xref: /openbmc/linux/io_uring/io_uring.c (revision fadbafc1)
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 
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 
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 
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 
179 static bool io_match_linked(struct io_kiocb *head)
180 {
181 	struct io_kiocb *req;
182 
183 	io_for_each_link(req, head) {
184 		if (req->flags & REQ_F_INFLIGHT)
185 			return true;
186 	}
187 	return false;
188 }
189 
190 /*
191  * As io_match_task() but protected against racing with linked timeouts.
192  * User must not hold timeout_lock.
193  */
194 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
195 			bool cancel_all)
196 {
197 	bool matched;
198 
199 	if (task && head->task != task)
200 		return false;
201 	if (cancel_all)
202 		return true;
203 
204 	if (head->flags & REQ_F_LINK_TIMEOUT) {
205 		struct io_ring_ctx *ctx = head->ctx;
206 
207 		/* protect against races with linked timeouts */
208 		spin_lock_irq(&ctx->timeout_lock);
209 		matched = io_match_linked(head);
210 		spin_unlock_irq(&ctx->timeout_lock);
211 	} else {
212 		matched = io_match_linked(head);
213 	}
214 	return matched;
215 }
216 
217 static inline void req_fail_link_node(struct io_kiocb *req, int res)
218 {
219 	req_set_fail(req);
220 	io_req_set_res(req, res, 0);
221 }
222 
223 static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
224 {
225 	wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
226 }
227 
228 static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
229 {
230 	struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
231 
232 	complete(&ctx->ref_comp);
233 }
234 
235 static __cold void io_fallback_req_func(struct work_struct *work)
236 {
237 	struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
238 						fallback_work.work);
239 	struct llist_node *node = llist_del_all(&ctx->fallback_llist);
240 	struct io_kiocb *req, *tmp;
241 	bool locked = false;
242 
243 	percpu_ref_get(&ctx->refs);
244 	llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
245 		req->io_task_work.func(req, &locked);
246 
247 	if (locked) {
248 		io_submit_flush_completions(ctx);
249 		mutex_unlock(&ctx->uring_lock);
250 	}
251 	percpu_ref_put(&ctx->refs);
252 }
253 
254 static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
255 {
256 	unsigned hash_buckets = 1U << bits;
257 	size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
258 
259 	table->hbs = kmalloc(hash_size, GFP_KERNEL);
260 	if (!table->hbs)
261 		return -ENOMEM;
262 
263 	table->hash_bits = bits;
264 	init_hash_table(table, hash_buckets);
265 	return 0;
266 }
267 
268 static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
269 {
270 	struct io_ring_ctx *ctx;
271 	int hash_bits;
272 
273 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
274 	if (!ctx)
275 		return NULL;
276 
277 	xa_init(&ctx->io_bl_xa);
278 
279 	/*
280 	 * Use 5 bits less than the max cq entries, that should give us around
281 	 * 32 entries per hash list if totally full and uniformly spread, but
282 	 * don't keep too many buckets to not overconsume memory.
283 	 */
284 	hash_bits = ilog2(p->cq_entries) - 5;
285 	hash_bits = clamp(hash_bits, 1, 8);
286 	if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
287 		goto err;
288 	if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
289 		goto err;
290 
291 	ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
292 	if (!ctx->dummy_ubuf)
293 		goto err;
294 	/* set invalid range, so io_import_fixed() fails meeting it */
295 	ctx->dummy_ubuf->ubuf = -1UL;
296 
297 	if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
298 			    0, GFP_KERNEL))
299 		goto err;
300 
301 	ctx->flags = p->flags;
302 	init_waitqueue_head(&ctx->sqo_sq_wait);
303 	INIT_LIST_HEAD(&ctx->sqd_list);
304 	INIT_LIST_HEAD(&ctx->cq_overflow_list);
305 	INIT_LIST_HEAD(&ctx->io_buffers_cache);
306 	io_alloc_cache_init(&ctx->apoll_cache);
307 	io_alloc_cache_init(&ctx->netmsg_cache);
308 	init_completion(&ctx->ref_comp);
309 	xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
310 	mutex_init(&ctx->uring_lock);
311 	init_waitqueue_head(&ctx->cq_wait);
312 	spin_lock_init(&ctx->completion_lock);
313 	spin_lock_init(&ctx->timeout_lock);
314 	INIT_WQ_LIST(&ctx->iopoll_list);
315 	INIT_LIST_HEAD(&ctx->io_buffers_pages);
316 	INIT_LIST_HEAD(&ctx->io_buffers_comp);
317 	INIT_LIST_HEAD(&ctx->defer_list);
318 	INIT_LIST_HEAD(&ctx->timeout_list);
319 	INIT_LIST_HEAD(&ctx->ltimeout_list);
320 	spin_lock_init(&ctx->rsrc_ref_lock);
321 	INIT_LIST_HEAD(&ctx->rsrc_ref_list);
322 	INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
323 	init_llist_head(&ctx->rsrc_put_llist);
324 	init_llist_head(&ctx->work_llist);
325 	INIT_LIST_HEAD(&ctx->tctx_list);
326 	ctx->submit_state.free_list.next = NULL;
327 	INIT_WQ_LIST(&ctx->locked_free_list);
328 	INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
329 	INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
