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" |
43e0bbbd | 92 | #include "net.h" |
eb42cebb | 93 | #include "notif.h" |
e27f928e | 94 | |
59915143 | 95 | #include "timeout.h" |
329061d3 | 96 | #include "poll.h" |
9b797a37 | 97 | #include "alloc_cache.h" |
5e2a18d9 | 98 | |
5277deaa | 99 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 100 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 | 101 | |
21b55dbc SG |
102 | #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \ |
103 | IORING_REGISTER_LAST + IORING_OP_LAST) | |
2b188cc1 | 104 | |
68fe256a PB |
105 | #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \ |
106 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) | |
107 | ||
5562a8d7 PB |
108 | #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \ |
109 | IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS) | |
68fe256a | 110 | |
c854357b | 111 | #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \ |
9cae36a0 JA |
112 | REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \ |
113 | REQ_F_ASYNC_DATA) | |
b16fed66 | 114 | |
a538be5b PB |
115 | #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\ |
116 | IO_REQ_CLEAN_FLAGS) | |
117 | ||
09899b19 PB |
118 | #define IO_TCTX_REFS_CACHE_NR (1U << 10) |
119 | ||
6dd0be1e | 120 | #define IO_COMPL_BATCH 32 |
bf019da7 | 121 | #define IO_REQ_ALLOC_BATCH 8 |
258b29a9 | 122 | |
10988a0a DY |
123 | enum { |
124 | IO_CHECK_CQ_OVERFLOW_BIT, | |
155bc950 | 125 | IO_CHECK_CQ_DROPPED_BIT, |
10988a0a DY |
126 | }; |
127 | ||
27dc8338 PB |
128 | struct io_defer_entry { |
129 | struct list_head list; | |
130 | struct io_kiocb *req; | |
9cf7c104 | 131 | u32 seq; |
2b188cc1 JA |
132 | }; |
133 | ||
0756a869 PB |
134 | /* requests with any of those set should undergo io_disarm_next() */ |
135 | #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL) | |
da1a08c5 | 136 | #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK) |
0756a869 | 137 | |
affa87db | 138 | static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
9936c7c2 | 139 | struct task_struct *task, |
3dd0c97a | 140 | bool cancel_all); |
1ffc5422 | 141 | |
c7dae4ba | 142 | static void io_dismantle_req(struct io_kiocb *req); |
68fb8979 | 143 | static void io_clean_op(struct io_kiocb *req); |
cbc2e203 | 144 | static void io_queue_sqe(struct io_kiocb *req); |
de0617e4 | 145 | |
c450178d | 146 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx); |
de0617e4 | 147 | |
2b188cc1 JA |
148 | static struct kmem_cache *req_cachep; |
149 | ||
2b188cc1 JA |
150 | struct sock *io_uring_get_socket(struct file *file) |
151 | { | |
152 | #if defined(CONFIG_UNIX) | |
cd40cae2 | 153 | if (io_is_uring_fops(file)) { |
2b188cc1 JA |
154 | struct io_ring_ctx *ctx = file->private_data; |
155 | ||
156 | return ctx->ring_sock->sk; | |
157 | } | |
158 | #endif | |
159 | return NULL; | |
160 | } | |
161 | EXPORT_SYMBOL(io_uring_get_socket); | |
162 | ||
c450178d PB |
163 | static inline void io_submit_flush_completions(struct io_ring_ctx *ctx) |
164 | { | |
6f33b0bc | 165 | if (!wq_list_empty(&ctx->submit_state.compl_reqs)) |
c450178d PB |
166 | __io_submit_flush_completions(ctx); |
167 | } | |
168 | ||
faf88dde PB |
169 | static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx) |
170 | { | |
171 | return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head); | |
172 | } | |
173 | ||
9cae36a0 JA |
174 | static bool io_match_linked(struct io_kiocb *head) |
175 | { | |
176 | struct io_kiocb *req; | |
177 | ||
178 | io_for_each_link(req, head) { | |
179 | if (req->flags & REQ_F_INFLIGHT) | |
180 | return true; | |
181 | } | |
182 | return false; | |
6af3f48b PB |
183 | } |
184 | ||
185 | /* | |
186 | * As io_match_task() but protected against racing with linked timeouts. | |
187 | * User must not hold timeout_lock. | |
188 | */ | |
329061d3 JA |
189 | bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, |
190 | bool cancel_all) | |
6af3f48b | 191 | { |
9cae36a0 JA |
192 | bool matched; |
193 | ||
6af3f48b PB |
194 | if (task && head->task != task) |
195 | return false; | |
9cae36a0 JA |
196 | if (cancel_all) |
197 | return true; | |
198 | ||
199 | if (head->flags & REQ_F_LINK_TIMEOUT) { | |
200 | struct io_ring_ctx *ctx = head->ctx; | |
201 | ||
202 | /* protect against races with linked timeouts */ | |
203 | spin_lock_irq(&ctx->timeout_lock); | |
204 | matched = io_match_linked(head); | |
205 | spin_unlock_irq(&ctx->timeout_lock); | |
206 | } else { | |
207 | matched = io_match_linked(head); | |
208 | } | |
209 | return matched; | |
6af3f48b PB |
210 | } |
211 | ||
a8295b98 HX |
212 | static inline void req_fail_link_node(struct io_kiocb *req, int res) |
213 | { | |
214 | req_set_fail(req); | |
97b388d7 | 215 | io_req_set_res(req, res, 0); |
a8295b98 HX |
216 | } |
217 | ||
fa05457a PB |
218 | static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx) |
219 | { | |
220 | wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list); | |
a8295b98 HX |
221 | } |
222 | ||
c072481d | 223 | static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref) |
2b188cc1 JA |
224 | { |
225 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
226 | ||
0f158b4c | 227 | complete(&ctx->ref_comp); |
2b188cc1 JA |
228 | } |
229 | ||
c072481d | 230 | static __cold void io_fallback_req_func(struct work_struct *work) |
f56165e6 PB |
231 | { |
232 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, | |
233 | fallback_work.work); | |
234 | struct llist_node *node = llist_del_all(&ctx->fallback_llist); | |
235 | struct io_kiocb *req, *tmp; | |
f237c30a | 236 | bool locked = false; |
f56165e6 PB |
237 | |
238 | percpu_ref_get(&ctx->refs); | |
3218e5d3 | 239 | llist_for_each_entry_safe(req, tmp, node, io_task_work.node) |
f237c30a | 240 | req->io_task_work.func(req, &locked); |
5636c00d | 241 | |
f237c30a | 242 | if (locked) { |
c450178d | 243 | io_submit_flush_completions(ctx); |
f237c30a PB |
244 | mutex_unlock(&ctx->uring_lock); |
245 | } | |
f56165e6 PB |
246 | percpu_ref_put(&ctx->refs); |
247 | } | |
248 | ||
e6f89be6 PB |
249 | static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits) |
250 | { | |
251 | unsigned hash_buckets = 1U << bits; | |
252 | size_t hash_size = hash_buckets * sizeof(table->hbs[0]); | |
253 | ||
254 | table->hbs = kmalloc(hash_size, GFP_KERNEL); | |
255 | if (!table->hbs) | |
256 | return -ENOMEM; | |
257 | ||
258 | table->hash_bits = bits; | |
259 | init_hash_table(table, hash_buckets); | |
260 | return 0; | |
261 | } | |
262 | ||
c072481d | 263 | static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
2b188cc1 JA |
264 | { |
265 | struct io_ring_ctx *ctx; | |
9cfc7e94 | 266 | int hash_bits; |
2b188cc1 JA |
267 | |
268 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
269 | if (!ctx) | |
270 | return NULL; | |
271 | ||
9cfc7e94 JA |
272 | xa_init(&ctx->io_bl_xa); |
273 | ||
78076bb6 JA |
274 | /* |
275 | * Use 5 bits less than the max cq entries, that should give us around | |
4a07723f PB |
276 | * 32 entries per hash list if totally full and uniformly spread, but |
277 | * don't keep too many buckets to not overconsume memory. | |
78076bb6 | 278 | */ |
4a07723f PB |
279 | hash_bits = ilog2(p->cq_entries) - 5; |
280 | hash_bits = clamp(hash_bits, 1, 8); | |
e6f89be6 | 281 | if (io_alloc_hash_table(&ctx->cancel_table, hash_bits)) |
78076bb6 | 282 | goto err; |
9ca9fb24 PB |
283 | if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits)) |
284 | goto err; | |
38513c46 | 285 | |
6224843d PB |
286 | ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL); |
287 | if (!ctx->dummy_ubuf) | |
288 | goto err; | |
289 | /* set invalid range, so io_import_fixed() fails meeting it */ | |
290 | ctx->dummy_ubuf->ubuf = -1UL; | |
291 | ||
21482896 | 292 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
48904229 | 293 | 0, GFP_KERNEL)) |
206aefde | 294 | goto err; |
2b188cc1 JA |
295 | |
296 | ctx->flags = p->flags; | |
90554200 | 297 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 298 | INIT_LIST_HEAD(&ctx->sqd_list); |
1d7bb1d5 | 299 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
cc3cec83 | 300 | INIT_LIST_HEAD(&ctx->io_buffers_cache); |
9b797a37 | 301 | io_alloc_cache_init(&ctx->apoll_cache); |
43e0bbbd | 302 | io_alloc_cache_init(&ctx->netmsg_cache); |
0f158b4c | 303 | init_completion(&ctx->ref_comp); |
61cf9370 | 304 | xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1); |
2b188cc1 | 305 | mutex_init(&ctx->uring_lock); |
311997b3 | 306 | init_waitqueue_head(&ctx->cq_wait); |
2b188cc1 | 307 | spin_lock_init(&ctx->completion_lock); |
89850fce | 308 | spin_lock_init(&ctx->timeout_lock); |
5eef4e87 | 309 | INIT_WQ_LIST(&ctx->iopoll_list); |
cc3cec83 JA |
310 | INIT_LIST_HEAD(&ctx->io_buffers_pages); |
311 | INIT_LIST_HEAD(&ctx->io_buffers_comp); | |
de0617e4 | 312 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 313 | INIT_LIST_HEAD(&ctx->timeout_list); |
ef9dd637 | 314 | INIT_LIST_HEAD(&ctx->ltimeout_list); |
d67d2263 BM |
315 | spin_lock_init(&ctx->rsrc_ref_lock); |
316 | INIT_LIST_HEAD(&ctx->rsrc_ref_list); | |
269bbe5f BM |
317 | INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work); |
318 | init_llist_head(&ctx->rsrc_put_llist); | |
13bf43f5 | 319 | INIT_LIST_HEAD(&ctx->tctx_list); |
c2b6c6bc PB |
320 | ctx->submit_state.free_list.next = NULL; |
321 | INIT_WQ_LIST(&ctx->locked_free_list); | |
9011bf9a | 322 | INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func); |
6f33b0bc | 323 | INIT_WQ_LIST(&ctx->submit_state.compl_reqs); |
2b188cc1 | 324 | return ctx; |
206aefde | 325 | err: |
6224843d | 326 | kfree(ctx->dummy_ubuf); |
e6f89be6 | 327 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 328 | kfree(ctx->cancel_table_locked.hbs); |
9cfc7e94 JA |
329 | kfree(ctx->io_bl); |
330 | xa_destroy(&ctx->io_bl_xa); | |
206aefde JA |
331 | kfree(ctx); |
332 | return NULL; | |
2b188cc1 JA |
333 | } |
334 | ||
8f6ed49a PB |
335 | static void io_account_cq_overflow(struct io_ring_ctx *ctx) |
336 | { | |
337 | struct io_rings *r = ctx->rings; | |
338 | ||
339 | WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1); | |
340 | ctx->cq_extra--; | |
341 | } | |
342 | ||
9cf7c104 | 343 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 344 | { |
2bc9930e JA |
345 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
346 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 347 | |
8f6ed49a | 348 | return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail; |
2bc9930e | 349 | } |
de0617e4 | 350 | |
9d858b21 | 351 | return false; |
de0617e4 JA |
352 | } |
353 | ||
9cae36a0 JA |
354 | static inline void io_req_track_inflight(struct io_kiocb *req) |
355 | { | |
356 | if (!(req->flags & REQ_F_INFLIGHT)) { | |
357 | req->flags |= REQ_F_INFLIGHT; | |
386e4fb6 | 358 | atomic_inc(&req->task->io_uring->inflight_tracked); |
9cae36a0 JA |
359 | } |
360 | } | |
361 | ||
fd08e530 PB |
362 | static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req) |
363 | { | |
906c6caa PB |
364 | if (WARN_ON_ONCE(!req->link)) |
365 | return NULL; | |
366 | ||
4d13d1a4 PB |
367 | req->flags &= ~REQ_F_ARM_LTIMEOUT; |
368 | req->flags |= REQ_F_LINK_TIMEOUT; | |
fd08e530 PB |
369 | |
370 | /* linked timeouts should have two refs once prep'ed */ | |
48dcd38d | 371 | io_req_set_refcount(req); |
4d13d1a4 PB |
372 | __io_req_set_refcount(req->link, 2); |
373 | return req->link; | |
fd08e530 PB |
374 | } |
375 | ||
376 | static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) | |
377 | { | |
4d13d1a4 | 378 | if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT))) |
fd08e530 PB |
379 | return NULL; |
380 | return __io_prep_linked_timeout(req); | |
381 | } | |
382 | ||
cb2d344c PB |
383 | static noinline void __io_arm_ltimeout(struct io_kiocb *req) |
384 | { | |
385 | io_queue_linked_timeout(__io_prep_linked_timeout(req)); | |
386 | } | |
387 | ||
388 | static inline void io_arm_ltimeout(struct io_kiocb *req) | |
389 | { | |
390 | if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT)) | |
391 | __io_arm_ltimeout(req); | |
392 | } | |
393 | ||
1e6fa521 JA |
394 | static void io_prep_async_work(struct io_kiocb *req) |
395 | { | |
396 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
1e6fa521 JA |
397 | struct io_ring_ctx *ctx = req->ctx; |
398 | ||
b8e64b53 PB |
399 | if (!(req->flags & REQ_F_CREDS)) { |
400 | req->flags |= REQ_F_CREDS; | |
c10d1f98 | 401 | req->creds = get_current_cred(); |
b8e64b53 | 402 | } |
003e8dcc | 403 | |
e1d675df PB |
404 | req->work.list.next = NULL; |
405 | req->work.flags = 0; | |
8e29da69 | 406 | req->work.cancel_seq = atomic_read(&ctx->cancel_seq); |
feaadc4f PB |
407 | if (req->flags & REQ_F_FORCE_ASYNC) |
408 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
409 | ||
f6b543fd JA |
410 | if (req->file && !io_req_ffs_set(req)) |
411 | req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT; | |
412 | ||
1e6fa521 JA |
413 | if (req->flags & REQ_F_ISREG) { |
414 | if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL)) | |
415 | io_wq_hash_work(&req->work, file_inode(req->file)); | |
4b982bd0 | 416 | } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) { |
1e6fa521 JA |
417 | if (def->unbound_nonreg_file) |
418 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
419 | } | |
561fb04a | 420 | } |
cccf0ee8 | 421 | |
cbdcb435 | 422 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 423 | { |
cbdcb435 | 424 | struct io_kiocb *cur; |
54a91f3b | 425 | |
44eff40a PB |
426 | if (req->flags & REQ_F_LINK_TIMEOUT) { |
427 | struct io_ring_ctx *ctx = req->ctx; | |
428 | ||
674ee8e1 | 429 | spin_lock_irq(&ctx->timeout_lock); |
44eff40a PB |
430 | io_for_each_link(cur, req) |
431 | io_prep_async_work(cur); | |
674ee8e1 | 432 | spin_unlock_irq(&ctx->timeout_lock); |
44eff40a PB |
433 | } else { |
434 | io_for_each_link(cur, req) | |
435 | io_prep_async_work(cur); | |
436 | } | |
561fb04a JA |
437 | } |
438 | ||
f3b44f92 | 439 | void io_queue_iowq(struct io_kiocb *req, bool *dont_use) |
561fb04a | 440 | { |
cbdcb435 | 441 | struct io_kiocb *link = io_prep_linked_timeout(req); |
5aa75ed5 | 442 | struct io_uring_task *tctx = req->task->io_uring; |
561fb04a | 443 | |
3bfe6106 JA |
444 | BUG_ON(!tctx); |
445 | BUG_ON(!tctx->io_wq); | |
561fb04a | 446 | |
cbdcb435 PB |
447 | /* init ->work of the whole link before punting */ |
448 | io_prep_async_link(req); | |
991468dc JA |
449 | |
450 | /* | |
451 | * Not expected to happen, but if we do have a bug where this _can_ | |
452 | * happen, catch it here and ensure the request is marked as | |
453 | * canceled. That will make io-wq go through the usual work cancel | |
454 | * procedure rather than attempt to run this request (or create a new | |
455 | * worker for it). | |
456 | */ | |
457 | if (WARN_ON_ONCE(!same_thread_group(req->task, current))) | |
458 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
459 | ||
48863ffd | 460 | trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work)); |
ebf93667 | 461 | io_wq_enqueue(tctx->io_wq, &req->work); |
7271ef3a JA |
462 | if (link) |
463 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
464 | } |
465 | ||
c072481d | 466 | static __cold void io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 467 | { |
441b8a78 | 468 | while (!list_empty(&ctx->defer_list)) { |
27dc8338 PB |
469 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
470 | struct io_defer_entry, list); | |
de0617e4 | 471 | |
9cf7c104 | 472 | if (req_need_defer(de->req, de->seq)) |
04518945 | 473 | break; |
27dc8338 | 474 | list_del_init(&de->list); |
907d1df3 | 475 | io_req_task_queue(de->req); |
27dc8338 | 476 | kfree(de); |
441b8a78 | 477 | } |
04518945 PB |
478 | } |
479 | ||
77bc59b4 | 480 | static void io_eventfd_signal(struct io_ring_ctx *ctx) |
f2842ab5 | 481 | { |
77bc59b4 | 482 | struct io_ev_fd *ev_fd; |
305bef98 PB |
483 | bool skip; |
484 | ||
485 | spin_lock(&ctx->completion_lock); | |
486 | /* | |
487 | * Eventfd should only get triggered when at least one event has been | |
488 | * posted. Some applications rely on the eventfd notification count only | |
489 | * changing IFF a new CQE has been added to the CQ ring. There's no | |
490 | * depedency on 1:1 relationship between how many times this function is | |
491 | * called (and hence the eventfd count) and number of CQEs posted to the | |
492 | * CQ ring. | |
493 | */ | |
494 | skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail; | |
495 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
496 | spin_unlock(&ctx->completion_lock); | |
497 | if (skip) | |
498 | return; | |
77bc59b4 | 499 | |
77bc59b4 UA |
500 | rcu_read_lock(); |
501 | /* | |
502 | * rcu_dereference ctx->io_ev_fd once and use it for both for checking | |
503 | * and eventfd_signal | |
504 | */ | |
505 | ev_fd = rcu_dereference(ctx->io_ev_fd); | |
506 | ||
507 | /* | |
508 | * Check again if ev_fd exists incase an io_eventfd_unregister call | |
509 | * completed between the NULL check of ctx->io_ev_fd at the start of | |
510 | * the function and rcu_read_lock. | |
511 | */ | |
512 | if (unlikely(!ev_fd)) | |
513 | goto out; | |
7e55a19c | 514 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
77bc59b4 UA |
515 | goto out; |
516 | ||
c75312dd | 517 | if (!ev_fd->eventfd_async || io_wq_current_is_worker()) |
77bc59b4 | 518 | eventfd_signal(ev_fd->cq_ev_fd, 1); |
77bc59b4 UA |
519 | out: |
520 | rcu_read_unlock(); | |
f2842ab5 JA |
521 | } |
522 | ||
a830ffd2 PB |
523 | void __io_commit_cqring_flush(struct io_ring_ctx *ctx) |
524 | { | |
525 | if (ctx->off_timeout_used || ctx->drain_active) { | |
526 | spin_lock(&ctx->completion_lock); | |
527 | if (ctx->off_timeout_used) | |
528 | io_flush_timeouts(ctx); | |
529 | if (ctx->drain_active) | |
530 | io_queue_deferred(ctx); | |
531 | spin_unlock(&ctx->completion_lock); | |
532 | } | |
533 | if (ctx->has_evfd) | |
534 | io_eventfd_signal(ctx); | |
535 | } | |
536 | ||
25399321 | 537 | static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 | 538 | { |
46929b08 | 539 | io_commit_cqring_flush(ctx); |
9aa8dfde | 540 | io_cqring_wake(ctx); |
1d7bb1d5 JA |
541 | } |
542 | ||
25399321 PB |
543 | static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx) |
544 | __releases(ctx->completion_lock) | |
545 | { | |
546 | io_commit_cqring(ctx); | |
547 | spin_unlock(&ctx->completion_lock); | |
548 | io_cqring_ev_posted(ctx); | |
549 | } | |
550 | ||
551 | void io_cq_unlock_post(struct io_ring_ctx *ctx) | |
552 | { | |
553 | __io_cq_unlock_post(ctx); | |
554 | } | |
555 | ||
c4a2ed72 | 556 | /* Returns true if there are no backlogged entries after the flush */ |
6c2450ae | 557 | static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) |
1d7bb1d5 | 558 | { |
305bef98 | 559 | bool all_flushed; |
e45a3e05 | 560 | size_t cqe_size = sizeof(struct io_uring_cqe); |
1d7bb1d5 | 561 | |
a566c556 | 562 | if (!force && __io_cqring_events(ctx) == ctx->cq_entries) |
e23de15f | 563 | return false; |
1d7bb1d5 | 564 | |
e45a3e05 SR |
565 | if (ctx->flags & IORING_SETUP_CQE32) |
566 | cqe_size <<= 1; | |
567 | ||
25399321 | 568 | io_cq_lock(ctx); |
6c2450ae | 569 | while (!list_empty(&ctx->cq_overflow_list)) { |
