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 | 475 | struct io_ev_fd *ev_fd; |
305bef98 PB |
476 | bool skip; |
477 | ||
478 | spin_lock(&ctx->completion_lock); | |
479 | /* | |
480 | * Eventfd should only get triggered when at least one event has been | |
481 | * posted. Some applications rely on the eventfd notification count only | |
482 | * changing IFF a new CQE has been added to the CQ ring. There's no | |
483 | * depedency on 1:1 relationship between how many times this function is | |
484 | * called (and hence the eventfd count) and number of CQEs posted to the | |
485 | * CQ ring. | |
486 | */ | |
487 | skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
488 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
489 | spin_unlock(&ctx->completion_lock); | |
490 | if (skip) | |
491 | return; | |
77bc59b4 | 492 | |
77bc59b4 UA |
493 | rcu_read_lock(); |
494 | /* | |
495 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
496 | * and eventfd_signal | |
497 | */ | |
498 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
499 | ||
500 | /* | |
501 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
502 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
503 | * the function and rcu_read_lock. | |
504 | */ | |
505 | if (unlikely(!ev_fd)) | |
506 | goto out; | |
7e55a19c | 507 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 UA |
508 | goto out; |
509 | ||
c75312dd | 510 | if (!ev_fd->eventfd_async || io_wq_current_is_worker()) |
77bc59b4 | 511 | eventfd_signal(ev_fd->cq_ev_fd, 1); |
77bc59b4 UA |
512 | out: |
513 | rcu_read_unlock(); | |
f2842ab5 JA |
514 | } |
515 | ||
a830ffd2 PB |
516 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
517 | { | |
518 | if (ctx->off_timeout_used || ctx->drain_active) { | |
519 | spin_lock(&ctx->completion_lock); | |
520 | if (ctx->off_timeout_used) | |
521 | io_flush_timeouts(ctx); | |
522 | if (ctx->drain_active) | |
523 | io_queue_deferred(ctx); | |
524 | spin_unlock(&ctx->completion_lock); | |
525 | } | |
526 | if (ctx->has_evfd) | |
527 | io_eventfd_signal(ctx); | |
528 | } | |
529 | ||
25399321 | 530 | static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 | 531 | { |
46929b08 | 532 | io_commit_cqring_flush(ctx); |
9aa8dfde | 533 | io_cqring_wake(ctx); |
1d7bb1d5 JA |
534 | } |
535 | ||
25399321 PB |
536 | static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) |
537 | __releases(ctx->completion_lock) | |
538 | { | |
539 | io_commit_cqring(ctx); | |
540 | spin_unlock(&ctx->completion_lock); | |
541 | io_cqring_ev_posted(ctx); | |
542 | } | |
543 | ||
544 | void io_cq_unlock_post(struct io_ring_ctx *ctx) | |
545 | { | |
546 | __io_cq_unlock_post(ctx); | |
547 | } | |
548 | ||
c4a2ed72 | 549 | /* Returns true if there are no backlogged entries after the flush */ |
6c2450ae | 550 | static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) |
1d7bb1d5 | 551 | { |
305bef98 | 552 | bool all_flushed; |
e45a3e05 | 553 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 554 | |
a566c556 | 555 | if (!force && __io_cqring_events(ctx) == ctx->cq_entries) |
e23de15f | 556 | return false; |
1d7bb1d5 | 557 | |
e45a3e05 SR |
558 | if (ctx->flags & IORING_SETUP_CQE32) |
559 | cqe_size <<= 1; | |
560 | ||
25399321 | 561 | io_cq_lock(ctx); |
6c2450ae | 562 | while (!list_empty(&ctx->cq_overflow_list)) { |
d068b506 | 563 | struct io_uring_cqe *cqe = io_get_cqe(ctx); |
6c2450ae | 564 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 565 | |
1d7bb1d5 JA |
566 | if (!cqe && !force) |
567 | break; | |
6c2450ae PB |
568 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
569 | struct io_overflow_cqe, list); | |
570 | if (cqe) | |
e45a3e05 | 571 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae | 572 | else |
8f6ed49a PB |
573 | io_account_cq_overflow(ctx); |
574 | ||
6c2450ae PB |
575 | list_del(&ocqe->list); |
576 | kfree(ocqe); | |
1d7bb1d5 JA |
577 | } |
578 | ||
09e88404 PB |
579 | all_flushed = list_empty(&ctx->cq_overflow_list); |
580 | if (all_flushed) { | |
10988a0a | 581 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 582 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 583 | } |
46930143 | 584 | |
25399321 | 585 | io_cq_unlock_post(ctx); |
09e88404 | 586 | return all_flushed; |
1d7bb1d5 JA |
587 | } |
588 | ||
90f67366 | 589 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 590 | { |
ca0a2651 JA |
591 | bool ret = true; |
592 | ||
10988a0a | 593 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
6c503150 PB |
594 | /* iopoll syncs against uring_lock, not completion_lock */ |
595 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
596 | mutex_lock(&ctx->uring_lock); | |
90f67366 | 597 | ret = __io_cqring_overflow_flush(ctx, false); |
6c503150 PB |
598 | if (ctx->flags & IORING_SETUP_IOPOLL) |
599 | mutex_unlock(&ctx->uring_lock); | |
600 | } | |
ca0a2651 JA |
601 | |
602 | return ret; | |
6c503150 PB |
603 | } |
604 | ||
9d170164 | 605 | static void __io_put_task(struct task_struct *task, int nr) |
6a290a14 PB |
606 | { |
607 | struct io_uring_task *tctx = task->io_uring; | |
608 | ||
9d170164 PB |
609 | percpu_counter_sub(&tctx->inflight, nr); |
610 | if (unlikely(atomic_read(&tctx->in_idle))) | |
611 | wake_up(&tctx->wait); | |
612 | put_task_struct_many(task, nr); | |
613 | } | |
614 | ||
615 | /* must to be called somewhat shortly after putting a request */ | |
616 | static inline void io_put_task(struct task_struct *task, int nr) | |
617 | { | |
618 | if (likely(task == current)) | |
619 | task->io_uring->cached_refs += nr; | |
620 | else | |
621 | __io_put_task(task, nr); | |
6a290a14 PB |
622 | } |
623 | ||
9a10867a PB |
624 | static void io_task_refs_refill(struct io_uring_task *tctx) |
625 | { | |
626 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
627 | ||
628 | percpu_counter_add(&tctx->inflight, refill); | |
629 | refcount_add(refill, ¤t->usage); | |
630 | tctx->cached_refs += refill; | |
631 | } | |
632 | ||
633 | static inline void io_get_task_refs(int nr) | |
634 | { | |
635 | struct io_uring_task *tctx = current->io_uring; | |
636 | ||
637 | tctx->cached_refs -= nr; | |
638 | if (unlikely(tctx->cached_refs < 0)) | |
639 | io_task_refs_refill(tctx); | |
640 | } | |
641 | ||
3cc7fdb9 PB |
642 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
643 | { | |
644 | struct io_uring_task *tctx = task->io_uring; | |
645 | unsigned int refs = tctx->cached_refs; | |
646 | ||
647 | if (refs) { | |
648 | tctx->cached_refs = 0; | |
649 | percpu_counter_sub(&tctx->inflight, refs); | |
650 | put_task_struct_many(task, refs); | |
651 | } | |
652 | } | |
653 | ||
68494a65 PB |
654 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
655 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 656 | { |
cce4b8b0 | 657 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
658 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
659 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 660 | |
e45a3e05 SR |
661 | if (is_cqe32) |
662 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 663 | |
e45a3e05 | 664 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 665 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
666 | if (!ocqe) { |
667 | /* | |
668 | * If we're in ring overflow flush mode, or in task cancel mode, | |
669 | * or cannot allocate an overflow entry, then we need to drop it | |
670 | * on the floor. | |
671 | */ | |
8f6ed49a | 672 | io_account_cq_overflow(ctx); |
155bc950 | 673 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 674 | return false; |
2b188cc1 | 675 | } |
cce4b8b0 | 676 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 677 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 678 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 679 | |
cce4b8b0 | 680 | } |
d4d19c19 | 681 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
682 | ocqe->cqe.res = res; |
683 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
684 | if (is_cqe32) { |
685 | ocqe->cqe.big_cqe[0] = extra1; | |
686 | ocqe->cqe.big_cqe[1] = extra2; | |
687 | } | |
cce4b8b0 PB |
688 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
689 | return true; | |
2b188cc1 JA |
690 | } |
691 | ||
68494a65 PB |
692 | bool io_req_cqe_overflow(struct io_kiocb *req) |
693 | { | |
694 | if (!(req->flags & REQ_F_CQE32_INIT)) { | |
695 | req->extra1 = 0; | |
696 | req->extra2 = 0; | |
697 | } | |
698 | return io_cqring_event_overflow(req->ctx, req->cqe.user_data, | |
699 | req->cqe.res, req->cqe.flags, | |
700 | req->extra1, req->extra2); | |
701 | } | |
702 | ||
faf88dde PB |
703 | /* |
704 | * writes to the cq entry need to come after reading head; the | |
705 | * control dependency is enough as we're using WRITE_ONCE to | |
706 | * fill the cq entry | |
707 | */ | |
708 | struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx) | |
709 | { | |
710 | struct io_rings *rings = ctx->rings; | |
711 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
712 | unsigned int free, queued, len; |
713 | ||
faf88dde PB |
714 | |
715 | /* userspace may cheat modifying the tail, be safe and do min */ | |
716 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
717 | free = ctx->cq_entries - queued; | |
718 | /* we need a contiguous range, limit based on the current array offset */ | |
719 | len = min(free, ctx->cq_entries - off); | |
720 | if (!len) | |
721 | return NULL; | |
722 | ||
b3659a65 PB |
723 | if (ctx->flags & IORING_SETUP_CQE32) { |
724 | off <<= 1; | |
725 | len <<= 1; | |
726 | } | |
727 | ||
faf88dde PB |
728 | ctx->cqe_cached = &rings->cqes[off]; |
729 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
b3659a65 PB |
730 | |
731 | ctx->cached_cq_tail++; | |
faf88dde | 732 | ctx->cqe_cached++; |
b3659a65 PB |
733 | if (ctx->flags & IORING_SETUP_CQE32) |
734 | ctx->cqe_cached++; | |
735 | return &rings->cqes[off]; | |
faf88dde PB |
736 | } |
737 | ||
d245bca6 PB |
738 | static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, |
739 | u64 user_data, s32 res, u32 cflags) | |
bcda7baa | 740 | { |
cd94903d PB |
741 | struct io_uring_cqe *cqe; |
742 | ||
913a571a | 743 | ctx->cq_extra++; |
c4bb964f | 744 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
cd94903d PB |
745 | |
746 | /* | |
747 | * If we can't get a cq entry, userspace overflowed the | |
748 | * submission (by quite a lot). Increment the overflow count in | |
749 | * the ring. | |
750 | */ | |
751 | cqe = io_get_cqe(ctx); | |
752 | if (likely(cqe)) { | |
753 | WRITE_ONCE(cqe->user_data, user_data); | |
754 | WRITE_ONCE(cqe->res, res); | |
755 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
756 | |
757 | if (ctx->flags & IORING_SETUP_CQE32) { | |
758 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
759 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
760 | } | |
cd94903d PB |
761 | return true; |
762 | } | |
763 | return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); | |
bcda7baa JA |
764 | } |
765 | ||
d245bca6 PB |
766 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, |
767 | u64 user_data, s32 res, u32 cflags) | |
768 | { | |
769 | bool filled; | |
770 | ||
25399321 | 771 | io_cq_lock(ctx); |
d245bca6 | 772 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags); |
25399321 | 773 | io_cq_unlock_post(ctx); |
d245bca6 PB |
774 | return filled; |
775 | } | |
776 | ||
effcf8bd | 777 | static void __io_req_complete_put(struct io_kiocb *req) |
2b188cc1 | 778 | { |
c7dae4ba JA |
779 | /* |
780 | * If we're the last reference to this request, add to our locked | |
781 | * free_list cache. | |
782 | */ | |
de9b4cca | 783 | if (req_ref_put_and_test(req)) { |
effcf8bd SR |
784 | struct io_ring_ctx *ctx = req->ctx; |
785 | ||
da1a08c5 | 786 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 787 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
788 | io_disarm_next(req); |
789 | if (req->link) { | |
790 | io_req_task_queue(req->link); | |
791 | req->link = NULL; | |
792 | } | |
793 | } | |
7ac1edc4 | 794 | io_req_put_rsrc(req); |
8197b053 PB |
795 | /* |
796 | * Selected buffer deallocation in io_clean_op() assumes that | |
797 | * we don't hold ->completion_lock. Clean them here to avoid | |
798 | * deadlocks. | |
799 | */ | |
800 | io_put_kbuf_comp(req); | |
c7dae4ba JA |
801 | io_dismantle_req(req); |
802 | io_put_task(req->task, 1); | |
c2b6c6bc | 803 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 804 | ctx->locked_free_nr++; |
180f829f | 805 | } |
a37fae8a HX |
806 | } |
807 | ||
59915143 | 808 | void __io_req_complete_post(struct io_kiocb *req) |
effcf8bd | 809 | { |
97b388d7 | 810 | if (!(req->flags & REQ_F_CQE_SKIP)) |
91ef75a7 | 811 | __io_fill_cqe_req(req->ctx, req); |
effcf8bd SR |
812 | __io_req_complete_put(req); |
813 | } | |
814 | ||
59915143 | 815 | void io_req_complete_post(struct io_kiocb *req) |
a37fae8a HX |
816 | { |
817 | struct io_ring_ctx *ctx = req->ctx; | |
818 | ||
25399321 | 819 | io_cq_lock(ctx); |
97b388d7 | 820 | __io_req_complete_post(req); |
25399321 | 821 | io_cq_unlock_post(ctx); |
4e3d9ff9 JA |
822 | } |
823 | ||
99f15d8d | 824 | inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags) |
bcda7baa | 825 | { |
75d7b3ae | 826 | io_req_complete_post(req); |
0ddf92e8 JA |
827 | } |
828 | ||
329061d3 | 829 | void io_req_complete_failed(struct io_kiocb *req, s32 res) |
f41db273 | 830 | { |
93d2bcd2 | 831 | req_set_fail(req); |
97b388d7 JA |
832 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
833 | io_req_complete_post(req); | |
f41db273 PB |
834 | } |
835 | ||
864ea921 PB |
836 | /* |
837 | * Don't initialise the fields below on every allocation, but do that in | |
838 | * advance and keep them valid across allocations. | |
839 | */ | |
840 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
841 | { | |
842 | req->ctx = ctx; | |
843 | req->link = NULL; | |
844 | req->async_data = NULL; | |
845 | /* not necessary, but safer to zero */ | |
cef216fc | 846 | req->cqe.res = 0; |
864ea921 PB |
847 | } |
848 | ||
dac7a098 | 849 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 850 | struct io_submit_state *state) |
dac7a098 | 851 | { |
79ebeaee | 852 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 853 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 854 | ctx->locked_free_nr = 0; |
79ebeaee | 855 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
856 | } |
857 | ||
88ab95be | 858 | static inline bool io_req_cache_empty(struct io_ring_ctx *ctx) |
0ddf92e8 | 859 | { |
88ab95be | 860 | return !ctx->submit_state.free_list.next; |
0ddf92e8 JA |
861 | } |
862 | ||
5d5901a3 PB |
863 | /* |
864 | * A request might get retired back into the request caches even before opcode | |
865 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
866 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
867 | * and with extra caution to not get a request that is still worked on. | |
868 | */ | |
c072481d | 869 | static __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 870 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 871 | { |
864ea921 | 872 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 873 | void *reqs[IO_REQ_ALLOC_BATCH]; |
864ea921 | 874 | int ret, i; |
e5d1bc0a | 875 | |
23a5c43b PB |
876 | /* |
877 | * If we have more than a batch's worth of requests in our IRQ side | |
878 | * locked cache, grab the lock and move them over to our submission | |
879 | * side cache. | |
880 | */ | |
a6d97a8a | 881 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 882 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 883 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
884 | return true; |
885 | } | |
e5d1bc0a | 886 | |
3ab665b7 | 887 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 888 | |
864ea921 PB |
889 | /* |
890 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
891 | * retry single alloc to be on the safe side. | |
892 | */ | |
893 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
894 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
895 | if (!reqs[0]) | |
a33ae9ce | 896 | return false; |
864ea921 | 897 | ret = 1; |
2b188cc1 | 898 | } |
864ea921 | 899 | |
37f0e767 | 900 | percpu_ref_get_many(&ctx->refs, ret); |
3ab665b7 | 901 | for (i = 0; i < ret; i++) { |
23a5c43b | 902 | struct io_kiocb *req = reqs[i]; |
3ab665b7 PB |
903 | |
904 | io_preinit_req(req, ctx); | |
fa05457a | 905 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 906 | } |
a33ae9ce PB |
907 | return true; |
908 | } | |
909 | ||
910 | static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx) | |
911 | { | |
88ab95be | 912 | if (unlikely(io_req_cache_empty(ctx))) |
a33ae9ce PB |
913 | return __io_alloc_req_refill(ctx); |
914 | return true; | |
915 | } | |
