Commit | Line | Data |
---|---|---|
2b188cc1 JA |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* | |
3 | * Shared application/kernel submission and completion ring pairs, for | |
4 | * supporting fast/efficient IO. | |
5 | * | |
6 | * A note on the read/write ordering memory barriers that are matched between | |
1e84b97b SB |
7 | * the application and kernel side. |
8 | * | |
9 | * After the application reads the CQ ring tail, it must use an | |
10 | * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses | |
11 | * before writing the tail (using smp_load_acquire to read the tail will | |
12 | * do). It also needs a smp_mb() before updating CQ head (ordering the | |
13 | * entry load(s) with the head store), pairing with an implicit barrier | |
d068b506 | 14 | * through a control-dependency in io_get_cqe (smp_store_release to |
1e84b97b SB |
15 | * store head will do). Failure to do so could lead to reading invalid |
16 | * CQ entries. | |
17 | * | |
18 | * Likewise, the application must use an appropriate smp_wmb() before | |
19 | * writing the SQ tail (ordering SQ entry stores with the tail store), | |
20 | * which pairs with smp_load_acquire in io_get_sqring (smp_store_release | |
21 | * to store the tail will do). And it needs a barrier ordering the SQ | |
22 | * head load before writing new SQ entries (smp_load_acquire to read | |
23 | * head will do). | |
24 | * | |
25 | * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application | |
26 | * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after* | |
27 | * updating the SQ tail; a full memory barrier smp_mb() is needed | |
28 | * between. | |
2b188cc1 JA |
29 | * |
30 | * Also see the examples in the liburing library: | |
31 | * | |
32 | * git://git.kernel.dk/liburing | |
33 | * | |
34 | * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens | |
35 | * from data shared between the kernel and application. This is done both | |
36 | * for ordering purposes, but also to ensure that once a value is loaded from | |
37 | * data that the application could potentially modify, it remains stable. | |
38 | * | |
39 | * Copyright (C) 2018-2019 Jens Axboe | |
c992fe29 | 40 | * Copyright (c) 2018-2019 Christoph Hellwig |
2b188cc1 JA |
41 | */ |
42 | #include <linux/kernel.h> | |
43 | #include <linux/init.h> | |
44 | #include <linux/errno.h> | |
45 | #include <linux/syscalls.h> | |
52de1fe1 | 46 | #include <net/compat.h> |
2b188cc1 JA |
47 | #include <linux/refcount.h> |
48 | #include <linux/uio.h> | |
6b47ee6e | 49 | #include <linux/bits.h> |
2b188cc1 JA |
50 | |
51 | #include <linux/sched/signal.h> | |
52 | #include <linux/fs.h> | |
53 | #include <linux/file.h> | |
54 | #include <linux/fdtable.h> | |
55 | #include <linux/mm.h> | |
56 | #include <linux/mman.h> | |
2b188cc1 JA |
57 | #include <linux/percpu.h> |
58 | #include <linux/slab.h> | |
edafccee | 59 | #include <linux/bvec.h> |
2b188cc1 JA |
60 | #include <linux/net.h> |
61 | #include <net/sock.h> | |
62 | #include <net/af_unix.h> | |
6b06314c | 63 | #include <net/scm.h> |
2b188cc1 JA |
64 | #include <linux/anon_inodes.h> |
65 | #include <linux/sched/mm.h> | |
66 | #include <linux/uaccess.h> | |
67 | #include <linux/nospec.h> | |
aa4c3967 | 68 | #include <linux/highmem.h> |
15b71abe | 69 | #include <linux/fsnotify.h> |
4840e418 | 70 | #include <linux/fadvise.h> |
b41e9852 | 71 | #include <linux/task_work.h> |
0f212204 | 72 | #include <linux/io_uring.h> |
5bd2182d | 73 | #include <linux/audit.h> |
cdc1404a | 74 | #include <linux/security.h> |
2b188cc1 | 75 | |
c826bd7a DD |
76 | #define CREATE_TRACE_POINTS |
77 | #include <trace/events/io_uring.h> | |
78 | ||
2b188cc1 JA |
79 | #include <uapi/linux/io_uring.h> |
80 | ||
561fb04a | 81 | #include "io-wq.h" |
2b188cc1 | 82 | |
de23077e | 83 | #include "io_uring.h" |
329061d3 | 84 | #include "opdef.h" |
e418bbc9 | 85 | #include "refs.h" |
c9f06aa7 | 86 | #include "tctx.h" |
17437f31 | 87 | #include "sqpoll.h" |
a4ad4f74 | 88 | #include "fdinfo.h" |
3b77495a | 89 | #include "kbuf.h" |
73572984 | 90 | #include "rsrc.h" |
38513c46 | 91 | #include "cancel.h" |
e27f928e | 92 | |
59915143 | 93 | #include "timeout.h" |
329061d3 | 94 | #include "poll.h" |
5e2a18d9 | 95 | |
5277deaa | 96 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 97 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 | 98 | |
21b55dbc SG |
99 | #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \ |
100 | IORING_REGISTER_LAST + IORING_OP_LAST) | |
2b188cc1 | 101 | |
68fe256a PB |
102 | #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \ |
103 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) | |
104 | ||
5562a8d7 PB |
105 | #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \ |
106 | IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS) | |
68fe256a | 107 | |
c854357b | 108 | #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ |
9cae36a0 JA |
109 | REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \ |
110 | REQ_F_ASYNC_DATA) | |
b16fed66 | 111 | |
a538be5b PB |
112 | #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\ |
113 | IO_REQ_CLEAN_FLAGS) | |
114 | ||
09899b19 PB |
115 | #define IO_TCTX_REFS_CACHE_NR (1U << 10) |
116 | ||
6dd0be1e | 117 | #define IO_COMPL_BATCH 32 |
bf019da7 | 118 | #define IO_REQ_ALLOC_BATCH 8 |
258b29a9 | 119 | |
10988a0a DY |
120 | enum { |
121 | IO_CHECK_CQ_OVERFLOW_BIT, | |
155bc950 | 122 | IO_CHECK_CQ_DROPPED_BIT, |
10988a0a DY |
123 | }; |
124 | ||
27dc8338 PB |
125 | struct io_defer_entry { |
126 | struct list_head list; | |
127 | struct io_kiocb *req; | |
9cf7c104 | 128 | u32 seq; |
2b188cc1 JA |
129 | }; |
130 | ||
0756a869 PB |
131 | /* requests with any of those set should undergo io_disarm_next() */ |
132 | #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) | |
da1a08c5 | 133 | #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK) |
0756a869 | 134 | |
affa87db | 135 | static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
9936c7c2 | 136 | struct task_struct *task, |
3dd0c97a | 137 | bool cancel_all); |
1ffc5422 | 138 | |
c7dae4ba | 139 | static void io_dismantle_req(struct io_kiocb *req); |
68fb8979 | 140 | static void io_clean_op(struct io_kiocb *req); |
cbc2e203 | 141 | static void io_queue_sqe(struct io_kiocb *req); |
de0617e4 | 142 | |
c450178d | 143 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx); |
de0617e4 | 144 | |
2b188cc1 JA |
145 | static struct kmem_cache *req_cachep; |
146 | ||
2b188cc1 JA |
147 | struct sock *io_uring_get_socket(struct file *file) |
148 | { | |
149 | #if defined(CONFIG_UNIX) | |
cd40cae2 | 150 | if (io_is_uring_fops(file)) { |
2b188cc1 JA |
151 | struct io_ring_ctx *ctx = file->private_data; |
152 | ||
153 | return ctx->ring_sock->sk; | |
154 | } | |
155 | #endif | |
156 | return NULL; | |
157 | } | |
158 | EXPORT_SYMBOL(io_uring_get_socket); | |
159 | ||
c450178d PB |
160 | static inline void io_submit_flush_completions(struct io_ring_ctx *ctx) |
161 | { | |
6f33b0bc | 162 | if (!wq_list_empty(&ctx->submit_state.compl_reqs)) |
c450178d PB |
163 | __io_submit_flush_completions(ctx); |
164 | } | |
165 | ||
faf88dde PB |
166 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
167 | { | |
168 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
169 | } | |
170 | ||
9cae36a0 JA |
171 | static bool io_match_linked(struct io_kiocb *head) |
172 | { | |
173 | struct io_kiocb *req; | |
174 | ||
175 | io_for_each_link(req, head) { | |
176 | if (req->flags & REQ_F_INFLIGHT) | |
177 | return true; | |
178 | } | |
179 | return false; | |
6af3f48b PB |
180 | } |
181 | ||
182 | /* | |
183 | * As io_match_task() but protected against racing with linked timeouts. | |
184 | * User must not hold timeout_lock. | |
185 | */ | |
329061d3 JA |
186 | bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, |
187 | bool cancel_all) | |
6af3f48b | 188 | { |
9cae36a0 JA |
189 | bool matched; |
190 | ||
6af3f48b PB |
191 | if (task && head->task != task) |
192 | return false; | |
9cae36a0 JA |
193 | if (cancel_all) |
194 | return true; | |
195 | ||
196 | if (head->flags & REQ_F_LINK_TIMEOUT) { | |
197 | struct io_ring_ctx *ctx = head->ctx; | |
198 | ||
199 | /* protect against races with linked timeouts */ | |
200 | spin_lock_irq(&ctx->timeout_lock); | |
201 | matched = io_match_linked(head); | |
202 | spin_unlock_irq(&ctx->timeout_lock); | |
203 | } else { | |
204 | matched = io_match_linked(head); | |
205 | } | |
206 | return matched; | |
6af3f48b PB |
207 | } |
208 | ||
a8295b98 HX |
209 | static inline void req_fail_link_node(struct io_kiocb *req, int res) |
210 | { | |
211 | req_set_fail(req); | |
97b388d7 | 212 | io_req_set_res(req, res, 0); |
a8295b98 HX |
213 | } |
214 | ||
fa05457a PB |
215 | static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) |
216 | { | |
217 | wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); | |
a8295b98 HX |
218 | } |
219 | ||
c072481d | 220 | static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) |
2b188cc1 JA |
221 | { |
222 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
223 | ||
0f158b4c | 224 | complete(&ctx->ref_comp); |
2b188cc1 JA |
225 | } |
226 | ||
c072481d | 227 | static __cold void io_fallback_req_func(struct work_struct *work) |
f56165e6 PB |
228 | { |
229 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
230 | fallback_work.work); | |
231 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
232 | struct io_kiocb *req, *tmp; | |
f237c30a | 233 | bool locked = false; |
f56165e6 PB |
234 | |
235 | percpu_ref_get(&ctx->refs); | |
236 | llist_for_each_entry_safe(req, tmp, node, io_task_work.fallback_node) | |
f237c30a | 237 | req->io_task_work.func(req, &locked); |
5636c00d | 238 | |
f237c30a | 239 | if (locked) { |
c450178d | 240 | io_submit_flush_completions(ctx); |
f237c30a PB |
241 | mutex_unlock(&ctx->uring_lock); |
242 | } | |
f56165e6 PB |
243 | percpu_ref_put(&ctx->refs); |
244 | } | |
245 | ||
e6f89be6 PB |
246 | static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) |
247 | { | |
248 | unsigned hash_buckets = 1U << bits; | |
249 | size_t hash_size = hash_buckets * sizeof(table->hbs[0]); | |
250 | ||
251 | table->hbs = kmalloc(hash_size, GFP_KERNEL); | |
252 | if (!table->hbs) | |
253 | return -ENOMEM; | |
254 | ||
255 | table->hash_bits = bits; | |
256 | init_hash_table(table, hash_buckets); | |
257 | return 0; | |
258 | } | |
259 | ||
c072481d | 260 | static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
2b188cc1 JA |
261 | { |
262 | struct io_ring_ctx *ctx; | |
9cfc7e94 | 263 | int hash_bits; |
2b188cc1 JA |
264 | |
265 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
266 | if (!ctx) | |
267 | return NULL; | |
268 | ||
9cfc7e94 JA |
269 | xa_init(&ctx->io_bl_xa); |
270 | ||
78076bb6 JA |
271 | /* |
272 | * Use 5 bits less than the max cq entries, that should give us around | |
4a07723f PB |
273 | * 32 entries per hash list if totally full and uniformly spread, but |
274 | * don't keep too many buckets to not overconsume memory. | |
78076bb6 | 275 | */ |
4a07723f PB |
276 | hash_bits = ilog2(p->cq_entries) - 5; |
277 | hash_bits = clamp(hash_bits, 1, 8); | |
e6f89be6 | 278 | if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) |
78076bb6 | 279 | goto err; |
9ca9fb24 PB |
280 | if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) |
281 | goto err; | |
38513c46 | 282 | |
6224843d PB |
283 | ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL); |
284 | if (!ctx->dummy_ubuf) | |
285 | goto err; | |
286 | /* set invalid range, so io_import_fixed() fails meeting it */ | |
287 | ctx->dummy_ubuf->ubuf = -1UL; | |
288 | ||
21482896 | 289 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
290 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
291 | goto err; | |
2b188cc1 JA |
292 | |
293 | ctx->flags = p->flags; | |
90554200 | 294 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 295 | INIT_LIST_HEAD(&ctx->sqd_list); |
1d7bb1d5 | 296 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
cc3cec83 | 297 | INIT_LIST_HEAD(&ctx->io_buffers_cache); |
4d9237e3 | 298 | INIT_LIST_HEAD(&ctx->apoll_cache); |
0f158b4c | 299 | init_completion(&ctx->ref_comp); |
61cf9370 | 300 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 301 | mutex_init(&ctx->uring_lock); |
311997b3 | 302 | init_waitqueue_head(&ctx->cq_wait); |
2b188cc1 | 303 | spin_lock_init(&ctx->completion_lock); |
89850fce | 304 | spin_lock_init(&ctx->timeout_lock); |
5eef4e87 | 305 | INIT_WQ_LIST(&ctx->iopoll_list); |
cc3cec83 JA |
306 | INIT_LIST_HEAD(&ctx->io_buffers_pages); |
307 | INIT_LIST_HEAD(&ctx->io_buffers_comp); | |
de0617e4 | 308 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 309 | INIT_LIST_HEAD(&ctx->timeout_list); |
ef9dd637 | 310 | INIT_LIST_HEAD(&ctx->ltimeout_list); |
d67d2263 BM |
311 | spin_lock_init(&ctx->rsrc_ref_lock); |
312 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); | |
269bbe5f BM |
313 | INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work); |
314 | init_llist_head(&ctx->rsrc_put_llist); | |
13bf43f5 | 315 | INIT_LIST_HEAD(&ctx->tctx_list); |
c2b6c6bc PB |
316 | ctx->submit_state.free_list.next = NULL; |
317 | INIT_WQ_LIST(&ctx->locked_free_list); | |
9011bf9a | 318 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
6f33b0bc | 319 | INIT_WQ_LIST(&ctx->submit_state.compl_reqs); |
2b188cc1 | 320 | return ctx; |
206aefde | 321 | err: |
6224843d | 322 | kfree(ctx->dummy_ubuf); |
e6f89be6 | 323 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 324 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 JA |
325 | kfree(ctx->io_bl); |
326 | xa_destroy(&ctx->io_bl_xa); | |
206aefde JA |
327 | kfree(ctx); |
328 | return NULL; | |
2b188cc1 JA |
329 | } |
330 | ||
8f6ed49a PB |
331 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
332 | { | |
333 | struct io_rings *r = ctx->rings; | |
334 | ||
335 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
336 | ctx->cq_extra--; | |
337 | } | |
338 | ||
9cf7c104 | 339 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 340 | { |
2bc9930e JA |
341 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
342 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 343 | |
8f6ed49a | 344 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 345 | } |
de0617e4 | 346 | |
9d858b21 | 347 | return false; |
de0617e4 JA |
348 | } |
349 | ||
9cae36a0 JA |
350 | static inline void io_req_track_inflight(struct io_kiocb *req) |
351 | { | |
352 | if (!(req->flags & REQ_F_INFLIGHT)) { | |
353 | req->flags |= REQ_F_INFLIGHT; | |
386e4fb6 | 354 | atomic_inc(&req->task->io_uring->inflight_tracked); |
9cae36a0 JA |
355 | } |
356 | } | |
357 | ||
fd08e530 PB |
358 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
359 | { | |
906c6caa PB |
360 | if (WARN_ON_ONCE(!req->link)) |
361 | return NULL; | |
362 | ||
4d13d1a4 PB |
363 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
364 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
365 | |
366 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 367 | io_req_set_refcount(req); |
4d13d1a4 PB |
368 | __io_req_set_refcount(req->link, 2); |
369 | return req->link; | |
fd08e530 PB |
370 | } |
371 | ||
372 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
373 | { | |
4d13d1a4 | 374 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
375 | return NULL; |
376 | return __io_prep_linked_timeout(req); | |
377 | } | |
378 | ||
cb2d344c PB |
379 | static noinline void __io_arm_ltimeout(struct io_kiocb *req) |
380 | { | |
381 | io_queue_linked_timeout(__io_prep_linked_timeout(req)); | |
382 | } | |
383 | ||
384 | static inline void io_arm_ltimeout(struct io_kiocb *req) | |
385 | { | |
386 | if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) | |
387 | __io_arm_ltimeout(req); | |
388 | } | |
389 | ||
1e6fa521 JA |
390 | static void io_prep_async_work(struct io_kiocb *req) |
391 | { | |
392 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
1e6fa521 JA |
393 | struct io_ring_ctx *ctx = req->ctx; |
394 | ||
b8e64b53 PB |
395 | if (!(req->flags & REQ_F_CREDS)) { |
396 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 397 | req->creds = get_current_cred(); |
b8e64b53 | 398 | } |
003e8dcc | 399 | |
e1d675df PB |
400 | req->work.list.next = NULL; |
401 | req->work.flags = 0; | |
8e29da69 | 402 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); |
feaadc4f PB |
403 | if (req->flags & REQ_F_FORCE_ASYNC) |
404 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
405 | ||
1e6fa521 JA |
406 | if (req->flags & REQ_F_ISREG) { |
407 | if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL)) | |
408 | io_wq_hash_work(&req->work, file_inode(req->file)); | |
4b982bd0 | 409 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
410 | if (def->unbound_nonreg_file) |
411 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
412 | } | |
561fb04a | 413 | } |
cccf0ee8 | 414 | |
cbdcb435 | 415 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 416 | { |
cbdcb435 | 417 | struct io_kiocb *cur; |
54a91f3b | 418 | |
44eff40a PB |
419 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
420 | struct io_ring_ctx *ctx = req->ctx; | |
421 | ||
674ee8e1 | 422 | spin_lock_irq(&ctx->timeout_lock); |
44eff40a PB |
423 | io_for_each_link(cur, req) |
424 | io_prep_async_work(cur); | |
674ee8e1 | 425 | spin_unlock_irq(&ctx->timeout_lock); |
44eff40a PB |
426 | } else { |
427 | io_for_each_link(cur, req) | |
428 | io_prep_async_work(cur); | |
429 | } | |
561fb04a JA |
430 | } |
431 | ||
f3b44f92 | 432 | void io_queue_iowq(struct io_kiocb *req, bool *dont_use) |
561fb04a | 433 | { |
cbdcb435 | 434 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 435 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 436 | |
3bfe6106 JA |
437 | BUG_ON(!tctx); |
438 | BUG_ON(!tctx->io_wq); | |
561fb04a | 439 | |
cbdcb435 PB |
440 | /* init ->work of the whole link before punting */ |
441 | io_prep_async_link(req); | |
991468dc JA |
442 | |
443 | /* | |
444 | * Not expected to happen, but if we do have a bug where this _can_ | |
445 | * happen, catch it here and ensure the request is marked as | |
446 | * canceled. That will make io-wq go through the usual work cancel | |
447 | * procedure rather than attempt to run this request (or create a new | |
448 | * worker for it). | |
449 | */ | |
450 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
451 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
452 | ||
48863ffd | 453 | trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); |
ebf93667 | 454 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
455 | if (link) |
456 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
457 | } |
458 | ||
c072481d | 459 | static __cold void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 460 | { |
441b8a78 | 461 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
462 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
463 | struct io_defer_entry, list); | |
de0617e4 | 464 | |
9cf7c104 | 465 | if (req_need_defer(de->req, de->seq)) |
04518945 | 466 | break; |
27dc8338 | 467 | list_del_init(&de->list); |
907d1df3 | 468 | io_req_task_queue(de->req); |
27dc8338 | 469 | kfree(de); |
441b8a78 | 470 | } |
04518945 PB |
471 | } |
472 | ||
77bc59b4 | 473 | static void io_eventfd_signal(struct io_ring_ctx *ctx) |
f2842ab5 | 474 | { |
77bc59b4 UA |
475 | struct io_ev_fd *ev_fd; |
476 | ||
77bc59b4 UA |