d068b506 | 570 | struct io_uring_cqe *cqe = io_get_cqe(ctx); |
6c2450ae | 571 | struct io_overflow_cqe *ocqe; |
e6c8aa9a | 572 | |
1d7bb1d5 JA |
573 | if (!cqe && !force) |
574 | break; | |
6c2450ae PB |
575 | ocqe = list_first_entry(&ctx->cq_overflow_list, |
576 | struct io_overflow_cqe, list); | |
577 | if (cqe) | |
e45a3e05 | 578 | memcpy(cqe, &ocqe->cqe, cqe_size); |
6c2450ae | 579 | else |
8f6ed49a PB |
580 | io_account_cq_overflow(ctx); |
581 | ||
6c2450ae PB |
582 | list_del(&ocqe->list); |
583 | kfree(ocqe); | |
1d7bb1d5 JA |
584 | } |
585 | ||
09e88404 PB |
586 | all_flushed = list_empty(&ctx->cq_overflow_list); |
587 | if (all_flushed) { | |
10988a0a | 588 | clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 589 | atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
09e88404 | 590 | } |
46930143 | 591 | |
25399321 | 592 | io_cq_unlock_post(ctx); |
09e88404 | 593 | return all_flushed; |
1d7bb1d5 JA |
594 | } |
595 | ||
90f67366 | 596 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx) |
6c503150 | 597 | { |
ca0a2651 JA |
598 | bool ret = true; |
599 | ||
10988a0a | 600 | if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) { |
6c503150 PB |
601 | /* iopoll syncs against uring_lock, not completion_lock */ |
602 | if (ctx->flags & IORING_SETUP_IOPOLL) | |
603 | mutex_lock(&ctx->uring_lock); | |
90f67366 | 604 | ret = __io_cqring_overflow_flush(ctx, false); |
6c503150 PB |
605 | if (ctx->flags & IORING_SETUP_IOPOLL) |
606 | mutex_unlock(&ctx->uring_lock); | |
607 | } | |
ca0a2651 JA |
608 | |
609 | return ret; | |
6c503150 PB |
610 | } |
611 | ||
e70cb608 | 612 | void __io_put_task(struct task_struct *task, int nr) |
6a290a14 PB |
613 | { |
614 | struct io_uring_task *tctx = task->io_uring; | |
615 | ||
9d170164 PB |
616 | percpu_counter_sub(&tctx->inflight, nr); |
617 | if (unlikely(atomic_read(&tctx->in_idle))) | |
618 | wake_up(&tctx->wait); | |
619 | put_task_struct_many(task, nr); | |
620 | } | |
621 | ||
63809137 | 622 | void io_task_refs_refill(struct io_uring_task *tctx) |
9a10867a PB |
623 | { |
624 | unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR; | |
625 | ||
626 | percpu_counter_add(&tctx->inflight, refill); | |
627 | refcount_add(refill, ¤t->usage); | |
628 | tctx->cached_refs += refill; | |
629 | } | |
630 | ||
3cc7fdb9 PB |
631 | static __cold void io_uring_drop_tctx_refs(struct task_struct *task) |
632 | { | |
633 | struct io_uring_task *tctx = task->io_uring; | |
634 | unsigned int refs = tctx->cached_refs; | |
635 | ||
636 | if (refs) { | |
637 | tctx->cached_refs = 0; | |
638 | percpu_counter_sub(&tctx->inflight, refs); | |
639 | put_task_struct_many(task, refs); | |
640 | } | |
641 | } | |
642 | ||
68494a65 PB |
643 | static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data, |
644 | s32 res, u32 cflags, u64 extra1, u64 extra2) | |
2b188cc1 | 645 | { |
cce4b8b0 | 646 | struct io_overflow_cqe *ocqe; |
e45a3e05 SR |
647 | size_t ocq_size = sizeof(struct io_overflow_cqe); |
648 | bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32); | |
2b188cc1 | 649 | |
e45a3e05 SR |
650 | if (is_cqe32) |
651 | ocq_size += sizeof(struct io_uring_cqe); | |
2b188cc1 | 652 | |
e45a3e05 | 653 | ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT); |
08dcd028 | 654 | trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe); |
cce4b8b0 PB |
655 | if (!ocqe) { |
656 | /* | |
657 | * If we're in ring overflow flush mode, or in task cancel mode, | |
658 | * or cannot allocate an overflow entry, then we need to drop it | |
659 | * on the floor. | |
660 | */ | |
8f6ed49a | 661 | io_account_cq_overflow(ctx); |
155bc950 | 662 | set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq); |
cce4b8b0 | 663 | return false; |
2b188cc1 | 664 | } |
cce4b8b0 | 665 | if (list_empty(&ctx->cq_overflow_list)) { |
10988a0a | 666 | set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq); |
3a4b89a2 | 667 | atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags); |
20c0b380 | 668 | |
cce4b8b0 | 669 | } |
d4d19c19 | 670 | ocqe->cqe.user_data = user_data; |
cce4b8b0 PB |
671 | ocqe->cqe.res = res; |
672 | ocqe->cqe.flags = cflags; | |
e45a3e05 SR |
673 | if (is_cqe32) { |
674 | ocqe->cqe.big_cqe[0] = extra1; | |
675 | ocqe->cqe.big_cqe[1] = extra2; | |
676 | } | |
cce4b8b0 PB |
677 | list_add_tail(&ocqe->list, &ctx->cq_overflow_list); |
678 | return true; | |
2b188cc1 JA |
679 | } |
680 | ||
68494a65 PB |
681 | bool io_req_cqe_overflow(struct io_kiocb *req) |
682 | { | |
683 | if (!(req->flags & REQ_F_CQE32_INIT)) { | |
684 | req->extra1 = 0; | |
685 | req->extra2 = 0; | |
686 | } | |
687 | return io_cqring_event_overflow(req->ctx, req->cqe.user_data, | |
688 | req->cqe.res, req->cqe.flags, | |
689 | req->extra1, req->extra2); | |
690 | } | |
691 | ||
faf88dde PB |
692 | /* |
693 | * writes to the cq entry need to come after reading head; the | |
694 | * control dependency is enough as we're using WRITE_ONCE to | |
695 | * fill the cq entry | |
696 | */ | |
697 | struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx) | |
698 | { | |
699 | struct io_rings *rings = ctx->rings; | |
700 | unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1); | |
faf88dde PB |
701 | unsigned int free, queued, len; |
702 | ||
faf88dde PB |
703 | |
704 | /* userspace may cheat modifying the tail, be safe and do min */ | |
705 | queued = min(__io_cqring_events(ctx), ctx->cq_entries); | |
706 | free = ctx->cq_entries - queued; | |
707 | /* we need a contiguous range, limit based on the current array offset */ | |
708 | len = min(free, ctx->cq_entries - off); | |
709 | if (!len) | |
710 | return NULL; | |
711 | ||
b3659a65 PB |
712 | if (ctx->flags & IORING_SETUP_CQE32) { |
713 | off <<= 1; | |
714 | len <<= 1; | |
715 | } | |
716 | ||
faf88dde PB |
717 | ctx->cqe_cached = &rings->cqes[off]; |
718 | ctx->cqe_sentinel = ctx->cqe_cached + len; | |
b3659a65 PB |
719 | |
720 | ctx->cached_cq_tail++; | |
faf88dde | 721 | ctx->cqe_cached++; |
b3659a65 PB |
722 | if (ctx->flags & IORING_SETUP_CQE32) |
723 | ctx->cqe_cached++; | |
724 | return &rings->cqes[off]; | |
faf88dde PB |
725 | } |
726 | ||
eb42cebb PB |
727 | bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, |
728 | bool allow_overflow) | |
bcda7baa | 729 | { |
cd94903d PB |
730 | struct io_uring_cqe *cqe; |
731 | ||
913a571a | 732 | ctx->cq_extra++; |
cd94903d PB |
733 | |
734 | /* | |
735 | * If we can't get a cq entry, userspace overflowed the | |
736 | * submission (by quite a lot). Increment the overflow count in | |
737 | * the ring. | |
738 | */ | |
739 | cqe = io_get_cqe(ctx); | |
740 | if (likely(cqe)) { | |
e0486f3f DY |
741 | trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0); |
742 | ||
cd94903d PB |
743 | WRITE_ONCE(cqe->user_data, user_data); |
744 | WRITE_ONCE(cqe->res, res); | |
745 | WRITE_ONCE(cqe->flags, cflags); | |
c5595975 PB |
746 | |
747 | if (ctx->flags & IORING_SETUP_CQE32) { | |
748 | WRITE_ONCE(cqe->big_cqe[0], 0); | |
749 | WRITE_ONCE(cqe->big_cqe[1], 0); | |
750 | } | |
cd94903d PB |
751 | return true; |
752 | } | |
52120f0f DY |
753 | |
754 | if (allow_overflow) | |
755 | return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0); | |
756 | ||
757 | return false; | |
bcda7baa JA |
758 | } |
759 | ||
d245bca6 | 760 | bool io_post_aux_cqe(struct io_ring_ctx *ctx, |
52120f0f DY |
761 | u64 user_data, s32 res, u32 cflags, |
762 | bool allow_overflow) | |
d245bca6 PB |
763 | { |
764 | bool filled; | |
765 | ||
25399321 | 766 | io_cq_lock(ctx); |
52120f0f | 767 | filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow); |
25399321 | 768 | io_cq_unlock_post(ctx); |
d245bca6 PB |
769 | return filled; |
770 | } | |
771 | ||
effcf8bd | 772 | static void __io_req_complete_put(struct io_kiocb *req) |
2b188cc1 | 773 | { |
c7dae4ba JA |
774 | /* |
775 | * If we're the last reference to this request, add to our locked | |
776 | * free_list cache. | |
777 | */ | |
de9b4cca | 778 | if (req_ref_put_and_test(req)) { |
effcf8bd SR |
779 | struct io_ring_ctx *ctx = req->ctx; |
780 | ||
da1a08c5 | 781 | if (req->flags & IO_REQ_LINK_FLAGS) { |
0756a869 | 782 | if (req->flags & IO_DISARM_MASK) |
7a612350 PB |
783 | io_disarm_next(req); |
784 | if (req->link) { | |
785 | io_req_task_queue(req->link); | |
786 | req->link = NULL; | |
787 | } | |
788 | } | |
7ac1edc4 | 789 | io_req_put_rsrc(req); |
8197b053 PB |
790 | /* |
791 | * Selected buffer deallocation in io_clean_op() assumes that | |
792 | * we don't hold ->completion_lock. Clean them here to avoid | |
793 | * deadlocks. | |
794 | */ | |
795 | io_put_kbuf_comp(req); | |
c7dae4ba JA |
796 | io_dismantle_req(req); |
797 | io_put_task(req->task, 1); | |
c2b6c6bc | 798 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
d0acdee2 | 799 | ctx->locked_free_nr++; |
180f829f | 800 | } |
a37fae8a HX |
801 | } |
802 | ||
59915143 | 803 | void __io_req_complete_post(struct io_kiocb *req) |
effcf8bd | 804 | { |
97b388d7 | 805 | if (!(req->flags & REQ_F_CQE_SKIP)) |
91ef75a7 | 806 | __io_fill_cqe_req(req->ctx, req); |
effcf8bd SR |
807 | __io_req_complete_put(req); |
808 | } | |
809 | ||
59915143 | 810 | void io_req_complete_post(struct io_kiocb *req) |
a37fae8a HX |
811 | { |
812 | struct io_ring_ctx *ctx = req->ctx; | |
813 | ||
25399321 | 814 | io_cq_lock(ctx); |
97b388d7 | 815 | __io_req_complete_post(req); |
25399321 | 816 | io_cq_unlock_post(ctx); |
4e3d9ff9 JA |
817 | } |
818 | ||
99f15d8d | 819 | inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags) |
bcda7baa | 820 | { |
75d7b3ae | 821 | io_req_complete_post(req); |
0ddf92e8 JA |
822 | } |
823 | ||
329061d3 | 824 | void io_req_complete_failed(struct io_kiocb *req, s32 res) |
f41db273 | 825 | { |
93d2bcd2 | 826 | req_set_fail(req); |
97b388d7 JA |
827 | io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED)); |
828 | io_req_complete_post(req); | |
f41db273 PB |
829 | } |
830 | ||
864ea921 PB |
831 | /* |
832 | * Don't initialise the fields below on every allocation, but do that in | |
833 | * advance and keep them valid across allocations. | |
834 | */ | |
835 | static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx) | |
836 | { | |
837 | req->ctx = ctx; | |
838 | req->link = NULL; | |
839 | req->async_data = NULL; | |
840 | /* not necessary, but safer to zero */ | |
cef216fc | 841 | req->cqe.res = 0; |
864ea921 PB |
842 | } |
843 | ||
dac7a098 | 844 | static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx, |
cd0ca2e0 | 845 | struct io_submit_state *state) |
dac7a098 | 846 | { |
79ebeaee | 847 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 848 | wq_list_splice(&ctx->locked_free_list, &state->free_list); |
d0acdee2 | 849 | ctx->locked_free_nr = 0; |
79ebeaee | 850 | spin_unlock(&ctx->completion_lock); |
dac7a098 PB |
851 | } |
852 | ||
5d5901a3 PB |
853 | /* |
854 | * A request might get retired back into the request caches even before opcode | |
855 | * handlers and io_issue_sqe() are done with it, e.g. inline completion path. | |
856 | * Because of that, io_alloc_req() should be called only under ->uring_lock | |
857 | * and with extra caution to not get a request that is still worked on. | |
858 | */ | |
bd1a3783 | 859 | __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx) |
5d5901a3 | 860 | __must_hold(&ctx->uring_lock) |
2b188cc1 | 861 | { |
864ea921 | 862 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
3ab665b7 | 863 | void *reqs[IO_REQ_ALLOC_BATCH]; |
864ea921 | 864 | int ret, i; |
e5d1bc0a | 865 | |
23a5c43b PB |
866 | /* |
867 | * If we have more than a batch's worth of requests in our IRQ side | |
868 | * locked cache, grab the lock and move them over to our submission | |
869 | * side cache. | |
870 | */ | |
a6d97a8a | 871 | if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) { |
23a5c43b | 872 | io_flush_cached_locked_reqs(ctx, &ctx->submit_state); |
88ab95be | 873 | if (!io_req_cache_empty(ctx)) |
23a5c43b PB |
874 | return true; |
875 | } | |
e5d1bc0a | 876 | |
3ab665b7 | 877 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs); |
fd6fab2c | 878 | |
864ea921 PB |
879 | /* |
880 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
881 | * retry single alloc to be on the safe side. | |
882 | */ | |
883 | if (unlikely(ret <= 0)) { | |
3ab665b7 PB |
884 | reqs[0] = kmem_cache_alloc(req_cachep, gfp); |
885 | if (!reqs[0]) | |
a33ae9ce | 886 | return false; |
864ea921 | 887 | ret = 1; |
2b188cc1 | 888 | } |
864ea921 | 889 | |
37f0e767 | 890 | percpu_ref_get_many(&ctx->refs, ret); |
3ab665b7 | 891 | for (i = 0; i < ret; i++) { |
23a5c43b | 892 | struct io_kiocb *req = reqs[i]; |
3ab665b7 PB |
893 | |
894 | io_preinit_req(req, ctx); | |
fa05457a | 895 | io_req_add_to_cache(req, ctx); |
3ab665b7 | 896 | } |
a33ae9ce PB |
897 | return true; |
898 | } | |
899 | ||
6b639522 | 900 | static inline void io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 901 | { |
094bae49 | 902 | unsigned int flags = req->flags; |
929a3af9 | 903 | |
867f8fa5 | 904 | if (unlikely(flags & IO_REQ_CLEAN_FLAGS)) |
3a0a6902 | 905 | io_clean_op(req); |
e1d767f0 PB |
906 | if (!(flags & REQ_F_FIXED_FILE)) |
907 | io_put_file(req->file); | |
e65ef56d JA |
908 | } |
909 | ||
59915143 | 910 | __cold void io_free_req(struct io_kiocb *req) |
c6ca97b3 | 911 | { |
51a4cc11 | 912 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 913 | |
7ac1edc4 | 914 | io_req_put_rsrc(req); |
216578e5 | 915 | io_dismantle_req(req); |
7c660731 | 916 | io_put_task(req->task, 1); |
c6ca97b3 | 917 | |
79ebeaee | 918 | spin_lock(&ctx->completion_lock); |
c2b6c6bc | 919 | wq_list_add_head(&req->comp_list, &ctx->locked_free_list); |
c34b025f | 920 | ctx->locked_free_nr++; |
79ebeaee | 921 | spin_unlock(&ctx->completion_lock); |
e65ef56d JA |
922 | } |
923 | ||
d81499bf PB |
924 | static void __io_req_find_next_prep(struct io_kiocb *req) |
925 | { | |
926 | struct io_ring_ctx *ctx = req->ctx; | |
d81499bf | 927 | |
25399321 | 928 | io_cq_lock(ctx); |
305bef98 | 929 | io_disarm_next(req); |
25399321 | 930 | io_cq_unlock_post(ctx); |
d81499bf PB |
931 | } |
932 | ||
933 | static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req) | |
c69f8dbe | 934 | { |
33cc89a9 | 935 | struct io_kiocb *nxt; |
944e58bf | 936 | |
9e645e11 JA |
937 | /* |
938 | * If LINK is set, we have dependent requests in this chain. If we | |
939 | * didn't fail this request, queue the first one up, moving any other | |
940 | * dependencies to the next request. In case of failure, fail the rest | |
941 | * of the chain. | |
942 | */ | |
d81499bf PB |
943 | if (unlikely(req->flags & IO_DISARM_MASK)) |
944 | __io_req_find_next_prep(req); | |
33cc89a9 PB |
945 | nxt = req->link; |
946 | req->link = NULL; | |
947 | return nxt; | |
4d7dd462 | 948 | } |
9e645e11 | 949 | |
f237c30a | 950 | static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked) |
2c32395d PB |
951 | { |
952 | if (!ctx) | |
953 | return; | |
ef060ea9 JA |
954 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
955 | atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
f237c30a | 956 | if (*locked) { |
c450178d | 957 | io_submit_flush_completions(ctx); |
2c32395d | 958 | mutex_unlock(&ctx->uring_lock); |
f237c30a | 959 | *locked = false; |
2c32395d PB |
960 | } |
961 | percpu_ref_put(&ctx->refs); | |
962 | } | |
963 | ||
c6dd763c DY |
964 | static unsigned int handle_tw_list(struct llist_node *node, |
965 | struct io_ring_ctx **ctx, bool *locked, | |
966 | struct llist_node *last) | |
9f8d032a | 967 | { |
c6dd763c DY |
968 | unsigned int count = 0; |
969 | ||
3a0c037b | 970 | while (node != last) { |
f88262e6 | 971 | struct llist_node *next = node->next; |
9f8d032a HX |
972 | struct io_kiocb *req = container_of(node, struct io_kiocb, |
973 | io_task_work.node); | |
974 | ||
34d2bfe7 JA |
975 | prefetch(container_of(next, struct io_kiocb, io_task_work.node)); |
976 | ||
9f8d032a HX |
977 | if (req->ctx != *ctx) { |
978 | ctx_flush_and_put(*ctx, locked); | |
979 | *ctx = req->ctx; | |
980 | /* if not contended, grab and improve batching */ | |
981 | *locked = mutex_trylock(&(*ctx)->uring_lock); | |
982 | percpu_ref_get(&(*ctx)->refs); | |
983 | } | |
984 | req->io_task_work.func(req, locked); | |
985 | node = next; | |
c6dd763c | 986 | count++; |
3a0c037b | 987 | } |
c6dd763c DY |
988 | |
989 | return count; | |
9f8d032a HX |
990 | } |
991 | ||
923d1592 DY |
992 | /** |
993 | * io_llist_xchg - swap all entries in a lock-less list | |
994 | * @head: the head of lock-less list to delete all entries | |
995 | * @new: new entry as the head of the list | |
996 | * | |
997 | * If list is empty, return NULL, otherwise, return the pointer to the first entry. | |
998 | * The order of entries returned is from the newest to the oldest added one. | |
999 | */ | |
1000 | static inline struct llist_node *io_llist_xchg(struct llist_head *head, | |
1001 | struct llist_node *new) | |
1002 | { | |
1003 | return xchg(&head->first, new); | |
1004 | } | |
1005 | ||
1006 | /** | |
1007 | * io_llist_cmpxchg - possibly swap all entries in a lock-less list | |
1008 | * @head: the head of lock-less list to delete all entries | |
1009 | * @old: expected old value of the first entry of the list | |
1010 | * @new: new entry as the head of the list | |
1011 | * | |
1012 | * perform a cmpxchg on the first entry of the list. | |
1013 | */ | |
1014 | ||
1015 | static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head, | |
1016 | struct llist_node *old, | |
1017 | struct llist_node *new) | |
1018 | { | |
1019 | return cmpxchg(&head->first, old, new); | |
1020 | } | |
1021 | ||
c9f06aa7 | 1022 | void tctx_task_work(struct callback_head *cb) |
c40f6379 | 1023 | { |
f28c240e | 1024 | bool uring_locked = false; |
ebd0df2e | 1025 | struct io_ring_ctx *ctx = NULL; |
3f18407d PB |
1026 | struct io_uring_task *tctx = container_of(cb, struct io_uring_task, |
1027 | task_work); | |
3a0c037b DY |
1028 | struct llist_node fake = {}; |
1029 | struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake); | |
c6dd763c DY |
1030 | unsigned int loops = 1; |
1031 | unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL); | |
3a0c037b | 1032 | |
3a0c037b DY |
1033 | node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); |
1034 | while (node != &fake) { | |
c6dd763c | 1035 | loops++; |
3a0c037b | 1036 | node = io_llist_xchg(&tctx->task_list, &fake); |
c6dd763c | 1037 | count += handle_tw_list(node, &ctx, &uring_locked, &fake); |
3a0c037b | 1038 | node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL); |
3f18407d | 1039 | } |
ebd0df2e | 1040 | |
f28c240e | 1041 | ctx_flush_and_put(ctx, &uring_locked); |
3cc7fdb9 PB |
1042 | |
1043 | /* relaxed read is enough as only the task itself sets ->in_idle */ | |
1044 | if (unlikely(atomic_read(&tctx->in_idle))) | |
1045 | io_uring_drop_tctx_refs(current); | |
c6dd763c DY |
1046 | |
1047 | trace_io_uring_task_work_run(tctx, count, loops); | |
7cbf1722 JA |
1048 | } |
1049 | ||
c34398a8 | 1050 | void io_req_task_work_add(struct io_kiocb *req) |
7cbf1722 | 1051 | { |
c34398a8 | 1052 | struct io_uring_task *tctx = req->task->io_uring; |
9f010507 | 1053 | struct io_ring_ctx *ctx = req->ctx; |
