcf19d5 PB |
2588 | |
2589 | check_cq = READ_ONCE(ctx->check_cq); | |
2590 | if (unlikely(check_cq)) { | |
2591 | /* let the caller flush overflows, retry */ | |
326a9e48 | 2592 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) |
3fcf19d5 | 2593 | io_cqring_do_overflow_flush(ctx); |
3fcf19d5 PB |
2594 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) { |
2595 | ret = -EBADR; | |
2596 | break; | |
2597 | } | |
2598 | } | |
2599 | ||
846072f1 PB |
2600 | if (io_should_wake(&iowq)) { |
2601 | ret = 0; | |
35d90f95 | 2602 | break; |
846072f1 | 2603 | } |
ca0a2651 | 2604 | cond_resched(); |
846072f1 | 2605 | } while (1); |
bda52162 | 2606 | |
130bd686 PB |
2607 | if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) |
2608 | finish_wait(&ctx->cq_wait, &iowq.wq); | |
b7db41c9 | 2609 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2610 | |
75b28aff | 2611 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2612 | } |
2613 | ||
73572984 | 2614 | static void io_mem_free(void *ptr) |
b9bd2bea | 2615 | { |
73572984 | 2616 | struct page *page; |
b36a2050 | 2617 | |
73572984 JA |
2618 | if (!ptr) |
2619 | return; | |
b9bd2bea | 2620 | |
73572984 JA |
2621 | page = virt_to_head_page(ptr); |
2622 | if (put_page_testzero(page)) | |
2623 | free_compound_page(page); | |
b9bd2bea PB |
2624 | } |
2625 | ||
03d89a2d JA |
2626 | static void io_pages_free(struct page ***pages, int npages) |
2627 | { | |
2628 | struct page **page_array; | |
2629 | int i; | |
2630 | ||
2631 | if (!pages) | |
2632 | return; | |
2633 | page_array = *pages; | |
2634 | for (i = 0; i < npages; i++) | |
2635 | unpin_user_page(page_array[i]); | |
2636 | kvfree(page_array); | |
2637 | *pages = NULL; | |
2638 | } | |
2639 | ||
2640 | static void *__io_uaddr_map(struct page ***pages, unsigned short *npages, | |
2641 | unsigned long uaddr, size_t size) | |
2642 | { | |
2643 | struct page **page_array; | |
2644 | unsigned int nr_pages; | |
2645 | int ret; | |
2646 | ||
2647 | *npages = 0; | |
2648 | ||
2649 | if (uaddr & (PAGE_SIZE - 1) || !size) | |
2650 | return ERR_PTR(-EINVAL); | |
2651 | ||
2652 | nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2653 | if (nr_pages > USHRT_MAX) | |
2654 | return ERR_PTR(-EINVAL); | |
2655 | page_array = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); | |
2656 | if (!page_array) | |
2657 | return ERR_PTR(-ENOMEM); | |
2658 | ||
2659 | ret = pin_user_pages_fast(uaddr, nr_pages, FOLL_WRITE | FOLL_LONGTERM, | |
2660 | page_array); | |
2661 | if (ret != nr_pages) { | |
2662 | err: | |
2663 | io_pages_free(&page_array, ret > 0 ? ret : 0); | |
2664 | return ret < 0 ? ERR_PTR(ret) : ERR_PTR(-EFAULT); | |
2665 | } | |
2666 | /* | |
2667 | * Should be a single page. If the ring is small enough that we can | |
2668 | * use a normal page, that is fine. If we need multiple pages, then | |
2669 | * userspace should use a huge page. That's the only way to guarantee | |
2670 | * that we get contigious memory, outside of just being lucky or | |
2671 | * (currently) having low memory fragmentation. | |
2672 | */ | |
2673 | if (page_array[0] != page_array[ret - 1]) | |
2674 | goto err; | |
2675 | *pages = page_array; | |
2676 | *npages = nr_pages; | |
2677 | return page_to_virt(page_array[0]); | |
2678 | } | |
2679 | ||
2680 | static void *io_rings_map(struct io_ring_ctx *ctx, unsigned long uaddr, | |
2681 | size_t size) | |
2682 | { | |
2683 | return __io_uaddr_map(&ctx->ring_pages, &ctx->n_ring_pages, uaddr, | |
2684 | size); | |
2685 | } | |
2686 | ||
2687 | static void *io_sqes_map(struct io_ring_ctx *ctx, unsigned long uaddr, | |
2688 | size_t size) | |
2689 | { | |
2690 | return __io_uaddr_map(&ctx->sqe_pages, &ctx->n_sqe_pages, uaddr, | |
2691 | size); | |
2692 | } | |
2693 | ||
9c189eee JA |
2694 | static void io_rings_free(struct io_ring_ctx *ctx) |
2695 | { | |
03d89a2d JA |
2696 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) { |
2697 | io_mem_free(ctx->rings); | |
2698 | io_mem_free(ctx->sq_sqes); | |
2699 | ctx->rings = NULL; | |
2700 | ctx->sq_sqes = NULL; | |
2701 | } else { | |
2702 | io_pages_free(&ctx->ring_pages, ctx->n_ring_pages); | |
2703 | io_pages_free(&ctx->sqe_pages, ctx->n_sqe_pages); | |
2704 | } | |
9c189eee JA |
2705 | } |
2706 | ||
73572984 | 2707 | static void *io_mem_alloc(size_t size) |
b9bd2bea | 2708 | { |
73572984 | 2709 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; |
e27cef86 | 2710 | void *ret; |
b9bd2bea | 2711 | |
e27cef86 JA |
2712 | ret = (void *) __get_free_pages(gfp, get_order(size)); |
2713 | if (ret) | |
2714 | return ret; | |
2715 | return ERR_PTR(-ENOMEM); | |
b9bd2bea PB |
2716 | } |
2717 | ||
73572984 JA |
2718 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2719 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2720 | { |
73572984 JA |
2721 | struct io_rings *rings; |
2722 | size_t off, sq_array_size; | |
6b06314c | 2723 | |
73572984 JA |
2724 | off = struct_size(rings, cqes, cq_entries); |
2725 | if (off == SIZE_MAX) | |
2726 | return SIZE_MAX; | |
2727 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2728 | if (check_shl_overflow(off, 1, &off)) | |
2729 | return SIZE_MAX; | |
2730 | } | |
ab409402 | 2731 | |
73572984 JA |
2732 | #ifdef CONFIG_SMP |
2733 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2734 | if (off == 0) | |
2735 | return SIZE_MAX; | |
2736 | #endif | |
82fbcfa9 | 2737 | |
73572984 JA |
2738 | if (sq_offset) |
2739 | *sq_offset = off; | |
82fbcfa9 | 2740 | |
73572984 JA |
2741 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2742 | if (sq_array_size == SIZE_MAX) | |
2743 | return SIZE_MAX; | |
6b06314c | 2744 | |
73572984 JA |
2745 | if (check_add_overflow(off, sq_array_size, &off)) |
2746 | return SIZE_MAX; | |
8bad28d8 | 2747 | |
73572984 | 2748 | return off; |
8bad28d8 HX |
2749 | } |
2750 | ||
73572984 JA |
2751 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, |
2752 | unsigned int eventfd_async) | |
8bad28d8 | 2753 | { |
73572984 JA |
2754 | struct io_ev_fd *ev_fd; |
2755 | __s32 __user *fds = arg; | |
2756 | int fd; | |
f2303b1f | 2757 | |
73572984 JA |
2758 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, |
2759 | lockdep_is_held(&ctx->uring_lock)); | |
2760 | if (ev_fd) | |
2761 | return -EBUSY; | |
8bad28d8 | 2762 | |
73572984 JA |
2763 | if (copy_from_user(&fd, fds, sizeof(*fds))) |
2764 | return -EFAULT; | |
8dd03afe | 2765 | |
73572984 JA |
2766 | ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); |
2767 | if (!ev_fd) | |
2768 | return -ENOMEM; | |
05f3fb3c | 2769 | |
73572984 JA |
2770 | ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); |
2771 | if (IS_ERR(ev_fd->cq_ev_fd)) { | |
2772 | int ret = PTR_ERR(ev_fd->cq_ev_fd); | |
2773 | kfree(ev_fd); | |
2774 | return ret; | |
2775 | } | |
305bef98 PB |
2776 | |
2777 | spin_lock(&ctx->completion_lock); | |
2778 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
2779 | spin_unlock(&ctx->completion_lock); | |
2780 | ||
73572984 JA |
2781 | ev_fd->eventfd_async = eventfd_async; |
2782 | ctx->has_evfd = true; | |
2783 | rcu_assign_pointer(ctx->io_ev_fd, ev_fd); | |
21a091b9 DY |
2784 | atomic_set(&ev_fd->refs, 1); |
2785 | atomic_set(&ev_fd->ops, 0); | |
73572984 | 2786 | return 0; |
d7954b2b BM |
2787 | } |
2788 | ||
73572984 | 2789 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 2790 | { |
73572984 JA |
2791 | struct io_ev_fd *ev_fd; |
2792 | ||
2793 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, | |
2794 | lockdep_is_held(&ctx->uring_lock)); | |
2795 | if (ev_fd) { | |
2796 | ctx->has_evfd = false; | |
2797 | rcu_assign_pointer(ctx->io_ev_fd, NULL); | |
21a091b9 DY |
2798 | if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops)) |
2799 | call_rcu(&ev_fd->rcu, io_eventfd_ops); | |
73572984 JA |
2800 | return 0; |
2801 | } | |
2d091d62 | 2802 | |
73572984 | 2803 | return -ENXIO; |
44b31f2f PB |
2804 | } |
2805 | ||
73572984 | 2806 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2807 | { |
c8576f3e | 2808 | struct io_kiocb *req; |
37f0e767 | 2809 | int nr = 0; |
bf019da7 | 2810 | |
9a4fdbd8 | 2811 | mutex_lock(&ctx->uring_lock); |
34f0bc42 | 2812 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
9a4fdbd8 | 2813 | |
88ab95be | 2814 | while (!io_req_cache_empty(ctx)) { |
c8576f3e | 2815 | req = io_extract_req(ctx); |
c2b6c6bc | 2816 | kmem_cache_free(req_cachep, req); |
37f0e767 | 2817 | nr++; |
c2b6c6bc | 2818 | } |
37f0e767 PB |
2819 | if (nr) |
2820 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2821 | mutex_unlock(&ctx->uring_lock); |
2822 | } | |
2823 | ||
9eae8655 PB |
2824 | static void io_rsrc_node_cache_free(struct io_cache_entry *entry) |
2825 | { | |
2826 | kfree(container_of(entry, struct io_rsrc_node, cache)); | |
2827 | } | |
2828 | ||
c072481d | 2829 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2830 | { |
37d1e2e3 | 2831 | io_sq_thread_finish(ctx); |
43597aac | 2832 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
0b222eeb PB |
2833 | if (WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list))) |
2834 | return; | |
43597aac | 2835 | |
8bad28d8 | 2836 | mutex_lock(&ctx->uring_lock); |
43597aac | 2837 | if (ctx->buf_data) |
bd54b6fe | 2838 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2839 | if (ctx->file_data) |
08480400 | 2840 | __io_sqe_files_unregister(ctx); |
a85381d8 | 2841 | io_cqring_overflow_kill(ctx); |
9b402849 | 2842 | io_eventfd_unregister(ctx); |
9b797a37 | 2843 | io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); |
43e0bbbd | 2844 | io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); |
5a2e745d | 2845 | io_destroy_buffers(ctx); |
b4a72c05 | 2846 | mutex_unlock(&ctx->uring_lock); |