477 | rcu_read_lock(); |
478 | /* | |
479 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
480 | * and eventfd_signal | |
481 | */ | |
482 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
483 | ||
484 | /* | |
485 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
486 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
487 | * the function and rcu_read_lock. | |
488 | */ | |
489 | if (unlikely(!ev_fd)) | |
490 | goto out; | |
7e55a19c | 491 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 UA |
492 | goto out; |
493 | ||
c75312dd | 494 | if (!ev_fd->eventfd_async || io_wq_current_is_worker()) |
77bc59b4 | 495 | eventfd_signal(ev_fd->cq_ev_fd, 1); |
77bc59b4 UA |
496 | out: |
497 | rcu_read_unlock(); | |
f2842ab5 JA |
498 | } |
499 | ||
a830ffd2 PB |
500 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
501 | { | |
502 | if (ctx->off_timeout_used || ctx->drain_active) { | |
503 | spin_lock(&ctx->completion_lock); | |
504 | if (ctx->off_timeout_used) | |
505 | io_flush_timeouts(ctx); | |
506 | if (ctx->drain_active) | |
507 | io_queue_deferred(ctx); | |
508 | spin_unlock(&ctx->completion_lock); | |
509 | } | |
510 | if (ctx->has_evfd) | |
511 | io_eventfd_signal(ctx); | |
512 | } | |
513 | ||
2c5d763c JA |
514 | /* |
515 | * This should only get called when at least one event has been posted. | |
516 | * Some applications rely on the eventfd notification count only changing | |
517 | * IFF a new CQE has been added to the CQ ring. There's no depedency on | |
518 | * 1:1 relationship between how many times this function is called (and | |
519 | * hence the eventfd count) and number of CQEs posted to the CQ ring. | |
520 | */ | |
f9ead18c | 521 | void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 | 522 | { |
9aa8dfde PB |
523 | if (unlikely(ctx->off_timeout_used || ctx->drain_active || |
524 | ctx->has_evfd)) | |
9333f6b4 PB |
525 | __io_commit_cqring_flush(ctx); |
526 | ||
9aa8dfde | 527 | io_cqring_wake(ctx); |
1d7bb1d5 JA |
528 | } |
529 | ||
c4a2ed72 | 530 | /* Returns true if there are no backlogged entries after the flush */ |
6c2450ae | 531 | static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) |
1d7bb1d5 | 532 | { |
b18032bb | 533 | bool all_flushed, posted; |
e45a3e05 | 534 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 535 | |
a566c556 | 536 | if (!force && __io_cqring_events(ctx) == ctx->cq_entries) |
e23de15f | 537 | return false; |
1d7bb1d5 | 538 | |
e45a3e05 SR |
539 | if (ctx->flags & IORING_SETUP_CQE32) |
540 | cqe_size <<= 1; | |
541 | ||
b18032bb | 542 | posted = false; |
79ebeaee | 543 | spin_lock(&ctx->completion_lock); |
6c2450ae | 544 | while (!list_empty(&ctx->cq_overflow_list)) { |
d068b506 | 545 | struct io_uring_cqe *cqe = io_get_cqe(ctx); |
6c2450ae | 546 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 547 | |
1d7bb1d5 JA |
548 | if (!cqe && !force) |
549 | break; | |
6c2450ae PB |
550 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
551 | struct io_overflow_cqe, list); | |
552 | if (cqe) | |
e45a3e05 | 553 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae | 554 | else |
8f6ed49a PB |
555 | io_account_cq_overflow(ctx); |
556 | ||
b18032bb | 557 | posted = true; |
6c2450ae PB |
558 | list_del(&ocqe->list); |
559 | kfree(ocqe); | |
1d7bb1d5 JA |
560 | } |
561 | ||
09e88404 PB |
562 | all_flushed = list_empty(&ctx->cq_overflow_list); |
563 | if (all_flushed) { | |
10988a0a | 564 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 565 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 566 | } |
46930143 | 567 | |
60053be8 | 568 | io_commit_cqring(ctx); |
79ebeaee | 569 | spin_unlock(&ctx->completion_lock); |
b18032bb JA |
570 | if (posted) |
571 | io_cqring_ev_posted(ctx); | |
09e88404 | 572 | return all_flushed; |
1d7bb1d5 JA |
573 | } |
574 | ||
90f67366 | 575 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 576 | { |
ca0a2651 JA |
577 | bool ret = true; |
578 | ||
10988a0a | 579 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
6c503150 PB |
580 | /* iopoll syncs against uring_lock, not completion_lock */ |
581 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
582 | mutex_lock(&ctx->uring_lock); | |
90f67366 | 583 | ret = __io_cqring_overflow_flush(ctx, false); |
6c503150 PB |
584 | if (ctx->flags & IORING_SETUP_IOPOLL) |
585 | mutex_unlock(&ctx->uring_lock); | |
586 | } | |
ca0a2651 JA |
587 | |
588 | return ret; | |
6c503150 PB |
589 | } |
590 | ||
9d170164 | 591 | static void __io_put_task(struct task_struct *task, int nr) |
6a290a14 PB |
592 | { |
593 | struct io_uring_task *tctx = task->io_uring; | |
594 | ||
9d170164 PB |
595 | percpu_counter_sub(&tctx->inflight, nr); |
596 | if (unlikely(atomic_read(&tctx->in_idle))) | |
597 | wake_up(&tctx->wait); | |
598 | put_task_struct_many(task, nr); | |
599 | } | |
600 | ||
601 | /* must to be called somewhat shortly after putting a request */ | |
602 | static inline void io_put_task(struct task_struct *task, int nr) | |
603 | { | |
604 | if (likely(task == current)) | |
605 | task->io_uring->cached_refs += nr; | |
606 | else | |
607 | __io_put_task(task, nr); | |
6a290a14 PB |
608 | } |
609 | ||
9a10867a PB |
610 | static void io_task_refs_refill(struct io_uring_task *tctx) |
611 | { | |
612 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
613 | ||
614 | percpu_counter_add(&tctx->inflight, refill); | |
615 | refcount_add(refill, ¤t->usage); | |
616 | tctx->cached_refs += refill; | |
617 | } | |
618 | ||
619 | static inline void io_get_task_refs(int nr) | |
620 | { | |
621 | struct io_uring_task *tctx = current->io_uring; | |
622 | ||
623 | tctx->cached_refs -= nr; | |
624 | if (unlikely(tctx->cached_refs < 0)) | |
625 | io_task_refs_refill(tctx); | |
626 | } | |
627 | ||
3cc7fdb9 PB |
628 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
629 | { | |
630 | struct io_uring_task *tctx = task->io_uring; | |
631 | unsigned int refs = tctx->cached_refs; | |
632 | ||
633 | if (refs) { | |
634 | tctx->cached_refs = 0; | |
635 | percpu_counter_sub(&tctx->inflight, refs); | |
636 | put_task_struct_many(task, refs); | |
637 | } | |
638 | } | |
639 | ||
68494a65 PB |
640 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
641 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 642 | { |
cce4b8b0 | 643 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
644 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
645 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 646 | |
e45a3e05 SR |
647 | if (is_cqe32) |
648 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 649 | |
e45a3e05 | 650 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 651 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
652 | if (!ocqe) { |
653 | /* | |
654 | * If we're in ring overflow flush mode, or in task cancel mode, | |
655 | * or cannot allocate an overflow entry, then we need to drop it | |
656 | * on the floor. | |
657 | */ | |
8f6ed49a | 658 | io_account_cq_overflow(ctx); |
155bc950 | 659 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 660 | return false; |
2b188cc1 | 661 | } |
cce4b8b0 | 662 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 663 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 664 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 665 | |
cce4b8b0 | 666 | } |
d4d19c19 | 667 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
668 | ocqe->cqe.res = res; |
669 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
670 | if (is_cqe32) { |
671 | ocqe->cqe.big_cqe[0] = extra1; | |
672 | ocqe->cqe.big_cqe[1] = extra2; | |
673 | } | |
cce4b8b0 PB |
674 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
675 | return true; | |
2b188cc1 JA |
676 | } |
677 | ||
68494a65 PB |
678 | bool io_req_cqe_overflow(struct io_kiocb *req) |
679 | { | |
680 | if (!(req->flags & REQ_F_CQE32_INIT)) { | |
681 | req->extra1 = 0; | |
682 | req->extra2 = 0; | |
683 | } | |
684 | return io_cqring_event_overflow(req->ctx, req->cqe.user_data, | |
685 | req->cqe.res, req->cqe.flags, | |
686 | req->extra1, req->extra2); | |
687 | } | |
688 | ||
faf88dde PB |
689 | /* |
690 | * writes to the cq entry need to come after reading head; the | |
691 | * control dependency is enough as we're using WRITE_ONCE to | |
692 | * fill the cq entry | |
693 | */ | |
694 | struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx) | |
695 | { | |
696 | struct io_rings *rings = ctx->rings; | |
697 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
698 | unsigned int free, queued, len; |
699 | ||
faf88dde PB |
700 | |
701 | /* userspace may cheat modifying the tail, be safe and do min */ | |
702 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
703 | free = ctx->cq_entries - queued; | |
704 | /* we need a contiguous range, limit based on the current array offset */ | |
705 | len = min(free, ctx->cq_entries - off); | |
706 | if (!len) | |
707 | return NULL; | |
708 | ||
b3659a65 PB |
709 | if (ctx->flags & IORING_SETUP_CQE32) { |
710 | off <<= 1; | |
711 | len <<= 1; | |
712 | } | |
713 | ||
faf88dde PB |
714 | ctx->cqe_cached = &rings->cqes[off]; |
715 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
b3659a65 PB |
716 | |
717 | ctx->cached_cq_tail++; | |
faf88dde | 718 | ctx->cqe_cached++; |
b3659a65 PB |
719 | if (ctx->flags & IORING_SETUP_CQE32) |
720 | ctx->cqe_cached++; | |
721 | return &rings->cqes[off]; | |
faf88dde PB |
722 | } |
723 | ||
d245bca6 PB |
724 | static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, |
725 | u64 user_data, s32 res, u32 cflags) | |
bcda7baa | 726 | { |
cd94903d PB |
727 | struct io_uring_cqe *cqe; |
728 | ||
913a571a | 729 | ctx->cq_extra++; |
c4bb964f | 730 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
cd94903d PB |
731 | |
732 | /* | |
733 | * If we can't get a cq entry, userspace overflowed the | |
734 | * submission (by quite a lot). Increment the overflow count in | |
735 | * the ring. | |
736 | */ | |
737 | cqe = io_get_cqe(ctx); | |
738 | if (likely(cqe)) { | |
739 | WRITE_ONCE(cqe->user_data, user_data); | |
740 | WRITE_ONCE(cqe->res, res); | |
741 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
742 | |
743 | if (ctx->flags & IORING_SETUP_CQE32) { | |
744 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
745 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
746 | } | |
cd94903d PB |
747 | return true; |
748 | } | |
749 | return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); | |
bcda7baa JA |
750 | } |
751 | ||
d245bca6 PB |
752 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, |
753 | u64 user_data, s32 res, u32 cflags) | |
754 | { | |
755 | bool filled; | |
756 | ||
757 | spin_lock(&ctx->completion_lock); | |
758 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags); | |
759 | io_commit_cqring(ctx); | |
760 | spin_unlock(&ctx->completion_lock); | |
761 | if (filled) | |
762 | io_cqring_ev_posted(ctx); | |
763 | return filled; | |
764 | } | |
765 | ||
effcf8bd | 766 | static void __io_req_complete_put(struct io_kiocb *req) |
2b188cc1 | 767 | { |
c7dae4ba JA |
768 | /* |
769 | * If we're the last reference to this request, add to our locked | |
770 | * free_list cache. | |
771 | */ | |
de9b4cca | 772 | if (req_ref_put_and_test(req)) { |
effcf8bd SR |
773 | struct io_ring_ctx *ctx = req->ctx; |
774 | ||
da1a08c5 | 775 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 776 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
777 | io_disarm_next(req); |
778 | if (req->link) { | |
779 | io_req_task_queue(req->link); | |
780 | req->link = NULL; | |
781 | } | |
782 | } | |
7ac1edc4 | 783 | io_req_put_rsrc(req); |
8197b053 PB |
784 | /* |
785 | * Selected buffer deallocation in io_clean_op() assumes that | |
786 | * we don't hold ->completion_lock. Clean them here to avoid | |
787 | * deadlocks. | |
788 | */ | |
789 | io_put_kbuf_comp(req); | |
c7dae4ba JA |
790 | io_dismantle_req(req); |
791 | io_put_task(req->task, 1); | |
c2b6c6bc | 792 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 793 | ctx->locked_free_nr++; |
180f829f | 794 | } |
a37fae8a HX |
795 | } |
796 | ||
59915143 | 797 | void __io_req_complete_post(struct io_kiocb *req) |
effcf8bd | 798 | { |
97b388d7 | 799 | if (!(req->flags & REQ_F_CQE_SKIP)) |
91ef75a7 | 800 | __io_fill_cqe_req(req->ctx, req); |
effcf8bd SR |
801 | __io_req_complete_put(req); |
802 | } | |
803 | ||
59915143 | 804 | void io_req_complete_post(struct io_kiocb *req) |
a37fae8a HX |
805 | { |
806 | struct io_ring_ctx *ctx = req->ctx; | |
807 | ||
808 | spin_lock(&ctx->completion_lock); | |
97b388d7 | 809 | __io_req_complete_post(req); |
7a612350 | 810 | io_commit_cqring(ctx); |
79ebeaee | 811 | spin_unlock(&ctx->completion_lock); |
a3f34907 | 812 | io_cqring_ev_posted(ctx); |
4e3d9ff9 JA |
813 | } |
814 | ||
99f15d8d | 815 | inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags) |
bcda7baa | 816 | { |
75d7b3ae | 817 | io_req_complete_post(req); |
0ddf92e8 JA |
818 | } |
819 | ||
329061d3 | 820 | void io_req_complete_failed(struct io_kiocb *req, s32 res) |
f41db273 | 821 | { |
93d2bcd2 | 822 | req_set_fail(req); |
97b388d7 JA |
823 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
824 | io_req_complete_post(req); | |
f41db273 PB |
825 | } |
826 | ||
864ea921 PB |
827 | /* |
828 | * Don't initialise the fields below on every allocation, but do that in | |
829 | * advance and keep them valid across allocations. | |
830 | */ | |
831 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
832 | { | |
833 | req->ctx = ctx; | |
834 | req->link = NULL; | |
835 | req->async_data = NULL; | |
836 | /* not necessary, but safer to zero */ | |
cef216fc | 837 | req->cqe.res = 0; |
864ea921 PB |
838 | } |
839 | ||
dac7a098 | 840 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 841 | struct io_submit_state *state) |
dac7a098 | 842 | { |
79ebeaee | 843 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 844 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 845 | ctx->locked_free_nr = 0; |
79ebeaee | 846 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
847 | } |
848 | ||
88ab95be | 849 | static inline bool io_req_cache_empty(struct io_ring_ctx *ctx) |
0ddf92e8 | 850 | { |
88ab95be | 851 | return !ctx->submit_state.free_list.next; |
0ddf92e8 JA |
852 | } |
853 | ||
5d5901a3 PB |
854 | /* |
855 | * A request might get retired back into the request caches even before opcode | |
856 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
857 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
858 | * and with extra caution to not get a request that is still worked on. | |
859 | */ | |
c072481d | 860 | static __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 861 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 862 | { |
864ea921 | 863 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 864 | void *reqs[IO_REQ_ALLOC_BATCH]; |
864ea921 | 865 | int ret, i; |
e5d1bc0a | 866 | |
23a5c43b PB |
867 | /* |
868 | * If we have more than a batch's worth of requests in our IRQ side | |
869 | * locked cache, grab the lock and move them over to our submission | |
870 | * side cache. | |
871 | */ | |
a6d97a8a | 872 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 873 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 874 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
875 | return true; |
876 | } | |
e5d1bc0a | 877 | |
3ab665b7 | 878 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 879 | |
864ea921 PB |
880 | /* |
881 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
882 | * retry single alloc to be on the safe side. | |
883 | */ | |
884 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
885 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
886 | if (!reqs[0]) | |
a33ae9ce | 887 | return false; |
864ea921 | 888 | ret = 1; |
2b188cc1 | 889 | } |
864ea921 | 890 | |
37f0e767 | 891 | percpu_ref_get_many(&ctx->refs, ret); |
3ab665b7 | 892 | for (i = 0; i < ret; i++) { |
23a5c43b | 893 | struct io_kiocb *req = reqs[i]; |
3ab665b7 PB |
894 | |
895 | io_preinit_req(req, ctx); | |
fa05457a | 896 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 897 | } |
a33ae9ce PB |
898 | return true; |
899 | } | |
900 | ||
901 | static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx) | |
902 | { | |
88ab95be | 903 | if (unlikely(io_req_cache_empty(ctx))) |
a33ae9ce PB |
904 | return __io_alloc_req_refill(ctx); |
905 | return true; | |
906 | } | |
907 | ||
908 | static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx) | |
909 | { | |
910 | struct io_wq_work_node *node; | |
911 | ||
912 | node = wq_stack_extract(&ctx->submit_state.free_list); | |
c2b6c6bc | 913 | return container_of(node, struct io_kiocb, comp_list); |
2b188cc1 JA |
914 | } |
915 | ||
6b639522 | 916 | static inline void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 917 | { |
094bae49 | 918 | unsigned int flags = req->flags; |
929a3af9 | 919 | |
867f8fa5 | 920 | if (unlikely(flags & IO_REQ_CLEAN_FLAGS)) |
3a0a6902 | 921 | io_clean_op(req); |
e1d767f0 PB |
922 | if (!(flags & REQ_F_FIXED_FILE)) |
923 | io_put_file(req->file); | |
e65ef56d JA |
924 | } |
925 | ||
59915143 | 926 | __cold void io_free_req(struct io_kiocb *req) |
c6ca97b3 | 927 | { |
51a4cc11 | 928 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 929 | |
7ac1edc4 | 930 | io_req_put_rsrc(req); |
216578e5 | 931 | io_dismantle_req(req); |
7c660731 | 932 | io_put_task(req->task, 1); |
c6ca97b3 | 933 | |
79ebeaee | 934 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 935 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
c34b025f | 936 | ctx->locked_free_nr++; |
79ebeaee | 937 | spin_unlock(&ctx->completion_lock); |
e65ef56d JA |
938 | } |
939 | ||
d81499bf PB |
940 | static void __io_req_find_next_prep(struct io_kiocb *req) |
941 | { | |
942 | struct io_ring_ctx *ctx = req->ctx; | |
943 | bool posted; | |
944 | ||
945 | spin_lock(&ctx->completion_lock); | |
946 | posted = io_disarm_next(req); | |
60053be8 | 947 | io_commit_cqring(ctx); |
d81499bf PB |
948 | spin_unlock(&ctx->completion_lock); |
949 | if (posted) | |
950 | io_cqring_ev_posted(ctx); | |
951 | } | |
952 | ||
953 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 954 | { |
33cc89a9 | 955 | struct io_kiocb *nxt; |
944e58bf | 956 | |
9e645e11 JA |
957 | /* |
958 | * If LINK is set, we have dependent requests in this chain. If we | |
959 | * didn't fail this request, queue the first one up, moving any other | |
960 | * dependencies to the next request. In case of failure, fail the rest | |
961 | * of the chain. | |
962 | */ | |
d81499bf PB |
963 | if (unlikely(req->flags & IO_DISARM_MASK)) |
964 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
965 | nxt = req->link; |
966 | req->link = NULL; | |
967 | return nxt; | |
4d7dd462 | 968 | } |
9e645e11 | 969 | |
f237c30a | 970 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked) |
2c32395d PB |
971 | { |
972 | if (!ctx) | |
973 | return; | |
ef060ea9 JA |
974 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
975 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
f237c30a | 976 | if (*locked) { |
c450178d | 977 | io_submit_flush_completions(ctx); |
2c32395d | 978 | mutex_unlock(&ctx->uring_lock); |
f237c30a | 979 | *locked = false; |
2c32395d PB |
980 | } |
981 | percpu_ref_put(&ctx->refs); | |
982 | } | |
983 | ||
f28c240e HX |
984 | static inline void ctx_commit_and_unlock(struct io_ring_ctx *ctx) |