330 	return ctx;
331 err:
332 	kfree(ctx->dummy_ubuf);
333 	kfree(ctx->cancel_table.hbs);
334 	kfree(ctx->cancel_table_locked.hbs);
335 	kfree(ctx->io_bl);
336 	xa_destroy(&ctx->io_bl_xa);
337 	kfree(ctx);
338 	return NULL;
339 }
340 
341 static void io_account_cq_overflow(struct io_ring_ctx *ctx)
342 {
343 	struct io_rings *r = ctx->rings;
344 
345 	WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
346 	ctx->cq_extra--;
347 }
348 
349 static bool req_need_defer(struct io_kiocb *req, u32 seq)
350 {
351 	if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
352 		struct io_ring_ctx *ctx = req->ctx;
353 
354 		return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
355 	}
356 
357 	return false;
358 }
359 
360 static inline void io_req_track_inflight(struct io_kiocb *req)
361 {
362 	if (!(req->flags & REQ_F_INFLIGHT)) {
363 		req->flags |= REQ_F_INFLIGHT;
364 		atomic_inc(&req->task->io_uring->inflight_tracked);
365 	}
366 }
367 
368 static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
369 {
370 	if (WARN_ON_ONCE(!req->link))
371 		return NULL;
372 
373 	req->flags &= ~REQ_F_ARM_LTIMEOUT;
374 	req->flags |= REQ_F_LINK_TIMEOUT;
375 
376 	/* linked timeouts should have two refs once prep'ed */
377 	io_req_set_refcount(req);
378 	__io_req_set_refcount(req->link, 2);
379 	return req->link;
380 }
381 
382 static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
383 {
384 	if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
385 		return NULL;
386 	return __io_prep_linked_timeout(req);
387 }
388 
389 static noinline void __io_arm_ltimeout(struct io_kiocb *req)
390 {
391 	io_queue_linked_timeout(__io_prep_linked_timeout(req));
392 }
393 
394 static inline void io_arm_ltimeout(struct io_kiocb *req)
395 {
396 	if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
397 		__io_arm_ltimeout(req);
398 }
399 
400 static void io_prep_async_work(struct io_kiocb *req)
401 {
402 	const struct io_op_def *def = &io_op_defs[req->opcode];
403 	struct io_ring_ctx *ctx = req->ctx;
404 
405 	if (!(req->flags & REQ_F_CREDS)) {
406 		req->flags |= REQ_F_CREDS;
407 		req->creds = get_current_cred();
408 	}
409 
410 	req->work.list.next = NULL;
411 	req->work.flags = 0;
412 	req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
413 	if (req->flags & REQ_F_FORCE_ASYNC)
414 		req->work.flags |= IO_WQ_WORK_CONCURRENT;
415 
416 	if (req->file && !io_req_ffs_set(req))
417 		req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
418 
419 	if (req->flags & REQ_F_ISREG) {
420 		if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
421 			io_wq_hash_work(&req->work, file_inode(req->file));
422 	} else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
423 		if (def->unbound_nonreg_file)
424 			req->work.flags |= IO_WQ_WORK_UNBOUND;
425 	}
426 }
427 
428 static void io_prep_async_link(struct io_kiocb *req)
429 {
430 	struct io_kiocb *cur;
431 
432 	if (req->flags & REQ_F_LINK_TIMEOUT) {
433 		struct io_ring_ctx *ctx = req->ctx;
434 
435 		spin_lock_irq(&ctx->timeout_lock);
436 		io_for_each_link(cur, req)
437 			io_prep_async_work(cur);
438 		spin_unlock_irq(&ctx->timeout_lock);
439 	} else {
440 		io_for_each_link(cur, req)
441 			io_prep_async_work(cur);
442 	}
443 }
444 
445 void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
446 {
447 	struct io_kiocb *link = io_prep_linked_timeout(req);
448 	struct io_uring_task *tctx = req->task->io_uring;
449 
450 	BUG_ON(!tctx);
451 	BUG_ON(!tctx->io_wq);
452 
453 	/* init ->work of the whole link before punting */
454 	io_prep_async_link(req);
455 
456 	/*
457 	 * Not expected to happen, but if we do have a bug where this _can_
458 	 * happen, catch it here and ensure the request is marked as
459 	 * canceled. That will make io-wq go through the usual work cancel
460 	 * procedure rather than attempt to run this request (or create a new
461 	 * worker for it).
462 	 */
463 	if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
464 		req->work.flags |= IO_WQ_WORK_CANCEL;
465 
466 	trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
467 	io_wq_enqueue(tctx->io_wq, &req->work);
468 	if (link)
469 		io_queue_linked_timeout(link);
470 }
471 
472 static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
473 {
474 	while (!list_empty(&ctx->defer_list)) {
475 		struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
476 						struct io_defer_entry, list);
477 
478 		if (req_need_defer(de->req, de->seq))
479 			break;
480 		list_del_init(&de->list);
481 		io_req_task_queue(de->req);
482 		kfree(de);
483 	}
484 }
485 
486 
487 static void io_eventfd_ops(struct rcu_head *rcu)
488 {
489 	struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
490 	int ops = atomic_xchg(&ev_fd->ops, 0);
491 
492 	if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
493 		eventfd_signal(ev_fd->cq_ev_fd, 1);
494 
495 	/* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
496 	 * ordering in a race but if references are 0 we know we have to free
497 	 * it regardless.