916 | ||
917 | static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx) | |
918 | { | |
919 | struct io_wq_work_node *node; | |
920 | ||
921 | node = wq_stack_extract(&ctx->submit_state.free_list); | |
c2b6c6bc | 922 | return container_of(node, struct io_kiocb, comp_list); |
2b188cc1 JA |
923 | } |
924 | ||
6b639522 | 925 | static inline void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 926 | { |
094bae49 | 927 | unsigned int flags = req->flags; |
929a3af9 | 928 | |
867f8fa5 | 929 | if (unlikely(flags & IO_REQ_CLEAN_FLAGS)) |
3a0a6902 | 930 | io_clean_op(req); |
e1d767f0 PB |
931 | if (!(flags & REQ_F_FIXED_FILE)) |
932 | io_put_file(req->file); | |
e65ef56d JA |
933 | } |
934 | ||
59915143 | 935 | __cold void io_free_req(struct io_kiocb *req) |
c6ca97b3 | 936 | { |
51a4cc11 | 937 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 938 | |
7ac1edc4 | 939 | io_req_put_rsrc(req); |
216578e5 | 940 | io_dismantle_req(req); |
7c660731 | 941 | io_put_task(req->task, 1); |
c6ca97b3 | 942 | |
79ebeaee | 943 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 944 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
c34b025f | 945 | ctx->locked_free_nr++; |
79ebeaee | 946 | spin_unlock(&ctx->completion_lock); |
e65ef56d JA |
947 | } |
948 | ||
d81499bf PB |
949 | static void __io_req_find_next_prep(struct io_kiocb *req) |
950 | { | |
951 | struct io_ring_ctx *ctx = req->ctx; | |
d81499bf | 952 | |
25399321 | 953 | io_cq_lock(ctx); |
305bef98 | 954 | io_disarm_next(req); |
25399321 | 955 | io_cq_unlock_post(ctx); |
d81499bf PB |
956 | } |
957 | ||
958 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 959 | { |
33cc89a9 | 960 | struct io_kiocb *nxt; |
944e58bf | 961 | |
9e645e11 JA |
962 | /* |
963 | * If LINK is set, we have dependent requests in this chain. If we | |
964 | * didn't fail this request, queue the first one up, moving any other | |
965 | * dependencies to the next request. In case of failure, fail the rest | |
966 | * of the chain. | |
967 | */ | |
d81499bf PB |
968 | if (unlikely(req->flags & IO_DISARM_MASK)) |
969 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
970 | nxt = req->link; |
971 | req->link = NULL; | |
972 | return nxt; | |
4d7dd462 | 973 | } |
9e645e11 | 974 | |
f237c30a | 975 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked) |
2c32395d PB |
976 | { |
977 | if (!ctx) | |
978 | return; | |
ef060ea9 JA |
979 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
980 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
f237c30a | 981 | if (*locked) { |
c450178d | 982 | io_submit_flush_completions(ctx); |
2c32395d | 983 | mutex_unlock(&ctx->uring_lock); |
f237c30a | 984 | *locked = false; |
2c32395d PB |
985 | } |
986 | percpu_ref_put(&ctx->refs); | |
987 | } | |
988 | ||
f28c240e HX |
989 | static void handle_tw_list(struct io_wq_work_node *node, |
990 | struct io_ring_ctx **ctx, bool *locked) | |
9f8d032a HX |
991 | { |
992 | do { | |
993 | struct io_wq_work_node *next = node->next; | |
994 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
995 | io_task_work.node); | |
996 | ||
34d2bfe7 JA |
997 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); |
998 | ||
9f8d032a HX |
999 | if (req->ctx != *ctx) { |
1000 | ctx_flush_and_put(*ctx, locked); | |
1001 | *ctx = req->ctx; | |
1002 | /* if not contended, grab and improve batching */ | |
1003 | *locked = mutex_trylock(&(*ctx)->uring_lock); | |
1004 | percpu_ref_get(&(*ctx)->refs); | |
1005 | } | |
1006 | req->io_task_work.func(req, locked); | |
1007 | node = next; | |
1008 | } while (node); | |
1009 | } | |
1010 | ||
c9f06aa7 | 1011 | void tctx_task_work(struct callback_head *cb) |
c40f6379 | 1012 | { |
f28c240e | 1013 | bool uring_locked = false; |
ebd0df2e | 1014 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
1015 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
1016 | task_work); | |
c40f6379 | 1017 | |
16f72070 | 1018 | while (1) { |
ed5ccb3b | 1019 | struct io_wq_work_node *node; |
3f18407d PB |
1020 | |
1021 | spin_lock_irq(&tctx->task_lock); | |
ed5ccb3b | 1022 | node = tctx->task_list.first; |
3f18407d | 1023 | INIT_WQ_LIST(&tctx->task_list); |
ed5ccb3b | 1024 | if (!node) |
6294f368 | 1025 | tctx->task_running = false; |
3f18407d | 1026 | spin_unlock_irq(&tctx->task_lock); |
ed5ccb3b | 1027 | if (!node) |
6294f368 | 1028 | break; |
ed5ccb3b | 1029 | handle_tw_list(node, &ctx, &uring_locked); |
7cbf1722 | 1030 | cond_resched(); |
68ca8fc0 | 1031 | |
ed5ccb3b | 1032 | if (data_race(!tctx->task_list.first) && uring_locked) |
68ca8fc0 | 1033 | io_submit_flush_completions(ctx); |
3f18407d | 1034 | } |
ebd0df2e | 1035 | |
f28c240e | 1036 | ctx_flush_and_put(ctx, &uring_locked); |
3cc7fdb9 PB |
1037 | |
1038 | /* relaxed read is enough as only the task itself sets ->in_idle */ | |
1039 | if (unlikely(atomic_read(&tctx->in_idle))) | |
1040 | io_uring_drop_tctx_refs(current); | |
7cbf1722 JA |
1041 | } |
1042 | ||
3fe07bcd | 1043 | static void __io_req_task_work_add(struct io_kiocb *req, |
ed5ccb3b | 1044 | struct io_uring_task *tctx) |
7cbf1722 | 1045 | { |
9f010507 | 1046 | struct io_ring_ctx *ctx = req->ctx; |
e09ee510 | 1047 | struct io_wq_work_node *node; |
0b81e80c | 1048 | unsigned long flags; |
6294f368 | 1049 | bool running; |
7cbf1722 | 1050 | |
0b81e80c | 1051 | spin_lock_irqsave(&tctx->task_lock, flags); |
ed5ccb3b | 1052 | wq_list_add_tail(&req->io_task_work.node, &tctx->task_list); |
6294f368 PB |
1053 | running = tctx->task_running; |
1054 | if (!running) | |
1055 | tctx->task_running = true; | |
0b81e80c | 1056 | spin_unlock_irqrestore(&tctx->task_lock, flags); |
7cbf1722 JA |
1057 | |
1058 | /* task_work already pending, we're done */ | |
6294f368 | 1059 | if (running) |
e09ee510 | 1060 | return; |
7cbf1722 | 1061 | |
ef060ea9 JA |
1062 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1063 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1064 | ||
3fe07bcd | 1065 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1066 | return; |
2215bed9 | 1067 | |
0b81e80c | 1068 | spin_lock_irqsave(&tctx->task_lock, flags); |
6294f368 | 1069 | tctx->task_running = false; |
ed5ccb3b DY |
1070 | node = tctx->task_list.first; |
1071 | INIT_WQ_LIST(&tctx->task_list); | |
0b81e80c | 1072 | spin_unlock_irqrestore(&tctx->task_lock, flags); |
7cbf1722 | 1073 | |
e09ee510 PB |
1074 | while (node) { |
1075 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1076 | node = node->next; | |
1077 | if (llist_add(&req->io_task_work.fallback_node, | |
1078 | &req->ctx->fallback_llist)) | |
1079 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1080 | } | |
eab30c4d PB |
1081 | } |
1082 | ||
99f15d8d | 1083 | void io_req_task_work_add(struct io_kiocb *req) |
3fe07bcd JA |
1084 | { |
1085 | struct io_uring_task *tctx = req->task->io_uring; | |
1086 | ||
ed5ccb3b | 1087 | __io_req_task_work_add(req, tctx); |
3fe07bcd JA |
1088 | } |
1089 | ||
4e118cd9 | 1090 | static void io_req_tw_post(struct io_kiocb *req, bool *locked) |
c40f6379 | 1091 | { |
97b388d7 | 1092 | io_req_complete_post(req); |
4e118cd9 | 1093 | } |
c40f6379 | 1094 | |
59915143 | 1095 | void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags) |
4e118cd9 | 1096 | { |
97b388d7 | 1097 | io_req_set_res(req, res, cflags); |
4e118cd9 | 1098 | req->io_task_work.func = io_req_tw_post; |
3fe07bcd | 1099 | io_req_task_work_add(req); |
4e118cd9 PB |
1100 | } |
1101 | ||
f237c30a | 1102 | static void io_req_task_cancel(struct io_kiocb *req, bool *locked) |
c40f6379 | 1103 | { |
b18a1a45 | 1104 | /* not needed for normal modes, but SQPOLL depends on it */ |
971cf9c1 | 1105 | io_tw_lock(req->ctx, locked); |
cef216fc | 1106 | io_req_complete_failed(req, req->cqe.res); |
c40f6379 JA |
1107 | } |
1108 | ||
329061d3 | 1109 | void io_req_task_submit(struct io_kiocb *req, bool *locked) |
c40f6379 | 1110 | { |
971cf9c1 | 1111 | io_tw_lock(req->ctx, locked); |
316319e8 | 1112 | /* req->task == current here, checking PF_EXITING is safe */ |
af066f31 | 1113 | if (likely(!(req->task->flags & PF_EXITING))) |
cbc2e203 | 1114 | io_queue_sqe(req); |
81b6d05c | 1115 | else |
2593553a | 1116 | io_req_complete_failed(req, -EFAULT); |
c40f6379 JA |
1117 | } |
1118 | ||
59915143 | 1119 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1120 | { |
97b388d7 | 1121 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1122 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1123 | io_req_task_work_add(req); |
c40f6379 JA |
1124 | } |
1125 | ||
f3b44f92 | 1126 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1127 | { |
5b0a6acc | 1128 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1129 | io_req_task_work_add(req); |
a3df7698 PB |
1130 | } |
1131 | ||
59915143 | 1132 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1133 | { |
9b5f7bd9 | 1134 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1135 | |
1136 | if (nxt) | |
906a8c3f | 1137 | io_req_task_queue(nxt); |
c69f8dbe JL |
1138 | } |
1139 | ||
f3b44f92 | 1140 | void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node) |
3aa83bfb | 1141 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1142 | { |
d4b7a5ef | 1143 | struct task_struct *task = NULL; |
37f0e767 | 1144 | int task_refs = 0; |
5af1d13e | 1145 | |
3aa83bfb PB |
1146 | do { |
1147 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1148 | comp_list); | |
2d6500d4 | 1149 | |
a538be5b PB |
1150 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1151 | if (req->flags & REQ_F_REFCOUNT) { | |
1152 | node = req->comp_list.next; | |
1153 | if (!req_ref_put_and_test(req)) | |
1154 | continue; | |
1155 | } | |
b605a7fa PB |
1156 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1157 | struct async_poll *apoll = req->apoll; | |
1158 | ||
1159 | if (apoll->double_poll) | |
1160 | kfree(apoll->double_poll); | |
1161 | list_add(&apoll->poll.wait.entry, | |
1162 | &ctx->apoll_cache); | |
1163 | req->flags &= ~REQ_F_POLLED; | |
1164 | } | |
da1a08c5 | 1165 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1166 | io_queue_next(req); |
a538be5b PB |
1167 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1168 | io_clean_op(req); | |
c1e53a69 | 1169 | } |
a538be5b PB |
1170 | if (!(req->flags & REQ_F_FIXED_FILE)) |
1171 | io_put_file(req->file); | |
2d6500d4 | 1172 | |
ab409402 | 1173 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1174 | |
d4b7a5ef PB |
1175 | if (req->task != task) { |
1176 | if (task) | |
1177 | io_put_task(task, task_refs); | |
1178 | task = req->task; | |
1179 | task_refs = 0; | |
1180 | } | |
1181 | task_refs++; | |
c1e53a69 | 1182 | node = req->comp_list.next; |
fa05457a | 1183 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1184 | } while (node); |
d4b7a5ef | 1185 | |
d4b7a5ef PB |
1186 | if (task) |
1187 | io_put_task(task, task_refs); | |
7a743e22 PB |
1188 | } |
1189 | ||
c450178d | 1190 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1191 | __must_hold(&ctx->uring_lock) |
905c172f | 1192 | { |
6f33b0bc | 1193 | struct io_wq_work_node *node, *prev; |
cd0ca2e0 | 1194 | struct io_submit_state *state = &ctx->submit_state; |
905c172f | 1195 | |
d9dee430 PB |
1196 | spin_lock(&ctx->completion_lock); |
1197 | wq_list_for_each(node, prev, &state->compl_reqs) { | |
1198 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1199 | comp_list); | |
3d4aeb9f | 1200 | |
d9dee430 PB |
1201 | if (!(req->flags & REQ_F_CQE_SKIP)) |
1202 | __io_fill_cqe_req(ctx, req); | |
905c172f | 1203 | } |
25399321 | 1204 | __io_cq_unlock_post(ctx); |
d9dee430 | 1205 | |
1cce17ac | 1206 | io_free_batch_list(ctx, state->compl_reqs.first); |
6f33b0bc | 1207 | INIT_WQ_LIST(&state->compl_reqs); |
7a743e22 PB |
1208 | } |
1209 | ||
ba816ad6 JA |
1210 | /* |
1211 | * Drop reference to request, return next in chain (if there is one) if this | |
1212 | * was the last reference to this request. | |
1213 | */ | |
0d85035a | 1214 | static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 1215 | { |
9b5f7bd9 PB |
1216 | struct io_kiocb *nxt = NULL; |
1217 | ||
de9b4cca | 1218 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 1219 | if (unlikely(req->flags & IO_REQ_LINK_FLAGS)) |
7819a1f6 | 1220 | nxt = io_req_find_next(req); |
f5c6cf2a | 1221 | io_free_req(req); |
2a44f467 | 1222 | } |
9b5f7bd9 | 1223 | return nxt; |
2b188cc1 JA |
1224 | } |
1225 | ||
6c503150 | 1226 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1227 | { |
1228 | /* See comment at the top of this file */ | |
1229 | smp_rmb(); | |
e23de15f | 1230 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1231 | } |
1232 | ||
def596e9 JA |
1233 | /* |
1234 | * We can't just wait for polled events to come to us, we have to actively | |
1235 | * find and complete them. | |
1236 | */ | |
c072481d | 1237 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1238 | { |
1239 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1240 | return; | |
1241 | ||
1242 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1243 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1244 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1245 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1246 | break; |
08f5439f JA |
1247 | /* |
1248 | * Ensure we allow local-to-the-cpu processing to take place, | |
1249 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1250 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1251 | */ |
3fcee5a6 PB |
1252 | if (need_resched()) { |
1253 | mutex_unlock(&ctx->uring_lock); | |
1254 | cond_resched(); | |
1255 | mutex_lock(&ctx->uring_lock); | |
1256 | } | |
def596e9 JA |
1257 | } |
1258 | mutex_unlock(&ctx->uring_lock); | |
1259 | } | |
1260 | ||
7668b92a | 1261 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1262 | { |
7668b92a | 1263 | unsigned int nr_events = 0; |
e9979b36 | 1264 | int ret = 0; |
155bc950 | 1265 | unsigned long check_cq; |
500f9fba | 1266 | |
3a08576b PB |
1267 | check_cq = READ_ONCE(ctx->check_cq); |
1268 | if (unlikely(check_cq)) { | |
1269 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
1270 | __io_cqring_overflow_flush(ctx, false); | |
1271 | /* | |
1272 | * Similarly do not spin if we have not informed the user of any | |
1273 | * dropped CQE. | |
1274 | */ | |
1275 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1276 | return -EBADR; | |
1277 | } | |
f39c8a5b PB |
1278 | /* |
1279 | * Don't enter poll loop if we already have events pending. | |
1280 | * If we do, we can potentially be spinning for commands that | |
1281 | * already triggered a CQE (eg in error). | |
1282 | */ | |
f39c8a5b | 1283 | if (io_cqring_events(ctx)) |
d487b43c | 1284 | return 0; |
155bc950 | 1285 | |
def596e9 | 1286 | do { |
500f9fba JA |
1287 | /* |
1288 | * If a submit got punted to a workqueue, we can have the | |
1289 | * application entering polling for a command before it gets | |
1290 | * issued. That app will hold the uring_lock for the duration | |
1291 | * of the poll right here, so we need to take a breather every | |
1292 | * now and then to ensure that the issue has a chance to add | |
1293 | * the poll to the issued list. Otherwise we can spin here | |
1294 | * forever, while the workqueue is stuck trying to acquire the | |
1295 | * very same mutex. | |
1296 | */ | |
5eef4e87 | 1297 | if (wq_list_empty(&ctx->iopoll_list)) { |
8f487ef2 PB |
1298 | u32 tail = ctx->cached_cq_tail; |
1299 | ||
500f9fba | 1300 | mutex_unlock(&ctx->uring_lock); |
4c6e277c | 1301 | io_run_task_work(); |
500f9fba | 1302 | mutex_lock(&ctx->uring_lock); |
def596e9 | 1303 | |
8f487ef2 PB |
1304 | /* some requests don't go through iopoll_list */ |
1305 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1306 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1307 | break; |
500f9fba | 1308 | } |
5ba3c874 PB |
1309 | ret = io_do_iopoll(ctx, !min); |
1310 | if (ret < 0) | |
1311 | break; | |
1312 | nr_events += ret; | |
1313 | ret = 0; | |
1314 | } while (nr_events < min && !need_resched()); | |
d487b43c | 1315 | |
def596e9 JA |
1316 | return ret; |
1317 | } | |
7012c815 PB |
1318 | |
1319 | void io_req_task_complete(struct io_kiocb *req, bool *locked) | |
8ef12efe | 1320 | { |
7012c815 PB |
1321 | if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) { |
1322 | unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED; | |
1323 | ||
1324 | req->cqe.flags |= io_put_kbuf(req, issue_flags); | |
1325 | } | |
1326 | ||
1327 | if (*locked) | |
9da070b1 | 1328 | io_req_complete_defer(req); |
7012c815 | 1329 | else |
97b388d7 | 1330 | io_req_complete_post(req); |
8ef12efe JA |
1331 | } |
1332 | ||
def596e9 JA |
1333 | /* |
1334 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1335 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1336 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1337 | * accessing the kiocb cookie. |