985 | { | |
986 | io_commit_cqring(ctx); | |
987 | spin_unlock(&ctx->completion_lock); | |
988 | io_cqring_ev_posted(ctx); | |
989 | } | |
990 | ||
991 | static void handle_prev_tw_list(struct io_wq_work_node *node, | |
992 | struct io_ring_ctx **ctx, bool *uring_locked) | |
993 | { | |
994 | if (*ctx && !*uring_locked) | |
995 | spin_lock(&(*ctx)->completion_lock); | |
996 | ||
997 | do { | |
998 | struct io_wq_work_node *next = node->next; | |
999 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1000 | io_task_work.node); | |
1001 | ||
34d2bfe7 JA |
1002 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); |
1003 | ||
f28c240e HX |
1004 | if (req->ctx != *ctx) { |
1005 | if (unlikely(!*uring_locked && *ctx)) | |
1006 | ctx_commit_and_unlock(*ctx); | |
1007 | ||
1008 | ctx_flush_and_put(*ctx, uring_locked); | |
1009 | *ctx = req->ctx; | |
1010 | /* if not contended, grab and improve batching */ | |
1011 | *uring_locked = mutex_trylock(&(*ctx)->uring_lock); | |
1012 | percpu_ref_get(&(*ctx)->refs); | |
1013 | if (unlikely(!*uring_locked)) | |
1014 | spin_lock(&(*ctx)->completion_lock); | |
1015 | } | |
97b388d7 | 1016 | if (likely(*uring_locked)) { |
f28c240e | 1017 | req->io_task_work.func(req, uring_locked); |
97b388d7 JA |
1018 | } else { |
1019 | req->cqe.flags = io_put_kbuf_comp(req); | |
1020 | __io_req_complete_post(req); | |
1021 | } | |
f28c240e HX |
1022 | node = next; |
1023 | } while (node); | |
1024 | ||
1025 | if (unlikely(!*uring_locked)) | |
1026 | ctx_commit_and_unlock(*ctx); | |
1027 | } | |
1028 | ||
1029 | static void handle_tw_list(struct io_wq_work_node *node, | |
1030 | struct io_ring_ctx **ctx, bool *locked) | |
9f8d032a HX |
1031 | { |
1032 | do { | |
1033 | struct io_wq_work_node *next = node->next; | |
1034 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1035 | io_task_work.node); | |
1036 | ||
34d2bfe7 JA |
1037 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); |
1038 | ||
9f8d032a HX |
1039 | if (req->ctx != *ctx) { |
1040 | ctx_flush_and_put(*ctx, locked); | |
1041 | *ctx = req->ctx; | |
1042 | /* if not contended, grab and improve batching */ | |
1043 | *locked = mutex_trylock(&(*ctx)->uring_lock); | |
1044 | percpu_ref_get(&(*ctx)->refs); | |
1045 | } | |
1046 | req->io_task_work.func(req, locked); | |
1047 | node = next; | |
1048 | } while (node); | |
1049 | } | |
1050 | ||
c9f06aa7 | 1051 | void tctx_task_work(struct callback_head *cb) |
c40f6379 | 1052 | { |
f28c240e | 1053 | bool uring_locked = false; |
ebd0df2e | 1054 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
1055 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
1056 | task_work); | |
c40f6379 | 1057 | |
16f72070 | 1058 | while (1) { |
f28c240e | 1059 | struct io_wq_work_node *node1, *node2; |
3f18407d PB |
1060 | |
1061 | spin_lock_irq(&tctx->task_lock); | |
3fe07bcd | 1062 | node1 = tctx->prio_task_list.first; |
f28c240e | 1063 | node2 = tctx->task_list.first; |
3f18407d | 1064 | INIT_WQ_LIST(&tctx->task_list); |
3fe07bcd | 1065 | INIT_WQ_LIST(&tctx->prio_task_list); |
f28c240e | 1066 | if (!node2 && !node1) |
6294f368 | 1067 | tctx->task_running = false; |
3f18407d | 1068 | spin_unlock_irq(&tctx->task_lock); |
f28c240e | 1069 | if (!node2 && !node1) |
6294f368 | 1070 | break; |
3f18407d | 1071 | |
f28c240e HX |
1072 | if (node1) |
1073 | handle_prev_tw_list(node1, &ctx, &uring_locked); | |
f28c240e HX |
1074 | if (node2) |
1075 | handle_tw_list(node2, &ctx, &uring_locked); | |
7cbf1722 | 1076 | cond_resched(); |
68ca8fc0 | 1077 | |
a6d97a8a | 1078 | if (data_race(!tctx->task_list.first) && |
3fe07bcd | 1079 | data_race(!tctx->prio_task_list.first) && uring_locked) |
68ca8fc0 | 1080 | io_submit_flush_completions(ctx); |
3f18407d | 1081 | } |
ebd0df2e | 1082 | |
f28c240e | 1083 | ctx_flush_and_put(ctx, &uring_locked); |
3cc7fdb9 PB |
1084 | |
1085 | /* relaxed read is enough as only the task itself sets ->in_idle */ | |
1086 | if (unlikely(atomic_read(&tctx->in_idle))) | |
1087 | io_uring_drop_tctx_refs(current); | |
7cbf1722 JA |
1088 | } |
1089 | ||
3fe07bcd JA |
1090 | static void __io_req_task_work_add(struct io_kiocb *req, |
1091 | struct io_uring_task *tctx, | |
1092 | struct io_wq_work_list *list) | |
7cbf1722 | 1093 | { |
9f010507 | 1094 | struct io_ring_ctx *ctx = req->ctx; |
e09ee510 | 1095 | struct io_wq_work_node *node; |
0b81e80c | 1096 | unsigned long flags; |
6294f368 | 1097 | bool running; |
7cbf1722 | 1098 | |
0b81e80c | 1099 | spin_lock_irqsave(&tctx->task_lock, flags); |
3fe07bcd | 1100 | wq_list_add_tail(&req->io_task_work.node, list); |
6294f368 PB |
1101 | running = tctx->task_running; |
1102 | if (!running) | |
1103 | tctx->task_running = true; | |
0b81e80c | 1104 | spin_unlock_irqrestore(&tctx->task_lock, flags); |
7cbf1722 JA |
1105 | |
1106 | /* task_work already pending, we're done */ | |
6294f368 | 1107 | if (running) |
e09ee510 | 1108 | return; |
7cbf1722 | 1109 | |
ef060ea9 JA |
1110 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1111 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1112 | ||
3fe07bcd | 1113 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1114 | return; |
2215bed9 | 1115 | |
0b81e80c | 1116 | spin_lock_irqsave(&tctx->task_lock, flags); |
6294f368 | 1117 | tctx->task_running = false; |
3fe07bcd | 1118 | node = wq_list_merge(&tctx->prio_task_list, &tctx->task_list); |
0b81e80c | 1119 | spin_unlock_irqrestore(&tctx->task_lock, flags); |
7cbf1722 | 1120 | |
e09ee510 PB |
1121 | while (node) { |
1122 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1123 | node = node->next; | |
1124 | if (llist_add(&req->io_task_work.fallback_node, | |
1125 | &req->ctx->fallback_llist)) | |
1126 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1127 | } | |
eab30c4d PB |
1128 | } |
1129 | ||
99f15d8d | 1130 | void io_req_task_work_add(struct io_kiocb *req) |
3fe07bcd JA |
1131 | { |
1132 | struct io_uring_task *tctx = req->task->io_uring; | |
1133 | ||
1134 | __io_req_task_work_add(req, tctx, &tctx->task_list); | |
1135 | } | |
1136 | ||
f3b44f92 | 1137 | void io_req_task_prio_work_add(struct io_kiocb *req) |
3fe07bcd JA |
1138 | { |
1139 | struct io_uring_task *tctx = req->task->io_uring; | |
1140 | ||
1141 | if (req->ctx->flags & IORING_SETUP_SQPOLL) | |
1142 | __io_req_task_work_add(req, tctx, &tctx->prio_task_list); | |
1143 | else | |
1144 | __io_req_task_work_add(req, tctx, &tctx->task_list); | |
1145 | } | |
1146 | ||
4e118cd9 | 1147 | static void io_req_tw_post(struct io_kiocb *req, bool *locked) |
c40f6379 | 1148 | { |
97b388d7 | 1149 | io_req_complete_post(req); |
4e118cd9 | 1150 | } |
c40f6379 | 1151 | |
59915143 | 1152 | void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags) |
4e118cd9 | 1153 | { |
97b388d7 | 1154 | io_req_set_res(req, res, cflags); |
4e118cd9 | 1155 | req->io_task_work.func = io_req_tw_post; |
3fe07bcd | 1156 | io_req_task_work_add(req); |
4e118cd9 PB |
1157 | } |
1158 | ||
f237c30a | 1159 | static void io_req_task_cancel(struct io_kiocb *req, bool *locked) |
c40f6379 | 1160 | { |
b18a1a45 | 1161 | /* not needed for normal modes, but SQPOLL depends on it */ |
971cf9c1 | 1162 | io_tw_lock(req->ctx, locked); |
cef216fc | 1163 | io_req_complete_failed(req, req->cqe.res); |
c40f6379 JA |
1164 | } |
1165 | ||
329061d3 | 1166 | void io_req_task_submit(struct io_kiocb *req, bool *locked) |
c40f6379 | 1167 | { |
971cf9c1 | 1168 | io_tw_lock(req->ctx, locked); |
316319e8 | 1169 | /* req->task == current here, checking PF_EXITING is safe */ |
af066f31 | 1170 | if (likely(!(req->task->flags & PF_EXITING))) |
cbc2e203 | 1171 | io_queue_sqe(req); |
81b6d05c | 1172 | else |
2593553a | 1173 | io_req_complete_failed(req, -EFAULT); |
c40f6379 JA |
1174 | } |
1175 | ||
59915143 | 1176 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1177 | { |
97b388d7 | 1178 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1179 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1180 | io_req_task_work_add(req); |
c40f6379 JA |
1181 | } |
1182 | ||
f3b44f92 | 1183 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1184 | { |
5b0a6acc | 1185 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1186 | io_req_task_work_add(req); |
a3df7698 PB |
1187 | } |
1188 | ||
59915143 | 1189 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1190 | { |
9b5f7bd9 | 1191 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1192 | |
1193 | if (nxt) | |
906a8c3f | 1194 | io_req_task_queue(nxt); |
c69f8dbe JL |
1195 | } |
1196 | ||
f3b44f92 | 1197 | void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node) |
3aa83bfb | 1198 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1199 | { |
d4b7a5ef | 1200 | struct task_struct *task = NULL; |
37f0e767 | 1201 | int task_refs = 0; |
5af1d13e | 1202 | |
3aa83bfb PB |
1203 | do { |
1204 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1205 | comp_list); | |
2d6500d4 | 1206 | |
a538be5b PB |
1207 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1208 | if (req->flags & REQ_F_REFCOUNT) { | |
1209 | node = req->comp_list.next; | |
1210 | if (!req_ref_put_and_test(req)) | |
1211 | continue; | |
1212 | } | |
b605a7fa PB |
1213 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1214 | struct async_poll *apoll = req->apoll; | |
1215 | ||
1216 | if (apoll->double_poll) | |
1217 | kfree(apoll->double_poll); | |
1218 | list_add(&apoll->poll.wait.entry, | |
1219 | &ctx->apoll_cache); | |
1220 | req->flags &= ~REQ_F_POLLED; | |
1221 | } | |
da1a08c5 | 1222 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1223 | io_queue_next(req); |
a538be5b PB |
1224 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1225 | io_clean_op(req); | |
c1e53a69 | 1226 | } |
a538be5b PB |
1227 | if (!(req->flags & REQ_F_FIXED_FILE)) |
1228 | io_put_file(req->file); | |
2d6500d4 | 1229 | |
ab409402 | 1230 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1231 | |
d4b7a5ef PB |
1232 | if (req->task != task) { |
1233 | if (task) | |
1234 | io_put_task(task, task_refs); | |
1235 | task = req->task; | |
1236 | task_refs = 0; | |
1237 | } | |
1238 | task_refs++; | |
c1e53a69 | 1239 | node = req->comp_list.next; |
fa05457a | 1240 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1241 | } while (node); |
d4b7a5ef | 1242 | |
d4b7a5ef PB |
1243 | if (task) |
1244 | io_put_task(task, task_refs); | |
7a743e22 PB |
1245 | } |
1246 | ||
c450178d | 1247 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1248 | __must_hold(&ctx->uring_lock) |
905c172f | 1249 | { |
6f33b0bc | 1250 | struct io_wq_work_node *node, *prev; |
cd0ca2e0 | 1251 | struct io_submit_state *state = &ctx->submit_state; |
905c172f | 1252 | |
d9dee430 PB |
1253 | spin_lock(&ctx->completion_lock); |
1254 | wq_list_for_each(node, prev, &state->compl_reqs) { | |
1255 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1256 | comp_list); | |
3d4aeb9f | 1257 | |
d9dee430 PB |
1258 | if (!(req->flags & REQ_F_CQE_SKIP)) |
1259 | __io_fill_cqe_req(ctx, req); | |
905c172f | 1260 | } |
5182ed2e | 1261 | |
d9dee430 PB |
1262 | io_commit_cqring(ctx); |
1263 | spin_unlock(&ctx->completion_lock); | |
1264 | io_cqring_ev_posted(ctx); | |
1265 | ||
1cce17ac | 1266 | io_free_batch_list(ctx, state->compl_reqs.first); |
6f33b0bc | 1267 | INIT_WQ_LIST(&state->compl_reqs); |
7a743e22 PB |
1268 | } |
1269 | ||
ba816ad6 JA |
1270 | /* |
1271 | * Drop reference to request, return next in chain (if there is one) if this | |
1272 | * was the last reference to this request. | |
1273 | */ | |
0d85035a | 1274 | static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 1275 | { |
9b5f7bd9 PB |
1276 | struct io_kiocb *nxt = NULL; |
1277 | ||
de9b4cca | 1278 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 1279 | if (unlikely(req->flags & IO_REQ_LINK_FLAGS)) |
7819a1f6 | 1280 | nxt = io_req_find_next(req); |
f5c6cf2a | 1281 | io_free_req(req); |
2a44f467 | 1282 | } |
9b5f7bd9 | 1283 | return nxt; |
2b188cc1 JA |
1284 | } |
1285 | ||
6c503150 | 1286 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1287 | { |
1288 | /* See comment at the top of this file */ | |
1289 | smp_rmb(); | |
e23de15f | 1290 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1291 | } |
1292 | ||
def596e9 JA |
1293 | /* |
1294 | * We can't just wait for polled events to come to us, we have to actively | |
1295 | * find and complete them. | |
1296 | */ | |
c072481d | 1297 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1298 | { |
1299 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1300 | return; | |
1301 | ||
1302 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1303 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1304 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1305 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1306 | break; |
08f5439f JA |
1307 | /* |
1308 | * Ensure we allow local-to-the-cpu processing to take place, | |
1309 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1310 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1311 | */ |
3fcee5a6 PB |
1312 | if (need_resched()) { |
1313 | mutex_unlock(&ctx->uring_lock); | |
1314 | cond_resched(); | |
1315 | mutex_lock(&ctx->uring_lock); | |
1316 | } | |
def596e9 JA |
1317 | } |
1318 | mutex_unlock(&ctx->uring_lock); | |
1319 | } | |
1320 | ||
7668b92a | 1321 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1322 | { |
7668b92a | 1323 | unsigned int nr_events = 0; |
e9979b36 | 1324 | int ret = 0; |
155bc950 | 1325 | unsigned long check_cq; |
500f9fba | 1326 | |
3a08576b PB |
1327 | check_cq = READ_ONCE(ctx->check_cq); |
1328 | if (unlikely(check_cq)) { | |
1329 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
1330 | __io_cqring_overflow_flush(ctx, false); | |
1331 | /* | |
1332 | * Similarly do not spin if we have not informed the user of any | |
1333 | * dropped CQE. | |
1334 | */ | |
1335 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1336 | return -EBADR; | |
1337 | } | |
f39c8a5b PB |
1338 | /* |
1339 | * Don't enter poll loop if we already have events pending. | |
1340 | * If we do, we can potentially be spinning for commands that | |
1341 | * already triggered a CQE (eg in error). | |
1342 | */ | |
f39c8a5b | 1343 | if (io_cqring_events(ctx)) |
d487b43c | 1344 | return 0; |
155bc950 | 1345 | |
def596e9 | 1346 | do { |
500f9fba JA |
1347 | /* |
1348 | * If a submit got punted to a workqueue, we can have the | |
1349 | * application entering polling for a command before it gets | |
1350 | * issued. That app will hold the uring_lock for the duration | |
1351 | * of the poll right here, so we need to take a breather every | |
1352 | * now and then to ensure that the issue has a chance to add | |
1353 | * the poll to the issued list. Otherwise we can spin here | |
1354 | * forever, while the workqueue is stuck trying to acquire the | |
1355 | * very same mutex. | |
1356 | */ | |
5eef4e87 | 1357 | if (wq_list_empty(&ctx->iopoll_list)) { |
8f487ef2 PB |
1358 | u32 tail = ctx->cached_cq_tail; |
1359 | ||
500f9fba | 1360 | mutex_unlock(&ctx->uring_lock); |
4c6e277c | 1361 | io_run_task_work(); |
500f9fba | 1362 | mutex_lock(&ctx->uring_lock); |
def596e9 | 1363 | |
8f487ef2 PB |
1364 | /* some requests don't go through iopoll_list */ |
1365 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1366 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1367 | break; |
500f9fba | 1368 | } |
5ba3c874 PB |
1369 | ret = io_do_iopoll(ctx, !min); |
1370 | if (ret < 0) | |
1371 | break; | |
1372 | nr_events += ret; | |
1373 | ret = 0; | |
1374 | } while (nr_events < min && !need_resched()); | |
d487b43c | 1375 | |
def596e9 JA |
1376 | return ret; |
1377 | } | |
7012c815 PB |
1378 | |
1379 | void io_req_task_complete(struct io_kiocb *req, bool *locked) | |
8ef12efe | 1380 | { |
7012c815 PB |
1381 | if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) { |
1382 | unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED; | |
1383 | ||
1384 | req->cqe.flags |= io_put_kbuf(req, issue_flags); | |
1385 | } | |
1386 | ||
1387 | if (*locked) | |
fff4e40e | 1388 | io_req_add_compl_list(req); |
7012c815 | 1389 | else |
97b388d7 | 1390 | io_req_complete_post(req); |
8ef12efe JA |
1391 | } |
1392 | ||
def596e9 JA |
1393 | /* |
1394 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1395 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1396 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1397 | * accessing the kiocb cookie. |
1398 | */ | |
9882131c | 1399 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1400 | { |
1401 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1402 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1403 | |
1404 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1405 | if (unlikely(needs_lock)) |
cb3d8972 | 1406 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1407 | |
1408 | /* | |
1409 | * Track whether we have multiple files in our lists. This will impact | |
1410 | * how we do polling eventually, not spinning if we're on potentially | |
1411 | * different devices. | |
1412 | */ | |
5eef4e87 | 1413 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1414 | ctx->poll_multi_queue = false; |
1415 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1416 | struct io_kiocb *list_req; |
1417 | ||
5eef4e87 PB |
1418 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1419 | comp_list); | |
30da1b45 | 1420 | if (list_req->file != req->file) |
915b3dde | 1421 | ctx->poll_multi_queue = true; |
def596e9 JA |
1422 | } |
1423 | ||
1424 | /* | |
1425 | * For fast devices, IO may have already completed. If it has, add | |
1426 | * it to the front so we find it first. | |
1427 | */ | |
65a6543d | 1428 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1429 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1430 | else |
5eef4e87 | 1431 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1432 | |
3b44b371 | 1433 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1434 | /* |
1435 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1436 | * in sq thread task context or in io worker task context. If | |
1437 | * current task context is sq thread, we don't need to check | |
1438 | * whether should wake up sq thread. | |
1439 | */ | |
1440 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1441 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1442 | wake_up(&ctx->sq_data->wait); | |
1443 | ||
1444 | mutex_unlock(&ctx->uring_lock); | |
1445 | } | |
def596e9 JA |
1446 | } |
1447 | ||
4503b767 JA |
1448 | static bool io_bdev_nowait(struct block_device *bdev) |
1449 | { | |
9ba0d0c8 | 1450 | return !bdev || blk_queue_nowait(bdev_get_queue(bdev)); |
4503b767 JA |
1451 | } |
1452 | ||
2b188cc1 JA |
1453 | /* |
1454 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1455 | * any file. For now, just ensure that anything potentially problematic is done | |
1456 | * inline. | |
1457 | */ | |
88459b50 | 1458 | static bool __io_file_supports_nowait(struct file *file, umode_t mode) |
2b188cc1 | 1459 | { |
4503b767 | 1460 | if (S_ISBLK(mode)) { |
4e7b5671 CH |
1461 | if (IS_ENABLED(CONFIG_BLOCK) && |
1462 | io_bdev_nowait(I_BDEV(file->f_mapping->host))) | |
4503b767 JA |
1463 | return true; |
1464 | return false; | |
1465 | } | |
976517f1 | 1466 | if (S_ISSOCK(mode)) |
2b188cc1 | 1467 | return true; |
4503b767 | 1468 | if (S_ISREG(mode)) { |
4e7b5671 CH |
1469 | if (IS_ENABLED(CONFIG_BLOCK) && |