498 	 */
499 	if (atomic_dec_and_test(&ev_fd->refs)) {
500 		eventfd_ctx_put(ev_fd->cq_ev_fd);
501 		kfree(ev_fd);
502 	}
503 }
504 
505 static void io_eventfd_signal(struct io_ring_ctx *ctx)
506 {
507 	struct io_ev_fd *ev_fd = NULL;
508 
509 	rcu_read_lock();
510 	/*
511 	 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
512 	 * and eventfd_signal
513 	 */
514 	ev_fd = rcu_dereference(ctx->io_ev_fd);
515 
516 	/*
517 	 * Check again if ev_fd exists incase an io_eventfd_unregister call
518 	 * completed between the NULL check of ctx->io_ev_fd at the start of
519 	 * the function and rcu_read_lock.
520 	 */
521 	if (unlikely(!ev_fd))
522 		goto out;
523 	if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
524 		goto out;
525 	if (ev_fd->eventfd_async && !io_wq_current_is_worker())
526 		goto out;
527 
528 	if (likely(eventfd_signal_allowed())) {
529 		eventfd_signal(ev_fd->cq_ev_fd, 1);
530 	} else {
531 		atomic_inc(&ev_fd->refs);
532 		if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
533 			call_rcu(&ev_fd->rcu, io_eventfd_ops);
534 		else
535 			atomic_dec(&ev_fd->refs);
536 	}
537 
538 out:
539 	rcu_read_unlock();
540 }
541 
542 static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
543 {
544 	bool skip;
545 
546 	spin_lock(&ctx->completion_lock);
547 
548 	/*
549 	 * Eventfd should only get triggered when at least one event has been
550 	 * posted. Some applications rely on the eventfd notification count
551 	 * only changing IFF a new CQE has been added to the CQ ring. There's
552 	 * no depedency on 1:1 relationship between how many times this
553 	 * function is called (and hence the eventfd count) and number of CQEs
554 	 * posted to the CQ ring.
555 	 */
556 	skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
557 	ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
558 	spin_unlock(&ctx->completion_lock);
559 	if (skip)
560 		return;
561 
562 	io_eventfd_signal(ctx);
563 }
564 
565 void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
566 {
567 	if (ctx->off_timeout_used || ctx->drain_active) {
568 		spin_lock(&ctx->completion_lock);
569 		if (ctx->off_timeout_used)
570 			io_flush_timeouts(ctx);
571 		if (ctx->drain_active)
572 			io_queue_deferred(ctx);
573 		spin_unlock(&ctx->completion_lock);
574 	}
575 	if (ctx->has_evfd)
576 		io_eventfd_flush_signal(ctx);
577 }
578 
579 static inline void io_cqring_ev_posted(struct io_ring_ctx *ctx)
580 {
581 	io_commit_cqring_flush(ctx);
582 	io_cqring_wake(ctx);
583 }
584 
585 static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
586 	__releases(ctx->completion_lock)
587 {
588 	io_commit_cqring(ctx);
589 	spin_unlock(&ctx->completion_lock);
590 	io_cqring_ev_posted(ctx);
591 }
592 
593 void io_cq_unlock_post(struct io_ring_ctx *ctx)
594 {
595 	__io_cq_unlock_post(ctx);
596 }
597 
598 /* Returns true if there are no backlogged entries after the flush */
599 static bool __io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
600 {
601 	bool all_flushed;
602 	size_t cqe_size = sizeof(struct io_uring_cqe);
603 
604 	if (!force && __io_cqring_events(ctx) == ctx->cq_entries)
605 		return false;
606 
607 	if (ctx->flags & IORING_SETUP_CQE32)
608 		cqe_size <<= 1;
609 
610 	io_cq_lock(ctx);
611 	while (!list_empty(&ctx->cq_overflow_list)) {
612 		struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
613 		struct io_overflow_cqe *ocqe;
614 
615 		if (!cqe && !force)
616 			break;
617 		ocqe = list_first_entry(&ctx->cq_overflow_list,
618 					struct io_overflow_cqe, list);
619 		if (cqe)
620 			memcpy(cqe, &ocqe->cqe, cqe_size);
621 		else
622 			io_account_cq_overflow(ctx);
623 
624 		list_del(&ocqe->list);
625 		kfree(ocqe);
626 	}
627 
628 	all_flushed = list_empty(&ctx->cq_overflow_list);
629 	if (all_flushed) {
630 		clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
631 		atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
632 	}
633 
634 	io_cq_unlock_post(ctx);
635 	return all_flushed;
636 }
637 
638 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx)
639 {
640 	bool ret = true;
641 
642 	if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
643 		/* iopoll syncs against uring_lock, not completion_lock */
644 		if (ctx->flags & IORING_SETUP_IOPOLL)
645 			mutex_lock(&ctx->uring_lock);
646 		ret = __io_cqring_overflow_flush(ctx, false);
647 		if (ctx->flags & IORING_SETUP_IOPOLL)
648 			mutex_unlock(&ctx->uring_lock);
649 	}
650 
651 	return ret;
652 }
653 
654 void __io_put_task(struct task_struct *task, int nr)
655 {
656 	struct io_uring_task *tctx = task->io_uring;
657 
658 	percpu_counter_sub(&tctx->inflight, nr);
659 	if (unlikely(atomic_read(&tctx->in_idle)))
660 		wake_up(&tctx->wait);
661 	put_task_struct_many(task, nr);
662 }
663 
664 void io_task_refs_refill(struct io_uring_task *tctx)
665 {
666 	unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
667 
668 	percpu_counter_add(&tctx->inflight, refill);
669 	refcount_add(refill, &current->usage);
670 	tctx->cached_refs += refill;
671 }
672 
673 static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
674 {
675 	struct io_uring_task *tctx = task->io_uring;
676 	unsigned int refs = tctx->cached_refs;
677 
678 	if (refs) {
679 		tctx->cached_refs = 0;
680 		percpu_counter_sub(&tctx->inflight, refs);
681 		put_task_struct_many(task, refs);
682 	}
683 }
684 
685 static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
686 				     s32 res, u32 cflags, u64 extra1, u64 extra2)
687 {
688 	struct io_overflow_cqe *ocqe;
689 	size_t ocq_size = sizeof(struct io_overflow_cqe);
690 	bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
691 
692 	if (is_cqe32)
693 		ocq_size += sizeof(struct io_uring_cqe);
694 
695 	ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
696 	trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
697 	if (!ocqe) {
698 		/*
699 		 * If we're in ring overflow flush mode, or in task cancel mode,
700 		 * or cannot allocate an overflow entry, then we need to drop it
701 		 * on the floor.