1338 | */ | |
9882131c | 1339 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1340 | { |
1341 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1342 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1343 | |
1344 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1345 | if (unlikely(needs_lock)) |
cb3d8972 | 1346 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1347 | |
1348 | /* | |
1349 | * Track whether we have multiple files in our lists. This will impact | |
1350 | * how we do polling eventually, not spinning if we're on potentially | |
1351 | * different devices. | |
1352 | */ | |
5eef4e87 | 1353 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1354 | ctx->poll_multi_queue = false; |
1355 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1356 | struct io_kiocb *list_req; |
1357 | ||
5eef4e87 PB |
1358 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1359 | comp_list); | |
30da1b45 | 1360 | if (list_req->file != req->file) |
915b3dde | 1361 | ctx->poll_multi_queue = true; |
def596e9 JA |
1362 | } |
1363 | ||
1364 | /* | |
1365 | * For fast devices, IO may have already completed. If it has, add | |
1366 | * it to the front so we find it first. | |
1367 | */ | |
65a6543d | 1368 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1369 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1370 | else |
5eef4e87 | 1371 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1372 | |
3b44b371 | 1373 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1374 | /* |
1375 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1376 | * in sq thread task context or in io worker task context. If | |
1377 | * current task context is sq thread, we don't need to check | |
1378 | * whether should wake up sq thread. | |
1379 | */ | |
1380 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1381 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1382 | wake_up(&ctx->sq_data->wait); | |
1383 | ||
1384 | mutex_unlock(&ctx->uring_lock); | |
1385 | } | |
def596e9 JA |
1386 | } |
1387 | ||
4503b767 JA |
1388 | static bool io_bdev_nowait(struct block_device *bdev) |
1389 | { | |
9ba0d0c8 | 1390 | return !bdev || blk_queue_nowait(bdev_get_queue(bdev)); |
4503b767 JA |
1391 | } |
1392 | ||
2b188cc1 JA |
1393 | /* |
1394 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1395 | * any file. For now, just ensure that anything potentially problematic is done | |
1396 | * inline. | |
1397 | */ | |
88459b50 | 1398 | static bool __io_file_supports_nowait(struct file *file, umode_t mode) |
2b188cc1 | 1399 | { |
4503b767 | 1400 | if (S_ISBLK(mode)) { |
4e7b5671 CH |
1401 | if (IS_ENABLED(CONFIG_BLOCK) && |
1402 | io_bdev_nowait(I_BDEV(file->f_mapping->host))) | |
4503b767 JA |
1403 | return true; |
1404 | return false; | |
1405 | } | |
976517f1 | 1406 | if (S_ISSOCK(mode)) |
2b188cc1 | 1407 | return true; |
4503b767 | 1408 | if (S_ISREG(mode)) { |
4e7b5671 CH |
1409 | if (IS_ENABLED(CONFIG_BLOCK) && |
1410 | io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
e5550a14 | 1411 | !io_is_uring_fops(file)) |
4503b767 JA |
1412 | return true; |
1413 | return false; | |
1414 | } | |
2b188cc1 | 1415 | |
c5b85625 JA |
1416 | /* any ->read/write should understand O_NONBLOCK */ |
1417 | if (file->f_flags & O_NONBLOCK) | |
1418 | return true; | |
35645ac3 | 1419 | return file->f_mode & FMODE_NOWAIT; |
2b188cc1 | 1420 | } |
c5b85625 | 1421 | |
88459b50 PB |
1422 | /* |
1423 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1424 | * any file. For now, just ensure that anything potentially problematic is done | |
1425 | * inline. | |
1426 | */ | |
a4ad4f74 | 1427 | unsigned int io_file_get_flags(struct file *file) |
88459b50 PB |
1428 | { |
1429 | umode_t mode = file_inode(file)->i_mode; | |
1430 | unsigned int res = 0; | |
af197f50 | 1431 | |
88459b50 PB |
1432 | if (S_ISREG(mode)) |
1433 | res |= FFS_ISREG; | |
1434 | if (__io_file_supports_nowait(file, mode)) | |
1435 | res |= FFS_NOWAIT; | |
5e45690a JA |
1436 | if (io_file_need_scm(file)) |
1437 | res |= FFS_SCM; | |
88459b50 | 1438 | return res; |
2b188cc1 JA |
1439 | } |
1440 | ||
99f15d8d | 1441 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1442 | { |
e8c2bc1f JA |
1443 | WARN_ON_ONCE(!io_op_defs[req->opcode].async_size); |
1444 | req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL); | |
d886e185 PB |
1445 | if (req->async_data) { |
1446 | req->flags |= REQ_F_ASYNC_DATA; | |
1447 | return false; | |
1448 | } | |
1449 | return true; | |
3d9932a8 XW |
1450 | } |
1451 | ||
f3b44f92 | 1452 | int io_req_prep_async(struct io_kiocb *req) |
f67676d1 | 1453 | { |
0702e536 JA |
1454 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
1455 | ||
1456 | /* assign early for deferred execution for non-fixed file */ | |
1457 | if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE)) | |
1458 | req->file = io_file_get_normal(req, req->cqe.fd); | |
dc919caf | 1459 | if (!def->prep_async) |
0702e536 JA |
1460 | return 0; |
1461 | if (WARN_ON_ONCE(req_has_async_data(req))) | |
1462 | return -EFAULT; | |
1463 | if (io_alloc_async_data(req)) | |
1464 | return -EAGAIN; | |
1465 | ||
dc919caf | 1466 | return def->prep_async(req); |
bfe76559 PB |
1467 | } |
1468 | ||
9cf7c104 PB |
1469 | static u32 io_get_sequence(struct io_kiocb *req) |
1470 | { | |
a3dbdf54 | 1471 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1472 | struct io_kiocb *cur; |
9cf7c104 | 1473 | |
a3dbdf54 | 1474 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1475 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1476 | seq--; |
1477 | return seq; | |
9cf7c104 PB |
1478 | } |
1479 | ||
c072481d | 1480 | static __cold void io_drain_req(struct io_kiocb *req) |
de0617e4 | 1481 | { |
a197f664 | 1482 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1483 | struct io_defer_entry *de; |
f67676d1 | 1484 | int ret; |
e0eb71dc | 1485 | u32 seq = io_get_sequence(req); |
3c19966d | 1486 | |
9d858b21 | 1487 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1488 | spin_lock(&ctx->completion_lock); |
5e371265 | 1489 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1490 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1491 | queue: |
10c66904 | 1492 | ctx->drain_active = false; |
e0eb71dc PB |
1493 | io_req_task_queue(req); |
1494 | return; | |
10c66904 | 1495 | } |
e302f104 | 1496 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1497 | |
b7e298d2 | 1498 | ret = io_req_prep_async(req); |
e0eb71dc PB |
1499 | if (ret) { |
1500 | fail: | |
1501 | io_req_complete_failed(req, ret); | |
1502 | return; | |
1503 | } | |
cbdcb435 | 1504 | io_prep_async_link(req); |
27dc8338 | 1505 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1506 | if (!de) { |
1b48773f | 1507 | ret = -ENOMEM; |
e0eb71dc | 1508 | goto fail; |
76cc33d7 | 1509 | } |
2d28390a | 1510 | |
79ebeaee | 1511 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1512 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1513 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1514 | kfree(de); |
e0eb71dc | 1515 | goto queue; |
de0617e4 JA |
1516 | } |
1517 | ||
48863ffd | 1518 | trace_io_uring_defer(req); |
27dc8338 | 1519 | de->req = req; |
9cf7c104 | 1520 | de->seq = seq; |
27dc8338 | 1521 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1522 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1523 | } |
1524 | ||
68fb8979 | 1525 | static void io_clean_op(struct io_kiocb *req) |
99bc4c38 | 1526 | { |
8197b053 PB |
1527 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
1528 | spin_lock(&req->ctx->completion_lock); | |
cc3cec83 | 1529 | io_put_kbuf_comp(req); |
8197b053 PB |
1530 | spin_unlock(&req->ctx->completion_lock); |
1531 | } | |
99bc4c38 | 1532 | |
0e1b6fe3 | 1533 | if (req->flags & REQ_F_NEED_CLEANUP) { |
4d4c9cff | 1534 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
bb040a21 | 1535 | |
4d4c9cff JA |
1536 | if (def->cleanup) |
1537 | def->cleanup(req); | |
99bc4c38 | 1538 | } |
75652a30 JA |
1539 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1540 | kfree(req->apoll->double_poll); | |
1541 | kfree(req->apoll); | |
1542 | req->apoll = NULL; | |
1543 | } | |
9cae36a0 JA |
1544 | if (req->flags & REQ_F_INFLIGHT) { |
1545 | struct io_uring_task *tctx = req->task->io_uring; | |
1546 | ||
1547 | atomic_dec(&tctx->inflight_tracked); | |
1548 | } | |
c854357b | 1549 | if (req->flags & REQ_F_CREDS) |
b8e64b53 | 1550 | put_cred(req->creds); |
d886e185 PB |
1551 | if (req->flags & REQ_F_ASYNC_DATA) { |
1552 | kfree(req->async_data); | |
1553 | req->async_data = NULL; | |
1554 | } | |
c854357b | 1555 | req->flags &= ~IO_REQ_CLEAN_FLAGS; |
99bc4c38 PB |
1556 | } |
1557 | ||
6bf9c47a JA |
1558 | static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags) |
1559 | { | |
1560 | if (req->file || !io_op_defs[req->opcode].needs_file) | |
1561 | return true; | |
1562 | ||
1563 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1564 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1565 | else |
cef216fc | 1566 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1567 | |
772f5e00 | 1568 | return !!req->file; |
6bf9c47a JA |
1569 | } |
1570 | ||
889fca73 | 1571 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1572 | { |
fcde59fe | 1573 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
5730b27e | 1574 | const struct cred *creds = NULL; |
d625c6ee | 1575 | int ret; |
2b188cc1 | 1576 | |
70152140 JA |
1577 | if (unlikely(!io_assign_file(req, issue_flags))) |
1578 | return -EBADF; | |
1579 | ||
6878b40e | 1580 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1581 | creds = override_creds(req->creds); |
5730b27e | 1582 | |
fcde59fe | 1583 | if (!def->audit_skip) |
5bd2182d PM |
1584 | audit_uring_entry(req->opcode); |
1585 | ||
0702e536 | 1586 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1587 | |
fcde59fe | 1588 | if (!def->audit_skip) |
5bd2182d PM |
1589 | audit_uring_exit(!ret, ret); |
1590 | ||
5730b27e JA |
1591 | if (creds) |
1592 | revert_creds(creds); | |
97b388d7 | 1593 | |
75d7b3ae PB |
1594 | if (ret == IOU_OK) { |
1595 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
9da070b1 | 1596 | io_req_complete_defer(req); |
75d7b3ae PB |
1597 | else |
1598 | io_req_complete_post(req); | |
1599 | } else if (ret != IOU_ISSUE_SKIP_COMPLETE) | |
def596e9 | 1600 | return ret; |
97b388d7 | 1601 | |
b532576e | 1602 | /* If the op doesn't have a file, we're not polling for it */ |
9983028e | 1603 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file) |
9882131c | 1604 | io_iopoll_req_issued(req, issue_flags); |
def596e9 JA |
1605 | |
1606 | return 0; | |
2b188cc1 JA |
1607 | } |
1608 | ||
329061d3 JA |
1609 | int io_poll_issue(struct io_kiocb *req, bool *locked) |
1610 | { | |
1611 | io_tw_lock(req->ctx, locked); | |
1612 | if (unlikely(req->task->flags & PF_EXITING)) | |
1613 | return -EFAULT; | |
aeaa72c6 | 1614 | return io_issue_sqe(req, IO_URING_F_NONBLOCK); |
329061d3 JA |
1615 | } |
1616 | ||
c9f06aa7 | 1617 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1618 | { |
1619 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1620 | ||
1621 | req = io_put_req_find_next(req); | |
1622 | return req ? &req->work : NULL; | |
1623 | } | |
1624 | ||
c9f06aa7 | 1625 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1626 | { |
1627 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6bf9c47a | 1628 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
d01905db PB |
1629 | unsigned int issue_flags = IO_URING_F_UNLOCKED; |
1630 | bool needs_poll = false; | |
6bf9c47a | 1631 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1632 | |
48dcd38d PB |
1633 | /* one will be dropped by ->io_free_work() after returning to io-wq */ |
1634 | if (!(req->flags & REQ_F_REFCOUNT)) | |
1635 | __io_req_set_refcount(req, 2); | |
1636 | else | |
1637 | req_ref_get(req); | |
5d5901a3 | 1638 | |
cb2d344c | 1639 | io_arm_ltimeout(req); |
6bf9c47a | 1640 | |
dadebc35 | 1641 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1642 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1643 | fail: |
6bf9c47a | 1644 | io_req_task_queue_fail(req, err); |
d01905db PB |
1645 | return; |
1646 | } | |
0f8da75b PB |
1647 | if (!io_assign_file(req, issue_flags)) { |
1648 | err = -EBADF; | |
1649 | work->flags |= IO_WQ_WORK_CANCEL; | |
1650 | goto fail; | |
1651 | } | |
31b51510 | 1652 | |
d01905db | 1653 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
1654 | bool opcode_poll = def->pollin || def->pollout; |
1655 | ||
1656 | if (opcode_poll && file_can_poll(req->file)) { | |
1657 | needs_poll = true; | |
d01905db | 1658 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 1659 | } |
561fb04a | 1660 | } |
31b51510 | 1661 | |
d01905db PB |
1662 | do { |
1663 | ret = io_issue_sqe(req, issue_flags); | |
1664 | if (ret != -EAGAIN) | |
1665 | break; | |
1666 | /* | |
1667 | * We can get EAGAIN for iopolled IO even though we're | |
1668 | * forcing a sync submission from here, since we can't | |
1669 | * wait for request slots on the block side. | |
1670 | */ | |
1671 | if (!needs_poll) { | |
e0deb6a0 PB |
1672 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
1673 | break; | |
d01905db PB |
1674 | cond_resched(); |
1675 | continue; | |
90fa0288 HX |
1676 | } |
1677 | ||
4d9237e3 | 1678 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
1679 | return; |
1680 | /* aborted or ready, in either case retry blocking */ | |
1681 | needs_poll = false; | |
1682 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
1683 | } while (1); | |
31b51510 | 1684 | |
a3df7698 | 1685 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 1686 | if (ret < 0) |
a3df7698 | 1687 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
1688 | } |
1689 | ||
531113bb JA |
1690 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
1691 | unsigned int issue_flags) | |
09bb8394 | 1692 | { |
5106dd6e JA |
1693 | struct io_ring_ctx *ctx = req->ctx; |
1694 | struct file *file = NULL; | |
ac177053 | 1695 | unsigned long file_ptr; |
09bb8394 | 1696 | |
93f052cb | 1697 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 1698 | |
ac177053 | 1699 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 1700 | goto out; |
ac177053 PB |
1701 | fd = array_index_nospec(fd, ctx->nr_user_files); |
1702 | file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; | |
1703 | file = (struct file *) (file_ptr & FFS_MASK); | |
1704 | file_ptr &= ~FFS_MASK; | |
1705 | /* mask in overlapping REQ_F and FFS bits */ | |
35645ac3 | 1706 | req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT); |
5106dd6e | 1707 | io_req_set_rsrc_node(req, ctx, 0); |
d78bd8ad | 1708 | WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap)); |
5106dd6e | 1709 | out: |
93f052cb | 1710 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
1711 | return file; |
1712 | } | |
d44f554e | 1713 | |
531113bb | 1714 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 1715 | { |
62906e89 | 1716 | struct file *file = fget(fd); |
ac177053 | 1717 | |
48863ffd | 1718 | trace_io_uring_file_get(req, fd); |
09bb8394 | 1719 | |
ac177053 | 1720 | /* we don't allow fixed io_uring files */ |
e5550a14 | 1721 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 1722 | io_req_track_inflight(req); |
8371adf5 | 1723 | return file; |
09bb8394 JA |
1724 | } |
1725 | ||
7bfa9bad | 1726 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
1727 | __must_hold(&req->ctx->uring_lock) |
1728 | { | |
7bfa9bad PB |
1729 | struct io_kiocb *linked_timeout; |
1730 | ||
1731 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
1732 | io_req_complete_failed(req, ret); | |
1733 | return; | |
1734 | } | |
1735 | ||
1736 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 1737 | |
4d9237e3 | 1738 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 1739 | case IO_APOLL_READY: |
d475a9a6 PB |
1740 | io_req_task_queue(req); |
1741 | break; | |
1742 | case IO_APOLL_ABORTED: | |
1743 | /* | |
1744 | * Queued up for async execution, worker will release | |
1745 | * submit reference when the iocb is actually submitted. | |
1746 | */ | |
6436c770 | 1747 | io_kbuf_recycle(req, 0); |
77955efb | 1748 | io_queue_iowq(req, NULL); |
d475a9a6 | 1749 | break; |
b1c62645 | 1750 | case IO_APOLL_OK: |
b1c62645 | 1751 | break; |
d475a9a6 PB |
1752 | } |
1753 | ||
1754 | if (linked_timeout) | |
1755 | io_queue_linked_timeout(linked_timeout); | |
1756 | } | |
1757 | ||
cbc2e203 | 1758 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 1759 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 1760 | { |
e0c5c576 | 1761 | int ret; |
2b188cc1 | 1762 | |