07db298a PB |
2847 | if (ctx->sq_creds) |
2848 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2849 | if (ctx->submitter_task) |
2850 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2851 | |
a7f0ed5a PB |
2852 | /* there are no registered resources left, nobody uses it */ |
2853 | if (ctx->rsrc_node) | |
9eae8655 | 2854 | io_rsrc_node_destroy(ctx, ctx->rsrc_node); |
a7f0ed5a PB |
2855 | |
2856 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
def596e9 | 2857 | |
2b188cc1 | 2858 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
2859 | if (ctx->ring_sock) { |
2860 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 2861 | sock_release(ctx->ring_sock); |
355e8d26 | 2862 | } |
2b188cc1 | 2863 | #endif |
ef9dd637 | 2864 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
2b188cc1 | 2865 | |
9eae8655 | 2866 | io_alloc_cache_free(&ctx->rsrc_node_cache, io_rsrc_node_cache_free); |
42b6419d PB |
2867 | if (ctx->mm_account) { |
2868 | mmdrop(ctx->mm_account); | |
2869 | ctx->mm_account = NULL; | |
2870 | } | |
9c189eee | 2871 | io_rings_free(ctx); |
2b188cc1 JA |
2872 | |
2873 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2874 | free_uid(ctx->user); |
4010fec4 | 2875 | io_req_caches_free(ctx); |
e941894e JA |
2876 | if (ctx->hash_map) |
2877 | io_wq_put_hash(ctx->hash_map); | |
e6f89be6 | 2878 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2879 | kfree(ctx->cancel_table_locked.hbs); |
6224843d | 2880 | kfree(ctx->dummy_ubuf); |
9cfc7e94 JA |
2881 | kfree(ctx->io_bl); |
2882 | xa_destroy(&ctx->io_bl_xa); | |
2b188cc1 JA |
2883 | kfree(ctx); |
2884 | } | |
2885 | ||
bca39f39 PB |
2886 | static __cold void io_activate_pollwq_cb(struct callback_head *cb) |
2887 | { | |
2888 | struct io_ring_ctx *ctx = container_of(cb, struct io_ring_ctx, | |
2889 | poll_wq_task_work); | |
2890 | ||
2891 | mutex_lock(&ctx->uring_lock); | |
2892 | ctx->poll_activated = true; | |
2893 | mutex_unlock(&ctx->uring_lock); | |
2894 | ||
2895 | /* | |
2896 | * Wake ups for some events between start of polling and activation | |
2897 | * might've been lost due to loose synchronisation. | |
2898 | */ | |
2899 | wake_up_all(&ctx->poll_wq); | |
2900 | percpu_ref_put(&ctx->refs); | |
2901 | } | |
2902 | ||
2903 | static __cold void io_activate_pollwq(struct io_ring_ctx *ctx) | |
2904 | { | |
2905 | spin_lock(&ctx->completion_lock); | |
2906 | /* already activated or in progress */ | |
2907 | if (ctx->poll_activated || ctx->poll_wq_task_work.func) | |
2908 | goto out; | |
2909 | if (WARN_ON_ONCE(!ctx->task_complete)) | |
2910 | goto out; | |
2911 | if (!ctx->submitter_task) | |
2912 | goto out; | |
2913 | /* | |
2914 | * with ->submitter_task only the submitter task completes requests, we | |
2915 | * only need to sync with it, which is done by injecting a tw | |
2916 | */ | |
2917 | init_task_work(&ctx->poll_wq_task_work, io_activate_pollwq_cb); | |
2918 | percpu_ref_get(&ctx->refs); | |
2919 | if (task_work_add(ctx->submitter_task, &ctx->poll_wq_task_work, TWA_SIGNAL)) | |
2920 | percpu_ref_put(&ctx->refs); | |
2921 | out: | |
2922 | spin_unlock(&ctx->completion_lock); | |
2923 | } | |
2924 | ||
2b188cc1 JA |
2925 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) |
2926 | { | |
2927 | struct io_ring_ctx *ctx = file->private_data; | |
2928 | __poll_t mask = 0; | |
2929 | ||
bca39f39 PB |
2930 | if (unlikely(!ctx->poll_activated)) |
2931 | io_activate_pollwq(ctx); | |
2932 | ||
7b235dd8 | 2933 | poll_wait(file, &ctx->poll_wq, wait); |
4f7067c3 SB |
2934 | /* |
2935 | * synchronizes with barrier from wq_has_sleeper call in | |
2936 | * io_commit_cqring | |
2937 | */ | |
2b188cc1 | 2938 | smp_rmb(); |
90554200 | 2939 | if (!io_sqring_full(ctx)) |
2b188cc1 | 2940 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
2941 | |
2942 | /* | |
2943 | * Don't flush cqring overflow list here, just do a simple check. | |
2944 | * Otherwise there could possible be ABBA deadlock: | |
2945 | * CPU0 CPU1 | |
2946 | * ---- ---- | |
2947 | * lock(&ctx->uring_lock); | |
2948 | * lock(&ep->mtx); | |
2949 | * lock(&ctx->uring_lock); | |
2950 | * lock(&ep->mtx); | |
2951 | * | |
2952 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
10d8bc35 | 2953 | * pushes them to do the flush. |
ed670c3f | 2954 | */ |
b4c98d59 | 2955 | |
c10bb646 | 2956 | if (__io_cqring_events_user(ctx) || io_has_work(ctx)) |
2b188cc1 JA |
2957 | mask |= EPOLLIN | EPOLLRDNORM; |
2958 | ||
2959 | return mask; | |
2960 | } | |
2961 | ||
0bead8cd | 2962 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 2963 | { |
4379bf8b | 2964 | const struct cred *creds; |
071698e1 | 2965 | |
61cf9370 | 2966 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
2967 | if (creds) { |
2968 | put_cred(creds); | |
0bead8cd | 2969 | return 0; |
1e6fa521 | 2970 | } |
0bead8cd YD |
2971 | |
2972 | return -EINVAL; | |
2973 | } | |
2974 | ||
d56d938b PB |
2975 | struct io_tctx_exit { |
2976 | struct callback_head task_work; | |
2977 | struct completion completion; | |
baf186c4 | 2978 | struct io_ring_ctx *ctx; |
d56d938b PB |
2979 | }; |
2980 | ||
c072481d | 2981 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
2982 | { |
2983 | struct io_uring_task *tctx = current->io_uring; | |
2984 | struct io_tctx_exit *work; | |
2985 | ||
2986 | work = container_of(cb, struct io_tctx_exit, task_work); | |
2987 | /* | |
8d664282 | 2988 | * When @in_cancel, we're in cancellation and it's racy to remove the |
d56d938b | 2989 | * node. It'll be removed by the end of cancellation, just ignore it. |
998b30c3 HM |
2990 | * tctx can be NULL if the queueing of this task_work raced with |
2991 | * work cancelation off the exec path. | |
d56d938b | 2992 | */ |
8d664282 | 2993 | if (tctx && !atomic_read(&tctx->in_cancel)) |
eef51daa | 2994 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
2995 | complete(&work->completion); |
2996 | } | |
2997 | ||
c072481d | 2998 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
2999 | { |
3000 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
3001 | ||
3002 | return req->ctx == data; | |
3003 | } | |
3004 | ||
c072481d | 3005 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 3006 | { |
d56d938b | 3007 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 3008 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 3009 | unsigned long interval = HZ / 20; |
d56d938b PB |
3010 | struct io_tctx_exit exit; |
3011 | struct io_tctx_node *node; | |
3012 | int ret; | |
85faa7b8 | 3013 | |
56952e91 JA |
3014 | /* |
3015 | * If we're doing polled IO and end up having requests being | |
3016 | * submitted async (out-of-line), then completions can come in while | |
3017 | * we're waiting for refs to drop. We need to reap these manually, | |
3018 | * as nobody else will be looking for them. | |
3019 | */ | |
b2edc0a7 | 3020 | do { |
a85381d8 PB |
3021 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
3022 | mutex_lock(&ctx->uring_lock); | |
3023 | io_cqring_overflow_kill(ctx); | |
3024 | mutex_unlock(&ctx->uring_lock); | |
3025 | } | |
3026 | ||
c0e0d6ba DY |
3027 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) |
3028 | io_move_task_work_from_local(ctx); | |
3029 | ||
affa87db PB |
3030 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
3031 | cond_resched(); | |
3032 | ||
28090c13 PB |
3033 | if (ctx->sq_data) { |
3034 | struct io_sq_data *sqd = ctx->sq_data; | |
3035 | struct task_struct *tsk; | |
3036 | ||
3037 | io_sq_thread_park(sqd); | |
3038 | tsk = sqd->thread; | |
3039 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
3040 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
3041 | io_cancel_ctx_cb, ctx, true); | |
3042 | io_sq_thread_unpark(sqd); | |
3043 | } | |
b5bb3a24 | 3044 | |
37f0e767 PB |
3045 | io_req_caches_free(ctx); |
3046 | ||
58d3be2c PB |
3047 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
3048 | /* there is little hope left, don't run it too often */ | |
3049 | interval = HZ * 60; | |
3050 | } | |
4826c594 JA |
3051 | /* |
3052 | * This is really an uninterruptible wait, as it has to be | |
3053 | * complete. But it's also run from a kworker, which doesn't | |
3054 | * take signals, so it's fine to make it interruptible. This | |
3055 | * avoids scenarios where we knowingly can wait much longer | |
3056 | * on completions, for example if someone does a SIGSTOP on | |
3057 | * a task that needs to finish task_work to make this loop | |
3058 | * complete. That's a synthetic situation that should not | |
3059 | * cause a stuck task backtrace, and hence a potential panic | |
3060 | * on stuck tasks if that is enabled. | |
3061 | */ | |
3062 | } while (!wait_for_completion_interruptible_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 3063 | |
7f00651a PB |
3064 | init_completion(&exit.completion); |
3065 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
3066 | exit.ctx = ctx; | |
89b5066e PB |
3067 | /* |
3068 | * Some may use context even when all refs and requests have been put, | |
3069 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 3070 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
3071 | * this lock/unlock section also waits them to finish. |
3072 | */ | |
d56d938b PB |
3073 | mutex_lock(&ctx->uring_lock); |
3074 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
3075 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
3076 | ||
d56d938b PB |
3077 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
3078 | ctx_node); | |
7f00651a PB |
3079 | /* don't spin on a single task if cancellation failed */ |
3080 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
3081 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