702 		 */
703 		io_account_cq_overflow(ctx);
704 		set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
705 		return false;
706 	}
707 	if (list_empty(&ctx->cq_overflow_list)) {
708 		set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
709 		atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
710 
711 	}
712 	ocqe->cqe.user_data = user_data;
713 	ocqe->cqe.res = res;
714 	ocqe->cqe.flags = cflags;
715 	if (is_cqe32) {
716 		ocqe->cqe.big_cqe[0] = extra1;
717 		ocqe->cqe.big_cqe[1] = extra2;
718 	}
719 	list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
720 	return true;
721 }
722 
723 bool io_req_cqe_overflow(struct io_kiocb *req)
724 {
725 	if (!(req->flags & REQ_F_CQE32_INIT)) {
726 		req->extra1 = 0;
727 		req->extra2 = 0;
728 	}
729 	return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
730 					req->cqe.res, req->cqe.flags,
731 					req->extra1, req->extra2);
732 }
733 
734 /*
735  * writes to the cq entry need to come after reading head; the
736  * control dependency is enough as we're using WRITE_ONCE to
737  * fill the cq entry
738  */
739 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
740 {
741 	struct io_rings *rings = ctx->rings;
742 	unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
743 	unsigned int free, queued, len;
744 
745 	/*
746 	 * Posting into the CQ when there are pending overflowed CQEs may break
747 	 * ordering guarantees, which will affect links, F_MORE users and more.
748 	 * Force overflow the completion.
749 	 */
750 	if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
751 		return NULL;
752 
753 	/* userspace may cheat modifying the tail, be safe and do min */
754 	queued = min(__io_cqring_events(ctx), ctx->cq_entries);
755 	free = ctx->cq_entries - queued;
756 	/* we need a contiguous range, limit based on the current array offset */
757 	len = min(free, ctx->cq_entries - off);
758 	if (!len)
759 		return NULL;
760 
761 	if (ctx->flags & IORING_SETUP_CQE32) {
762 		off <<= 1;
763 		len <<= 1;
764 	}
765 
766 	ctx->cqe_cached = &rings->cqes[off];
767 	ctx->cqe_sentinel = ctx->cqe_cached + len;
768 
769 	ctx->cached_cq_tail++;
770 	ctx->cqe_cached++;
771 	if (ctx->flags & IORING_SETUP_CQE32)
772 		ctx->cqe_cached++;
773 	return &rings->cqes[off];
774 }
775 
776 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
777 		     bool allow_overflow)
778 {
779 	struct io_uring_cqe *cqe;
780 
781 	ctx->cq_extra++;
782 
783 	/*
784 	 * If we can't get a cq entry, userspace overflowed the
785 	 * submission (by quite a lot). Increment the overflow count in
786 	 * the ring.
787 	 */
788 	cqe = io_get_cqe(ctx);
789 	if (likely(cqe)) {
790 		trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
791 
792 		WRITE_ONCE(cqe->user_data, user_data);
793 		WRITE_ONCE(cqe->res, res);
794 		WRITE_ONCE(cqe->flags, cflags);
795 
796 		if (ctx->flags & IORING_SETUP_CQE32) {
797 			WRITE_ONCE(cqe->big_cqe[0], 0);
798 			WRITE_ONCE(cqe->big_cqe[1], 0);
799 		}
800 		return true;
801 	}
802 
803 	if (allow_overflow)
804 		return io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
805 
806 	return false;
807 }
808 
809 bool io_post_aux_cqe(struct io_ring_ctx *ctx,
810 		     u64 user_data, s32 res, u32 cflags,
811 		     bool allow_overflow)
812 {
813 	bool filled;
814 
815 	io_cq_lock(ctx);
816 	filled = io_fill_cqe_aux(ctx, user_data, res, cflags, allow_overflow);
817 	io_cq_unlock_post(ctx);
818 	return filled;
819 }
820 
821 static void __io_req_complete_put(struct io_kiocb *req)
822 {
823 	/*
824 	 * If we're the last reference to this request, add to our locked
825 	 * free_list cache.
