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
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
14 * through a control-dependency in io_get_cqe (smp_store_release to
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.
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
40 * Copyright (c) 2018-2019 Christoph Hellwig
41 */
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <net/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
49 #include <linux/bits.h>
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>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/bvec.h>
60 #include <linux/net.h>
61 #include <net/sock.h>
62 #include <net/af_unix.h>
63 #include <net/scm.h>
64 #include <linux/anon_inodes.h>
65 #include <linux/sched/mm.h>
66 #include <linux/uaccess.h>
67 #include <linux/nospec.h>
68 #include <linux/highmem.h>
69 #include <linux/fsnotify.h>
70 #include <linux/fadvise.h>
71 #include <linux/task_work.h>
72 #include <linux/io_uring.h>
73 #include <linux/audit.h>
74 #include <linux/security.h>
75
76 #define CREATE_TRACE_POINTS
77 #include <trace/events/io_uring.h>
78
79 #include <uapi/linux/io_uring.h>
80
81 #include "io-wq.h"
82
83 #include "io_uring.h"
84 #include "opdef.h"
85 #include "refs.h"
86 #include "tctx.h"
87 #include "sqpoll.h"
88 #include "fdinfo.h"
89 #include "kbuf.h"
90 #include "rsrc.h"
91 #include "cancel.h"
92 #include "net.h"
93 #include "notif.h"
94
95 #include "timeout.h"
96 #include "poll.h"
97 #include "alloc_cache.h"
98
99 #define IORING_MAX_ENTRIES 32768
100 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
101
102 #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \
103 IORING_REGISTER_LAST + IORING_OP_LAST)
104
105 #define SQE_COMMON_FLAGS (IOSQE_FIXED_FILE | IOSQE_IO_LINK | \
106 IOSQE_IO_HARDLINK | IOSQE_ASYNC)
107
108 #define SQE_VALID_FLAGS (SQE_COMMON_FLAGS | IOSQE_BUFFER_SELECT | \
109 IOSQE_IO_DRAIN | IOSQE_CQE_SKIP_SUCCESS)
110
111 #define IO_REQ_CLEAN_FLAGS (REQ_F_BUFFER_SELECTED | REQ_F_NEED_CLEANUP | \
112 REQ_F_POLLED | REQ_F_INFLIGHT | REQ_F_CREDS | \
113 REQ_F_ASYNC_DATA)
114
115 #define IO_REQ_CLEAN_SLOW_FLAGS (REQ_F_REFCOUNT | REQ_F_LINK | REQ_F_HARDLINK |\
116 IO_REQ_CLEAN_FLAGS)
117
118 #define IO_TCTX_REFS_CACHE_NR (1U << 10)
119
120 #define IO_COMPL_BATCH 32
121 #define IO_REQ_ALLOC_BATCH 8
122
123 enum {
124 IO_CHECK_CQ_OVERFLOW_BIT,
125 IO_CHECK_CQ_DROPPED_BIT,
126 };
127
128 enum {
129 IO_EVENTFD_OP_SIGNAL_BIT,
130 IO_EVENTFD_OP_FREE_BIT,
131 };
132
133 struct io_defer_entry {
134 struct list_head list;
135 struct io_kiocb *req;
136 u32 seq;
137 };
138
139 /* requests with any of those set should undergo io_disarm_next() */
140 #define IO_DISARM_MASK (REQ_F_ARM_LTIMEOUT | REQ_F_LINK_TIMEOUT | REQ_F_FAIL)
141 #define IO_REQ_LINK_FLAGS (REQ_F_LINK | REQ_F_HARDLINK)
142
143 static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
144 struct task_struct *task,
145 bool cancel_all);
146
147 static void io_dismantle_req(struct io_kiocb *req);
148 static void io_clean_op(struct io_kiocb *req);
149 static void io_queue_sqe(struct io_kiocb *req);
150 static void io_move_task_work_from_local(struct io_ring_ctx *ctx);
151 static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
152
153 static struct kmem_cache *req_cachep;
154
io_uring_get_socket(struct file * file)155 struct sock *io_uring_get_socket(struct file *file)
156 {
157 #if defined(CONFIG_UNIX)
158 if (io_is_uring_fops(file)) {
159 struct io_ring_ctx *ctx = file->private_data;
160
161 return ctx->ring_sock->sk;
162 }
163 #endif
164 return NULL;
165 }
166 EXPORT_SYMBOL(io_uring_get_socket);
167
io_submit_flush_completions(struct io_ring_ctx * ctx)168 static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
169 {
170 if (!wq_list_empty(&ctx->submit_state.compl_reqs))
171 __io_submit_flush_completions(ctx);
172 }
173
__io_cqring_events(struct io_ring_ctx * ctx)174 static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
175 {
176 return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
177 }
178
__io_cqring_events_user(struct io_ring_ctx * ctx)179 static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
180 {
181 return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
182 }
183
io_match_linked(struct io_kiocb * head)184 static bool io_match_linked(struct io_kiocb *head)
185 {
186 struct io_kiocb *req;
187
188 io_for_each_link(req, head) {
189 if (req->flags & REQ_F_INFLIGHT)
190 return true;
191 }
192 return false;
193 }
194
195 /*
196 * As io_match_task() but protected against racing with linked timeouts.
197 * User must not hold timeout_lock.
198 */
io_match_task_safe(struct io_kiocb * head,struct task_struct * task,bool cancel_all)199 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
200 bool cancel_all)
201 {
202 bool matched;
203
204 if (task && head->task != task)
205 return false;
206 if (cancel_all)
207 return true;
208
209 if (head->flags & REQ_F_LINK_TIMEOUT) {
210 struct io_ring_ctx *ctx = head->ctx;
211
212 /* protect against races with linked timeouts */
213 spin_lock_irq(&ctx->timeout_lock);
214 matched = io_match_linked(head);
215 spin_unlock_irq(&ctx->timeout_lock);
216 } else {
217 matched = io_match_linked(head);
218 }
219 return matched;
220 }
221
req_fail_link_node(struct io_kiocb * req,int res)222 static inline void req_fail_link_node(struct io_kiocb *req, int res)
223 {
224 req_set_fail(req);
225 io_req_set_res(req, res, 0);
226 }
227
io_req_add_to_cache(struct io_kiocb * req,struct io_ring_ctx * ctx)228 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
229 {
230 wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
231 }
232
io_ring_ctx_ref_free(struct percpu_ref * ref)233 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
234 {
235 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
236
237 complete(&ctx->ref_comp);
238 }
239
io_fallback_req_func(struct work_struct * work)240 static __cold void io_fallback_req_func(struct work_struct *work)
241 {
242 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
243 fallback_work.work);
244 struct llist_node *node = llist_del_all(&ctx->fallback_llist);
245 struct io_kiocb *req, *tmp;
246 bool locked = false;
247
248 percpu_ref_get(&ctx->refs);
249 llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
250 req->io_task_work.func(req, &locked);
251
252 if (locked) {
253 io_submit_flush_completions(ctx);
254 mutex_unlock(&ctx->uring_lock);
255 }
256 percpu_ref_put(&ctx->refs);
257 }
258
io_alloc_hash_table(struct io_hash_table * table,unsigned bits)259 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
260 {
261 unsigned hash_buckets = 1U << bits;
262 size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
263
264 table->hbs = kmalloc(hash_size, GFP_KERNEL);
265 if (!table->hbs)
266 return -ENOMEM;
267
268 table->hash_bits = bits;
269 init_hash_table(table, hash_buckets);
270 return 0;
271 }
272
io_ring_ctx_alloc(struct io_uring_params * p)273 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
274 {
275 struct io_ring_ctx *ctx;
276 int hash_bits;
277
278 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
279 if (!ctx)
280 return NULL;
281
282 xa_init(&ctx->io_bl_xa);
283
284 /*
285 * Use 5 bits less than the max cq entries, that should give us around
286 * 32 entries per hash list if totally full and uniformly spread, but
287 * don't keep too many buckets to not overconsume memory.
288 */
289 hash_bits = ilog2(p->cq_entries) - 5;
290 hash_bits = clamp(hash_bits, 1, 8);
291 if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
292 goto err;
293 if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
294 goto err;
295
296 ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
297 if (!ctx->dummy_ubuf)
298 goto err;
299 /* set invalid range, so io_import_fixed() fails meeting it */
300 ctx->dummy_ubuf->ubuf = -1UL;
301
302 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
303 0, GFP_KERNEL))
304 goto err;
305
306 ctx->flags = p->flags;
307 init_waitqueue_head(&ctx->sqo_sq_wait);
308 INIT_LIST_HEAD(&ctx->sqd_list);
309 INIT_LIST_HEAD(&ctx->cq_overflow_list);
310 INIT_LIST_HEAD(&ctx->io_buffers_cache);
311 io_alloc_cache_init(&ctx->apoll_cache);
312 io_alloc_cache_init(&ctx->netmsg_cache);
313 init_completion(&ctx->ref_comp);
314 xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
315 mutex_init(&ctx->uring_lock);
316 init_waitqueue_head(&ctx->cq_wait);
317 spin_lock_init(&ctx->completion_lock);
318 spin_lock_init(&ctx->timeout_lock);
319 INIT_WQ_LIST(&ctx->iopoll_list);
320 INIT_LIST_HEAD(&ctx->io_buffers_pages);
321 INIT_LIST_HEAD(&ctx->io_buffers_comp);
322 INIT_LIST_HEAD(&ctx->defer_list);
323 INIT_LIST_HEAD(&ctx->timeout_list);
324 INIT_LIST_HEAD(&ctx->ltimeout_list);
325 spin_lock_init(&ctx->rsrc_ref_lock);
326 INIT_LIST_HEAD(&ctx->rsrc_ref_list);
327 INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
328 init_llist_head(&ctx->rsrc_put_llist);
329 init_llist_head(&ctx->work_llist);
330 INIT_LIST_HEAD(&ctx->tctx_list);
331 ctx->submit_state.free_list.next = NULL;
332 INIT_WQ_LIST(&ctx->locked_free_list);
333 INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
334 INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
335 return ctx;
336 err:
337 kfree(ctx->dummy_ubuf);
338 kfree(ctx->cancel_table.hbs);
339 kfree(ctx->cancel_table_locked.hbs);
340 kfree(ctx->io_bl);
341 xa_destroy(&ctx->io_bl_xa);
342 kfree(ctx);
343 return NULL;
344 }
345
io_account_cq_overflow(struct io_ring_ctx * ctx)346 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
347 {
348 struct io_rings *r = ctx->rings;
349
350 WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
351 ctx->cq_extra--;
352 }
353
req_need_defer(struct io_kiocb * req,u32 seq)354 static bool req_need_defer(struct io_kiocb *req, u32 seq)
355 {
356 if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
357 struct io_ring_ctx *ctx = req->ctx;
358
359 return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
360 }
361
362 return false;
363 }
364
io_req_track_inflight(struct io_kiocb * req)365 static inline void io_req_track_inflight(struct io_kiocb *req)
366 {
367 if (!(req->flags & REQ_F_INFLIGHT)) {
368 req->flags |= REQ_F_INFLIGHT;
369 atomic_inc(&req->task->io_uring->inflight_tracked);
370 }
371 }
372
__io_prep_linked_timeout(struct io_kiocb * req)373 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
374 {
375 if (WARN_ON_ONCE(!req->link))
376 return NULL;
377
378 req->flags &= ~REQ_F_ARM_LTIMEOUT;
379 req->flags |= REQ_F_LINK_TIMEOUT;
380
381 /* linked timeouts should have two refs once prep'ed */
382 io_req_set_refcount(req);
383 __io_req_set_refcount(req->link, 2);
384 return req->link;
385 }
386
io_prep_linked_timeout(struct io_kiocb * req)387 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
388 {
389 if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
390 return NULL;
391 return __io_prep_linked_timeout(req);
392 }
393
__io_arm_ltimeout(struct io_kiocb * req)394 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
395 {
396 io_queue_linked_timeout(__io_prep_linked_timeout(req));
397 }
398
io_arm_ltimeout(struct io_kiocb * req)399 static inline void io_arm_ltimeout(struct io_kiocb *req)
400 {
401 if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
402 __io_arm_ltimeout(req);
403 }
404
io_prep_async_work(struct io_kiocb * req)405 static void io_prep_async_work(struct io_kiocb *req)
406 {
407 const struct io_op_def *def = &io_op_defs[req->opcode];
408 struct io_ring_ctx *ctx = req->ctx;
409
410 if (!(req->flags & REQ_F_CREDS)) {
411 req->flags |= REQ_F_CREDS;
412 req->creds = get_current_cred();
413 }
414
415 req->work.list.next = NULL;
416 req->work.flags = 0;
417 req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
418 if (req->flags & REQ_F_FORCE_ASYNC)
419 req->work.flags |= IO_WQ_WORK_CONCURRENT;
420
421 if (req->file && !io_req_ffs_set(req))
422 req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
423
424 if (req->flags & REQ_F_ISREG) {
425 if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
426 io_wq_hash_work(&req->work, file_inode(req->file));
427 } else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
428 if (def->unbound_nonreg_file)
429 req->work.flags |= IO_WQ_WORK_UNBOUND;
430 }
431 }
432
io_prep_async_link(struct io_kiocb * req)433 static void io_prep_async_link(struct io_kiocb *req)
434 {
435 struct io_kiocb *cur;
436
437 if (req->flags & REQ_F_LINK_TIMEOUT) {
438 struct io_ring_ctx *ctx = req->ctx;
439
440 spin_lock_irq(&ctx->timeout_lock);
441 io_for_each_link(cur, req)
442 io_prep_async_work(cur);
443 spin_unlock_irq(&ctx->timeout_lock);
444 } else {
445 io_for_each_link(cur, req)
446 io_prep_async_work(cur);
447 }
448 }
449
io_queue_iowq(struct io_kiocb * req,bool * dont_use)450 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
451 {
452 struct io_kiocb *link = io_prep_linked_timeout(req);
453 struct io_uring_task *tctx = req->task->io_uring;
454
455 BUG_ON(!tctx);
456 BUG_ON(!tctx->io_wq);
457
458 /* init ->work of the whole link before punting */
459 io_prep_async_link(req);
460
461 /*
462 * Not expected to happen, but if we do have a bug where this _can_
463 * happen, catch it here and ensure the request is marked as
464 * canceled. That will make io-wq go through the usual work cancel
465 * procedure rather than attempt to run this request (or create a new
466 * worker for it).
467 */
468 if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
469 req->work.flags |= IO_WQ_WORK_CANCEL;
470
471 trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
472 io_wq_enqueue(tctx->io_wq, &req->work);
473 if (link)
474 io_queue_linked_timeout(link);
475 }
476
io_queue_deferred(struct io_ring_ctx * ctx)477 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
478 {
479 while (!list_empty(&ctx->defer_list)) {
480 struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
481 struct io_defer_entry, list);
482
483 if (req_need_defer(de->req, de->seq))
484 break;
485 list_del_init(&de->list);
486 io_req_task_queue(de->req);
487 kfree(de);
488 }
489 }
490
491
io_eventfd_ops(struct rcu_head * rcu)492 static void io_eventfd_ops(struct rcu_head *rcu)
493 {
494 struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
495 int ops = atomic_xchg(&ev_fd->ops, 0);
496
497 if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
498 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
499
500 /* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
501 * ordering in a race but if references are 0 we know we have to free
502 * it regardless.
503 */
504 if (atomic_dec_and_test(&ev_fd->refs)) {
505 eventfd_ctx_put(ev_fd->cq_ev_fd);
506 kfree(ev_fd);
507 }
508 }
509
io_eventfd_signal(struct io_ring_ctx * ctx)510 static void io_eventfd_signal(struct io_ring_ctx *ctx)
511 {
512 struct io_ev_fd *ev_fd = NULL;
513
514 rcu_read_lock();
515 /*
516 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
517 * and eventfd_signal
518 */
519 ev_fd = rcu_dereference(ctx->io_ev_fd);
520
521 /*
522 * Check again if ev_fd exists incase an io_eventfd_unregister call
523 * completed between the NULL check of ctx->io_ev_fd at the start of
524 * the function and rcu_read_lock.
525 */
526 if (unlikely(!ev_fd))
527 goto out;
528 if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
529 goto out;
530 if (ev_fd->eventfd_async && !io_wq_current_is_worker())
531 goto out;
532
533 if (likely(eventfd_signal_allowed())) {
534 eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
535 } else {
536 atomic_inc(&ev_fd->refs);
537 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
538 call_rcu(&ev_fd->rcu, io_eventfd_ops);
539 else
540 atomic_dec(&ev_fd->refs);
541 }
542
543 out:
544 rcu_read_unlock();
545 }
546
io_eventfd_flush_signal(struct io_ring_ctx * ctx)547 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
548 {
549 bool skip;
550
551 spin_lock(&ctx->completion_lock);
552
553 /*
554 * Eventfd should only get triggered when at least one event has been
555 * posted. Some applications rely on the eventfd notification count
556 * only changing IFF a new CQE has been added to the CQ ring. There's
557 * no depedency on 1:1 relationship between how many times this
558 * function is called (and hence the eventfd count) and number of CQEs
559 * posted to the CQ ring.