3082 | if (WARN_ON_ONCE(ret)) | |
3083 | continue; | |
d56d938b PB |
3084 | |
3085 | mutex_unlock(&ctx->uring_lock); | |
4826c594 JA |
3086 | /* |
3087 | * See comment above for | |
3088 | * wait_for_completion_interruptible_timeout() on why this | |
3089 | * wait is marked as interruptible. | |
3090 | */ | |
3091 | wait_for_completion_interruptible(&exit.completion); | |
d56d938b PB |
3092 | mutex_lock(&ctx->uring_lock); |
3093 | } | |
3094 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
3095 | spin_lock(&ctx->completion_lock); |
3096 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 3097 | |
d73a572d PB |
3098 | /* pairs with RCU read section in io_req_local_work_add() */ |
3099 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3100 | synchronize_rcu(); | |
3101 | ||
85faa7b8 JA |
3102 | io_ring_ctx_free(ctx); |
3103 | } | |
3104 | ||
c072481d | 3105 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 3106 | { |
61cf9370 MWO |
3107 | unsigned long index; |
3108 | struct creds *creds; | |
3109 | ||
2b188cc1 JA |
3110 | mutex_lock(&ctx->uring_lock); |
3111 | percpu_ref_kill(&ctx->refs); | |
61cf9370 MWO |
3112 | xa_for_each(&ctx->personalities, index, creds) |
3113 | io_unregister_personality(ctx, index); | |
9ca9fb24 PB |
3114 | if (ctx->rings) |
3115 | io_poll_remove_all(ctx, NULL, true); | |
2b188cc1 JA |
3116 | mutex_unlock(&ctx->uring_lock); |
3117 | ||
02bac94b PB |
3118 | /* |
3119 | * If we failed setting up the ctx, we might not have any rings | |
3120 | * and therefore did not submit any requests | |
3121 | */ | |
3122 | if (ctx->rings) | |
60053be8 | 3123 | io_kill_timeouts(ctx, NULL, true); |
309fc03a | 3124 | |
dfbe5561 JA |
3125 | flush_delayed_work(&ctx->fallback_work); |
3126 | ||
85faa7b8 | 3127 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
3128 | /* |
3129 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
3130 | * if we're exiting a ton of rings at the same time. It just adds | |
3131 | * noise and overhead, there's no discernable change in runtime | |
3132 | * over using system_wq. | |
3133 | */ | |
3134 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
3135 | } |
3136 | ||
3137 | static int io_uring_release(struct inode *inode, struct file *file) | |
3138 | { | |
3139 | struct io_ring_ctx *ctx = file->private_data; | |
3140 | ||
3141 | file->private_data = NULL; | |
3142 | io_ring_ctx_wait_and_kill(ctx); | |
3143 | return 0; | |
3144 | } | |
3145 | ||
f6edbabb PB |
3146 | struct io_task_cancel { |
3147 | struct task_struct *task; | |
3dd0c97a | 3148 | bool all; |
f6edbabb | 3149 | }; |
f254ac04 | 3150 | |
f6edbabb | 3151 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 3152 | { |
9a472ef7 | 3153 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 3154 | struct io_task_cancel *cancel = data; |
9a472ef7 | 3155 | |
6af3f48b | 3156 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
3157 | } |
3158 | ||
c072481d PB |
3159 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
3160 | struct task_struct *task, | |
3161 | bool cancel_all) | |
b7ddce3c | 3162 | { |
e1915f76 | 3163 | struct io_defer_entry *de; |
b7ddce3c PB |
3164 | LIST_HEAD(list); |
3165 | ||
79ebeaee | 3166 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 3167 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 3168 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
3169 | list_cut_position(&list, &ctx->defer_list, &de->list); |
3170 | break; | |
3171 | } | |
3172 | } | |
79ebeaee | 3173 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
3174 | if (list_empty(&list)) |
3175 | return false; | |
b7ddce3c PB |
3176 | |
3177 | while (!list_empty(&list)) { | |
3178 | de = list_first_entry(&list, struct io_defer_entry, list); | |
3179 | list_del_init(&de->list); | |
e276ae34 | 3180 | io_req_task_queue_fail(de->req, -ECANCELED); |
b7ddce3c PB |
3181 | kfree(de); |
3182 | } | |
e1915f76 | 3183 | return true; |
b7ddce3c PB |
3184 | } |
3185 | ||
c072481d | 3186 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
3187 | { |
3188 | struct io_tctx_node *node; | |
3189 | enum io_wq_cancel cret; | |
3190 | bool ret = false; | |
3191 | ||
3192 | mutex_lock(&ctx->uring_lock); | |
3193 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3194 | struct io_uring_task *tctx = node->task->io_uring; | |
3195 | ||
3196 | /* | |
3197 | * io_wq will stay alive while we hold uring_lock, because it's | |
3198 | * killed after ctx nodes, which requires to take the lock. | |
3199 | */ | |
3200 | if (!tctx || !tctx->io_wq) | |
3201 | continue; | |
3202 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
3203 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3204 | } | |
3205 | mutex_unlock(&ctx->uring_lock); | |
3206 | ||
3207 | return ret; | |
3208 | } | |
3209 | ||
affa87db | 3210 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
3211 | struct task_struct *task, |
3212 | bool cancel_all) | |
9936c7c2 | 3213 | { |
3dd0c97a | 3214 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 3215 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
3216 | enum io_wq_cancel cret; |
3217 | bool ret = false; | |
9936c7c2 | 3218 | |
360cd42c PB |
3219 | /* set it so io_req_local_work_add() would wake us up */ |
3220 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) { | |
3221 | atomic_set(&ctx->cq_wait_nr, 1); | |
3222 | smp_mb(); | |
3223 | } | |
3224 | ||
60053be8 PB |
3225 | /* failed during ring init, it couldn't have issued any requests */ |
3226 | if (!ctx->rings) | |
affa87db | 3227 | return false; |
60053be8 | 3228 | |
affa87db PB |
3229 | if (!task) { |
3230 | ret |= io_uring_try_cancel_iowq(ctx); | |
3231 | } else if (tctx && tctx->io_wq) { | |
3232 | /* | |
3233 | * Cancels requests of all rings, not only @ctx, but | |
3234 | * it's fine as the task is in exit/exec. | |
3235 | */ | |
3236 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
3237 | &cancel, true); | |
3238 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
3239 | } | |
9936c7c2 | 3240 | |
affa87db PB |
3241 | /* SQPOLL thread does its own polling */ |
3242 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
3243 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
3244 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
3245 | io_iopoll_try_reap_events(ctx); | |
3246 | ret = true; | |
fcc926bb | 3247 | cond_resched(); |
9936c7c2 | 3248 | } |
9936c7c2 | 3249 | } |
affa87db | 3250 | |
140102ae PB |
3251 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3252 | io_allowed_defer_tw_run(ctx)) | |
c0e0d6ba | 3253 | ret |= io_run_local_work(ctx) > 0; |
affa87db PB |
3254 | ret |= io_cancel_defer_files(ctx, task, cancel_all); |
3255 | mutex_lock(&ctx->uring_lock); | |
3256 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
3257 | mutex_unlock(&ctx->uring_lock); | |
3258 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
3259 | if (task) | |
c0e0d6ba | 3260 | ret |= io_run_task_work() > 0; |
affa87db | 3261 | return ret; |
9936c7c2 PB |
3262 | } |
3263 | ||
3f48cf18 | 3264 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 3265 | { |
3f48cf18 | 3266 | if (tracked) |
9cae36a0 | 3267 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
3268 | return percpu_counter_sum(&tctx->inflight); |
3269 | } | |
3270 | ||
78cc687b PB |
3271 | /* |
3272 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 3273 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 3274 | */ |
17437f31 | 3275 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 3276 | { |
521d6a73 | 3277 | struct io_uring_task *tctx = current->io_uring; |
734551df | 3278 | struct io_ring_ctx *ctx; |
360cd42c PB |
3279 | struct io_tctx_node *node; |
3280 | unsigned long index; | |
0e9ddb39 PB |
3281 | s64 inflight; |
3282 | DEFINE_WAIT(wait); | |
fdaf083c | 3283 | |
78cc687b PB |
3284 | WARN_ON_ONCE(sqd && sqd->thread != current); |
3285 | ||
6d042ffb PO |
3286 | if (!current->io_uring) |
3287 | return; | |
17a91051 PB |
3288 | if (tctx->io_wq) |
3289 | io_wq_exit_start(tctx->io_wq); | |
3290 | ||
8d664282 | 3291 | atomic_inc(&tctx->in_cancel); |
0e9ddb39 | 3292 | do { |
affa87db PB |
3293 | bool loop = false; |
3294 | ||
e9dbe221 | 3295 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 3296 | /* read completions before cancelations */ |
78cc687b | 3297 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
3298 | if (!inflight) |
3299 | break; | |
fdaf083c | 3300 | |
78cc687b | 3301 | if (!sqd) { |
78cc687b PB |
3302 | xa_for_each(&tctx->xa, index, node) { |
3303 | /* sqpoll task will cancel all its requests */ | |
3304 | if (node->ctx->sq_data) | |
3305 | continue; | |
affa87db PB |
3306 | loop |= io_uring_try_cancel_requests(node->ctx, |
3307 | current, cancel_all); | |
78cc687b PB |
3308 | } |
3309 | } else { | |
3310 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
3311 | loop |= io_uring_try_cancel_requests(ctx, |
3312 | current, | |
3313 | cancel_all); | |
3314 | } | |
3315 | ||
3316 | if (loop) { | |
3317 | cond_resched(); | |
3318 | continue; | |
78cc687b | 3319 | } |
17a91051 | 3320 | |
78a78060 JA |
3321 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
3322 | io_run_task_work(); | |
e9dbe221 | 3323 | io_uring_drop_tctx_refs(current); |
360cd42c PB |
3324 | xa_for_each(&tctx->xa, index, node) { |
3325 | if (!llist_empty(&node->ctx->work_llist)) { | |
3326 | WARN_ON_ONCE(node->ctx->submitter_task && | |
3327 | node->ctx->submitter_task != current); | |
3328 | goto end_wait; | |
3329 | } | |
3330 | } | |
0f212204 | 3331 | /* |
a1bb3cd5 PB |
3332 | * If we've seen completions, retry without waiting. This |
3333 | * avoids a race where a completion comes in before we did | |