826 	 */
827 	if (req_ref_put_and_test(req)) {
828 		struct io_ring_ctx *ctx = req->ctx;
829 
830 		if (req->flags & IO_REQ_LINK_FLAGS) {
831 			if (req->flags & IO_DISARM_MASK)
832 				io_disarm_next(req);
833 			if (req->link) {
834 				io_req_task_queue(req->link);
835 				req->link = NULL;
836 			}
837 		}
838 		io_req_put_rsrc(req);
839 		/*
840 		 * Selected buffer deallocation in io_clean_op() assumes that
841 		 * we don't hold ->completion_lock. Clean them here to avoid
842 		 * deadlocks.
843 		 */
844 		io_put_kbuf_comp(req);
845 		io_dismantle_req(req);
846 		io_put_task(req->task, 1);
847 		wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
848 		ctx->locked_free_nr++;
849 	}
850 }
851 
852 void __io_req_complete_post(struct io_kiocb *req)
853 {
854 	if (!(req->flags & REQ_F_CQE_SKIP))
855 		__io_fill_cqe_req(req->ctx, req);
856 	__io_req_complete_put(req);
857 }
858 
859 void io_req_complete_post(struct io_kiocb *req)
860 {
861 	struct io_ring_ctx *ctx = req->ctx;
862 
863 	io_cq_lock(ctx);
864 	__io_req_complete_post(req);
865 	io_cq_unlock_post(ctx);
866 }
867 
868 inline void __io_req_complete(struct io_kiocb *req, unsigned issue_flags)
869 {
870 	io_req_complete_post(req);
871 }
872 
873 void io_req_complete_failed(struct io_kiocb *req, s32 res)
874 {
875 	const struct io_op_def *def = &io_op_defs[req->opcode];
876 
877 	req_set_fail(req);
878 	io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
879 	if (def->fail)
880 		def->fail(req);
881 	io_req_complete_post(req);
882 }
883 
884 /*
885  * Don't initialise the fields below on every allocation, but do that in
886  * advance and keep them valid across allocations.
887  */
888 static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
889 {
890 	req->ctx = ctx;
891 	req->link = NULL;
892 	req->async_data = NULL;
893 	/* not necessary, but safer to zero */
894 	req->cqe.res = 0;
895 }
896 
897 static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
898 					struct io_submit_state *state)
899 {
900 	spin_lock(&ctx->completion_lock);
901 	wq_list_splice(&ctx->locked_free_list, &state->free_list);
902 	ctx->locked_free_nr = 0;
903 	spin_unlock(&ctx->completion_lock);
904 }
905 
906 /*
907  * A request might get retired back into the request caches even before opcode
908  * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
909  * Because of that, io_alloc_req() should be called only under ->uring_lock
910  * and with extra caution to not get a request that is still worked on.
911  */
912 __cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
913 	__must_hold(&ctx->uring_lock)
914 {
915 	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
916 	void *reqs[IO_REQ_ALLOC_BATCH];
917 	int ret, i;
918 
919 	/*
920 	 * If we have more than a batch's worth of requests in our IRQ side
921 	 * locked cache, grab the lock and move them over to our submission
922 	 * side cache.
923 	 */
924 	if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
925 		io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
926 		if (!io_req_cache_empty(ctx))
927 			return true;
928 	}
929 
930 	ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
931 
932 	/*
933 	 * Bulk alloc is all-or-nothing. If we fail to get a batch,
934 	 * retry single alloc to be on the safe side.
935 	 */
936 	if (unlikely(ret <= 0)) {
937 		reqs[0] = kmem_cache_alloc(req_cachep, gfp);
938 		if (!reqs[0])
939 			return false;
940 		ret = 1;
941 	}
942 
943 	percpu_ref_get_many(&ctx->refs, ret);
944 	for (i = 0; i < ret; i++) {
945 		struct io_kiocb *req = reqs[i];
946 
947 		io_preinit_req(req, ctx);
948 		io_req_add_to_cache(req, ctx);
949 	}
950 	return true;
951 }
952 
953 static inline void io_dismantle_req(struct io_kiocb *req)
954 {
955 	unsigned int flags = req->flags;
956 
957 	if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
958 		io_clean_op(req);
959 	if (!(flags & REQ_F_FIXED_FILE))
960 		io_put_file(req->file);
961 }
962 
963 __cold void io_free_req(struct io_kiocb *req)
964 {
965 	struct io_ring_ctx *ctx = req->ctx;
966 
967 	io_req_put_rsrc(req);
968 	io_dismantle_req(req);
969 	io_put_task(req->task, 1);
970 
971 	spin_lock(&ctx->completion_lock);
972 	wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
973 	ctx->locked_free_nr++;
974 	spin_unlock(&ctx->completion_lock);
975 }
976 
977 static void __io_req_find_next_prep(struct io_kiocb *req)
978 {
979 	struct io_ring_ctx *ctx = req->ctx;
980 
981 	io_cq_lock(ctx);
982 	io_disarm_next(req);
983 	io_cq_unlock_post(ctx);
984 }
985 
986 static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
987 {
988 	struct io_kiocb *nxt;
989 
990 	/*
991 	 * If LINK is set, we have dependent requests in this chain. If we
992 	 * didn't fail this request, queue the first one up, moving any other
993 	 * dependencies to the next request. In case of failure, fail the rest
994 	 * of the chain.