c5eef2b9 | 1763 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 1764 | |
491381ce JA |
1765 | /* |
1766 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
1767 | * doesn't support non-blocking read/write attempts | |
1768 | */ | |
7bfa9bad | 1769 | if (likely(!ret)) |
cb2d344c | 1770 | io_arm_ltimeout(req); |
7bfa9bad PB |
1771 | else |
1772 | io_queue_async(req, ret); | |
2b188cc1 JA |
1773 | } |
1774 | ||
4652fe3f | 1775 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 1776 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 1777 | { |
17b147f6 PB |
1778 | if (unlikely(req->flags & REQ_F_FAIL)) { |
1779 | /* | |
1780 | * We don't submit, fail them all, for that replace hardlinks | |
1781 | * with normal links. Extra REQ_F_LINK is tolerated. | |
1782 | */ | |
1783 | req->flags &= ~REQ_F_HARDLINK; | |
1784 | req->flags |= REQ_F_LINK; | |
1785 | io_req_complete_failed(req, req->cqe.res); | |
e0eb71dc PB |
1786 | } else if (unlikely(req->ctx->drain_active)) { |
1787 | io_drain_req(req); | |
76cc33d7 PB |
1788 | } else { |
1789 | int ret = io_req_prep_async(req); | |
1790 | ||
1791 | if (unlikely(ret)) | |
1792 | io_req_complete_failed(req, ret); | |
1793 | else | |
77955efb | 1794 | io_queue_iowq(req, NULL); |
ce35a47a | 1795 | } |
4fe2c963 JL |
1796 | } |
1797 | ||
b16fed66 PB |
1798 | /* |
1799 | * Check SQE restrictions (opcode and flags). | |
1800 | * | |
1801 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
1802 | */ | |
1803 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
1804 | struct io_kiocb *req, | |
1805 | unsigned int sqe_flags) | |
4fe2c963 | 1806 | { |
b16fed66 PB |
1807 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
1808 | return false; | |
1809 | ||
1810 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
1811 | ctx->restrictions.sqe_flags_required) | |
1812 | return false; | |
1813 | ||
1814 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
1815 | ctx->restrictions.sqe_flags_required)) | |
1816 | return false; | |
1817 | ||
1818 | return true; | |
4fe2c963 JL |
1819 | } |
1820 | ||
22b2ca31 PB |
1821 | static void io_init_req_drain(struct io_kiocb *req) |
1822 | { | |
1823 | struct io_ring_ctx *ctx = req->ctx; | |
1824 | struct io_kiocb *head = ctx->submit_state.link.head; | |
1825 | ||
1826 | ctx->drain_active = true; | |
1827 | if (head) { | |
1828 | /* | |
1829 | * If we need to drain a request in the middle of a link, drain | |
1830 | * the head request and the next request/link after the current | |
1831 | * link. Considering sequential execution of links, | |
b6c7db32 | 1832 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
1833 | * link. |
1834 | */ | |
b6c7db32 | 1835 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
1836 | ctx->drain_next = true; |
1837 | } | |
1838 | } | |
1839 | ||
b16fed66 PB |
1840 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1841 | const struct io_uring_sqe *sqe) | |
282cdc86 | 1842 | __must_hold(&ctx->uring_lock) |
b16fed66 | 1843 | { |
fcde59fe | 1844 | const struct io_op_def *def; |
b16fed66 | 1845 | unsigned int sqe_flags; |
fc0ae024 | 1846 | int personality; |
4a04d1d1 | 1847 | u8 opcode; |
b16fed66 | 1848 | |
864ea921 | 1849 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 1850 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 PB |
1851 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
1852 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
cef216fc | 1853 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 1854 | req->file = NULL; |
c1bdf8ed | 1855 | req->rsrc_node = NULL; |
b16fed66 | 1856 | req->task = current; |
b16fed66 | 1857 | |
4a04d1d1 PB |
1858 | if (unlikely(opcode >= IORING_OP_LAST)) { |
1859 | req->opcode = 0; | |
b16fed66 | 1860 | return -EINVAL; |
4a04d1d1 | 1861 | } |
fcde59fe | 1862 | def = &io_op_defs[opcode]; |
68fe256a PB |
1863 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
1864 | /* enforce forwards compatibility on users */ | |
1865 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
1866 | return -EINVAL; | |
4e906702 | 1867 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 1868 | if (!def->buffer_select) |
4e906702 JA |
1869 | return -EOPNOTSUPP; |
1870 | req->buf_index = READ_ONCE(sqe->buf_group); | |
1871 | } | |
5562a8d7 PB |
1872 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
1873 | ctx->drain_disabled = true; | |
1874 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
1875 | if (ctx->drain_disabled) | |
1876 | return -EOPNOTSUPP; | |
22b2ca31 | 1877 | io_init_req_drain(req); |
5562a8d7 | 1878 | } |
2a56a9bd PB |
1879 | } |
1880 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
1881 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
1882 | return -EACCES; | |
1883 | /* knock it to the slow queue path, will be drained there */ | |
1884 | if (ctx->drain_active) | |
1885 | req->flags |= REQ_F_FORCE_ASYNC; | |
1886 | /* if there is no link, we're at "next" request and need to drain */ | |
1887 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
1888 | ctx->drain_next = false; | |
1889 | ctx->drain_active = true; | |
b6c7db32 | 1890 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 1891 | } |
68fe256a | 1892 | } |
b16fed66 | 1893 | |
fcde59fe | 1894 | if (!def->ioprio && sqe->ioprio) |
73911426 | 1895 | return -EINVAL; |
fcde59fe | 1896 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
73911426 JA |
1897 | return -EINVAL; |
1898 | ||
fcde59fe | 1899 | if (def->needs_file) { |
6d63416d PB |
1900 | struct io_submit_state *state = &ctx->submit_state; |
1901 | ||
cef216fc | 1902 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 1903 | |
6d63416d PB |
1904 | /* |
1905 | * Plug now if we have more than 2 IO left after this, and the | |
1906 | * target is potentially a read/write to block based storage. | |
1907 | */ | |
fcde59fe | 1908 | if (state->need_plug && def->plug) { |
6d63416d PB |
1909 | state->plug_started = true; |
1910 | state->need_plug = false; | |
5ca7a8b3 | 1911 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 1912 | } |
b16fed66 | 1913 | } |
863e0560 | 1914 | |
003e8dcc JA |
1915 | personality = READ_ONCE(sqe->personality); |
1916 | if (personality) { | |
cdab10bf LT |
1917 | int ret; |
1918 | ||
c10d1f98 PB |
1919 | req->creds = xa_load(&ctx->personalities, personality); |
1920 | if (!req->creds) | |
003e8dcc | 1921 | return -EINVAL; |
c10d1f98 | 1922 | get_cred(req->creds); |
cdc1404a PM |
1923 | ret = security_uring_override_creds(req->creds); |
1924 | if (ret) { | |
1925 | put_cred(req->creds); | |
1926 | return ret; | |
1927 | } | |
b8e64b53 | 1928 | req->flags |= REQ_F_CREDS; |
003e8dcc | 1929 | } |
b16fed66 | 1930 | |
0702e536 | 1931 | return def->prep(req, sqe); |
b16fed66 PB |
1932 | } |
1933 | ||
df3becde PB |
1934 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
1935 | struct io_kiocb *req, int ret) | |
1936 | { | |
1937 | struct io_ring_ctx *ctx = req->ctx; | |
1938 | struct io_submit_link *link = &ctx->submit_state.link; | |
1939 | struct io_kiocb *head = link->head; | |
1940 | ||
48863ffd | 1941 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
1942 | |
1943 | /* | |
1944 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
1945 | * unusable. Instead of failing eagerly, continue assembling the link if | |
1946 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
1947 | * should find the flag and handle the rest. | |
1948 | */ | |
1949 | req_fail_link_node(req, ret); | |
1950 | if (head && !(head->flags & REQ_F_FAIL)) | |
1951 | req_fail_link_node(head, -ECANCELED); | |
1952 | ||
1953 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
1954 | if (head) { | |
1955 | link->last->link = req; | |
1956 | link->head = NULL; | |
1957 | req = head; | |
1958 | } | |
1959 | io_queue_sqe_fallback(req); | |
1960 | return ret; | |
1961 | } | |
1962 | ||
1963 | if (head) | |
1964 | link->last->link = req; | |
1965 | else | |
1966 | link->head = req; | |
1967 | link->last = req; | |
1968 | return 0; | |
1969 | } | |
1970 | ||
1971 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 1972 | const struct io_uring_sqe *sqe) |
282cdc86 | 1973 | __must_hold(&ctx->uring_lock) |
9e645e11 | 1974 | { |
a1ab7b35 | 1975 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 1976 | int ret; |
9e645e11 | 1977 | |
a6b8cadc | 1978 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
1979 | if (unlikely(ret)) |
1980 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 1981 | |
be7053b7 | 1982 | /* don't need @sqe from now on */ |
48863ffd | 1983 | trace_io_uring_submit_sqe(req, true); |
a6b8cadc | 1984 | |
9e645e11 JA |
1985 | /* |
1986 | * If we already have a head request, queue this one for async | |
1987 | * submittal once the head completes. If we don't have a head but | |
1988 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
1989 | * submitted sync once the chain is complete. If none of those | |
1990 | * conditions are true (normal request), then just queue it. | |
1991 | */ | |
924a07e4 | 1992 | if (unlikely(link->head)) { |
df3becde PB |
1993 | ret = io_req_prep_async(req); |
1994 | if (unlikely(ret)) | |
1995 | return io_submit_fail_init(sqe, req, ret); | |
1996 | ||
48863ffd | 1997 | trace_io_uring_link(req, link->head); |
f2f87370 | 1998 | link->last->link = req; |
863e0560 | 1999 | link->last = req; |
32fe525b | 2000 | |
da1a08c5 | 2001 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 2002 | return 0; |
df3becde PB |
2003 | /* last request of the link, flush it */ |
2004 | req = link->head; | |
f15a3431 | 2005 | link->head = NULL; |
924a07e4 PB |
2006 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
2007 | goto fallback; | |
2008 | ||
2009 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2010 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2011 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2012 | link->head = req; | |
2013 | link->last = req; | |
2014 | } else { | |
2015 | fallback: | |
2016 | io_queue_sqe_fallback(req); | |
2017 | } | |
f15a3431 | 2018 | return 0; |
9e645e11 | 2019 | } |
2e6e1fde | 2020 | |
f15a3431 | 2021 | io_queue_sqe(req); |
1d4240cc | 2022 | return 0; |
9e645e11 JA |
2023 | } |
2024 | ||
9a56a232 JA |
2025 | /* |
2026 | * Batched submission is done, ensure local IO is flushed out. | |
2027 | */ | |
553deffd | 2028 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2029 | { |
553deffd PB |
2030 | struct io_submit_state *state = &ctx->submit_state; |
2031 | ||
e126391c PB |
2032 | if (unlikely(state->link.head)) |
2033 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2034 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2035 | io_submit_flush_completions(ctx); |
27926b68 JA |
2036 | if (state->plug_started) |
2037 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2038 | } |
2039 | ||
2040 | /* | |
2041 | * Start submission side cache. | |
2042 | */ | |
2043 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2044 | unsigned int max_ios) |
9a56a232 | 2045 | { |
27926b68 | 2046 | state->plug_started = false; |
4b628aeb | 2047 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2048 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2049 | /* set only head, no need to init link_last in advance */ |
2050 | state->link.head = NULL; | |
9a56a232 JA |
2051 | } |
2052 | ||
2b188cc1 JA |
2053 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2054 | { | |
75b28aff | 2055 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2056 | |
caf582c6 PB |
2057 | /* |
2058 | * Ensure any loads from the SQEs are done at this point, | |
2059 | * since once we write the new head, the application could | |
2060 | * write new data to them. | |
2061 | */ | |
2062 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2063 | } |
2064 | ||
2b188cc1 | 2065 | /* |
dd9ae8a0 | 2066 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2067 | * that is mapped by userspace. This means that care needs to be taken to |
2068 | * ensure that reads are stable, as we cannot rely on userspace always | |
2069 | * being a good citizen. If members of the sqe are validated and then later | |
2070 | * used, it's important that those reads are done through READ_ONCE() to | |
2071 | * prevent a re-load down the line. | |
2072 | */ | |
709b302f | 2073 | static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) |
2b188cc1 | 2074 | { |
ea5ab3b5 | 2075 | unsigned head, mask = ctx->sq_entries - 1; |
17d3aeb3 | 2076 | unsigned sq_idx = ctx->cached_sq_head++ & mask; |
2b188cc1 JA |
2077 | |
2078 | /* | |
2079 | * The cached sq head (or cq tail) serves two purposes: | |
2080 | * | |
2081 | * 1) allows us to batch the cost of updating the user visible | |
2082 | * head updates. | |
2083 | * 2) allows the kernel side to track the head on its own, even | |
2084 | * though the application is the one updating it. | |
2085 | */ | |
17d3aeb3 | 2086 | head = READ_ONCE(ctx->sq_array[sq_idx]); |
ebdeb7c0 JA |
2087 | if (likely(head < ctx->sq_entries)) { |
2088 | /* double index for 128-byte SQEs, twice as long */ | |
2089 | if (ctx->flags & IORING_SETUP_SQE128) | |
2090 | head <<= 1; | |
709b302f | 2091 | return &ctx->sq_sqes[head]; |
ebdeb7c0 | 2092 | } |
2b188cc1 JA |
2093 | |
2094 | /* drop invalid entries */ | |
15641e42 PB |
2095 | ctx->cq_extra--; |
2096 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2097 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
709b302f PB |
2098 | return NULL; |
2099 | } | |
2100 | ||
17437f31 | 2101 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2102 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2103 | { |
69629809 | 2104 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2105 | unsigned int left; |
2106 | int ret; | |
6c271ce2 | 2107 | |
51d48dab | 2108 | if (unlikely(!entries)) |
69629809 | 2109 | return 0; |
ee7d46d9 | 2110 | /* make sure SQ entry isn't read before tail */ |
8e6971a8 PB |
2111 | ret = left = min3(nr, ctx->sq_entries, entries); |
2112 | io_get_task_refs(left); | |
2113 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2114 | |
69629809 | 2115 | do { |
3529d8c2 | 2116 | const struct io_uring_sqe *sqe; |
196be95c | 2117 | struct io_kiocb *req; |
fb5ccc98 | 2118 | |
8e6971a8 | 2119 | if (unlikely(!io_alloc_req_refill(ctx))) |
fb5ccc98 | 2120 | break; |
a33ae9ce | 2121 | req = io_alloc_req(ctx); |
4fccfcbb PB |
2122 | sqe = io_get_sqe(ctx); |
2123 | if (unlikely(!sqe)) { | |
fa05457a | 2124 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2125 | break; |
2126 | } | |
6c271ce2 | 2127 | |
1cd15904 PB |
2128 | /* |
2129 | * Continue submitting even for sqe failure if the | |
2130 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2131 | */ | |
2132 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2133 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2134 | left--; | |
2135 | break; | |
bcbb7bf6 | 2136 | } |
1cd15904 | 2137 | } while (--left); |
9466f437 | 2138 | |
8e6971a8 PB |
2139 | if (unlikely(left)) { |
2140 | ret -= left; | |
2141 | /* try again if it submitted nothing and can't allocate a req */ | |
2142 | if (!ret && io_req_cache_empty(ctx)) | |
2143 | ret = -EAGAIN; | |
2144 | current->io_uring->cached_refs += left; | |
9466f437 | 2145 | } |
6c271ce2 | 2146 | |
553deffd | 2147 | io_submit_state_end(ctx); |
ae9428ca PB |
2148 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2149 | io_commit_sqring(ctx); | |
8e6971a8 | 2150 | return ret; |
6c271ce2 JA |
2151 | } |
2152 | ||
bda52162 JA |
2153 | struct io_wait_queue { |
2154 | struct wait_queue_entry wq; | |
2155 | struct io_ring_ctx *ctx; | |
5fd46178 | 2156 | unsigned cq_tail; |
bda52162 JA |
2157 | unsigned nr_timeouts; |
2158 | }; | |
2159 | ||
6c503150 | 2160 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
2161 | { |
2162 | struct io_ring_ctx *ctx = iowq->ctx; | |
5fd46178 | 2163 | int dist = ctx->cached_cq_tail - (int) iowq->cq_tail; |
bda52162 JA |
2164 | |
2165 | /* | |
d195a66e | 2166 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
2167 | * started waiting. For timeouts, we always want to return to userspace, |
2168 | * regardless of event count. | |
2169 | */ | |
5fd46178 | 2170 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
2171 | } |
2172 | ||
2173 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2174 | int wake_flags, void *key) | |
2175 | { | |
2176 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
2177 | wq); | |
2178 | ||
6c503150 PB |
2179 | /* |
2180 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2181 | * the task, and the next invocation will do it. | |
2182 | */ | |
10988a0a DY |
2183 | if (io_should_wake(iowq) || |
2184 | test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq)) | |