560 */
561 skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
562 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
563 spin_unlock(&ctx->completion_lock);
564 if (skip)
565 return;
566
567 io_eventfd_signal(ctx);
568 }
569
__io_commit_cqring_flush(struct io_ring_ctx * ctx)570 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
571 {
572 if (ctx->off_timeout_used || ctx->drain_active) {
573 spin_lock(&ctx->completion_lock);
574 if (ctx->off_timeout_used)
575 io_flush_timeouts(ctx);
576 if (ctx->drain_active)
577 io_queue_deferred(ctx);
578 spin_unlock(&ctx->completion_lock);
579 }
580 if (ctx->has_evfd)
581 io_eventfd_flush_signal(ctx);
582 }
583
io_cqring_ev_posted(struct io_ring_ctx * ctx)584 static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx)
585 {
586 io_commit_cqring_flush(ctx);
587 io_cqring_wake(ctx);
588 }
589
__io_cq_unlock_post(struct io_ring_ctx * ctx)590 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
591 __releases(ctx->completion_lock)
592 {
593 io_commit_cqring(ctx);
594 spin_unlock(&ctx->completion_lock);
595 io_cqring_ev_posted(ctx);
596 }
597
io_cq_unlock_post(struct io_ring_ctx * ctx)598 void io_cq_unlock_post(struct io_ring_ctx *ctx)
599 {
600 __io_cq_unlock_post(ctx);
601 }
602
603 /* Returns true if there are no backlogged entries after the flush */
__io_cqring_overflow_flush(struct io_ring_ctx * ctx,bool force)604 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
605 {
606 bool all_flushed;
607 size_t cqe_size = sizeof(struct io_uring_cqe);
608
609 if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
610 return false;
611
612 if (ctx->flags & IORING_SETUP_CQE32)
613 cqe_size <<= 1;
614
615 io_cq_lock(ctx);
616 while (!list_empty(&ctx->cq_overflow_list)) {
617 struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
618 struct io_overflow_cqe *ocqe;
619
620 if (!cqe && !force)
621 break;
622 ocqe = list_first_entry(&ctx->cq_overflow_list,
623 struct io_overflow_cqe, list);
624 if (cqe)
625 memcpy(cqe, &ocqe->cqe, cqe_size);
626 else
627 io_account_cq_overflow(ctx);
628
629 list_del(&ocqe->list);
630 kfree(ocqe);
631 }
632
633 all_flushed = list_empty(&ctx->cq_overflow_list);
634 if (all_flushed) {
635 clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
636 atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
637 }
638
639 io_cq_unlock_post(ctx);
640 return all_flushed;
641 }
642
io_cqring_overflow_flush(struct io_ring_ctx * ctx)643 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
644 {
645 bool ret = true;
646
647 if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
648 /* iopoll syncs against uring_lock, not completion_lock */
649 if (ctx->flags & IORING_SETUP_IOPOLL)
650 mutex_lock(&ctx->uring_lock);
651 ret = __io_cqring_overflow_flush(ctx, false);
652 if (ctx->flags & IORING_SETUP_IOPOLL)
653 mutex_unlock(&ctx->uring_lock);
654 }
655
656 return ret;
657 }
658
__io_put_task(struct task_struct * task,int nr)659 void __io_put_task(struct task_struct *task, int nr)
660 {
661 struct io_uring_task *tctx = task->io_uring;
662
663 percpu_counter_sub(&tctx->inflight, nr);
664 if (unlikely(atomic_read(&tctx->in_idle)))
665 wake_up(&tctx->wait);
666 put_task_struct_many(task, nr);
667 }
668
io_task_refs_refill(struct io_uring_task * tctx)669 void io_task_refs_refill(struct io_uring_task *tctx)
670 {
671 unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
672
673 percpu_counter_add(&tctx->inflight, refill);
674 refcount_add(refill, ¤t->usage);
675 tctx->cached_refs += refill;
676 }
677
io_uring_drop_tctx_refs(struct task_struct * task)678 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
679 {
680 struct io_uring_task *tctx = task->io_uring;
681 unsigned int refs = tctx->cached_refs;
682
683 if (refs) {
684 tctx->cached_refs = 0;
685 percpu_counter_sub(&tctx->inflight, refs);
686 put_task_struct_many(task, refs);
687 }
688 }
689
io_cqring_event_overflow(struct io_ring_ctx * ctx,u64 user_data,s32 res,u32 cflags,u64 extra1,u64 extra2)690 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
691 s32 res, u32 cflags, u64 extra1, u64 extra2)
692 {
693 struct io_overflow_cqe *ocqe;
694 size_t ocq_size = sizeof(struct io_overflow_cqe);
695 bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
696
697 if (is_cqe32)
698 ocq_size += sizeof(struct io_uring_cqe);
699
700 ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
701 trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
702 if (!ocqe) {
703 /*
704 * If we're in ring overflow flush mode, or in task cancel mode,
705 * or cannot allocate an overflow entry, then we need to drop it
706 * on the floor.
707 */
708 io_account_cq_overflow(ctx);
709 set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
710 return false;
711 }
712 if (list_empty(&ctx->cq_overflow_list)) {
713 set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
714 atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
715
716 }
717 ocqe->cqe.user_data = user_data;
718 ocqe->cqe.res = res;
719 ocqe->cqe.flags = cflags;
720 if (is_cqe32) {
721 ocqe->cqe.big_cqe[0] = extra1;
722 ocqe->cqe.big_cqe[1] = extra2;
723 }
724 list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
725 return true;
726 }
727
io_req_cqe_overflow(struct io_kiocb * req)728 bool io_req_cqe_overflow(struct io_kiocb *req)
729 {
730 if (!(req->flags & REQ_F_CQE32_INIT)) {
731 req->extra1 = 0;
732 req->extra2 = 0;
733 }
734 return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
735 req->cqe.res, req->cqe.flags,
736 req->extra1, req->extra2);
737 }
738
739 /*
740 * writes to the cq entry need to come after reading head; the
741 * control dependency is enough as we're using WRITE_ONCE to
742 * fill the cq entry
743 */
__io_get_cqe(struct io_ring_ctx * ctx,bool overflow)744 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
745 {
746 struct io_rings *rings = ctx->rings;
747 unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
748 unsigned int free, queued, len;
749
750 /*
751 * Posting into the CQ when there are pending overflowed CQEs may break
752 * ordering guarantees, which will affect links, F_MORE users and more.
753 * Force overflow the completion.
754 */
755 if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
756 return NULL;
757
758 /* userspace may cheat modifying the tail, be safe and do min */
759 queued = min(__io_cqring_events(ctx), ctx->cq_entries);
760 free = ctx->cq_entries - queued;
761 /* we need a contiguous range, limit based on the current array offset */
762 len = min(free, ctx->cq_entries - off);
763 if (!len)
764 return NULL;
765
766 if (ctx->flags & IORING_SETUP_CQE32) {
767 off <<= 1;
768 len <<= 1;
769 }
770
771 ctx->cqe_cached = &rings->cqes[off];
772 ctx->cqe_sentinel = ctx->cqe_cached + len;
773
774 ctx->cached_cq_tail++;
775 ctx->cqe_cached++;
776 if (ctx->flags & IORING_SETUP_CQE32)
777 ctx->cqe_cached++;
778 return &rings->cqes[off];
779 }
780
io_fill_cqe_aux(struct io_ring_ctx * ctx,u64 user_data,s32 res,u32 cflags,bool allow_overflow)781 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
782 bool allow_overflow)
783 {
784 struct io_uring_cqe *cqe;
785
786 ctx->cq_extra++;
787
788 /*
789 * If we can't get a cq entry, userspace overflowed the
790 * submission (by quite a lot). Increment the overflow count in
791 * the ring.
792 */
793 cqe = io_get_cqe(ctx);
794 if (likely(cqe)) {
795 trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
796
797 WRITE_ONCE(cqe->user_data, user_data);
798 WRITE_ONCE(cqe->res, res);
799 WRITE_ONCE(cqe->flags, cflags);
800
801 if (ctx->flags & IORING_SETUP_CQE32) {
802 WRITE_ONCE(cqe->big_cqe[0], 0);
803 WRITE_ONCE(cqe->big_cqe[1], 0);
804 }
805 return true;
806 }
807
808 if (allow_overflow)
809 return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
810
811 return false;
812 }
813
io_post_aux_cqe(struct io_ring_ctx * ctx,u64 user_data,s32 res,u32 cflags,bool allow_overflow)814 bool io_post_aux_cqe(struct io_ring_ctx *ctx,
815 u64 user_data, s32 res, u32 cflags,
816 bool allow_overflow)
817 {
818 bool filled;
819
820 io_cq_lock(ctx);
821 filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow);
822 io_cq_unlock_post(ctx);
823 return filled;
824 }
825
io_req_complete_post(struct io_kiocb * req)826 void io_req_complete_post(struct io_kiocb *req)
827 {
828 struct io_ring_ctx *ctx = req->ctx;
829
830 io_cq_lock(ctx);
831 if (!(req->flags & REQ_F_CQE_SKIP))
832 __io_fill_cqe_req(ctx, req);
833
834 /*
835 * If we're the last reference to this request, add to our locked
836 * free_list cache.
837 */
838 if (req_ref_put_and_test(req)) {
839 if (req->flags & IO_REQ_LINK_FLAGS) {
840 if (req->flags & IO_DISARM_MASK)
841 io_disarm_next(req);
842 if (req->link) {
843 io_req_task_queue(req->link);
844 req->link = NULL;
845 }
846 }
847 io_req_put_rsrc(req);
848 /*
849 * Selected buffer deallocation in io_clean_op() assumes that
850 * we don't hold ->completion_lock. Clean them here to avoid
851 * deadlocks.
852 */
853 io_put_kbuf_comp(req);
854 io_dismantle_req(req);
855 io_put_task(req->task, 1);
856 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
857 ctx->locked_free_nr++;
858 }
859 io_cq_unlock_post(ctx);
860 }
861
__io_req_complete(struct io_kiocb * req,unsigned issue_flags)862 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
863 {
864 io_req_complete_post(req);
865 }
866
io_req_complete_failed(struct io_kiocb * req,s32 res)867 void io_req_complete_failed(struct io_kiocb *req, s32 res)
868 __must_hold(&ctx->uring_lock)
869 {
870 const struct io_op_def *def = &io_op_defs[req->opcode];
871
872 lockdep_assert_held(&req->ctx->uring_lock);
873
874 req_set_fail(req);
875 io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
876 if (def->fail)
877 def->fail(req);
878 io_req_complete_post(req);
879 }
880
881 /*
882 * Don't initialise the fields below on every allocation, but do that in
883 * advance and keep them valid across allocations.
884 */
io_preinit_req(struct io_kiocb * req,struct io_ring_ctx * ctx)885 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
886 {
887 req->ctx = ctx;
888 req->link = NULL;
889 req->async_data = NULL;
890 /* not necessary, but safer to zero */
891 req->cqe.res = 0;
892 }
893
io_flush_cached_locked_reqs(struct io_ring_ctx * ctx,struct io_submit_state * state)894 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
895 struct io_submit_state *state)
896 {
897 spin_lock(&ctx->completion_lock);
898 wq_list_splice(&ctx->locked_free_list, &state->free_list);
899 ctx->locked_free_nr = 0;
900 spin_unlock(&ctx->completion_lock);
901 }
902
903 /*
904 * A request might get retired back into the request caches even before opcode
905 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
906 * Because of that, io_alloc_req() should be called only under ->uring_lock
907 * and with extra caution to not get a request that is still worked on.
908 */
__io_alloc_req_refill(struct io_ring_ctx * ctx)909 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
910 __must_hold(&ctx->uring_lock)
911 {
912 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
913 void *reqs[IO_REQ_ALLOC_BATCH];
914 int ret, i;
915
916 /*
917 * If we have more than a batch's worth of requests in our IRQ side
918 * locked cache, grab the lock and move them over to our submission
919 * side cache.
920 */
921 if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
922 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
923 if (!io_req_cache_empty(ctx))
924 return true;
925 }
926
927 ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
928
929 /*
930 * Bulk alloc is all-or-nothing. If we fail to get a batch,
931 * retry single alloc to be on the safe side.
932 */
933 if (unlikely(ret <= 0)) {
934 reqs[0] = kmem_cache_alloc(req_cachep, gfp);
935 if (!reqs[0])
936 return false;
937 ret = 1;
938 }
939
940 percpu_ref_get_many(&ctx->refs, ret);
941 for (i = 0; i < ret; i++) {
942 struct io_kiocb *req = reqs[i];
943
944 io_preinit_req(req, ctx);
945 io_req_add_to_cache(req, ctx);
946 }
947 return true;
948 }
949
io_dismantle_req(struct io_kiocb * req)950 static inline void io_dismantle_req(struct io_kiocb *req)
951 {
952 unsigned int flags = req->flags;
953
954 if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
955 io_clean_op(req);
956 if (!(flags & REQ_F_FIXED_FILE))
957 io_put_file(req->file);
958 }
959
io_free_req(struct io_kiocb * req)960 __cold void io_free_req(struct io_kiocb *req)
961 {
962 struct io_ring_ctx *ctx = req->ctx;
963
964 io_req_put_rsrc(req);
965 io_dismantle_req(req);
966 io_put_task(req->task, 1);
967
968 spin_lock(&ctx->completion_lock);
969 wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
970 ctx->locked_free_nr++;
971 spin_unlock(&ctx->completion_lock);
972 }
973
__io_req_find_next_prep(struct io_kiocb * req)974 static void __io_req_find_next_prep(struct io_kiocb *req)
975 {
976 struct io_ring_ctx *ctx = req->ctx;
977
978 io_cq_lock(ctx);
979 io_disarm_next(req);
980 io_cq_unlock_post(ctx);
981 }
982
io_req_find_next(struct io_kiocb * req)983 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
984 {
985 struct io_kiocb *nxt;
986
987 /*
988 * If LINK is set, we have dependent requests in this chain. If we
989 * didn't fail this request, queue the first one up, moving any other
990 * dependencies to the next request. In case of failure, fail the rest
991 * of the chain.
992 */
993 if (unlikely(req->flags & IO_DISARM_MASK))
994 __io_req_find_next_prep(req);
995 nxt = req->link;
996 req->link = NULL;
997 return nxt;
998 }
999
ctx_flush_and_put(struct io_ring_ctx * ctx,bool * locked)1000 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
1001 {
1002 if (!ctx)
1003 return;
1004 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1005 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1006 if (*locked) {
1007 io_submit_flush_completions(ctx);
1008 mutex_unlock(&ctx->uring_lock);
1009 *locked = false;
1010 }
1011 percpu_ref_put(&ctx->refs);
1012 }
1013
handle_tw_list(struct llist_node * node,struct io_ring_ctx ** ctx,bool * locked,struct llist_node * last)1014 static unsigned int handle_tw_list(struct llist_node *node,
1015 struct io_ring_ctx **ctx, bool *locked,
1016 struct llist_node *last)
1017 {
1018 unsigned int count = 0;
1019
1020 while (node != last) {
1021 struct llist_node *next = node->next;
1022 struct io_kiocb *req = container_of(node, struct io_kiocb,
1023 io_task_work.node);
1024
1025 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1026
1027 if (req->ctx != *ctx) {
1028 ctx_flush_and_put(*ctx, locked);
1029 *ctx = req->ctx;
1030 /* if not contended, grab and improve batching */
1031 *locked = mutex_trylock(&(*ctx)->uring_lock);
1032 percpu_ref_get(&(*ctx)->refs);
1033 }
1034 req->io_task_work.func(req, locked);
1035 node = next;
1036 count++;
1037 }
1038
1039 return count;
1040 }
1041
1042 /**
1043 * io_llist_xchg - swap all entries in a lock-less list
1044 * @head: the head of lock-less list to delete all entries
1045 * @new: new entry as the head of the list
1046 *
1047 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1048 * The order of entries returned is from the newest to the oldest added one.
1049 */
io_llist_xchg(struct llist_head * head,struct llist_node * new)1050 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1051 struct llist_node *new)
1052 {
1053 return xchg(&head->first, new);
1054 }
1055
1056 /**
1057 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1058 * @head: the head of lock-less list to delete all entries
1059 * @old: expected old value of the first entry of the list
1060 * @new: new entry as the head of the list
1061 *
1062 * perform a cmpxchg on the first entry of the list.