995 	 */
996 	if (unlikely(req->flags & IO_DISARM_MASK))
997 		__io_req_find_next_prep(req);
998 	nxt = req->link;
999 	req->link = NULL;
1000 	return nxt;
1001 }
1002 
1003 static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
1004 {
1005 	if (!ctx)
1006 		return;
1007 	if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1008 		atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1009 	if (*locked) {
1010 		io_submit_flush_completions(ctx);
1011 		mutex_unlock(&ctx->uring_lock);
1012 		*locked = false;
1013 	}
1014 	percpu_ref_put(&ctx->refs);
1015 }
1016 
1017 static unsigned int handle_tw_list(struct llist_node *node,
1018 				   struct io_ring_ctx **ctx, bool *locked,
1019 				   struct llist_node *last)
1020 {
1021 	unsigned int count = 0;
1022 
1023 	while (node != last) {
1024 		struct llist_node *next = node->next;
1025 		struct io_kiocb *req = container_of(node, struct io_kiocb,
1026 						    io_task_work.node);
1027 
1028 		prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1029 
1030 		if (req->ctx != *ctx) {
1031 			ctx_flush_and_put(*ctx, locked);
1032 			*ctx = req->ctx;
1033 			/* if not contended, grab and improve batching */
1034 			*locked = mutex_trylock(&(*ctx)->uring_lock);
1035 			percpu_ref_get(&(*ctx)->refs);
1036 		}
1037 		req->io_task_work.func(req, locked);
1038 		node = next;
1039 		count++;
1040 	}
1041 
1042 	return count;
1043 }
1044 
1045 /**
1046  * io_llist_xchg - swap all entries in a lock-less list
1047  * @head:	the head of lock-less list to delete all entries
1048  * @new:	new entry as the head of the list
1049  *
1050  * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1051  * The order of entries returned is from the newest to the oldest added one.
1052  */
1053 static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1054 					       struct llist_node *new)
1055 {
1056 	return xchg(&head->first, new);
1057 }
1058 
1059 /**
1060  * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1061  * @head:	the head of lock-less list to delete all entries
1062  * @old:	expected old value of the first entry of the list
1063  * @new:	new entry as the head of the list
1064  *
1065  * perform a cmpxchg on the first entry of the list.
1066  */
1067 
1068 static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1069 						  struct llist_node *old,
1070 						  struct llist_node *new)
1071 {
1072 	return cmpxchg(&head->first, old, new);
1073 }
1074 
1075 void tctx_task_work(struct callback_head *cb)
1076 {
1077 	bool uring_locked = false;
1078 	struct io_ring_ctx *ctx = NULL;
1079 	struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1080 						  task_work);
1081 	struct llist_node fake = {};
1082 	struct llist_node *node = io_llist_xchg(&tctx->task_list, &fake);
1083 	unsigned int loops = 1;
1084 	unsigned int count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1085 
1086 	node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1087 	while (node != &fake) {
1088 		loops++;
1089 		node = io_llist_xchg(&tctx->task_list, &fake);
1090 		count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1091 		node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1092 	}
1093 
1094 	ctx_flush_and_put(ctx, &uring_locked);
1095 
1096 	/* relaxed read is enough as only the task itself sets ->in_idle */
1097 	if (unlikely(atomic_read(&tctx->in_idle)))
1098 		io_uring_drop_tctx_refs(current);
1099 
1100 	trace_io_uring_task_work_run(tctx, count, loops);
1101 }
1102 
1103 static void io_req_local_work_add(struct io_kiocb *req)
1104 {
1105 	struct io_ring_ctx *ctx = req->ctx;
1106 
1107 	if (!llist_add(&req->io_task_work.node, &ctx->work_llist))
1108 		return;
1109 	/* need it for the following io_cqring_wake() */
1110 	smp_mb__after_atomic();
1111 
1112 	if (unlikely(atomic_read(&req->task->io_uring->in_idle))) {
1113 		io_move_task_work_from_local(ctx);
1114 		return;
1115 	}
1116 
1117 	if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1118 		atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1119 
1120 	if (ctx->has_evfd)
1121 		io_eventfd_signal(ctx);
1122 	__io_cqring_wake(ctx);
1123 }
1124 
1125 static inline void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1126 {
1127 	struct io_uring_task *tctx = req->task->io_uring;
1128 	struct io_ring_ctx *ctx = req->ctx;
1129 	struct llist_node *node;
1130 
1131 	if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1132 		io_req_local_work_add(req);
1133 		return;
1134 	}
1135 
1136 	/* task_work already pending, we're done */
1137 	if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1138 		return;
1139 
1140 	if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1141 		atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1142 