6c503150 PB |
2185 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2186 | return -1; | |
bda52162 JA |
2187 | } |
2188 | ||
73572984 | 2189 | int io_run_task_work_sig(void) |
af9c1a44 JA |
2190 | { |
2191 | if (io_run_task_work()) | |
2192 | return 1; | |
0b8cfa97 | 2193 | if (test_thread_flag(TIF_NOTIFY_SIGNAL)) |
792ee0f6 | 2194 | return -ERESTARTSYS; |
c5020bc8 OL |
2195 | if (task_sigpending(current)) |
2196 | return -EINTR; | |
2197 | return 0; | |
af9c1a44 JA |
2198 | } |
2199 | ||
eeb60b9a PB |
2200 | /* when returns >0, the caller should retry */ |
2201 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
2202 | struct io_wait_queue *iowq, | |
22833966 | 2203 | ktime_t timeout) |
eeb60b9a PB |
2204 | { |
2205 | int ret; | |
155bc950 | 2206 | unsigned long check_cq; |
eeb60b9a PB |
2207 | |
2208 | /* make sure we run task_work before checking for signals */ | |
2209 | ret = io_run_task_work_sig(); | |
2210 | if (ret || io_should_wake(iowq)) | |
2211 | return ret; | |
3a08576b | 2212 | |
155bc950 | 2213 | check_cq = READ_ONCE(ctx->check_cq); |
3a08576b PB |
2214 | if (unlikely(check_cq)) { |
2215 | /* let the caller flush overflows, retry */ | |
2216 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
2217 | return 1; | |
2218 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
2219 | return -EBADR; | |
2220 | } | |
22833966 JA |
2221 | if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS)) |
2222 | return -ETIME; | |
2223 | return 1; | |
eeb60b9a PB |
2224 | } |
2225 | ||
2b188cc1 JA |
2226 | /* |
2227 | * Wait until events become available, if we don't already have some. The | |
2228 | * application must reap them itself, as they reside on the shared cq ring. | |
2229 | */ | |
2230 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2231 | const sigset_t __user *sig, size_t sigsz, |
2232 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2233 | { |
90291099 | 2234 | struct io_wait_queue iowq; |
75b28aff | 2235 | struct io_rings *rings = ctx->rings; |
22833966 | 2236 | ktime_t timeout = KTIME_MAX; |
c1d5a224 | 2237 | int ret; |
2b188cc1 | 2238 | |
b41e9852 | 2239 | do { |
90f67366 | 2240 | io_cqring_overflow_flush(ctx); |
6c503150 | 2241 | if (io_cqring_events(ctx) >= min_events) |
b41e9852 | 2242 | return 0; |
4c6e277c | 2243 | if (!io_run_task_work()) |
b41e9852 | 2244 | break; |
b41e9852 | 2245 | } while (1); |
2b188cc1 JA |
2246 | |
2247 | if (sig) { | |
9e75ad5d AB |
2248 | #ifdef CONFIG_COMPAT |
2249 | if (in_compat_syscall()) | |
2250 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2251 | sigsz); |
9e75ad5d AB |
2252 | else |
2253 | #endif | |
b772434b | 2254 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2255 | |
2b188cc1 JA |
2256 | if (ret) |
2257 | return ret; | |
2258 | } | |
2259 | ||
950e79dd OL |
2260 | if (uts) { |
2261 | struct timespec64 ts; | |
2262 | ||
2263 | if (get_timespec64(&ts, uts)) | |
2264 | return -EFAULT; | |
2265 | timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); | |
2266 | } | |
2267 | ||
90291099 PB |
2268 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
2269 | iowq.wq.private = current; | |
2270 | INIT_LIST_HEAD(&iowq.wq.entry); | |
2271 | iowq.ctx = ctx; | |
bda52162 | 2272 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 2273 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
90291099 | 2274 | |
c826bd7a | 2275 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2276 | do { |
ca0a2651 | 2277 | /* if we can't even flush overflow, don't wait for more */ |
90f67366 | 2278 | if (!io_cqring_overflow_flush(ctx)) { |
ca0a2651 JA |
2279 | ret = -EBUSY; |
2280 | break; | |
2281 | } | |
311997b3 | 2282 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, |
bda52162 | 2283 | TASK_INTERRUPTIBLE); |
22833966 | 2284 | ret = io_cqring_wait_schedule(ctx, &iowq, timeout); |
ca0a2651 | 2285 | cond_resched(); |
eeb60b9a | 2286 | } while (ret > 0); |
bda52162 | 2287 | |
b4f20bb4 | 2288 | finish_wait(&ctx->cq_wait, &iowq.wq); |
b7db41c9 | 2289 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2290 | |
75b28aff | 2291 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2292 | } |
2293 | ||
73572984 | 2294 | static void io_mem_free(void *ptr) |
b9bd2bea | 2295 | { |
73572984 | 2296 | struct page *page; |
b36a2050 | 2297 | |
73572984 JA |
2298 | if (!ptr) |
2299 | return; | |
b9bd2bea | 2300 | |
73572984 JA |
2301 | page = virt_to_head_page(ptr); |
2302 | if (put_page_testzero(page)) | |
2303 | free_compound_page(page); | |
b9bd2bea PB |
2304 | } |
2305 | ||
73572984 | 2306 | static void *io_mem_alloc(size_t size) |
b9bd2bea | 2307 | { |
73572984 | 2308 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; |
b9bd2bea | 2309 | |
73572984 | 2310 | return (void *) __get_free_pages(gfp, get_order(size)); |
b9bd2bea PB |
2311 | } |
2312 | ||
73572984 JA |
2313 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2314 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2315 | { |
73572984 JA |
2316 | struct io_rings *rings; |
2317 | size_t off, sq_array_size; | |
6b06314c | 2318 | |
73572984 JA |
2319 | off = struct_size(rings, cqes, cq_entries); |
2320 | if (off == SIZE_MAX) | |
2321 | return SIZE_MAX; | |
2322 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2323 | if (check_shl_overflow(off, 1, &off)) | |
2324 | return SIZE_MAX; | |
2325 | } | |
ab409402 | 2326 | |
73572984 JA |
2327 | #ifdef CONFIG_SMP |
2328 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2329 | if (off == 0) | |
2330 | return SIZE_MAX; | |
2331 | #endif | |
82fbcfa9 | 2332 | |
73572984 JA |
2333 | if (sq_offset) |
2334 | *sq_offset = off; | |
82fbcfa9 | 2335 | |
73572984 JA |
2336 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2337 | if (sq_array_size == SIZE_MAX) | |
2338 | return SIZE_MAX; | |
6b06314c | 2339 | |
73572984 JA |
2340 | if (check_add_overflow(off, sq_array_size, &off)) |
2341 | return SIZE_MAX; | |
8bad28d8 | 2342 | |
73572984 | 2343 | return off; |
8bad28d8 HX |
2344 | } |
2345 | ||
73572984 JA |
2346 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, |
2347 | unsigned int eventfd_async) | |
8bad28d8 | 2348 | { |
73572984 JA |
2349 | struct io_ev_fd *ev_fd; |
2350 | __s32 __user *fds = arg; | |
2351 | int fd; | |
f2303b1f | 2352 | |
73572984 JA |
2353 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, |
2354 | lockdep_is_held(&ctx->uring_lock)); | |
2355 | if (ev_fd) | |
2356 | return -EBUSY; | |
8bad28d8 | 2357 | |
73572984 JA |
2358 | if (copy_from_user(&fd, fds, sizeof(*fds))) |
2359 | return -EFAULT; | |
8dd03afe | 2360 | |
73572984 JA |
2361 | ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); |
2362 | if (!ev_fd) | |
2363 | return -ENOMEM; | |
05f3fb3c | 2364 | |
73572984 JA |
2365 | ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); |
2366 | if (IS_ERR(ev_fd->cq_ev_fd)) { | |
2367 | int ret = PTR_ERR(ev_fd->cq_ev_fd); | |
2368 | kfree(ev_fd); | |
2369 | return ret; | |
2370 | } | |
305bef98 PB |
2371 | |
2372 | spin_lock(&ctx->completion_lock); | |
2373 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
2374 | spin_unlock(&ctx->completion_lock); | |
2375 | ||
73572984 JA |
2376 | ev_fd->eventfd_async = eventfd_async; |
2377 | ctx->has_evfd = true; | |
2378 | rcu_assign_pointer(ctx->io_ev_fd, ev_fd); | |
2379 | return 0; | |
d7954b2b BM |
2380 | } |
2381 | ||
73572984 | 2382 | static void io_eventfd_put(struct rcu_head *rcu) |
2d091d62 | 2383 | { |
73572984 | 2384 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
2d091d62 | 2385 | |
73572984 JA |
2386 | eventfd_ctx_put(ev_fd->cq_ev_fd); |
2387 | kfree(ev_fd); | |
2d091d62 PB |
2388 | } |
2389 | ||
73572984 | 2390 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 2391 | { |
73572984 JA |
2392 | struct io_ev_fd *ev_fd; |
2393 | ||
2394 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, | |
2395 | lockdep_is_held(&ctx->uring_lock)); | |
2396 | if (ev_fd) { | |
2397 | ctx->has_evfd = false; | |
2398 | rcu_assign_pointer(ctx->io_ev_fd, NULL); | |
2399 | call_rcu(&ev_fd->rcu, io_eventfd_put); | |
2400 | return 0; | |
2401 | } | |
2d091d62 | 2402 | |
73572984 | 2403 | return -ENXIO; |
44b31f2f PB |
2404 | } |
2405 | ||
73572984 | 2406 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2407 | { |
cd0ca2e0 | 2408 | struct io_submit_state *state = &ctx->submit_state; |
37f0e767 | 2409 | int nr = 0; |
bf019da7 | 2410 | |
9a4fdbd8 | 2411 | mutex_lock(&ctx->uring_lock); |
cd0ca2e0 | 2412 | io_flush_cached_locked_reqs(ctx, state); |
9a4fdbd8 | 2413 | |
88ab95be | 2414 | while (!io_req_cache_empty(ctx)) { |
c2b6c6bc PB |
2415 | struct io_wq_work_node *node; |
2416 | struct io_kiocb *req; | |
9a4fdbd8 | 2417 | |
c2b6c6bc PB |
2418 | node = wq_stack_extract(&state->free_list); |
2419 | req = container_of(node, struct io_kiocb, comp_list); | |
2420 | kmem_cache_free(req_cachep, req); | |
37f0e767 | 2421 | nr++; |
c2b6c6bc | 2422 | } |
37f0e767 PB |
2423 | if (nr) |
2424 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2425 | mutex_unlock(&ctx->uring_lock); |
2426 | } | |
2427 | ||
4d9237e3 JA |
2428 | static void io_flush_apoll_cache(struct io_ring_ctx *ctx) |
2429 | { | |
2430 | struct async_poll *apoll; | |
2431 | ||
2432 | while (!list_empty(&ctx->apoll_cache)) { | |
2433 | apoll = list_first_entry(&ctx->apoll_cache, struct async_poll, | |
2434 | poll.wait.entry); | |
2435 | list_del(&apoll->poll.wait.entry); | |
2436 | kfree(apoll); | |
2437 | } | |
2438 | } | |
2439 | ||
c072481d | 2440 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2441 | { |
37d1e2e3 | 2442 | io_sq_thread_finish(ctx); |
2aede0e4 | 2443 | |
37d1e2e3 | 2444 | if (ctx->mm_account) { |
2aede0e4 JA |
2445 | mmdrop(ctx->mm_account); |
2446 | ctx->mm_account = NULL; | |
30975825 | 2447 | } |
def596e9 | 2448 | |
ab409402 | 2449 | io_rsrc_refs_drop(ctx); |
43597aac PB |
2450 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
2451 | io_wait_rsrc_data(ctx->buf_data); | |
2452 | io_wait_rsrc_data(ctx->file_data); | |
2453 | ||
8bad28d8 | 2454 | mutex_lock(&ctx->uring_lock); |
43597aac | 2455 | if (ctx->buf_data) |
bd54b6fe | 2456 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2457 | if (ctx->file_data) |
08480400 | 2458 | __io_sqe_files_unregister(ctx); |
c4ea060e PB |
2459 | if (ctx->rings) |
2460 | __io_cqring_overflow_flush(ctx, true); | |
9b402849 | 2461 | io_eventfd_unregister(ctx); |
4d9237e3 | 2462 | io_flush_apoll_cache(ctx); |
77bc59b4 | 2463 | mutex_unlock(&ctx->uring_lock); |
5a2e745d | 2464 | io_destroy_buffers(ctx); |
07db298a PB |
2465 | if (ctx->sq_creds) |
2466 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2467 | if (ctx->submitter_task) |
2468 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2469 | |
a7f0ed5a PB |
2470 | /* there are no registered resources left, nobody uses it */ |
2471 | if (ctx->rsrc_node) | |
2472 | io_rsrc_node_destroy(ctx->rsrc_node); | |
8dd03afe | 2473 | if (ctx->rsrc_backup_node) |
b895c9a6 | 2474 | io_rsrc_node_destroy(ctx->rsrc_backup_node); |
a7f0ed5a | 2475 | flush_delayed_work(&ctx->rsrc_put_work); |
756ab7c0 | 2476 | flush_delayed_work(&ctx->fallback_work); |
a7f0ed5a PB |
2477 | |
2478 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
2479 | WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist)); | |
def596e9 | 2480 | |
2b188cc1 | 2481 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
2482 | if (ctx->ring_sock) { |
2483 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 2484 | sock_release(ctx->ring_sock); |
355e8d26 | 2485 | } |
2b188cc1 | 2486 | #endif |
ef9dd637 | 2487 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
2b188cc1 | 2488 | |
75b28aff | 2489 | io_mem_free(ctx->rings); |
2b188cc1 | 2490 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
2491 | |
2492 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2493 | free_uid(ctx->user); |
4010fec4 | 2494 | io_req_caches_free(ctx); |
e941894e JA |
2495 | if (ctx->hash_map) |
2496 | io_wq_put_hash(ctx->hash_map); | |
e6f89be6 | 2497 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2498 | kfree(ctx->cancel_table_locked.hbs); |
6224843d | 2499 | kfree(ctx->dummy_ubuf); |
9cfc7e94 JA |
2500 | kfree(ctx->io_bl); |
2501 | xa_destroy(&ctx->io_bl_xa); | |
2b188cc1 JA |
2502 | kfree(ctx); |
2503 | } | |
2504 | ||
2505 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
2506 | { | |
2507 | struct io_ring_ctx *ctx = file->private_data; | |
2508 | __poll_t mask = 0; | |
2509 | ||
d60aa65b | 2510 | poll_wait(file, &ctx->cq_wait, wait); |
4f7067c3 SB |
2511 | /* |
2512 | * synchronizes with barrier from wq_has_sleeper call in | |
2513 | * io_commit_cqring | |
2514 | */ | |
2b188cc1 | 2515 | smp_rmb(); |
90554200 | 2516 | if (!io_sqring_full(ctx)) |
2b188cc1 | 2517 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
2518 | |
2519 | /* | |
2520 | * Don't flush cqring overflow list here, just do a simple check. | |
2521 | * Otherwise there could possible be ABBA deadlock: | |
2522 | * CPU0 CPU1 | |
2523 | * ---- ---- | |
2524 | * lock(&ctx->uring_lock); | |
2525 | * lock(&ep->mtx); | |
2526 | * lock(&ctx->uring_lock); | |
2527 | * lock(&ep->mtx); | |
2528 | * | |
2529 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
2530 | * pushs them to do the flush. | |
2531 | */ | |
10988a0a DY |
2532 | if (io_cqring_events(ctx) || |
2533 | test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) | |
2b188cc1 JA |
2534 | mask |= EPOLLIN | EPOLLRDNORM; |
2535 | ||
2536 | return mask; | |
2537 | } | |
2538 | ||
0bead8cd | 2539 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 2540 | { |
4379bf8b | 2541 | const struct cred *creds; |
071698e1 | 2542 | |
61cf9370 | 2543 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
2544 | if (creds) { |
2545 | put_cred(creds); | |
0bead8cd | 2546 | return 0; |
1e6fa521 | 2547 | } |
0bead8cd YD |
2548 | |
2549 | return -EINVAL; | |
2550 | } | |
2551 | ||
d56d938b PB |
2552 | struct io_tctx_exit { |
2553 | struct callback_head task_work; | |
2554 | struct completion completion; | |
baf186c4 | 2555 | struct io_ring_ctx *ctx; |
d56d938b PB |
2556 | }; |
2557 | ||
c072481d | 2558 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
2559 | { |
2560 | struct io_uring_task *tctx = current->io_uring; | |
2561 | struct io_tctx_exit *work; | |
2562 | ||
2563 | work = container_of(cb, struct io_tctx_exit, task_work); | |
2564 | /* | |
2565 | * When @in_idle, we're in cancellation and it's racy to remove the | |
2566 | * node. It'll be removed by the end of cancellation, just ignore it. | |
2567 | */ | |
2568 | if (!atomic_read(&tctx->in_idle)) | |
eef51daa | 2569 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
2570 | complete(&work->completion); |
2571 | } | |
2572 | ||
c072481d | 2573 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
2574 | { |
2575 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2576 | ||
2577 | return req->ctx == data; | |
2578 | } | |
2579 | ||
c072481d | 2580 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 2581 | { |
d56d938b | 2582 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 2583 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 2584 | unsigned long interval = HZ / 20; |
d56d938b PB |
2585 | struct io_tctx_exit exit; |
2586 | struct io_tctx_node *node; | |
2587 | int ret; | |
85faa7b8 | 2588 | |
56952e91 JA |
2589 | /* |
2590 | * If we're doing polled IO and end up having requests being | |
2591 | * submitted async (out-of-line), then completions can come in while | |
2592 | * we're waiting for refs to drop. We need to reap these manually, | |
2593 | * as nobody else will be looking for them. | |
2594 | */ | |
b2edc0a7 | 2595 | do { |
affa87db PB |
2596 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
2597 | cond_resched(); | |
2598 | ||
28090c13 PB |
2599 | if (ctx->sq_data) { |
2600 | struct io_sq_data *sqd = ctx->sq_data; | |
2601 | struct task_struct *tsk; | |
2602 | ||
2603 | io_sq_thread_park(sqd); | |
2604 | tsk = sqd->thread; | |
2605 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
2606 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
2607 | io_cancel_ctx_cb, ctx, true); | |
2608 | io_sq_thread_unpark(sqd); | |
2609 | } | |
b5bb3a24 | 2610 | |
37f0e767 PB |
2611 | io_req_caches_free(ctx); |
2612 | ||
58d3be2c PB |
2613 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
2614 | /* there is little hope left, don't run it too often */ | |
2615 | interval = HZ * 60; | |
2616 | } | |
2617 | } while (!wait_for_completion_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 2618 | |
7f00651a PB |
2619 | init_completion(&exit.completion); |
2620 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
2621 | exit.ctx = ctx; | |
89b5066e PB |