f88262e6 | 1054 | struct llist_node *node; |
6294f368 | 1055 | bool running; |
7cbf1722 | 1056 | |
f88262e6 | 1057 | running = !llist_add(&req->io_task_work.node, &tctx->task_list); |
7cbf1722 JA |
1058 | |
1059 | /* task_work already pending, we're done */ | |
6294f368 | 1060 | if (running) |
e09ee510 | 1061 | return; |
7cbf1722 | 1062 | |
ef060ea9 JA |
1063 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
1064 | atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags); | |
1065 | ||
3fe07bcd | 1066 | if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method))) |
e09ee510 | 1067 | return; |
2215bed9 | 1068 | |
f88262e6 | 1069 | node = llist_del_all(&tctx->task_list); |
7cbf1722 | 1070 | |
e09ee510 PB |
1071 | while (node) { |
1072 | req = container_of(node, struct io_kiocb, io_task_work.node); | |
1073 | node = node->next; | |
3218e5d3 | 1074 | if (llist_add(&req->io_task_work.node, |
e09ee510 PB |
1075 | &req->ctx->fallback_llist)) |
1076 | schedule_delayed_work(&req->ctx->fallback_work, 1); | |
1077 | } | |
eab30c4d PB |
1078 | } |
1079 | ||
4e118cd9 | 1080 | static void io_req_tw_post(struct io_kiocb *req, bool *locked) |
c40f6379 | 1081 | { |
97b388d7 | 1082 | io_req_complete_post(req); |
4e118cd9 | 1083 | } |
c40f6379 | 1084 | |
59915143 | 1085 | void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags) |
4e118cd9 | 1086 | { |
97b388d7 | 1087 | io_req_set_res(req, res, cflags); |
4e118cd9 | 1088 | req->io_task_work.func = io_req_tw_post; |
3fe07bcd | 1089 | io_req_task_work_add(req); |
4e118cd9 PB |
1090 | } |
1091 | ||
f237c30a | 1092 | static void io_req_task_cancel(struct io_kiocb *req, bool *locked) |
c40f6379 | 1093 | { |
b18a1a45 | 1094 | /* not needed for normal modes, but SQPOLL depends on it */ |
971cf9c1 | 1095 | io_tw_lock(req->ctx, locked); |
cef216fc | 1096 | io_req_complete_failed(req, req->cqe.res); |
c40f6379 JA |
1097 | } |
1098 | ||
329061d3 | 1099 | void io_req_task_submit(struct io_kiocb *req, bool *locked) |
c40f6379 | 1100 | { |
971cf9c1 | 1101 | io_tw_lock(req->ctx, locked); |
316319e8 | 1102 | /* req->task == current here, checking PF_EXITING is safe */ |
af066f31 | 1103 | if (likely(!(req->task->flags & PF_EXITING))) |
cbc2e203 | 1104 | io_queue_sqe(req); |
81b6d05c | 1105 | else |
2593553a | 1106 | io_req_complete_failed(req, -EFAULT); |
c40f6379 JA |
1107 | } |
1108 | ||
59915143 | 1109 | void io_req_task_queue_fail(struct io_kiocb *req, int ret) |
c40f6379 | 1110 | { |
97b388d7 | 1111 | io_req_set_res(req, ret, 0); |
5b0a6acc | 1112 | req->io_task_work.func = io_req_task_cancel; |
3fe07bcd | 1113 | io_req_task_work_add(req); |
c40f6379 JA |
1114 | } |
1115 | ||
f3b44f92 | 1116 | void io_req_task_queue(struct io_kiocb *req) |
a3df7698 | 1117 | { |
5b0a6acc | 1118 | req->io_task_work.func = io_req_task_submit; |
3fe07bcd | 1119 | io_req_task_work_add(req); |
a3df7698 PB |
1120 | } |
1121 | ||
59915143 | 1122 | void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1123 | { |
9b5f7bd9 | 1124 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1125 | |
1126 | if (nxt) | |
906a8c3f | 1127 | io_req_task_queue(nxt); |
c69f8dbe JL |
1128 | } |
1129 | ||
f3b44f92 | 1130 | void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node) |
3aa83bfb | 1131 | __must_hold(&ctx->uring_lock) |
5af1d13e | 1132 | { |
d4b7a5ef | 1133 | struct task_struct *task = NULL; |
37f0e767 | 1134 | int task_refs = 0; |
5af1d13e | 1135 | |
3aa83bfb PB |
1136 | do { |
1137 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1138 | comp_list); | |
2d6500d4 | 1139 | |
a538be5b PB |
1140 | if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) { |
1141 | if (req->flags & REQ_F_REFCOUNT) { | |
1142 | node = req->comp_list.next; | |
1143 | if (!req_ref_put_and_test(req)) | |
1144 | continue; | |
1145 | } | |
b605a7fa PB |
1146 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1147 | struct async_poll *apoll = req->apoll; | |
1148 | ||
1149 | if (apoll->double_poll) | |
1150 | kfree(apoll->double_poll); | |
9731bc98 JA |
1151 | if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache)) |
1152 | kfree(apoll); | |
b605a7fa PB |
1153 | req->flags &= ~REQ_F_POLLED; |
1154 | } | |
da1a08c5 | 1155 | if (req->flags & IO_REQ_LINK_FLAGS) |
57859f4d | 1156 | io_queue_next(req); |
a538be5b PB |
1157 | if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS)) |
1158 | io_clean_op(req); | |
c1e53a69 | 1159 | } |
a538be5b PB |
1160 | if (!(req->flags & REQ_F_FIXED_FILE)) |
1161 | io_put_file(req->file); | |
2d6500d4 | 1162 | |
ab409402 | 1163 | io_req_put_rsrc_locked(req, ctx); |
5af1d13e | 1164 | |
d4b7a5ef PB |
1165 | if (req->task != task) { |
1166 | if (task) | |
1167 | io_put_task(task, task_refs); | |
1168 | task = req->task; | |
1169 | task_refs = 0; | |
1170 | } | |
1171 | task_refs++; | |
c1e53a69 | 1172 | node = req->comp_list.next; |
fa05457a | 1173 | io_req_add_to_cache(req, ctx); |
3aa83bfb | 1174 | } while (node); |
d4b7a5ef | 1175 | |
d4b7a5ef PB |
1176 | if (task) |
1177 | io_put_task(task, task_refs); | |
7a743e22 PB |
1178 | } |
1179 | ||
c450178d | 1180 | static void __io_submit_flush_completions(struct io_ring_ctx *ctx) |
a141dd89 | 1181 | __must_hold(&ctx->uring_lock) |
905c172f | 1182 | { |
6f33b0bc | 1183 | struct io_wq_work_node *node, *prev; |
cd0ca2e0 | 1184 | struct io_submit_state *state = &ctx->submit_state; |
905c172f | 1185 | |
d9dee430 PB |
1186 | spin_lock(&ctx->completion_lock); |
1187 | wq_list_for_each(node, prev, &state->compl_reqs) { | |
1188 | struct io_kiocb *req = container_of(node, struct io_kiocb, | |
1189 | comp_list); | |
3d4aeb9f | 1190 | |
d9dee430 PB |
1191 | if (!(req->flags & REQ_F_CQE_SKIP)) |
1192 | __io_fill_cqe_req(ctx, req); | |
905c172f | 1193 | } |
25399321 | 1194 | __io_cq_unlock_post(ctx); |
d9dee430 | 1195 | |
1cce17ac | 1196 | io_free_batch_list(ctx, state->compl_reqs.first); |
6f33b0bc | 1197 | INIT_WQ_LIST(&state->compl_reqs); |
7a743e22 PB |
1198 | } |
1199 | ||
ba816ad6 JA |
1200 | /* |
1201 | * Drop reference to request, return next in chain (if there is one) if this | |
1202 | * was the last reference to this request. | |
1203 | */ | |
0d85035a | 1204 | static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 1205 | { |
9b5f7bd9 PB |
1206 | struct io_kiocb *nxt = NULL; |
1207 | ||
de9b4cca | 1208 | if (req_ref_put_and_test(req)) { |
da1a08c5 | 1209 | if (unlikely(req->flags & IO_REQ_LINK_FLAGS)) |
7819a1f6 | 1210 | nxt = io_req_find_next(req); |
f5c6cf2a | 1211 | io_free_req(req); |
2a44f467 | 1212 | } |
9b5f7bd9 | 1213 | return nxt; |
2b188cc1 JA |
1214 | } |
1215 | ||
6c503150 | 1216 | static unsigned io_cqring_events(struct io_ring_ctx *ctx) |
a3a0e43f JA |
1217 | { |
1218 | /* See comment at the top of this file */ | |
1219 | smp_rmb(); | |
e23de15f | 1220 | return __io_cqring_events(ctx); |
a3a0e43f JA |
1221 | } |
1222 | ||
def596e9 JA |
1223 | /* |
1224 | * We can't just wait for polled events to come to us, we have to actively | |
1225 | * find and complete them. | |
1226 | */ | |
c072481d | 1227 | static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
1228 | { |
1229 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1230 | return; | |
1231 | ||
1232 | mutex_lock(&ctx->uring_lock); | |
5eef4e87 | 1233 | while (!wq_list_empty(&ctx->iopoll_list)) { |
b2edc0a7 | 1234 | /* let it sleep and repeat later if can't complete a request */ |
5ba3c874 | 1235 | if (io_do_iopoll(ctx, true) == 0) |
b2edc0a7 | 1236 | break; |
08f5439f JA |
1237 | /* |
1238 | * Ensure we allow local-to-the-cpu processing to take place, | |
1239 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 1240 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 1241 | */ |
3fcee5a6 PB |
1242 | if (need_resched()) { |
1243 | mutex_unlock(&ctx->uring_lock); | |
1244 | cond_resched(); | |
1245 | mutex_lock(&ctx->uring_lock); | |
1246 | } | |
def596e9 JA |
1247 | } |
1248 | mutex_unlock(&ctx->uring_lock); | |
1249 | } | |
1250 | ||
7668b92a | 1251 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 1252 | { |
7668b92a | 1253 | unsigned int nr_events = 0; |
e9979b36 | 1254 | int ret = 0; |
155bc950 | 1255 | unsigned long check_cq; |
500f9fba | 1256 | |
3a08576b PB |
1257 | check_cq = READ_ONCE(ctx->check_cq); |
1258 | if (unlikely(check_cq)) { | |
1259 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
1260 | __io_cqring_overflow_flush(ctx, false); | |
1261 | /* | |
1262 | * Similarly do not spin if we have not informed the user of any | |
1263 | * dropped CQE. | |
1264 | */ | |
1265 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
1266 | return -EBADR; | |
1267 | } | |
f39c8a5b PB |
1268 | /* |
1269 | * Don't enter poll loop if we already have events pending. | |
1270 | * If we do, we can potentially be spinning for commands that | |
1271 | * already triggered a CQE (eg in error). | |
1272 | */ | |
f39c8a5b | 1273 | if (io_cqring_events(ctx)) |
d487b43c | 1274 | return 0; |
155bc950 | 1275 | |
def596e9 | 1276 | do { |
500f9fba JA |
1277 | /* |
1278 | * If a submit got punted to a workqueue, we can have the | |
1279 | * application entering polling for a command before it gets | |
1280 | * issued. That app will hold the uring_lock for the duration | |
1281 | * of the poll right here, so we need to take a breather every | |
1282 | * now and then to ensure that the issue has a chance to add | |
1283 | * the poll to the issued list. Otherwise we can spin here | |
1284 | * forever, while the workqueue is stuck trying to acquire the | |
1285 | * very same mutex. | |
1286 | */ | |
5eef4e87 | 1287 | if (wq_list_empty(&ctx->iopoll_list)) { |
8f487ef2 PB |
1288 | u32 tail = ctx->cached_cq_tail; |
1289 | ||
500f9fba | 1290 | mutex_unlock(&ctx->uring_lock); |
4c6e277c | 1291 | io_run_task_work(); |
500f9fba | 1292 | mutex_lock(&ctx->uring_lock); |
def596e9 | 1293 | |
8f487ef2 PB |
1294 | /* some requests don't go through iopoll_list */ |
1295 | if (tail != ctx->cached_cq_tail || | |
5eef4e87 | 1296 | wq_list_empty(&ctx->iopoll_list)) |
e9979b36 | 1297 | break; |
500f9fba | 1298 | } |
5ba3c874 PB |
1299 | ret = io_do_iopoll(ctx, !min); |
1300 | if (ret < 0) | |
1301 | break; | |
1302 | nr_events += ret; | |
1303 | ret = 0; | |
1304 | } while (nr_events < min && !need_resched()); | |
d487b43c | 1305 | |
def596e9 JA |
1306 | return ret; |
1307 | } | |
7012c815 PB |
1308 | |
1309 | void io_req_task_complete(struct io_kiocb *req, bool *locked) | |
8ef12efe | 1310 | { |
7012c815 PB |
1311 | if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) { |
1312 | unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED; | |
1313 | ||
1314 | req->cqe.flags |= io_put_kbuf(req, issue_flags); | |
1315 | } | |
1316 | ||
1317 | if (*locked) | |
9da070b1 | 1318 | io_req_complete_defer(req); |
7012c815 | 1319 | else |
97b388d7 | 1320 | io_req_complete_post(req); |
8ef12efe JA |
1321 | } |
1322 | ||
def596e9 JA |
1323 | /* |
1324 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1325 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
f39c8a5b | 1326 | * find it from a io_do_iopoll() thread before the issuer is done |
def596e9 JA |
1327 | * accessing the kiocb cookie. |
1328 | */ | |
9882131c | 1329 | static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags) |
def596e9 JA |
1330 | { |
1331 | struct io_ring_ctx *ctx = req->ctx; | |
3b44b371 | 1332 | const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED; |
cb3d8972 PB |
1333 | |
1334 | /* workqueue context doesn't hold uring_lock, grab it now */ | |
3b44b371 | 1335 | if (unlikely(needs_lock)) |
cb3d8972 | 1336 | mutex_lock(&ctx->uring_lock); |
def596e9 JA |
1337 | |
1338 | /* | |
1339 | * Track whether we have multiple files in our lists. This will impact | |
1340 | * how we do polling eventually, not spinning if we're on potentially | |
1341 | * different devices. | |
1342 | */ | |
5eef4e87 | 1343 | if (wq_list_empty(&ctx->iopoll_list)) { |
915b3dde HX |
1344 | ctx->poll_multi_queue = false; |
1345 | } else if (!ctx->poll_multi_queue) { | |
def596e9 JA |
1346 | struct io_kiocb *list_req; |
1347 | ||
5eef4e87 PB |
1348 | list_req = container_of(ctx->iopoll_list.first, struct io_kiocb, |
1349 | comp_list); | |
30da1b45 | 1350 | if (list_req->file != req->file) |
915b3dde | 1351 | ctx->poll_multi_queue = true; |
def596e9 JA |
1352 | } |
1353 | ||
1354 | /* | |
1355 | * For fast devices, IO may have already completed. If it has, add | |
1356 | * it to the front so we find it first. | |
1357 | */ | |
65a6543d | 1358 | if (READ_ONCE(req->iopoll_completed)) |
5eef4e87 | 1359 | wq_list_add_head(&req->comp_list, &ctx->iopoll_list); |
def596e9 | 1360 | else |
5eef4e87 | 1361 | wq_list_add_tail(&req->comp_list, &ctx->iopoll_list); |
bdcd3eab | 1362 | |
3b44b371 | 1363 | if (unlikely(needs_lock)) { |
cb3d8972 PB |
1364 | /* |
1365 | * If IORING_SETUP_SQPOLL is enabled, sqes are either handle | |
1366 | * in sq thread task context or in io worker task context. If | |
1367 | * current task context is sq thread, we don't need to check | |
1368 | * whether should wake up sq thread. | |
1369 | */ | |
1370 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1371 | wq_has_sleeper(&ctx->sq_data->wait)) | |
1372 | wake_up(&ctx->sq_data->wait); | |
1373 | ||
1374 | mutex_unlock(&ctx->uring_lock); | |
1375 | } | |
def596e9 JA |
1376 | } |
1377 | ||
4503b767 JA |
1378 | static bool io_bdev_nowait(struct block_device *bdev) |
1379 | { | |
9ba0d0c8 | 1380 | return !bdev || blk_queue_nowait(bdev_get_queue(bdev)); |
4503b767 JA |
1381 | } |
1382 | ||
2b188cc1 JA |
1383 | /* |
1384 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1385 | * any file. For now, just ensure that anything potentially problematic is done | |
1386 | * inline. | |
1387 | */ | |
88459b50 | 1388 | static bool __io_file_supports_nowait(struct file *file, umode_t mode) |
2b188cc1 | 1389 | { |
4503b767 | 1390 | if (S_ISBLK(mode)) { |
4e7b5671 CH |
1391 | if (IS_ENABLED(CONFIG_BLOCK) && |
1392 | io_bdev_nowait(I_BDEV(file->f_mapping->host))) | |
4503b767 JA |
1393 | return true; |
1394 | return false; | |
1395 | } | |
976517f1 | 1396 | if (S_ISSOCK(mode)) |
2b188cc1 | 1397 | return true; |
4503b767 | 1398 | if (S_ISREG(mode)) { |
4e7b5671 CH |
1399 | if (IS_ENABLED(CONFIG_BLOCK) && |
1400 | io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
e5550a14 | 1401 | !io_is_uring_fops(file)) |
4503b767 JA |
1402 | return true; |
1403 | return false; | |
1404 | } | |
2b188cc1 | 1405 | |
c5b85625 JA |
1406 | /* any ->read/write should understand O_NONBLOCK */ |
1407 | if (file->f_flags & O_NONBLOCK) | |
1408 | return true; | |
35645ac3 | 1409 | return file->f_mode & FMODE_NOWAIT; |
2b188cc1 | 1410 | } |
c5b85625 | 1411 | |
88459b50 PB |
1412 | /* |
1413 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1414 | * any file. For now, just ensure that anything potentially problematic is done | |
1415 | * inline. | |
1416 | */ | |
a4ad4f74 | 1417 | unsigned int io_file_get_flags(struct file *file) |
88459b50 PB |
1418 | { |
1419 | umode_t mode = file_inode(file)->i_mode; | |
1420 | unsigned int res = 0; | |
af197f50 | 1421 | |
88459b50 PB |
1422 | if (S_ISREG(mode)) |
1423 | res |= FFS_ISREG; | |
1424 | if (__io_file_supports_nowait(file, mode)) | |
1425 | res |= FFS_NOWAIT; | |
5e45690a JA |
1426 | if (io_file_need_scm(file)) |
1427 | res |= FFS_SCM; | |
88459b50 | 1428 | return res; |
2b188cc1 JA |
1429 | } |
1430 | ||
99f15d8d | 1431 | bool io_alloc_async_data(struct io_kiocb *req) |
3d9932a8 | 1432 | { |
e8c2bc1f JA |
1433 | WARN_ON_ONCE(!io_op_defs[req->opcode].async_size); |
1434 | req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL); | |
d886e185 PB |
1435 | if (req->async_data) { |
1436 | req->flags |= REQ_F_ASYNC_DATA; | |
1437 | return false; | |
1438 | } | |
1439 | return true; | |
3d9932a8 XW |
1440 | } |
1441 | ||
f3b44f92 | 1442 | int io_req_prep_async(struct io_kiocb *req) |
f67676d1 | 1443 | { |
0702e536 JA |
1444 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
1445 | ||
1446 | /* assign early for deferred execution for non-fixed file */ | |
1447 | if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE)) | |
1448 | req->file = io_file_get_normal(req, req->cqe.fd); | |
dc919caf | 1449 | if (!def->prep_async) |
0702e536 JA |
1450 | return 0; |
1451 | if (WARN_ON_ONCE(req_has_async_data(req))) | |
1452 | return -EFAULT; | |
1453 | if (io_alloc_async_data(req)) | |
1454 | return -EAGAIN; | |
1455 | ||
dc919caf | 1456 | return def->prep_async(req); |
bfe76559 PB |
1457 | } |
1458 | ||
9cf7c104 PB |
1459 | static u32 io_get_sequence(struct io_kiocb *req) |
1460 | { | |
a3dbdf54 | 1461 | u32 seq = req->ctx->cached_sq_head; |
963c6abb | 1462 | struct io_kiocb *cur; |
9cf7c104 | 1463 | |
a3dbdf54 | 1464 | /* need original cached_sq_head, but it was increased for each req */ |
963c6abb | 1465 | io_for_each_link(cur, req) |
a3dbdf54 PB |
1466 | seq--; |
1467 | return seq; | |
9cf7c104 PB |
1468 | } |
1469 | ||
c072481d | 1470 | static __cold void io_drain_req(struct io_kiocb *req) |
de0617e4 | 1471 | { |
a197f664 | 1472 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 1473 | struct io_defer_entry *de; |
f67676d1 | 1474 | int ret; |
e0eb71dc | 1475 | u32 seq = io_get_sequence(req); |
3c19966d | 1476 | |
9d858b21 | 1477 | /* Still need defer if there is pending req in defer list. */ |
e302f104 | 1478 | spin_lock(&ctx->completion_lock); |
5e371265 | 1479 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) { |
e302f104 | 1480 | spin_unlock(&ctx->completion_lock); |
e0eb71dc | 1481 | queue: |
10c66904 | 1482 | ctx->drain_active = false; |
e0eb71dc PB |
1483 | io_req_task_queue(req); |
1484 | return; | |
10c66904 | 1485 | } |
e302f104 | 1486 | spin_unlock(&ctx->completion_lock); |
9cf7c104 | 1487 | |
b7e298d2 | 1488 | ret = io_req_prep_async(req); |
e0eb71dc PB |
1489 | if (ret) { |
1490 | fail: | |
1491 | io_req_complete_failed(req, ret); | |
1492 | return; | |
1493 | } | |
cbdcb435 | 1494 | io_prep_async_link(req); |
27dc8338 | 1495 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
76cc33d7 | 1496 | if (!de) { |
1b48773f | 1497 | ret = -ENOMEM; |
e0eb71dc | 1498 | goto fail; |
76cc33d7 | 1499 | } |
2d28390a | 1500 | |
79ebeaee | 1501 | spin_lock(&ctx->completion_lock); |
9cf7c104 | 1502 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
79ebeaee | 1503 | spin_unlock(&ctx->completion_lock); |
27dc8338 | 1504 | kfree(de); |
e0eb71dc | 1505 | goto queue; |
de0617e4 JA |
1506 | } |
1507 | ||
48863ffd | 1508 | trace_io_uring_defer(req); |
27dc8338 | 1509 | de->req = req; |
9cf7c104 | 1510 | de->seq = seq; |
27dc8338 | 1511 | list_add_tail(&de->list, &ctx->defer_list); |
79ebeaee | 1512 | spin_unlock(&ctx->completion_lock); |
de0617e4 JA |
1513 | } |
1514 | ||
68fb8979 | 1515 | static void io_clean_op(struct io_kiocb *req) |
99bc4c38 | 1516 | { |
8197b053 PB |
1517 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
1518 | spin_lock(&req->ctx->completion_lock); | |
cc3cec83 | 1519 | io_put_kbuf_comp(req); |
8197b053 PB |
1520 | spin_unlock(&req->ctx->completion_lock); |
1521 | } | |
99bc4c38 | 1522 | |
0e1b6fe3 | 1523 | if (req->flags & REQ_F_NEED_CLEANUP) { |