1470 | io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
e5550a14 | 1471 | !io_is_uring_fops(file)) |
4503b767 JA |
1472 | return true; |
1473 | return false; | |
1474 | } | |
2b188cc1 | 1475 | |
c5b85625 JA |
1476 | /* any ->read/write should understand O_NONBLOCK */ |
1477 | if (file->f_flags & O_NONBLOCK) | |
1478 | return true; | |
35645ac3 | 1479 | return file->f_mode & FMODE_NOWAIT; |
2b188cc1 | 1480 | } |
c5b85625 | 1481 | |
88459b50 PB |
1482 | /* |
1483 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1484 | * any file. For now, just ensure that anything potentially problematic is done | |
1485 | * inline. | |
1486 | */ | |
a4ad4f74 | 1487 | unsigned int io_file_get_flags(struct file *file) |
88459b50 PB |
1488 | { |
1489 | umode_t mode = file_inode(file)->i_mode; | |
1490 | unsigned int res = 0; | |
af197f50 | 1491 | |
88459b50 PB |
1492 | if (S_ISREG(mode)) |
1493 | res |= FFS_ISREG; | |
1494 | if (__io_file_supports_nowait(file, mode)) | |
1495 | res |= FFS_NOWAIT; | |
5e45690a JA |
1496 | if (io_file_need_scm(file)) |
1497 | res |= FFS_SCM; | |
88459b50 | 1498 | return res; |
2b188cc1 JA |
1499 | } |
1500 | ||
99f15d8d | 1501 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1502 | { |
e8c2bc1f JA |
1503 | WARN_ON_ONCE(!io_op_defs[req->opcode].async_size); |
1504 | req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL); | |
d886e185 PB |
1505 | if (req->async_data) { |
1506 | req->flags |= REQ_F_ASYNC_DATA; | |
1507 | return false; | |
1508 | } | |
1509 | return true; | |
3d9932a8 XW |
1510 | } |
1511 | ||
f3b44f92 | 1512 | int io_req_prep_async(struct io_kiocb *req) |
f67676d1 | 1513 | { |
0702e536 JA |
1514 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
1515 | ||
1516 | /* assign early for deferred execution for non-fixed file */ | |
1517 | if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE)) | |
1518 | req->file = io_file_get_normal(req, req->cqe.fd); | |
dc919caf | 1519 | if (!def->prep_async) |
0702e536 JA |
1520 | return 0; |
1521 | if (WARN_ON_ONCE(req_has_async_data(req))) | |
1522 | return -EFAULT; | |
1523 | if (io_alloc_async_data(req)) | |
1524 | return -EAGAIN; | |
1525 | ||
dc919caf | 1526 | return def->prep_async(req); |
bfe76559 PB |
1527 | } |
1528 | ||
9cf7c104 PB |
1529 | static u32 io_get_sequence(struct io_kiocb *req) |
1530 | { | |
a3dbdf54 | 1531 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1532 | struct io_kiocb *cur; |
9cf7c104 | 1533 | |
a3dbdf54 | 1534 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1535 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1536 | seq--; |
1537 | return seq; | |
9cf7c104 PB |
1538 | } |
1539 | ||
c072481d | 1540 | static __cold void io_drain_req(struct io_kiocb *req) |
de0617e4 | 1541 | { |
a197f664 | 1542 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1543 | struct io_defer_entry *de; |
f67676d1 | 1544 | int ret; |
e0eb71dc | 1545 | u32 seq = io_get_sequence(req); |
3c19966d | 1546 | |
9d858b21 | 1547 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1548 | spin_lock(&ctx->completion_lock); |
5e371265 | 1549 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1550 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1551 | queue: |
10c66904 | 1552 | ctx->drain_active = false; |
e0eb71dc PB |
1553 | io_req_task_queue(req); |
1554 | return; | |
10c66904 | 1555 | } |
e302f104 | 1556 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1557 | |
b7e298d2 | 1558 | ret = io_req_prep_async(req); |
e0eb71dc PB |
1559 | if (ret) { |
1560 | fail: | |
1561 | io_req_complete_failed(req, ret); | |
1562 | return; | |
1563 | } | |
cbdcb435 | 1564 | io_prep_async_link(req); |
27dc8338 | 1565 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1566 | if (!de) { |
1b48773f | 1567 | ret = -ENOMEM; |
e0eb71dc | 1568 | goto fail; |
76cc33d7 | 1569 | } |
2d28390a | 1570 | |
79ebeaee | 1571 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1572 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1573 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1574 | kfree(de); |
e0eb71dc | 1575 | goto queue; |
de0617e4 JA |
1576 | } |
1577 | ||
48863ffd | 1578 | trace_io_uring_defer(req); |
27dc8338 | 1579 | de->req = req; |
9cf7c104 | 1580 | de->seq = seq; |
27dc8338 | 1581 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1582 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1583 | } |
1584 | ||
68fb8979 | 1585 | static void io_clean_op(struct io_kiocb *req) |
99bc4c38 | 1586 | { |
8197b053 PB |
1587 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
1588 | spin_lock(&req->ctx->completion_lock); | |
cc3cec83 | 1589 | io_put_kbuf_comp(req); |
8197b053 PB |
1590 | spin_unlock(&req->ctx->completion_lock); |
1591 | } | |
99bc4c38 | 1592 | |
0e1b6fe3 | 1593 | if (req->flags & REQ_F_NEED_CLEANUP) { |
4d4c9cff | 1594 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
bb040a21 | 1595 | |
4d4c9cff JA |
1596 | if (def->cleanup) |
1597 | def->cleanup(req); | |
99bc4c38 | 1598 | } |
75652a30 JA |
1599 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1600 | kfree(req->apoll->double_poll); | |
1601 | kfree(req->apoll); | |
1602 | req->apoll = NULL; | |
1603 | } | |
9cae36a0 JA |
1604 | if (req->flags & REQ_F_INFLIGHT) { |
1605 | struct io_uring_task *tctx = req->task->io_uring; | |
1606 | ||
1607 | atomic_dec(&tctx->inflight_tracked); | |
1608 | } | |
c854357b | 1609 | if (req->flags & REQ_F_CREDS) |
b8e64b53 | 1610 | put_cred(req->creds); |
d886e185 PB |
1611 | if (req->flags & REQ_F_ASYNC_DATA) { |
1612 | kfree(req->async_data); | |
1613 | req->async_data = NULL; | |
1614 | } | |
c854357b | 1615 | req->flags &= ~IO_REQ_CLEAN_FLAGS; |
99bc4c38 PB |
1616 | } |
1617 | ||
6bf9c47a JA |
1618 | static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags) |
1619 | { | |
1620 | if (req->file || !io_op_defs[req->opcode].needs_file) | |
1621 | return true; | |
1622 | ||
1623 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1624 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1625 | else |
cef216fc | 1626 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1627 | |
772f5e00 | 1628 | return !!req->file; |
6bf9c47a JA |
1629 | } |
1630 | ||
889fca73 | 1631 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1632 | { |
fcde59fe | 1633 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
5730b27e | 1634 | const struct cred *creds = NULL; |
d625c6ee | 1635 | int ret; |
2b188cc1 | 1636 | |
70152140 JA |
1637 | if (unlikely(!io_assign_file(req, issue_flags))) |
1638 | return -EBADF; | |
1639 | ||
6878b40e | 1640 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1641 | creds = override_creds(req->creds); |
5730b27e | 1642 | |
fcde59fe | 1643 | if (!def->audit_skip) |
5bd2182d PM |
1644 | audit_uring_entry(req->opcode); |
1645 | ||
0702e536 | 1646 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1647 | |
fcde59fe | 1648 | if (!def->audit_skip) |
5bd2182d PM |
1649 | audit_uring_exit(!ret, ret); |
1650 | ||
5730b27e JA |
1651 | if (creds) |
1652 | revert_creds(creds); | |
97b388d7 | 1653 | |
75d7b3ae PB |
1654 | if (ret == IOU_OK) { |
1655 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
1656 | io_req_add_compl_list(req); | |
1657 | else | |
1658 | io_req_complete_post(req); | |
1659 | } else if (ret != IOU_ISSUE_SKIP_COMPLETE) | |
def596e9 | 1660 | return ret; |
97b388d7 | 1661 | |
b532576e | 1662 | /* If the op doesn't have a file, we're not polling for it */ |
9983028e | 1663 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file) |
9882131c | 1664 | io_iopoll_req_issued(req, issue_flags); |
def596e9 JA |
1665 | |
1666 | return 0; | |
2b188cc1 JA |
1667 | } |
1668 | ||
329061d3 JA |
1669 | int io_poll_issue(struct io_kiocb *req, bool *locked) |
1670 | { | |
1671 | io_tw_lock(req->ctx, locked); | |
1672 | if (unlikely(req->task->flags & PF_EXITING)) | |
1673 | return -EFAULT; | |
aeaa72c6 | 1674 | return io_issue_sqe(req, IO_URING_F_NONBLOCK); |
329061d3 JA |
1675 | } |
1676 | ||
c9f06aa7 | 1677 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1678 | { |
1679 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1680 | ||
1681 | req = io_put_req_find_next(req); | |
1682 | return req ? &req->work : NULL; | |
1683 | } | |
1684 | ||
c9f06aa7 | 1685 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1686 | { |
1687 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6bf9c47a | 1688 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
d01905db PB |
1689 | unsigned int issue_flags = IO_URING_F_UNLOCKED; |
1690 | bool needs_poll = false; | |
6bf9c47a | 1691 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1692 | |
48dcd38d PB |
1693 | /* one will be dropped by ->io_free_work() after returning to io-wq */ |
1694 | if (!(req->flags & REQ_F_REFCOUNT)) | |
1695 | __io_req_set_refcount(req, 2); | |
1696 | else | |
1697 | req_ref_get(req); | |
5d5901a3 | 1698 | |
cb2d344c | 1699 | io_arm_ltimeout(req); |
6bf9c47a | 1700 | |
dadebc35 | 1701 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1702 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1703 | fail: |
6bf9c47a | 1704 | io_req_task_queue_fail(req, err); |
d01905db PB |
1705 | return; |
1706 | } | |
0f8da75b PB |
1707 | if (!io_assign_file(req, issue_flags)) { |
1708 | err = -EBADF; | |
1709 | work->flags |= IO_WQ_WORK_CANCEL; | |
1710 | goto fail; | |
1711 | } | |
31b51510 | 1712 | |
d01905db | 1713 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
1714 | bool opcode_poll = def->pollin || def->pollout; |
1715 | ||
1716 | if (opcode_poll && file_can_poll(req->file)) { | |
1717 | needs_poll = true; | |
d01905db | 1718 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 1719 | } |
561fb04a | 1720 | } |
31b51510 | 1721 | |
d01905db PB |
1722 | do { |
1723 | ret = io_issue_sqe(req, issue_flags); | |
1724 | if (ret != -EAGAIN) | |
1725 | break; | |
1726 | /* | |
1727 | * We can get EAGAIN for iopolled IO even though we're | |
1728 | * forcing a sync submission from here, since we can't | |
1729 | * wait for request slots on the block side. | |
1730 | */ | |
1731 | if (!needs_poll) { | |
e0deb6a0 PB |
1732 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
1733 | break; | |
d01905db PB |
1734 | cond_resched(); |
1735 | continue; | |
90fa0288 HX |
1736 | } |
1737 | ||
4d9237e3 | 1738 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
1739 | return; |
1740 | /* aborted or ready, in either case retry blocking */ | |
1741 | needs_poll = false; | |
1742 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
1743 | } while (1); | |
31b51510 | 1744 | |
a3df7698 | 1745 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 1746 | if (ret < 0) |
a3df7698 | 1747 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
1748 | } |
1749 | ||
531113bb JA |
1750 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
1751 | unsigned int issue_flags) | |
09bb8394 | 1752 | { |
5106dd6e JA |
1753 | struct io_ring_ctx *ctx = req->ctx; |
1754 | struct file *file = NULL; | |
ac177053 | 1755 | unsigned long file_ptr; |
09bb8394 | 1756 | |
93f052cb | 1757 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 1758 | |
ac177053 | 1759 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 1760 | goto out; |
ac177053 PB |
1761 | fd = array_index_nospec(fd, ctx->nr_user_files); |
1762 | file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; | |
1763 | file = (struct file *) (file_ptr & FFS_MASK); | |
1764 | file_ptr &= ~FFS_MASK; | |
1765 | /* mask in overlapping REQ_F and FFS bits */ | |
35645ac3 | 1766 | req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT); |
5106dd6e | 1767 | io_req_set_rsrc_node(req, ctx, 0); |
d78bd8ad | 1768 | WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap)); |
5106dd6e | 1769 | out: |
93f052cb | 1770 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
1771 | return file; |
1772 | } | |
d44f554e | 1773 | |
531113bb | 1774 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 1775 | { |
62906e89 | 1776 | struct file *file = fget(fd); |
ac177053 | 1777 | |
48863ffd | 1778 | trace_io_uring_file_get(req, fd); |
09bb8394 | 1779 | |
ac177053 | 1780 | /* we don't allow fixed io_uring files */ |
e5550a14 | 1781 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 1782 | io_req_track_inflight(req); |
8371adf5 | 1783 | return file; |
09bb8394 JA |
1784 | } |
1785 | ||
7bfa9bad | 1786 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
1787 | __must_hold(&req->ctx->uring_lock) |
1788 | { | |
7bfa9bad PB |
1789 | struct io_kiocb *linked_timeout; |
1790 | ||
1791 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
1792 | io_req_complete_failed(req, ret); | |
1793 | return; | |
1794 | } | |
1795 | ||
1796 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 1797 | |
4d9237e3 | 1798 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 1799 | case IO_APOLL_READY: |
d475a9a6 PB |
1800 | io_req_task_queue(req); |
1801 | break; | |
1802 | case IO_APOLL_ABORTED: | |
1803 | /* | |
1804 | * Queued up for async execution, worker will release | |
1805 | * submit reference when the iocb is actually submitted. | |
1806 | */ | |
6436c770 | 1807 | io_kbuf_recycle(req, 0); |
77955efb | 1808 | io_queue_iowq(req, NULL); |
d475a9a6 | 1809 | break; |
b1c62645 | 1810 | case IO_APOLL_OK: |
b1c62645 | 1811 | break; |
d475a9a6 PB |
1812 | } |
1813 | ||
1814 | if (linked_timeout) | |
1815 | io_queue_linked_timeout(linked_timeout); | |
1816 | } | |
1817 | ||
cbc2e203 | 1818 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 1819 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 1820 | { |
e0c5c576 | 1821 | int ret; |
2b188cc1 | 1822 | |
c5eef2b9 | 1823 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 1824 | |
491381ce JA |
1825 | /* |
1826 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
1827 | * doesn't support non-blocking read/write attempts | |
1828 | */ | |
7bfa9bad | 1829 | if (likely(!ret)) |
cb2d344c | 1830 | io_arm_ltimeout(req); |
7bfa9bad PB |
1831 | else |
1832 | io_queue_async(req, ret); | |
2b188cc1 JA |
1833 | } |
1834 | ||
4652fe3f | 1835 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 1836 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 1837 | { |
17b147f6 PB |
1838 | if (unlikely(req->flags & REQ_F_FAIL)) { |
1839 | /* | |
1840 | * We don't submit, fail them all, for that replace hardlinks | |
1841 | * with normal links. Extra REQ_F_LINK is tolerated. | |
1842 | */ | |
1843 | req->flags &= ~REQ_F_HARDLINK; | |
1844 | req->flags |= REQ_F_LINK; | |
1845 | io_req_complete_failed(req, req->cqe.res); | |
e0eb71dc PB |
1846 | } else if (unlikely(req->ctx->drain_active)) { |
1847 | io_drain_req(req); | |
76cc33d7 PB |
1848 | } else { |
1849 | int ret = io_req_prep_async(req); | |
1850 | ||
1851 | if (unlikely(ret)) | |
1852 | io_req_complete_failed(req, ret); | |
1853 | else | |
77955efb | 1854 | io_queue_iowq(req, NULL); |
ce35a47a | 1855 | } |
4fe2c963 JL |
1856 | } |
1857 | ||
b16fed66 PB |
1858 | /* |
1859 | * Check SQE restrictions (opcode and flags). | |
1860 | * | |
1861 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
1862 | */ | |
1863 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
1864 | struct io_kiocb *req, | |
1865 | unsigned int sqe_flags) | |
4fe2c963 | 1866 | { |
b16fed66 PB |
1867 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
1868 | return false; | |
1869 | ||
1870 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
1871 | ctx->restrictions.sqe_flags_required) | |
1872 | return false; | |
1873 | ||
1874 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
1875 | ctx->restrictions.sqe_flags_required)) | |
1876 | return false; | |
1877 | ||
1878 | return true; | |
4fe2c963 JL |
1879 | } |
1880 | ||
22b2ca31 PB |
1881 | static void io_init_req_drain(struct io_kiocb *req) |
1882 | { | |
1883 | struct io_ring_ctx *ctx = req->ctx; | |
1884 | struct io_kiocb *head = ctx->submit_state.link.head; | |
1885 | ||
1886 | ctx->drain_active = true; | |
1887 | if (head) { | |
1888 | /* | |
1889 | * If we need to drain a request in the middle of a link, drain | |
1890 | * the head request and the next request/link after the current | |
1891 | * link. Considering sequential execution of links, | |
b6c7db32 | 1892 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
1893 | * link. |
1894 | */ | |
b6c7db32 | 1895 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
1896 | ctx->drain_next = true; |
1897 | } | |
1898 | } | |
1899 | ||
b16fed66 PB |
1900 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1901 | const struct io_uring_sqe *sqe) | |
282cdc86 | 1902 | __must_hold(&ctx->uring_lock) |
b16fed66 | 1903 | { |
fcde59fe | 1904 | const struct io_op_def *def; |
b16fed66 | 1905 | unsigned int sqe_flags; |
fc0ae024 | 1906 | int personality; |
4a04d1d1 | 1907 | u8 opcode; |
b16fed66 | 1908 | |
864ea921 | 1909 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 1910 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 PB |
1911 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
1912 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
cef216fc | 1913 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 1914 | req->file = NULL; |
c1bdf8ed | 1915 | req->rsrc_node = NULL; |
b16fed66 | 1916 | req->task = current; |
b16fed66 | 1917 | |
4a04d1d1 PB |
1918 | if (unlikely(opcode >= IORING_OP_LAST)) { |
1919 | req->opcode = 0; | |
b16fed66 | 1920 | return -EINVAL; |
4a04d1d1 | 1921 | } |
fcde59fe | 1922 | def = &io_op_defs[opcode]; |
68fe256a PB |
1923 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
1924 | /* enforce forwards compatibility on users */ | |
1925 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
1926 | return -EINVAL; | |
4e906702 | 1927 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 1928 | if (!def->buffer_select) |
4e906702 JA |
1929 | return -EOPNOTSUPP; |
1930 | req->buf_index = READ_ONCE(sqe->buf_group); | |
1931 | } | |
5562a8d7 PB |
1932 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
1933 | ctx->drain_disabled = true; | |
1934 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
1935 | if (ctx->drain_disabled) | |
1936 | return -EOPNOTSUPP; | |
22b2ca31 | 1937 | io_init_req_drain(req); |
5562a8d7 | 1938 | } |
2a56a9bd PB |
1939 | } |
1940 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
1941 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
1942 | return -EACCES; | |
1943 | /* knock it to the slow queue path, will be drained there */ | |
1944 | if (ctx->drain_active) | |
1945 | req->flags |= REQ_F_FORCE_ASYNC; | |
1946 | /* if there is no link, we're at "next" request and need to drain */ | |
1947 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
1948 | ctx->drain_next = false; | |
1949 | ctx->drain_active = true; | |
b6c7db32 | 1950 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 1951 | } |
68fe256a | 1952 | } |
b16fed66 | 1953 | |
fcde59fe | 1954 | if (!def->ioprio && sqe->ioprio) |
73911426 | 1955 | return -EINVAL; |
fcde59fe | 1956 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
73911426 JA |
1957 | return -EINVAL; |
1958 | ||
fcde59fe | 1959 | if (def->needs_file) { |
6d63416d PB |
1960 | struct io_submit_state *state = &ctx->submit_state; |
1961 | ||
cef216fc | 1962 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 1963 | |
6d63416d PB |
1964 | /* |
1965 | * Plug now if we have more than 2 IO left after this, and the | |