1063 */
1064
io_llist_cmpxchg(struct llist_head * head,struct llist_node * old,struct llist_node * new)1065 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1066 struct llist_node *old,
1067 struct llist_node *new)
1068 {
1069 return cmpxchg(&head->first, old, new);
1070 }
1071
tctx_task_work(struct callback_head * cb)1072 void tctx_task_work(struct callback_head *cb)
1073 {
1074 bool uring_locked = false;
1075 struct io_ring_ctx *ctx = NULL;
1076 struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1077 task_work);
1078 struct llist_node fake = {};
1079 struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake);
1080 unsigned int loops = 1;
1081 unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1082
1083 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1084 while (node != &fake) {
1085 loops++;
1086 node = io_llist_xchg(&tctx->task_list, &fake);
1087 count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1088 node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1089 }
1090
1091 ctx_flush_and_put(ctx, &uring_locked);
1092
1093 /* relaxed read is enough as only the task itself sets ->in_idle */
1094 if (unlikely(atomic_read(&tctx->in_idle)))
1095 io_uring_drop_tctx_refs(current);
1096
1097 trace_io_uring_task_work_run(tctx, count, loops);
1098 }
1099
io_req_local_work_add(struct io_kiocb * req)1100 static void io_req_local_work_add(struct io_kiocb *req)
1101 {
1102 struct io_ring_ctx *ctx = req->ctx;
1103
1104 percpu_ref_get(&ctx->refs);
1105
1106 if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) {
1107 percpu_ref_put(&ctx->refs);
1108 return;
1109 }
1110 /* need it for the following io_cqring_wake() */
1111 smp_mb__after_atomic();
1112
1113 if (unlikely(atomic_read(&req->task->io_uring->in_idle))) {
1114 io_move_task_work_from_local(ctx);
1115 percpu_ref_put(&ctx->refs);
1116 return;
1117 }
1118
1119 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1120 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1121
1122 if (ctx->has_evfd)
1123 io_eventfd_signal(ctx);
1124 __io_cqring_wake(ctx);
1125 percpu_ref_put(&ctx->refs);
1126 }
1127
__io_req_task_work_add(struct io_kiocb * req,bool allow_local)1128 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1129 {
1130 struct io_uring_task *tctx = req->task->io_uring;
1131 struct io_ring_ctx *ctx = req->ctx;
1132 struct llist_node *node;
1133
1134 if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1135 io_req_local_work_add(req);
1136 return;
1137 }
1138
1139 /* task_work already pending, we're done */
1140 if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1141 return;
1142
1143 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1144 atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1145
1146 if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1147 return;
1148
1149 node = llist_del_all(&tctx->task_list);
1150
1151 while (node) {
1152 req = container_of(node, struct io_kiocb, io_task_work.node);
1153 node = node->next;
1154 if (llist_add(&req->io_task_work.node,
1155 &req->ctx->fallback_llist))
1156 schedule_delayed_work(&req->ctx->fallback_work, 1);
1157 }
1158 }
1159
io_move_task_work_from_local(struct io_ring_ctx * ctx)1160 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1161 {
1162 struct llist_node *node;
1163
1164 node = llist_del_all(&ctx->work_llist);
1165 while (node) {
1166 struct io_kiocb *req = container_of(node, struct io_kiocb,
1167 io_task_work.node);
1168
1169 node = node->next;
1170 __io_req_task_work_add(req, false);
1171 }
1172 }
1173
__io_run_local_work(struct io_ring_ctx * ctx,bool * locked)1174 int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked)
1175 {
1176 struct llist_node *node;
1177 struct llist_node fake;
1178 struct llist_node *current_final = NULL;
1179 int ret;
1180 unsigned int loops = 1;
1181
1182 if (unlikely(ctx->submitter_task != current))
1183 return -EEXIST;
1184
1185 node = io_llist_xchg(&ctx->work_llist, &fake);
1186 ret = 0;
1187 again:
1188 while (node != current_final) {
1189 struct llist_node *next = node->next;
1190 struct io_kiocb *req = container_of(node, struct io_kiocb,
1191 io_task_work.node);
1192 prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1193 req->io_task_work.func(req, locked);
1194 ret++;
1195 node = next;
1196 }
1197
1198 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1199 atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1200
1201 node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL);
1202 if (node != &fake) {
1203 loops++;
1204 current_final = &fake;
1205 node = io_llist_xchg(&ctx->work_llist, &fake);
1206 goto again;
1207 }
1208
1209 if (*locked)
1210 io_submit_flush_completions(ctx);
1211 trace_io_uring_local_work_run(ctx, ret, loops);
1212 return ret;
1213
1214 }
1215
io_run_local_work(struct io_ring_ctx * ctx)1216 int io_run_local_work(struct io_ring_ctx *ctx)
1217 {
1218 bool locked;
1219 int ret;
1220
1221 if (llist_empty(&ctx->work_llist))
1222 return 0;
1223
1224 __set_current_state(TASK_RUNNING);
1225 locked = mutex_trylock(&ctx->uring_lock);
1226 ret = __io_run_local_work(ctx, &locked);
1227 if (locked)
1228 mutex_unlock(&ctx->uring_lock);
1229
1230 return ret;
1231 }
1232
io_req_task_cancel(struct io_kiocb * req,bool * locked)1233 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1234 {
1235 /* not needed for normal modes, but SQPOLL depends on it */
1236 io_tw_lock(req->ctx, locked);
1237 io_req_complete_failed(req, req->cqe.res);
1238 }
1239
io_req_task_submit(struct io_kiocb * req,bool * locked)1240 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1241 {
1242 io_tw_lock(req->ctx, locked);
1243 /* req->task == current here, checking PF_EXITING is safe */
1244 if (likely(!(req->task->flags & PF_EXITING)))
1245 io_queue_sqe(req);
1246 else
1247 io_req_complete_failed(req, -EFAULT);
1248 }
1249
io_req_task_queue_fail(struct io_kiocb * req,int ret)1250 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1251 {
1252 io_req_set_res(req, ret, 0);
1253 req->io_task_work.func = io_req_task_cancel;
1254 io_req_task_work_add(req);
1255 }
1256
io_req_task_queue(struct io_kiocb * req)1257 void io_req_task_queue(struct io_kiocb *req)
1258 {
1259 req->io_task_work.func = io_req_task_submit;
1260 io_req_task_work_add(req);
1261 }
1262
io_queue_next(struct io_kiocb * req)1263 void io_queue_next(struct io_kiocb *req)
1264 {
1265 struct io_kiocb *nxt = io_req_find_next(req);
1266
1267 if (nxt)
1268 io_req_task_queue(nxt);
1269 }
1270
io_free_batch_list(struct io_ring_ctx * ctx,struct io_wq_work_node * node)1271 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1272 __must_hold(&ctx->uring_lock)
1273 {
1274 struct task_struct *task = NULL;
1275 int task_refs = 0;
1276
1277 do {
1278 struct io_kiocb *req = container_of(node, struct io_kiocb,
1279 comp_list);
1280
1281 if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1282 if (req->flags & REQ_F_REFCOUNT) {
1283 node = req->comp_list.next;
1284 if (!req_ref_put_and_test(req))
1285 continue;
1286 }
1287 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1288 struct async_poll *apoll = req->apoll;
1289
1290 if (apoll->double_poll)
1291 kfree(apoll->double_poll);
1292 if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1293 kfree(apoll);
1294 req->flags &= ~REQ_F_POLLED;
1295 }
1296 if (req->flags & IO_REQ_LINK_FLAGS)
1297 io_queue_next(req);
1298 if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1299 io_clean_op(req);
1300 }
1301 if (!(req->flags & REQ_F_FIXED_FILE))
1302 io_put_file(req->file);
1303
1304 io_req_put_rsrc_locked(req, ctx);
1305
1306 if (req->task != task) {
1307 if (task)
1308 io_put_task(task, task_refs);
1309 task = req->task;
1310 task_refs = 0;
1311 }
1312 task_refs++;
1313 node = req->comp_list.next;
1314 io_req_add_to_cache(req, ctx);
1315 } while (node);
1316
1317 if (task)
1318 io_put_task(task, task_refs);
1319 }
1320
__io_submit_flush_completions(struct io_ring_ctx * ctx)1321 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1322 __must_hold(&ctx->uring_lock)
1323 {
1324 struct io_wq_work_node *node, *prev;
1325 struct io_submit_state *state = &ctx->submit_state;
1326
1327 io_cq_lock(ctx);
1328 wq_list_for_each(node, prev, &state->compl_reqs) {
1329 struct io_kiocb *req = container_of(node, struct io_kiocb,
1330 comp_list);
1331
1332 if (!(req->flags & REQ_F_CQE_SKIP))
1333 __io_fill_cqe_req(ctx, req);
1334 }
1335 __io_cq_unlock_post(ctx);
1336
1337 io_free_batch_list(ctx, state->compl_reqs.first);
1338 INIT_WQ_LIST(&state->compl_reqs);
1339 }
1340
1341 /*
1342 * Drop reference to request, return next in chain (if there is one) if this
1343 * was the last reference to this request.
1344 */
io_put_req_find_next(struct io_kiocb * req)1345 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1346 {
1347 struct io_kiocb *nxt = NULL;
1348
1349 if (req_ref_put_and_test(req)) {
1350 if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1351 nxt = io_req_find_next(req);
1352 io_free_req(req);
1353 }
1354 return nxt;
1355 }
1356
io_cqring_events(struct io_ring_ctx * ctx)1357 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1358 {
1359 /* See comment at the top of this file */
1360 smp_rmb();
1361 return __io_cqring_events(ctx);
1362 }
1363
1364 /*
1365 * We can't just wait for polled events to come to us, we have to actively
1366 * find and complete them.
1367 */
io_iopoll_try_reap_events(struct io_ring_ctx * ctx)1368 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1369 {
1370 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1371 return;
1372
1373 mutex_lock(&ctx->uring_lock);
1374 while (!wq_list_empty(&ctx->iopoll_list)) {
1375 /* let it sleep and repeat later if can't complete a request */
1376 if (io_do_iopoll(ctx, true) == 0)
1377 break;
1378 /*
1379 * Ensure we allow local-to-the-cpu processing to take place,
1380 * in this case we need to ensure that we reap all events.
1381 * Also let task_work, etc. to progress by releasing the mutex
1382 */
1383 if (need_resched()) {
1384 mutex_unlock(&ctx->uring_lock);
1385 cond_resched();
1386 mutex_lock(&ctx->uring_lock);
1387 }
1388 }
1389 mutex_unlock(&ctx->uring_lock);
1390 }
1391
io_iopoll_check(struct io_ring_ctx * ctx,long min)1392 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1393 {
1394 unsigned int nr_events = 0;
1395 int ret = 0;
1396 unsigned long check_cq;
1397
1398 if (!io_allowed_run_tw(ctx))
1399 return -EEXIST;
1400
1401 check_cq = READ_ONCE(ctx->check_cq);
1402 if (unlikely(check_cq)) {
1403 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1404 __io_cqring_overflow_flush(ctx, false);
1405 /*
1406 * Similarly do not spin if we have not informed the user of any
1407 * dropped CQE.
1408 */
1409 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1410 return -EBADR;
1411 }
1412 /*
1413 * Don't enter poll loop if we already have events pending.
1414 * If we do, we can potentially be spinning for commands that
1415 * already triggered a CQE (eg in error).
1416 */
1417 if (io_cqring_events(ctx))
1418 return 0;
1419
1420 do {
1421 /*
1422 * If a submit got punted to a workqueue, we can have the
1423 * application entering polling for a command before it gets
1424 * issued. That app will hold the uring_lock for the duration
1425 * of the poll right here, so we need to take a breather every
1426 * now and then to ensure that the issue has a chance to add
1427 * the poll to the issued list. Otherwise we can spin here
1428 * forever, while the workqueue is stuck trying to acquire the
1429 * very same mutex.
1430 */
1431 if (wq_list_empty(&ctx->iopoll_list) ||
1432 io_task_work_pending(ctx)) {
1433 u32 tail = ctx->cached_cq_tail;
1434
1435 (void) io_run_local_work_locked(ctx);
1436
1437 if (task_work_pending(current) ||
1438 wq_list_empty(&ctx->iopoll_list)) {
1439 mutex_unlock(&ctx->uring_lock);
1440 io_run_task_work();
1441 mutex_lock(&ctx->uring_lock);
1442 }
1443 /* some requests don't go through iopoll_list */
1444 if (tail != ctx->cached_cq_tail ||
1445 wq_list_empty(&ctx->iopoll_list))
1446 break;
1447 }
1448 ret = io_do_iopoll(ctx, !min);
1449 if (ret < 0)
1450 break;
1451 nr_events += ret;
1452 ret = 0;
1453 } while (nr_events < min && !need_resched());
1454
1455 return ret;
1456 }
1457
io_req_task_complete(struct io_kiocb * req,bool * locked)1458 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1459 {
1460 if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1461 unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1462
1463 req->cqe.flags |= io_put_kbuf(req, issue_flags);
1464 }
1465
1466 if (*locked)
1467 io_req_complete_defer(req);
1468 else
1469 io_req_complete_post(req);
1470 }
1471
1472 /*
1473 * After the iocb has been issued, it's safe to be found on the poll list.
1474 * Adding the kiocb to the list AFTER submission ensures that we don't
1475 * find it from a io_do_iopoll() thread before the issuer is done
1476 * accessing the kiocb cookie.
1477 */
io_iopoll_req_issued(struct io_kiocb * req,unsigned int issue_flags)1478 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1479 {
1480 struct io_ring_ctx *ctx = req->ctx;
1481 const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1482
1483 /* workqueue context doesn't hold uring_lock, grab it now */
1484 if (unlikely(needs_lock))
1485 mutex_lock(&ctx->uring_lock);
1486
1487 /*
1488 * Track whether we have multiple files in our lists. This will impact
1489 * how we do polling eventually, not spinning if we're on potentially
1490 * different devices.
1491 */
1492 if (wq_list_empty(&ctx->iopoll_list)) {
1493 ctx->poll_multi_queue = false;
1494 } else if (!ctx->poll_multi_queue) {
1495 struct io_kiocb *list_req;
1496
1497 list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1498 comp_list);
1499 if (list_req->file != req->file)
1500 ctx->poll_multi_queue = true;
1501 }
1502
1503 /*
1504 * For fast devices, IO may have already completed. If it has, add
1505 * it to the front so we find it first.
1506 */
1507 if (READ_ONCE(req->iopoll_completed))
1508 wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1509 else
1510 wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1511
1512 if (unlikely(needs_lock)) {
1513 /*
1514 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1515 * in sq thread task context or in io worker task context. If
1516 * current task context is sq thread, we don't need to check
1517 * whether should wake up sq thread.
1518 */
1519 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1520 wq_has_sleeper(&ctx->sq_data->wait))
1521 wake_up(&ctx->sq_data->wait);
1522
1523 mutex_unlock(&ctx->uring_lock);
1524 }
1525 }
1526
io_bdev_nowait(struct block_device * bdev)1527 static bool io_bdev_nowait(struct block_device *bdev)
1528 {
1529 return !bdev || bdev_nowait(bdev);
1530 }
1531
1532 /*
1533 * If we tracked the file through the SCM inflight mechanism, we could support
1534 * any file. For now, just ensure that anything potentially problematic is done
1535 * inline.
1536 */
__io_file_supports_nowait(struct file * file,umode_t mode)1537 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1538 {
1539 if (S_ISBLK(mode)) {
1540 if (IS_ENABLED(CONFIG_BLOCK) &&
1541 io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1542 return true;
1543 return false;
1544 }
1545 if (S_ISSOCK(mode))
1546 return true;
1547 if (S_ISREG(mode)) {
1548 if (IS_ENABLED(CONFIG_BLOCK) &&
1549 io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1550 !io_is_uring_fops(file))
1551 return true;
1552 return false;
1553 }
1554
1555 /* any ->read/write should understand O_NONBLOCK */
1556 if (file->f_flags & O_NONBLOCK)
1557 return true;
1558 return file->f_mode & FMODE_NOWAIT;
1559 }
1560
1561 /*
1562 * If we tracked the file through the SCM inflight mechanism, we could support
1563 * any file. For now, just ensure that anything potentially problematic is done
1564 * inline.