3334 | * prepare_to_wait(). | |
0f212204 | 3335 | */ |
3dd0c97a | 3336 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 3337 | schedule(); |
360cd42c | 3338 | end_wait: |
f57555ed | 3339 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 3340 | } while (1); |
de7f1d9e | 3341 | |
8452d4a6 | 3342 | io_uring_clean_tctx(tctx); |
3dd0c97a | 3343 | if (cancel_all) { |
3cc7fdb9 PB |
3344 | /* |
3345 | * We shouldn't run task_works after cancel, so just leave | |
8d664282 | 3346 | * ->in_cancel set for normal exit. |
3cc7fdb9 | 3347 | */ |
8d664282 | 3348 | atomic_dec(&tctx->in_cancel); |
3f48cf18 PB |
3349 | /* for exec all current's requests should be gone, kill tctx */ |
3350 | __io_uring_free(current); | |
3351 | } | |
44e728b8 PB |
3352 | } |
3353 | ||
f552a27a | 3354 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 3355 | { |
f552a27a | 3356 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
3357 | } |
3358 | ||
6c5c240e RP |
3359 | static void *io_uring_validate_mmap_request(struct file *file, |
3360 | loff_t pgoff, size_t sz) | |
2b188cc1 | 3361 | { |
2b188cc1 | 3362 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 3363 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
3364 | struct page *page; |
3365 | void *ptr; | |
3366 | ||
03d89a2d JA |
3367 | /* Don't allow mmap if the ring was setup without it */ |
3368 | if (ctx->flags & IORING_SETUP_NO_MMAP) | |
3369 | return ERR_PTR(-EINVAL); | |
3370 | ||
c56e022c | 3371 | switch (offset & IORING_OFF_MMAP_MASK) { |
2b188cc1 | 3372 | case IORING_OFF_SQ_RING: |
75b28aff HV |
3373 | case IORING_OFF_CQ_RING: |
3374 | ptr = ctx->rings; | |
2b188cc1 JA |
3375 | break; |
3376 | case IORING_OFF_SQES: | |
3377 | ptr = ctx->sq_sqes; | |
3378 | break; | |
c56e022c JA |
3379 | case IORING_OFF_PBUF_RING: { |
3380 | unsigned int bgid; | |
3381 | ||
3382 | bgid = (offset & ~IORING_OFF_MMAP_MASK) >> IORING_OFF_PBUF_SHIFT; | |
3383 | mutex_lock(&ctx->uring_lock); | |
3384 | ptr = io_pbuf_get_address(ctx, bgid); | |
3385 | mutex_unlock(&ctx->uring_lock); | |
3386 | if (!ptr) | |
3387 | return ERR_PTR(-EINVAL); | |
3388 | break; | |
3389 | } | |
2b188cc1 | 3390 | default: |
6c5c240e | 3391 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
3392 | } |
3393 | ||
3394 | page = virt_to_head_page(ptr); | |
a50b854e | 3395 | if (sz > page_size(page)) |
6c5c240e RP |
3396 | return ERR_PTR(-EINVAL); |
3397 | ||
3398 | return ptr; | |
3399 | } | |
3400 | ||
3401 | #ifdef CONFIG_MMU | |
3402 | ||
c072481d | 3403 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e RP |
3404 | { |
3405 | size_t sz = vma->vm_end - vma->vm_start; | |
3406 | unsigned long pfn; | |
3407 | void *ptr; | |
3408 | ||
3409 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
3410 | if (IS_ERR(ptr)) | |
3411 | return PTR_ERR(ptr); | |
2b188cc1 JA |
3412 | |
3413 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
3414 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
3415 | } | |
3416 | ||
d808459b HD |
3417 | static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp, |
3418 | unsigned long addr, unsigned long len, | |
3419 | unsigned long pgoff, unsigned long flags) | |
3420 | { | |
3421 | const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags); | |
3422 | struct vm_unmapped_area_info info; | |
3423 | void *ptr; | |
3424 | ||
3425 | /* | |
3426 | * Do not allow to map to user-provided address to avoid breaking the | |
3427 | * aliasing rules. Userspace is not able to guess the offset address of | |
3428 | * kernel kmalloc()ed memory area. | |
3429 | */ | |
3430 | if (addr) | |
3431 | return -EINVAL; | |
3432 | ||
3433 | ptr = io_uring_validate_mmap_request(filp, pgoff, len); | |
3434 | if (IS_ERR(ptr)) | |
3435 | return -ENOMEM; | |
3436 | ||
3437 | info.flags = VM_UNMAPPED_AREA_TOPDOWN; | |
3438 | info.length = len; | |
3439 | info.low_limit = max(PAGE_SIZE, mmap_min_addr); | |
3440 | info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base); | |
3441 | #ifdef SHM_COLOUR | |
3442 | info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL); | |
3443 | #else | |
3444 | info.align_mask = PAGE_MASK & (SHMLBA - 1UL); | |
3445 | #endif | |
3446 | info.align_offset = (unsigned long) ptr; | |
3447 | ||
3448 | /* | |
3449 | * A failed mmap() very likely causes application failure, | |
3450 | * so fall back to the bottom-up function here. This scenario | |
3451 | * can happen with large stack limits and large mmap() | |
3452 | * allocations. | |
3453 | */ | |
3454 | addr = vm_unmapped_area(&info); | |
3455 | if (offset_in_page(addr)) { | |
3456 | info.flags = 0; | |
3457 | info.low_limit = TASK_UNMAPPED_BASE; | |
3458 | info.high_limit = mmap_end; | |
3459 | addr = vm_unmapped_area(&info); | |
3460 | } | |
3461 | ||
3462 | return addr; | |
3463 | } | |
3464 | ||
6c5c240e RP |
3465 | #else /* !CONFIG_MMU */ |
3466 | ||
3467 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
3468 | { | |
fc4f4be9 | 3469 | return is_nommu_shared_mapping(vma->vm_flags) ? 0 : -EINVAL; |
6c5c240e RP |
3470 | } |
3471 | ||
3472 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
3473 | { | |
3474 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
3475 | } | |
3476 | ||
3477 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
3478 | unsigned long addr, unsigned long len, | |
3479 | unsigned long pgoff, unsigned long flags) | |
3480 | { | |
3481 | void *ptr; | |
3482 | ||
3483 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
3484 | if (IS_ERR(ptr)) | |
3485 | return PTR_ERR(ptr); | |
3486 | ||
3487 | return (unsigned long) ptr; | |
3488 | } | |
3489 | ||
3490 | #endif /* !CONFIG_MMU */ | |
3491 | ||
f81440d3 PB |
3492 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
3493 | { | |
3494 | if (flags & IORING_ENTER_EXT_ARG) { | |
3495 | struct io_uring_getevents_arg arg; | |
3496 | ||
3497 | if (argsz != sizeof(arg)) | |
3498 | return -EINVAL; | |
3499 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3500 | return -EFAULT; | |
3501 | } | |
3502 | return 0; | |
3503 | } | |
3504 | ||
c73ebb68 HX |
3505 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
3506 | struct __kernel_timespec __user **ts, | |
3507 | const sigset_t __user **sig) | |
3508 | { | |
3509 | struct io_uring_getevents_arg arg; | |
3510 | ||
3511 | /* | |
3512 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
3513 | * is just a pointer to the sigset_t. | |
3514 | */ | |
3515 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
3516 | *sig = (const sigset_t __user *) argp; | |
3517 | *ts = NULL; | |
3518 | return 0; | |
3519 | } | |
3520 | ||
3521 | /* | |
3522 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
3523 | * timespec and sigset_t pointers if good. | |
3524 | */ | |
3525 | if (*argsz != sizeof(arg)) | |
3526 | return -EINVAL; | |
3527 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3528 | return -EFAULT; | |
d2347b96 DY |
3529 | if (arg.pad) |
3530 | return -EINVAL; | |
c73ebb68 HX |
3531 | *sig = u64_to_user_ptr(arg.sigmask); |
3532 | *argsz = arg.sigmask_sz; | |
3533 | *ts = u64_to_user_ptr(arg.ts); | |
3534 | return 0; | |
3535 | } | |
3536 | ||
2b188cc1 | 3537 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
3538 | u32, min_complete, u32, flags, const void __user *, argp, |
3539 | size_t, argsz) | |
2b188cc1 JA |
3540 | { |
3541 | struct io_ring_ctx *ctx; | |
2b188cc1 | 3542 | struct fd f; |
33f993da | 3543 | long ret; |
2b188cc1 | 3544 | |
33f993da | 3545 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
3546 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
3547 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
3548 | return -EINVAL; |
3549 | ||
e7a6c00d JA |
3550 | /* |
3551 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3552 | * need only dereference our task private array to find it. | |
3553 | */ | |
3554 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3555 | struct io_uring_task *tctx = current->io_uring; | |
3556 | ||
3273c440 | 3557 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
e7a6c00d JA |
3558 | return -EINVAL; |
3559 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
3560 | f.file = tctx->registered_rings[fd]; | |
4329490a | 3561 | f.flags = 0; |
3273c440 PB |
3562 | if (unlikely(!f.file)) |
3563 | return -EBADF; | |
e7a6c00d JA |
3564 | } else { |
3565 | f = fdget(fd); | |
3273c440 PB |
3566 | if (unlikely(!f.file)) |
3567 | return -EBADF; | |
3568 | ret = -EOPNOTSUPP; | |
3569 | if (unlikely(!io_is_uring_fops(f.file))) | |
fbb8bb02 | 3570 | goto out; |
e7a6c00d | 3571 | } |
2b188cc1 | 3572 | |
2b188cc1 | 3573 | ctx = f.file->private_data; |
7e84e1c7 | 3574 | ret = -EBADFD; |
33f993da | 3575 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3576 | goto out; |
3577 | ||
6c271ce2 JA |
3578 | /* |
3579 | * For SQ polling, the thread will do all submissions and completions. | |
3580 | * Just return the requested submit count, and wake the thread if | |
3581 | * we were asked to. | |
3582 | */ | |
b2a9eada | 3583 | ret = 0; |
6c271ce2 | 3584 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3585 | io_cqring_overflow_flush(ctx); |
89448c47 | 3586 | |
21f96522 JA |
3587 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3588 | ret = -EOWNERDEAD; | |
04147488 | 3589 | goto out; |
21f96522 | 3590 | } |
6c271ce2 | 3591 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3592 | wake_up(&ctx->sq_data->wait); |
88b80534 QF |
3593 | if (flags & IORING_ENTER_SQ_WAIT) |
3594 | io_sqpoll_wait_sq(ctx); | |
3595 | ||
3e813c90 | 3596 | ret = to_submit; |
b2a9eada | 3597 | } else if (to_submit) { |
eef51daa | 3598 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3599 | if (unlikely(ret)) |
3600 | goto out; | |