4d4c9cff | 1524 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
bb040a21 | 1525 | |
4d4c9cff JA |
1526 | if (def->cleanup) |
1527 | def->cleanup(req); | |
99bc4c38 | 1528 | } |
75652a30 JA |
1529 | if ((req->flags & REQ_F_POLLED) && req->apoll) { |
1530 | kfree(req->apoll->double_poll); | |
1531 | kfree(req->apoll); | |
1532 | req->apoll = NULL; | |
1533 | } | |
9cae36a0 JA |
1534 | if (req->flags & REQ_F_INFLIGHT) { |
1535 | struct io_uring_task *tctx = req->task->io_uring; | |
1536 | ||
1537 | atomic_dec(&tctx->inflight_tracked); | |
1538 | } | |
c854357b | 1539 | if (req->flags & REQ_F_CREDS) |
b8e64b53 | 1540 | put_cred(req->creds); |
d886e185 PB |
1541 | if (req->flags & REQ_F_ASYNC_DATA) { |
1542 | kfree(req->async_data); | |
1543 | req->async_data = NULL; | |
1544 | } | |
c854357b | 1545 | req->flags &= ~IO_REQ_CLEAN_FLAGS; |
99bc4c38 PB |
1546 | } |
1547 | ||
6bf9c47a JA |
1548 | static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags) |
1549 | { | |
1550 | if (req->file || !io_op_defs[req->opcode].needs_file) | |
1551 | return true; | |
1552 | ||
1553 | if (req->flags & REQ_F_FIXED_FILE) | |
cef216fc | 1554 | req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags); |
6bf9c47a | 1555 | else |
cef216fc | 1556 | req->file = io_file_get_normal(req, req->cqe.fd); |
6bf9c47a | 1557 | |
772f5e00 | 1558 | return !!req->file; |
6bf9c47a JA |
1559 | } |
1560 | ||
889fca73 | 1561 | static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags) |
2b188cc1 | 1562 | { |
fcde59fe | 1563 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
5730b27e | 1564 | const struct cred *creds = NULL; |
d625c6ee | 1565 | int ret; |
2b188cc1 | 1566 | |
70152140 JA |
1567 | if (unlikely(!io_assign_file(req, issue_flags))) |
1568 | return -EBADF; | |
1569 | ||
6878b40e | 1570 | if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred())) |
c10d1f98 | 1571 | creds = override_creds(req->creds); |
5730b27e | 1572 | |
fcde59fe | 1573 | if (!def->audit_skip) |
5bd2182d PM |
1574 | audit_uring_entry(req->opcode); |
1575 | ||
0702e536 | 1576 | ret = def->issue(req, issue_flags); |
2b188cc1 | 1577 | |
fcde59fe | 1578 | if (!def->audit_skip) |
5bd2182d PM |
1579 | audit_uring_exit(!ret, ret); |
1580 | ||
5730b27e JA |
1581 | if (creds) |
1582 | revert_creds(creds); | |
97b388d7 | 1583 | |
75d7b3ae PB |
1584 | if (ret == IOU_OK) { |
1585 | if (issue_flags & IO_URING_F_COMPLETE_DEFER) | |
9da070b1 | 1586 | io_req_complete_defer(req); |
75d7b3ae PB |
1587 | else |
1588 | io_req_complete_post(req); | |
1589 | } else if (ret != IOU_ISSUE_SKIP_COMPLETE) | |
def596e9 | 1590 | return ret; |
97b388d7 | 1591 | |
b532576e | 1592 | /* If the op doesn't have a file, we're not polling for it */ |
9983028e | 1593 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file) |
9882131c | 1594 | io_iopoll_req_issued(req, issue_flags); |
def596e9 JA |
1595 | |
1596 | return 0; | |
2b188cc1 JA |
1597 | } |
1598 | ||
329061d3 JA |
1599 | int io_poll_issue(struct io_kiocb *req, bool *locked) |
1600 | { | |
1601 | io_tw_lock(req->ctx, locked); | |
1602 | if (unlikely(req->task->flags & PF_EXITING)) | |
1603 | return -EFAULT; | |
aeaa72c6 | 1604 | return io_issue_sqe(req, IO_URING_F_NONBLOCK); |
329061d3 JA |
1605 | } |
1606 | ||
c9f06aa7 | 1607 | struct io_wq_work *io_wq_free_work(struct io_wq_work *work) |
ebc11b6c PB |
1608 | { |
1609 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
1610 | ||
1611 | req = io_put_req_find_next(req); | |
1612 | return req ? &req->work : NULL; | |
1613 | } | |
1614 | ||
c9f06aa7 | 1615 | void io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
1616 | { |
1617 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6bf9c47a | 1618 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
d01905db PB |
1619 | unsigned int issue_flags = IO_URING_F_UNLOCKED; |
1620 | bool needs_poll = false; | |
6bf9c47a | 1621 | int ret = 0, err = -ECANCELED; |
2b188cc1 | 1622 | |
48dcd38d PB |
1623 | /* one will be dropped by ->io_free_work() after returning to io-wq */ |
1624 | if (!(req->flags & REQ_F_REFCOUNT)) | |
1625 | __io_req_set_refcount(req, 2); | |
1626 | else | |
1627 | req_ref_get(req); | |
5d5901a3 | 1628 | |
cb2d344c | 1629 | io_arm_ltimeout(req); |
6bf9c47a | 1630 | |
dadebc35 | 1631 | /* either cancelled or io-wq is dying, so don't touch tctx->iowq */ |
d01905db | 1632 | if (work->flags & IO_WQ_WORK_CANCEL) { |
0f8da75b | 1633 | fail: |
6bf9c47a | 1634 | io_req_task_queue_fail(req, err); |
d01905db PB |
1635 | return; |
1636 | } | |
0f8da75b PB |
1637 | if (!io_assign_file(req, issue_flags)) { |
1638 | err = -EBADF; | |
1639 | work->flags |= IO_WQ_WORK_CANCEL; | |
1640 | goto fail; | |
1641 | } | |
31b51510 | 1642 | |
d01905db | 1643 | if (req->flags & REQ_F_FORCE_ASYNC) { |
afb7f56f PB |
1644 | bool opcode_poll = def->pollin || def->pollout; |
1645 | ||
1646 | if (opcode_poll && file_can_poll(req->file)) { | |
1647 | needs_poll = true; | |
d01905db | 1648 | issue_flags |= IO_URING_F_NONBLOCK; |
afb7f56f | 1649 | } |
561fb04a | 1650 | } |
31b51510 | 1651 | |
d01905db PB |
1652 | do { |
1653 | ret = io_issue_sqe(req, issue_flags); | |
1654 | if (ret != -EAGAIN) | |
1655 | break; | |
1656 | /* | |
1657 | * We can get EAGAIN for iopolled IO even though we're | |
1658 | * forcing a sync submission from here, since we can't | |
1659 | * wait for request slots on the block side. | |
1660 | */ | |
1661 | if (!needs_poll) { | |
e0deb6a0 PB |
1662 | if (!(req->ctx->flags & IORING_SETUP_IOPOLL)) |
1663 | break; | |
d01905db PB |
1664 | cond_resched(); |
1665 | continue; | |
90fa0288 HX |
1666 | } |
1667 | ||
4d9237e3 | 1668 | if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK) |
d01905db PB |
1669 | return; |
1670 | /* aborted or ready, in either case retry blocking */ | |
1671 | needs_poll = false; | |
1672 | issue_flags &= ~IO_URING_F_NONBLOCK; | |
1673 | } while (1); | |
31b51510 | 1674 | |
a3df7698 | 1675 | /* avoid locking problems by failing it from a clean context */ |
97b388d7 | 1676 | if (ret < 0) |
a3df7698 | 1677 | io_req_task_queue_fail(req, ret); |
2b188cc1 JA |
1678 | } |
1679 | ||
531113bb JA |
1680 | inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd, |
1681 | unsigned int issue_flags) | |
09bb8394 | 1682 | { |
5106dd6e JA |
1683 | struct io_ring_ctx *ctx = req->ctx; |
1684 | struct file *file = NULL; | |
ac177053 | 1685 | unsigned long file_ptr; |
09bb8394 | 1686 | |
93f052cb | 1687 | io_ring_submit_lock(ctx, issue_flags); |
5106dd6e | 1688 | |
ac177053 | 1689 | if (unlikely((unsigned int)fd >= ctx->nr_user_files)) |
5106dd6e | 1690 | goto out; |
ac177053 PB |
1691 | fd = array_index_nospec(fd, ctx->nr_user_files); |
1692 | file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr; | |
1693 | file = (struct file *) (file_ptr & FFS_MASK); | |
1694 | file_ptr &= ~FFS_MASK; | |
1695 | /* mask in overlapping REQ_F and FFS bits */ | |
35645ac3 | 1696 | req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT); |
5106dd6e | 1697 | io_req_set_rsrc_node(req, ctx, 0); |
d78bd8ad | 1698 | WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap)); |
5106dd6e | 1699 | out: |
93f052cb | 1700 | io_ring_submit_unlock(ctx, issue_flags); |
ac177053 PB |
1701 | return file; |
1702 | } | |
d44f554e | 1703 | |
531113bb | 1704 | struct file *io_file_get_normal(struct io_kiocb *req, int fd) |
ac177053 | 1705 | { |
62906e89 | 1706 | struct file *file = fget(fd); |
ac177053 | 1707 | |
48863ffd | 1708 | trace_io_uring_file_get(req, fd); |
09bb8394 | 1709 | |
ac177053 | 1710 | /* we don't allow fixed io_uring files */ |
e5550a14 | 1711 | if (file && io_is_uring_fops(file)) |
9cae36a0 | 1712 | io_req_track_inflight(req); |
8371adf5 | 1713 | return file; |
09bb8394 JA |
1714 | } |
1715 | ||
7bfa9bad | 1716 | static void io_queue_async(struct io_kiocb *req, int ret) |
d475a9a6 PB |
1717 | __must_hold(&req->ctx->uring_lock) |
1718 | { | |
7bfa9bad PB |
1719 | struct io_kiocb *linked_timeout; |
1720 | ||
1721 | if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) { | |
1722 | io_req_complete_failed(req, ret); | |
1723 | return; | |
1724 | } | |
1725 | ||
1726 | linked_timeout = io_prep_linked_timeout(req); | |
d475a9a6 | 1727 | |
4d9237e3 | 1728 | switch (io_arm_poll_handler(req, 0)) { |
d475a9a6 | 1729 | case IO_APOLL_READY: |
d475a9a6 PB |
1730 | io_req_task_queue(req); |
1731 | break; | |
1732 | case IO_APOLL_ABORTED: | |
1733 | /* | |
1734 | * Queued up for async execution, worker will release | |
1735 | * submit reference when the iocb is actually submitted. | |
1736 | */ | |
6436c770 | 1737 | io_kbuf_recycle(req, 0); |
77955efb | 1738 | io_queue_iowq(req, NULL); |
d475a9a6 | 1739 | break; |
b1c62645 | 1740 | case IO_APOLL_OK: |
b1c62645 | 1741 | break; |
d475a9a6 PB |
1742 | } |
1743 | ||
1744 | if (linked_timeout) | |
1745 | io_queue_linked_timeout(linked_timeout); | |
1746 | } | |
1747 | ||
cbc2e203 | 1748 | static inline void io_queue_sqe(struct io_kiocb *req) |
282cdc86 | 1749 | __must_hold(&req->ctx->uring_lock) |
2b188cc1 | 1750 | { |
e0c5c576 | 1751 | int ret; |
2b188cc1 | 1752 | |
c5eef2b9 | 1753 | ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER); |
193155c8 | 1754 | |
491381ce JA |
1755 | /* |
1756 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
1757 | * doesn't support non-blocking read/write attempts | |
1758 | */ | |
7bfa9bad | 1759 | if (likely(!ret)) |
cb2d344c | 1760 | io_arm_ltimeout(req); |
7bfa9bad PB |
1761 | else |
1762 | io_queue_async(req, ret); | |
2b188cc1 JA |
1763 | } |
1764 | ||
4652fe3f | 1765 | static void io_queue_sqe_fallback(struct io_kiocb *req) |
282cdc86 | 1766 | __must_hold(&req->ctx->uring_lock) |
4fe2c963 | 1767 | { |
17b147f6 PB |
1768 | if (unlikely(req->flags & REQ_F_FAIL)) { |
1769 | /* | |
1770 | * We don't submit, fail them all, for that replace hardlinks | |
1771 | * with normal links. Extra REQ_F_LINK is tolerated. | |
1772 | */ | |
1773 | req->flags &= ~REQ_F_HARDLINK; | |
1774 | req->flags |= REQ_F_LINK; | |
1775 | io_req_complete_failed(req, req->cqe.res); | |
e0eb71dc PB |
1776 | } else if (unlikely(req->ctx->drain_active)) { |
1777 | io_drain_req(req); | |
76cc33d7 PB |
1778 | } else { |
1779 | int ret = io_req_prep_async(req); | |
1780 | ||
1781 | if (unlikely(ret)) | |
1782 | io_req_complete_failed(req, ret); | |
1783 | else | |
77955efb | 1784 | io_queue_iowq(req, NULL); |
ce35a47a | 1785 | } |
4fe2c963 JL |
1786 | } |
1787 | ||
b16fed66 PB |
1788 | /* |
1789 | * Check SQE restrictions (opcode and flags). | |
1790 | * | |
1791 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
1792 | */ | |
1793 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
1794 | struct io_kiocb *req, | |
1795 | unsigned int sqe_flags) | |
4fe2c963 | 1796 | { |
b16fed66 PB |
1797 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) |
1798 | return false; | |
1799 | ||
1800 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
1801 | ctx->restrictions.sqe_flags_required) | |
1802 | return false; | |
1803 | ||
1804 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
1805 | ctx->restrictions.sqe_flags_required)) | |
1806 | return false; | |
1807 | ||
1808 | return true; | |
4fe2c963 JL |
1809 | } |
1810 | ||
22b2ca31 PB |
1811 | static void io_init_req_drain(struct io_kiocb *req) |
1812 | { | |
1813 | struct io_ring_ctx *ctx = req->ctx; | |
1814 | struct io_kiocb *head = ctx->submit_state.link.head; | |
1815 | ||
1816 | ctx->drain_active = true; | |
1817 | if (head) { | |
1818 | /* | |
1819 | * If we need to drain a request in the middle of a link, drain | |
1820 | * the head request and the next request/link after the current | |
1821 | * link. Considering sequential execution of links, | |
b6c7db32 | 1822 | * REQ_F_IO_DRAIN will be maintained for every request of our |
22b2ca31 PB |
1823 | * link. |
1824 | */ | |
b6c7db32 | 1825 | head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
22b2ca31 PB |
1826 | ctx->drain_next = true; |
1827 | } | |
1828 | } | |
1829 | ||
b16fed66 PB |
1830 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, |
1831 | const struct io_uring_sqe *sqe) | |
282cdc86 | 1832 | __must_hold(&ctx->uring_lock) |
b16fed66 | 1833 | { |
fcde59fe | 1834 | const struct io_op_def *def; |
b16fed66 | 1835 | unsigned int sqe_flags; |
fc0ae024 | 1836 | int personality; |
4a04d1d1 | 1837 | u8 opcode; |
b16fed66 | 1838 | |
864ea921 | 1839 | /* req is partially pre-initialised, see io_preinit_req() */ |
4a04d1d1 | 1840 | req->opcode = opcode = READ_ONCE(sqe->opcode); |
b16fed66 PB |
1841 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
1842 | req->flags = sqe_flags = READ_ONCE(sqe->flags); | |
cef216fc | 1843 | req->cqe.user_data = READ_ONCE(sqe->user_data); |
b16fed66 | 1844 | req->file = NULL; |
c1bdf8ed | 1845 | req->rsrc_node = NULL; |
b16fed66 | 1846 | req->task = current; |
b16fed66 | 1847 | |
4a04d1d1 PB |
1848 | if (unlikely(opcode >= IORING_OP_LAST)) { |
1849 | req->opcode = 0; | |
b16fed66 | 1850 | return -EINVAL; |
4a04d1d1 | 1851 | } |
fcde59fe | 1852 | def = &io_op_defs[opcode]; |
68fe256a PB |
1853 | if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) { |
1854 | /* enforce forwards compatibility on users */ | |
1855 | if (sqe_flags & ~SQE_VALID_FLAGS) | |
1856 | return -EINVAL; | |
4e906702 | 1857 | if (sqe_flags & IOSQE_BUFFER_SELECT) { |
fcde59fe | 1858 | if (!def->buffer_select) |
4e906702 JA |
1859 | return -EOPNOTSUPP; |
1860 | req->buf_index = READ_ONCE(sqe->buf_group); | |
1861 | } | |
5562a8d7 PB |
1862 | if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS) |
1863 | ctx->drain_disabled = true; | |
1864 | if (sqe_flags & IOSQE_IO_DRAIN) { | |
1865 | if (ctx->drain_disabled) | |
1866 | return -EOPNOTSUPP; | |
22b2ca31 | 1867 | io_init_req_drain(req); |
5562a8d7 | 1868 | } |
2a56a9bd PB |
1869 | } |
1870 | if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) { | |
1871 | if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags)) | |
1872 | return -EACCES; | |
1873 | /* knock it to the slow queue path, will be drained there */ | |
1874 | if (ctx->drain_active) | |
1875 | req->flags |= REQ_F_FORCE_ASYNC; | |
1876 | /* if there is no link, we're at "next" request and need to drain */ | |
1877 | if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) { | |
1878 | ctx->drain_next = false; | |
1879 | ctx->drain_active = true; | |
b6c7db32 | 1880 | req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC; |
2a56a9bd | 1881 | } |
68fe256a | 1882 | } |
b16fed66 | 1883 | |
fcde59fe | 1884 | if (!def->ioprio && sqe->ioprio) |
73911426 | 1885 | return -EINVAL; |
fcde59fe | 1886 | if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL)) |
73911426 JA |
1887 | return -EINVAL; |
1888 | ||
fcde59fe | 1889 | if (def->needs_file) { |
6d63416d PB |
1890 | struct io_submit_state *state = &ctx->submit_state; |
1891 | ||
cef216fc | 1892 | req->cqe.fd = READ_ONCE(sqe->fd); |
6bf9c47a | 1893 | |
6d63416d PB |
1894 | /* |
1895 | * Plug now if we have more than 2 IO left after this, and the | |
1896 | * target is potentially a read/write to block based storage. | |
1897 | */ | |
fcde59fe | 1898 | if (state->need_plug && def->plug) { |
6d63416d PB |
1899 | state->plug_started = true; |
1900 | state->need_plug = false; | |
5ca7a8b3 | 1901 | blk_start_plug_nr_ios(&state->plug, state->submit_nr); |
6d63416d | 1902 | } |
b16fed66 | 1903 | } |
863e0560 | 1904 | |
003e8dcc JA |
1905 | personality = READ_ONCE(sqe->personality); |
1906 | if (personality) { | |
cdab10bf LT |
1907 | int ret; |
1908 | ||
c10d1f98 PB |
1909 | req->creds = xa_load(&ctx->personalities, personality); |
1910 | if (!req->creds) | |
003e8dcc | 1911 | return -EINVAL; |
c10d1f98 | 1912 | get_cred(req->creds); |
cdc1404a PM |
1913 | ret = security_uring_override_creds(req->creds); |
1914 | if (ret) { | |
1915 | put_cred(req->creds); | |
1916 | return ret; | |
1917 | } | |
b8e64b53 | 1918 | req->flags |= REQ_F_CREDS; |
003e8dcc | 1919 | } |
b16fed66 | 1920 | |
0702e536 | 1921 | return def->prep(req, sqe); |
b16fed66 PB |
1922 | } |
1923 | ||
df3becde PB |
1924 | static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe, |
1925 | struct io_kiocb *req, int ret) | |
1926 | { | |
1927 | struct io_ring_ctx *ctx = req->ctx; | |
1928 | struct io_submit_link *link = &ctx->submit_state.link; | |
1929 | struct io_kiocb *head = link->head; | |
1930 | ||
48863ffd | 1931 | trace_io_uring_req_failed(sqe, req, ret); |
df3becde PB |
1932 | |
1933 | /* | |
1934 | * Avoid breaking links in the middle as it renders links with SQPOLL | |
1935 | * unusable. Instead of failing eagerly, continue assembling the link if | |
1936 | * applicable and mark the head with REQ_F_FAIL. The link flushing code | |
1937 | * should find the flag and handle the rest. | |
1938 | */ | |
1939 | req_fail_link_node(req, ret); | |
1940 | if (head && !(head->flags & REQ_F_FAIL)) | |
1941 | req_fail_link_node(head, -ECANCELED); | |
1942 | ||
1943 | if (!(req->flags & IO_REQ_LINK_FLAGS)) { | |
1944 | if (head) { | |
1945 | link->last->link = req; | |
1946 | link->head = NULL; | |
1947 | req = head; | |
1948 | } | |
1949 | io_queue_sqe_fallback(req); | |
1950 | return ret; | |
1951 | } | |
1952 | ||
1953 | if (head) | |
1954 | link->last->link = req; | |
1955 | else | |
1956 | link->head = req; | |
1957 | link->last = req; | |
1958 | return 0; | |
1959 | } | |
1960 | ||
1961 | static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
a1ab7b35 | 1962 | const struct io_uring_sqe *sqe) |
282cdc86 | 1963 | __must_hold(&ctx->uring_lock) |
9e645e11 | 1964 | { |
a1ab7b35 | 1965 | struct io_submit_link *link = &ctx->submit_state.link; |
ef4ff581 | 1966 | int ret; |
9e645e11 | 1967 | |
a6b8cadc | 1968 | ret = io_init_req(ctx, req, sqe); |
df3becde PB |
1969 | if (unlikely(ret)) |
1970 | return io_submit_fail_init(sqe, req, ret); | |
441b8a78 | 1971 | |
be7053b7 | 1972 | /* don't need @sqe from now on */ |
48863ffd | 1973 | trace_io_uring_submit_sqe(req, true); |
a6b8cadc | 1974 | |
9e645e11 JA |
1975 | /* |
1976 | * If we already have a head request, queue this one for async | |
1977 | * submittal once the head completes. If we don't have a head but | |
1978 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
1979 | * submitted sync once the chain is complete. If none of those | |
1980 | * conditions are true (normal request), then just queue it. | |
1981 | */ | |
924a07e4 | 1982 | if (unlikely(link->head)) { |
df3becde PB |
1983 | ret = io_req_prep_async(req); |
1984 | if (unlikely(ret)) | |
1985 | return io_submit_fail_init(sqe, req, ret); | |
1986 | ||
48863ffd | 1987 | trace_io_uring_link(req, link->head); |
f2f87370 | 1988 | link->last->link = req; |
863e0560 | 1989 | link->last = req; |
32fe525b | 1990 | |
da1a08c5 | 1991 | if (req->flags & IO_REQ_LINK_FLAGS) |
f15a3431 | 1992 | return 0; |
df3becde PB |
1993 | /* last request of the link, flush it */ |
1994 | req = link->head; | |
f15a3431 | 1995 | link->head = NULL; |
924a07e4 PB |
1996 | if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL)) |
1997 | goto fallback; | |
1998 | ||
1999 | } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS | | |
2000 | REQ_F_FORCE_ASYNC | REQ_F_FAIL))) { | |
2001 | if (req->flags & IO_REQ_LINK_FLAGS) { | |
2002 | link->head = req; | |
2003 | link->last = req; | |
2004 | } else { | |
2005 | fallback: | |
2006 | io_queue_sqe_fallback(req); | |
2007 | } | |
f15a3431 | 2008 | return 0; |
9e645e11 | 2009 | } |
2e6e1fde | 2010 | |
f15a3431 | 2011 | io_queue_sqe(req); |
1d4240cc | 2012 | return 0; |
9e645e11 JA |
2013 | } |
2014 | ||
9a56a232 JA |
2015 | /* |
2016 | * Batched submission is done, ensure local IO is flushed out. | |
2017 | */ | |
553deffd | 2018 | static void io_submit_state_end(struct io_ring_ctx *ctx) |
9a56a232 | 2019 | { |
553deffd PB |
2020 | struct io_submit_state *state = &ctx->submit_state; |
2021 | ||
e126391c PB |
2022 | if (unlikely(state->link.head)) |
2023 | io_queue_sqe_fallback(state->link.head); | |
553deffd | 2024 | /* flush only after queuing links as they can generate completions */ |
c450178d | 2025 | io_submit_flush_completions(ctx); |
27926b68 JA |
2026 | if (state->plug_started) |
2027 | blk_finish_plug(&state->plug); | |
9a56a232 JA |
2028 | } |
2029 | ||
2030 | /* | |
2031 | * Start submission side cache. | |
2032 | */ | |
2033 | static void io_submit_state_start(struct io_submit_state *state, | |
ba88ff11 | 2034 | unsigned int max_ios) |
9a56a232 | 2035 | { |
27926b68 | 2036 | state->plug_started = false; |
4b628aeb | 2037 | state->need_plug = max_ios > 2; |
5ca7a8b3 | 2038 | state->submit_nr = max_ios; |
a1ab7b35 PB |
2039 | /* set only head, no need to init link_last in advance */ |
2040 | state->link.head = NULL; | |
9a56a232 JA |
2041 | } |
2042 | ||
2b188cc1 JA |
2043 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
2044 | { | |
75b28aff | 2045 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 2046 | |
caf582c6 PB |
2047 | /* |
2048 | * Ensure any loads from the SQEs are done at this point, | |
2049 | * since once we write the new head, the application could | |
2050 | * write new data to them. | |
2051 | */ | |
2052 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
2053 | } |
2054 | ||
2b188cc1 | 2055 | /* |
dd9ae8a0 | 2056 | * Fetch an sqe, if one is available. Note this returns a pointer to memory |
2b188cc1 JA |
2057 | * that is mapped by userspace. This means that care needs to be taken to |
2058 | * ensure that reads are stable, as we cannot rely on userspace always | |
2059 | * being a good citizen. If members of the sqe are validated and then later | |
2060 | * used, it's important that those reads are done through READ_ONCE() to | |
2061 | * prevent a re-load down the line. | |
2062 | */ | |
709b302f | 2063 | static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) |
2b188cc1 | 2064 | { |
ea5ab3b5 | 2065 | unsigned head, mask = ctx->sq_entries - 1; |
17d3aeb3 | 2066 | unsigned sq_idx = ctx->cached_sq_head++ & mask; |
2b188cc1 JA |
2067 | |
2068 | /* | |
2069 | * The cached sq head (or cq tail) serves two purposes: | |
2070 | * | |
2071 | * 1) allows us to batch the cost of updating the user visible | |
2072 | * head updates. | |
2073 | * 2) allows the kernel side to track the head on its own, even | |
2074 | * though the application is the one updating it. | |
2075 | */ | |
17d3aeb3 | 2076 | head = READ_ONCE(ctx->sq_array[sq_idx]); |
ebdeb7c0 JA |
2077 | if (likely(head < ctx->sq_entries)) { |
2078 | /* double index for 128-byte SQEs, twice as long */ | |
2079 | if (ctx->flags & IORING_SETUP_SQE128) | |
2080 | head <<= 1; | |
709b302f | 2081 | return &ctx->sq_sqes[head]; |
ebdeb7c0 | 2082 | } |
2b188cc1 JA |
2083 | |
2084 | /* drop invalid entries */ | |
15641e42 PB |
2085 | ctx->cq_extra--; |
2086 | WRITE_ONCE(ctx->rings->sq_dropped, | |
2087 | READ_ONCE(ctx->rings->sq_dropped) + 1); | |
709b302f PB |
2088 | return NULL; |
2089 | } | |
2090 | ||
17437f31 | 2091 | int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
282cdc86 | 2092 | __must_hold(&ctx->uring_lock) |
6c271ce2 | 2093 | { |
69629809 | 2094 | unsigned int entries = io_sqring_entries(ctx); |
8e6971a8 PB |
2095 | unsigned int left; |
2096 | int ret; | |
6c271ce2 | 2097 | |
51d48dab | 2098 | if (unlikely(!entries)) |
69629809 | 2099 | return 0; |
ee7d46d9 | 2100 | /* make sure SQ entry isn't read before tail */ |
8e6971a8 PB |
2101 | ret = left = min3(nr, ctx->sq_entries, entries); |
2102 | io_get_task_refs(left); | |
2103 | io_submit_state_start(&ctx->submit_state, left); | |
6c271ce2 | 2104 | |
69629809 | 2105 | do { |
3529d8c2 | 2106 | const struct io_uring_sqe *sqe; |
196be95c | 2107 | struct io_kiocb *req; |
fb5ccc98 | 2108 | |
8e6971a8 | 2109 | if (unlikely(!io_alloc_req_refill(ctx))) |
fb5ccc98 | 2110 | break; |
a33ae9ce | 2111 | req = io_alloc_req(ctx); |
4fccfcbb PB |
2112 | sqe = io_get_sqe(ctx); |
2113 | if (unlikely(!sqe)) { | |
fa05457a | 2114 | io_req_add_to_cache(req, ctx); |
4fccfcbb PB |
2115 | break; |
2116 | } | |
6c271ce2 | 2117 | |
1cd15904 PB |
2118 | /* |
2119 | * Continue submitting even for sqe failure if the | |
2120 | * ring was setup with IORING_SETUP_SUBMIT_ALL | |
2121 | */ | |
2122 | if (unlikely(io_submit_sqe(ctx, req, sqe)) && | |
2123 | !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) { | |
2124 | left--; | |
2125 | break; | |
bcbb7bf6 | 2126 | } |
1cd15904 | 2127 | } while (--left); |
9466f437 | 2128 | |
8e6971a8 PB |
2129 | if (unlikely(left)) { |
2130 | ret -= left; | |
2131 | /* try again if it submitted nothing and can't allocate a req */ | |
2132 | if (!ret && io_req_cache_empty(ctx)) | |
2133 | ret = -EAGAIN; | |
2134 | current->io_uring->cached_refs += left; | |
9466f437 | 2135 | } |
6c271ce2 | 2136 | |
553deffd | 2137 | io_submit_state_end(ctx); |
ae9428ca PB |
2138 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
2139 | io_commit_sqring(ctx); | |
8e6971a8 | 2140 | return ret; |
6c271ce2 JA |
2141 | } |
2142 | ||
bda52162 JA |
2143 | struct io_wait_queue { |
2144 | struct wait_queue_entry wq; | |
2145 | struct io_ring_ctx *ctx; | |
5fd46178 | 2146 | unsigned cq_tail; |
bda52162 JA |
2147 | unsigned nr_timeouts; |
2148 | }; | |
2149 | ||
6c503150 | 2150 | static inline bool io_should_wake(struct io_wait_queue *iowq) |
bda52162 JA |
2151 | { |
2152 | struct io_ring_ctx *ctx = iowq->ctx; | |
5fd46178 | 2153 | int dist = ctx->cached_cq_tail - (int) iowq->cq_tail; |
bda52162 JA |
2154 | |
2155 | /* | |
d195a66e | 2156 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
2157 | * started waiting. For timeouts, we always want to return to userspace, |
2158 | * regardless of event count. | |
2159 | */ | |
5fd46178 | 2160 | return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
bda52162 JA |
2161 | } |
2162 | ||
2163 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
2164 | int wake_flags, void *key) | |
2165 | { | |
2166 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
2167 | wq); | |
2168 | ||
6c503150 PB |
2169 | /* |
2170 | * Cannot safely flush overflowed CQEs from here, ensure we wake up | |
2171 | * the task, and the next invocation will do it. | |
2172 | */ | |
10988a0a DY |
2173 | if (io_should_wake(iowq) || |
2174 | test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &iowq->ctx->check_cq)) | |
6c503150 PB |
2175 | return autoremove_wake_function(curr, mode, wake_flags, key); |
2176 | return -1; | |
bda52162 JA |
2177 | } |
2178 | ||
73572984 | 2179 | int io_run_task_work_sig(void) |
af9c1a44 JA |
2180 | { |
2181 | if (io_run_task_work()) | |
2182 | return 1; | |
c5020bc8 OL |
2183 | if (task_sigpending(current)) |
2184 | return -EINTR; | |
2185 | return 0; | |
af9c1a44 JA |
2186 | } |
2187 | ||
eeb60b9a PB |
2188 | /* when returns >0, the caller should retry */ |
2189 | static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, | |
2190 | struct io_wait_queue *iowq, | |
22833966 | 2191 | ktime_t timeout) |
eeb60b9a PB |
2192 | { |
2193 | int ret; | |
155bc950 | 2194 | unsigned long check_cq; |
eeb60b9a PB |
2195 | |
2196 | /* make sure we run task_work before checking for signals */ | |
2197 | ret = io_run_task_work_sig(); | |
2198 | if (ret || io_should_wake(iowq)) | |
2199 | return ret; | |
3a08576b | 2200 | |
155bc950 | 2201 | check_cq = READ_ONCE(ctx->check_cq); |
3a08576b PB |
2202 | if (unlikely(check_cq)) { |
2203 | /* let the caller flush overflows, retry */ | |
2204 | if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)) | |
2205 | return 1; | |
2206 | if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT)) | |
2207 | return -EBADR; | |
2208 | } | |
22833966 JA |
2209 | if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS)) |
2210 | return -ETIME; | |
2211 | return 1; | |
eeb60b9a PB |
2212 | } |
2213 | ||
2b188cc1 JA |
2214 | /* |
2215 | * Wait until events become available, if we don't already have some. The | |
2216 | * application must reap them itself, as they reside on the shared cq ring. | |
2217 | */ | |
2218 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
c73ebb68 HX |
2219 | const sigset_t __user *sig, size_t sigsz, |
2220 | struct __kernel_timespec __user *uts) | |
2b188cc1 | 2221 | { |
90291099 | 2222 | struct io_wait_queue iowq; |
75b28aff | 2223 | struct io_rings *rings = ctx->rings; |
22833966 | 2224 | ktime_t timeout = KTIME_MAX; |
c1d5a224 | 2225 | int ret; |
2b188cc1 | 2226 | |
b41e9852 | 2227 | do { |
90f67366 | 2228 | io_cqring_overflow_flush(ctx); |
6c503150 | 2229 | if (io_cqring_events(ctx) >= min_events) |
b41e9852 | 2230 | return 0; |
4c6e277c | 2231 | if (!io_run_task_work()) |
b41e9852 | 2232 | break; |
b41e9852 | 2233 | } while (1); |
2b188cc1 JA |
2234 | |
2235 | if (sig) { | |
9e75ad5d AB |
2236 | #ifdef CONFIG_COMPAT |
2237 | if (in_compat_syscall()) | |
2238 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 2239 | sigsz); |
9e75ad5d AB |
2240 | else |
2241 | #endif | |
b772434b | 2242 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 2243 | |
2b188cc1 JA |
2244 | if (ret) |
2245 | return ret; | |
2246 | } | |
2247 | ||
950e79dd OL |
2248 | if (uts) { |
2249 | struct timespec64 ts; | |
2250 | ||
2251 | if (get_timespec64(&ts, uts)) | |
2252 | return -EFAULT; | |
2253 | timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns()); | |
2254 | } | |
2255 | ||
90291099 PB |
2256 | init_waitqueue_func_entry(&iowq.wq, io_wake_function); |
2257 | iowq.wq.private = current; | |
2258 | INIT_LIST_HEAD(&iowq.wq.entry); | |
2259 | iowq.ctx = ctx; | |
bda52162 | 2260 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
5fd46178 | 2261 | iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events; |
90291099 | 2262 | |
c826bd7a | 2263 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 | 2264 | do { |
ca0a2651 | 2265 | /* if we can't even flush overflow, don't wait for more */ |
90f67366 | 2266 | if (!io_cqring_overflow_flush(ctx)) { |
ca0a2651 JA |
2267 | ret = -EBUSY; |
2268 | break; | |
2269 | } | |
311997b3 | 2270 | prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq, |
bda52162 | 2271 | TASK_INTERRUPTIBLE); |
22833966 | 2272 | ret = io_cqring_wait_schedule(ctx, &iowq, timeout); |
ca0a2651 | 2273 | cond_resched(); |
eeb60b9a | 2274 | } while (ret > 0); |
bda52162 | 2275 | |
b4f20bb4 | 2276 | finish_wait(&ctx->cq_wait, &iowq.wq); |
b7db41c9 | 2277 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 2278 | |
75b28aff | 2279 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
2280 | } |
2281 | ||
73572984 | 2282 | static void io_mem_free(void *ptr) |
b9bd2bea | 2283 | { |
73572984 | 2284 | struct page *page; |
b36a2050 | 2285 | |
73572984 JA |
2286 | if (!ptr) |
2287 | return; | |
b9bd2bea | 2288 | |
73572984 JA |
2289 | page = virt_to_head_page(ptr); |
2290 | if (put_page_testzero(page)) | |
2291 | free_compound_page(page); | |
b9bd2bea PB |
2292 | } |
2293 | ||
73572984 | 2294 | static void *io_mem_alloc(size_t size) |
b9bd2bea | 2295 | { |
73572984 | 2296 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP; |
b9bd2bea | 2297 | |
73572984 | 2298 | return (void *) __get_free_pages(gfp, get_order(size)); |
b9bd2bea PB |
2299 | } |
2300 | ||
73572984 JA |
2301 | static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries, |
2302 | unsigned int cq_entries, size_t *sq_offset) | |
6b06314c | 2303 | { |
73572984 JA |
2304 | struct io_rings *rings; |
2305 | size_t off, sq_array_size; | |
6b06314c | 2306 | |
73572984 JA |
2307 | off = struct_size(rings, cqes, cq_entries); |
2308 | if (off == SIZE_MAX) | |
2309 | return SIZE_MAX; | |
2310 | if (ctx->flags & IORING_SETUP_CQE32) { | |
2311 | if (check_shl_overflow(off, 1, &off)) | |
2312 | return SIZE_MAX; | |
2313 | } | |
ab409402 | 2314 | |
73572984 JA |
2315 | #ifdef CONFIG_SMP |
2316 | off = ALIGN(off, SMP_CACHE_BYTES); | |
2317 | if (off == 0) | |
2318 | return SIZE_MAX; | |
2319 | #endif | |
82fbcfa9 | 2320 | |
73572984 JA |
2321 | if (sq_offset) |
2322 | *sq_offset = off; | |
82fbcfa9 | 2323 | |
73572984 JA |
2324 | sq_array_size = array_size(sizeof(u32), sq_entries); |
2325 | if (sq_array_size == SIZE_MAX) | |
2326 | return SIZE_MAX; | |
6b06314c | 2327 | |
73572984 JA |
2328 | if (check_add_overflow(off, sq_array_size, &off)) |
2329 | return SIZE_MAX; | |
8bad28d8 | 2330 | |
73572984 | 2331 | return off; |
8bad28d8 HX |
2332 | } |
2333 | ||
73572984 JA |
2334 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg, |
2335 | unsigned int eventfd_async) | |
8bad28d8 | 2336 | { |
73572984 JA |
2337 | struct io_ev_fd *ev_fd; |
2338 | __s32 __user *fds = arg; | |
2339 | int fd; | |
f2303b1f | 2340 | |
73572984 JA |
2341 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, |
2342 | lockdep_is_held(&ctx->uring_lock)); | |
2343 | if (ev_fd) | |
2344 | return -EBUSY; | |
8bad28d8 | 2345 | |
73572984 JA |
2346 | if (copy_from_user(&fd, fds, sizeof(*fds))) |
2347 | return -EFAULT; | |
8dd03afe | 2348 | |
73572984 JA |
2349 | ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL); |
2350 | if (!ev_fd) | |
2351 | return -ENOMEM; | |
05f3fb3c | 2352 | |
73572984 JA |
2353 | ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd); |
2354 | if (IS_ERR(ev_fd->cq_ev_fd)) { | |
2355 | int ret = PTR_ERR(ev_fd->cq_ev_fd); | |
2356 | kfree(ev_fd); | |
2357 | return ret; | |
2358 | } | |
305bef98 PB |
2359 | |
2360 | spin_lock(&ctx->completion_lock); | |
2361 | ctx->evfd_last_cq_tail = ctx->cached_cq_tail; | |
2362 | spin_unlock(&ctx->completion_lock); | |
2363 | ||
73572984 JA |
2364 | ev_fd->eventfd_async = eventfd_async; |
2365 | ctx->has_evfd = true; | |
2366 | rcu_assign_pointer(ctx->io_ev_fd, ev_fd); | |
2367 | return 0; | |
d7954b2b BM |
2368 | } |
2369 | ||
73572984 | 2370 | static void io_eventfd_put(struct rcu_head *rcu) |
2d091d62 | 2371 | { |
73572984 | 2372 | struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu); |
2d091d62 | 2373 | |
73572984 JA |
2374 | eventfd_ctx_put(ev_fd->cq_ev_fd); |
2375 | kfree(ev_fd); | |
2d091d62 PB |
2376 | } |
2377 | ||
73572984 | 2378 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) |
1ad555c6 | 2379 | { |
73572984 JA |
2380 | struct io_ev_fd *ev_fd; |
2381 | ||
2382 | ev_fd = rcu_dereference_protected(ctx->io_ev_fd, | |
2383 | lockdep_is_held(&ctx->uring_lock)); | |
2384 | if (ev_fd) { | |
2385 | ctx->has_evfd = false; | |
2386 | rcu_assign_pointer(ctx->io_ev_fd, NULL); | |
2387 | call_rcu(&ev_fd->rcu, io_eventfd_put); | |
2388 | return 0; | |
2389 | } | |
2d091d62 | 2390 | |
73572984 | 2391 | return -ENXIO; |
44b31f2f PB |
2392 | } |
2393 | ||
73572984 | 2394 | static void io_req_caches_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2395 | { |
cd0ca2e0 | 2396 | struct io_submit_state *state = &ctx->submit_state; |
37f0e767 | 2397 | int nr = 0; |
bf019da7 | 2398 | |
9a4fdbd8 | 2399 | mutex_lock(&ctx->uring_lock); |
cd0ca2e0 | 2400 | io_flush_cached_locked_reqs(ctx, state); |
9a4fdbd8 | 2401 | |
88ab95be | 2402 | while (!io_req_cache_empty(ctx)) { |
c2b6c6bc PB |
2403 | struct io_wq_work_node *node; |
2404 | struct io_kiocb *req; | |
9a4fdbd8 | 2405 | |
c2b6c6bc PB |
2406 | node = wq_stack_extract(&state->free_list); |
2407 | req = container_of(node, struct io_kiocb, comp_list); | |
2408 | kmem_cache_free(req_cachep, req); | |
37f0e767 | 2409 | nr++; |
c2b6c6bc | 2410 | } |
37f0e767 PB |
2411 | if (nr) |
2412 | percpu_ref_put_many(&ctx->refs, nr); | |
9a4fdbd8 JA |
2413 | mutex_unlock(&ctx->uring_lock); |
2414 | } | |
2415 | ||
c072481d | 2416 | static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx) |
2b188cc1 | 2417 | { |
37d1e2e3 | 2418 | io_sq_thread_finish(ctx); |
2aede0e4 | 2419 | |
37d1e2e3 | 2420 | if (ctx->mm_account) { |
2aede0e4 JA |
2421 | mmdrop(ctx->mm_account); |
2422 | ctx->mm_account = NULL; | |
30975825 | 2423 | } |
def596e9 | 2424 | |
ab409402 | 2425 | io_rsrc_refs_drop(ctx); |
43597aac PB |
2426 | /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */ |
2427 | io_wait_rsrc_data(ctx->buf_data); | |
2428 | io_wait_rsrc_data(ctx->file_data); | |
2429 | ||
8bad28d8 | 2430 | mutex_lock(&ctx->uring_lock); |
43597aac | 2431 | if (ctx->buf_data) |
bd54b6fe | 2432 | __io_sqe_buffers_unregister(ctx); |
43597aac | 2433 | if (ctx->file_data) |
08480400 | 2434 | __io_sqe_files_unregister(ctx); |
c4ea060e PB |
2435 | if (ctx->rings) |
2436 | __io_cqring_overflow_flush(ctx, true); | |
9b402849 | 2437 | io_eventfd_unregister(ctx); |
9b797a37 | 2438 | io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free); |
43e0bbbd | 2439 | io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free); |
77bc59b4 | 2440 | mutex_unlock(&ctx->uring_lock); |
5a2e745d | 2441 | io_destroy_buffers(ctx); |
07db298a PB |
2442 | if (ctx->sq_creds) |
2443 | put_cred(ctx->sq_creds); | |
97bbdc06 PB |
2444 | if (ctx->submitter_task) |
2445 | put_task_struct(ctx->submitter_task); | |
def596e9 | 2446 | |
a7f0ed5a PB |
2447 | /* there are no registered resources left, nobody uses it */ |
2448 | if (ctx->rsrc_node) | |
2449 | io_rsrc_node_destroy(ctx->rsrc_node); | |
8dd03afe | 2450 | if (ctx->rsrc_backup_node) |
b895c9a6 | 2451 | io_rsrc_node_destroy(ctx->rsrc_backup_node); |
a7f0ed5a | 2452 | flush_delayed_work(&ctx->rsrc_put_work); |
756ab7c0 | 2453 | flush_delayed_work(&ctx->fallback_work); |
a7f0ed5a PB |
2454 | |
2455 | WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list)); | |
2456 | WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist)); | |
def596e9 | 2457 | |
2b188cc1 | 2458 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
2459 | if (ctx->ring_sock) { |
2460 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 2461 | sock_release(ctx->ring_sock); |
355e8d26 | 2462 | } |
2b188cc1 | 2463 | #endif |
ef9dd637 | 2464 | WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list)); |
eb42cebb | 2465 | WARN_ON_ONCE(ctx->notif_slots || ctx->nr_notif_slots); |
2b188cc1 | 2466 | |
75b28aff | 2467 | io_mem_free(ctx->rings); |
2b188cc1 | 2468 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
2469 | |
2470 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 2471 | free_uid(ctx->user); |
4010fec4 | 2472 | io_req_caches_free(ctx); |
e941894e JA |
2473 | if (ctx->hash_map) |
2474 | io_wq_put_hash(ctx->hash_map); | |
e6f89be6 | 2475 | kfree(ctx->cancel_table.hbs); |
9ca9fb24 | 2476 | kfree(ctx->cancel_table_locked.hbs); |