1565 */
io_file_get_flags(struct file * file)1566 unsigned int io_file_get_flags(struct file *file)
1567 {
1568 umode_t mode = file_inode(file)->i_mode;
1569 unsigned int res = 0;
1570
1571 if (S_ISREG(mode))
1572 res |= FFS_ISREG;
1573 if (__io_file_supports_nowait(file, mode))
1574 res |= FFS_NOWAIT;
1575 return res;
1576 }
1577
io_alloc_async_data(struct io_kiocb * req)1578 bool io_alloc_async_data(struct io_kiocb *req)
1579 {
1580 WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1581 req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1582 if (req->async_data) {
1583 req->flags |= REQ_F_ASYNC_DATA;
1584 return false;
1585 }
1586 return true;
1587 }
1588
io_req_prep_async(struct io_kiocb * req)1589 int io_req_prep_async(struct io_kiocb *req)
1590 {
1591 const struct io_op_def *def = &io_op_defs[req->opcode];
1592
1593 /* assign early for deferred execution for non-fixed file */
1594 if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1595 req->file = io_file_get_normal(req, req->cqe.fd);
1596 if (!def->prep_async)
1597 return 0;
1598 if (WARN_ON_ONCE(req_has_async_data(req)))
1599 return -EFAULT;
1600 if (!io_op_defs[req->opcode].manual_alloc) {
1601 if (io_alloc_async_data(req))
1602 return -EAGAIN;
1603 }
1604 return def->prep_async(req);
1605 }
1606
io_get_sequence(struct io_kiocb * req)1607 static u32 io_get_sequence(struct io_kiocb *req)
1608 {
1609 u32 seq = req->ctx->cached_sq_head;
1610 struct io_kiocb *cur;
1611
1612 /* need original cached_sq_head, but it was increased for each req */
1613 io_for_each_link(cur, req)
1614 seq--;
1615 return seq;
1616 }
1617
io_drain_req(struct io_kiocb * req)1618 static __cold void io_drain_req(struct io_kiocb *req)
1619 __must_hold(&ctx->uring_lock)
1620 {
1621 struct io_ring_ctx *ctx = req->ctx;
1622 struct io_defer_entry *de;
1623 int ret;
1624 u32 seq = io_get_sequence(req);
1625
1626 /* Still need defer if there is pending req in defer list. */
1627 spin_lock(&ctx->completion_lock);
1628 if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1629 spin_unlock(&ctx->completion_lock);
1630 queue:
1631 ctx->drain_active = false;
1632 io_req_task_queue(req);
1633 return;
1634 }
1635 spin_unlock(&ctx->completion_lock);
1636
1637 io_prep_async_link(req);
1638 de = kmalloc(sizeof(*de), GFP_KERNEL);
1639 if (!de) {
1640 ret = -ENOMEM;
1641 io_req_complete_failed(req, ret);
1642 return;
1643 }
1644
1645 spin_lock(&ctx->completion_lock);
1646 if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1647 spin_unlock(&ctx->completion_lock);
1648 kfree(de);
1649 goto queue;
1650 }
1651
1652 trace_io_uring_defer(req);
1653 de->req = req;
1654 de->seq = seq;
1655 list_add_tail(&de->list, &ctx->defer_list);
1656 spin_unlock(&ctx->completion_lock);
1657 }
1658
io_clean_op(struct io_kiocb * req)1659 static void io_clean_op(struct io_kiocb *req)
1660 {
1661 if (req->flags & REQ_F_BUFFER_SELECTED) {
1662 spin_lock(&req->ctx->completion_lock);
1663 io_put_kbuf_comp(req);
1664 spin_unlock(&req->ctx->completion_lock);
1665 }
1666
1667 if (req->flags & REQ_F_NEED_CLEANUP) {
1668 const struct io_op_def *def = &io_op_defs[req->opcode];
1669
1670 if (def->cleanup)
1671 def->cleanup(req);
1672 }
1673 if ((req->flags & REQ_F_POLLED) && req->apoll) {
1674 kfree(req->apoll->double_poll);
1675 kfree(req->apoll);
1676 req->apoll = NULL;
1677 }
1678 if (req->flags & REQ_F_INFLIGHT) {
1679 struct io_uring_task *tctx = req->task->io_uring;
1680
1681 atomic_dec(&tctx->inflight_tracked);
1682 }
1683 if (req->flags & REQ_F_CREDS)
1684 put_cred(req->creds);
1685 if (req->flags & REQ_F_ASYNC_DATA) {
1686 kfree(req->async_data);
1687 req->async_data = NULL;
1688 }
1689 req->flags &= ~IO_REQ_CLEAN_FLAGS;
1690 }
1691
io_assign_file(struct io_kiocb * req,unsigned int issue_flags)1692 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1693 {
1694 if (req->file || !io_op_defs[req->opcode].needs_file)
1695 return true;
1696
1697 if (req->flags & REQ_F_FIXED_FILE)
1698 req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1699 else
1700 req->file = io_file_get_normal(req, req->cqe.fd);
1701
1702 return !!req->file;
1703 }
1704
io_issue_sqe(struct io_kiocb * req,unsigned int issue_flags)1705 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1706 {
1707 const struct io_op_def *def = &io_op_defs[req->opcode];
1708 const struct cred *creds = NULL;
1709 int ret;
1710
1711 if (unlikely(!io_assign_file(req, issue_flags)))
1712 return -EBADF;
1713
1714 if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1715 creds = override_creds(req->creds);
1716
1717 if (!def->audit_skip)
1718 audit_uring_entry(req->opcode);
1719
1720 ret = def->issue(req, issue_flags);
1721
1722 if (!def->audit_skip)
1723 audit_uring_exit(!ret, ret);
1724
1725 if (creds)
1726 revert_creds(creds);
1727
1728 if (ret == IOU_OK) {
1729 if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1730 io_req_complete_defer(req);
1731 else
1732 io_req_complete_post(req);
1733 } else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1734 return ret;
1735
1736 /* If the op doesn't have a file, we're not polling for it */
1737 if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1738 io_iopoll_req_issued(req, issue_flags);
1739
1740 return 0;
1741 }
1742
io_poll_issue(struct io_kiocb * req,bool * locked)1743 int io_poll_issue(struct io_kiocb *req, bool *locked)
1744 {
1745 io_tw_lock(req->ctx, locked);
1746 if (unlikely(req->task->flags & PF_EXITING))
1747 return -EFAULT;
1748 return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT);
1749 }
1750
io_wq_free_work(struct io_wq_work * work)1751 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1752 {
1753 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1754
1755 req = io_put_req_find_next(req);
1756 return req ? &req->work : NULL;
1757 }
1758
io_wq_submit_work(struct io_wq_work * work)1759 void io_wq_submit_work(struct io_wq_work *work)
1760 {
1761 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1762 const struct io_op_def *def = &io_op_defs[req->opcode];
1763 unsigned int issue_flags = IO_URING_F_UNLOCKED;
1764 bool needs_poll = false;
1765 int ret = 0, err = -ECANCELED;
1766
1767 /* one will be dropped by ->io_free_work() after returning to io-wq */
1768 if (!(req->flags & REQ_F_REFCOUNT))
1769 __io_req_set_refcount(req, 2);
1770 else
1771 req_ref_get(req);
1772
1773 io_arm_ltimeout(req);
1774
1775 /* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1776 if (work->flags & IO_WQ_WORK_CANCEL) {
1777 fail:
1778 io_req_task_queue_fail(req, err);
1779 return;
1780 }
1781 if (!io_assign_file(req, issue_flags)) {
1782 err = -EBADF;
1783 work->flags |= IO_WQ_WORK_CANCEL;
1784 goto fail;
1785 }
1786
1787 if (req->flags & REQ_F_FORCE_ASYNC) {
1788 bool opcode_poll = def->pollin || def->pollout;
1789
1790 if (opcode_poll && file_can_poll(req->file)) {
1791 needs_poll = true;
1792 issue_flags |= IO_URING_F_NONBLOCK;
1793 }
1794 }
1795
1796 do {
1797 ret = io_issue_sqe(req, issue_flags);
1798 if (ret != -EAGAIN)
1799 break;
1800 /*
1801 * We can get EAGAIN for iopolled IO even though we're
1802 * forcing a sync submission from here, since we can't
1803 * wait for request slots on the block side.
1804 */
1805 if (!needs_poll) {
1806 if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1807 break;
1808 cond_resched();
1809 continue;
1810 }
1811
1812 if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1813 return;
1814 /* aborted or ready, in either case retry blocking */
1815 needs_poll = false;
1816 issue_flags &= ~IO_URING_F_NONBLOCK;
1817 } while (1);
1818
1819 /* avoid locking problems by failing it from a clean context */
1820 if (ret < 0)
1821 io_req_task_queue_fail(req, ret);
1822 }
1823
io_file_get_fixed(struct io_kiocb * req,int fd,unsigned int issue_flags)1824 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1825 unsigned int issue_flags)
1826 {
1827 struct io_ring_ctx *ctx = req->ctx;
1828 struct file *file = NULL;
1829 unsigned long file_ptr;
1830
1831 io_ring_submit_lock(ctx, issue_flags);
1832
1833 if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1834 goto out;
1835 fd = array_index_nospec(fd, ctx->nr_user_files);
1836 file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1837 file = (struct file *) (file_ptr & FFS_MASK);
1838 file_ptr &= ~FFS_MASK;
1839 /* mask in overlapping REQ_F and FFS bits */
1840 req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1841 io_req_set_rsrc_node(req, ctx, 0);
1842 out:
1843 io_ring_submit_unlock(ctx, issue_flags);
1844 return file;
1845 }
1846
io_file_get_normal(struct io_kiocb * req,int fd)1847 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1848 {
1849 struct file *file = fget(fd);
1850
1851 trace_io_uring_file_get(req, fd);
1852
1853 /* we don't allow fixed io_uring files */
1854 if (file && io_is_uring_fops(file))
1855 io_req_track_inflight(req);
1856 return file;
1857 }
1858
io_queue_async(struct io_kiocb * req,int ret)1859 static void io_queue_async(struct io_kiocb *req, int ret)
1860 __must_hold(&req->ctx->uring_lock)
1861 {
1862 struct io_kiocb *linked_timeout;
1863
1864 if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1865 io_req_complete_failed(req, ret);
1866 return;
1867 }
1868
1869 linked_timeout = io_prep_linked_timeout(req);
1870
1871 switch (io_arm_poll_handler(req, 0)) {
1872 case IO_APOLL_READY:
1873 io_kbuf_recycle(req, 0);
1874 io_req_task_queue(req);
1875 break;
1876 case IO_APOLL_ABORTED:
1877 io_kbuf_recycle(req, 0);
1878 io_queue_iowq(req, NULL);
1879 break;
1880 case IO_APOLL_OK:
1881 break;
1882 }
1883
1884 if (linked_timeout)
1885 io_queue_linked_timeout(linked_timeout);
1886 }
1887
io_queue_sqe(struct io_kiocb * req)1888 static inline void io_queue_sqe(struct io_kiocb *req)
1889 __must_hold(&req->ctx->uring_lock)
1890 {
1891 int ret;
1892
1893 ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1894
1895 /*
1896 * We async punt it if the file wasn't marked NOWAIT, or if the file
1897 * doesn't support non-blocking read/write attempts
1898 */
1899 if (likely(!ret))
1900 io_arm_ltimeout(req);
1901 else
1902 io_queue_async(req, ret);
1903 }
1904
io_queue_sqe_fallback(struct io_kiocb * req)1905 static void io_queue_sqe_fallback(struct io_kiocb *req)
1906 __must_hold(&req->ctx->uring_lock)
1907 {
1908 if (unlikely(req->flags & REQ_F_FAIL)) {
1909 /*
1910 * We don't submit, fail them all, for that replace hardlinks
1911 * with normal links. Extra REQ_F_LINK is tolerated.
1912 */
1913 req->flags &= ~REQ_F_HARDLINK;
1914 req->flags |= REQ_F_LINK;
1915 io_req_complete_failed(req, req->cqe.res);
1916 } else {
1917 int ret = io_req_prep_async(req);
1918
1919 if (unlikely(ret)) {
1920 io_req_complete_failed(req, ret);
1921 return;
1922 }
1923
1924 if (unlikely(req->ctx->drain_active))
1925 io_drain_req(req);
1926 else
1927 io_queue_iowq(req, NULL);
1928 }
1929 }
1930
1931 /*
1932 * Check SQE restrictions (opcode and flags).
1933 *
1934 * Returns 'true' if SQE is allowed, 'false' otherwise.
1935 */
io_check_restriction(struct io_ring_ctx * ctx,struct io_kiocb * req,unsigned int sqe_flags)1936 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1937 struct io_kiocb *req,
1938 unsigned int sqe_flags)
1939 {
1940 if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1941 return false;
1942
1943 if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1944 ctx->restrictions.sqe_flags_required)
1945 return false;
1946
1947 if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1948 ctx->restrictions.sqe_flags_required))
1949 return false;
1950
1951 return true;
1952 }
1953
io_init_req_drain(struct io_kiocb * req)1954 static void io_init_req_drain(struct io_kiocb *req)
1955 {
1956 struct io_ring_ctx *ctx = req->ctx;
1957 struct io_kiocb *head = ctx->submit_state.link.head;
1958
1959 ctx->drain_active = true;
1960 if (head) {
1961 /*
1962 * If we need to drain a request in the middle of a link, drain
1963 * the head request and the next request/link after the current
1964 * link. Considering sequential execution of links,
1965 * REQ_F_IO_DRAIN will be maintained for every request of our
1966 * link.
1967 */
1968 head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1969 ctx->drain_next = true;
1970 }
1971 }
1972
io_init_req(struct io_ring_ctx * ctx,struct io_kiocb * req,const struct io_uring_sqe * sqe)1973 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1974 const struct io_uring_sqe *sqe)
1975 __must_hold(&ctx->uring_lock)
1976 {
1977 const struct io_op_def *def;
1978 unsigned int sqe_flags;
1979 int personality;
1980 u8 opcode;
1981
1982 /* req is partially pre-initialised, see io_preinit_req() */
1983 req->opcode = opcode = READ_ONCE(sqe->opcode);
1984 /* same numerical values with corresponding REQ_F_*, safe to copy */
1985 req->flags = sqe_flags = READ_ONCE(sqe->flags);
1986 req->cqe.user_data = READ_ONCE(sqe->user_data);
1987 req->file = NULL;
1988 req->rsrc_node = NULL;
1989 req->task = current;
1990
1991 if (unlikely(opcode >= IORING_OP_LAST)) {
1992 req->opcode = 0;
1993 return -EINVAL;
1994 }
1995 def = &io_op_defs[opcode];
1996 if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
1997 /* enforce forwards compatibility on users */
1998 if (sqe_flags & ~SQE_VALID_FLAGS)
1999 return -EINVAL;
2000 if (sqe_flags & IOSQE_BUFFER_SELECT) {
2001 if (!def->buffer_select)
2002 return -EOPNOTSUPP;
2003 req->buf_index = READ_ONCE(sqe->buf_group);
2004 }
2005 if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2006 ctx->drain_disabled = true;
2007 if (sqe_flags & IOSQE_IO_DRAIN) {
2008 if (ctx->drain_disabled)
2009 return -EOPNOTSUPP;
2010 io_init_req_drain(req);
2011 }
2012 }
2013 if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2014 if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2015 return -EACCES;
2016 /* knock it to the slow queue path, will be drained there */
2017 if (ctx->drain_active)
2018 req->flags |= REQ_F_FORCE_ASYNC;
2019 /* if there is no link, we're at "next" request and need to drain */
2020 if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2021 ctx->drain_next = false;
2022 ctx->drain_active = true;
2023 req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2024 }
2025 }
2026
2027 if (!def->ioprio && sqe->ioprio)
2028 return -EINVAL;
2029 if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2030 return -EINVAL;
2031
2032 if (def->needs_file) {
2033 struct io_submit_state *state = &ctx->submit_state;
2034
2035 req->cqe.fd = READ_ONCE(sqe->fd);
2036
2037 /*
2038 * Plug now if we have more than 2 IO left after this, and the
2039 * target is potentially a read/write to block based storage.
2040 */
2041 if (state->need_plug && def->plug) {
2042 state->plug_started = true;
2043 state->need_plug = false;
2044 blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2045 }
2046 }
2047
2048 personality = READ_ONCE(sqe->personality);
2049 if (personality) {
2050 int ret;
2051
2052 req->creds = xa_load(&ctx->personalities, personality);
2053 if (!req->creds)
2054 return -EINVAL;
2055 get_cred(req->creds);
2056 ret = security_uring_override_creds(req->creds);
2057 if (ret) {
2058 put_cred(req->creds);
2059 return ret;
2060 }
2061 req->flags |= REQ_F_CREDS;
2062 }
2063
2064 return def->prep(req, sqe);
2065 }
2066
io_submit_fail_init(const struct io_uring_sqe * sqe,struct io_kiocb * req,int ret)2067 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2068 struct io_kiocb *req, int ret)
2069 {
2070 struct io_ring_ctx *ctx = req->ctx;
2071 struct io_submit_link *link = &ctx->submit_state.link;
2072 struct io_kiocb *head = link->head;
2073
2074 trace_io_uring_req_failed(sqe, req, ret);
2075
2076 /*
2077 * Avoid breaking links in the middle as it renders links with SQPOLL
2078 * unusable. Instead of failing eagerly, continue assembling the link if
2079 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2080 * should find the flag and handle the rest.