7c504e65 | 3601 | |
2b188cc1 | 3602 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3603 | ret = io_submit_sqes(ctx, to_submit); |
3604 | if (ret != to_submit) { | |
d487b43c | 3605 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3606 | goto out; |
d487b43c | 3607 | } |
44f87745 PB |
3608 | if (flags & IORING_ENTER_GETEVENTS) { |
3609 | if (ctx->syscall_iopoll) | |
3610 | goto iopoll_locked; | |
3611 | /* | |
3612 | * Ignore errors, we'll soon call io_cqring_wait() and | |
3613 | * it should handle ownership problems if any. | |
3614 | */ | |
3615 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) | |
3616 | (void)io_run_local_work_locked(ctx); | |
3617 | } | |
d487b43c | 3618 | mutex_unlock(&ctx->uring_lock); |
2b188cc1 | 3619 | } |
c0e0d6ba | 3620 | |
2b188cc1 | 3621 | if (flags & IORING_ENTER_GETEVENTS) { |
3e813c90 | 3622 | int ret2; |
c0e0d6ba | 3623 | |
773697b6 | 3624 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3625 | /* |
3626 | * We disallow the app entering submit/complete with | |
3627 | * polling, but we still need to lock the ring to | |
3628 | * prevent racing with polled issue that got punted to | |
3629 | * a workqueue. | |
3630 | */ | |
3631 | mutex_lock(&ctx->uring_lock); | |
3632 | iopoll_locked: | |
3e813c90 DY |
3633 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3634 | if (likely(!ret2)) { | |
3635 | min_complete = min(min_complete, | |
3636 | ctx->cq_entries); | |
3637 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3638 | } |
3639 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3640 | } else { |
f81440d3 PB |
3641 | const sigset_t __user *sig; |
3642 | struct __kernel_timespec __user *ts; | |
3643 | ||
3e813c90 DY |
3644 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3645 | if (likely(!ret2)) { | |
3646 | min_complete = min(min_complete, | |
3647 | ctx->cq_entries); | |
3648 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3649 | argsz, ts); | |
3650 | } | |
def596e9 | 3651 | } |
c73ebb68 | 3652 | |
155bc950 | 3653 | if (!ret) { |
3e813c90 | 3654 | ret = ret2; |
2b188cc1 | 3655 | |
155bc950 DY |
3656 | /* |
3657 | * EBADR indicates that one or more CQE were dropped. | |
3658 | * Once the user has been informed we can clear the bit | |
3659 | * as they are obviously ok with those drops. | |
3660 | */ | |
3661 | if (unlikely(ret2 == -EBADR)) | |
3662 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3663 | &ctx->check_cq); | |
def596e9 | 3664 | } |
2b188cc1 | 3665 | } |
7c504e65 | 3666 | out: |
4329490a | 3667 | fdput(f); |
3e813c90 | 3668 | return ret; |
2b188cc1 JA |
3669 | } |
3670 | ||
3671 | static const struct file_operations io_uring_fops = { | |
3672 | .release = io_uring_release, | |
3673 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3674 | #ifndef CONFIG_MMU |
3675 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3676 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
d808459b HD |
3677 | #else |
3678 | .get_unmapped_area = io_uring_mmu_get_unmapped_area, | |
6c5c240e | 3679 | #endif |
2b188cc1 | 3680 | .poll = io_uring_poll, |
bebdb65e | 3681 | #ifdef CONFIG_PROC_FS |
87ce955b | 3682 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3683 | #endif |
2b188cc1 JA |
3684 | }; |
3685 | ||
92ac8bea JA |
3686 | bool io_is_uring_fops(struct file *file) |
3687 | { | |
3688 | return file->f_op == &io_uring_fops; | |
3689 | } | |
3690 | ||
c072481d PB |
3691 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3692 | struct io_uring_params *p) | |
2b188cc1 | 3693 | { |
75b28aff HV |
3694 | struct io_rings *rings; |
3695 | size_t size, sq_array_offset; | |
e27cef86 | 3696 | void *ptr; |
2b188cc1 | 3697 | |
bd740481 JA |
3698 | /* make sure these are sane, as we already accounted them */ |
3699 | ctx->sq_entries = p->sq_entries; | |
3700 | ctx->cq_entries = p->cq_entries; | |
3701 | ||
baf9cb64 | 3702 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3703 | if (size == SIZE_MAX) |
3704 | return -EOVERFLOW; | |
3705 | ||
03d89a2d JA |
3706 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3707 | rings = io_mem_alloc(size); | |
3708 | else | |
3709 | rings = io_rings_map(ctx, p->cq_off.user_addr, size); | |
3710 | ||
e27cef86 JA |
3711 | if (IS_ERR(rings)) |
3712 | return PTR_ERR(rings); | |
2b188cc1 | 3713 | |
75b28aff HV |
3714 | ctx->rings = rings; |
3715 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
3716 | rings->sq_ring_mask = p->sq_entries - 1; | |
3717 | rings->cq_ring_mask = p->cq_entries - 1; | |
3718 | rings->sq_ring_entries = p->sq_entries; | |
3719 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3720 | |
ebdeb7c0 JA |
3721 | if (p->flags & IORING_SETUP_SQE128) |
3722 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3723 | else | |
3724 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 | 3725 | if (size == SIZE_MAX) { |
9c189eee | 3726 | io_rings_free(ctx); |
2b188cc1 | 3727 | return -EOVERFLOW; |
eb065d30 | 3728 | } |
2b188cc1 | 3729 | |
03d89a2d JA |
3730 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3731 | ptr = io_mem_alloc(size); | |
3732 | else | |
3733 | ptr = io_sqes_map(ctx, p->sq_off.user_addr, size); | |
3734 | ||
e27cef86 | 3735 | if (IS_ERR(ptr)) { |
9c189eee | 3736 | io_rings_free(ctx); |
e27cef86 | 3737 | return PTR_ERR(ptr); |
eb065d30 | 3738 | } |
2b188cc1 | 3739 | |
e27cef86 | 3740 | ctx->sq_sqes = ptr; |
2b188cc1 JA |
3741 | return 0; |
3742 | } | |
3743 | ||
6e76ac59 | 3744 | static int io_uring_install_fd(struct file *file) |
9faadcc8 | 3745 | { |
6e76ac59 | 3746 | int fd; |
9faadcc8 PB |
3747 | |
3748 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3749 | if (fd < 0) | |
3750 | return fd; | |
9faadcc8 PB |
3751 | fd_install(fd, file); |
3752 | return fd; | |
3753 | } | |
3754 | ||
2b188cc1 JA |
3755 | /* |
3756 | * Allocate an anonymous fd, this is what constitutes the application | |
3757 | * visible backing of an io_uring instance. The application mmaps this | |
3758 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3759 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3760 | */ | |
9faadcc8 | 3761 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
3762 | { |
3763 | struct file *file; | |
9faadcc8 | 3764 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
3765 | int ret; |
3766 | ||
2b188cc1 JA |
3767 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
3768 | &ctx->ring_sock); | |
3769 | if (ret) | |
9faadcc8 | 3770 | return ERR_PTR(ret); |
2b188cc1 JA |
3771 | #endif |
3772 | ||
91a9ab7c PM |
3773 | file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, |
3774 | O_RDWR | O_CLOEXEC, NULL); | |
2b188cc1 | 3775 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
3776 | if (IS_ERR(file)) { |
3777 | sock_release(ctx->ring_sock); | |
3778 | ctx->ring_sock = NULL; | |
3779 | } else { | |
3780 | ctx->ring_sock->file = file; | |
0f212204 | 3781 | } |
2b188cc1 | 3782 | #endif |
9faadcc8 | 3783 | return file; |
2b188cc1 JA |
3784 | } |
3785 | ||
c072481d PB |
3786 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3787 | struct io_uring_params __user *params) | |
2b188cc1 | 3788 | { |
2b188cc1 | 3789 | struct io_ring_ctx *ctx; |
6e76ac59 | 3790 | struct io_uring_task *tctx; |
9faadcc8 | 3791 | struct file *file; |
2b188cc1 JA |
3792 | int ret; |
3793 | ||
8110c1a6 | 3794 | if (!entries) |
2b188cc1 | 3795 | return -EINVAL; |
8110c1a6 JA |
3796 | if (entries > IORING_MAX_ENTRIES) { |
3797 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3798 | return -EINVAL; | |
3799 | entries = IORING_MAX_ENTRIES; | |
3800 | } | |
2b188cc1 | 3801 | |
6e76ac59 JT |
3802 | if ((p->flags & IORING_SETUP_REGISTERED_FD_ONLY) |
3803 | && !(p->flags & IORING_SETUP_NO_MMAP)) | |
3804 | return -EINVAL; | |
3805 | ||
2b188cc1 JA |
3806 | /* |
3807 | * Use twice as many entries for the CQ ring. It's possible for the | |
3808 | * application to drive a higher depth than the size of the SQ ring, | |
3809 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3810 | * some flexibility in overcommitting a bit. If the application has |
3811 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3812 | * of CQ ring entries manually. | |
2b188cc1 JA |
3813 | */ |
3814 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3815 | if (p->flags & IORING_SETUP_CQSIZE) { |
3816 | /* | |
3817 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3818 | * to a power-of-two, if it isn't already. We do NOT impose | |
3819 | * any cq vs sq ring sizing. | |
3820 | */ | |
eb2667b3 | 3821 | if (!p->cq_entries) |
33a107f0 | 3822 | return -EINVAL; |
8110c1a6 JA |
3823 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3824 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3825 | return -EINVAL; | |
3826 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3827 | } | |
eb2667b3 JQ |
3828 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3829 | if (p->cq_entries < p->sq_entries) | |
3830 | return -EINVAL; | |
33a107f0 JA |
3831 | } else { |
3832 | p->cq_entries = 2 * p->sq_entries; | |
3833 | } | |
2b188cc1 | 3834 | |
2b188cc1 | 3835 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3836 | if (!ctx) |
2b188cc1 | 3837 | return -ENOMEM; |
773697b6 | 3838 | |
e6aeb272 PB |
3839 | if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) && |
3840 | !(ctx->flags & IORING_SETUP_IOPOLL) && | |
3841 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3842 | ctx->task_complete = true; | |
3843 | ||
bca39f39 PB |
3844 | /* |
3845 | * lazy poll_wq activation relies on ->task_complete for synchronisation | |
3846 | * purposes, see io_activate_pollwq() | |
3847 | */ | |
3848 | if (!ctx->task_complete) | |
3849 | ctx->poll_activated = true; | |