1966 | * target is potentially a read/write to block based storage. | |
1967 | */ | |
fcde59fe | 1968 | if (state->need_plug && def->plug) { |
6d63416d PB |
1969 | state->plug_started = true; |
1970 | state->need_plug = false; | |
5ca7a8b3 | 1971 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 1972 | } |
b16fed66 | 1973 | } |
863e0560 | 1974 | |
003e8dcc JA |
1975 | personality = READ_ONCE(sqe->personality); |
1976 | if (personality) { | |
cdab10bf LT |
1977 | int ret; |
1978 | ||
c10d1f98 PB |
1979 | req->creds = xa_load(&ctx->personalities, personality); |
1980 | if (!req->creds) | |
003e8dcc | 1981 | return -EINVAL; |
c10d1f98 | 1982 | get_cred(req->creds); |
cdc1404a PM |
1983 | ret = security_uring_override_creds(req->creds); |
1984 | if (ret) { | |
1985 | put_cred(req->creds); | |
1986 | return ret; | |
1987 | } | |
b8e64b53 | 1988 | req->flags |= REQ_F_CREDS; |
003e8dcc | 1989 | } |
b16fed66 | 1990 | |
0702e536 | 1991 | return def->prep(req, sqe); |
b16fed66 PB |
1992 | } |
1993 | ||
df3becde PB |
1994 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
1995 | struct io_kiocb *req, int ret) | |
1996 | { | |
1997 | struct io_ring_ctx *ctx = req->ctx; | |
1998 | struct io_submit_link *link = &ctx->submit_state.link; | |
1999 | struct io_kiocb *head = link->head; | |
2000 | ||
48863ffd | 2001 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
2002 | |
2003 | /* | |
2004 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
2005 | * unusable. Instead of failing eagerly, continue assembling the link if | |
2006 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
2007 | * should find the flag and handle the rest. | |
2008 | */ | |
2009 | req_fail_link_node(req, ret); | |
2010 | if (head && !(head->flags & REQ_F_FAIL)) | |
2011 | req_fail_link_node(head, -ECANCELED); | |
2012 | ||
2013 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
2014 | if (head) { | |
2015 | link->last->link = req; | |
2016 | link->head = NULL; | |
2017 | req = head; | |
2018 | } | |
2019 | io_queue_sqe_fallback(req); | |
2020 | return ret; | |
2021 | } | |
2022 | ||
2023 | if (head) | |
2024 | link->last->link = req; | |
2025 | else | |
2026 | link->head = req; | |
2027 | link->last = req; | |
2028 | return 0; | |
2029 | } | |
2030 | ||
2031 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 2032 | const struct io_uring_sqe *sqe) |
282cdc86 | 2033 | __must_hold(&ctx->uring_lock) |
9e645e11 | 2034 | { |
a1ab7b35 | 2035 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 2036 | int ret; |
9e645e11 | 2037 | |
a6b8cadc | 2038 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
2039 | if (unlikely(ret)) |
2040 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 2041 | |
be7053b7 | 2042 | /* don't need @sqe from now on */ |
48863ffd | 2043 | trace_io_uring_submit_sqe(req, true); |
a6b8cadc | 2044 | |
9e645e11 JA |
2045 | /* |
2046 | * If we already have a head request, queue this one for async | |
2047 | * submittal once the head completes. If we don't have a head but | |
2048 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
2049 | * submitted sync once the chain is complete. If none of those | |
2050 | * conditions are true (normal request), then just queue it. | |
2051 | */ | |
924a07e4 | 2052 | if (unlikely(link->head)) { |
df3becde PB |
2053 | ret = io_req_prep_async(req); |
2054 | if (unlikely(ret)) | |
2055 | return io_submit_fail_init(sqe, req, ret); | |
2056 | ||
48863ffd | 2057 | trace_io_uring_link(req, link->head); |
f2f87370 | 2058 | link->last->link = req; |
863e0560 | 2059 | link->last = req; |
32fe525b | 2060 | |
da1a08c5 | 2061 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 2062 | return 0; |
df3becde PB |
2063 | /* last request of the link, flush it */ |
2064 | req = link->head; | |
f15a3431 | 2065 | link->head = NULL; |
924a07e4 PB |
2066 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
2067 | goto fallback; | |
2068 | ||
2069 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2070 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2071 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2072 | link->head = req; | |
2073 | link->last = req; | |
2074 | } else { | |
2075 | fallback: | |
2076 | io_queue_sqe_fallback(req); | |
2077 | } | |
f15a3431 | 2078 | return 0; |
9e645e11 | 2079 | } |
2e6e1fde | 2080 | |
f15a3431 | 2081 | io_queue_sqe(req); |
1d4240cc | 2082 | return 0; |
9e645e11 JA |
2083 | } |
2084 | ||
9a56a232 JA |
2085 | /* |
2086 | * Batched submission is done, ensure local IO is flushed out. | |
2087 | */ | |
553deffd | 2088 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2089 | { |
553deffd PB |
2090 | struct io_submit_state *state = &ctx->submit_state; |
2091 | ||
e126391c PB |
2092 | if (unlikely(state->link.head)) |
2093 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2094 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2095 | io_submit_flush_completions(ctx); |
27926b68 JA |
2096 | if (state->plug_started) |
2097 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2098 | } |
2099 | ||
2100 | /* | |
2101 | * Start submission side cache. | |
2102 | */ | |
2103 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2104 | unsigned int max_ios) |
9a56a232 | 2105 | { |
27926b68 | 2106 | state->plug_started = false; |
4b628aeb | 2107 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2108 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2109 | /* set only head, no need to init link_last in advance */ |
2110 | state->link.head = NULL; | |
9a56a232 JA |
2111 | } |
2112 | ||
2b188cc1 JA |
2113 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2114 | { | |
75b28aff | 2115 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2116 | |
caf582c6 PB |
2117 | /* |
2118 | * Ensure any loads from the SQEs are done at this point, | |
2119 | * since once we write the new head, the application could | |
2120 | * write new data to them. | |
2121 | */ | |
2122 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2123 | } |
2124 | ||
2b188cc1 | 2125 | /* |
dd9ae8a0 | 2126 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2127 | * that is mapped by userspace. This means that care needs to be taken to |
2128 | * ensure that reads are stable, as we cannot rely on userspace always | |
2129 | * being a good citizen. If members of the sqe are validated and then later | |
2130 | * used, it's important that those reads are done through READ_ONCE() to | |
2131 | * prevent a re-load down the line. | |
2132 | */ | |
709b302f | 2133 | static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) |
2b188cc1 | 2134 | { |
ea5ab3b5 | 2135 | unsigned head, mask = ctx->sq_entries - 1; |
17d3aeb3 | 2136 | unsigned sq_idx = ctx->cached_sq_head++ & mask; |
2b188cc1 JA |
2137 | |
2138 | /* | |
2139 | * The cached sq head (or cq tail) serves two purposes: | |
2140 | * | |
2141 | * 1) allows us to batch the cost of updating the user visible | |
2142 | * head updates. | |
2143 | * 2) allows the kernel side to track the head on its own, even | |
2144 | * though the application is the one updating it. | |
2145 | */ | |
17d3aeb3 | 2146 | head = READ_ONCE(ctx->sq_array[sq_idx]); |
ebdeb7c0 JA |
2147 | if (likely(head < ctx->sq_entries)) { |
2148 | /* double index for 128-byte SQEs, twice as long */ | |
2149 | if (ctx->flags & IORING_SETUP_SQE128) | |
2150 | head <<= 1; | |
709b302f | 2151 | return &ctx->sq_sqes[head]; |
ebdeb7c0 | 2152 | } |
2b188cc1 JA |
2153 | |
2154 | /* drop invalid entries */ | |
15641e42 PB |
2155 | ctx->cq_extra--; |
2156 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2157 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
709b302f PB |
2158 | return NULL; |
2159 | } | |
2160 | ||
17437f31 | 2161 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2162 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2163 | { |
69629809 | 2164 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2165 | unsigned int left; |
2166 | int ret; | |
6c271ce2 | 2167 | |
51d48dab | 2168 | if (unlikely(!entries)) |
69629809 | 2169 | return 0; |
ee7d46d9 | 2170 | /* make sure SQ entry isn't read before tail */ |
8e6971a8 PB |
2171 | ret = left = min3(nr, ctx->sq_entries, entries); |
2172 | io_get_task_refs(left); | |
2173 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2174 | |
69629809 | 2175 | do { |
3529d8c2 | 2176 | const struct io_uring_sqe *sqe; |
196be95c | 2177 | struct io_kiocb *req; |
fb5ccc98 | 2178 | |
8e6971a8 | 2179 | if (unlikely(!io_alloc_req_refill(ctx))) |
fb5ccc98 | 2180 | break; |
a33ae9ce | 2181 | req = io_alloc_req(ctx); |
4fccfcbb PB |
2182 | sqe = io_get_sqe(ctx); |
2183 | if (unlikely(!sqe)) { | |
fa05457a | 2184 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2185 | break; |
2186 | } | |
6c271ce2 | 2187 | |
1cd15904 PB |
2188 | /* |
2189 | * Continue submitting even for sqe failure if the | |
2190 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2191 | */ | |
2192 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2193 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2194 | left--; | |
2195 | break; | |
bcbb7bf6 | 2196 | } |
1cd15904 | 2197 | } while (--left); |
9466f437 | 2198 | |
8e6971a8 PB |
2199 | if (unlikely(left)) { |
2200 | ret -= left; | |
2201 | /* try again if it submitted nothing and can't allocate a req */ | |
2202 | if (!ret && io_req_cache_empty(ctx)) | |
2203 | ret = -EAGAIN; | |
2204 | current->io_uring->cached_refs += left; | |
9466f437 | 2205 | } |
6c271ce2 | 2206 | |
553deffd | 2207 | io_submit_state_end(ctx); |
ae9428ca PB |
2208 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2209 | io_commit_sqring(ctx); | |
8e6971a8 | 2210 | return ret; |
6c271ce2 JA |
2211 | } |
2212 | ||
bda52162 JA |
2213 | struct io_wait_queue { |
2214 | struct wait_queue_entry wq; | |
2215 | struct io_ring_ctx *ctx; | |
5fd46178 | 2216 | unsigned cq_tail; |
bda52162 JA |
2217 | unsigned nr_timeouts; |
2218 | }; | |
2219 | ||
6c503150 | 2220 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
2221 | { |
2222 | struct io_ring_ctx *ctx = iowq->ctx; | |
5fd46178 | 2223 | int dist = ctx->cached_cq_tail - (int) iowq->cq_tail; |
bda52162 JA |
2224 | |
2225 | /* | |
d195a66e | 2226 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
2227 | * started waiting. For timeouts, we always want to return to userspace, |
2228 | * regardless of event count. | |
2229 | */ | |
5fd46178 | 2230 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
2231 | } |
2232 | ||
2233 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2234 | int wake_flags, void *key) | |
2235 | { | |
2236 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
2237 | wq); | |
2238 | ||
6c503150 PB |
2239 | /* |
2240 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2241 | * the task, and the next invocation will do it. | |
2242 | */ | |
10988a0a DY |
2243 | if (io_should_wake(iowq) || |
2244 | test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq)) | |
6c503150 PB |
2245 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2246 | return -1; | |
bda52162 JA |
2247 | } |
2248 | ||
73572984 | 2249 | int io_run_task_work_sig(void) |
af9c1a44 JA |
2250 | { |
2251 | if (io_run_task_work()) | |
2252 | return 1; | |
0b8cfa97 | 2253 | if (test_thread_flag(TIF_NOTIFY_SIGNAL)) |
792ee0f6 | 2254 | return -ERESTARTSYS; |
c5020bc8 OL |
2255 | if (task_sigpending(current)) |
2256 | return -EINTR; | |
2257 | return 0; | |
af9c1a44 JA |
2258 | } |
2259 | ||
eeb60b9a PB |
2260 | /* when returns >0, the caller should retry */ |
2261 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
2262 | struct io_wait_queue *iowq, | |
22833966 | 2263 | ktime_t timeout) |
eeb60b9a PB |
2264 | { |
2265 | int ret; | |
155bc950 | 2266 | unsigned long check_cq; |
eeb60b9a PB |
2267 | |
2268 | /* make sure we run task_work before checking for signals */ | |
2269 | ret = io_run_task_work_sig(); | |
2270 | if (ret || io_should_wake(iowq)) | |
2271 | return ret; | |
3a08576b | 2272 | |
155bc950 | 2273 | check_cq = READ_ONCE(ctx->check_cq); |
3a08576b PB |
2274 | if (unlikely(check_cq)) { |
2275 | /* let the caller flush overflows, retry */ | |
2276 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
2277 | return 1; | |
2278 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
2279 | return -EBADR; | |
2280 | } | |
22833966 JA |
2281 | if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS)) |
2282 | return -ETIME; | |
2283 | return 1; | |
eeb60b9a PB |
2284 | } |
2285 | ||
2b188cc1 JA |
2286 | /* |
2287 | * Wait until events become available, if we don't already have some. The | |
2288 | * application must reap them itself, as they reside on the shared cq ring. | |
2289 | */ | |
2290 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2291 | const sigset_t __user *sig, size_t sigsz, |
2292 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2293 | { |
90291099 | 2294 | struct io_wait_queue iowq; |
75b28aff | 2295 | struct io_rings *rings = ctx->rings; |
22833966 | 2296 | ktime_t timeout = KTIME_MAX; |
c1d5a224 | 2297 | int ret; |
2b188cc1 | 2298 | |
b41e9852 | 2299 | do { |
90f67366 | 2300 | io_cqring_overflow_flush(ctx); |
6c503150 | 2301 | if (io_cqring_events(ctx) >= min_events) |
b41e9852 | 2302 | return 0; |
4c6e277c | 2303 | if (!io_run_task_work()) |
b41e9852 | 2304 | break; |
b41e9852 | 2305 | } while (1); |
2b188cc1 JA |
2306 | |
2307 | if (sig) { | |
9e75ad5d AB |
2308 | #ifdef CONFIG_COMPAT |
2309 | if (in_compat_syscall()) | |
2310 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2311 | sigsz); |
9e75ad5d AB |
2312 | else |
2313 | #endif | |
b772434b | 2314 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2315 | |
2b188cc1 JA |
2316 | if (ret) |
2317 | return ret; | |
2318 | } | |
2319 | ||
950e79dd OL |
2320 | if (uts) { |
2321 | struct timespec64 ts; | |
2322 | ||
2323 | if (get_timespec64(&ts, uts)) | |
2324 | return -EFAULT; | |
2325 | timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); | |
2326 | } | |
2327 | ||
90291099 PB |
2328 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
2329 | iowq.wq.private = current; | |
2330 | INIT_LIST_HEAD(&iowq.wq.entry); | |
2331 | iowq.ctx = ctx; | |
bda52162 | 2332 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 2333 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
90291099 | 2334 | |
c826bd7a | 2335 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2336 | do { |
ca0a2651 | 2337 | /* if we can't even flush overflow, don't wait for more */ |
90f67366 | 2338 | if (!io_cqring_overflow_flush(ctx)) { |
ca0a2651 JA |
2339 | ret = -EBUSY; |
2340 | break; | |
2341 | } | |
311997b3 | 2342 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, |
bda52162 | 2343 | TASK_INTERRUPTIBLE); |
22833966 | 2344 | ret = io_cqring_wait_schedule(ctx, &iowq, timeout); |
ca0a2651 | 2345 | cond_resched(); |
eeb60b9a | 2346 | } while (ret > 0); |
bda52162 | 2347 | |
b4f20bb4 | 2348 | finish_wait(&ctx->cq_wait, &iowq.wq); |
b7db41c9 | 2349 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2350 | |
75b28aff | 2351 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2352 | } |
2353 | ||
73572984 | 2354 | static void io_mem_free(void *ptr) |
b9bd2bea | 2355 | { |
73572984 | 2356 | struct page *page; |
b36a2050 | 2357 | |
73572984 JA |
2358 | if (!ptr) |
2359 | return; | |
b9bd2bea | 2360 | |
73572984 JA |
2361 | page = virt_to_head_page(ptr); |
2362 | if (put_page_testzero(page)) | |
2363 | free_compound_page(page); | |
b9bd2bea PB |
2364 | } |
2365 | ||
73572984 | 2366 | static void *io_mem_alloc(size_t size) |
b9bd2bea | 2367 | { |
73572984 | 2368 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; |
b9bd2bea | 2369 | |
73572984 | 2370 | return (void *) __get_free_pages(gfp, get_order(size)); |
b9bd2bea PB |
2371 | } |
2372 | ||
73572984 JA |
2373 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2374 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2375 | { |
73572984 JA |
2376 | struct io_rings *rings; |
2377 | size_t off, sq_array_size; | |
6b06314c | 2378 | |
73572984 JA |
2379 | off = struct_size(rings, cqes, cq_entries); |
2380 | if (off == SIZE_MAX) | |
2381 | return SIZE_MAX; | |
2382 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2383 | if (check_shl_overflow(off, 1, &off)) | |
2384 | return SIZE_MAX; | |
2385 | } | |
ab409402 | 2386 | |
73572984 JA |
2387 | #ifdef CONFIG_SMP |
2388 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2389 | if (off == 0) | |
2390 | return SIZE_MAX; | |
2391 | #endif | |
82fbcfa9 | 2392 | |
73572984 JA |
2393 | if (sq_offset) |
2394 | *sq_offset = off; | |
82fbcfa9 | 2395 | |
73572984 JA |
2396 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2397 | if (sq_array_size == SIZE_MAX) | |
2398 | return SIZE_MAX; | |
6b06314c | 2399 | |
73572984 JA |
2400 | if (check_add_overflow(off, sq_array_size, &off)) |
2401 | return SIZE_MAX; | |
8bad28d8 | 2402 | |
73572984 | 2403 | return off; |
8bad28d8 HX |
2404 | } |
2405 | ||
73572984 JA |
2406 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, |
2407 | unsigned int eventfd_async) | |
8bad28d8 | 2408 | { |
73572984 JA |
2409 | struct io_ev_fd *ev_fd; |
2410 | __s32 __user *fds = arg; | |
2411 | int fd; | |
f2303b1f | 2412 | |
73572984 JA |
2413 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, |
2414 | lockdep_is_held(&ctx->uring_lock)); | |
2415 | if (ev_fd) | |
2416 | return -EBUSY; | |
8bad28d8 | 2417 | |
73572984 JA |
2418 | if (copy_from_user(&fd, fds, sizeof(*fds))) |
2419 | return -EFAULT; | |
8dd03afe | 2420 | |
73572984 JA |
2421 | ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); |
2422 | if (!ev_fd) | |
2423 | return -ENOMEM; | |
05f3fb3c | 2424 | |
73572984 JA |
2425 | ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); |
2426 | if (IS_ERR(ev_fd->cq_ev_fd)) { | |
2427 | int ret = PTR_ERR(ev_fd->cq_ev_fd); | |
2428 | kfree(ev_fd); | |
2429 | return ret; | |
2430 | } | |
2431 | ev_fd->eventfd_async = eventfd_async; | |
2432 | ctx->has_evfd = true; | |
2433 | rcu_assign_pointer(ctx->io_ev_fd, ev_fd); | |
2434 | return 0; | |
d7954b2b BM |
2435 | } |
2436 | ||
73572984 | 2437 | static void io_eventfd_put(struct rcu_head *rcu) |
2d091d62 | 2438 | { |
73572984 | 2439 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
2d091d62 | 2440 | |
73572984 JA |
2441 | eventfd_ctx_put(ev_fd->cq_ev_fd); |
2442 | kfree(ev_fd); | |
2d091d62 PB |
2443 | } |
2444 | ||
73572984 | 2445 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 2446 | { |
73572984 JA |
2447 | struct io_ev_fd *ev_fd; |
2448 | ||
2449 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, | |
2450 | lockdep_is_held(&ctx->uring_lock)); | |
2451 | if (ev_fd) { | |
2452 | ctx->has_evfd = false; | |
2453 | rcu_assign_pointer(ctx->io_ev_fd, NULL); | |
2454 | call_rcu(&ev_fd->rcu, io_eventfd_put); | |
2455 | return 0; | |
2456 | } | |
2d091d62 | 2457 | |
73572984 | 2458 | return -ENXIO; |
44b31f2f PB |
2459 | } |
2460 | ||
73572984 | 2461 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2462 | { |
cd0ca2e0 | 2463 | struct io_submit_state *state = &ctx->submit_state; |
37f0e767 | 2464 | int nr = 0; |
bf019da7 | 2465 | |
9a4fdbd8 | 2466 | mutex_lock(&ctx->uring_lock); |
cd0ca2e0 | 2467 | io_flush_cached_locked_reqs(ctx, state); |
9a4fdbd8 | 2468 | |
88ab95be | 2469 | while (!io_req_cache_empty(ctx)) { |
c2b6c6bc PB |
2470 | struct io_wq_work_node *node; |