2081 */
2082 req_fail_link_node(req, ret);
2083 if (head && !(head->flags & REQ_F_FAIL))
2084 req_fail_link_node(head, -ECANCELED);
2085
2086 if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2087 if (head) {
2088 link->last->link = req;
2089 link->head = NULL;
2090 req = head;
2091 }
2092 io_queue_sqe_fallback(req);
2093 return ret;
2094 }
2095
2096 if (head)
2097 link->last->link = req;
2098 else
2099 link->head = req;
2100 link->last = req;
2101 return 0;
2102 }
2103
io_submit_sqe(struct io_ring_ctx * ctx,struct io_kiocb * req,const struct io_uring_sqe * sqe)2104 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2105 const struct io_uring_sqe *sqe)
2106 __must_hold(&ctx->uring_lock)
2107 {
2108 struct io_submit_link *link = &ctx->submit_state.link;
2109 int ret;
2110
2111 ret = io_init_req(ctx, req, sqe);
2112 if (unlikely(ret))
2113 return io_submit_fail_init(sqe, req, ret);
2114
2115 /* don't need @sqe from now on */
2116 trace_io_uring_submit_sqe(req, true);
2117
2118 /*
2119 * If we already have a head request, queue this one for async
2120 * submittal once the head completes. If we don't have a head but
2121 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2122 * submitted sync once the chain is complete. If none of those
2123 * conditions are true (normal request), then just queue it.
2124 */
2125 if (unlikely(link->head)) {
2126 ret = io_req_prep_async(req);
2127 if (unlikely(ret))
2128 return io_submit_fail_init(sqe, req, ret);
2129
2130 trace_io_uring_link(req, link->head);
2131 link->last->link = req;
2132 link->last = req;
2133
2134 if (req->flags & IO_REQ_LINK_FLAGS)
2135 return 0;
2136 /* last request of the link, flush it */
2137 req = link->head;
2138 link->head = NULL;
2139 if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2140 goto fallback;
2141
2142 } else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2143 REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2144 if (req->flags & IO_REQ_LINK_FLAGS) {
2145 link->head = req;
2146 link->last = req;
2147 } else {
2148 fallback:
2149 io_queue_sqe_fallback(req);
2150 }
2151 return 0;
2152 }
2153
2154 io_queue_sqe(req);
2155 return 0;
2156 }
2157
2158 /*
2159 * Batched submission is done, ensure local IO is flushed out.
2160 */
io_submit_state_end(struct io_ring_ctx * ctx)2161 static void io_submit_state_end(struct io_ring_ctx *ctx)
2162 {
2163 struct io_submit_state *state = &ctx->submit_state;
2164
2165 if (unlikely(state->link.head))
2166 io_queue_sqe_fallback(state->link.head);
2167 /* flush only after queuing links as they can generate completions */
2168 io_submit_flush_completions(ctx);
2169 if (state->plug_started)
2170 blk_finish_plug(&state->plug);
2171 }
2172
2173 /*
2174 * Start submission side cache.
2175 */
io_submit_state_start(struct io_submit_state * state,unsigned int max_ios)2176 static void io_submit_state_start(struct io_submit_state *state,
2177 unsigned int max_ios)
2178 {
2179 state->plug_started = false;
2180 state->need_plug = max_ios > 2;
2181 state->submit_nr = max_ios;
2182 /* set only head, no need to init link_last in advance */
2183 state->link.head = NULL;
2184 }
2185
io_commit_sqring(struct io_ring_ctx * ctx)2186 static void io_commit_sqring(struct io_ring_ctx *ctx)
2187 {
2188 struct io_rings *rings = ctx->rings;
2189
2190 /*
2191 * Ensure any loads from the SQEs are done at this point,
2192 * since once we write the new head, the application could
2193 * write new data to them.
2194 */
2195 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2196 }
2197
2198 /*
2199 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2200 * that is mapped by userspace. This means that care needs to be taken to
2201 * ensure that reads are stable, as we cannot rely on userspace always
2202 * being a good citizen. If members of the sqe are validated and then later
2203 * used, it's important that those reads are done through READ_ONCE() to
2204 * prevent a re-load down the line.
2205 */
io_get_sqe(struct io_ring_ctx * ctx)2206 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2207 {
2208 unsigned head, mask = ctx->sq_entries - 1;
2209 unsigned sq_idx = ctx->cached_sq_head++ & mask;
2210
2211 /*
2212 * The cached sq head (or cq tail) serves two purposes:
2213 *
2214 * 1) allows us to batch the cost of updating the user visible
2215 * head updates.
2216 * 2) allows the kernel side to track the head on its own, even
2217 * though the application is the one updating it.
2218 */
2219 head = READ_ONCE(ctx->sq_array[sq_idx]);
2220 if (likely(head < ctx->sq_entries)) {
2221 /* double index for 128-byte SQEs, twice as long */
2222 if (ctx->flags & IORING_SETUP_SQE128)
2223 head <<= 1;
2224 return &ctx->sq_sqes[head];
2225 }
2226
2227 /* drop invalid entries */
2228 ctx->cq_extra--;
2229 WRITE_ONCE(ctx->rings->sq_dropped,
2230 READ_ONCE(ctx->rings->sq_dropped) + 1);
2231 return NULL;
2232 }
2233
io_submit_sqes(struct io_ring_ctx * ctx,unsigned int nr)2234 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2235 __must_hold(&ctx->uring_lock)
2236 {
2237 unsigned int entries = io_sqring_entries(ctx);
2238 unsigned int left;
2239 int ret;
2240
2241 if (unlikely(!entries))
2242 return 0;
2243 /* make sure SQ entry isn't read before tail */
2244 ret = left = min3(nr, ctx->sq_entries, entries);
2245 io_get_task_refs(left);
2246 io_submit_state_start(&ctx->submit_state, left);
2247
2248 do {
2249 const struct io_uring_sqe *sqe;
2250 struct io_kiocb *req;
2251
2252 if (unlikely(!io_alloc_req_refill(ctx)))
2253 break;
2254 req = io_alloc_req(ctx);
2255 sqe = io_get_sqe(ctx);
2256 if (unlikely(!sqe)) {
2257 io_req_add_to_cache(req, ctx);
2258 break;
2259 }
2260
2261 /*
2262 * Continue submitting even for sqe failure if the
2263 * ring was setup with IORING_SETUP_SUBMIT_ALL
2264 */
2265 if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2266 !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2267 left--;
2268 break;
2269 }
2270 } while (--left);
2271
2272 if (unlikely(left)) {
2273 ret -= left;
2274 /* try again if it submitted nothing and can't allocate a req */
2275 if (!ret && io_req_cache_empty(ctx))
2276 ret = -EAGAIN;
2277 current->io_uring->cached_refs += left;
2278 }
2279
2280 io_submit_state_end(ctx);
2281 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2282 io_commit_sqring(ctx);
2283 return ret;
2284 }
2285
2286 struct io_wait_queue {
2287 struct wait_queue_entry wq;
2288 struct io_ring_ctx *ctx;
2289 unsigned cq_tail;
2290 unsigned nr_timeouts;
2291 };
2292
io_has_work(struct io_ring_ctx * ctx)2293 static inline bool io_has_work(struct io_ring_ctx *ctx)
2294 {
2295 return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2296 ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
2297 !llist_empty(&ctx->work_llist));
2298 }
2299
io_should_wake(struct io_wait_queue * iowq)2300 static inline bool io_should_wake(struct io_wait_queue *iowq)
2301 {
2302 struct io_ring_ctx *ctx = iowq->ctx;
2303 int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2304
2305 /*
2306 * Wake up if we have enough events, or if a timeout occurred since we
2307 * started waiting. For timeouts, we always want to return to userspace,
2308 * regardless of event count.
2309 */
2310 return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2311 }
2312
io_wake_function(struct wait_queue_entry * curr,unsigned int mode,int wake_flags,void * key)2313 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2314 int wake_flags, void *key)
2315 {
2316 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2317 wq);
2318 struct io_ring_ctx *ctx = iowq->ctx;
2319
2320 /*
2321 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2322 * the task, and the next invocation will do it.
2323 */
2324 if (io_should_wake(iowq) || io_has_work(ctx))
2325 return autoremove_wake_function(curr, mode, wake_flags, key);
2326 return -1;
2327 }
2328
io_run_task_work_sig(struct io_ring_ctx * ctx)2329 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2330 {
2331 if (io_run_task_work_ctx(ctx) > 0)
2332 return 1;
2333 if (task_sigpending(current))
2334 return -EINTR;
2335 return 0;
2336 }
2337
2338 /* when returns >0, the caller should retry */
io_cqring_wait_schedule(struct io_ring_ctx * ctx,struct io_wait_queue * iowq,ktime_t * timeout)2339 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2340 struct io_wait_queue *iowq,
2341 ktime_t *timeout)
2342 {
2343 int ret;
2344 unsigned long check_cq;
2345
2346 /* make sure we run task_work before checking for signals */
2347 ret = io_run_task_work_sig(ctx);
2348 if (ret || io_should_wake(iowq))
2349 return ret;
2350
2351 check_cq = READ_ONCE(ctx->check_cq);
2352 if (unlikely(check_cq)) {
2353 /* let the caller flush overflows, retry */
2354 if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2355 return 1;
2356 if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2357 return -EBADR;
2358 }
2359 if (!schedule_hrtimeout(timeout, HRTIMER_MODE_ABS))
2360 return -ETIME;
2361 return 1;
2362 }
2363
2364 /*
2365 * Wait until events become available, if we don't already have some. The
2366 * application must reap them itself, as they reside on the shared cq ring.
2367 */
io_cqring_wait(struct io_ring_ctx * ctx,int min_events,const sigset_t __user * sig,size_t sigsz,struct __kernel_timespec __user * uts)2368 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2369 const sigset_t __user *sig, size_t sigsz,
2370 struct __kernel_timespec __user *uts)
2371 {
2372 struct io_wait_queue iowq;
2373 struct io_rings *rings = ctx->rings;
2374 ktime_t timeout = KTIME_MAX;
2375 int ret;
2376
2377 if (!io_allowed_run_tw(ctx))
2378 return -EEXIST;
2379
2380 do {
2381 /* always run at least 1 task work to process local work */
2382 ret = io_run_task_work_ctx(ctx);
2383 if (ret < 0)
2384 return ret;
2385 io_cqring_overflow_flush(ctx);
2386
2387 /* if user messes with these they will just get an early return */
2388 if (__io_cqring_events_user(ctx) >= min_events)
2389 return 0;
2390 } while (ret > 0);
2391
2392 if (sig) {
2393 #ifdef CONFIG_COMPAT
2394 if (in_compat_syscall())
2395 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2396 sigsz);
2397 else
2398 #endif
2399 ret = set_user_sigmask(sig, sigsz);
2400
2401 if (ret)
2402 return ret;
2403 }
2404
2405 if (uts) {
2406 struct timespec64 ts;
2407
2408 if (get_timespec64(&ts, uts))
2409 return -EFAULT;
2410 timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2411 }
2412
2413 init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2414 iowq.wq.private = current;
2415 INIT_LIST_HEAD(&iowq.wq.entry);
2416 iowq.ctx = ctx;
2417 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2418 iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2419
2420 trace_io_uring_cqring_wait(ctx, min_events);
2421 do {
2422 /* if we can't even flush overflow, don't wait for more */
2423 if (!io_cqring_overflow_flush(ctx)) {
2424 ret = -EBUSY;
2425 break;
2426 }
2427 prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2428 TASK_INTERRUPTIBLE);
2429 ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
2430 cond_resched();
2431 } while (ret > 0);
2432
2433 finish_wait(&ctx->cq_wait, &iowq.wq);
2434 restore_saved_sigmask_unless(ret == -EINTR);
2435
2436 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2437 }
2438
io_mem_free(void * ptr)2439 static void io_mem_free(void *ptr)
2440 {
2441 struct page *page;
2442
2443 if (!ptr)
2444 return;
2445
2446 page = virt_to_head_page(ptr);
2447 if (put_page_testzero(page))
2448 free_compound_page(page);
2449 }
2450
io_mem_alloc(size_t size)2451 static void *io_mem_alloc(size_t size)
2452 {
2453 gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2454
2455 return (void *) __get_free_pages(gfp, get_order(size));
2456 }
2457
rings_size(struct io_ring_ctx * ctx,unsigned int sq_entries,unsigned int cq_entries,size_t * sq_offset)2458 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2459 unsigned int cq_entries, size_t *sq_offset)
2460 {
2461 struct io_rings *rings;
2462 size_t off, sq_array_size;
2463
2464 off = struct_size(rings, cqes, cq_entries);
2465 if (off == SIZE_MAX)
2466 return SIZE_MAX;
2467 if (ctx->flags & IORING_SETUP_CQE32) {
2468 if (check_shl_overflow(off, 1, &off))
2469 return SIZE_MAX;
2470 }
2471
2472 #ifdef CONFIG_SMP
2473 off = ALIGN(off, SMP_CACHE_BYTES);
2474 if (off == 0)
2475 return SIZE_MAX;
2476 #endif
2477
2478 if (sq_offset)
2479 *sq_offset = off;
2480
2481 sq_array_size = array_size(sizeof(u32), sq_entries);
2482 if (sq_array_size == SIZE_MAX)
2483 return SIZE_MAX;
2484
2485 if (check_add_overflow(off, sq_array_size, &off))
2486 return SIZE_MAX;
2487
2488 return off;
2489 }
2490
io_eventfd_register(struct io_ring_ctx * ctx,void __user * arg,unsigned int eventfd_async)2491 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2492 unsigned int eventfd_async)
2493 {
2494 struct io_ev_fd *ev_fd;
2495 __s32 __user *fds = arg;
2496 int fd;
2497
2498 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2499 lockdep_is_held(&ctx->uring_lock));
2500 if (ev_fd)
2501 return -EBUSY;
2502
2503 if (copy_from_user(&fd, fds, sizeof(*fds)))
2504 return -EFAULT;
2505
2506 ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2507 if (!ev_fd)
2508 return -ENOMEM;
2509
2510 ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2511 if (IS_ERR(ev_fd->cq_ev_fd)) {
2512 int ret = PTR_ERR(ev_fd->cq_ev_fd);
2513 kfree(ev_fd);
2514 return ret;
2515 }
2516
2517 spin_lock(&ctx->completion_lock);
2518 ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2519 spin_unlock(&ctx->completion_lock);
2520
2521 ev_fd->eventfd_async = eventfd_async;
2522 ctx->has_evfd = true;
2523 rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2524 atomic_set(&ev_fd->refs, 1);
2525 atomic_set(&ev_fd->ops, 0);
2526 return 0;
2527 }
2528
io_eventfd_unregister(struct io_ring_ctx * ctx)2529 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2530 {
2531 struct io_ev_fd *ev_fd;
2532
2533 ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2534 lockdep_is_held(&ctx->uring_lock));
2535 if (ev_fd) {
2536 ctx->has_evfd = false;
2537 rcu_assign_pointer(ctx->io_ev_fd, NULL);
2538 if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2539 call_rcu(&ev_fd->rcu, io_eventfd_ops);
2540 return 0;
2541 }
2542
2543 return -ENXIO;
2544 }
2545
io_req_caches_free(struct io_ring_ctx * ctx)2546 static void io_req_caches_free(struct io_ring_ctx *ctx)
2547 {
2548 int nr = 0;
2549
2550 mutex_lock(&ctx->uring_lock);
2551 io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2552
2553 while (!io_req_cache_empty(ctx)) {
2554 struct io_kiocb *req = io_alloc_req(ctx);
2555
2556 kmem_cache_free(req_cachep, req);
2557 nr++;
2558 }
2559 if (nr)
2560 percpu_ref_put_many(&ctx->refs, nr);
2561 mutex_unlock(&ctx->uring_lock);
2562 }
2563
io_ring_ctx_free(struct io_ring_ctx * ctx)2564 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2565 {
2566 io_sq_thread_finish(ctx);
2567 io_rsrc_refs_drop(ctx);
2568 /* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2569 io_wait_rsrc_data(ctx->buf_data);
2570 io_wait_rsrc_data(ctx->file_data);
2571
2572 mutex_lock(&ctx->uring_lock);
2573 if (ctx->buf_data)
2574 __io_sqe_buffers_unregister(ctx);
2575 if (ctx->file_data)
2576 __io_sqe_files_unregister(ctx);
2577 if (ctx->rings)
2578 __io_cqring_overflow_flush(ctx, true);
2579 io_eventfd_unregister(ctx);
2580 io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2581 io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2582 mutex_unlock(&ctx->uring_lock);
2583 io_destroy_buffers(ctx);
2584 if (ctx->sq_creds)
2585 put_cred(ctx->sq_creds);
2586 if (ctx->submitter_task)
2587 put_task_struct(ctx->submitter_task);
2588
2589 /* there are no registered resources left, nobody uses it */
2590 if (ctx->rsrc_node)
2591 io_rsrc_node_destroy(ctx->rsrc_node);
2592 if (ctx->rsrc_backup_node)
2593 io_rsrc_node_destroy(ctx->rsrc_backup_node);
2594 flush_delayed_work(&ctx->rsrc_put_work);
2595 flush_delayed_work(&ctx->fallback_work);
2596
2597 WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2598 WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2599
2600 #if defined(CONFIG_UNIX)
2601 if (ctx->ring_sock) {
2602 ctx->ring_sock->file = NULL; /* so that iput() is called */
2603 sock_release(ctx->ring_sock);
2604 }
2605 #endif
2606 WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2607
2608 if (ctx->mm_account) {
2609 mmdrop(ctx->mm_account);
2610 ctx->mm_account = NULL;
2611 }
2612 io_mem_free(ctx->rings);
2613 io_mem_free(ctx->sq_sqes);
2614
2615 percpu_ref_exit(&ctx->refs);
2616 free_uid(ctx->user);
2617 io_req_caches_free(ctx);
2618 if (ctx->hash_map)
2619 io_wq_put_hash(ctx->hash_map);
2620 kfree(ctx->cancel_table.hbs);
2621 kfree(ctx->cancel_table_locked.hbs);
2622 kfree(ctx->dummy_ubuf);
2623 kfree(ctx->io_bl);
2624 xa_destroy(&ctx->io_bl_xa);
2625 kfree(ctx);
2626 }
2627
io_uring_poll(struct file * file,poll_table * wait)2628 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2629 {
2630 struct io_ring_ctx *ctx = file->private_data;
2631 __poll_t mask = 0;
2632
2633 poll_wait(file, &ctx->cq_wait, wait);
2634 /*
2635 * synchronizes with barrier from wq_has_sleeper call in
2636 * io_commit_cqring
2637 */
2638 smp_rmb();
2639 if (!io_sqring_full(ctx))
2640 mask |= EPOLLOUT | EPOLLWRNORM;
2641
2642 /*
2643 * Don't flush cqring overflow list here, just do a simple check.