3850 | ||
773697b6 PB |
3851 | /* |
3852 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3853 | * space applications don't need to do io completion events | |
3854 | * polling again, they can rely on io_sq_thread to do polling | |
3855 | * work, which can reduce cpu usage and uring_lock contention. | |
3856 | */ | |
3857 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3858 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3859 | ctx->syscall_iopoll = 1; | |
3860 | ||
2b188cc1 | 3861 | ctx->compat = in_compat_syscall(); |
62e398be JA |
3862 | if (!capable(CAP_IPC_LOCK)) |
3863 | ctx->user = get_uid(current_user()); | |
2aede0e4 | 3864 | |
9f010507 | 3865 | /* |
e1169f06 JA |
3866 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3867 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3868 | */ |
e1169f06 JA |
3869 | ret = -EINVAL; |
3870 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3871 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 | 3872 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
c0e0d6ba DY |
3873 | IORING_SETUP_TASKRUN_FLAG | |
3874 | IORING_SETUP_DEFER_TASKRUN)) | |
e1169f06 | 3875 | goto err; |
9f010507 | 3876 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3877 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3878 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3879 | } else { | |
c0e0d6ba DY |
3880 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG && |
3881 | !(ctx->flags & IORING_SETUP_DEFER_TASKRUN)) | |
ef060ea9 | 3882 | goto err; |
9f010507 | 3883 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3884 | } |
9f010507 | 3885 | |
c0e0d6ba DY |
3886 | /* |
3887 | * For DEFER_TASKRUN we require the completion task to be the same as the | |
3888 | * submission task. This implies that there is only one submitter, so enforce | |
3889 | * that. | |
3890 | */ | |
3891 | if (ctx->flags & IORING_SETUP_DEFER_TASKRUN && | |
3892 | !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) { | |
3893 | goto err; | |
3894 | } | |
3895 | ||
2aede0e4 JA |
3896 | /* |
3897 | * This is just grabbed for accounting purposes. When a process exits, | |
3898 | * the mm is exited and dropped before the files, hence we need to hang | |
3899 | * on to this mm purely for the purposes of being able to unaccount | |
3900 | * memory (locked/pinned vm). It's not used for anything else. | |
3901 | */ | |
6b7898eb | 3902 | mmgrab(current->mm); |
2aede0e4 | 3903 | ctx->mm_account = current->mm; |
6b7898eb | 3904 | |
2b188cc1 JA |
3905 | ret = io_allocate_scq_urings(ctx, p); |
3906 | if (ret) | |
3907 | goto err; | |
3908 | ||
7e84e1c7 | 3909 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
3910 | if (ret) |
3911 | goto err; | |
2933ae6e PB |
3912 | |
3913 | ret = io_rsrc_init(ctx); | |
47b228ce PB |
3914 | if (ret) |
3915 | goto err; | |
2b188cc1 | 3916 | |
75b28aff HV |
3917 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3918 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3919 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3920 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3921 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3922 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
3923 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
9b1b58ca | 3924 | p->sq_off.resv1 = 0; |
03d89a2d JA |
3925 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3926 | p->sq_off.user_addr = 0; | |
2b188cc1 | 3927 | |
75b28aff HV |
3928 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3929 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3930 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3931 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3932 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3933 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 3934 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
9b1b58ca | 3935 | p->cq_off.resv1 = 0; |
03d89a2d JA |
3936 | if (!(ctx->flags & IORING_SETUP_NO_MMAP)) |
3937 | p->cq_off.user_addr = 0; | |
ac90f249 | 3938 | |
7f13657d XW |
3939 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
3940 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 3941 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 3942 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 3943 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e | 3944 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
7d3fd88d | 3945 | IORING_FEAT_LINKED_FILE | IORING_FEAT_REG_REG_RING; |
7f13657d XW |
3946 | |
3947 | if (copy_to_user(params, p, sizeof(*p))) { | |
3948 | ret = -EFAULT; | |
3949 | goto err; | |
3950 | } | |
d1719f70 | 3951 | |
7cae596b DY |
3952 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER |
3953 | && !(ctx->flags & IORING_SETUP_R_DISABLED)) | |
8579538c | 3954 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
7cae596b | 3955 | |
9faadcc8 PB |
3956 | file = io_uring_get_file(ctx); |
3957 | if (IS_ERR(file)) { | |
3958 | ret = PTR_ERR(file); | |
3959 | goto err; | |
3960 | } | |
3961 | ||
6e76ac59 JT |
3962 | ret = __io_uring_add_tctx_node(ctx); |
3963 | if (ret) | |
3964 | goto err_fput; | |
3965 | tctx = current->io_uring; | |
3966 | ||
044c1ab3 JA |
3967 | /* |
3968 | * Install ring fd as the very last thing, so we don't risk someone | |
3969 | * having closed it before we finish setup | |
3970 | */ | |
6e76ac59 JT |
3971 | if (p->flags & IORING_SETUP_REGISTERED_FD_ONLY) |
3972 | ret = io_ring_add_registered_file(tctx, file, 0, IO_RINGFD_REG_MAX); | |
3973 | else | |
3974 | ret = io_uring_install_fd(file); | |
3975 | if (ret < 0) | |
3976 | goto err_fput; | |
044c1ab3 | 3977 | |
c826bd7a | 3978 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
3979 | return ret; |
3980 | err: | |
3981 | io_ring_ctx_wait_and_kill(ctx); | |
3982 | return ret; | |
6e76ac59 JT |
3983 | err_fput: |
3984 | fput(file); | |
3985 | return ret; | |
2b188cc1 JA |
3986 | } |
3987 | ||
3988 | /* | |
3989 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3990 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3991 | * params structure passed in. | |
3992 | */ | |
3993 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3994 | { | |
3995 | struct io_uring_params p; | |
2b188cc1 JA |
3996 | int i; |
3997 | ||
3998 | if (copy_from_user(&p, params, sizeof(p))) | |
3999 | return -EFAULT; | |
4000 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
4001 | if (p.resv[i]) | |
4002 | return -EINVAL; | |
4003 | } | |
4004 | ||
6c271ce2 | 4005 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 4006 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 4007 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 4008 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 4009 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 | 4010 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
03d89a2d | 4011 | IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN | |
6e76ac59 | 4012 | IORING_SETUP_NO_MMAP | IORING_SETUP_REGISTERED_FD_ONLY)) |
2b188cc1 JA |
4013 | return -EINVAL; |
4014 | ||
ef060ea9 | 4015 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
4016 | } |
4017 | ||
4018 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
4019 | struct io_uring_params __user *, params) | |
4020 | { | |
4021 | return io_uring_setup(entries, params); | |
4022 | } | |
4023 | ||
c072481d PB |
4024 | static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, |
4025 | unsigned nr_args) | |
66f4af93 JA |
4026 | { |
4027 | struct io_uring_probe *p; | |
4028 | size_t size; | |
4029 | int i, ret; | |
4030 | ||
4031 | size = struct_size(p, ops, nr_args); | |
4032 | if (size == SIZE_MAX) | |
4033 | return -EOVERFLOW; | |
4034 | p = kzalloc(size, GFP_KERNEL); | |
4035 | if (!p) | |
4036 | return -ENOMEM; | |
4037 | ||
4038 | ret = -EFAULT; | |
4039 | if (copy_from_user(p, arg, size)) | |
4040 | goto out; | |
4041 | ret = -EINVAL; | |
4042 | if (memchr_inv(p, 0, size)) | |
4043 | goto out; | |
4044 | ||
4045 | p->last_op = IORING_OP_LAST - 1; | |
4046 | if (nr_args > IORING_OP_LAST) | |
4047 | nr_args = IORING_OP_LAST; | |
4048 | ||
4049 | for (i = 0; i < nr_args; i++) { | |
4050 | p->ops[i].op = i; | |
a7dd2782 | 4051 | if (!io_issue_defs[i].not_supported) |
66f4af93 JA |
4052 | p->ops[i].flags = IO_URING_OP_SUPPORTED; |
4053 | } | |
4054 | p->ops_len = i; | |
4055 | ||
4056 | ret = 0; | |
4057 | if (copy_to_user(arg, p, size)) | |
4058 | ret = -EFAULT; | |
4059 | out: | |
4060 | kfree(p); | |
4061 | return ret; | |
4062 | } | |
4063 | ||
071698e1 JA |
4064 | static int io_register_personality(struct io_ring_ctx *ctx) |
4065 | { | |
4379bf8b | 4066 | const struct cred *creds; |
61cf9370 | 4067 | u32 id; |
1e6fa521 | 4068 | int ret; |
071698e1 | 4069 | |
4379bf8b | 4070 | creds = get_current_cred(); |
1e6fa521 | 4071 | |
61cf9370 MWO |
4072 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
4073 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
4074 | if (ret < 0) { |
4075 | put_cred(creds); | |
4076 | return ret; | |
4077 | } | |
4078 | return id; | |
071698e1 JA |
4079 | } |
4080 | ||
c072481d PB |
4081 | static __cold int io_register_restrictions(struct io_ring_ctx *ctx, |
4082 | void __user *arg, unsigned int nr_args) | |
21b55dbc SG |
4083 | { |
4084 | struct io_uring_restriction *res; | |
4085 | size_t size; | |
4086 | int i, ret; | |
4087 | ||
7e84e1c7 SG |
4088 | /* Restrictions allowed only if rings started disabled */ |
4089 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
4090 | return -EBADFD; | |
4091 | ||
21b55dbc | 4092 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 4093 | if (ctx->restrictions.registered) |
21b55dbc SG |