1143 	if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1144 		return;
1145 
1146 	node = llist_del_all(&tctx->task_list);
1147 
1148 	while (node) {
1149 		req = container_of(node, struct io_kiocb, io_task_work.node);
1150 		node = node->next;
1151 		if (llist_add(&req->io_task_work.node,
1152 			      &req->ctx->fallback_llist))
1153 			schedule_delayed_work(&req->ctx->fallback_work, 1);
1154 	}
1155 }
1156 
1157 void io_req_task_work_add(struct io_kiocb *req)
1158 {
1159 	__io_req_task_work_add(req, true);
1160 }
1161 
1162 static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1163 {
1164 	struct llist_node *node;
1165 
1166 	node = llist_del_all(&ctx->work_llist);
1167 	while (node) {
1168 		struct io_kiocb *req = container_of(node, struct io_kiocb,
1169 						    io_task_work.node);
1170 
1171 		node = node->next;
1172 		__io_req_task_work_add(req, false);
1173 	}
1174 }
1175 
1176 int __io_run_local_work(struct io_ring_ctx *ctx, bool locked)
1177 {
1178 	struct llist_node *node;
1179 	struct llist_node fake;
1180 	struct llist_node *current_final = NULL;
1181 	int ret;
1182 	unsigned int loops = 1;
1183 
1184 	if (unlikely(ctx->submitter_task != current))
1185 		return -EEXIST;
1186 
1187 	node = io_llist_xchg(&ctx->work_llist, &fake);
1188 	ret = 0;
1189 again:
1190 	while (node != current_final) {
1191 		struct llist_node *next = node->next;
1192 		struct io_kiocb *req = container_of(node, struct io_kiocb,
1193 						    io_task_work.node);
1194 		prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1195 		req->io_task_work.func(req, &locked);
1196 		ret++;
1197 		node = next;
1198 	}
1199 
1200 	if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1201 		atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1202 
1203 	node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL);
1204 	if (node != &fake) {
1205 		loops++;
1206 		current_final = &fake;
1207 		node = io_llist_xchg(&ctx->work_llist, &fake);
1208 		goto again;
1209 	}
1210 
1211 	if (locked)
1212 		io_submit_flush_completions(ctx);
1213 	trace_io_uring_local_work_run(ctx, ret, loops);
1214 	return ret;
1215 
1216 }
1217 
1218 int io_run_local_work(struct io_ring_ctx *ctx)
1219 {
1220 	bool locked;
1221 	int ret;
1222 
1223 	if (llist_empty(&ctx->work_llist))
1224 		return 0;
1225 
1226 	__set_current_state(TASK_RUNNING);
1227 	locked = mutex_trylock(&ctx->uring_lock);
1228 	ret = __io_run_local_work(ctx, locked);
1229 	if (locked)
1230 		mutex_unlock(&ctx->uring_lock);
1231 
1232 	return ret;
1233 }
1234 
1235 static void io_req_tw_post(struct io_kiocb *req, bool *locked)
1236 {
1237 	io_req_complete_post(req);
1238 }
1239 
1240 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags)
1241 {
1242 	io_req_set_res(req, res, cflags);
1243 	req->io_task_work.func = io_req_tw_post;
1244 	io_req_task_work_add(req);
1245 }
1246 
1247 static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1248 {
1249 	/* not needed for normal modes, but SQPOLL depends on it */
1250 	io_tw_lock(req->ctx, locked);
1251 	io_req_complete_failed(req, req->cqe.res);
1252 }
1253 
1254 void io_req_task_submit(struct io_kiocb *req, bool *locked)
1255 {
1256 	io_tw_lock(req->ctx, locked);
1257 	/* req->task == current here, checking PF_EXITING is safe */
1258 	if (likely(!(req->task->flags & PF_EXITING)))
1259 		io_queue_sqe(req);
1260 	else
1261 		io_req_complete_failed(req, -EFAULT);
1262 }
1263 
1264 void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1265 {
1266 	io_req_set_res(req, ret, 0);
1267 	req->io_task_work.func = io_req_task_cancel;
1268 	io_req_task_work_add(req);
1269 }
1270 
1271 void io_req_task_queue(struct io_kiocb *req)
1272 {
1273 	req->io_task_work.func = io_req_task_submit;
1274 	io_req_task_work_add(req);
1275 }
1276 
1277 void io_queue_next(struct io_kiocb *req)
1278 {
1279 	struct io_kiocb *nxt = io_req_find_next(req);
1280 
1281 	if (nxt)
1282 		io_req_task_queue(nxt);
1283 }
1284 
1285 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1286 	__must_hold(&ctx->uring_lock)
1287 {
1288 	struct task_struct *task = NULL;
1289 	int task_refs = 0;
1290 
1291 	do {
1292 		struct io_kiocb *req = container_of(node, struct io_kiocb,
1293 						    comp_list);
1294 
1295 		if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1296 			if (req->flags & REQ_F_REFCOUNT) {
1297 				node = req->comp_list.next;
1298 				if (!req_ref_put_and_test(req))
1299 					continue;
1300 			}
1301 			if ((req->flags & REQ_F_POLLED) && req->apoll) {
1302 				struct async_poll *apoll = req->apoll;
1303 
1304 				if (apoll->double_poll)
1305 					kfree(apoll->double_poll);
1306 				if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1307 					kfree(apoll);
1308 				req->flags &= ~REQ_F_POLLED;
1309 			}