2644 * Otherwise there could possible be ABBA deadlock:
2645 * CPU0 CPU1
2646 * ---- ----
2647 * lock(&ctx->uring_lock);
2648 * lock(&ep->mtx);
2649 * lock(&ctx->uring_lock);
2650 * lock(&ep->mtx);
2651 *
2652 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2653 * pushs them to do the flush.
2654 */
2655
2656 if (io_cqring_events(ctx) || io_has_work(ctx))
2657 mask |= EPOLLIN | EPOLLRDNORM;
2658
2659 return mask;
2660 }
2661
io_unregister_personality(struct io_ring_ctx * ctx,unsigned id)2662 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2663 {
2664 const struct cred *creds;
2665
2666 creds = xa_erase(&ctx->personalities, id);
2667 if (creds) {
2668 put_cred(creds);
2669 return 0;
2670 }
2671
2672 return -EINVAL;
2673 }
2674
2675 struct io_tctx_exit {
2676 struct callback_head task_work;
2677 struct completion completion;
2678 struct io_ring_ctx *ctx;
2679 };
2680
io_tctx_exit_cb(struct callback_head * cb)2681 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2682 {
2683 struct io_uring_task *tctx = current->io_uring;
2684 struct io_tctx_exit *work;
2685
2686 work = container_of(cb, struct io_tctx_exit, task_work);
2687 /*
2688 * When @in_idle, we're in cancellation and it's racy to remove the
2689 * node. It'll be removed by the end of cancellation, just ignore it.
2690 * tctx can be NULL if the queueing of this task_work raced with
2691 * work cancelation off the exec path.
2692 */
2693 if (tctx && !atomic_read(&tctx->in_idle))
2694 io_uring_del_tctx_node((unsigned long)work->ctx);
2695 complete(&work->completion);
2696 }
2697
io_cancel_ctx_cb(struct io_wq_work * work,void * data)2698 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2699 {
2700 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2701
2702 return req->ctx == data;
2703 }
2704
io_ring_exit_work(struct work_struct * work)2705 static __cold void io_ring_exit_work(struct work_struct *work)
2706 {
2707 struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2708 unsigned long timeout = jiffies + HZ * 60 * 5;
2709 unsigned long interval = HZ / 20;
2710 struct io_tctx_exit exit;
2711 struct io_tctx_node *node;
2712 int ret;
2713
2714 /*
2715 * If we're doing polled IO and end up having requests being
2716 * submitted async (out-of-line), then completions can come in while
2717 * we're waiting for refs to drop. We need to reap these manually,
2718 * as nobody else will be looking for them.
2719 */
2720 do {
2721 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2722 io_move_task_work_from_local(ctx);
2723
2724 while (io_uring_try_cancel_requests(ctx, NULL, true))
2725 cond_resched();
2726
2727 if (ctx->sq_data) {
2728 struct io_sq_data *sqd = ctx->sq_data;
2729 struct task_struct *tsk;
2730
2731 io_sq_thread_park(sqd);
2732 tsk = sqd->thread;
2733 if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2734 io_wq_cancel_cb(tsk->io_uring->io_wq,
2735 io_cancel_ctx_cb, ctx, true);
2736 io_sq_thread_unpark(sqd);
2737 }
2738
2739 io_req_caches_free(ctx);
2740
2741 if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2742 /* there is little hope left, don't run it too often */
2743 interval = HZ * 60;
2744 }
2745 } while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2746
2747 init_completion(&exit.completion);
2748 init_task_work(&exit.task_work, io_tctx_exit_cb);
2749 exit.ctx = ctx;
2750 /*
2751 * Some may use context even when all refs and requests have been put,
2752 * and they are free to do so while still holding uring_lock or
2753 * completion_lock, see io_req_task_submit(). Apart from other work,
2754 * this lock/unlock section also waits them to finish.
2755 */
2756 mutex_lock(&ctx->uring_lock);
2757 while (!list_empty(&ctx->tctx_list)) {
2758 WARN_ON_ONCE(time_after(jiffies, timeout));
2759
2760 node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2761 ctx_node);
2762 /* don't spin on a single task if cancellation failed */
2763 list_rotate_left(&ctx->tctx_list);
2764 ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2765 if (WARN_ON_ONCE(ret))
2766 continue;
2767
2768 mutex_unlock(&ctx->uring_lock);
2769 wait_for_completion(&exit.completion);
2770 mutex_lock(&ctx->uring_lock);
2771 }
2772 mutex_unlock(&ctx->uring_lock);
2773 spin_lock(&ctx->completion_lock);
2774 spin_unlock(&ctx->completion_lock);
2775
2776 io_ring_ctx_free(ctx);
2777 }
2778
io_ring_ctx_wait_and_kill(struct io_ring_ctx * ctx)2779 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2780 {
2781 unsigned long index;
2782 struct creds *creds;
2783
2784 mutex_lock(&ctx->uring_lock);
2785 percpu_ref_kill(&ctx->refs);
2786 if (ctx->rings)
2787 __io_cqring_overflow_flush(ctx, true);
2788 xa_for_each(&ctx->personalities, index, creds)
2789 io_unregister_personality(ctx, index);
2790 if (ctx->rings)
2791 io_poll_remove_all(ctx, NULL, true);
2792 mutex_unlock(&ctx->uring_lock);
2793
2794 /*
2795 * If we failed setting up the ctx, we might not have any rings
2796 * and therefore did not submit any requests
2797 */
2798 if (ctx->rings)
2799 io_kill_timeouts(ctx, NULL, true);
2800
2801 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2802 /*
2803 * Use system_unbound_wq to avoid spawning tons of event kworkers
2804 * if we're exiting a ton of rings at the same time. It just adds
2805 * noise and overhead, there's no discernable change in runtime
2806 * over using system_wq.
2807 */
2808 queue_work(system_unbound_wq, &ctx->exit_work);
2809 }
2810
io_uring_release(struct inode * inode,struct file * file)2811 static int io_uring_release(struct inode *inode, struct file *file)
2812 {
2813 struct io_ring_ctx *ctx = file->private_data;
2814
2815 file->private_data = NULL;
2816 io_ring_ctx_wait_and_kill(ctx);
2817 return 0;
2818 }
2819
2820 struct io_task_cancel {
2821 struct task_struct *task;
2822 bool all;
2823 };
2824
io_cancel_task_cb(struct io_wq_work * work,void * data)2825 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2826 {
2827 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2828 struct io_task_cancel *cancel = data;
2829
2830 return io_match_task_safe(req, cancel->task, cancel->all);
2831 }
2832
io_cancel_defer_files(struct io_ring_ctx * ctx,struct task_struct * task,bool cancel_all)2833 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2834 struct task_struct *task,
2835 bool cancel_all)
2836 {
2837 struct io_defer_entry *de;
2838 LIST_HEAD(list);
2839
2840 spin_lock(&ctx->completion_lock);
2841 list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2842 if (io_match_task_safe(de->req, task, cancel_all)) {
2843 list_cut_position(&list, &ctx->defer_list, &de->list);
2844 break;
2845 }
2846 }
2847 spin_unlock(&ctx->completion_lock);
2848 if (list_empty(&list))
2849 return false;
2850
2851 while (!list_empty(&list)) {
2852 de = list_first_entry(&list, struct io_defer_entry, list);
2853 list_del_init(&de->list);
2854 io_req_task_queue_fail(de->req, -ECANCELED);
2855 kfree(de);
2856 }
2857 return true;
2858 }
2859
io_uring_try_cancel_iowq(struct io_ring_ctx * ctx)2860 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2861 {
2862 struct io_tctx_node *node;
2863 enum io_wq_cancel cret;
2864 bool ret = false;
2865
2866 mutex_lock(&ctx->uring_lock);
2867 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2868 struct io_uring_task *tctx = node->task->io_uring;
2869
2870 /*
2871 * io_wq will stay alive while we hold uring_lock, because it's
2872 * killed after ctx nodes, which requires to take the lock.
2873 */
2874 if (!tctx || !tctx->io_wq)
2875 continue;
2876 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2877 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2878 }
2879 mutex_unlock(&ctx->uring_lock);
2880
2881 return ret;
2882 }
2883
io_uring_try_cancel_requests(struct io_ring_ctx * ctx,struct task_struct * task,bool cancel_all)2884 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2885 struct task_struct *task,
2886 bool cancel_all)
2887 {
2888 struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2889 struct io_uring_task *tctx = task ? task->io_uring : NULL;
2890 enum io_wq_cancel cret;
2891 bool ret = false;
2892
2893 /* failed during ring init, it couldn't have issued any requests */
2894 if (!ctx->rings)
2895 return false;
2896
2897 if (!task) {
2898 ret |= io_uring_try_cancel_iowq(ctx);
2899 } else if (tctx && tctx->io_wq) {
2900 /*
2901 * Cancels requests of all rings, not only @ctx, but
2902 * it's fine as the task is in exit/exec.
2903 */
2904 cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2905 &cancel, true);
2906 ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2907 }
2908
2909 /* SQPOLL thread does its own polling */
2910 if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2911 (ctx->sq_data && ctx->sq_data->thread == current)) {
2912 while (!wq_list_empty(&ctx->iopoll_list)) {
2913 io_iopoll_try_reap_events(ctx);
2914 ret = true;
2915 }
2916 }
2917
2918 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2919 ret |= io_run_local_work(ctx) > 0;
2920 ret |= io_cancel_defer_files(ctx, task, cancel_all);
2921 mutex_lock(&ctx->uring_lock);
2922 ret |= io_poll_remove_all(ctx, task, cancel_all);
2923 mutex_unlock(&ctx->uring_lock);
2924 ret |= io_kill_timeouts(ctx, task, cancel_all);
2925 if (task)
2926 ret |= io_run_task_work() > 0;
2927 return ret;
2928 }
2929
tctx_inflight(struct io_uring_task * tctx,bool tracked)2930 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2931 {
2932 if (tracked)
2933 return atomic_read(&tctx->inflight_tracked);
2934 return percpu_counter_sum(&tctx->inflight);
2935 }
2936
2937 /*
2938 * Find any io_uring ctx that this task has registered or done IO on, and cancel
2939 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2940 */
io_uring_cancel_generic(bool cancel_all,struct io_sq_data * sqd)2941 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2942 {
2943 struct io_uring_task *tctx = current->io_uring;
2944 struct io_ring_ctx *ctx;
2945 s64 inflight;
2946 DEFINE_WAIT(wait);
2947
2948 WARN_ON_ONCE(sqd && sqd->thread != current);
2949
2950 if (!current->io_uring)
2951 return;
2952 if (tctx->io_wq)
2953 io_wq_exit_start(tctx->io_wq);
2954
2955 atomic_inc(&tctx->in_idle);
2956 do {
2957 bool loop = false;
2958
2959 io_uring_drop_tctx_refs(current);
2960 /* read completions before cancelations */
2961 inflight = tctx_inflight(tctx, !cancel_all);
2962 if (!inflight)
2963 break;
2964
2965 if (!sqd) {
2966 struct io_tctx_node *node;
2967 unsigned long index;
2968
2969 xa_for_each(&tctx->xa, index, node) {
2970 /* sqpoll task will cancel all its requests */
2971 if (node->ctx->sq_data)
2972 continue;
2973 loop |= io_uring_try_cancel_requests(node->ctx,
2974 current, cancel_all);
2975 }
2976 } else {
2977 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
2978 loop |= io_uring_try_cancel_requests(ctx,
2979 current,
2980 cancel_all);
2981 }
2982
2983 if (loop) {
2984 cond_resched();
2985 continue;
2986 }
2987
2988 prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
2989 io_run_task_work();
2990 io_uring_drop_tctx_refs(current);
2991
2992 /*
2993 * If we've seen completions, retry without waiting. This
2994 * avoids a race where a completion comes in before we did
2995 * prepare_to_wait().
2996 */
2997 if (inflight == tctx_inflight(tctx, !cancel_all))
2998 schedule();
2999 finish_wait(&tctx->wait, &wait);
3000 } while (1);
3001
3002 io_uring_clean_tctx(tctx);
3003 if (cancel_all) {
3004 /*
3005 * We shouldn't run task_works after cancel, so just leave
3006 * ->in_idle set for normal exit.
3007 */
3008 atomic_dec(&tctx->in_idle);
3009 /* for exec all current's requests should be gone, kill tctx */
3010 __io_uring_free(current);
3011 }
3012 }
3013
__io_uring_cancel(bool cancel_all)3014 void __io_uring_cancel(bool cancel_all)
3015 {
3016 io_uring_cancel_generic(cancel_all, NULL);
3017 }
3018
io_uring_validate_mmap_request(struct file * file,loff_t pgoff,size_t sz)3019 static void *io_uring_validate_mmap_request(struct file *file,
3020 loff_t pgoff, size_t sz)
3021 {
3022 struct io_ring_ctx *ctx = file->private_data;
3023 loff_t offset = pgoff << PAGE_SHIFT;
3024 struct page *page;
3025 void *ptr;
3026
3027 switch (offset) {
3028 case IORING_OFF_SQ_RING:
3029 case IORING_OFF_CQ_RING:
3030 ptr = ctx->rings;
3031 break;
3032 case IORING_OFF_SQES:
3033 ptr = ctx->sq_sqes;
3034 break;
3035 default:
3036 return ERR_PTR(-EINVAL);
3037 }
3038
3039 page = virt_to_head_page(ptr);
3040 if (sz > page_size(page))
3041 return ERR_PTR(-EINVAL);
3042
3043 return ptr;
3044 }
3045
3046 #ifdef CONFIG_MMU
3047
io_uring_mmap(struct file * file,struct vm_area_struct * vma)3048 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3049 {
3050 size_t sz = vma->vm_end - vma->vm_start;
3051 unsigned long pfn;
3052 void *ptr;
3053
3054 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3055 if (IS_ERR(ptr))
3056 return PTR_ERR(ptr);
3057
3058 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3059 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3060 }
3061
3062 #else /* !CONFIG_MMU */
3063
io_uring_mmap(struct file * file,struct vm_area_struct * vma)3064 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3065 {
3066 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
3067 }
3068
io_uring_nommu_mmap_capabilities(struct file * file)3069 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3070 {
3071 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3072 }
3073
io_uring_nommu_get_unmapped_area(struct file * file,unsigned long addr,unsigned long len,unsigned long pgoff,unsigned long flags)3074 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3075 unsigned long addr, unsigned long len,
3076 unsigned long pgoff, unsigned long flags)
3077 {
3078 void *ptr;
3079
3080 ptr = io_uring_validate_mmap_request(file, pgoff, len);
3081 if (IS_ERR(ptr))
3082 return PTR_ERR(ptr);
3083
3084 return (unsigned long) ptr;
3085 }
3086
3087 #endif /* !CONFIG_MMU */
3088
io_validate_ext_arg(unsigned flags,const void __user * argp,size_t argsz)3089 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3090 {
3091 if (flags & IORING_ENTER_EXT_ARG) {
3092 struct io_uring_getevents_arg arg;
3093
3094 if (argsz != sizeof(arg))
3095 return -EINVAL;
3096 if (copy_from_user(&arg, argp, sizeof(arg)))
3097 return -EFAULT;
3098 }
3099 return 0;
3100 }
3101
io_get_ext_arg(unsigned flags,const void __user * argp,size_t * argsz,struct __kernel_timespec __user ** ts,const sigset_t __user ** sig)3102 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3103 struct __kernel_timespec __user **ts,
3104 const sigset_t __user **sig)
3105 {
3106 struct io_uring_getevents_arg arg;
3107
3108 /*
3109 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3110 * is just a pointer to the sigset_t.