6224843d | 2477 | kfree(ctx->dummy_ubuf); |
9cfc7e94 JA |
2478 | kfree(ctx->io_bl); |
2479 | xa_destroy(&ctx->io_bl_xa); | |
2b188cc1 JA |
2480 | kfree(ctx); |
2481 | } | |
2482 | ||
2483 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
2484 | { | |
2485 | struct io_ring_ctx *ctx = file->private_data; | |
2486 | __poll_t mask = 0; | |
2487 | ||
d60aa65b | 2488 | poll_wait(file, &ctx->cq_wait, wait); |
4f7067c3 SB |
2489 | /* |
2490 | * synchronizes with barrier from wq_has_sleeper call in | |
2491 | * io_commit_cqring | |
2492 | */ | |
2b188cc1 | 2493 | smp_rmb(); |
90554200 | 2494 | if (!io_sqring_full(ctx)) |
2b188cc1 | 2495 | mask |= EPOLLOUT | EPOLLWRNORM; |
ed670c3f HX |
2496 | |
2497 | /* | |
2498 | * Don't flush cqring overflow list here, just do a simple check. | |
2499 | * Otherwise there could possible be ABBA deadlock: | |
2500 | * CPU0 CPU1 | |
2501 | * ---- ---- | |
2502 | * lock(&ctx->uring_lock); | |
2503 | * lock(&ep->mtx); | |
2504 | * lock(&ctx->uring_lock); | |
2505 | * lock(&ep->mtx); | |
2506 | * | |
2507 | * Users may get EPOLLIN meanwhile seeing nothing in cqring, this | |
2508 | * pushs them to do the flush. | |
2509 | */ | |
10988a0a DY |
2510 | if (io_cqring_events(ctx) || |
2511 | test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) | |
2b188cc1 JA |
2512 | mask |= EPOLLIN | EPOLLRDNORM; |
2513 | ||
2514 | return mask; | |
2515 | } | |
2516 | ||
0bead8cd | 2517 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) |
071698e1 | 2518 | { |
4379bf8b | 2519 | const struct cred *creds; |
071698e1 | 2520 | |
61cf9370 | 2521 | creds = xa_erase(&ctx->personalities, id); |
4379bf8b JA |
2522 | if (creds) { |
2523 | put_cred(creds); | |
0bead8cd | 2524 | return 0; |
1e6fa521 | 2525 | } |
0bead8cd YD |
2526 | |
2527 | return -EINVAL; | |
2528 | } | |
2529 | ||
d56d938b PB |
2530 | struct io_tctx_exit { |
2531 | struct callback_head task_work; | |
2532 | struct completion completion; | |
baf186c4 | 2533 | struct io_ring_ctx *ctx; |
d56d938b PB |
2534 | }; |
2535 | ||
c072481d | 2536 | static __cold void io_tctx_exit_cb(struct callback_head *cb) |
d56d938b PB |
2537 | { |
2538 | struct io_uring_task *tctx = current->io_uring; | |
2539 | struct io_tctx_exit *work; | |
2540 | ||
2541 | work = container_of(cb, struct io_tctx_exit, task_work); | |
2542 | /* | |
2543 | * When @in_idle, we're in cancellation and it's racy to remove the | |
2544 | * node. It'll be removed by the end of cancellation, just ignore it. | |
2545 | */ | |
2546 | if (!atomic_read(&tctx->in_idle)) | |
eef51daa | 2547 | io_uring_del_tctx_node((unsigned long)work->ctx); |
d56d938b PB |
2548 | complete(&work->completion); |
2549 | } | |
2550 | ||
c072481d | 2551 | static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data) |
28090c13 PB |
2552 | { |
2553 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
2554 | ||
2555 | return req->ctx == data; | |
2556 | } | |
2557 | ||
c072481d | 2558 | static __cold void io_ring_exit_work(struct work_struct *work) |
85faa7b8 | 2559 | { |
d56d938b | 2560 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work); |
b5bb3a24 | 2561 | unsigned long timeout = jiffies + HZ * 60 * 5; |
58d3be2c | 2562 | unsigned long interval = HZ / 20; |
d56d938b PB |
2563 | struct io_tctx_exit exit; |
2564 | struct io_tctx_node *node; | |
2565 | int ret; | |
85faa7b8 | 2566 | |
56952e91 JA |
2567 | /* |
2568 | * If we're doing polled IO and end up having requests being | |
2569 | * submitted async (out-of-line), then completions can come in while | |
2570 | * we're waiting for refs to drop. We need to reap these manually, | |
2571 | * as nobody else will be looking for them. | |
2572 | */ | |
b2edc0a7 | 2573 | do { |
affa87db PB |
2574 | while (io_uring_try_cancel_requests(ctx, NULL, true)) |
2575 | cond_resched(); | |
2576 | ||
28090c13 PB |
2577 | if (ctx->sq_data) { |
2578 | struct io_sq_data *sqd = ctx->sq_data; | |
2579 | struct task_struct *tsk; | |
2580 | ||
2581 | io_sq_thread_park(sqd); | |
2582 | tsk = sqd->thread; | |
2583 | if (tsk && tsk->io_uring && tsk->io_uring->io_wq) | |
2584 | io_wq_cancel_cb(tsk->io_uring->io_wq, | |
2585 | io_cancel_ctx_cb, ctx, true); | |
2586 | io_sq_thread_unpark(sqd); | |
2587 | } | |
b5bb3a24 | 2588 | |
37f0e767 PB |
2589 | io_req_caches_free(ctx); |
2590 | ||
58d3be2c PB |
2591 | if (WARN_ON_ONCE(time_after(jiffies, timeout))) { |
2592 | /* there is little hope left, don't run it too often */ | |
2593 | interval = HZ * 60; | |
2594 | } | |
2595 | } while (!wait_for_completion_timeout(&ctx->ref_comp, interval)); | |
d56d938b | 2596 | |
7f00651a PB |
2597 | init_completion(&exit.completion); |
2598 | init_task_work(&exit.task_work, io_tctx_exit_cb); | |
2599 | exit.ctx = ctx; | |
89b5066e PB |
2600 | /* |
2601 | * Some may use context even when all refs and requests have been put, | |
2602 | * and they are free to do so while still holding uring_lock or | |
5b0a6acc | 2603 | * completion_lock, see io_req_task_submit(). Apart from other work, |
89b5066e PB |
2604 | * this lock/unlock section also waits them to finish. |
2605 | */ | |
d56d938b PB |
2606 | mutex_lock(&ctx->uring_lock); |
2607 | while (!list_empty(&ctx->tctx_list)) { | |
b5bb3a24 PB |
2608 | WARN_ON_ONCE(time_after(jiffies, timeout)); |
2609 | ||
d56d938b PB |
2610 | node = list_first_entry(&ctx->tctx_list, struct io_tctx_node, |
2611 | ctx_node); | |
7f00651a PB |
2612 | /* don't spin on a single task if cancellation failed */ |
2613 | list_rotate_left(&ctx->tctx_list); | |
d56d938b PB |
2614 | ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL); |
2615 | if (WARN_ON_ONCE(ret)) | |
2616 | continue; | |
d56d938b PB |
2617 | |
2618 | mutex_unlock(&ctx->uring_lock); | |
2619 | wait_for_completion(&exit.completion); | |
d56d938b PB |
2620 | mutex_lock(&ctx->uring_lock); |
2621 | } | |
2622 | mutex_unlock(&ctx->uring_lock); | |
79ebeaee JA |
2623 | spin_lock(&ctx->completion_lock); |
2624 | spin_unlock(&ctx->completion_lock); | |
d56d938b | 2625 | |
85faa7b8 JA |
2626 | io_ring_ctx_free(ctx); |
2627 | } | |
2628 | ||
c072481d | 2629 | static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
2b188cc1 | 2630 | { |
61cf9370 MWO |
2631 | unsigned long index; |
2632 | struct creds *creds; | |
2633 | ||
2b188cc1 JA |
2634 | mutex_lock(&ctx->uring_lock); |
2635 | percpu_ref_kill(&ctx->refs); | |
634578f8 | 2636 | if (ctx->rings) |
6c2450ae | 2637 | __io_cqring_overflow_flush(ctx, true); |
61cf9370 MWO |
2638 | xa_for_each(&ctx->personalities, index, creds) |
2639 | io_unregister_personality(ctx, index); | |
9ca9fb24 PB |
2640 | if (ctx->rings) |
2641 | io_poll_remove_all(ctx, NULL, true); | |
eb42cebb | 2642 | io_notif_unregister(ctx); |
2b188cc1 JA |
2643 | mutex_unlock(&ctx->uring_lock); |
2644 | ||
60053be8 PB |
2645 | /* failed during ring init, it couldn't have issued any requests */ |
2646 | if (ctx->rings) { | |
2647 | io_kill_timeouts(ctx, NULL, true); | |
60053be8 PB |
2648 | /* if we failed setting up the ctx, we might not have any rings */ |
2649 | io_iopoll_try_reap_events(ctx); | |
2650 | } | |
309fc03a | 2651 | |
85faa7b8 | 2652 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
2653 | /* |
2654 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
2655 | * if we're exiting a ton of rings at the same time. It just adds | |
2656 | * noise and overhead, there's no discernable change in runtime | |
2657 | * over using system_wq. | |
2658 | */ | |
2659 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
2660 | } |
2661 | ||
2662 | static int io_uring_release(struct inode *inode, struct file *file) | |
2663 | { | |
2664 | struct io_ring_ctx *ctx = file->private_data; | |
2665 | ||
2666 | file->private_data = NULL; | |
2667 | io_ring_ctx_wait_and_kill(ctx); | |
2668 | return 0; | |
2669 | } | |
2670 | ||
f6edbabb PB |
2671 | struct io_task_cancel { |
2672 | struct task_struct *task; | |
3dd0c97a | 2673 | bool all; |
f6edbabb | 2674 | }; |
f254ac04 | 2675 | |
f6edbabb | 2676 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
b711d4ea | 2677 | { |
9a472ef7 | 2678 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
f6edbabb | 2679 | struct io_task_cancel *cancel = data; |
9a472ef7 | 2680 | |
6af3f48b | 2681 | return io_match_task_safe(req, cancel->task, cancel->all); |
b711d4ea JA |
2682 | } |
2683 | ||
c072481d PB |
2684 | static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx, |
2685 | struct task_struct *task, | |
2686 | bool cancel_all) | |
b7ddce3c | 2687 | { |
e1915f76 | 2688 | struct io_defer_entry *de; |
b7ddce3c PB |
2689 | LIST_HEAD(list); |
2690 | ||
79ebeaee | 2691 | spin_lock(&ctx->completion_lock); |
b7ddce3c | 2692 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { |
6af3f48b | 2693 | if (io_match_task_safe(de->req, task, cancel_all)) { |
b7ddce3c PB |
2694 | list_cut_position(&list, &ctx->defer_list, &de->list); |
2695 | break; | |
2696 | } | |
2697 | } | |
79ebeaee | 2698 | spin_unlock(&ctx->completion_lock); |
e1915f76 PB |
2699 | if (list_empty(&list)) |
2700 | return false; | |
b7ddce3c PB |
2701 | |
2702 | while (!list_empty(&list)) { | |
2703 | de = list_first_entry(&list, struct io_defer_entry, list); | |
2704 | list_del_init(&de->list); | |
f41db273 | 2705 | io_req_complete_failed(de->req, -ECANCELED); |
b7ddce3c PB |
2706 | kfree(de); |
2707 | } | |
e1915f76 | 2708 | return true; |
b7ddce3c PB |
2709 | } |
2710 | ||
c072481d | 2711 | static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx) |
1b00764f PB |
2712 | { |
2713 | struct io_tctx_node *node; | |
2714 | enum io_wq_cancel cret; | |
2715 | bool ret = false; | |
2716 | ||
2717 | mutex_lock(&ctx->uring_lock); | |
2718 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
2719 | struct io_uring_task *tctx = node->task->io_uring; | |
2720 | ||
2721 | /* | |
2722 | * io_wq will stay alive while we hold uring_lock, because it's | |
2723 | * killed after ctx nodes, which requires to take the lock. | |
2724 | */ | |
2725 | if (!tctx || !tctx->io_wq) | |
2726 | continue; | |
2727 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true); | |
2728 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
2729 | } | |
2730 | mutex_unlock(&ctx->uring_lock); | |
2731 | ||
2732 | return ret; | |
2733 | } | |
2734 | ||
affa87db | 2735 | static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx, |
c072481d PB |
2736 | struct task_struct *task, |
2737 | bool cancel_all) | |
9936c7c2 | 2738 | { |
3dd0c97a | 2739 | struct io_task_cancel cancel = { .task = task, .all = cancel_all, }; |
1b00764f | 2740 | struct io_uring_task *tctx = task ? task->io_uring : NULL; |
affa87db PB |
2741 | enum io_wq_cancel cret; |
2742 | bool ret = false; | |
9936c7c2 | 2743 | |
60053be8 PB |
2744 | /* failed during ring init, it couldn't have issued any requests */ |
2745 | if (!ctx->rings) | |
affa87db | 2746 | return false; |
60053be8 | 2747 | |
affa87db PB |
2748 | if (!task) { |
2749 | ret |= io_uring_try_cancel_iowq(ctx); | |
2750 | } else if (tctx && tctx->io_wq) { | |
2751 | /* | |
2752 | * Cancels requests of all rings, not only @ctx, but | |
2753 | * it's fine as the task is in exit/exec. | |
2754 | */ | |
2755 | cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb, | |
2756 | &cancel, true); | |
2757 | ret |= (cret != IO_WQ_CANCEL_NOTFOUND); | |
2758 | } | |
9936c7c2 | 2759 | |
affa87db PB |
2760 | /* SQPOLL thread does its own polling */ |
2761 | if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) || | |
2762 | (ctx->sq_data && ctx->sq_data->thread == current)) { | |
2763 | while (!wq_list_empty(&ctx->iopoll_list)) { | |
2764 | io_iopoll_try_reap_events(ctx); | |
2765 | ret = true; | |
9936c7c2 | 2766 | } |
9936c7c2 | 2767 | } |
affa87db PB |
2768 | |
2769 | ret |= io_cancel_defer_files(ctx, task, cancel_all); | |
2770 | mutex_lock(&ctx->uring_lock); | |
2771 | ret |= io_poll_remove_all(ctx, task, cancel_all); | |
2772 | mutex_unlock(&ctx->uring_lock); | |
2773 | ret |= io_kill_timeouts(ctx, task, cancel_all); | |
2774 | if (task) | |
2775 | ret |= io_run_task_work(); | |
2776 | return ret; | |
9936c7c2 PB |
2777 | } |
2778 | ||
3f48cf18 | 2779 | static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked) |
521d6a73 | 2780 | { |
3f48cf18 | 2781 | if (tracked) |
9cae36a0 | 2782 | return atomic_read(&tctx->inflight_tracked); |
521d6a73 PB |
2783 | return percpu_counter_sum(&tctx->inflight); |
2784 | } | |
2785 | ||
78cc687b PB |
2786 | /* |
2787 | * Find any io_uring ctx that this task has registered or done IO on, and cancel | |
78a78060 | 2788 | * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation. |
78cc687b | 2789 | */ |
17437f31 | 2790 | __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd) |
0e9ddb39 | 2791 | { |
521d6a73 | 2792 | struct io_uring_task *tctx = current->io_uring; |
734551df | 2793 | struct io_ring_ctx *ctx; |
0e9ddb39 PB |
2794 | s64 inflight; |
2795 | DEFINE_WAIT(wait); | |
fdaf083c | 2796 | |
78cc687b PB |
2797 | WARN_ON_ONCE(sqd && sqd->thread != current); |
2798 | ||
6d042ffb PO |
2799 | if (!current->io_uring) |
2800 | return; | |
17a91051 PB |
2801 | if (tctx->io_wq) |
2802 | io_wq_exit_start(tctx->io_wq); | |
2803 | ||
0e9ddb39 PB |
2804 | atomic_inc(&tctx->in_idle); |
2805 | do { | |
affa87db PB |
2806 | bool loop = false; |
2807 | ||
e9dbe221 | 2808 | io_uring_drop_tctx_refs(current); |
0e9ddb39 | 2809 | /* read completions before cancelations */ |
78cc687b | 2810 | inflight = tctx_inflight(tctx, !cancel_all); |
0e9ddb39 PB |
2811 | if (!inflight) |
2812 | break; | |
fdaf083c | 2813 | |
78cc687b PB |
2814 | if (!sqd) { |
2815 | struct io_tctx_node *node; | |
2816 | unsigned long index; | |
0f212204 | 2817 | |
78cc687b PB |
2818 | xa_for_each(&tctx->xa, index, node) { |
2819 | /* sqpoll task will cancel all its requests */ | |
2820 | if (node->ctx->sq_data) | |
2821 | continue; | |
affa87db PB |
2822 | loop |= io_uring_try_cancel_requests(node->ctx, |
2823 | current, cancel_all); | |
78cc687b PB |
2824 | } |
2825 | } else { | |
2826 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
affa87db PB |
2827 | loop |= io_uring_try_cancel_requests(ctx, |
2828 | current, | |
2829 | cancel_all); | |
2830 | } | |
2831 | ||
2832 | if (loop) { | |
2833 | cond_resched(); | |
2834 | continue; | |
78cc687b | 2835 | } |
17a91051 | 2836 | |
78a78060 JA |
2837 | prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE); |
2838 | io_run_task_work(); | |
e9dbe221 | 2839 | io_uring_drop_tctx_refs(current); |
78a78060 | 2840 | |
0f212204 | 2841 | /* |
a1bb3cd5 PB |
2842 | * If we've seen completions, retry without waiting. This |
2843 | * avoids a race where a completion comes in before we did | |
2844 | * prepare_to_wait(). | |
0f212204 | 2845 | */ |
3dd0c97a | 2846 | if (inflight == tctx_inflight(tctx, !cancel_all)) |
a1bb3cd5 | 2847 | schedule(); |
f57555ed | 2848 | finish_wait(&tctx->wait, &wait); |
d8a6df10 | 2849 | } while (1); |
de7f1d9e | 2850 | |
8452d4a6 | 2851 | io_uring_clean_tctx(tctx); |
3dd0c97a | 2852 | if (cancel_all) { |
3cc7fdb9 PB |
2853 | /* |
2854 | * We shouldn't run task_works after cancel, so just leave | |
2855 | * ->in_idle set for normal exit. | |
2856 | */ | |
2857 | atomic_dec(&tctx->in_idle); | |
3f48cf18 PB |
2858 | /* for exec all current's requests should be gone, kill tctx */ |
2859 | __io_uring_free(current); | |
2860 | } | |
44e728b8 PB |
2861 | } |
2862 | ||
f552a27a | 2863 | void __io_uring_cancel(bool cancel_all) |
78cc687b | 2864 | { |
f552a27a | 2865 | io_uring_cancel_generic(cancel_all, NULL); |
78cc687b PB |
2866 | } |
2867 | ||
6c5c240e RP |
2868 | static void *io_uring_validate_mmap_request(struct file *file, |
2869 | loff_t pgoff, size_t sz) | |
2b188cc1 | 2870 | { |
2b188cc1 | 2871 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 2872 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
2873 | struct page *page; |
2874 | void *ptr; | |
2875 | ||
2876 | switch (offset) { | |
2877 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
2878 | case IORING_OFF_CQ_RING: |
2879 | ptr = ctx->rings; | |
2b188cc1 JA |
2880 | break; |
2881 | case IORING_OFF_SQES: | |
2882 | ptr = ctx->sq_sqes; | |
2883 | break; | |
2b188cc1 | 2884 | default: |
6c5c240e | 2885 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
2886 | } |
2887 | ||
2888 | page = virt_to_head_page(ptr); | |
a50b854e | 2889 | if (sz > page_size(page)) |
6c5c240e RP |
2890 | return ERR_PTR(-EINVAL); |
2891 | ||
2892 | return ptr; | |
2893 | } | |
2894 | ||
2895 | #ifdef CONFIG_MMU | |
2896 | ||
c072481d | 2897 | static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma) |
6c5c240e RP |
2898 | { |
2899 | size_t sz = vma->vm_end - vma->vm_start; | |
2900 | unsigned long pfn; | |
2901 | void *ptr; | |
2902 | ||
2903 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
2904 | if (IS_ERR(ptr)) | |
2905 | return PTR_ERR(ptr); | |
2b188cc1 JA |
2906 | |
2907 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
2908 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
2909 | } | |
2910 | ||
6c5c240e RP |
2911 | #else /* !CONFIG_MMU */ |
2912 | ||
2913 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
2914 | { | |
2915 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
2916 | } | |
2917 | ||
2918 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
2919 | { | |
2920 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
2921 | } | |
2922 | ||
2923 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
2924 | unsigned long addr, unsigned long len, | |
2925 | unsigned long pgoff, unsigned long flags) | |
2926 | { | |
2927 | void *ptr; | |
2928 | ||
2929 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
2930 | if (IS_ERR(ptr)) | |
2931 | return PTR_ERR(ptr); | |
2932 | ||
2933 | return (unsigned long) ptr; | |
2934 | } | |
2935 | ||
2936 | #endif /* !CONFIG_MMU */ | |
2937 | ||
f81440d3 PB |
2938 | static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz) |
2939 | { | |
2940 | if (flags & IORING_ENTER_EXT_ARG) { | |
2941 | struct io_uring_getevents_arg arg; | |
2942 | ||
2943 | if (argsz != sizeof(arg)) | |
2944 | return -EINVAL; | |
2945 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
2946 | return -EFAULT; | |
2947 | } | |
2948 | return 0; | |
2949 | } | |
2950 | ||
c73ebb68 HX |
2951 | static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz, |
2952 | struct __kernel_timespec __user **ts, | |
2953 | const sigset_t __user **sig) | |
2954 | { | |
2955 | struct io_uring_getevents_arg arg; | |
2956 | ||
2957 | /* | |
2958 | * If EXT_ARG isn't set, then we have no timespec and the argp pointer | |
2959 | * is just a pointer to the sigset_t. | |
2960 | */ | |
2961 | if (!(flags & IORING_ENTER_EXT_ARG)) { | |
2962 | *sig = (const sigset_t __user *) argp; | |
2963 | *ts = NULL; | |
2964 | return 0; | |
2965 | } | |
2966 | ||
2967 | /* | |
2968 | * EXT_ARG is set - ensure we agree on the size of it and copy in our | |
2969 | * timespec and sigset_t pointers if good. | |
2970 | */ | |
2971 | if (*argsz != sizeof(arg)) | |
2972 | return -EINVAL; | |
2973 | if (copy_from_user(&arg, argp, sizeof(arg))) | |
2974 | return -EFAULT; | |
d2347b96 DY |
2975 | if (arg.pad) |
2976 | return -EINVAL; | |