2471 | struct io_kiocb *req; | |
9a4fdbd8 | 2472 | |
c2b6c6bc PB |
2473 | node = wq_stack_extract(&state->free_list); |
2474 | req = container_of(node, struct io_kiocb, comp_list); | |
2475 | kmem_cache_free(req_cachep, req); | |
37f0e767 | 2476 | nr++; |
c2b6c6bc | 2477 | } |
37f0e767 PB |
2478 | if (nr) |
2479 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2480 | mutex_unlock(&ctx->uring_lock); |
2481 | } | |
2482 | ||
4d9237e3 JA |
2483 | static void io_flush_apoll_cache(struct io_ring_ctx *ctx) |
2484 | { | |
2485 | struct async_poll *apoll; | |
2486 | ||
2487 | while (!list_empty(&ctx->apoll_cache)) { | |
2488 | apoll = list_first_entry(&ctx->apoll_cache, struct async_poll, | |
2489 | poll.wait.entry); | |
2490 | list_del(&apoll->poll.wait.entry); | |
2491 | kfree(apoll); | |
2492 | } | |
2493 | } | |
2494 | ||
c072481d | 2495 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2496 | { |
37d1e2e3 | 2497 | io_sq_thread_finish(ctx); |
2aede0e4 | 2498 | |
37d1e2e3 | 2499 | if (ctx->mm_account) { |
2aede0e4 JA |
2500 | mmdrop(ctx->mm_account); |
2501 | ctx->mm_account = NULL; | |
30975825 | 2502 | } |
def596e9 | 2503 | |
ab409402 | 2504 | io_rsrc_refs_drop(ctx); |
43597aac PB |
2505 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
2506 | io_wait_rsrc_data(ctx->buf_data); | |
2507 | io_wait_rsrc_data(ctx->file_data); | |
2508 | ||
8bad28d8 | 2509 | mutex_lock(&ctx->uring_lock); |
43597aac | 2510 | if (ctx->buf_data) |
bd54b6fe | 2511 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2512 | if (ctx->file_data) |
08480400 | 2513 | __io_sqe_files_unregister(ctx); |
c4ea060e PB |
2514 | if (ctx->rings) |
2515 | __io_cqring_overflow_flush(ctx, true); | |
9b402849 | 2516 | io_eventfd_unregister(ctx); |
4d9237e3 | 2517 | io_flush_apoll_cache(ctx); |
77bc59b4 | 2518 | mutex_unlock(&ctx->uring_lock); |
5a2e745d | 2519 | io_destroy_buffers(ctx); |
07db298a PB |
2520 | if (ctx->sq_creds) |
2521 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2522 | if (ctx->submitter_task) |
2523 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2524 | |
a7f0ed5a PB |
2525 | /* there are no registered resources left, nobody uses it */ |
2526 | if (ctx->rsrc_node) | |
2527 | io_rsrc_node_destroy(ctx->rsrc_node); | |
8dd03afe | 2528 | if (ctx->rsrc_backup_node) |
b895c9a6 | 2529 | io_rsrc_node_destroy(ctx->rsrc_backup_node); |
a7f0ed5a | 2530 | flush_delayed_work(&ctx->rsrc_put_work); |
756ab7c0 | 2531 | flush_delayed_work(&ctx->fallback_work); |
a7f0ed5a PB |
2532 | |
2533 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
2534 | WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist)); | |
def596e9 | 2535 | |
2b188cc1 | 2536 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
2537 | if (ctx->ring_sock) { |
2538 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 2539 | sock_release(ctx->ring_sock); |
355e8d26 | 2540 | } |
2b188cc1 | 2541 | #endif |
ef9dd637 | 2542 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
2b188cc1 | 2543 | |
75b28aff | 2544 | io_mem_free(ctx->rings); |
2b188cc1 | 2545 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
2546 | |
2547 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2548 | free_uid(ctx->user); |
4010fec4 | 2549 | io_req_caches_free(ctx); |
e941894e JA |
2550 | if (ctx->hash_map) |
2551 | io_wq_put_hash(ctx->hash_map); | |
e6f89be6 | 2552 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2553 | kfree(ctx->cancel_table_locked.hbs); |
6224843d | 2554 | kfree(ctx->dummy_ubuf); |
9cfc7e94 JA |
2555 | kfree(ctx->io_bl); |
2556 | xa_destroy(&ctx->io_bl_xa); | |
2b188cc1 JA |
2557 | kfree(ctx); |
2558 | } | |
2559 | ||
2560 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
2561 | { | |
2562 | struct io_ring_ctx *ctx = file->private_data; | |
2563 | __poll_t mask = 0; | |
2564 | ||
d60aa65b | 2565 | poll_wait(file, &ctx->cq_wait, wait); |
4f7067c3 SB |
2566 | /* |
2567 | * synchronizes with barrier from wq_has_sleeper call in | |
2568 | * io_commit_cqring | |
2569 | */ | |
2b188cc1 | 2570 | smp_rmb(); |
90554200 | 2571 | if (!io_sqring_full(ctx)) |
2b188cc1 | 2572 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
2573 | |
2574 | /* | |
2575 | * Don't flush cqring overflow list here, just do a simple check. | |
2576 | * Otherwise there could possible be ABBA deadlock: | |
2577 | * CPU0 CPU1 | |
2578 | * ---- ---- | |
2579 | * lock(&ctx->uring_lock); | |
2580 | * lock(&ep->mtx); | |
2581 | * lock(&ctx->uring_lock); | |
2582 | * lock(&ep->mtx); | |
2583 | * | |
2584 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
2585 | * pushs them to do the flush. | |
2586 | */ | |
10988a0a DY |
2587 | if (io_cqring_events(ctx) || |
2588 | test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) | |
2b188cc1 JA |
2589 | mask |= EPOLLIN | EPOLLRDNORM; |
2590 | ||
2591 | return mask; | |
2592 | } | |
2593 | ||
0bead8cd | 2594 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 2595 | { |
4379bf8b | 2596 | const struct cred *creds; |
071698e1 | 2597 | |
61cf9370 | 2598 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
2599 | if (creds) { |
2600 | put_cred(creds); | |
0bead8cd | 2601 | return 0; |
1e6fa521 | 2602 | } |
0bead8cd YD |
2603 | |
2604 | return -EINVAL; | |
2605 | } | |
2606 | ||
d56d938b PB |
2607 | struct io_tctx_exit { |
2608 | struct callback_head task_work; | |
2609 | struct completion completion; | |
baf186c4 | 2610 | struct io_ring_ctx *ctx; |
d56d938b PB |
2611 | }; |
2612 | ||
c072481d | 2613 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
2614 | { |
2615 | struct io_uring_task *tctx = current->io_uring; | |
2616 | struct io_tctx_exit *work; | |
2617 | ||
2618 | work = container_of(cb, struct io_tctx_exit, task_work); | |
2619 | /* | |
2620 | * When @in_idle, we're in cancellation and it's racy to remove the | |
2621 | * node. It'll be removed by the end of cancellation, just ignore it. | |
2622 | */ | |
2623 | if (!atomic_read(&tctx->in_idle)) | |
eef51daa | 2624 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
2625 | complete(&work->completion); |
2626 | } | |
2627 | ||
c072481d | 2628 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
2629 | { |
2630 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2631 | ||
2632 | return req->ctx == data; | |
2633 | } | |
2634 | ||
c072481d | 2635 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 2636 | { |
d56d938b | 2637 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 2638 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 2639 | unsigned long interval = HZ / 20; |
d56d938b PB |
2640 | struct io_tctx_exit exit; |
2641 | struct io_tctx_node *node; | |
2642 | int ret; | |
85faa7b8 | 2643 | |
56952e91 JA |
2644 | /* |
2645 | * If we're doing polled IO and end up having requests being | |
2646 | * submitted async (out-of-line), then completions can come in while | |
2647 | * we're waiting for refs to drop. We need to reap these manually, | |
2648 | * as nobody else will be looking for them. | |
2649 | */ | |
b2edc0a7 | 2650 | do { |
affa87db PB |
2651 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
2652 | cond_resched(); | |
2653 | ||
28090c13 PB |
2654 | if (ctx->sq_data) { |
2655 | struct io_sq_data *sqd = ctx->sq_data; | |
2656 | struct task_struct *tsk; | |
2657 | ||
2658 | io_sq_thread_park(sqd); | |
2659 | tsk = sqd->thread; | |
2660 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
2661 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
2662 | io_cancel_ctx_cb, ctx, true); | |
2663 | io_sq_thread_unpark(sqd); | |
2664 | } | |
b5bb3a24 | 2665 | |
37f0e767 PB |
2666 | io_req_caches_free(ctx); |
2667 | ||
58d3be2c PB |
2668 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
2669 | /* there is little hope left, don't run it too often */ | |
2670 | interval = HZ * 60; | |
2671 | } | |
2672 | } while (!wait_for_completion_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 2673 | |
7f00651a PB |
2674 | init_completion(&exit.completion); |
2675 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
2676 | exit.ctx = ctx; | |
89b5066e PB |
2677 | /* |
2678 | * Some may use context even when all refs and requests have been put, | |
2679 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 2680 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
2681 | * this lock/unlock section also waits them to finish. |
2682 | */ | |
d56d938b PB |
2683 | mutex_lock(&ctx->uring_lock); |
2684 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
2685 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
2686 | ||
d56d938b PB |
2687 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
2688 | ctx_node); | |
7f00651a PB |
2689 | /* don't spin on a single task if cancellation failed */ |
2690 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
2691 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
2692 | if (WARN_ON_ONCE(ret)) | |
2693 | continue; | |
d56d938b PB |
2694 | |
2695 | mutex_unlock(&ctx->uring_lock); | |
2696 | wait_for_completion(&exit.completion); | |
d56d938b PB |
2697 | mutex_lock(&ctx->uring_lock); |
2698 | } | |
2699 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
2700 | spin_lock(&ctx->completion_lock); |
2701 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 2702 | |
85faa7b8 JA |
2703 | io_ring_ctx_free(ctx); |
2704 | } | |
2705 | ||
c072481d | 2706 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 2707 | { |
61cf9370 MWO |
2708 | unsigned long index; |
2709 | struct creds *creds; | |
2710 | ||
2b188cc1 JA |
2711 | mutex_lock(&ctx->uring_lock); |
2712 | percpu_ref_kill(&ctx->refs); | |
634578f8 | 2713 | if (ctx->rings) |
6c2450ae | 2714 | __io_cqring_overflow_flush(ctx, true); |
61cf9370 MWO |
2715 | xa_for_each(&ctx->personalities, index, creds) |
2716 | io_unregister_personality(ctx, index); | |
9ca9fb24 PB |
2717 | if (ctx->rings) |
2718 | io_poll_remove_all(ctx, NULL, true); | |
2b188cc1 JA |
2719 | mutex_unlock(&ctx->uring_lock); |
2720 | ||
60053be8 PB |
2721 | /* failed during ring init, it couldn't have issued any requests */ |
2722 | if (ctx->rings) { | |
2723 | io_kill_timeouts(ctx, NULL, true); | |
60053be8 PB |
2724 | /* if we failed setting up the ctx, we might not have any rings */ |
2725 | io_iopoll_try_reap_events(ctx); | |
2726 | } | |
309fc03a | 2727 | |
85faa7b8 | 2728 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
2729 | /* |
2730 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
2731 | * if we're exiting a ton of rings at the same time. It just adds | |
2732 | * noise and overhead, there's no discernable change in runtime | |
2733 | * over using system_wq. | |
2734 | */ | |
2735 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
2736 | } |
2737 | ||
2738 | static int io_uring_release(struct inode *inode, struct file *file) | |
2739 | { | |
2740 | struct io_ring_ctx *ctx = file->private_data; | |
2741 | ||
2742 | file->private_data = NULL; | |
2743 | io_ring_ctx_wait_and_kill(ctx); | |
2744 | return 0; | |
2745 | } | |
2746 | ||
f6edbabb PB |
2747 | struct io_task_cancel { |
2748 | struct task_struct *task; | |
3dd0c97a | 2749 | bool all; |
f6edbabb | 2750 | }; |
f254ac04 | 2751 | |
f6edbabb | 2752 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 2753 | { |
9a472ef7 | 2754 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 2755 | struct io_task_cancel *cancel = data; |
9a472ef7 | 2756 | |
6af3f48b | 2757 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
2758 | } |
2759 | ||
c072481d PB |
2760 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
2761 | struct task_struct *task, | |
2762 | bool cancel_all) | |
b7ddce3c | 2763 | { |
e1915f76 | 2764 | struct io_defer_entry *de; |
b7ddce3c PB |
2765 | LIST_HEAD(list); |
2766 | ||
79ebeaee | 2767 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 2768 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 2769 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
2770 | list_cut_position(&list, &ctx->defer_list, &de->list); |
2771 | break; | |
2772 | } | |
2773 | } | |
79ebeaee | 2774 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
2775 | if (list_empty(&list)) |
2776 | return false; | |
b7ddce3c PB |
2777 | |
2778 | while (!list_empty(&list)) { | |
2779 | de = list_first_entry(&list, struct io_defer_entry, list); | |
2780 | list_del_init(&de->list); | |
f41db273 | 2781 | io_req_complete_failed(de->req, -ECANCELED); |
b7ddce3c PB |
2782 | kfree(de); |
2783 | } | |
e1915f76 | 2784 | return true; |
b7ddce3c PB |
2785 | } |
2786 | ||
c072481d | 2787 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
2788 | { |
2789 | struct io_tctx_node *node; | |
2790 | enum io_wq_cancel cret; | |
2791 | bool ret = false; | |
2792 | ||
2793 | mutex_lock(&ctx->uring_lock); | |
2794 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
2795 | struct io_uring_task *tctx = node->task->io_uring; | |
2796 | ||
2797 | /* | |
2798 | * io_wq will stay alive while we hold uring_lock, because it's | |
2799 | * killed after ctx nodes, which requires to take the lock. | |
2800 | */ | |
2801 | if (!tctx || !tctx->io_wq) | |
2802 | continue; | |
2803 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
2804 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
2805 | } | |
2806 | mutex_unlock(&ctx->uring_lock); | |
2807 | ||
2808 | return ret; | |
2809 | } | |
2810 | ||
affa87db | 2811 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
2812 | struct task_struct *task, |
2813 | bool cancel_all) | |
9936c7c2 | 2814 | { |
3dd0c97a | 2815 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 2816 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
2817 | enum io_wq_cancel cret; |
2818 | bool ret = false; | |
9936c7c2 | 2819 | |
60053be8 PB |
2820 | /* failed during ring init, it couldn't have issued any requests */ |
2821 | if (!ctx->rings) | |
affa87db | 2822 | return false; |
60053be8 | 2823 | |
affa87db PB |
2824 | if (!task) { |
2825 | ret |= io_uring_try_cancel_iowq(ctx); | |
2826 | } else if (tctx && tctx->io_wq) { | |
2827 | /* | |
2828 | * Cancels requests of all rings, not only @ctx, but | |
2829 | * it's fine as the task is in exit/exec. | |
2830 | */ | |
2831 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
2832 | &cancel, true); | |
2833 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
2834 | } | |
9936c7c2 | 2835 | |
affa87db PB |
2836 | /* SQPOLL thread does its own polling */ |
2837 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
2838 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
2839 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
2840 | io_iopoll_try_reap_events(ctx); | |
2841 | ret = true; | |
9936c7c2 | 2842 | } |
9936c7c2 | 2843 | } |
affa87db PB |
2844 | |
2845 | ret |= io_cancel_defer_files(ctx, task, cancel_all); | |
2846 | mutex_lock(&ctx->uring_lock); | |
2847 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
2848 | mutex_unlock(&ctx->uring_lock); | |
2849 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
2850 | if (task) | |
2851 | ret |= io_run_task_work(); | |
2852 | return ret; | |
9936c7c2 PB |
2853 | } |
2854 | ||
3f48cf18 | 2855 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 2856 | { |
3f48cf18 | 2857 | if (tracked) |
9cae36a0 | 2858 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
2859 | return percpu_counter_sum(&tctx->inflight); |
2860 | } | |
2861 | ||
78cc687b PB |
2862 | /* |
2863 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 2864 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 2865 | */ |
17437f31 | 2866 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 2867 | { |
521d6a73 | 2868 | struct io_uring_task *tctx = current->io_uring; |
734551df | 2869 | struct io_ring_ctx *ctx; |
0e9ddb39 PB |
2870 | s64 inflight; |
2871 | DEFINE_WAIT(wait); | |
fdaf083c | 2872 | |
78cc687b PB |
2873 | WARN_ON_ONCE(sqd && sqd->thread != current); |
2874 | ||
6d042ffb PO |
2875 | if (!current->io_uring) |
2876 | return; | |
17a91051 PB |
2877 | if (tctx->io_wq) |
2878 | io_wq_exit_start(tctx->io_wq); | |
2879 | ||
0e9ddb39 PB |
2880 | atomic_inc(&tctx->in_idle); |
2881 | do { | |
affa87db PB |
2882 | bool loop = false; |
2883 | ||
e9dbe221 | 2884 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 2885 | /* read completions before cancelations */ |
78cc687b | 2886 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
2887 | if (!inflight) |
2888 | break; | |
fdaf083c | 2889 | |
78cc687b PB |
2890 | if (!sqd) { |
2891 | struct io_tctx_node *node; | |
2892 | unsigned long index; | |
0f212204 | 2893 | |
78cc687b PB |
2894 | xa_for_each(&tctx->xa, index, node) { |
2895 | /* sqpoll task will cancel all its requests */ | |
2896 | if (node->ctx->sq_data) | |
2897 | continue; | |
affa87db PB |
2898 | loop |= io_uring_try_cancel_requests(node->ctx, |
2899 | current, cancel_all); | |
78cc687b PB |
2900 | } |
2901 | } else { | |
2902 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
2903 | loop |= io_uring_try_cancel_requests(ctx, |
2904 | current, | |
2905 | cancel_all); | |
2906 | } | |
2907 | ||
2908 | if (loop) { | |
2909 | cond_resched(); | |
2910 | continue; | |
78cc687b | 2911 | } |
17a91051 | 2912 | |
78a78060 JA |
2913 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
2914 | io_run_task_work(); | |
e9dbe221 | 2915 | io_uring_drop_tctx_refs(current); |
78a78060 | 2916 | |
0f212204 | 2917 | /* |
a1bb3cd5 PB |
2918 | * If we've seen completions, retry without waiting. This |
2919 | * avoids a race where a completion comes in before we did | |
2920 | * prepare_to_wait(). | |
0f212204 | 2921 | */ |
3dd0c97a | 2922 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 2923 | schedule(); |
f57555ed | 2924 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 2925 | } while (1); |
de7f1d9e | 2926 | |
8452d4a6 | 2927 | io_uring_clean_tctx(tctx); |
3dd0c97a | 2928 | if (cancel_all) { |
3cc7fdb9 PB |
2929 | /* |
2930 | * We shouldn't run task_works after cancel, so just leave | |
2931 | * ->in_idle set for normal exit. | |
2932 | */ | |
2933 | atomic_dec(&tctx->in_idle); | |
3f48cf18 PB |
2934 | /* for exec all current's requests should be gone, kill tctx */ |
2935 | __io_uring_free(current); | |
2936 | } | |
44e728b8 PB |
2937 | } |
2938 | ||
f552a27a | 2939 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 2940 | { |
f552a27a | 2941 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
2942 | } |
2943 | ||
6c5c240e RP |
2944 | static void *io_uring_validate_mmap_request(struct file *file, |
2945 | loff_t pgoff, size_t sz) | |
2b188cc1 | 2946 | { |
2b188cc1 | 2947 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 2948 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
2949 | struct page *page; |
2950 | void *ptr; | |
2951 | ||
2952 | switch (offset) { | |
2953 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
2954 | case IORING_OFF_CQ_RING: |
2955 | ptr = ctx->rings; | |
2b188cc1 JA |
2956 | break; |
2957 | case IORING_OFF_SQES: | |
2958 | ptr = ctx->sq_sqes; | |
2959 | break; | |
2b188cc1 | 2960 | default: |
6c5c240e | 2961 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
2962 | } |
2963 | ||
2964 | page = virt_to_head_page(ptr); | |
a50b854e | 2965 | if (sz > page_size(page)) |
6c5c240e RP |
2966 | return ERR_PTR(-EINVAL); |
2967 | ||
2968 | return ptr; | |
2969 | } | |
2970 | ||
2971 | #ifdef CONFIG_MMU | |
2972 | ||
c072481d | 2973 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e RP |