2622 | /* |
2623 | * Some may use context even when all refs and requests have been put, | |
2624 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 2625 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
2626 | * this lock/unlock section also waits them to finish. |
2627 | */ | |
d56d938b PB |
2628 | mutex_lock(&ctx->uring_lock); |
2629 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
2630 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
2631 | ||
d56d938b PB |
2632 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
2633 | ctx_node); | |
7f00651a PB |
2634 | /* don't spin on a single task if cancellation failed */ |
2635 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
2636 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
2637 | if (WARN_ON_ONCE(ret)) | |
2638 | continue; | |
d56d938b PB |
2639 | |
2640 | mutex_unlock(&ctx->uring_lock); | |
2641 | wait_for_completion(&exit.completion); | |
d56d938b PB |
2642 | mutex_lock(&ctx->uring_lock); |
2643 | } | |
2644 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
2645 | spin_lock(&ctx->completion_lock); |
2646 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 2647 | |
85faa7b8 JA |
2648 | io_ring_ctx_free(ctx); |
2649 | } | |
2650 | ||
c072481d | 2651 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 2652 | { |
61cf9370 MWO |
2653 | unsigned long index; |
2654 | struct creds *creds; | |
2655 | ||
2b188cc1 JA |
2656 | mutex_lock(&ctx->uring_lock); |
2657 | percpu_ref_kill(&ctx->refs); | |
634578f8 | 2658 | if (ctx->rings) |
6c2450ae | 2659 | __io_cqring_overflow_flush(ctx, true); |
61cf9370 MWO |
2660 | xa_for_each(&ctx->personalities, index, creds) |
2661 | io_unregister_personality(ctx, index); | |
9ca9fb24 PB |
2662 | if (ctx->rings) |
2663 | io_poll_remove_all(ctx, NULL, true); | |
2b188cc1 JA |
2664 | mutex_unlock(&ctx->uring_lock); |
2665 | ||
60053be8 PB |
2666 | /* failed during ring init, it couldn't have issued any requests */ |
2667 | if (ctx->rings) { | |
2668 | io_kill_timeouts(ctx, NULL, true); | |
60053be8 PB |
2669 | /* if we failed setting up the ctx, we might not have any rings */ |
2670 | io_iopoll_try_reap_events(ctx); | |
2671 | } | |
309fc03a | 2672 | |
85faa7b8 | 2673 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
2674 | /* |
2675 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
2676 | * if we're exiting a ton of rings at the same time. It just adds | |
2677 | * noise and overhead, there's no discernable change in runtime | |
2678 | * over using system_wq. | |
2679 | */ | |
2680 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
2681 | } |
2682 | ||
2683 | static int io_uring_release(struct inode *inode, struct file *file) | |
2684 | { | |
2685 | struct io_ring_ctx *ctx = file->private_data; | |
2686 | ||
2687 | file->private_data = NULL; | |
2688 | io_ring_ctx_wait_and_kill(ctx); | |
2689 | return 0; | |
2690 | } | |
2691 | ||
f6edbabb PB |
2692 | struct io_task_cancel { |
2693 | struct task_struct *task; | |
3dd0c97a | 2694 | bool all; |
f6edbabb | 2695 | }; |
f254ac04 | 2696 | |
f6edbabb | 2697 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 2698 | { |
9a472ef7 | 2699 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 2700 | struct io_task_cancel *cancel = data; |
9a472ef7 | 2701 | |
6af3f48b | 2702 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
2703 | } |
2704 | ||
c072481d PB |
2705 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
2706 | struct task_struct *task, | |
2707 | bool cancel_all) | |
b7ddce3c | 2708 | { |
e1915f76 | 2709 | struct io_defer_entry *de; |
b7ddce3c PB |
2710 | LIST_HEAD(list); |
2711 | ||
79ebeaee | 2712 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 2713 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 2714 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
2715 | list_cut_position(&list, &ctx->defer_list, &de->list); |
2716 | break; | |
2717 | } | |
2718 | } | |
79ebeaee | 2719 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
2720 | if (list_empty(&list)) |
2721 | return false; | |
b7ddce3c PB |
2722 | |
2723 | while (!list_empty(&list)) { | |
2724 | de = list_first_entry(&list, struct io_defer_entry, list); | |
2725 | list_del_init(&de->list); | |
f41db273 | 2726 | io_req_complete_failed(de->req, -ECANCELED); |
b7ddce3c PB |
2727 | kfree(de); |
2728 | } | |
e1915f76 | 2729 | return true; |
b7ddce3c PB |
2730 | } |
2731 | ||
c072481d | 2732 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
2733 | { |
2734 | struct io_tctx_node *node; | |
2735 | enum io_wq_cancel cret; | |
2736 | bool ret = false; | |
2737 | ||
2738 | mutex_lock(&ctx->uring_lock); | |
2739 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
2740 | struct io_uring_task *tctx = node->task->io_uring; | |
2741 | ||
2742 | /* | |
2743 | * io_wq will stay alive while we hold uring_lock, because it's | |
2744 | * killed after ctx nodes, which requires to take the lock. | |
2745 | */ | |
2746 | if (!tctx || !tctx->io_wq) | |
2747 | continue; | |
2748 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
2749 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
2750 | } | |
2751 | mutex_unlock(&ctx->uring_lock); | |
2752 | ||
2753 | return ret; | |
2754 | } | |
2755 | ||
affa87db | 2756 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
2757 | struct task_struct *task, |
2758 | bool cancel_all) | |
9936c7c2 | 2759 | { |
3dd0c97a | 2760 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 2761 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
2762 | enum io_wq_cancel cret; |
2763 | bool ret = false; | |
9936c7c2 | 2764 | |
60053be8 PB |
2765 | /* failed during ring init, it couldn't have issued any requests */ |
2766 | if (!ctx->rings) | |
affa87db | 2767 | return false; |
60053be8 | 2768 | |
affa87db PB |
2769 | if (!task) { |
2770 | ret |= io_uring_try_cancel_iowq(ctx); | |
2771 | } else if (tctx && tctx->io_wq) { | |
2772 | /* | |
2773 | * Cancels requests of all rings, not only @ctx, but | |
2774 | * it's fine as the task is in exit/exec. | |
2775 | */ | |
2776 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
2777 | &cancel, true); | |
2778 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
2779 | } | |
9936c7c2 | 2780 | |
affa87db PB |
2781 | /* SQPOLL thread does its own polling */ |
2782 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
2783 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
2784 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
2785 | io_iopoll_try_reap_events(ctx); | |
2786 | ret = true; | |
9936c7c2 | 2787 | } |
9936c7c2 | 2788 | } |
affa87db PB |
2789 | |
2790 | ret |= io_cancel_defer_files(ctx, task, cancel_all); | |
2791 | mutex_lock(&ctx->uring_lock); | |
2792 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
2793 | mutex_unlock(&ctx->uring_lock); | |
2794 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
2795 | if (task) | |
2796 | ret |= io_run_task_work(); | |
2797 | return ret; | |
9936c7c2 PB |
2798 | } |
2799 | ||
3f48cf18 | 2800 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 2801 | { |
3f48cf18 | 2802 | if (tracked) |
9cae36a0 | 2803 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
2804 | return percpu_counter_sum(&tctx->inflight); |
2805 | } | |
2806 | ||
78cc687b PB |
2807 | /* |
2808 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 2809 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 2810 | */ |
17437f31 | 2811 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 2812 | { |
521d6a73 | 2813 | struct io_uring_task *tctx = current->io_uring; |
734551df | 2814 | struct io_ring_ctx *ctx; |
0e9ddb39 PB |
2815 | s64 inflight; |
2816 | DEFINE_WAIT(wait); | |
fdaf083c | 2817 | |
78cc687b PB |
2818 | WARN_ON_ONCE(sqd && sqd->thread != current); |
2819 | ||
6d042ffb PO |
2820 | if (!current->io_uring) |
2821 | return; | |
17a91051 PB |
2822 | if (tctx->io_wq) |
2823 | io_wq_exit_start(tctx->io_wq); | |
2824 | ||
0e9ddb39 PB |
2825 | atomic_inc(&tctx->in_idle); |
2826 | do { | |
affa87db PB |
2827 | bool loop = false; |
2828 | ||
e9dbe221 | 2829 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 2830 | /* read completions before cancelations */ |
78cc687b | 2831 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
2832 | if (!inflight) |
2833 | break; | |
fdaf083c | 2834 | |
78cc687b PB |
2835 | if (!sqd) { |
2836 | struct io_tctx_node *node; | |
2837 | unsigned long index; | |
0f212204 | 2838 | |
78cc687b PB |
2839 | xa_for_each(&tctx->xa, index, node) { |
2840 | /* sqpoll task will cancel all its requests */ | |
2841 | if (node->ctx->sq_data) | |
2842 | continue; | |
affa87db PB |
2843 | loop |= io_uring_try_cancel_requests(node->ctx, |
2844 | current, cancel_all); | |
78cc687b PB |
2845 | } |
2846 | } else { | |
2847 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
2848 | loop |= io_uring_try_cancel_requests(ctx, |
2849 | current, | |
2850 | cancel_all); | |
2851 | } | |
2852 | ||
2853 | if (loop) { | |
2854 | cond_resched(); | |
2855 | continue; | |
78cc687b | 2856 | } |
17a91051 | 2857 | |
78a78060 JA |
2858 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
2859 | io_run_task_work(); | |
e9dbe221 | 2860 | io_uring_drop_tctx_refs(current); |
78a78060 | 2861 | |
0f212204 | 2862 | /* |
a1bb3cd5 PB |
2863 | * If we've seen completions, retry without waiting. This |
2864 | * avoids a race where a completion comes in before we did | |
2865 | * prepare_to_wait(). | |
0f212204 | 2866 | */ |
3dd0c97a | 2867 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 2868 | schedule(); |
f57555ed | 2869 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 2870 | } while (1); |
de7f1d9e | 2871 | |
8452d4a6 | 2872 | io_uring_clean_tctx(tctx); |
3dd0c97a | 2873 | if (cancel_all) { |
3cc7fdb9 PB |
2874 | /* |
2875 | * We shouldn't run task_works after cancel, so just leave | |
2876 | * ->in_idle set for normal exit. | |
2877 | */ | |
2878 | atomic_dec(&tctx->in_idle); | |
3f48cf18 PB |
2879 | /* for exec all current's requests should be gone, kill tctx */ |
2880 | __io_uring_free(current); | |
2881 | } | |
44e728b8 PB |
2882 | } |
2883 | ||
f552a27a | 2884 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 2885 | { |
f552a27a | 2886 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
2887 | } |
2888 | ||
6c5c240e RP |
2889 | static void *io_uring_validate_mmap_request(struct file *file, |
2890 | loff_t pgoff, size_t sz) | |
2b188cc1 | 2891 | { |
2b188cc1 | 2892 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 2893 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
2894 | struct page *page; |
2895 | void *ptr; | |
2896 | ||
2897 | switch (offset) { | |
2898 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
2899 | case IORING_OFF_CQ_RING: |
2900 | ptr = ctx->rings; | |
2b188cc1 JA |
2901 | break; |
2902 | case IORING_OFF_SQES: | |
2903 | ptr = ctx->sq_sqes; | |
2904 | break; | |
2b188cc1 | 2905 | default: |
6c5c240e | 2906 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
2907 | } |
2908 | ||
2909 | page = virt_to_head_page(ptr); | |
a50b854e | 2910 | if (sz > page_size(page)) |
6c5c240e RP |
2911 | return ERR_PTR(-EINVAL); |
2912 | ||
2913 | return ptr; | |
2914 | } | |
2915 | ||
2916 | #ifdef CONFIG_MMU | |
2917 | ||
c072481d | 2918 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e RP |
2919 | { |
2920 | size_t sz = vma->vm_end - vma->vm_start; | |
2921 | unsigned long pfn; | |
2922 | void *ptr; | |
2923 | ||
2924 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
2925 | if (IS_ERR(ptr)) | |
2926 | return PTR_ERR(ptr); | |
2b188cc1 JA |
2927 | |
2928 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
2929 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
2930 | } | |
2931 | ||
6c5c240e RP |
2932 | #else /* !CONFIG_MMU */ |
2933 | ||
2934 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
2935 | { | |
2936 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
2937 | } | |
2938 | ||
2939 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
2940 | { | |
2941 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
2942 | } | |
2943 | ||
2944 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
2945 | unsigned long addr, unsigned long len, | |
2946 | unsigned long pgoff, unsigned long flags) | |
2947 | { | |
2948 | void *ptr; | |
2949 | ||
2950 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
2951 | if (IS_ERR(ptr)) | |
2952 | return PTR_ERR(ptr); | |
2953 | ||
2954 | return (unsigned long) ptr; | |
2955 | } | |
2956 | ||
2957 | #endif /* !CONFIG_MMU */ | |
2958 | ||
f81440d3 PB |
2959 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
2960 | { | |
2961 | if (flags & IORING_ENTER_EXT_ARG) { | |
2962 | struct io_uring_getevents_arg arg; | |
2963 | ||
2964 | if (argsz != sizeof(arg)) | |
2965 | return -EINVAL; | |
2966 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
2967 | return -EFAULT; | |
2968 | } | |
2969 | return 0; | |
2970 | } | |
2971 | ||
c73ebb68 HX |
2972 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
2973 | struct __kernel_timespec __user **ts, | |
2974 | const sigset_t __user **sig) | |
2975 | { | |
2976 | struct io_uring_getevents_arg arg; | |
2977 | ||
2978 | /* | |
2979 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
2980 | * is just a pointer to the sigset_t. | |
2981 | */ | |
2982 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
2983 | *sig = (const sigset_t __user *) argp; | |
2984 | *ts = NULL; | |
2985 | return 0; | |
2986 | } | |
2987 | ||
2988 | /* | |
2989 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
2990 | * timespec and sigset_t pointers if good. | |
2991 | */ | |
2992 | if (*argsz != sizeof(arg)) | |
2993 | return -EINVAL; | |
2994 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
2995 | return -EFAULT; | |
d2347b96 DY |
2996 | if (arg.pad) |
2997 | return -EINVAL; | |
c73ebb68 HX |
2998 | *sig = u64_to_user_ptr(arg.sigmask); |
2999 | *argsz = arg.sigmask_sz; | |
3000 | *ts = u64_to_user_ptr(arg.ts); | |
3001 | return 0; | |
3002 | } | |
3003 | ||
2b188cc1 | 3004 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
3005 | u32, min_complete, u32, flags, const void __user *, argp, |
3006 | size_t, argsz) | |
2b188cc1 JA |
3007 | { |
3008 | struct io_ring_ctx *ctx; | |
2b188cc1 | 3009 | struct fd f; |
33f993da | 3010 | long ret; |
2b188cc1 | 3011 | |
4c6e277c | 3012 | io_run_task_work(); |
b41e9852 | 3013 | |
33f993da | 3014 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
3015 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
3016 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
3017 | return -EINVAL; |
3018 | ||
e7a6c00d JA |
3019 | /* |
3020 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3021 | * need only dereference our task private array to find it. | |
3022 | */ | |
3023 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3024 | struct io_uring_task *tctx = current->io_uring; | |
3025 | ||
3026 | if (!tctx || fd >= IO_RINGFD_REG_MAX) | |
3027 | return -EINVAL; | |
3028 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
3029 | f.file = tctx->registered_rings[fd]; | |
4329490a | 3030 | f.flags = 0; |
e7a6c00d JA |
3031 | } else { |
3032 | f = fdget(fd); | |
e7a6c00d | 3033 | } |
2b188cc1 | 3034 | |
4329490a AV |
3035 | if (unlikely(!f.file)) |
3036 | return -EBADF; | |
3037 | ||
2b188cc1 | 3038 | ret = -EOPNOTSUPP; |
e5550a14 | 3039 | if (unlikely(!io_is_uring_fops(f.file))) |
2b188cc1 JA |
3040 | goto out_fput; |
3041 | ||
3042 | ret = -ENXIO; | |
3043 | ctx = f.file->private_data; | |
33f993da | 3044 | if (unlikely(!percpu_ref_tryget(&ctx->refs))) |
2b188cc1 JA |
3045 | goto out_fput; |
3046 | ||
7e84e1c7 | 3047 | ret = -EBADFD; |
33f993da | 3048 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3049 | goto out; |
3050 | ||
6c271ce2 JA |
3051 | /* |
3052 | * For SQ polling, the thread will do all submissions and completions. | |
3053 | * Just return the requested submit count, and wake the thread if | |
3054 | * we were asked to. | |
3055 | */ | |
b2a9eada | 3056 | ret = 0; |
6c271ce2 | 3057 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3058 | io_cqring_overflow_flush(ctx); |
89448c47 | 3059 | |
21f96522 JA |
3060 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3061 | ret = -EOWNERDEAD; | |
04147488 | 3062 | goto out; |
21f96522 | 3063 | } |
6c271ce2 | 3064 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3065 | wake_up(&ctx->sq_data->wait); |
d9d05217 PB |
3066 | if (flags & IORING_ENTER_SQ_WAIT) { |
3067 | ret = io_sqpoll_wait_sq(ctx); | |
3068 | if (ret) | |
3069 | goto out; | |
3070 | } | |
3e813c90 | 3071 | ret = to_submit; |
b2a9eada | 3072 | } else if (to_submit) { |