c73ebb68 HX |
2977 | *sig = u64_to_user_ptr(arg.sigmask); |
2978 | *argsz = arg.sigmask_sz; | |
2979 | *ts = u64_to_user_ptr(arg.ts); | |
2980 | return 0; | |
2981 | } | |
2982 | ||
2b188cc1 | 2983 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
c73ebb68 HX |
2984 | u32, min_complete, u32, flags, const void __user *, argp, |
2985 | size_t, argsz) | |
2b188cc1 JA |
2986 | { |
2987 | struct io_ring_ctx *ctx; | |
2b188cc1 | 2988 | struct fd f; |
33f993da | 2989 | long ret; |
2b188cc1 | 2990 | |
4c6e277c | 2991 | io_run_task_work(); |
b41e9852 | 2992 | |
33f993da | 2993 | if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
e7a6c00d JA |
2994 | IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG | |
2995 | IORING_ENTER_REGISTERED_RING))) | |
2b188cc1 JA |
2996 | return -EINVAL; |
2997 | ||
e7a6c00d JA |
2998 | /* |
2999 | * Ring fd has been registered via IORING_REGISTER_RING_FDS, we | |
3000 | * need only dereference our task private array to find it. | |
3001 | */ | |
3002 | if (flags & IORING_ENTER_REGISTERED_RING) { | |
3003 | struct io_uring_task *tctx = current->io_uring; | |
3004 | ||
3273c440 | 3005 | if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX)) |
e7a6c00d JA |
3006 | return -EINVAL; |
3007 | fd = array_index_nospec(fd, IO_RINGFD_REG_MAX); | |
3008 | f.file = tctx->registered_rings[fd]; | |
4329490a | 3009 | f.flags = 0; |
3273c440 PB |
3010 | if (unlikely(!f.file)) |
3011 | return -EBADF; | |
e7a6c00d JA |
3012 | } else { |
3013 | f = fdget(fd); | |
3273c440 PB |
3014 | if (unlikely(!f.file)) |
3015 | return -EBADF; | |
3016 | ret = -EOPNOTSUPP; | |
3017 | if (unlikely(!io_is_uring_fops(f.file))) | |
fbb8bb02 | 3018 | goto out; |
e7a6c00d | 3019 | } |
2b188cc1 | 3020 | |
2b188cc1 | 3021 | ctx = f.file->private_data; |
7e84e1c7 | 3022 | ret = -EBADFD; |
33f993da | 3023 | if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED)) |
7e84e1c7 SG |
3024 | goto out; |
3025 | ||
6c271ce2 JA |
3026 | /* |
3027 | * For SQ polling, the thread will do all submissions and completions. | |
3028 | * Just return the requested submit count, and wake the thread if | |
3029 | * we were asked to. | |
3030 | */ | |
b2a9eada | 3031 | ret = 0; |
6c271ce2 | 3032 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
90f67366 | 3033 | io_cqring_overflow_flush(ctx); |
89448c47 | 3034 | |
21f96522 JA |
3035 | if (unlikely(ctx->sq_data->thread == NULL)) { |
3036 | ret = -EOWNERDEAD; | |
04147488 | 3037 | goto out; |
21f96522 | 3038 | } |
6c271ce2 | 3039 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 3040 | wake_up(&ctx->sq_data->wait); |
d9d05217 PB |
3041 | if (flags & IORING_ENTER_SQ_WAIT) { |
3042 | ret = io_sqpoll_wait_sq(ctx); | |
3043 | if (ret) | |
3044 | goto out; | |
3045 | } | |
3e813c90 | 3046 | ret = to_submit; |
b2a9eada | 3047 | } else if (to_submit) { |
eef51daa | 3048 | ret = io_uring_add_tctx_node(ctx); |
0f212204 JA |
3049 | if (unlikely(ret)) |
3050 | goto out; | |
7c504e65 | 3051 | |
2b188cc1 | 3052 | mutex_lock(&ctx->uring_lock); |
3e813c90 DY |
3053 | ret = io_submit_sqes(ctx, to_submit); |
3054 | if (ret != to_submit) { | |
d487b43c | 3055 | mutex_unlock(&ctx->uring_lock); |
7c504e65 | 3056 | goto out; |
d487b43c PB |
3057 | } |
3058 | if ((flags & IORING_ENTER_GETEVENTS) && ctx->syscall_iopoll) | |
3059 | goto iopoll_locked; | |
3060 | mutex_unlock(&ctx->uring_lock); | |
2b188cc1 JA |
3061 | } |
3062 | if (flags & IORING_ENTER_GETEVENTS) { | |
3e813c90 | 3063 | int ret2; |
773697b6 | 3064 | if (ctx->syscall_iopoll) { |
d487b43c PB |
3065 | /* |
3066 | * We disallow the app entering submit/complete with | |
3067 | * polling, but we still need to lock the ring to | |
3068 | * prevent racing with polled issue that got punted to | |
3069 | * a workqueue. | |
3070 | */ | |
3071 | mutex_lock(&ctx->uring_lock); | |
3072 | iopoll_locked: | |
3e813c90 DY |
3073 | ret2 = io_validate_ext_arg(flags, argp, argsz); |
3074 | if (likely(!ret2)) { | |
3075 | min_complete = min(min_complete, | |
3076 | ctx->cq_entries); | |
3077 | ret2 = io_iopoll_check(ctx, min_complete); | |
d487b43c PB |
3078 | } |
3079 | mutex_unlock(&ctx->uring_lock); | |
def596e9 | 3080 | } else { |
f81440d3 PB |
3081 | const sigset_t __user *sig; |
3082 | struct __kernel_timespec __user *ts; | |
3083 | ||
3e813c90 DY |
3084 | ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig); |
3085 | if (likely(!ret2)) { | |
3086 | min_complete = min(min_complete, | |
3087 | ctx->cq_entries); | |
3088 | ret2 = io_cqring_wait(ctx, min_complete, sig, | |
3089 | argsz, ts); | |
3090 | } | |
def596e9 | 3091 | } |
c73ebb68 | 3092 | |
155bc950 | 3093 | if (!ret) { |
3e813c90 | 3094 | ret = ret2; |
2b188cc1 | 3095 | |
155bc950 DY |
3096 | /* |
3097 | * EBADR indicates that one or more CQE were dropped. | |
3098 | * Once the user has been informed we can clear the bit | |
3099 | * as they are obviously ok with those drops. | |
3100 | */ | |
3101 | if (unlikely(ret2 == -EBADR)) | |
3102 | clear_bit(IO_CHECK_CQ_DROPPED_BIT, | |
3103 | &ctx->check_cq); | |
def596e9 | 3104 | } |
2b188cc1 | 3105 | } |
7c504e65 | 3106 | out: |
4329490a | 3107 | fdput(f); |
3e813c90 | 3108 | return ret; |
2b188cc1 JA |
3109 | } |
3110 | ||
3111 | static const struct file_operations io_uring_fops = { | |
3112 | .release = io_uring_release, | |
3113 | .mmap = io_uring_mmap, | |
6c5c240e RP |
3114 | #ifndef CONFIG_MMU |
3115 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
3116 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
3117 | #endif | |
2b188cc1 | 3118 | .poll = io_uring_poll, |
bebdb65e | 3119 | #ifdef CONFIG_PROC_FS |
87ce955b | 3120 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 3121 | #endif |
2b188cc1 JA |
3122 | }; |
3123 | ||
92ac8bea JA |
3124 | bool io_is_uring_fops(struct file *file) |
3125 | { | |
3126 | return file->f_op == &io_uring_fops; | |
3127 | } | |
3128 | ||
c072481d PB |
3129 | static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx, |
3130 | struct io_uring_params *p) | |
2b188cc1 | 3131 | { |
75b28aff HV |
3132 | struct io_rings *rings; |
3133 | size_t size, sq_array_offset; | |
2b188cc1 | 3134 | |
bd740481 JA |
3135 | /* make sure these are sane, as we already accounted them */ |
3136 | ctx->sq_entries = p->sq_entries; | |
3137 | ctx->cq_entries = p->cq_entries; | |
3138 | ||
baf9cb64 | 3139 | size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset); |
75b28aff HV |
3140 | if (size == SIZE_MAX) |
3141 | return -EOVERFLOW; | |
3142 | ||
3143 | rings = io_mem_alloc(size); | |
3144 | if (!rings) | |
2b188cc1 JA |
3145 | return -ENOMEM; |
3146 | ||
75b28aff HV |
3147 | ctx->rings = rings; |
3148 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
3149 | rings->sq_ring_mask = p->sq_entries - 1; | |
3150 | rings->cq_ring_mask = p->cq_entries - 1; | |
3151 | rings->sq_ring_entries = p->sq_entries; | |
3152 | rings->cq_ring_entries = p->cq_entries; | |
2b188cc1 | 3153 | |
ebdeb7c0 JA |
3154 | if (p->flags & IORING_SETUP_SQE128) |
3155 | size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries); | |
3156 | else | |
3157 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
3158 | if (size == SIZE_MAX) { |
3159 | io_mem_free(ctx->rings); | |
3160 | ctx->rings = NULL; | |
2b188cc1 | 3161 | return -EOVERFLOW; |
eb065d30 | 3162 | } |
2b188cc1 JA |
3163 | |
3164 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
3165 | if (!ctx->sq_sqes) { |
3166 | io_mem_free(ctx->rings); | |
3167 | ctx->rings = NULL; | |
2b188cc1 | 3168 | return -ENOMEM; |
eb065d30 | 3169 | } |
2b188cc1 | 3170 | |
2b188cc1 JA |
3171 | return 0; |
3172 | } | |
3173 | ||
9faadcc8 PB |
3174 | static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file) |
3175 | { | |
3176 | int ret, fd; | |
3177 | ||
3178 | fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
3179 | if (fd < 0) | |
3180 | return fd; | |
3181 | ||
97bbdc06 | 3182 | ret = __io_uring_add_tctx_node(ctx, false); |
9faadcc8 PB |
3183 | if (ret) { |
3184 | put_unused_fd(fd); | |
3185 | return ret; | |
3186 | } | |
3187 | fd_install(fd, file); | |
3188 | return fd; | |
3189 | } | |
3190 | ||
2b188cc1 JA |
3191 | /* |
3192 | * Allocate an anonymous fd, this is what constitutes the application | |
3193 | * visible backing of an io_uring instance. The application mmaps this | |
3194 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
3195 | * we have to tie this fd to a socket for file garbage collection purposes. | |
3196 | */ | |
9faadcc8 | 3197 | static struct file *io_uring_get_file(struct io_ring_ctx *ctx) |
2b188cc1 JA |
3198 | { |
3199 | struct file *file; | |
9faadcc8 | 3200 | #if defined(CONFIG_UNIX) |
2b188cc1 JA |
3201 | int ret; |
3202 | ||
2b188cc1 JA |
3203 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, |
3204 | &ctx->ring_sock); | |
3205 | if (ret) | |
9faadcc8 | 3206 | return ERR_PTR(ret); |
2b188cc1 JA |
3207 | #endif |
3208 | ||
91a9ab7c PM |
3209 | file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx, |
3210 | O_RDWR | O_CLOEXEC, NULL); | |
2b188cc1 | 3211 | #if defined(CONFIG_UNIX) |
9faadcc8 PB |
3212 | if (IS_ERR(file)) { |
3213 | sock_release(ctx->ring_sock); | |
3214 | ctx->ring_sock = NULL; | |
3215 | } else { | |
3216 | ctx->ring_sock->file = file; | |
0f212204 | 3217 | } |
2b188cc1 | 3218 | #endif |
9faadcc8 | 3219 | return file; |
2b188cc1 JA |
3220 | } |
3221 | ||
c072481d PB |
3222 | static __cold int io_uring_create(unsigned entries, struct io_uring_params *p, |
3223 | struct io_uring_params __user *params) | |
2b188cc1 | 3224 | { |
2b188cc1 | 3225 | struct io_ring_ctx *ctx; |
9faadcc8 | 3226 | struct file *file; |
2b188cc1 JA |
3227 | int ret; |
3228 | ||
8110c1a6 | 3229 | if (!entries) |
2b188cc1 | 3230 | return -EINVAL; |
8110c1a6 JA |
3231 | if (entries > IORING_MAX_ENTRIES) { |
3232 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3233 | return -EINVAL; | |
3234 | entries = IORING_MAX_ENTRIES; | |
3235 | } | |
2b188cc1 JA |
3236 | |
3237 | /* | |
3238 | * Use twice as many entries for the CQ ring. It's possible for the | |
3239 | * application to drive a higher depth than the size of the SQ ring, | |
3240 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
3241 | * some flexibility in overcommitting a bit. If the application has |
3242 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
3243 | * of CQ ring entries manually. | |
2b188cc1 JA |
3244 | */ |
3245 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
3246 | if (p->flags & IORING_SETUP_CQSIZE) { |
3247 | /* | |
3248 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
3249 | * to a power-of-two, if it isn't already. We do NOT impose | |
3250 | * any cq vs sq ring sizing. | |
3251 | */ | |
eb2667b3 | 3252 | if (!p->cq_entries) |
33a107f0 | 3253 | return -EINVAL; |
8110c1a6 JA |
3254 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
3255 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
3256 | return -EINVAL; | |
3257 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
3258 | } | |
eb2667b3 JQ |
3259 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
3260 | if (p->cq_entries < p->sq_entries) | |
3261 | return -EINVAL; | |
33a107f0 JA |
3262 | } else { |
3263 | p->cq_entries = 2 * p->sq_entries; | |
3264 | } | |
2b188cc1 | 3265 | |
2b188cc1 | 3266 | ctx = io_ring_ctx_alloc(p); |
62e398be | 3267 | if (!ctx) |
2b188cc1 | 3268 | return -ENOMEM; |
773697b6 PB |
3269 | |
3270 | /* | |
3271 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
3272 | * space applications don't need to do io completion events | |
3273 | * polling again, they can rely on io_sq_thread to do polling | |
3274 | * work, which can reduce cpu usage and uring_lock contention. | |
3275 | */ | |
3276 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
3277 | !(ctx->flags & IORING_SETUP_SQPOLL)) | |
3278 | ctx->syscall_iopoll = 1; | |
3279 | ||
2b188cc1 | 3280 | ctx->compat = in_compat_syscall(); |
62e398be JA |
3281 | if (!capable(CAP_IPC_LOCK)) |
3282 | ctx->user = get_uid(current_user()); | |
2aede0e4 | 3283 | |
9f010507 | 3284 | /* |
e1169f06 JA |
3285 | * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if |
3286 | * COOP_TASKRUN is set, then IPIs are never needed by the app. | |
9f010507 | 3287 | */ |
e1169f06 JA |
3288 | ret = -EINVAL; |
3289 | if (ctx->flags & IORING_SETUP_SQPOLL) { | |
3290 | /* IPI related flags don't make sense with SQPOLL */ | |
ef060ea9 JA |
3291 | if (ctx->flags & (IORING_SETUP_COOP_TASKRUN | |
3292 | IORING_SETUP_TASKRUN_FLAG)) | |
e1169f06 | 3293 | goto err; |
9f010507 | 3294 | ctx->notify_method = TWA_SIGNAL_NO_IPI; |
e1169f06 JA |
3295 | } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) { |
3296 | ctx->notify_method = TWA_SIGNAL_NO_IPI; | |
3297 | } else { | |
ef060ea9 JA |
3298 | if (ctx->flags & IORING_SETUP_TASKRUN_FLAG) |
3299 | goto err; | |
9f010507 | 3300 | ctx->notify_method = TWA_SIGNAL; |
e1169f06 | 3301 | } |
9f010507 | 3302 | |
2aede0e4 JA |
3303 | /* |
3304 | * This is just grabbed for accounting purposes. When a process exits, | |
3305 | * the mm is exited and dropped before the files, hence we need to hang | |
3306 | * on to this mm purely for the purposes of being able to unaccount | |
3307 | * memory (locked/pinned vm). It's not used for anything else. | |
3308 | */ | |
6b7898eb | 3309 | mmgrab(current->mm); |
2aede0e4 | 3310 | ctx->mm_account = current->mm; |
6b7898eb | 3311 | |
2b188cc1 JA |
3312 | ret = io_allocate_scq_urings(ctx, p); |
3313 | if (ret) | |
3314 | goto err; | |
3315 | ||
7e84e1c7 | 3316 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
3317 | if (ret) |
3318 | goto err; | |
eae071c9 | 3319 | /* always set a rsrc node */ |
47b228ce PB |
3320 | ret = io_rsrc_node_switch_start(ctx); |
3321 | if (ret) | |
3322 | goto err; | |
eae071c9 | 3323 | io_rsrc_node_switch(ctx, NULL); |
2b188cc1 | 3324 | |
2b188cc1 | 3325 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
3326 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
3327 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
3328 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
3329 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
3330 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
3331 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
3332 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
3333 | |
3334 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
3335 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
3336 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
3337 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
3338 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
3339 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
3340 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 3341 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 3342 | |
7f13657d XW |
3343 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
3344 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 | 3345 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
c73ebb68 | 3346 | IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED | |
9690557e | 3347 | IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS | |
c4212f3e JA |
3348 | IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP | |
3349 | IORING_FEAT_LINKED_FILE; | |
7f13657d XW |
3350 | |
3351 | if (copy_to_user(params, p, sizeof(*p))) { | |
3352 | ret = -EFAULT; | |
3353 | goto err; | |
3354 | } | |
d1719f70 | 3355 | |
9faadcc8 PB |
3356 | file = io_uring_get_file(ctx); |
3357 | if (IS_ERR(file)) { | |
3358 | ret = PTR_ERR(file); | |
3359 | goto err; | |
3360 | } | |
3361 | ||
044c1ab3 JA |
3362 | /* |
3363 | * Install ring fd as the very last thing, so we don't risk someone | |
3364 | * having closed it before we finish setup | |
3365 | */ | |
9faadcc8 PB |
3366 | ret = io_uring_install_fd(ctx, file); |
3367 | if (ret < 0) { | |
3368 | /* fput will clean it up */ | |
3369 | fput(file); | |
3370 | return ret; | |
3371 | } | |
044c1ab3 | 3372 | |
c826bd7a | 3373 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
3374 | return ret; |
3375 | err: | |
3376 | io_ring_ctx_wait_and_kill(ctx); | |
3377 | return ret; | |
3378 | } | |
3379 | ||
3380 | /* | |
3381 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
3382 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
3383 | * params structure passed in. | |
3384 | */ | |
3385 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
3386 | { | |
3387 | struct io_uring_params p; | |
2b188cc1 JA |
3388 | int i; |
3389 | ||
3390 | if (copy_from_user(&p, params, sizeof(p))) | |
3391 | return -EFAULT; | |
3392 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
3393 | if (p.resv[i]) | |
3394 | return -EINVAL; | |
3395 | } | |
3396 | ||
6c271ce2 | 3397 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 3398 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 | 3399 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
e1169f06 | 3400 | IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL | |
ebdeb7c0 | 3401 | IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG | |
97bbdc06 PB |
3402 | IORING_SETUP_SQE128 | IORING_SETUP_CQE32 | |
3403 | IORING_SETUP_SINGLE_ISSUER)) | |
2b188cc1 JA |
3404 | return -EINVAL; |
3405 | ||
ef060ea9 | 3406 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
3407 | } |
3408 | ||
3409 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
3410 | struct io_uring_params __user *, params) | |
3411 | { | |
3412 | return io_uring_setup(entries, params); | |
3413 | } | |
3414 | ||
c072481d PB |
3415 | static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg, |
3416 | unsigned nr_args) | |
66f4af93 JA |
3417 | { |
3418 | struct io_uring_probe *p; | |
3419 | size_t size; | |
3420 | int i, ret; | |
3421 | ||
3422 | size = struct_size(p, ops, nr_args); | |
3423 | if (size == SIZE_MAX) | |
3424 | return -EOVERFLOW; | |
3425 | p = kzalloc(size, GFP_KERNEL); | |
3426 | if (!p) | |
3427 | return -ENOMEM; | |
3428 | ||
3429 | ret = -EFAULT; | |
3430 | if (copy_from_user(p, arg, size)) | |
3431 | goto out; | |
3432 | ret = -EINVAL; | |
3433 | if (memchr_inv(p, 0, size)) | |
3434 | goto out; | |
3435 | ||
3436 | p->last_op = IORING_OP_LAST - 1; | |
3437 | if (nr_args > IORING_OP_LAST) | |
3438 | nr_args = IORING_OP_LAST; | |
3439 | ||
3440 | for (i = 0; i < nr_args; i++) { | |
3441 | p->ops[i].op = i; | |
3442 | if (!io_op_defs[i].not_supported) | |
3443 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
3444 | } | |
3445 | p->ops_len = i; | |
3446 | ||
3447 | ret = 0; | |
3448 | if (copy_to_user(arg, p, size)) | |
3449 | ret = -EFAULT; | |
3450 | out: | |
3451 | kfree(p); | |
3452 | return ret; | |
3453 | } | |
3454 | ||
071698e1 JA |
3455 | static int io_register_personality(struct io_ring_ctx *ctx) |
3456 | { | |
4379bf8b | 3457 | const struct cred *creds; |
61cf9370 | 3458 | u32 id; |
1e6fa521 | 3459 | int ret; |
071698e1 | 3460 | |
4379bf8b | 3461 | creds = get_current_cred(); |
1e6fa521 | 3462 | |
61cf9370 MWO |
3463 | ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds, |
3464 | XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL); | |
a30f895a JA |
3465 | if (ret < 0) { |