2974 | { |
2975 | size_t sz = vma->vm_end - vma->vm_start; | |
2976 | unsigned long pfn; | |
2977 | void *ptr; | |
2978 | ||
2979 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
2980 | if (IS_ERR(ptr)) | |
2981 | return PTR_ERR(ptr); | |
2b188cc1 JA |
2982 | |
2983 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
2984 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
2985 | } | |
2986 | ||
6c5c240e RP |
2987 | #else /* !CONFIG_MMU */ |
2988 | ||
2989 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
2990 | { | |
2991 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
2992 | } | |
2993 | ||
2994 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
2995 | { | |
2996 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
2997 | } | |
2998 | ||
2999 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
3000 | unsigned long addr, unsigned long len, | |
3001 | unsigned long pgoff, unsigned long flags) | |
3002 | { | |
3003 | void *ptr; | |
3004 | ||
3005 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
3006 | if (IS_ERR(ptr)) | |
3007 | return PTR_ERR(ptr); | |
3008 | ||
3009 | return (unsigned long) ptr; | |
3010 | } | |
3011 | ||
3012 | #endif /* !CONFIG_MMU */ | |
3013 | ||
f81440d3 PB |
3014 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
3015 | { | |
3016 | if (flags & IORING_ENTER_EXT_ARG) { | |
3017 | struct io_uring_getevents_arg arg; | |
3018 | ||
3019 | if (argsz != sizeof(arg)) | |
3020 | return -EINVAL; | |
3021 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3022 | return -EFAULT; | |
3023 | } | |
3024 | return 0; | |
3025 | } | |
3026 | ||
c73ebb68 HX |
3027 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
3028 | struct __kernel_timespec __user **ts, | |
3029 | const sigset_t __user **sig) | |
3030 | { | |
3031 | struct io_uring_getevents_arg arg; | |
3032 | ||
3033 | /* | |
3034 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
3035 | * is just a pointer to the sigset_t. | |
3036 | */ | |
3037 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
3038 | *sig = (const sigset_t __user *) argp; | |
3039 | *ts = NULL; | |
3040 | return 0; | |
3041 | } | |
3042 | ||
3043 | /* | |
3044 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
3045 | * timespec and sigset_t pointers if good. | |
3046 | */ | |
3047 | if (*argsz != sizeof(arg)) | |
3048 | return -EINVAL; | |
3049 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
3050 | return -EFAULT; | |
d2347b96 DY |
3051 | if (arg.pad) |
3052 | return -EINVAL; | |
c73ebb68 HX |
3053 | *sig = u64_to_user_ptr(arg.sigmask); |
3054 | *argsz = arg.sigmask_sz; | |
3055 | *ts = u64_to_user_ptr(arg.ts); | |
3056 | return 0; | |
3057 | } | |
3058 | ||
2b188cc1 | 3059 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
3060 | u32, min_complete, u32, flags, const void __user *, argp, |
3061 | size_t, argsz) | |
2b188cc1 JA |
3062 | { |
3063 | struct io_ring_ctx *ctx; | |
2b188cc1 | 3064 | struct fd f; |
33f993da | 3065 | long ret; |
2b188cc1 | 3066 | |
4c6e277c | 3067 | io_run_task_work(); |
b41e9852 | 3068 | |
33f993da | 3069 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
3070 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
3071 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
3072 | return -EINVAL; |
3073 | ||
e7a6c00d JA |
3074 | /* |
3075 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3076 | * need only dereference our task private array to find it. | |
3077 | */ | |
3078 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3079 | struct io_uring_task *tctx = current->io_uring; | |
3080 | ||
3081 | if (!tctx || fd >= IO_RINGFD_REG_MAX) | |
3082 | return -EINVAL; | |
3083 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
3084 | f.file = tctx->registered_rings[fd]; | |
4329490a | 3085 | f.flags = 0; |
e7a6c00d JA |
3086 | } else { |
3087 | f = fdget(fd); | |
e7a6c00d | 3088 | } |
2b188cc1 | 3089 | |
4329490a AV |
3090 | if (unlikely(!f.file)) |
3091 | return -EBADF; | |
3092 | ||
2b188cc1 | 3093 | ret = -EOPNOTSUPP; |
e5550a14 | 3094 | if (unlikely(!io_is_uring_fops(f.file))) |
2b188cc1 JA |
3095 | goto out_fput; |
3096 | ||
3097 | ret = -ENXIO; | |
3098 | ctx = f.file->private_data; | |
33f993da | 3099 | if (unlikely(!percpu_ref_tryget(&ctx->refs))) |
2b188cc1 JA |
3100 | goto out_fput; |
3101 | ||
7e84e1c7 | 3102 | ret = -EBADFD; |
33f993da | 3103 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3104 | goto out; |
3105 | ||
6c271ce2 JA |
3106 | /* |
3107 | * For SQ polling, the thread will do all submissions and completions. | |
3108 | * Just return the requested submit count, and wake the thread if | |
3109 | * we were asked to. | |
3110 | */ | |
b2a9eada | 3111 | ret = 0; |
6c271ce2 | 3112 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3113 | io_cqring_overflow_flush(ctx); |
89448c47 | 3114 | |
21f96522 JA |
3115 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3116 | ret = -EOWNERDEAD; | |
04147488 | 3117 | goto out; |
21f96522 | 3118 | } |
6c271ce2 | 3119 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3120 | wake_up(&ctx->sq_data->wait); |
d9d05217 PB |
3121 | if (flags & IORING_ENTER_SQ_WAIT) { |
3122 | ret = io_sqpoll_wait_sq(ctx); | |
3123 | if (ret) | |
3124 | goto out; | |
3125 | } | |
3e813c90 | 3126 | ret = to_submit; |
b2a9eada | 3127 | } else if (to_submit) { |
eef51daa | 3128 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3129 | if (unlikely(ret)) |
3130 | goto out; | |
7c504e65 | 3131 | |
2b188cc1 | 3132 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3133 | ret = io_submit_sqes(ctx, to_submit); |
3134 | if (ret != to_submit) { | |
d487b43c | 3135 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3136 | goto out; |
d487b43c PB |
3137 | } |
3138 | if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll) | |
3139 | goto iopoll_locked; | |
3140 | mutex_unlock(&ctx->uring_lock); | |
2b188cc1 JA |
3141 | } |
3142 | if (flags & IORING_ENTER_GETEVENTS) { | |
3e813c90 | 3143 | int ret2; |
773697b6 | 3144 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3145 | /* |
3146 | * We disallow the app entering submit/complete with | |
3147 | * polling, but we still need to lock the ring to | |
3148 | * prevent racing with polled issue that got punted to | |
3149 | * a workqueue. | |
3150 | */ | |
3151 | mutex_lock(&ctx->uring_lock); | |
3152 | iopoll_locked: | |
3e813c90 DY |
3153 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3154 | if (likely(!ret2)) { | |
3155 | min_complete = min(min_complete, | |
3156 | ctx->cq_entries); | |
3157 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3158 | } |
3159 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3160 | } else { |
f81440d3 PB |
3161 | const sigset_t __user *sig; |
3162 | struct __kernel_timespec __user *ts; | |
3163 | ||
3e813c90 DY |
3164 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3165 | if (likely(!ret2)) { | |
3166 | min_complete = min(min_complete, | |
3167 | ctx->cq_entries); | |
3168 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3169 | argsz, ts); | |
3170 | } | |
def596e9 | 3171 | } |
c73ebb68 | 3172 | |
155bc950 | 3173 | if (!ret) { |
3e813c90 | 3174 | ret = ret2; |
2b188cc1 | 3175 | |
155bc950 DY |
3176 | /* |
3177 | * EBADR indicates that one or more CQE were dropped. | |
3178 | * Once the user has been informed we can clear the bit | |
3179 | * as they are obviously ok with those drops. | |
3180 | */ | |
3181 | if (unlikely(ret2 == -EBADR)) | |
3182 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3183 | &ctx->check_cq); | |
def596e9 | 3184 | } |
2b188cc1 JA |
3185 | } |
3186 | ||
7c504e65 | 3187 | out: |
6805b32e | 3188 | percpu_ref_put(&ctx->refs); |
2b188cc1 | 3189 | out_fput: |
4329490a | 3190 | fdput(f); |
3e813c90 | 3191 | return ret; |
2b188cc1 JA |
3192 | } |
3193 | ||
3194 | static const struct file_operations io_uring_fops = { | |
3195 | .release = io_uring_release, | |
3196 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3197 | #ifndef CONFIG_MMU |
3198 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3199 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
3200 | #endif | |
2b188cc1 | 3201 | .poll = io_uring_poll, |
bebdb65e | 3202 | #ifdef CONFIG_PROC_FS |
87ce955b | 3203 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3204 | #endif |
2b188cc1 JA |
3205 | }; |
3206 | ||
92ac8bea JA |
3207 | bool io_is_uring_fops(struct file *file) |
3208 | { | |
3209 | return file->f_op == &io_uring_fops; | |
3210 | } | |
3211 | ||
c072481d PB |
3212 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3213 | struct io_uring_params *p) | |
2b188cc1 | 3214 | { |
75b28aff HV |
3215 | struct io_rings *rings; |
3216 | size_t size, sq_array_offset; | |
2b188cc1 | 3217 | |
bd740481 JA |
3218 | /* make sure these are sane, as we already accounted them */ |
3219 | ctx->sq_entries = p->sq_entries; | |
3220 | ctx->cq_entries = p->cq_entries; | |
3221 | ||
baf9cb64 | 3222 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3223 | if (size == SIZE_MAX) |
3224 | return -EOVERFLOW; | |
3225 | ||
3226 | rings = io_mem_alloc(size); | |
3227 | if (!rings) | |
2b188cc1 JA |
3228 | return -ENOMEM; |
3229 | ||
75b28aff HV |
3230 | ctx->rings = rings; |
3231 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
3232 | rings->sq_ring_mask = p->sq_entries - 1; | |
3233 | rings->cq_ring_mask = p->cq_entries - 1; | |
3234 | rings->sq_ring_entries = p->sq_entries; | |
3235 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3236 | |
ebdeb7c0 JA |
3237 | if (p->flags & IORING_SETUP_SQE128) |
3238 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3239 | else | |
3240 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
3241 | if (size == SIZE_MAX) { |
3242 | io_mem_free(ctx->rings); | |
3243 | ctx->rings = NULL; | |
2b188cc1 | 3244 | return -EOVERFLOW; |
eb065d30 | 3245 | } |
2b188cc1 JA |
3246 | |
3247 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
3248 | if (!ctx->sq_sqes) { |
3249 | io_mem_free(ctx->rings); | |
3250 | ctx->rings = NULL; | |
2b188cc1 | 3251 | return -ENOMEM; |
eb065d30 | 3252 | } |
2b188cc1 | 3253 | |
2b188cc1 JA |
3254 | return 0; |
3255 | } | |
3256 | ||
9faadcc8 PB |
3257 | static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) |
3258 | { | |
3259 | int ret, fd; | |
3260 | ||
3261 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3262 | if (fd < 0) | |
3263 | return fd; | |
3264 | ||
97bbdc06 | 3265 | ret = __io_uring_add_tctx_node(ctx, false); |
9faadcc8 PB |
3266 | if (ret) { |
3267 | put_unused_fd(fd); | |
3268 | return ret; | |
3269 | } | |
3270 | fd_install(fd, file); | |
3271 | return fd; | |
3272 | } | |
3273 | ||
2b188cc1 JA |
3274 | /* |
3275 | * Allocate an anonymous fd, this is what constitutes the application | |
3276 | * visible backing of an io_uring instance. The application mmaps this | |
3277 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3278 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3279 | */ | |
9faadcc8 | 3280 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
3281 | { |
3282 | struct file *file; | |
9faadcc8 | 3283 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
3284 | int ret; |
3285 | ||
2b188cc1 JA |
3286 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
3287 | &ctx->ring_sock); | |
3288 | if (ret) | |
9faadcc8 | 3289 | return ERR_PTR(ret); |
2b188cc1 JA |
3290 | #endif |
3291 | ||
91a9ab7c PM |
3292 | file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, |
3293 | O_RDWR | O_CLOEXEC, NULL); | |
2b188cc1 | 3294 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
3295 | if (IS_ERR(file)) { |
3296 | sock_release(ctx->ring_sock); | |
3297 | ctx->ring_sock = NULL; | |
3298 | } else { | |
3299 | ctx->ring_sock->file = file; | |
0f212204 | 3300 | } |
2b188cc1 | 3301 | #endif |
9faadcc8 | 3302 | return file; |
2b188cc1 JA |
3303 | } |
3304 | ||
c072481d PB |
3305 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3306 | struct io_uring_params __user *params) | |
2b188cc1 | 3307 | { |
2b188cc1 | 3308 | struct io_ring_ctx *ctx; |
9faadcc8 | 3309 | struct file *file; |
2b188cc1 JA |
3310 | int ret; |
3311 | ||
8110c1a6 | 3312 | if (!entries) |
2b188cc1 | 3313 | return -EINVAL; |
8110c1a6 JA |
3314 | if (entries > IORING_MAX_ENTRIES) { |
3315 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3316 | return -EINVAL; | |
3317 | entries = IORING_MAX_ENTRIES; | |
3318 | } | |
2b188cc1 JA |
3319 | |
3320 | /* | |
3321 | * Use twice as many entries for the CQ ring. It's possible for the | |
3322 | * application to drive a higher depth than the size of the SQ ring, | |
3323 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3324 | * some flexibility in overcommitting a bit. If the application has |
3325 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3326 | * of CQ ring entries manually. | |
2b188cc1 JA |
3327 | */ |
3328 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3329 | if (p->flags & IORING_SETUP_CQSIZE) { |
3330 | /* | |
3331 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3332 | * to a power-of-two, if it isn't already. We do NOT impose | |
3333 | * any cq vs sq ring sizing. | |
3334 | */ | |
eb2667b3 | 3335 | if (!p->cq_entries) |
33a107f0 | 3336 | return -EINVAL; |
8110c1a6 JA |
3337 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3338 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3339 | return -EINVAL; | |
3340 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3341 | } | |
eb2667b3 JQ |
3342 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3343 | if (p->cq_entries < p->sq_entries) | |
3344 | return -EINVAL; | |
33a107f0 JA |
3345 | } else { |
3346 | p->cq_entries = 2 * p->sq_entries; | |
3347 | } | |
2b188cc1 | 3348 | |
2b188cc1 | 3349 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3350 | if (!ctx) |
2b188cc1 | 3351 | return -ENOMEM; |
773697b6 PB |
3352 | |
3353 | /* | |
3354 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3355 | * space applications don't need to do io completion events | |
3356 | * polling again, they can rely on io_sq_thread to do polling | |
3357 | * work, which can reduce cpu usage and uring_lock contention. | |
3358 | */ | |
3359 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3360 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3361 | ctx->syscall_iopoll = 1; | |
3362 | ||
2b188cc1 | 3363 | ctx->compat = in_compat_syscall(); |
62e398be JA |
3364 | if (!capable(CAP_IPC_LOCK)) |
3365 | ctx->user = get_uid(current_user()); | |
2aede0e4 | 3366 | |
9f010507 | 3367 | /* |
e1169f06 JA |
3368 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3369 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3370 | */ |
e1169f06 JA |
3371 | ret = -EINVAL; |
3372 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3373 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 JA |
3374 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
3375 | IORING_SETUP_TASKRUN_FLAG)) | |
e1169f06 | 3376 | goto err; |
9f010507 | 3377 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3378 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3379 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3380 | } else { | |
ef060ea9 JA |
3381 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
3382 | goto err; | |
9f010507 | 3383 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3384 | } |
9f010507 | 3385 | |
2aede0e4 JA |
3386 | /* |
3387 | * This is just grabbed for accounting purposes. When a process exits, | |
3388 | * the mm is exited and dropped before the files, hence we need to hang | |
3389 | * on to this mm purely for the purposes of being able to unaccount | |
3390 | * memory (locked/pinned vm). It's not used for anything else. | |
3391 | */ | |
6b7898eb | 3392 | mmgrab(current->mm); |
2aede0e4 | 3393 | ctx->mm_account = current->mm; |
6b7898eb | 3394 | |
2b188cc1 JA |
3395 | ret = io_allocate_scq_urings(ctx, p); |
3396 | if (ret) | |
3397 | goto err; | |
3398 | ||
7e84e1c7 | 3399 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
3400 | if (ret) |
3401 | goto err; | |
eae071c9 | 3402 | /* always set a rsrc node */ |
47b228ce PB |
3403 | ret = io_rsrc_node_switch_start(ctx); |
3404 | if (ret) | |
3405 | goto err; | |
eae071c9 | 3406 | io_rsrc_node_switch(ctx, NULL); |
2b188cc1 | 3407 | |
2b188cc1 | 3408 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
3409 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3410 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3411 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3412 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3413 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3414 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
3415 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
3416 | |
3417 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
3418 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3419 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3420 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3421 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3422 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3423 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 3424 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 3425 | |
7f13657d XW |
3426 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
3427 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 3428 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 3429 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 3430 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e JA |
3431 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
3432 | IORING_FEAT_LINKED_FILE; | |
7f13657d XW |
3433 | |
3434 | if (copy_to_user(params, p, sizeof(*p))) { | |
3435 | ret = -EFAULT; | |
3436 | goto err; | |
3437 | } | |
d1719f70 | 3438 | |
9faadcc8 PB |
3439 | file = io_uring_get_file(ctx); |
3440 | if (IS_ERR(file)) { | |
3441 | ret = PTR_ERR(file); | |
3442 | goto err; | |
3443 | } | |
3444 | ||
044c1ab3 JA |
3445 | /* |
3446 | * Install ring fd as the very last thing, so we don't risk someone | |
3447 | * having closed it before we finish setup | |
3448 | */ | |
9faadcc8 PB |
3449 | ret = io_uring_install_fd(ctx, file); |
3450 | if (ret < 0) { | |
3451 | /* fput will clean it up */ | |
3452 | fput(file); | |
3453 | return ret; | |
3454 | } | |
044c1ab3 | 3455 | |
c826bd7a | 3456 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
3457 | return ret; |
3458 | err: | |
3459 | io_ring_ctx_wait_and_kill(ctx); | |
3460 | return ret; | |
3461 | } | |
3462 | ||
3463 | /* | |
3464 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3465 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3466 | * params structure passed in. | |
3467 | */ | |
3468 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3469 | { | |
3470 | struct io_uring_params p; | |
2b188cc1 JA |
3471 | int i; |
3472 | ||
3473 | if (copy_from_user(&p, params, sizeof(p))) | |
3474 | return -EFAULT; | |
3475 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3476 | if (p.resv[i]) | |
3477 | return -EINVAL; | |
3478 | } | |
3479 | ||
6c271ce2 | 3480 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 3481 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 3482 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 3483 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 3484 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 PB |