4094 | return -EBUSY; |
4095 | ||
4096 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
4097 | return -EINVAL; | |
4098 | ||
4099 | size = array_size(nr_args, sizeof(*res)); | |
4100 | if (size == SIZE_MAX) | |
4101 | return -EOVERFLOW; | |
4102 | ||
4103 | res = memdup_user(arg, size); | |
4104 | if (IS_ERR(res)) | |
4105 | return PTR_ERR(res); | |
4106 | ||
4107 | ret = 0; | |
4108 | ||
4109 | for (i = 0; i < nr_args; i++) { | |
4110 | switch (res[i].opcode) { | |
4111 | case IORING_RESTRICTION_REGISTER_OP: | |
4112 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
4113 | ret = -EINVAL; | |
4114 | goto out; | |
4115 | } | |
4116 | ||
4117 | __set_bit(res[i].register_op, | |
4118 | ctx->restrictions.register_op); | |
4119 | break; | |
4120 | case IORING_RESTRICTION_SQE_OP: | |
4121 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
4122 | ret = -EINVAL; | |
4123 | goto out; | |
4124 | } | |
4125 | ||
4126 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
4127 | break; | |
4128 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
4129 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
4130 | break; | |
4131 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
4132 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
4133 | break; | |
4134 | default: | |
4135 | ret = -EINVAL; | |
4136 | goto out; | |
4137 | } | |
4138 | } | |
4139 | ||
4140 | out: | |
4141 | /* Reset all restrictions if an error happened */ | |
4142 | if (ret != 0) | |
4143 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
4144 | else | |
7e84e1c7 | 4145 | ctx->restrictions.registered = true; |
21b55dbc SG |
4146 | |
4147 | kfree(res); | |
4148 | return ret; | |
4149 | } | |
4150 | ||
7e84e1c7 SG |
4151 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
4152 | { | |
4153 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
4154 | return -EBADFD; | |
4155 | ||
bca39f39 | 4156 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task) { |
8579538c | 4157 | WRITE_ONCE(ctx->submitter_task, get_task_struct(current)); |
bca39f39 PB |
4158 | /* |
4159 | * Lazy activation attempts would fail if it was polled before | |
4160 | * submitter_task is set. | |
4161 | */ | |
4162 | if (wq_has_sleeper(&ctx->poll_wq)) | |
4163 | io_activate_pollwq(ctx); | |
4164 | } | |
7cae596b | 4165 | |
7e84e1c7 SG |
4166 | if (ctx->restrictions.registered) |
4167 | ctx->restricted = 1; | |
4168 | ||
0298ef96 PB |
4169 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
4170 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
4171 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
4172 | return 0; |
4173 | } | |
4174 | ||
c072481d PB |
4175 | static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, |
4176 | void __user *arg, unsigned len) | |
fe76421d JA |
4177 | { |
4178 | struct io_uring_task *tctx = current->io_uring; | |
4179 | cpumask_var_t new_mask; | |
4180 | int ret; | |
4181 | ||
4182 | if (!tctx || !tctx->io_wq) | |
4183 | return -EINVAL; | |
4184 | ||
4185 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
4186 | return -ENOMEM; | |
4187 | ||
4188 | cpumask_clear(new_mask); | |
4189 | if (len > cpumask_size()) | |
4190 | len = cpumask_size(); | |
4191 | ||
0f5e4b83 ES |
4192 | if (in_compat_syscall()) { |
4193 | ret = compat_get_bitmap(cpumask_bits(new_mask), | |
4194 | (const compat_ulong_t __user *)arg, | |
4195 | len * 8 /* CHAR_BIT */); | |
4196 | } else { | |
4197 | ret = copy_from_user(new_mask, arg, len); | |
4198 | } | |
4199 | ||
4200 | if (ret) { | |
fe76421d JA |
4201 | free_cpumask_var(new_mask); |
4202 | return -EFAULT; | |
4203 | } | |
4204 | ||
4205 | ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); | |
4206 | free_cpumask_var(new_mask); | |
4207 | return ret; | |
4208 | } | |
4209 | ||
c072481d | 4210 | static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) |
fe76421d JA |
4211 | { |
4212 | struct io_uring_task *tctx = current->io_uring; | |
4213 | ||
4214 | if (!tctx || !tctx->io_wq) | |
4215 | return -EINVAL; | |
4216 | ||
4217 | return io_wq_cpu_affinity(tctx->io_wq, NULL); | |
4218 | } | |
4219 | ||
c072481d PB |
4220 | static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, |
4221 | void __user *arg) | |
b22fa62a | 4222 | __must_hold(&ctx->uring_lock) |
2e480058 | 4223 | { |
b22fa62a | 4224 | struct io_tctx_node *node; |
fa84693b JA |
4225 | struct io_uring_task *tctx = NULL; |
4226 | struct io_sq_data *sqd = NULL; | |
2e480058 JA |
4227 | __u32 new_count[2]; |
4228 | int i, ret; | |
4229 | ||
2e480058 JA |
4230 | if (copy_from_user(new_count, arg, sizeof(new_count))) |
4231 | return -EFAULT; | |
4232 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
4233 | if (new_count[i] > INT_MAX) | |
4234 | return -EINVAL; | |
4235 | ||
fa84693b JA |
4236 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
4237 | sqd = ctx->sq_data; | |
4238 | if (sqd) { | |
009ad9f0 JA |
4239 | /* |
4240 | * Observe the correct sqd->lock -> ctx->uring_lock | |
4241 | * ordering. Fine to drop uring_lock here, we hold | |
4242 | * a ref to the ctx. | |
4243 | */ | |
41d3a6bd | 4244 | refcount_inc(&sqd->refs); |
009ad9f0 | 4245 | mutex_unlock(&ctx->uring_lock); |
fa84693b | 4246 | mutex_lock(&sqd->lock); |
009ad9f0 | 4247 | mutex_lock(&ctx->uring_lock); |
41d3a6bd JA |
4248 | if (sqd->thread) |
4249 | tctx = sqd->thread->io_uring; | |
fa84693b JA |
4250 | } |
4251 | } else { | |
4252 | tctx = current->io_uring; | |
4253 | } | |
4254 | ||
e139a1ec | 4255 | BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); |
fa84693b | 4256 | |
bad119b9 PB |
4257 | for (i = 0; i < ARRAY_SIZE(new_count); i++) |
4258 | if (new_count[i]) | |
4259 | ctx->iowq_limits[i] = new_count[i]; | |
e139a1ec PB |
4260 | ctx->iowq_limits_set = true; |
4261 | ||
e139a1ec PB |
4262 | if (tctx && tctx->io_wq) { |
4263 | ret = io_wq_max_workers(tctx->io_wq, new_count); | |
4264 | if (ret) | |
4265 | goto err; | |
4266 | } else { | |
4267 | memset(new_count, 0, sizeof(new_count)); | |
4268 | } | |
fa84693b | 4269 | |
41d3a6bd | 4270 | if (sqd) { |
fa84693b | 4271 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
4272 | io_put_sq_data(sqd); |
4273 | } | |
2e480058 JA |
4274 | |
4275 | if (copy_to_user(arg, new_count, sizeof(new_count))) | |
4276 | return -EFAULT; | |
4277 | ||
b22fa62a PB |
4278 | /* that's it for SQPOLL, only the SQPOLL task creates requests */ |
4279 | if (sqd) | |
4280 | return 0; | |
4281 | ||
4282 | /* now propagate the restriction to all registered users */ | |
4283 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
4284 | struct io_uring_task *tctx = node->task->io_uring; | |
4285 | ||
4286 | if (WARN_ON_ONCE(!tctx->io_wq)) | |
4287 | continue; | |
4288 | ||
4289 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
4290 | new_count[i] = ctx->iowq_limits[i]; | |
4291 | /* ignore errors, it always returns zero anyway */ | |
4292 | (void)io_wq_max_workers(tctx->io_wq, new_count); | |
4293 | } | |
2e480058 | 4294 | return 0; |
fa84693b | 4295 | err: |
41d3a6bd | 4296 | if (sqd) { |
fa84693b | 4297 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
4298 | io_put_sq_data(sqd); |
4299 | } | |
fa84693b | 4300 | return ret; |
2e480058 JA |
4301 | } |
4302 | ||
edafccee JA |
4303 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
4304 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
4305 | __releases(ctx->uring_lock) |
4306 | __acquires(ctx->uring_lock) | |
edafccee JA |
4307 | { |
4308 | int ret; | |
4309 | ||
35fa71a0 | 4310 | /* |
fbb8bb02 PB |
4311 | * We don't quiesce the refs for register anymore and so it can't be |
4312 | * dying as we're holding a file ref here. | |
35fa71a0 | 4313 | */ |
fbb8bb02 | 4314 | if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs))) |
35fa71a0 JA |
4315 | return -ENXIO; |
4316 | ||
d7cce96c PB |
4317 | if (ctx->submitter_task && ctx->submitter_task != current) |
4318 | return -EEXIST; | |
4319 | ||
75c4021a | 4320 | if (ctx->restricted) { |
75c4021a PB |
4321 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); |
4322 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
4323 | return -EACCES; | |
4324 | } | |
4325 | ||
edafccee JA |
4326 | switch (opcode) { |
4327 | case IORING_REGISTER_BUFFERS: | |
0184f08e PB |
4328 | ret = -EFAULT; |
4329 | if (!arg) | |
4330 | break; | |
634d00df | 4331 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
4332 | break; |
4333 | case IORING_UNREGISTER_BUFFERS: | |
4334 | ret = -EINVAL; | |
4335 | if (arg || nr_args) | |
4336 | break; | |
0a96bbe4 | 4337 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 4338 | break; |
6b06314c | 4339 | case IORING_REGISTER_FILES: |
a8da73a3 JA |
4340 | ret = -EFAULT; |
4341 | if (!arg) | |
4342 | break; | |
792e3582 | 4343 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
4344 | break; |
4345 | case IORING_UNREGISTER_FILES: | |
4346 | ret = -EINVAL; | |
4347 | if (arg || nr_args) | |
4348 | break; | |
4349 | ret = io_sqe_files_unregister(ctx); | |
4350 | break; | |
c3a31e60 | 4351 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 4352 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 4353 | break; |
9b402849 JA |
4354 | case IORING_REGISTER_EVENTFD: |
4355 | ret = -EINVAL; | |
4356 | if (nr_args != 1) | |
4357 | break; | |
c75312dd UA |
4358 | ret = io_eventfd_register(ctx, arg, 0); |
4359 | break; | |
4360 | case IORING_REGISTER_EVENTFD_ASYNC: | |
4361 | ret = -EINVAL; | |
4362 | if (nr_args != 1) | |
f2842ab5 | 4363 | break; |
c75312dd | 4364 | ret = io_eventfd_register(ctx, arg, 1); |
9b402849 JA |
4365 | break; |
4366 | case IORING_UNREGISTER_EVENTFD: | |
4367 | ret = -EINVAL; | |
4368 | if (arg || nr_args) | |