3466 | put_cred(creds); | |
3467 | return ret; | |
3468 | } | |
3469 | return id; | |
071698e1 JA |
3470 | } |
3471 | ||
c072481d PB |
3472 | static __cold int io_register_restrictions(struct io_ring_ctx *ctx, |
3473 | void __user *arg, unsigned int nr_args) | |
21b55dbc SG |
3474 | { |
3475 | struct io_uring_restriction *res; | |
3476 | size_t size; | |
3477 | int i, ret; | |
3478 | ||
7e84e1c7 SG |
3479 | /* Restrictions allowed only if rings started disabled */ |
3480 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3481 | return -EBADFD; | |
3482 | ||
21b55dbc | 3483 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 3484 | if (ctx->restrictions.registered) |
21b55dbc SG |
3485 | return -EBUSY; |
3486 | ||
3487 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
3488 | return -EINVAL; | |
3489 | ||
3490 | size = array_size(nr_args, sizeof(*res)); | |
3491 | if (size == SIZE_MAX) | |
3492 | return -EOVERFLOW; | |
3493 | ||
3494 | res = memdup_user(arg, size); | |
3495 | if (IS_ERR(res)) | |
3496 | return PTR_ERR(res); | |
3497 | ||
3498 | ret = 0; | |
3499 | ||
3500 | for (i = 0; i < nr_args; i++) { | |
3501 | switch (res[i].opcode) { | |
3502 | case IORING_RESTRICTION_REGISTER_OP: | |
3503 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
3504 | ret = -EINVAL; | |
3505 | goto out; | |
3506 | } | |
3507 | ||
3508 | __set_bit(res[i].register_op, | |
3509 | ctx->restrictions.register_op); | |
3510 | break; | |
3511 | case IORING_RESTRICTION_SQE_OP: | |
3512 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
3513 | ret = -EINVAL; | |
3514 | goto out; | |
3515 | } | |
3516 | ||
3517 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
3518 | break; | |
3519 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
3520 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
3521 | break; | |
3522 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
3523 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
3524 | break; | |
3525 | default: | |
3526 | ret = -EINVAL; | |
3527 | goto out; | |
3528 | } | |
3529 | } | |
3530 | ||
3531 | out: | |
3532 | /* Reset all restrictions if an error happened */ | |
3533 | if (ret != 0) | |
3534 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
3535 | else | |
7e84e1c7 | 3536 | ctx->restrictions.registered = true; |
21b55dbc SG |
3537 | |
3538 | kfree(res); | |
3539 | return ret; | |
3540 | } | |
3541 | ||
7e84e1c7 SG |
3542 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
3543 | { | |
3544 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
3545 | return -EBADFD; | |
3546 | ||
3547 | if (ctx->restrictions.registered) | |
3548 | ctx->restricted = 1; | |
3549 | ||
0298ef96 PB |
3550 | ctx->flags &= ~IORING_SETUP_R_DISABLED; |
3551 | if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait)) | |
3552 | wake_up(&ctx->sq_data->wait); | |
7e84e1c7 SG |
3553 | return 0; |
3554 | } | |
3555 | ||
c072481d PB |
3556 | static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx, |
3557 | void __user *arg, unsigned len) | |
fe76421d JA |
3558 | { |
3559 | struct io_uring_task *tctx = current->io_uring; | |
3560 | cpumask_var_t new_mask; | |
3561 | int ret; | |
3562 | ||
3563 | if (!tctx || !tctx->io_wq) | |
3564 | return -EINVAL; | |
3565 | ||
3566 | if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) | |
3567 | return -ENOMEM; | |
3568 | ||
3569 | cpumask_clear(new_mask); | |
3570 | if (len > cpumask_size()) | |
3571 | len = cpumask_size(); | |
3572 | ||
0f5e4b83 ES |
3573 | if (in_compat_syscall()) { |
3574 | ret = compat_get_bitmap(cpumask_bits(new_mask), | |
3575 | (const compat_ulong_t __user *)arg, | |
3576 | len * 8 /* CHAR_BIT */); | |
3577 | } else { | |
3578 | ret = copy_from_user(new_mask, arg, len); | |
3579 | } | |
3580 | ||
3581 | if (ret) { | |
fe76421d JA |
3582 | free_cpumask_var(new_mask); |
3583 | return -EFAULT; | |
3584 | } | |
3585 | ||
3586 | ret = io_wq_cpu_affinity(tctx->io_wq, new_mask); | |
3587 | free_cpumask_var(new_mask); | |
3588 | return ret; | |
3589 | } | |
3590 | ||
c072481d | 3591 | static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx) |
fe76421d JA |
3592 | { |
3593 | struct io_uring_task *tctx = current->io_uring; | |
3594 | ||
3595 | if (!tctx || !tctx->io_wq) | |
3596 | return -EINVAL; | |
3597 | ||
3598 | return io_wq_cpu_affinity(tctx->io_wq, NULL); | |
3599 | } | |
3600 | ||
c072481d PB |
3601 | static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx, |
3602 | void __user *arg) | |
b22fa62a | 3603 | __must_hold(&ctx->uring_lock) |
2e480058 | 3604 | { |
b22fa62a | 3605 | struct io_tctx_node *node; |
fa84693b JA |
3606 | struct io_uring_task *tctx = NULL; |
3607 | struct io_sq_data *sqd = NULL; | |
2e480058 JA |
3608 | __u32 new_count[2]; |
3609 | int i, ret; | |
3610 | ||
2e480058 JA |
3611 | if (copy_from_user(new_count, arg, sizeof(new_count))) |
3612 | return -EFAULT; | |
3613 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
3614 | if (new_count[i] > INT_MAX) | |
3615 | return -EINVAL; | |
3616 | ||
fa84693b JA |
3617 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3618 | sqd = ctx->sq_data; | |
3619 | if (sqd) { | |
009ad9f0 JA |
3620 | /* |
3621 | * Observe the correct sqd->lock -> ctx->uring_lock | |
3622 | * ordering. Fine to drop uring_lock here, we hold | |
3623 | * a ref to the ctx. | |
3624 | */ | |
41d3a6bd | 3625 | refcount_inc(&sqd->refs); |
009ad9f0 | 3626 | mutex_unlock(&ctx->uring_lock); |
fa84693b | 3627 | mutex_lock(&sqd->lock); |
009ad9f0 | 3628 | mutex_lock(&ctx->uring_lock); |
41d3a6bd JA |
3629 | if (sqd->thread) |
3630 | tctx = sqd->thread->io_uring; | |
fa84693b JA |
3631 | } |
3632 | } else { | |
3633 | tctx = current->io_uring; | |
3634 | } | |
3635 | ||
e139a1ec | 3636 | BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits)); |
fa84693b | 3637 | |
bad119b9 PB |
3638 | for (i = 0; i < ARRAY_SIZE(new_count); i++) |
3639 | if (new_count[i]) | |
3640 | ctx->iowq_limits[i] = new_count[i]; | |
e139a1ec PB |
3641 | ctx->iowq_limits_set = true; |
3642 | ||
e139a1ec PB |
3643 | if (tctx && tctx->io_wq) { |
3644 | ret = io_wq_max_workers(tctx->io_wq, new_count); | |
3645 | if (ret) | |
3646 | goto err; | |
3647 | } else { | |
3648 | memset(new_count, 0, sizeof(new_count)); | |
3649 | } | |
fa84693b | 3650 | |
41d3a6bd | 3651 | if (sqd) { |
fa84693b | 3652 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
3653 | io_put_sq_data(sqd); |
3654 | } | |
2e480058 JA |
3655 | |
3656 | if (copy_to_user(arg, new_count, sizeof(new_count))) | |
3657 | return -EFAULT; | |
3658 | ||
b22fa62a PB |
3659 | /* that's it for SQPOLL, only the SQPOLL task creates requests */ |
3660 | if (sqd) | |
3661 | return 0; | |
3662 | ||
3663 | /* now propagate the restriction to all registered users */ | |
3664 | list_for_each_entry(node, &ctx->tctx_list, ctx_node) { | |
3665 | struct io_uring_task *tctx = node->task->io_uring; | |
3666 | ||
3667 | if (WARN_ON_ONCE(!tctx->io_wq)) | |
3668 | continue; | |
3669 | ||
3670 | for (i = 0; i < ARRAY_SIZE(new_count); i++) | |
3671 | new_count[i] = ctx->iowq_limits[i]; | |
3672 | /* ignore errors, it always returns zero anyway */ | |
3673 | (void)io_wq_max_workers(tctx->io_wq, new_count); | |
3674 | } | |
2e480058 | 3675 | return 0; |
fa84693b | 3676 | err: |
41d3a6bd | 3677 | if (sqd) { |
fa84693b | 3678 | mutex_unlock(&sqd->lock); |
41d3a6bd JA |
3679 | io_put_sq_data(sqd); |
3680 | } | |
fa84693b | 3681 | return ret; |
2e480058 JA |
3682 | } |
3683 | ||
edafccee JA |
3684 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
3685 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
3686 | __releases(ctx->uring_lock) |
3687 | __acquires(ctx->uring_lock) | |
edafccee JA |
3688 | { |
3689 | int ret; | |
3690 | ||
35fa71a0 | 3691 | /* |
fbb8bb02 PB |
3692 | * We don't quiesce the refs for register anymore and so it can't be |
3693 | * dying as we're holding a file ref here. | |
35fa71a0 | 3694 | */ |
fbb8bb02 | 3695 | if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs))) |
35fa71a0 JA |
3696 | return -ENXIO; |
3697 | ||
75c4021a PB |
3698 | if (ctx->restricted) { |
3699 | if (opcode >= IORING_REGISTER_LAST) | |
3700 | return -EINVAL; | |
3701 | opcode = array_index_nospec(opcode, IORING_REGISTER_LAST); | |
3702 | if (!test_bit(opcode, ctx->restrictions.register_op)) | |
3703 | return -EACCES; | |
3704 | } | |
3705 | ||
edafccee JA |
3706 | switch (opcode) { |
3707 | case IORING_REGISTER_BUFFERS: | |
0184f08e PB |
3708 | ret = -EFAULT; |
3709 | if (!arg) | |
3710 | break; | |
634d00df | 3711 | ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL); |
edafccee JA |
3712 | break; |
3713 | case IORING_UNREGISTER_BUFFERS: | |
3714 | ret = -EINVAL; | |
3715 | if (arg || nr_args) | |
3716 | break; | |
0a96bbe4 | 3717 | ret = io_sqe_buffers_unregister(ctx); |
edafccee | 3718 | break; |
6b06314c | 3719 | case IORING_REGISTER_FILES: |
a8da73a3 JA |
3720 | ret = -EFAULT; |
3721 | if (!arg) | |
3722 | break; | |
792e3582 | 3723 | ret = io_sqe_files_register(ctx, arg, nr_args, NULL); |
6b06314c JA |
3724 | break; |
3725 | case IORING_UNREGISTER_FILES: | |
3726 | ret = -EINVAL; | |
3727 | if (arg || nr_args) | |
3728 | break; | |
3729 | ret = io_sqe_files_unregister(ctx); | |
3730 | break; | |
c3a31e60 | 3731 | case IORING_REGISTER_FILES_UPDATE: |
c3bdad02 | 3732 | ret = io_register_files_update(ctx, arg, nr_args); |
c3a31e60 | 3733 | break; |
9b402849 JA |
3734 | case IORING_REGISTER_EVENTFD: |
3735 | ret = -EINVAL; | |
3736 | if (nr_args != 1) | |
3737 | break; | |
c75312dd UA |
3738 | ret = io_eventfd_register(ctx, arg, 0); |
3739 | break; | |
3740 | case IORING_REGISTER_EVENTFD_ASYNC: | |
3741 | ret = -EINVAL; | |
3742 | if (nr_args != 1) | |
f2842ab5 | 3743 | break; |
c75312dd | 3744 | ret = io_eventfd_register(ctx, arg, 1); |
9b402849 JA |
3745 | break; |
3746 | case IORING_UNREGISTER_EVENTFD: | |
3747 | ret = -EINVAL; | |
3748 | if (arg || nr_args) | |
3749 | break; | |
3750 | ret = io_eventfd_unregister(ctx); | |
3751 | break; | |
66f4af93 JA |
3752 | case IORING_REGISTER_PROBE: |
3753 | ret = -EINVAL; | |
3754 | if (!arg || nr_args > 256) | |
3755 | break; | |
3756 | ret = io_probe(ctx, arg, nr_args); | |
3757 | break; | |
071698e1 JA |
3758 | case IORING_REGISTER_PERSONALITY: |
3759 | ret = -EINVAL; | |
3760 | if (arg || nr_args) | |
3761 | break; | |
3762 | ret = io_register_personality(ctx); | |
3763 | break; | |
3764 | case IORING_UNREGISTER_PERSONALITY: | |
3765 | ret = -EINVAL; | |
3766 | if (arg) | |
3767 | break; | |
3768 | ret = io_unregister_personality(ctx, nr_args); | |
3769 | break; | |
7e84e1c7 SG |
3770 | case IORING_REGISTER_ENABLE_RINGS: |
3771 | ret = -EINVAL; | |
3772 | if (arg || nr_args) | |
3773 | break; | |
3774 | ret = io_register_enable_rings(ctx); | |
3775 | break; | |
21b55dbc SG |
3776 | case IORING_REGISTER_RESTRICTIONS: |
3777 | ret = io_register_restrictions(ctx, arg, nr_args); | |
3778 | break; | |
992da01a PB |
3779 | case IORING_REGISTER_FILES2: |
3780 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE); | |
3781 | break; | |
3782 | case IORING_REGISTER_FILES_UPDATE2: | |
3783 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
3784 | IORING_RSRC_FILE); | |
3785 | break; | |
3786 | case IORING_REGISTER_BUFFERS2: | |
3787 | ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER); | |
792e3582 | 3788 | break; |
992da01a PB |
3789 | case IORING_REGISTER_BUFFERS_UPDATE: |
3790 | ret = io_register_rsrc_update(ctx, arg, nr_args, | |
3791 | IORING_RSRC_BUFFER); | |
c3bdad02 | 3792 | break; |
fe76421d JA |
3793 | case IORING_REGISTER_IOWQ_AFF: |
3794 | ret = -EINVAL; | |
3795 | if (!arg || !nr_args) | |
3796 | break; | |
3797 | ret = io_register_iowq_aff(ctx, arg, nr_args); | |
3798 | break; | |
3799 | case IORING_UNREGISTER_IOWQ_AFF: | |
3800 | ret = -EINVAL; | |
3801 | if (arg || nr_args) | |
3802 | break; | |
3803 | ret = io_unregister_iowq_aff(ctx); | |
3804 | break; | |
2e480058 JA |
3805 | case IORING_REGISTER_IOWQ_MAX_WORKERS: |
3806 | ret = -EINVAL; | |
3807 | if (!arg || nr_args != 2) | |
3808 | break; | |
3809 | ret = io_register_iowq_max_workers(ctx, arg); | |
3810 | break; | |
e7a6c00d JA |
3811 | case IORING_REGISTER_RING_FDS: |
3812 | ret = io_ringfd_register(ctx, arg, nr_args); | |
3813 | break; | |
3814 | case IORING_UNREGISTER_RING_FDS: | |
3815 | ret = io_ringfd_unregister(ctx, arg, nr_args); | |
3816 | break; | |
c7fb1942 JA |
3817 | case IORING_REGISTER_PBUF_RING: |
3818 | ret = -EINVAL; | |
3819 | if (!arg || nr_args != 1) | |
3820 | break; | |
3821 | ret = io_register_pbuf_ring(ctx, arg); | |
3822 | break; | |
3823 | case IORING_UNREGISTER_PBUF_RING: | |
3824 | ret = -EINVAL; | |
3825 | if (!arg || nr_args != 1) | |
3826 | break; | |
3827 | ret = io_unregister_pbuf_ring(ctx, arg); | |
3828 | break; | |
78a861b9 JA |
3829 | case IORING_REGISTER_SYNC_CANCEL: |
3830 | ret = -EINVAL; | |
3831 | if (!arg || nr_args != 1) | |
3832 | break; | |
3833 | ret = io_sync_cancel(ctx, arg); | |
3834 | break; | |
6e73dffb PB |
3835 | case IORING_REGISTER_FILE_ALLOC_RANGE: |
3836 | ret = -EINVAL; | |
3837 | if (!arg || nr_args) | |
3838 | break; | |
3839 | ret = io_register_file_alloc_range(ctx, arg); | |
3840 | break; | |
bc24d6bd PB |
3841 | case IORING_REGISTER_NOTIFIERS: |
3842 | ret = io_notif_register(ctx, arg, nr_args); | |
3843 | break; | |
3844 | case IORING_UNREGISTER_NOTIFIERS: | |
3845 | ret = -EINVAL; | |
3846 | if (arg || nr_args) | |
3847 | break; | |
3848 | ret = io_notif_unregister(ctx); | |
3849 | break; | |
edafccee JA |
3850 | default: |
3851 | ret = -EINVAL; | |
3852 | break; | |
3853 | } | |
3854 | ||
edafccee JA |
3855 | return ret; |
3856 | } | |
3857 | ||
3858 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
3859 | void __user *, arg, unsigned int, nr_args) | |
3860 | { | |
3861 | struct io_ring_ctx *ctx; | |
3862 | long ret = -EBADF; | |
3863 | struct fd f; | |
3864 | ||
3865 | f = fdget(fd); | |
3866 | if (!f.file) | |
3867 | return -EBADF; | |
3868 | ||
3869 | ret = -EOPNOTSUPP; | |
e5550a14 | 3870 | if (!io_is_uring_fops(f.file)) |
edafccee JA |
3871 | goto out_fput; |
3872 | ||
3873 | ctx = f.file->private_data; | |
3874 | ||
b6c23dd5 PB |
3875 | io_run_task_work(); |
3876 | ||
edafccee JA |
3877 | mutex_lock(&ctx->uring_lock); |
3878 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
3879 | mutex_unlock(&ctx->uring_lock); | |
2757be22 | 3880 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret); |
edafccee JA |
3881 | out_fput: |
3882 | fdput(f); | |
3883 | return ret; | |
3884 | } | |
3885 | ||
2b188cc1 JA |
3886 | static int __init io_uring_init(void) |
3887 | { | |
9c71d39a | 3888 | #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \ |
d7f62e82 | 3889 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ |
9c71d39a | 3890 | BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \ |
d7f62e82 SM |
3891 | } while (0) |
3892 | ||
3893 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9c71d39a SM |
3894 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename) |
3895 | #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \ | |
3896 | __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename) | |
d7f62e82 SM |
3897 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); |
3898 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
3899 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
3900 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
3901 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
3902 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
3903 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9c71d39a SM |
3904 | BUILD_BUG_SQE_ELEM(8, __u32, cmd_op); |
3905 | BUILD_BUG_SQE_ELEM(12, __u32, __pad1); | |
d7f62e82 | 3906 | BUILD_BUG_SQE_ELEM(16, __u64, addr); |
7d67af2c | 3907 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
3908 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
3909 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
3910 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
3911 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
3912 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
3913 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
3914 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
3915 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
3916 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
3917 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
3918 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
3919 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
3920 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
3921 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
3922 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 3923 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
9c71d39a SM |
3924 | BUILD_BUG_SQE_ELEM(28, __u32, rename_flags); |
3925 | BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags); | |
3926 | BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags); | |
3927 | BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags); | |
3928 | BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags); | |
d7f62e82 SM |
3929 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
3930 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
16340eab | 3931 | BUILD_BUG_SQE_ELEM(40, __u16, buf_group); |
d7f62e82 | 3932 | BUILD_BUG_SQE_ELEM(42, __u16, personality); |
7d67af2c | 3933 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
b9445598 | 3934 | BUILD_BUG_SQE_ELEM(44, __u32, file_index); |
9c71d39a SM |
3935 | BUILD_BUG_SQE_ELEM(44, __u16, notification_idx); |
3936 | BUILD_BUG_SQE_ELEM(46, __u16, addr_len); | |
e9621e2b | 3937 | BUILD_BUG_SQE_ELEM(48, __u64, addr3); |
9c71d39a SM |
3938 | BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd); |
3939 | BUILD_BUG_SQE_ELEM(56, __u64, __pad2); | |
d7f62e82 | 3940 | |
b0d658ec PB |
3941 | BUILD_BUG_ON(sizeof(struct io_uring_files_update) != |
3942 | sizeof(struct io_uring_rsrc_update)); | |
3943 | BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) > | |
3944 | sizeof(struct io_uring_rsrc_update2)); | |
90499ad0 PB |
3945 | |
3946 | /* ->buf_index is u16 */ | |
c7fb1942 JA |
3947 | BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0); |
3948 | BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) != | |
3949 | offsetof(struct io_uring_buf_ring, tail)); | |
90499ad0 | 3950 | |
b0d658ec PB |
3951 | /* should fit into one byte */ |
3952 | BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8)); | |
68fe256a PB |
3953 | BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8)); |
3954 | BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS); | |
b0d658ec | 3955 | |
32c2d33e | 3956 | BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int)); |
16340eab | 3957 | |
3a4b89a2 JA |
3958 | BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32)); |
3959 | ||
d9b57aa3 | 3960 | io_uring_optable_init(); |
0702e536 | 3961 | |
91f245d5 JA |
3962 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC | |
3963 | SLAB_ACCOUNT); | |
2b188cc1 JA |
3964 | return 0; |
3965 | }; | |
3966 | __initcall(io_uring_init); |