eef51daa | 3073 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3074 | if (unlikely(ret)) |
3075 | goto out; | |
7c504e65 | 3076 | |
2b188cc1 | 3077 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3078 | ret = io_submit_sqes(ctx, to_submit); |
3079 | if (ret != to_submit) { | |
d487b43c | 3080 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3081 | goto out; |
d487b43c PB |
3082 | } |
3083 | if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll) | |
3084 | goto iopoll_locked; | |
3085 | mutex_unlock(&ctx->uring_lock); | |
2b188cc1 JA |
3086 | } |
3087 | if (flags & IORING_ENTER_GETEVENTS) { | |
3e813c90 | 3088 | int ret2; |
773697b6 | 3089 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3090 | /* |
3091 | * We disallow the app entering submit/complete with | |
3092 | * polling, but we still need to lock the ring to | |
3093 | * prevent racing with polled issue that got punted to | |
3094 | * a workqueue. | |
3095 | */ | |
3096 | mutex_lock(&ctx->uring_lock); | |
3097 | iopoll_locked: | |
3e813c90 DY |
3098 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3099 | if (likely(!ret2)) { | |
3100 | min_complete = min(min_complete, | |
3101 | ctx->cq_entries); | |
3102 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3103 | } |
3104 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3105 | } else { |
f81440d3 PB |
3106 | const sigset_t __user *sig; |
3107 | struct __kernel_timespec __user *ts; | |
3108 | ||
3e813c90 DY |
3109 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3110 | if (likely(!ret2)) { | |
3111 | min_complete = min(min_complete, | |
3112 | ctx->cq_entries); | |
3113 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3114 | argsz, ts); | |
3115 | } | |
def596e9 | 3116 | } |
c73ebb68 | 3117 | |
155bc950 | 3118 | if (!ret) { |
3e813c90 | 3119 | ret = ret2; |
2b188cc1 | 3120 | |
155bc950 DY |
3121 | /* |
3122 | * EBADR indicates that one or more CQE were dropped. | |
3123 | * Once the user has been informed we can clear the bit | |
3124 | * as they are obviously ok with those drops. | |
3125 | */ | |
3126 | if (unlikely(ret2 == -EBADR)) | |
3127 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3128 | &ctx->check_cq); | |
def596e9 | 3129 | } |
2b188cc1 JA |
3130 | } |
3131 | ||
7c504e65 | 3132 | out: |
6805b32e | 3133 | percpu_ref_put(&ctx->refs); |
2b188cc1 | 3134 | out_fput: |
4329490a | 3135 | fdput(f); |
3e813c90 | 3136 | return ret; |
2b188cc1 JA |
3137 | } |
3138 | ||
3139 | static const struct file_operations io_uring_fops = { | |
3140 | .release = io_uring_release, | |
3141 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3142 | #ifndef CONFIG_MMU |
3143 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3144 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
3145 | #endif | |
2b188cc1 | 3146 | .poll = io_uring_poll, |
bebdb65e | 3147 | #ifdef CONFIG_PROC_FS |
87ce955b | 3148 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3149 | #endif |
2b188cc1 JA |
3150 | }; |
3151 | ||
92ac8bea JA |
3152 | bool io_is_uring_fops(struct file *file) |
3153 | { | |
3154 | return file->f_op == &io_uring_fops; | |
3155 | } | |
3156 | ||
c072481d PB |
3157 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3158 | struct io_uring_params *p) | |
2b188cc1 | 3159 | { |
75b28aff HV |
3160 | struct io_rings *rings; |
3161 | size_t size, sq_array_offset; | |
2b188cc1 | 3162 | |
bd740481 JA |
3163 | /* make sure these are sane, as we already accounted them */ |
3164 | ctx->sq_entries = p->sq_entries; | |
3165 | ctx->cq_entries = p->cq_entries; | |
3166 | ||
baf9cb64 | 3167 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3168 | if (size == SIZE_MAX) |
3169 | return -EOVERFLOW; | |
3170 | ||
3171 | rings = io_mem_alloc(size); | |
3172 | if (!rings) | |
2b188cc1 JA |
3173 | return -ENOMEM; |
3174 | ||
75b28aff HV |
3175 | ctx->rings = rings; |
3176 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
3177 | rings->sq_ring_mask = p->sq_entries - 1; | |
3178 | rings->cq_ring_mask = p->cq_entries - 1; | |
3179 | rings->sq_ring_entries = p->sq_entries; | |
3180 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3181 | |
ebdeb7c0 JA |
3182 | if (p->flags & IORING_SETUP_SQE128) |
3183 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3184 | else | |
3185 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
3186 | if (size == SIZE_MAX) { |
3187 | io_mem_free(ctx->rings); | |
3188 | ctx->rings = NULL; | |
2b188cc1 | 3189 | return -EOVERFLOW; |
eb065d30 | 3190 | } |
2b188cc1 JA |
3191 | |
3192 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
3193 | if (!ctx->sq_sqes) { |
3194 | io_mem_free(ctx->rings); | |
3195 | ctx->rings = NULL; | |
2b188cc1 | 3196 | return -ENOMEM; |
eb065d30 | 3197 | } |
2b188cc1 | 3198 | |
2b188cc1 JA |
3199 | return 0; |
3200 | } | |
3201 | ||
9faadcc8 PB |
3202 | static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) |
3203 | { | |
3204 | int ret, fd; | |
3205 | ||
3206 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3207 | if (fd < 0) | |
3208 | return fd; | |
3209 | ||
97bbdc06 | 3210 | ret = __io_uring_add_tctx_node(ctx, false); |
9faadcc8 PB |
3211 | if (ret) { |
3212 | put_unused_fd(fd); | |
3213 | return ret; | |
3214 | } | |
3215 | fd_install(fd, file); | |
3216 | return fd; | |
3217 | } | |
3218 | ||
2b188cc1 JA |
3219 | /* |
3220 | * Allocate an anonymous fd, this is what constitutes the application | |
3221 | * visible backing of an io_uring instance. The application mmaps this | |
3222 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3223 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3224 | */ | |
9faadcc8 | 3225 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
3226 | { |
3227 | struct file *file; | |
9faadcc8 | 3228 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
3229 | int ret; |
3230 | ||
2b188cc1 JA |
3231 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
3232 | &ctx->ring_sock); | |
3233 | if (ret) | |
9faadcc8 | 3234 | return ERR_PTR(ret); |
2b188cc1 JA |
3235 | #endif |
3236 | ||
91a9ab7c PM |
3237 | file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, |
3238 | O_RDWR | O_CLOEXEC, NULL); | |
2b188cc1 | 3239 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
3240 | if (IS_ERR(file)) { |
3241 | sock_release(ctx->ring_sock); | |
3242 | ctx->ring_sock = NULL; | |
3243 | } else { | |
3244 | ctx->ring_sock->file = file; | |
0f212204 | 3245 | } |
2b188cc1 | 3246 | #endif |
9faadcc8 | 3247 | return file; |
2b188cc1 JA |
3248 | } |
3249 | ||
c072481d PB |
3250 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3251 | struct io_uring_params __user *params) | |
2b188cc1 | 3252 | { |
2b188cc1 | 3253 | struct io_ring_ctx *ctx; |
9faadcc8 | 3254 | struct file *file; |
2b188cc1 JA |
3255 | int ret; |
3256 | ||
8110c1a6 | 3257 | if (!entries) |
2b188cc1 | 3258 | return -EINVAL; |
8110c1a6 JA |
3259 | if (entries > IORING_MAX_ENTRIES) { |
3260 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3261 | return -EINVAL; | |
3262 | entries = IORING_MAX_ENTRIES; | |
3263 | } | |
2b188cc1 JA |
3264 | |
3265 | /* | |
3266 | * Use twice as many entries for the CQ ring. It's possible for the | |
3267 | * application to drive a higher depth than the size of the SQ ring, | |
3268 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3269 | * some flexibility in overcommitting a bit. If the application has |
3270 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3271 | * of CQ ring entries manually. | |
2b188cc1 JA |
3272 | */ |
3273 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3274 | if (p->flags & IORING_SETUP_CQSIZE) { |
3275 | /* | |
3276 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3277 | * to a power-of-two, if it isn't already. We do NOT impose | |
3278 | * any cq vs sq ring sizing. | |
3279 | */ | |
eb2667b3 | 3280 | if (!p->cq_entries) |
33a107f0 | 3281 | return -EINVAL; |
8110c1a6 JA |
3282 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3283 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3284 | return -EINVAL; | |
3285 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3286 | } | |
eb2667b3 JQ |
3287 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3288 | if (p->cq_entries < p->sq_entries) | |
3289 | return -EINVAL; | |
33a107f0 JA |
3290 | } else { |
3291 | p->cq_entries = 2 * p->sq_entries; | |
3292 | } | |
2b188cc1 | 3293 | |
2b188cc1 | 3294 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3295 | if (!ctx) |
2b188cc1 | 3296 | return -ENOMEM; |
773697b6 PB |
3297 | |
3298 | /* | |
3299 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3300 | * space applications don't need to do io completion events | |
3301 | * polling again, they can rely on io_sq_thread to do polling | |
3302 | * work, which can reduce cpu usage and uring_lock contention. | |
3303 | */ | |
3304 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3305 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3306 | ctx->syscall_iopoll = 1; | |
3307 | ||
2b188cc1 | 3308 | ctx->compat = in_compat_syscall(); |
62e398be JA |
3309 | if (!capable(CAP_IPC_LOCK)) |
3310 | ctx->user = get_uid(current_user()); | |
2aede0e4 | 3311 | |
9f010507 | 3312 | /* |
e1169f06 JA |
3313 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3314 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3315 | */ |
e1169f06 JA |
3316 | ret = -EINVAL; |
3317 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3318 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 JA |
3319 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
3320 | IORING_SETUP_TASKRUN_FLAG)) | |
e1169f06 | 3321 | goto err; |
9f010507 | 3322 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3323 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3324 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3325 | } else { | |
ef060ea9 JA |
3326 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
3327 | goto err; | |
9f010507 | 3328 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3329 | } |
9f010507 | 3330 | |
2aede0e4 JA |
3331 | /* |
3332 | * This is just grabbed for accounting purposes. When a process exits, | |
3333 | * the mm is exited and dropped before the files, hence we need to hang | |
3334 | * on to this mm purely for the purposes of being able to unaccount | |
3335 | * memory (locked/pinned vm). It's not used for anything else. | |
3336 | */ | |
6b7898eb | 3337 | mmgrab(current->mm); |
2aede0e4 | 3338 | ctx->mm_account = current->mm; |
6b7898eb | 3339 | |
2b188cc1 JA |
3340 | ret = io_allocate_scq_urings(ctx, p); |
3341 | if (ret) | |
3342 | goto err; | |
3343 | ||
7e84e1c7 | 3344 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
3345 | if (ret) |
3346 | goto err; | |
eae071c9 | 3347 | /* always set a rsrc node */ |
47b228ce PB |
3348 | ret = io_rsrc_node_switch_start(ctx); |
3349 | if (ret) | |
3350 | goto err; | |
eae071c9 | 3351 | io_rsrc_node_switch(ctx, NULL); |
2b188cc1 | 3352 | |
2b188cc1 | 3353 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
3354 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3355 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3356 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3357 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3358 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3359 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
3360 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
3361 | |
3362 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
3363 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3364 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3365 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3366 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3367 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3368 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 3369 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 3370 | |
7f13657d XW |
3371 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
3372 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 3373 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 3374 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 3375 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e JA |
3376 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
3377 | IORING_FEAT_LINKED_FILE; | |
7f13657d XW |
3378 | |
3379 | if (copy_to_user(params, p, sizeof(*p))) { | |
3380 | ret = -EFAULT; | |
3381 | goto err; | |
3382 | } | |
d1719f70 | 3383 | |
9faadcc8 PB |
3384 | file = io_uring_get_file(ctx); |
3385 | if (IS_ERR(file)) { | |
3386 | ret = PTR_ERR(file); | |
3387 | goto err; | |
3388 | } | |
3389 | ||
044c1ab3 JA |
3390 | /* |
3391 | * Install ring fd as the very last thing, so we don't risk someone | |
3392 | * having closed it before we finish setup | |
3393 | */ | |
9faadcc8 PB |
3394 | ret = io_uring_install_fd(ctx, file); |
3395 | if (ret < 0) { | |
3396 | /* fput will clean it up */ | |
3397 | fput(file); | |
3398 | return ret; | |
3399 | } | |
044c1ab3 | 3400 | |
c826bd7a | 3401 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
3402 | return ret; |
3403 | err: | |
3404 | io_ring_ctx_wait_and_kill(ctx); | |
3405 | return ret; | |
3406 | } | |
3407 | ||
3408 | /* | |
3409 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3410 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3411 | * params structure passed in. | |
3412 | */ | |
3413 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3414 | { | |
3415 | struct io_uring_params p; | |
2b188cc1 JA |
3416 | int i; |
3417 | ||
3418 | if (copy_from_user(&p, params, sizeof(p))) | |
3419 | return -EFAULT; | |
3420 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3421 | if (p.resv[i]) | |
3422 | return -EINVAL; | |
3423 | } | |
3424 | ||
6c271ce2 | 3425 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 3426 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 3427 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 3428 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 3429 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 PB |
3430 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
3431 | IORING_SETUP_SINGLE_ISSUER)) | |
2b188cc1 JA |
3432 | return -EINVAL; |
3433 | ||
ef060ea9 | 3434 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
3435 | } |
3436 | ||
3437 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3438 | struct io_uring_params __user *, params) | |
3439 | { | |
3440 | return io_uring_setup(entries, params); | |
3441 | } | |
3442 | ||
c072481d PB |
3443 | static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, |
3444 | unsigned nr_args) | |
66f4af93 JA |
3445 | { |
3446 | struct io_uring_probe *p; | |
3447 | size_t size; | |
3448 | int i, ret; | |
3449 | ||
3450 | size = struct_size(p, ops, nr_args); | |
3451 | if (size == SIZE_MAX) | |
3452 | return -EOVERFLOW; | |
3453 | p = kzalloc(size, GFP_KERNEL); | |
3454 | if (!p) | |
3455 | return -ENOMEM; | |
3456 | ||
3457 | ret = -EFAULT; | |
3458 | if (copy_from_user(p, arg, size)) | |
3459 | goto out; | |
3460 | ret = -EINVAL; | |
3461 | if (memchr_inv(p, 0, size)) | |
3462 | goto out; | |
3463 | ||
3464 | p->last_op = IORING_OP_LAST - 1; | |
3465 | if (nr_args > IORING_OP_LAST) | |
3466 | nr_args = IORING_OP_LAST; | |
3467 | ||
3468 | for (i = 0; i < nr_args; i++) { | |
3469 | p->ops[i].op = i; | |
3470 | if (!io_op_defs[i].not_supported) | |
3471 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
3472 | } | |
3473 | p->ops_len = i; | |
3474 | ||
3475 | ret = 0; | |
3476 | if (copy_to_user(arg, p, size)) | |
3477 | ret = -EFAULT; | |
3478 | out: | |
3479 | kfree(p); | |
3480 | return ret; | |
3481 | } | |
3482 | ||
071698e1 JA |
3483 | static int io_register_personality(struct io_ring_ctx *ctx) |
3484 | { | |
4379bf8b | 3485 | const struct cred *creds; |
61cf9370 | 3486 | u32 id; |
1e6fa521 | 3487 | int ret; |
071698e1 | 3488 | |
4379bf8b | 3489 | creds = get_current_cred(); |
1e6fa521 | 3490 | |
61cf9370 MWO |
3491 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
3492 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
3493 | if (ret < 0) { |
3494 | put_cred(creds); | |
3495 | return ret; | |
3496 | } | |
3497 | return id; | |
071698e1 JA |
3498 | } |
3499 | ||
c072481d PB |
3500 | static __cold int io_register_restrictions(struct io_ring_ctx *ctx, |
3501 | void __user *arg, unsigned int nr_args) | |
21b55dbc SG |
3502 | { |
3503 | struct io_uring_restriction *res; | |
3504 | size_t size; | |
3505 | int i, ret; | |
3506 | ||
7e84e1c7 SG |
3507 | /* Restrictions allowed only if rings started disabled */ |
3508 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3509 | return -EBADFD; | |
3510 | ||