4369 | break; | |
4370 | ret = io_eventfd_unregister(ctx); | |
4371 | break; | |
66f4af93 JA |
4372 | case IORING_REGISTER_PROBE: |
4373 | ret = -EINVAL; | |
4374 | if (!arg || nr_args > 256) | |
4375 | break; | |
4376 | ret = io_probe(ctx, arg, nr_args); | |
4377 | break; | |
071698e1 JA |
4378 | case IORING_REGISTER_PERSONALITY: |
4379 | ret = -EINVAL; | |
4380 | if (arg || nr_args) | |
4381 | break; | |
4382 | ret = io_register_personality(ctx); | |
4383 | break; | |
4384 | case IORING_UNREGISTER_PERSONALITY: | |
4385 | ret = -EINVAL; | |
4386 | if (arg) | |
4387 | break; | |
4388 | ret = io_unregister_personality(ctx, nr_args); | |
4389 | break; | |
7e84e1c7 SG |
4390 | case IORING_REGISTER_ENABLE_RINGS: |
4391 | ret = -EINVAL; | |
4392 | if (arg || nr_args) | |
4393 | break; | |
4394 | ret = io_register_enable_rings(ctx); | |
4395 | break; | |
21b55dbc SG |
4396 | case IORING_REGISTER_RESTRICTIONS: |
4397 | ret = io_register_restrictions(ctx, arg, nr_args); | |
4398 | break; | |
992da01a PB |
4399 | case IORING_REGISTER_FILES2: |
4400 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
4401 | break; | |
4402 | case IORING_REGISTER_FILES_UPDATE2: | |
4403 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
4404 | IORING_RSRC_FILE); | |
4405 | break; | |
4406 | case IORING_REGISTER_BUFFERS2: | |
4407 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 4408 | break; |
992da01a PB |
4409 | case IORING_REGISTER_BUFFERS_UPDATE: |
4410 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
4411 | IORING_RSRC_BUFFER); | |
c3bdad02 | 4412 | break; |
fe76421d JA |
4413 | case IORING_REGISTER_IOWQ_AFF: |
4414 | ret = -EINVAL; | |
4415 | if (!arg || !nr_args) | |
4416 | break; | |
4417 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
4418 | break; | |
4419 | case IORING_UNREGISTER_IOWQ_AFF: | |
4420 | ret = -EINVAL; | |
4421 | if (arg || nr_args) | |
4422 | break; | |
4423 | ret = io_unregister_iowq_aff(ctx); | |
4424 | break; | |
2e480058 JA |
4425 | case IORING_REGISTER_IOWQ_MAX_WORKERS: |
4426 | ret = -EINVAL; | |
4427 | if (!arg || nr_args != 2) | |
4428 | break; | |
4429 | ret = io_register_iowq_max_workers(ctx, arg); | |
4430 | break; | |
e7a6c00d JA |
4431 | case IORING_REGISTER_RING_FDS: |
4432 | ret = io_ringfd_register(ctx, arg, nr_args); | |
4433 | break; | |
4434 | case IORING_UNREGISTER_RING_FDS: | |
4435 | ret = io_ringfd_unregister(ctx, arg, nr_args); | |
4436 | break; | |
c7fb1942 JA |
4437 | case IORING_REGISTER_PBUF_RING: |
4438 | ret = -EINVAL; | |
4439 | if (!arg || nr_args != 1) | |
4440 | break; | |
4441 | ret = io_register_pbuf_ring(ctx, arg); | |
4442 | break; | |
4443 | case IORING_UNREGISTER_PBUF_RING: | |
4444 | ret = -EINVAL; | |
4445 | if (!arg || nr_args != 1) | |
4446 | break; | |
4447 | ret = io_unregister_pbuf_ring(ctx, arg); | |
4448 | break; | |
78a861b9 JA |
4449 | case IORING_REGISTER_SYNC_CANCEL: |
4450 | ret = -EINVAL; | |
4451 | if (!arg || nr_args != 1) | |
4452 | break; | |
4453 | ret = io_sync_cancel(ctx, arg); | |
4454 | break; | |
6e73dffb PB |
4455 | case IORING_REGISTER_FILE_ALLOC_RANGE: |
4456 | ret = -EINVAL; | |
4457 | if (!arg || nr_args) | |
4458 | break; | |
4459 | ret = io_register_file_alloc_range(ctx, arg); | |
4460 | break; | |
edafccee JA |
4461 | default: |
4462 | ret = -EINVAL; | |
4463 | break; | |
4464 | } | |
4465 | ||
edafccee JA |
4466 | return ret; |
4467 | } | |
4468 | ||
4469 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
4470 | void __user *, arg, unsigned int, nr_args) | |
4471 | { | |
4472 | struct io_ring_ctx *ctx; | |
4473 | long ret = -EBADF; | |
4474 | struct fd f; | |
7d3fd88d JT |
4475 | bool use_registered_ring; |
4476 | ||
4477 | use_registered_ring = !!(opcode & IORING_REGISTER_USE_REGISTERED_RING); | |
4478 | opcode &= ~IORING_REGISTER_USE_REGISTERED_RING; | |
edafccee | 4479 | |
34319084 JA |
4480 | if (opcode >= IORING_REGISTER_LAST) |
4481 | return -EINVAL; | |
4482 | ||
7d3fd88d JT |
4483 | if (use_registered_ring) { |
4484 | /* | |
4485 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
4486 | * need only dereference our task private array to find it. | |
4487 | */ | |
4488 | struct io_uring_task *tctx = current->io_uring; | |
edafccee | 4489 | |
7d3fd88d JT |
4490 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
4491 | return -EINVAL; | |
4492 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
4493 | f.file = tctx->registered_rings[fd]; | |
4494 | f.flags = 0; | |
4495 | if (unlikely(!f.file)) | |
4496 | return -EBADF; | |
4497 | } else { | |
4498 | f = fdget(fd); | |
4499 | if (unlikely(!f.file)) | |
4500 | return -EBADF; | |
4501 | ret = -EOPNOTSUPP; | |
4502 | if (!io_is_uring_fops(f.file)) | |
4503 | goto out_fput; | |
4504 | } | |
edafccee JA |
4505 | |
4506 | ctx = f.file->private_data; | |
4507 | ||
4508 | mutex_lock(&ctx->uring_lock); | |
4509 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
4510 | mutex_unlock(&ctx->uring_lock); | |
2757be22 | 4511 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); |
edafccee JA |
4512 | out_fput: |
4513 | fdput(f); | |
4514 | return ret; | |
4515 | } | |
4516 | ||
2b188cc1 JA |
4517 | static int __init io_uring_init(void) |
4518 | { | |
9c71d39a | 4519 | #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ |
d7f62e82 | 4520 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ |
9c71d39a | 4521 | BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ |
d7f62e82 SM |
4522 | } while (0) |
4523 | ||
4524 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9c71d39a SM |
4525 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) |
4526 | #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ | |
4527 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) | |
d7f62e82 SM |
4528 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); |
4529 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
4530 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
4531 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
4532 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
4533 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
4534 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9c71d39a SM |
4535 | BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); |
4536 | BUILD_BUG_SQE_ELEM(12, __u32, __pad1); | |
d7f62e82 | 4537 | BUILD_BUG_SQE_ELEM(16, __u64, addr); |
7d67af2c | 4538 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
4539 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
4540 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
4541 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
4542 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
4543 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
4544 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
4545 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
4546 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
4547 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
4548 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
4549 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
4550 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
4551 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
4552 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
4553 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 4554 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
9c71d39a SM |
4555 | BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); |
4556 | BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); | |
4557 | BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); | |
4558 | BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); | |
4559 | BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); | |
d7f62e82 SM |
4560 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
4561 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 4562 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 4563 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 4564 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 4565 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
b48c312b PB |
4566 | BUILD_BUG_SQE_ELEM(44, __u16, addr_len); |
4567 | BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]); | |
e9621e2b | 4568 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
9c71d39a SM |
4569 | BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); |
4570 | BUILD_BUG_SQE_ELEM(56, __u64, __pad2); | |
d7f62e82 | 4571 | |
b0d658ec PB |
4572 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
4573 | sizeof(struct io_uring_rsrc_update)); | |
4574 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
4575 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
4576 | |
4577 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
4578 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
4579 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
4580 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 4581 | |
b0d658ec PB |
4582 | /* should fit into one byte */ |
4583 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
4584 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
4585 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 4586 | |
32c2d33e | 4587 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); |
16340eab | 4588 | |
3a4b89a2 JA |
4589 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
4590 | ||
d9b57aa3 | 4591 | io_uring_optable_init(); |
0702e536 | 4592 | |
91f245d5 | 4593 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | |
8751d154 | 4594 | SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU); |
2b188cc1 JA |
4595 | return 0; |
4596 | }; | |
4597 | __initcall(io_uring_init); |