3485 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
3486 | IORING_SETUP_SINGLE_ISSUER)) | |
2b188cc1 JA |
3487 | return -EINVAL; |
3488 | ||
ef060ea9 | 3489 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
3490 | } |
3491 | ||
3492 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3493 | struct io_uring_params __user *, params) | |
3494 | { | |
3495 | return io_uring_setup(entries, params); | |
3496 | } | |
3497 | ||
c072481d PB |
3498 | static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, |
3499 | unsigned nr_args) | |
66f4af93 JA |
3500 | { |
3501 | struct io_uring_probe *p; | |
3502 | size_t size; | |
3503 | int i, ret; | |
3504 | ||
3505 | size = struct_size(p, ops, nr_args); | |
3506 | if (size == SIZE_MAX) | |
3507 | return -EOVERFLOW; | |
3508 | p = kzalloc(size, GFP_KERNEL); | |
3509 | if (!p) | |
3510 | return -ENOMEM; | |
3511 | ||
3512 | ret = -EFAULT; | |
3513 | if (copy_from_user(p, arg, size)) | |
3514 | goto out; | |
3515 | ret = -EINVAL; | |
3516 | if (memchr_inv(p, 0, size)) | |
3517 | goto out; | |
3518 | ||
3519 | p->last_op = IORING_OP_LAST - 1; | |
3520 | if (nr_args > IORING_OP_LAST) | |
3521 | nr_args = IORING_OP_LAST; | |
3522 | ||
3523 | for (i = 0; i < nr_args; i++) { | |
3524 | p->ops[i].op = i; | |
3525 | if (!io_op_defs[i].not_supported) | |
3526 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
3527 | } | |
3528 | p->ops_len = i; | |
3529 | ||
3530 | ret = 0; | |
3531 | if (copy_to_user(arg, p, size)) | |
3532 | ret = -EFAULT; | |
3533 | out: | |
3534 | kfree(p); | |
3535 | return ret; | |
3536 | } | |
3537 | ||
071698e1 JA |
3538 | static int io_register_personality(struct io_ring_ctx *ctx) |
3539 | { | |
4379bf8b | 3540 | const struct cred *creds; |
61cf9370 | 3541 | u32 id; |
1e6fa521 | 3542 | int ret; |
071698e1 | 3543 | |
4379bf8b | 3544 | creds = get_current_cred(); |
1e6fa521 | 3545 | |
61cf9370 MWO |
3546 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
3547 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
3548 | if (ret < 0) { |
3549 | put_cred(creds); | |
3550 | return ret; | |
3551 | } | |
3552 | return id; | |
071698e1 JA |
3553 | } |
3554 | ||
c072481d PB |
3555 | static __cold int io_register_restrictions(struct io_ring_ctx *ctx, |
3556 | void __user *arg, unsigned int nr_args) | |
21b55dbc SG |
3557 | { |
3558 | struct io_uring_restriction *res; | |
3559 | size_t size; | |
3560 | int i, ret; | |
3561 | ||
7e84e1c7 SG |
3562 | /* Restrictions allowed only if rings started disabled */ |
3563 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3564 | return -EBADFD; | |
3565 | ||
21b55dbc | 3566 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 3567 | if (ctx->restrictions.registered) |
21b55dbc SG |
3568 | return -EBUSY; |
3569 | ||
3570 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
3571 | return -EINVAL; | |
3572 | ||
3573 | size = array_size(nr_args, sizeof(*res)); | |
3574 | if (size == SIZE_MAX) | |
3575 | return -EOVERFLOW; | |
3576 | ||
3577 | res = memdup_user(arg, size); | |
3578 | if (IS_ERR(res)) | |
3579 | return PTR_ERR(res); | |
3580 | ||
3581 | ret = 0; | |
3582 | ||
3583 | for (i = 0; i < nr_args; i++) { | |
3584 | switch (res[i].opcode) { | |
3585 | case IORING_RESTRICTION_REGISTER_OP: | |
3586 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
3587 | ret = -EINVAL; | |
3588 | goto out; | |
3589 | } | |
3590 | ||
3591 | __set_bit(res[i].register_op, | |
3592 | ctx->restrictions.register_op); | |
3593 | break; | |
3594 | case IORING_RESTRICTION_SQE_OP: | |
3595 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
3596 | ret = -EINVAL; | |
3597 | goto out; | |
3598 | } | |
3599 | ||
3600 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
3601 | break; | |
3602 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
3603 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
3604 | break; | |
3605 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
3606 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
3607 | break; | |
3608 | default: | |
3609 | ret = -EINVAL; | |
3610 | goto out; | |
3611 | } | |
3612 | } | |
3613 | ||
3614 | out: | |
3615 | /* Reset all restrictions if an error happened */ | |
3616 | if (ret != 0) | |
3617 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
3618 | else | |
7e84e1c7 | 3619 | ctx->restrictions.registered = true; |
21b55dbc SG |
3620 | |
3621 | kfree(res); | |
3622 | return ret; | |
3623 | } | |
3624 | ||
7e84e1c7 SG |
3625 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
3626 | { | |
3627 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3628 | return -EBADFD; | |
3629 | ||
3630 | if (ctx->restrictions.registered) | |
3631 | ctx->restricted = 1; | |
3632 | ||
0298ef96 PB |
3633 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
3634 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
3635 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
3636 | return 0; |
3637 | } | |
3638 | ||
c072481d PB |
3639 | static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, |
3640 | void __user *arg, unsigned len) | |
fe76421d JA |
3641 | { |
3642 | struct io_uring_task *tctx = current->io_uring; | |
3643 | cpumask_var_t new_mask; | |
3644 | int ret; | |
3645 | ||
3646 | if (!tctx || !tctx->io_wq) | |
3647 | return -EINVAL; | |
3648 | ||
3649 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
3650 | return -ENOMEM; | |
3651 | ||
3652 | cpumask_clear(new_mask); | |
3653 | if (len > cpumask_size()) | |
3654 | len = cpumask_size(); | |
3655 | ||
0f5e4b83 ES |
3656 | if (in_compat_syscall()) { |
3657 | ret = compat_get_bitmap(cpumask_bits(new_mask), | |
3658 | (const compat_ulong_t __user *)arg, | |
3659 | len * 8 /* CHAR_BIT */); | |
3660 | } else { | |
3661 | ret = copy_from_user(new_mask, arg, len); | |
3662 | } | |
3663 | ||
3664 | if (ret) { | |
fe76421d JA |
3665 | free_cpumask_var(new_mask); |
3666 | return -EFAULT; | |
3667 | } | |
3668 | ||
3669 | ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); | |
3670 | free_cpumask_var(new_mask); | |
3671 | return ret; | |
3672 | } | |
3673 | ||
c072481d | 3674 | static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) |
fe76421d JA |
3675 | { |
3676 | struct io_uring_task *tctx = current->io_uring; | |
3677 | ||
3678 | if (!tctx || !tctx->io_wq) | |
3679 | return -EINVAL; | |
3680 | ||
3681 | return io_wq_cpu_affinity(tctx->io_wq, NULL); | |
3682 | } | |
3683 | ||
c072481d PB |
3684 | static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, |
3685 | void __user *arg) | |
b22fa62a | 3686 | __must_hold(&ctx->uring_lock) |
2e480058 | 3687 | { |
b22fa62a | 3688 | struct io_tctx_node *node; |
fa84693b JA |
3689 | struct io_uring_task *tctx = NULL; |
3690 | struct io_sq_data *sqd = NULL; | |
2e480058 JA |
3691 | __u32 new_count[2]; |
3692 | int i, ret; | |
3693 | ||
2e480058 JA |
3694 | if (copy_from_user(new_count, arg, sizeof(new_count))) |
3695 | return -EFAULT; | |
3696 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
3697 | if (new_count[i] > INT_MAX) | |
3698 | return -EINVAL; | |
3699 | ||
fa84693b JA |
3700 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3701 | sqd = ctx->sq_data; | |
3702 | if (sqd) { | |
009ad9f0 JA |
3703 | /* |
3704 | * Observe the correct sqd->lock -> ctx->uring_lock | |
3705 | * ordering. Fine to drop uring_lock here, we hold | |
3706 | * a ref to the ctx. | |
3707 | */ | |
41d3a6bd | 3708 | refcount_inc(&sqd->refs); |
009ad9f0 | 3709 | mutex_unlock(&ctx->uring_lock); |
fa84693b | 3710 | mutex_lock(&sqd->lock); |
009ad9f0 | 3711 | mutex_lock(&ctx->uring_lock); |
41d3a6bd JA |
3712 | if (sqd->thread) |
3713 | tctx = sqd->thread->io_uring; | |
fa84693b JA |
3714 | } |
3715 | } else { | |
3716 | tctx = current->io_uring; | |
3717 | } | |
3718 | ||
e139a1ec | 3719 | BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); |
fa84693b | 3720 | |
bad119b9 PB |
3721 | for (i = 0; i < ARRAY_SIZE(new_count); i++) |
3722 | if (new_count[i]) | |
3723 | ctx->iowq_limits[i] = new_count[i]; | |
e139a1ec PB |
3724 | ctx->iowq_limits_set = true; |
3725 | ||
e139a1ec PB |
3726 | if (tctx && tctx->io_wq) { |
3727 | ret = io_wq_max_workers(tctx->io_wq, new_count); | |
3728 | if (ret) | |
3729 | goto err; | |
3730 | } else { | |
3731 | memset(new_count, 0, sizeof(new_count)); | |
3732 | } | |
fa84693b | 3733 | |
41d3a6bd | 3734 | if (sqd) { |
fa84693b | 3735 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
3736 | io_put_sq_data(sqd); |
3737 | } | |
2e480058 JA |
3738 | |
3739 | if (copy_to_user(arg, new_count, sizeof(new_count))) | |
3740 | return -EFAULT; | |
3741 | ||
b22fa62a PB |
3742 | /* that's it for SQPOLL, only the SQPOLL task creates requests */ |
3743 | if (sqd) | |
3744 | return 0; | |
3745 | ||
3746 | /* now propagate the restriction to all registered users */ | |
3747 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3748 | struct io_uring_task *tctx = node->task->io_uring; | |
3749 | ||
3750 | if (WARN_ON_ONCE(!tctx->io_wq)) | |
3751 | continue; | |
3752 | ||
3753 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
3754 | new_count[i] = ctx->iowq_limits[i]; | |
3755 | /* ignore errors, it always returns zero anyway */ | |
3756 | (void)io_wq_max_workers(tctx->io_wq, new_count); | |
3757 | } | |
2e480058 | 3758 | return 0; |
fa84693b | 3759 | err: |
41d3a6bd | 3760 | if (sqd) { |
fa84693b | 3761 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
3762 | io_put_sq_data(sqd); |
3763 | } | |
fa84693b | 3764 | return ret; |
2e480058 JA |
3765 | } |
3766 | ||
edafccee JA |
3767 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
3768 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
3769 | __releases(ctx->uring_lock) |
3770 | __acquires(ctx->uring_lock) | |
edafccee JA |
3771 | { |
3772 | int ret; | |
3773 | ||
35fa71a0 JA |
3774 | /* |
3775 | * We're inside the ring mutex, if the ref is already dying, then | |
3776 | * someone else killed the ctx or is already going through | |
3777 | * io_uring_register(). | |
3778 | */ | |
3779 | if (percpu_ref_is_dying(&ctx->refs)) | |
3780 | return -ENXIO; | |
3781 | ||
75c4021a PB |
3782 | if (ctx->restricted) { |
3783 | if (opcode >= IORING_REGISTER_LAST) | |
3784 | return -EINVAL; | |
3785 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); | |
3786 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
3787 | return -EACCES; | |
3788 | } | |
3789 | ||
edafccee JA |
3790 | switch (opcode) { |
3791 | case IORING_REGISTER_BUFFERS: | |
0184f08e PB |
3792 | ret = -EFAULT; |
3793 | if (!arg) | |
3794 | break; | |
634d00df | 3795 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
3796 | break; |
3797 | case IORING_UNREGISTER_BUFFERS: | |
3798 | ret = -EINVAL; | |
3799 | if (arg || nr_args) | |
3800 | break; | |
0a96bbe4 | 3801 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 3802 | break; |
6b06314c | 3803 | case IORING_REGISTER_FILES: |
a8da73a3 JA |
3804 | ret = -EFAULT; |
3805 | if (!arg) | |
3806 | break; | |
792e3582 | 3807 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
3808 | break; |
3809 | case IORING_UNREGISTER_FILES: | |
3810 | ret = -EINVAL; | |
3811 | if (arg || nr_args) | |
3812 | break; | |
3813 | ret = io_sqe_files_unregister(ctx); | |
3814 | break; | |
c3a31e60 | 3815 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 3816 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 3817 | break; |
9b402849 JA |
3818 | case IORING_REGISTER_EVENTFD: |
3819 | ret = -EINVAL; | |
3820 | if (nr_args != 1) | |
3821 | break; | |
c75312dd UA |
3822 | ret = io_eventfd_register(ctx, arg, 0); |
3823 | break; | |
3824 | case IORING_REGISTER_EVENTFD_ASYNC: | |
3825 | ret = -EINVAL; | |
3826 | if (nr_args != 1) | |
f2842ab5 | 3827 | break; |
c75312dd | 3828 | ret = io_eventfd_register(ctx, arg, 1); |
9b402849 JA |
3829 | break; |
3830 | case IORING_UNREGISTER_EVENTFD: | |
3831 | ret = -EINVAL; | |
3832 | if (arg || nr_args) | |
3833 | break; | |
3834 | ret = io_eventfd_unregister(ctx); | |
3835 | break; | |
66f4af93 JA |
3836 | case IORING_REGISTER_PROBE: |
3837 | ret = -EINVAL; | |
3838 | if (!arg || nr_args > 256) | |
3839 | break; | |
3840 | ret = io_probe(ctx, arg, nr_args); | |
3841 | break; | |
071698e1 JA |
3842 | case IORING_REGISTER_PERSONALITY: |
3843 | ret = -EINVAL; | |
3844 | if (arg || nr_args) | |
3845 | break; | |
3846 | ret = io_register_personality(ctx); | |
3847 | break; | |
3848 | case IORING_UNREGISTER_PERSONALITY: | |
3849 | ret = -EINVAL; | |
3850 | if (arg) | |
3851 | break; | |
3852 | ret = io_unregister_personality(ctx, nr_args); | |
3853 | break; | |
7e84e1c7 SG |
3854 | case IORING_REGISTER_ENABLE_RINGS: |
3855 | ret = -EINVAL; | |
3856 | if (arg || nr_args) | |
3857 | break; | |
3858 | ret = io_register_enable_rings(ctx); | |
3859 | break; | |
21b55dbc SG |
3860 | case IORING_REGISTER_RESTRICTIONS: |
3861 | ret = io_register_restrictions(ctx, arg, nr_args); | |
3862 | break; | |
992da01a PB |
3863 | case IORING_REGISTER_FILES2: |
3864 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
3865 | break; | |
3866 | case IORING_REGISTER_FILES_UPDATE2: | |
3867 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
3868 | IORING_RSRC_FILE); | |
3869 | break; | |
3870 | case IORING_REGISTER_BUFFERS2: | |
3871 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 3872 | break; |
992da01a PB |
3873 | case IORING_REGISTER_BUFFERS_UPDATE: |
3874 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
3875 | IORING_RSRC_BUFFER); | |
c3bdad02 | 3876 | break; |
fe76421d JA |
3877 | case IORING_REGISTER_IOWQ_AFF: |
3878 | ret = -EINVAL; | |
3879 | if (!arg || !nr_args) | |
3880 | break; | |
3881 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
3882 | break; | |
3883 | case IORING_UNREGISTER_IOWQ_AFF: | |
3884 | ret = -EINVAL; | |
3885 | if (arg || nr_args) | |
3886 | break; | |
3887 | ret = io_unregister_iowq_aff(ctx); | |
3888 | break; | |
2e480058 JA |
3889 | case IORING_REGISTER_IOWQ_MAX_WORKERS: |
3890 | ret = -EINVAL; | |
3891 | if (!arg || nr_args != 2) | |
3892 | break; | |
3893 | ret = io_register_iowq_max_workers(ctx, arg); | |
3894 | break; | |
e7a6c00d JA |
3895 | case IORING_REGISTER_RING_FDS: |
3896 | ret = io_ringfd_register(ctx, arg, nr_args); | |
3897 | break; | |
3898 | case IORING_UNREGISTER_RING_FDS: | |
3899 | ret = io_ringfd_unregister(ctx, arg, nr_args); | |
3900 | break; | |
c7fb1942 JA |
3901 | case IORING_REGISTER_PBUF_RING: |
3902 | ret = -EINVAL; | |
3903 | if (!arg || nr_args != 1) | |
3904 | break; | |
3905 | ret = io_register_pbuf_ring(ctx, arg); | |
3906 | break; | |
3907 | case IORING_UNREGISTER_PBUF_RING: | |
3908 | ret = -EINVAL; | |
3909 | if (!arg || nr_args != 1) | |
3910 | break; | |
3911 | ret = io_unregister_pbuf_ring(ctx, arg); | |
3912 | break; | |
edafccee JA |
3913 | default: |
3914 | ret = -EINVAL; | |
3915 | break; | |
3916 | } | |
3917 | ||
edafccee JA |
3918 | return ret; |
3919 | } | |
3920 | ||
3921 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
3922 | void __user *, arg, unsigned int, nr_args) | |
3923 | { | |
3924 | struct io_ring_ctx *ctx; | |
3925 | long ret = -EBADF; | |
3926 | struct fd f; | |
3927 | ||
3928 | f = fdget(fd); | |
3929 | if (!f.file) | |
3930 | return -EBADF; | |
3931 | ||
3932 | ret = -EOPNOTSUPP; | |
e5550a14 | 3933 | if (!io_is_uring_fops(f.file)) |
edafccee JA |
3934 | goto out_fput; |
3935 | ||
3936 | ctx = f.file->private_data; | |
3937 | ||
b6c23dd5 PB |
3938 | io_run_task_work(); |
3939 | ||
edafccee JA |
3940 | mutex_lock(&ctx->uring_lock); |
3941 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
3942 | mutex_unlock(&ctx->uring_lock); | |
2757be22 | 3943 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); |
edafccee JA |
3944 | out_fput: |
3945 | fdput(f); | |
3946 | return ret; | |
3947 | } | |
3948 | ||
2b188cc1 JA |
3949 | static int __init io_uring_init(void) |
3950 | { | |
d7f62e82 SM |
3951 | #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \ |
3952 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ | |
3953 | BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \ | |
3954 | } while (0) | |
3955 | ||
3956 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
3957 | __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename) | |
3958 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); | |
3959 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
3960 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
3961 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
3962 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
3963 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
3964 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
3965 | BUILD_BUG_SQE_ELEM(16, __u64, addr); | |
7d67af2c | 3966 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
3967 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
3968 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
3969 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
3970 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
3971 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
3972 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
3973 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
3974 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
3975 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
3976 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
3977 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
3978 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
3979 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
3980 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
3981 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 3982 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
d7f62e82 SM |
3983 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
3984 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 3985 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 3986 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 3987 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 3988 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
e9621e2b | 3989 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
d7f62e82 | 3990 | |
b0d658ec PB |
3991 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
3992 | sizeof(struct io_uring_rsrc_update)); | |
3993 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
3994 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
3995 | |
3996 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
3997 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
3998 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
3999 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 4000 | |
b0d658ec PB |
4001 | /* should fit into one byte */ |
4002 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
4003 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
4004 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 4005 | |
32c2d33e | 4006 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); |
16340eab | 4007 | |
3a4b89a2 JA |
4008 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
4009 | ||
d9b57aa3 | 4010 | io_uring_optable_init(); |
0702e536 | 4011 | |
91f245d5 JA |
4012 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | |
4013 | SLAB_ACCOUNT); | |
2b188cc1 JA |
4014 | return 0; |
4015 | }; | |
4016 | __initcall(io_uring_init); |