21b55dbc | 3511 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 3512 | if (ctx->restrictions.registered) |
21b55dbc SG |
3513 | return -EBUSY; |
3514 | ||
3515 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
3516 | return -EINVAL; | |
3517 | ||
3518 | size = array_size(nr_args, sizeof(*res)); | |
3519 | if (size == SIZE_MAX) | |
3520 | return -EOVERFLOW; | |
3521 | ||
3522 | res = memdup_user(arg, size); | |
3523 | if (IS_ERR(res)) | |
3524 | return PTR_ERR(res); | |
3525 | ||
3526 | ret = 0; | |
3527 | ||
3528 | for (i = 0; i < nr_args; i++) { | |
3529 | switch (res[i].opcode) { | |
3530 | case IORING_RESTRICTION_REGISTER_OP: | |
3531 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
3532 | ret = -EINVAL; | |
3533 | goto out; | |
3534 | } | |
3535 | ||
3536 | __set_bit(res[i].register_op, | |
3537 | ctx->restrictions.register_op); | |
3538 | break; | |
3539 | case IORING_RESTRICTION_SQE_OP: | |
3540 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
3541 | ret = -EINVAL; | |
3542 | goto out; | |
3543 | } | |
3544 | ||
3545 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
3546 | break; | |
3547 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
3548 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
3549 | break; | |
3550 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
3551 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
3552 | break; | |
3553 | default: | |
3554 | ret = -EINVAL; | |
3555 | goto out; | |
3556 | } | |
3557 | } | |
3558 | ||
3559 | out: | |
3560 | /* Reset all restrictions if an error happened */ | |
3561 | if (ret != 0) | |
3562 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
3563 | else | |
7e84e1c7 | 3564 | ctx->restrictions.registered = true; |
21b55dbc SG |
3565 | |
3566 | kfree(res); | |
3567 | return ret; | |
3568 | } | |
3569 | ||
7e84e1c7 SG |
3570 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
3571 | { | |
3572 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3573 | return -EBADFD; | |
3574 | ||
3575 | if (ctx->restrictions.registered) | |
3576 | ctx->restricted = 1; | |
3577 | ||
0298ef96 PB |
3578 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
3579 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
3580 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
3581 | return 0; |
3582 | } | |
3583 | ||
c072481d PB |
3584 | static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, |
3585 | void __user *arg, unsigned len) | |
fe76421d JA |
3586 | { |
3587 | struct io_uring_task *tctx = current->io_uring; | |
3588 | cpumask_var_t new_mask; | |
3589 | int ret; | |
3590 | ||
3591 | if (!tctx || !tctx->io_wq) | |
3592 | return -EINVAL; | |
3593 | ||
3594 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
3595 | return -ENOMEM; | |
3596 | ||
3597 | cpumask_clear(new_mask); | |
3598 | if (len > cpumask_size()) | |
3599 | len = cpumask_size(); | |
3600 | ||
0f5e4b83 ES |
3601 | if (in_compat_syscall()) { |
3602 | ret = compat_get_bitmap(cpumask_bits(new_mask), | |
3603 | (const compat_ulong_t __user *)arg, | |
3604 | len * 8 /* CHAR_BIT */); | |
3605 | } else { | |
3606 | ret = copy_from_user(new_mask, arg, len); | |
3607 | } | |
3608 | ||
3609 | if (ret) { | |
fe76421d JA |
3610 | free_cpumask_var(new_mask); |
3611 | return -EFAULT; | |
3612 | } | |
3613 | ||
3614 | ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); | |
3615 | free_cpumask_var(new_mask); | |
3616 | return ret; | |
3617 | } | |
3618 | ||
c072481d | 3619 | static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) |
fe76421d JA |
3620 | { |
3621 | struct io_uring_task *tctx = current->io_uring; | |
3622 | ||
3623 | if (!tctx || !tctx->io_wq) | |
3624 | return -EINVAL; | |
3625 | ||
3626 | return io_wq_cpu_affinity(tctx->io_wq, NULL); | |
3627 | } | |
3628 | ||
c072481d PB |
3629 | static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, |
3630 | void __user *arg) | |
b22fa62a | 3631 | __must_hold(&ctx->uring_lock) |
2e480058 | 3632 | { |
b22fa62a | 3633 | struct io_tctx_node *node; |
fa84693b JA |
3634 | struct io_uring_task *tctx = NULL; |
3635 | struct io_sq_data *sqd = NULL; | |
2e480058 JA |
3636 | __u32 new_count[2]; |
3637 | int i, ret; | |
3638 | ||
2e480058 JA |
3639 | if (copy_from_user(new_count, arg, sizeof(new_count))) |
3640 | return -EFAULT; | |
3641 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
3642 | if (new_count[i] > INT_MAX) | |
3643 | return -EINVAL; | |
3644 | ||
fa84693b JA |
3645 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3646 | sqd = ctx->sq_data; | |
3647 | if (sqd) { | |
009ad9f0 JA |
3648 | /* |
3649 | * Observe the correct sqd->lock -> ctx->uring_lock | |
3650 | * ordering. Fine to drop uring_lock here, we hold | |
3651 | * a ref to the ctx. | |
3652 | */ | |
41d3a6bd | 3653 | refcount_inc(&sqd->refs); |
009ad9f0 | 3654 | mutex_unlock(&ctx->uring_lock); |
fa84693b | 3655 | mutex_lock(&sqd->lock); |
009ad9f0 | 3656 | mutex_lock(&ctx->uring_lock); |
41d3a6bd JA |
3657 | if (sqd->thread) |
3658 | tctx = sqd->thread->io_uring; | |
fa84693b JA |
3659 | } |
3660 | } else { | |
3661 | tctx = current->io_uring; | |
3662 | } | |
3663 | ||
e139a1ec | 3664 | BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); |
fa84693b | 3665 | |
bad119b9 PB |
3666 | for (i = 0; i < ARRAY_SIZE(new_count); i++) |
3667 | if (new_count[i]) | |
3668 | ctx->iowq_limits[i] = new_count[i]; | |
e139a1ec PB |
3669 | ctx->iowq_limits_set = true; |
3670 | ||
e139a1ec PB |
3671 | if (tctx && tctx->io_wq) { |
3672 | ret = io_wq_max_workers(tctx->io_wq, new_count); | |
3673 | if (ret) | |
3674 | goto err; | |
3675 | } else { | |
3676 | memset(new_count, 0, sizeof(new_count)); | |
3677 | } | |
fa84693b | 3678 | |
41d3a6bd | 3679 | if (sqd) { |
fa84693b | 3680 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
3681 | io_put_sq_data(sqd); |
3682 | } | |
2e480058 JA |
3683 | |
3684 | if (copy_to_user(arg, new_count, sizeof(new_count))) | |
3685 | return -EFAULT; | |
3686 | ||
b22fa62a PB |
3687 | /* that's it for SQPOLL, only the SQPOLL task creates requests */ |
3688 | if (sqd) | |
3689 | return 0; | |
3690 | ||
3691 | /* now propagate the restriction to all registered users */ | |
3692 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3693 | struct io_uring_task *tctx = node->task->io_uring; | |
3694 | ||
3695 | if (WARN_ON_ONCE(!tctx->io_wq)) | |
3696 | continue; | |
3697 | ||
3698 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
3699 | new_count[i] = ctx->iowq_limits[i]; | |
3700 | /* ignore errors, it always returns zero anyway */ | |
3701 | (void)io_wq_max_workers(tctx->io_wq, new_count); | |
3702 | } | |
2e480058 | 3703 | return 0; |
fa84693b | 3704 | err: |
41d3a6bd | 3705 | if (sqd) { |
fa84693b | 3706 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
3707 | io_put_sq_data(sqd); |
3708 | } | |
fa84693b | 3709 | return ret; |
2e480058 JA |
3710 | } |
3711 | ||
edafccee JA |
3712 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
3713 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
3714 | __releases(ctx->uring_lock) |
3715 | __acquires(ctx->uring_lock) | |
edafccee JA |
3716 | { |
3717 | int ret; | |
3718 | ||
35fa71a0 JA |
3719 | /* |
3720 | * We're inside the ring mutex, if the ref is already dying, then | |
3721 | * someone else killed the ctx or is already going through | |
3722 | * io_uring_register(). | |
3723 | */ | |
3724 | if (percpu_ref_is_dying(&ctx->refs)) | |
3725 | return -ENXIO; | |
3726 | ||
75c4021a PB |
3727 | if (ctx->restricted) { |
3728 | if (opcode >= IORING_REGISTER_LAST) | |
3729 | return -EINVAL; | |
3730 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); | |
3731 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
3732 | return -EACCES; | |
3733 | } | |
3734 | ||
edafccee JA |
3735 | switch (opcode) { |
3736 | case IORING_REGISTER_BUFFERS: | |
0184f08e PB |
3737 | ret = -EFAULT; |
3738 | if (!arg) | |
3739 | break; | |
634d00df | 3740 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
3741 | break; |
3742 | case IORING_UNREGISTER_BUFFERS: | |
3743 | ret = -EINVAL; | |
3744 | if (arg || nr_args) | |
3745 | break; | |
0a96bbe4 | 3746 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 3747 | break; |
6b06314c | 3748 | case IORING_REGISTER_FILES: |
a8da73a3 JA |
3749 | ret = -EFAULT; |
3750 | if (!arg) | |
3751 | break; | |
792e3582 | 3752 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
3753 | break; |
3754 | case IORING_UNREGISTER_FILES: | |
3755 | ret = -EINVAL; | |
3756 | if (arg || nr_args) | |
3757 | break; | |
3758 | ret = io_sqe_files_unregister(ctx); | |
3759 | break; | |
c3a31e60 | 3760 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 3761 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 3762 | break; |
9b402849 JA |
3763 | case IORING_REGISTER_EVENTFD: |
3764 | ret = -EINVAL; | |
3765 | if (nr_args != 1) | |
3766 | break; | |
c75312dd UA |
3767 | ret = io_eventfd_register(ctx, arg, 0); |
3768 | break; | |
3769 | case IORING_REGISTER_EVENTFD_ASYNC: | |
3770 | ret = -EINVAL; | |
3771 | if (nr_args != 1) | |
f2842ab5 | 3772 | break; |
c75312dd | 3773 | ret = io_eventfd_register(ctx, arg, 1); |
9b402849 JA |
3774 | break; |
3775 | case IORING_UNREGISTER_EVENTFD: | |
3776 | ret = -EINVAL; | |
3777 | if (arg || nr_args) | |
3778 | break; | |
3779 | ret = io_eventfd_unregister(ctx); | |
3780 | break; | |
66f4af93 JA |
3781 | case IORING_REGISTER_PROBE: |
3782 | ret = -EINVAL; | |
3783 | if (!arg || nr_args > 256) | |
3784 | break; | |
3785 | ret = io_probe(ctx, arg, nr_args); | |
3786 | break; | |
071698e1 JA |
3787 | case IORING_REGISTER_PERSONALITY: |
3788 | ret = -EINVAL; | |
3789 | if (arg || nr_args) | |
3790 | break; | |
3791 | ret = io_register_personality(ctx); | |
3792 | break; | |
3793 | case IORING_UNREGISTER_PERSONALITY: | |
3794 | ret = -EINVAL; | |
3795 | if (arg) | |
3796 | break; | |
3797 | ret = io_unregister_personality(ctx, nr_args); | |
3798 | break; | |
7e84e1c7 SG |
3799 | case IORING_REGISTER_ENABLE_RINGS: |
3800 | ret = -EINVAL; | |
3801 | if (arg || nr_args) | |
3802 | break; | |
3803 | ret = io_register_enable_rings(ctx); | |
3804 | break; | |
21b55dbc SG |
3805 | case IORING_REGISTER_RESTRICTIONS: |
3806 | ret = io_register_restrictions(ctx, arg, nr_args); | |
3807 | break; | |
992da01a PB |
3808 | case IORING_REGISTER_FILES2: |
3809 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
3810 | break; | |
3811 | case IORING_REGISTER_FILES_UPDATE2: | |
3812 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
3813 | IORING_RSRC_FILE); | |
3814 | break; | |
3815 | case IORING_REGISTER_BUFFERS2: | |
3816 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 3817 | break; |
992da01a PB |
3818 | case IORING_REGISTER_BUFFERS_UPDATE: |
3819 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
3820 | IORING_RSRC_BUFFER); | |
c3bdad02 | 3821 | break; |
fe76421d JA |
3822 | case IORING_REGISTER_IOWQ_AFF: |
3823 | ret = -EINVAL; | |
3824 | if (!arg || !nr_args) | |
3825 | break; | |
3826 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
3827 | break; | |
3828 | case IORING_UNREGISTER_IOWQ_AFF: | |
3829 | ret = -EINVAL; | |
3830 | if (arg || nr_args) | |
3831 | break; | |
3832 | ret = io_unregister_iowq_aff(ctx); | |
3833 | break; | |
2e480058 JA |
3834 | case IORING_REGISTER_IOWQ_MAX_WORKERS: |
3835 | ret = -EINVAL; | |
3836 | if (!arg || nr_args != 2) | |
3837 | break; | |
3838 | ret = io_register_iowq_max_workers(ctx, arg); | |
3839 | break; | |
e7a6c00d JA |
3840 | case IORING_REGISTER_RING_FDS: |
3841 | ret = io_ringfd_register(ctx, arg, nr_args); | |
3842 | break; | |
3843 | case IORING_UNREGISTER_RING_FDS: | |
3844 | ret = io_ringfd_unregister(ctx, arg, nr_args); | |
3845 | break; | |
c7fb1942 JA |
3846 | case IORING_REGISTER_PBUF_RING: |
3847 | ret = -EINVAL; | |
3848 | if (!arg || nr_args != 1) | |
3849 | break; | |
3850 | ret = io_register_pbuf_ring(ctx, arg); | |
3851 | break; | |
3852 | case IORING_UNREGISTER_PBUF_RING: | |
3853 | ret = -EINVAL; | |
3854 | if (!arg || nr_args != 1) | |
3855 | break; | |
3856 | ret = io_unregister_pbuf_ring(ctx, arg); | |
3857 | break; | |
edafccee JA |
3858 | default: |
3859 | ret = -EINVAL; | |
3860 | break; | |
3861 | } | |
3862 | ||
edafccee JA |
3863 | return ret; |
3864 | } | |
3865 | ||
3866 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
3867 | void __user *, arg, unsigned int, nr_args) | |
3868 | { | |
3869 | struct io_ring_ctx *ctx; | |
3870 | long ret = -EBADF; | |
3871 | struct fd f; | |
3872 | ||
3873 | f = fdget(fd); | |
3874 | if (!f.file) | |
3875 | return -EBADF; | |
3876 | ||
3877 | ret = -EOPNOTSUPP; | |
e5550a14 | 3878 | if (!io_is_uring_fops(f.file)) |
edafccee JA |
3879 | goto out_fput; |
3880 | ||
3881 | ctx = f.file->private_data; | |
3882 | ||
b6c23dd5 PB |
3883 | io_run_task_work(); |
3884 | ||
edafccee JA |
3885 | mutex_lock(&ctx->uring_lock); |
3886 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
3887 | mutex_unlock(&ctx->uring_lock); | |
2757be22 | 3888 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); |
edafccee JA |
3889 | out_fput: |
3890 | fdput(f); | |
3891 | return ret; | |
3892 | } | |
3893 | ||
2b188cc1 JA |
3894 | static int __init io_uring_init(void) |
3895 | { | |
d7f62e82 SM |
3896 | #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \ |
3897 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ | |
3898 | BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \ | |
3899 | } while (0) | |
3900 | ||
3901 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
3902 | __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename) | |
3903 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); | |
3904 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
3905 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
3906 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
3907 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
3908 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
3909 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
3910 | BUILD_BUG_SQE_ELEM(16, __u64, addr); | |
7d67af2c | 3911 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
3912 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
3913 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
3914 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
3915 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
3916 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
3917 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
3918 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
3919 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
3920 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
3921 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
3922 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
3923 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
3924 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
3925 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
3926 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 3927 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
d7f62e82 SM |
3928 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
3929 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 3930 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 3931 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 3932 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 3933 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
e9621e2b | 3934 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
d7f62e82 | 3935 | |
b0d658ec PB |
3936 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
3937 | sizeof(struct io_uring_rsrc_update)); | |
3938 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
3939 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
3940 | |
3941 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
3942 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
3943 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
3944 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 3945 | |
b0d658ec PB |
3946 | /* should fit into one byte */ |
3947 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
3948 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
3949 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 3950 | |
32c2d33e | 3951 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); |
16340eab | 3952 | |
3a4b89a2 JA |
3953 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
3954 | ||
d9b57aa3 | 3955 | io_uring_optable_init(); |
0702e536 | 3956 | |
91f245d5 JA |
3957 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | |
3958 | SLAB_ACCOUNT); | |
2b188cc1 JA |
3959 | return 0; |
3960 | }; | |
3961 | __initcall(io_uring_init); |