3111 */
3112 if (!(flags & IORING_ENTER_EXT_ARG)) {
3113 *sig = (const sigset_t __user *) argp;
3114 *ts = NULL;
3115 return 0;
3116 }
3117
3118 /*
3119 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3120 * timespec and sigset_t pointers if good.
3121 */
3122 if (*argsz != sizeof(arg))
3123 return -EINVAL;
3124 if (copy_from_user(&arg, argp, sizeof(arg)))
3125 return -EFAULT;
3126 if (arg.pad)
3127 return -EINVAL;
3128 *sig = u64_to_user_ptr(arg.sigmask);
3129 *argsz = arg.sigmask_sz;
3130 *ts = u64_to_user_ptr(arg.ts);
3131 return 0;
3132 }
3133
SYSCALL_DEFINE6(io_uring_enter,unsigned int,fd,u32,to_submit,u32,min_complete,u32,flags,const void __user *,argp,size_t,argsz)3134 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3135 u32, min_complete, u32, flags, const void __user *, argp,
3136 size_t, argsz)
3137 {
3138 struct io_ring_ctx *ctx;
3139 struct fd f;
3140 long ret;
3141
3142 if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3143 IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3144 IORING_ENTER_REGISTERED_RING)))
3145 return -EINVAL;
3146
3147 /*
3148 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3149 * need only dereference our task private array to find it.
3150 */
3151 if (flags & IORING_ENTER_REGISTERED_RING) {
3152 struct io_uring_task *tctx = current->io_uring;
3153
3154 if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3155 return -EINVAL;
3156 fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3157 f.file = tctx->registered_rings[fd];
3158 f.flags = 0;
3159 if (unlikely(!f.file))
3160 return -EBADF;
3161 } else {
3162 f = fdget(fd);
3163 if (unlikely(!f.file))
3164 return -EBADF;
3165 ret = -EOPNOTSUPP;
3166 if (unlikely(!io_is_uring_fops(f.file)))
3167 goto out;
3168 }
3169
3170 ctx = f.file->private_data;
3171 ret = -EBADFD;
3172 if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3173 goto out;
3174
3175 /*
3176 * For SQ polling, the thread will do all submissions and completions.
3177 * Just return the requested submit count, and wake the thread if
3178 * we were asked to.
3179 */
3180 ret = 0;
3181 if (ctx->flags & IORING_SETUP_SQPOLL) {
3182 io_cqring_overflow_flush(ctx);
3183
3184 if (unlikely(ctx->sq_data->thread == NULL)) {
3185 ret = -EOWNERDEAD;
3186 goto out;
3187 }
3188 if (flags & IORING_ENTER_SQ_WAKEUP)
3189 wake_up(&ctx->sq_data->wait);
3190 if (flags & IORING_ENTER_SQ_WAIT) {
3191 ret = io_sqpoll_wait_sq(ctx);
3192 if (ret)
3193 goto out;
3194 }
3195 ret = to_submit;
3196 } else if (to_submit) {
3197 ret = io_uring_add_tctx_node(ctx);
3198 if (unlikely(ret))
3199 goto out;
3200
3201 mutex_lock(&ctx->uring_lock);
3202 ret = io_submit_sqes(ctx, to_submit);
3203 if (ret != to_submit) {
3204 mutex_unlock(&ctx->uring_lock);
3205 goto out;
3206 }
3207 if (flags & IORING_ENTER_GETEVENTS) {
3208 if (ctx->syscall_iopoll)
3209 goto iopoll_locked;
3210 /*
3211 * Ignore errors, we'll soon call io_cqring_wait() and
3212 * it should handle ownership problems if any.
3213 */
3214 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3215 (void)io_run_local_work_locked(ctx);
3216 }
3217 mutex_unlock(&ctx->uring_lock);
3218 }
3219
3220 if (flags & IORING_ENTER_GETEVENTS) {
3221 int ret2;
3222
3223 if (ctx->syscall_iopoll) {
3224 /*
3225 * We disallow the app entering submit/complete with
3226 * polling, but we still need to lock the ring to
3227 * prevent racing with polled issue that got punted to
3228 * a workqueue.
3229 */
3230 mutex_lock(&ctx->uring_lock);
3231 iopoll_locked:
3232 ret2 = io_validate_ext_arg(flags, argp, argsz);
3233 if (likely(!ret2)) {
3234 min_complete = min(min_complete,
3235 ctx->cq_entries);
3236 ret2 = io_iopoll_check(ctx, min_complete);
3237 }
3238 mutex_unlock(&ctx->uring_lock);
3239 } else {
3240 const sigset_t __user *sig;
3241 struct __kernel_timespec __user *ts;
3242
3243 ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3244 if (likely(!ret2)) {
3245 min_complete = min(min_complete,
3246 ctx->cq_entries);
3247 ret2 = io_cqring_wait(ctx, min_complete, sig,
3248 argsz, ts);
3249 }
3250 }
3251
3252 if (!ret) {
3253 ret = ret2;
3254
3255 /*
3256 * EBADR indicates that one or more CQE were dropped.
3257 * Once the user has been informed we can clear the bit
3258 * as they are obviously ok with those drops.
3259 */
3260 if (unlikely(ret2 == -EBADR))
3261 clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3262 &ctx->check_cq);
3263 }
3264 }
3265 out:
3266 fdput(f);
3267 return ret;
3268 }
3269
3270 static const struct file_operations io_uring_fops = {
3271 .release = io_uring_release,
3272 .mmap = io_uring_mmap,
3273 #ifndef CONFIG_MMU
3274 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
3275 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
3276 #endif
3277 .poll = io_uring_poll,
3278 #ifdef CONFIG_PROC_FS
3279 .show_fdinfo = io_uring_show_fdinfo,
3280 #endif
3281 };
3282
io_is_uring_fops(struct file * file)3283 bool io_is_uring_fops(struct file *file)
3284 {
3285 return file->f_op == &io_uring_fops;
3286 }
3287
io_allocate_scq_urings(struct io_ring_ctx * ctx,struct io_uring_params * p)3288 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3289 struct io_uring_params *p)
3290 {
3291 struct io_rings *rings;
3292 size_t size, sq_array_offset;
3293
3294 /* make sure these are sane, as we already accounted them */
3295 ctx->sq_entries = p->sq_entries;
3296 ctx->cq_entries = p->cq_entries;
3297
3298 size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3299 if (size == SIZE_MAX)
3300 return -EOVERFLOW;
3301
3302 rings = io_mem_alloc(size);
3303 if (!rings)
3304 return -ENOMEM;
3305
3306 ctx->rings = rings;
3307 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3308 rings->sq_ring_mask = p->sq_entries - 1;
3309 rings->cq_ring_mask = p->cq_entries - 1;
3310 rings->sq_ring_entries = p->sq_entries;
3311 rings->cq_ring_entries = p->cq_entries;
3312
3313 if (p->flags & IORING_SETUP_SQE128)
3314 size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3315 else
3316 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3317 if (size == SIZE_MAX) {
3318 io_mem_free(ctx->rings);
3319 ctx->rings = NULL;
3320 return -EOVERFLOW;
3321 }
3322
3323 ctx->sq_sqes = io_mem_alloc(size);
3324 if (!ctx->sq_sqes) {
3325 io_mem_free(ctx->rings);
3326 ctx->rings = NULL;
3327 return -ENOMEM;
3328 }
3329
3330 return 0;
3331 }
3332
io_uring_install_fd(struct io_ring_ctx * ctx,struct file * file)3333 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3334 {
3335 int ret, fd;
3336
3337 fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3338 if (fd < 0)
3339 return fd;
3340
3341 ret = __io_uring_add_tctx_node(ctx);
3342 if (ret) {
3343 put_unused_fd(fd);
3344 return ret;
3345 }
3346 fd_install(fd, file);
3347 return fd;
3348 }
3349
3350 /*
3351 * Allocate an anonymous fd, this is what constitutes the application
3352 * visible backing of an io_uring instance. The application mmaps this
3353 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3354 * we have to tie this fd to a socket for file garbage collection purposes.
3355 */
io_uring_get_file(struct io_ring_ctx * ctx)3356 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3357 {
3358 struct file *file;
3359 #if defined(CONFIG_UNIX)
3360 int ret;
3361
3362 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3363 &ctx->ring_sock);
3364 if (ret)
3365 return ERR_PTR(ret);
3366 #endif
3367
3368 file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3369 O_RDWR | O_CLOEXEC, NULL);
3370 #if defined(CONFIG_UNIX)
3371 if (IS_ERR(file)) {
3372 sock_release(ctx->ring_sock);
3373 ctx->ring_sock = NULL;
3374 } else {
3375 ctx->ring_sock->file = file;
3376 }
3377 #endif
3378 return file;
3379 }
3380
io_uring_create(unsigned entries,struct io_uring_params * p,struct io_uring_params __user * params)3381 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3382 struct io_uring_params __user *params)
3383 {
3384 struct io_ring_ctx *ctx;
3385 struct file *file;
3386 int ret;
3387
3388 if (!entries)
3389 return -EINVAL;
3390 if (entries > IORING_MAX_ENTRIES) {
3391 if (!(p->flags & IORING_SETUP_CLAMP))
3392 return -EINVAL;
3393 entries = IORING_MAX_ENTRIES;
3394 }
3395
3396 /*
3397 * Use twice as many entries for the CQ ring. It's possible for the
3398 * application to drive a higher depth than the size of the SQ ring,
3399 * since the sqes are only used at submission time. This allows for
3400 * some flexibility in overcommitting a bit. If the application has
3401 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3402 * of CQ ring entries manually.
3403 */
3404 p->sq_entries = roundup_pow_of_two(entries);
3405 if (p->flags & IORING_SETUP_CQSIZE) {
3406 /*
3407 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3408 * to a power-of-two, if it isn't already. We do NOT impose
3409 * any cq vs sq ring sizing.
3410 */
3411 if (!p->cq_entries)
3412 return -EINVAL;
3413 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3414 if (!(p->flags & IORING_SETUP_CLAMP))
3415 return -EINVAL;
3416 p->cq_entries = IORING_MAX_CQ_ENTRIES;
3417 }
3418 p->cq_entries = roundup_pow_of_two(p->cq_entries);
3419 if (p->cq_entries < p->sq_entries)
3420 return -EINVAL;
3421 } else {
3422 p->cq_entries = 2 * p->sq_entries;
3423 }
3424
3425 ctx = io_ring_ctx_alloc(p);
3426 if (!ctx)
3427 return -ENOMEM;
3428
3429 /*
3430 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3431 * space applications don't need to do io completion events
3432 * polling again, they can rely on io_sq_thread to do polling
3433 * work, which can reduce cpu usage and uring_lock contention.
3434 */
3435 if (ctx->flags & IORING_SETUP_IOPOLL &&
3436 !(ctx->flags & IORING_SETUP_SQPOLL))
3437 ctx->syscall_iopoll = 1;
3438
3439 ctx->compat = in_compat_syscall();
3440 if (!capable(CAP_IPC_LOCK))
3441 ctx->user = get_uid(current_user());
3442
3443 /*
3444 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3445 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3446 */
3447 ret = -EINVAL;
3448 if (ctx->flags & IORING_SETUP_SQPOLL) {
3449 /* IPI related flags don't make sense with SQPOLL */
3450 if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3451 IORING_SETUP_TASKRUN_FLAG |
3452 IORING_SETUP_DEFER_TASKRUN))
3453 goto err;
3454 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3455 } else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3456 ctx->notify_method = TWA_SIGNAL_NO_IPI;
3457 } else {
3458 if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3459 !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3460 goto err;
3461 ctx->notify_method = TWA_SIGNAL;
3462 }
3463
3464 /*
3465 * For DEFER_TASKRUN we require the completion task to be the same as the
3466 * submission task. This implies that there is only one submitter, so enforce
3467 * that.
3468 */
3469 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3470 !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3471 goto err;
3472 }
3473
3474 /*
3475 * This is just grabbed for accounting purposes. When a process exits,
3476 * the mm is exited and dropped before the files, hence we need to hang
3477 * on to this mm purely for the purposes of being able to unaccount
3478 * memory (locked/pinned vm). It's not used for anything else.
3479 */
3480 mmgrab(current->mm);
3481 ctx->mm_account = current->mm;
3482
3483 ret = io_allocate_scq_urings(ctx, p);
3484 if (ret)
3485 goto err;
3486
3487 ret = io_sq_offload_create(ctx, p);
3488 if (ret)
3489 goto err;
3490 /* always set a rsrc node */
3491 ret = io_rsrc_node_switch_start(ctx);
3492 if (ret)
3493 goto err;
3494 io_rsrc_node_switch(ctx, NULL);
3495
3496 memset(&p->sq_off, 0, sizeof(p->sq_off));
3497 p->sq_off.head = offsetof(struct io_rings, sq.head);
3498 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3499 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3500 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3501 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3502 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3503 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3504
3505 memset(&p->cq_off, 0, sizeof(p->cq_off));
3506 p->cq_off.head = offsetof(struct io_rings, cq.head);
3507 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3508 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3509 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3510 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3511 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3512 p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3513
3514 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3515 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3516 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3517 IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3518 IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3519 IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3520 IORING_FEAT_LINKED_FILE;
3521
3522 if (copy_to_user(params, p, sizeof(*p))) {
3523 ret = -EFAULT;
3524 goto err;
3525 }
3526
3527 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3528 && !(ctx->flags & IORING_SETUP_R_DISABLED))
3529 ctx->submitter_task = get_task_struct(current);
3530
3531 file = io_uring_get_file(ctx);
3532 if (IS_ERR(file)) {
3533 ret = PTR_ERR(file);
3534 goto err;
3535 }
3536
3537 /*
3538 * Install ring fd as the very last thing, so we don't risk someone
3539 * having closed it before we finish setup
3540 */
3541 ret = io_uring_install_fd(ctx, file);
3542 if (ret < 0) {
3543 /* fput will clean it up */
3544 fput(file);
3545 return ret;
3546 }
3547
3548 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3549 return ret;
3550 err:
3551 io_ring_ctx_wait_and_kill(ctx);
3552 return ret;
3553 }
3554
3555 /*
3556 * Sets up an aio uring context, and returns the fd. Applications asks for a
3557 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3558 * params structure passed in.
3559 */
io_uring_setup(u32 entries,struct io_uring_params __user * params)3560 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3561 {
3562 struct io_uring_params p;
3563 int i;
3564
3565 if (copy_from_user(&p, params, sizeof(p)))
3566 return -EFAULT;
3567 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3568 if (p.resv[i])
3569 return -EINVAL;
3570 }
3571
3572 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3573 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3574 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3575 IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3576 IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3577 IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3578 IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3579 return -EINVAL;
3580
3581 return io_uring_create(entries, &p, params);
3582 }
3583
SYSCALL_DEFINE2(io_uring_setup,u32,entries,struct io_uring_params __user *,params)3584 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3585 struct io_uring_params __user *, params)
3586 {
3587 return io_uring_setup(entries, params);
3588 }
3589
io_probe(struct io_ring_ctx * ctx,void __user * arg,unsigned nr_args)3590 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3591 unsigned nr_args)
3592 {
3593 struct io_uring_probe *p;
3594 size_t size;
3595 int i, ret;
3596
3597 size = struct_size(p, ops, nr_args);
3598 if (size == SIZE_MAX)
3599 return -EOVERFLOW;
3600 p = kzalloc(size, GFP_KERNEL);
3601 if (!p)
3602 return -ENOMEM;
3603
3604 ret = -EFAULT;
3605 if (copy_from_user(p, arg, size))
3606 goto out;
3607 ret = -EINVAL;
3608 if (memchr_inv(p, 0, size))
3609 goto out;
3610
3611 p->last_op = IORING_OP_LAST - 1;
3612 if (nr_args > IORING_OP_LAST)
3613 nr_args = IORING_OP_LAST;
3614
3615 for (i = 0; i < nr_args; i++) {
3616 p->ops[i].op = i;
3617 if (!io_op_defs[i].not_supported)
3618 p->ops[i].flags = IO_URING_OP_SUPPORTED;
3619 }
3620 p->ops_len = i;
3621
3622 ret = 0;
3623 if (copy_to_user(arg, p, size))
3624 ret = -EFAULT;
3625 out:
3626 kfree(p);
3627 return ret;
3628 }
3629
io_register_personality(struct io_ring_ctx * ctx)3630 static int io_register_personality(struct io_ring_ctx *ctx)
3631 {
3632 const struct cred *creds;
3633 u32 id;
3634 int ret;
3635
3636 creds = get_current_cred();
3637
3638 ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3639 XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3640 if (ret < 0) {
3641 put_cred(creds);
3642 return ret;
3643 }
3644 return id;
3645 }
3646
io_register_restrictions(struct io_ring_ctx * ctx,void __user * arg,unsigned int nr_args)3647 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3648 void __user *arg, unsigned int nr_args)
3649 {
3650 struct io_uring_restriction *res;
3651 size_t size;
3652 int i, ret;
3653
3654 /* Restrictions allowed only if rings started disabled */
3655 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3656 return -EBADFD;
3657
3658 /* We allow only a single restrictions registration */
3659 if (ctx->restrictions.registered)
3660 return -EBUSY;
3661
3662 if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3663 return -EINVAL;
3664
3665 size = array_size(nr_args, sizeof(*res));
3666 if (size == SIZE_MAX)
3667 return -EOVERFLOW;
3668
3669 res = memdup_user(arg, size);
3670 if (IS_ERR(res))
3671 return PTR_ERR(res);
3672
3673 ret = 0;
3674
3675 for (i = 0; i < nr_args; i++) {
3676 switch (res[i].opcode) {
3677 case IORING_RESTRICTION_REGISTER_OP:
3678 if (res[i].register_op >= IORING_REGISTER_LAST) {
3679 ret = -EINVAL;
3680 goto out;
3681 }
3682
3683 __set_bit(res[i].register_op,
3684 ctx->restrictions.register_op);
3685 break;
3686 case IORING_RESTRICTION_SQE_OP:
3687 if (res[i].sqe_op >= IORING_OP_LAST) {
3688 ret = -EINVAL;
3689 goto out;
3690 }
3691
3692 __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3693 break;
3694 case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3695 ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3696 break;
3697 case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3698 ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3699 break;
3700 default:
3701 ret = -EINVAL;
3702 goto out;
3703 }
3704 }
3705
3706 out:
3707 /* Reset all restrictions if an error happened */
3708 if (ret != 0)
3709 memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3710 else
3711 ctx->restrictions.registered = true;
3712
3713 kfree(res);
3714 return ret;
3715 }
3716
io_register_enable_rings(struct io_ring_ctx * ctx)3717 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3718 {
3719 if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3720 return -EBADFD;
3721
3722 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task)
3723 ctx->submitter_task = get_task_struct(current);
3724
3725 if (ctx->restrictions.registered)
3726 ctx->restricted = 1;
3727
3728 ctx->flags &= ~IORING_SETUP_R_DISABLED;
3729 if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3730 wake_up(&ctx->sq_data->wait);
3731 return 0;
3732 }
3733
io_register_iowq_aff(struct io_ring_ctx * ctx,void __user * arg,unsigned len)3734 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3735 void __user *arg, unsigned len)
3736 {
3737 struct io_uring_task *tctx = current->io_uring;
3738 cpumask_var_t new_mask;
3739 int ret;
3740
3741 if (!tctx || !tctx->io_wq)
3742 return -EINVAL;
3743
3744 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3745 return -ENOMEM;
3746
3747 cpumask_clear(new_mask);
3748 if (len > cpumask_size())
3749 len = cpumask_size();
3750
3751 if (in_compat_syscall()) {
3752 ret = compat_get_bitmap(cpumask_bits(new_mask),
3753 (const compat_ulong_t __user *)arg,
3754 len * 8 /* CHAR_BIT */);
3755 } else {
3756 ret = copy_from_user(new_mask, arg, len);
3757 }
3758
3759 if (ret) {
3760 free_cpumask_var(new_mask);
3761 return -EFAULT;
3762 }
3763
3764 ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3765 free_cpumask_var(new_mask);
3766 return ret;
3767 }
3768
io_unregister_iowq_aff(struct io_ring_ctx * ctx)3769 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3770 {
3771 struct io_uring_task *tctx = current->io_uring;
3772
3773 if (!tctx || !tctx->io_wq)
3774 return -EINVAL;
3775
3776 return io_wq_cpu_affinity(tctx->io_wq, NULL);
3777 }
3778
io_register_iowq_max_workers(struct io_ring_ctx * ctx,void __user * arg)3779 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3780 void __user *arg)
3781 __must_hold(&ctx->uring_lock)
3782 {
3783 struct io_tctx_node *node;
3784 struct io_uring_task *tctx = NULL;
3785 struct io_sq_data *sqd = NULL;
3786 __u32 new_count[2];
3787 int i, ret;
3788
3789 if (copy_from_user(new_count, arg, sizeof(new_count)))
3790 return -EFAULT;
3791 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3792 if (new_count[i] > INT_MAX)
3793 return -EINVAL;
3794
3795 if (ctx->flags & IORING_SETUP_SQPOLL) {
3796 sqd = ctx->sq_data;
3797 if (sqd) {
3798 /*
3799 * Observe the correct sqd->lock -> ctx->uring_lock
3800 * ordering. Fine to drop uring_lock here, we hold
3801 * a ref to the ctx.
3802 */
3803 refcount_inc(&sqd->refs);
3804 mutex_unlock(&ctx->uring_lock);
3805 mutex_lock(&sqd->lock);
3806 mutex_lock(&ctx->uring_lock);
3807 if (sqd->thread)
3808 tctx = sqd->thread->io_uring;
3809 }
3810 } else {
3811 tctx = current->io_uring;
3812 }
3813
3814 BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3815
3816 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3817 if (new_count[i])
3818 ctx->iowq_limits[i] = new_count[i];
3819 ctx->iowq_limits_set = true;
3820
3821 if (tctx && tctx->io_wq) {
3822 ret = io_wq_max_workers(tctx->io_wq, new_count);
3823 if (ret)
3824 goto err;
3825 } else {
3826 memset(new_count, 0, sizeof(new_count));
3827 }
3828
3829 if (sqd) {
3830 mutex_unlock(&sqd->lock);
3831 io_put_sq_data(sqd);
3832 }
3833
3834 if (copy_to_user(arg, new_count, sizeof(new_count)))
3835 return -EFAULT;
3836
3837 /* that's it for SQPOLL, only the SQPOLL task creates requests */
3838 if (sqd)
3839 return 0;
3840
3841 /* now propagate the restriction to all registered users */
3842 list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3843 struct io_uring_task *tctx = node->task->io_uring;
3844
3845 if (WARN_ON_ONCE(!tctx->io_wq))
3846 continue;
3847
3848 for (i = 0; i < ARRAY_SIZE(new_count); i++)
3849 new_count[i] = ctx->iowq_limits[i];
3850 /* ignore errors, it always returns zero anyway */
3851 (void)io_wq_max_workers(tctx->io_wq, new_count);
3852 }
3853 return 0;
3854 err:
3855 if (sqd) {
3856 mutex_unlock(&sqd->lock);
3857 io_put_sq_data(sqd);
3858 }
3859 return ret;
3860 }
3861
__io_uring_register(struct io_ring_ctx * ctx,unsigned opcode,void __user * arg,unsigned nr_args)3862 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3863 void __user *arg, unsigned nr_args)
3864 __releases(ctx->uring_lock)
3865 __acquires(ctx->uring_lock)
3866 {
3867 int ret;
3868
3869 /*
3870 * We don't quiesce the refs for register anymore and so it can't be
3871 * dying as we're holding a file ref here.
3872 */
3873 if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
3874 return -ENXIO;
3875
3876 if (ctx->submitter_task && ctx->submitter_task != current)
3877 return -EEXIST;
3878
3879 if (ctx->restricted) {
3880 opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3881 if (!test_bit(opcode, ctx->restrictions.register_op))
3882 return -EACCES;
3883 }
3884
3885 switch (opcode) {
3886 case IORING_REGISTER_BUFFERS:
3887 ret = -EFAULT;
3888 if (!arg)
3889 break;
3890 ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3891 break;
3892 case IORING_UNREGISTER_BUFFERS:
3893 ret = -EINVAL;
3894 if (arg || nr_args)
3895 break;
3896 ret = io_sqe_buffers_unregister(ctx);
3897 break;
3898 case IORING_REGISTER_FILES:
3899 ret = -EFAULT;
3900 if (!arg)
3901 break;
3902 ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3903 break;
3904 case IORING_UNREGISTER_FILES:
3905 ret = -EINVAL;
3906 if (arg || nr_args)
3907 break;
3908 ret = io_sqe_files_unregister(ctx);
3909 break;
3910 case IORING_REGISTER_FILES_UPDATE:
3911 ret = io_register_files_update(ctx, arg, nr_args);
3912 break;
3913 case IORING_REGISTER_EVENTFD:
3914 ret = -EINVAL;
3915 if (nr_args != 1)
3916 break;
3917 ret = io_eventfd_register(ctx, arg, 0);
3918 break;
3919 case IORING_REGISTER_EVENTFD_ASYNC:
3920 ret = -EINVAL;
3921 if (nr_args != 1)
3922 break;
3923 ret = io_eventfd_register(ctx, arg, 1);
3924 break;
3925 case IORING_UNREGISTER_EVENTFD:
3926 ret = -EINVAL;
3927 if (arg || nr_args)
3928 break;
3929 ret = io_eventfd_unregister(ctx);
3930 break;
3931 case IORING_REGISTER_PROBE:
3932 ret = -EINVAL;
3933 if (!arg || nr_args > 256)
3934 break;
3935 ret = io_probe(ctx, arg, nr_args);
3936 break;
3937 case IORING_REGISTER_PERSONALITY:
3938 ret = -EINVAL;
3939 if (arg || nr_args)
3940 break;
3941 ret = io_register_personality(ctx);
3942 break;
3943 case IORING_UNREGISTER_PERSONALITY:
3944 ret = -EINVAL;
3945 if (arg)
3946 break;
3947 ret = io_unregister_personality(ctx, nr_args);
3948 break;
3949 case IORING_REGISTER_ENABLE_RINGS:
3950 ret = -EINVAL;
3951 if (arg || nr_args)
3952 break;
3953 ret = io_register_enable_rings(ctx);
3954 break;
3955 case IORING_REGISTER_RESTRICTIONS:
3956 ret = io_register_restrictions(ctx, arg, nr_args);
3957 break;
3958 case IORING_REGISTER_FILES2:
3959 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
3960 break;
3961 case IORING_REGISTER_FILES_UPDATE2:
3962 ret = io_register_rsrc_update(ctx, arg, nr_args,
3963 IORING_RSRC_FILE);
3964 break;
3965 case IORING_REGISTER_BUFFERS2:
3966 ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3967 break;
3968 case IORING_REGISTER_BUFFERS_UPDATE:
3969 ret = io_register_rsrc_update(ctx, arg, nr_args,
3970 IORING_RSRC_BUFFER);
3971 break;
3972 case IORING_REGISTER_IOWQ_AFF:
3973 ret = -EINVAL;
3974 if (!arg || !nr_args)
3975 break;
3976 ret = io_register_iowq_aff(ctx, arg, nr_args);
3977 break;
3978 case IORING_UNREGISTER_IOWQ_AFF:
3979 ret = -EINVAL;
3980 if (arg || nr_args)
3981 break;
3982 ret = io_unregister_iowq_aff(ctx);
3983 break;
3984 case IORING_REGISTER_IOWQ_MAX_WORKERS:
3985 ret = -EINVAL;
3986 if (!arg || nr_args != 2)
3987 break;
3988 ret = io_register_iowq_max_workers(ctx, arg);
3989 break;
3990 case IORING_REGISTER_RING_FDS:
3991 ret = io_ringfd_register(ctx, arg, nr_args);
3992 break;
3993 case IORING_UNREGISTER_RING_FDS:
3994 ret = io_ringfd_unregister(ctx, arg, nr_args);
3995 break;
3996 case IORING_REGISTER_PBUF_RING:
3997 ret = -EINVAL;
3998 if (!arg || nr_args != 1)
3999 break;
4000 ret = io_register_pbuf_ring(ctx, arg);
4001 break;
4002 case IORING_UNREGISTER_PBUF_RING:
4003 ret = -EINVAL;
4004 if (!arg || nr_args != 1)
4005 break;
4006 ret = io_unregister_pbuf_ring(ctx, arg);
4007 break;
4008 case IORING_REGISTER_SYNC_CANCEL:
4009 ret = -EINVAL;
4010 if (!arg || nr_args != 1)
4011 break;
4012 ret = io_sync_cancel(ctx, arg);
4013 break;
4014 case IORING_REGISTER_FILE_ALLOC_RANGE:
4015 ret = -EINVAL;
4016 if (!arg || nr_args)
4017 break;
4018 ret = io_register_file_alloc_range(ctx, arg);
4019 break;
4020 default:
4021 ret = -EINVAL;
4022 break;
4023 }
4024
4025 return ret;
4026 }
4027
SYSCALL_DEFINE4(io_uring_register,unsigned int,fd,unsigned int,opcode,void __user *,arg,unsigned int,nr_args)4028 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4029 void __user *, arg, unsigned int, nr_args)
4030 {
4031 struct io_ring_ctx *ctx;
4032 long ret = -EBADF;
4033 struct fd f;
4034
4035 if (opcode >= IORING_REGISTER_LAST)
4036 return -EINVAL;
4037
4038 f = fdget(fd);
4039 if (!f.file)
4040 return -EBADF;
4041
4042 ret = -EOPNOTSUPP;
4043 if (!io_is_uring_fops(f.file))
4044 goto out_fput;
4045
4046 ctx = f.file->private_data;
4047
4048 io_run_task_work_ctx(ctx);
4049
4050 mutex_lock(&ctx->uring_lock);
4051 ret = __io_uring_register(ctx, opcode, arg, nr_args);
4052 mutex_unlock(&ctx->uring_lock);
4053 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4054 out_fput:
4055 fdput(f);
4056 return ret;
4057 }
4058
io_uring_init(void)4059 static int __init io_uring_init(void)
4060 {
4061 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4062 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4063 BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4064 } while (0)
4065
4066 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4067 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4068 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4069 __BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4070 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4071 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
4072 BUILD_BUG_SQE_ELEM(1, __u8, flags);
4073 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
4074 BUILD_BUG_SQE_ELEM(4, __s32, fd);
4075 BUILD_BUG_SQE_ELEM(8, __u64, off);
4076 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
4077 BUILD_BUG_SQE_ELEM(8, __u32, cmd_op);
4078 BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4079 BUILD_BUG_SQE_ELEM(16, __u64, addr);
4080 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
4081 BUILD_BUG_SQE_ELEM(24, __u32, len);
4082 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
4083 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
4084 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4085 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
4086 BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events);
4087 BUILD_BUG_SQE_ELEM(28, __u32, poll32_events);
4088 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
4089 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
4090 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
4091 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
4092 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
4093 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
4094 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
4095 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
4096 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
4097 BUILD_BUG_SQE_ELEM(28, __u32, rename_flags);
4098 BUILD_BUG_SQE_ELEM(28, __u32, unlink_flags);
4099 BUILD_BUG_SQE_ELEM(28, __u32, hardlink_flags);
4100 BUILD_BUG_SQE_ELEM(28, __u32, xattr_flags);
4101 BUILD_BUG_SQE_ELEM(28, __u32, msg_ring_flags);
4102 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
4103 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
4104 BUILD_BUG_SQE_ELEM(40, __u16, buf_group);
4105 BUILD_BUG_SQE_ELEM(42, __u16, personality);
4106 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
4107 BUILD_BUG_SQE_ELEM(44, __u32, file_index);
4108 BUILD_BUG_SQE_ELEM(44, __u16, addr_len);
4109 BUILD_BUG_SQE_ELEM(46, __u16, __pad3[0]);
4110 BUILD_BUG_SQE_ELEM(48, __u64, addr3);
4111 BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4112 BUILD_BUG_SQE_ELEM(56, __u64, __pad2);
4113
4114 BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4115 sizeof(struct io_uring_rsrc_update));
4116 BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4117 sizeof(struct io_uring_rsrc_update2));
4118
4119 /* ->buf_index is u16 */
4120 BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4121 BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4122 offsetof(struct io_uring_buf_ring, tail));
4123
4124 /* should fit into one byte */
4125 BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4126 BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4127 BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4128
4129 BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4130
4131 BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4132
4133 io_uring_optable_init();
4134
4135 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4136 SLAB_ACCOUNT);
4137 return 0;
4138 };
4139 __initcall(io_uring_init);
4140