1310 			if (req->flags & IO_REQ_LINK_FLAGS)
1311 				io_queue_next(req);
1312 			if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1313 				io_clean_op(req);
1314 		}
1315 		if (!(req->flags & REQ_F_FIXED_FILE))
1316 			io_put_file(req->file);
1317 
1318 		io_req_put_rsrc_locked(req, ctx);
1319 
1320 		if (req->task != task) {
1321 			if (task)
1322 				io_put_task(task, task_refs);
1323 			task = req->task;
1324 			task_refs = 0;
1325 		}
1326 		task_refs++;
1327 		node = req->comp_list.next;
1328 		io_req_add_to_cache(req, ctx);
1329 	} while (node);
1330 
1331 	if (task)
1332 		io_put_task(task, task_refs);
1333 }
1334 
1335 static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1336 	__must_hold(&ctx->uring_lock)
1337 {
1338 	struct io_wq_work_node *node, *prev;
1339 	struct io_submit_state *state = &ctx->submit_state;
1340 
1341 	io_cq_lock(ctx);
1342 	wq_list_for_each(node, prev, &state->compl_reqs) {
1343 		struct io_kiocb *req = container_of(node, struct io_kiocb,
1344 					    comp_list);
1345 
1346 		if (!(req->flags & REQ_F_CQE_SKIP))
1347 			__io_fill_cqe_req(ctx, req);
1348 	}
1349 	__io_cq_unlock_post(ctx);
1350 
1351 	io_free_batch_list(ctx, state->compl_reqs.first);
1352 	INIT_WQ_LIST(&state->compl_reqs);
1353 }
1354 
1355 /*
1356  * Drop reference to request, return next in chain (if there is one) if this
1357  * was the last reference to this request.
1358  */
1359 static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1360 {
1361 	struct io_kiocb *nxt = NULL;
1362 
1363 	if (req_ref_put_and_test(req)) {
1364 		if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1365 			nxt = io_req_find_next(req);
1366 		io_free_req(req);
1367 	}
1368 	return nxt;
1369 }
1370 
1371 static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1372 {
1373 	/* See comment at the top of this file */
1374 	smp_rmb();
1375 	return __io_cqring_events(ctx);
1376 }
1377 
1378 /*
1379  * We can't just wait for polled events to come to us, we have to actively
1380  * find and complete them.
1381  */
1382 static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1383 {
1384 	if (!(ctx->flags & IORING_SETUP_IOPOLL))
1385 		return;
1386 
1387 	mutex_lock(&ctx->uring_lock);
1388 	while (!wq_list_empty(&ctx->iopoll_list)) {
1389 		/* let it sleep and repeat later if can't complete a request */
1390 		if (io_do_iopoll(ctx, true) == 0)
1391 			break;
1392 		/*
1393 		 * Ensure we allow local-to-the-cpu processing to take place,
1394 		 * in this case we need to ensure that we reap all events.
1395 		 * Also let task_work, etc. to progress by releasing the mutex
1396 		 */
1397 		if (need_resched()) {
1398 			mutex_unlock(&ctx->uring_lock);
1399 			cond_resched();
1400 			mutex_lock(&ctx->uring_lock);
1401 		}
1402 	}
1403 	mutex_unlock(&ctx->uring_lock);
1404 }
1405 
1406 static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1407 {
1408 	unsigned int nr_events = 0;
1409 	int ret = 0;
1410 	unsigned long check_cq;
1411 
1412 	if (!io_allowed_run_tw(ctx))
1413 		return -EEXIST;
1414 
1415 	check_cq = READ_ONCE(ctx->check_cq);
1416 	if (unlikely(check_cq)) {
1417 		if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1418 			__io_cqring_overflow_flush(ctx, false);
1419 		/*
1420 		 * Similarly do not spin if we have not informed the user of any
1421 		 * dropped CQE.
1422 		 */
1423 		if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1424 			return -EBADR;
1425 	}
1426 	/*
1427 	 * Don't enter poll loop if we already have events pending.
1428 	 * If we do, we can potentially be spinning for commands that
1429 	 * already triggered a CQE (eg in error).
1430 	 */
1431 	if (io_cqring_events(ctx))
1432 		return 0;
1433 
1434 	do {
1435 		/*
1436 		 * If a submit got punted to a workqueue, we can have the
1437 		 * application entering polling for a command before it gets
1438 		 * issued. That app will hold the uring_lock for the duration
1439 		 * of the poll right here, so we need to take a breather every
1440 		 * now and then to ensure that the issue has a chance to add
1441 		 * the poll to the issued list. Otherwise we can spin here
1442 		 * forever, while the workqueue is stuck trying to acquire the
1443 		 * very same mutex.
1444 		 */
1445 		if (wq_list_empty(&ctx->iopoll_list) ||
1446 		    io_task_work_pending(ctx)) {
1447 			u32 tail = ctx->cached_cq_tail;
1448 
1449 			if (!llist_empty(&ctx->work_llist))
1450 				__io_run_local_work(ctx, true);
1451 
1452 			if (task_work_pending(current) ||
1453 			    wq_list_empty(&ctx->iopoll_list)) {
1454 				mutex_unlock(&ctx->uring_lock);
1455 				io_run_task_work();
1456 				mutex_lock(&ctx->uring_lock);
1457 			}
1458 			/* some requests don't go through iopoll_list */
1459 			if (tail != ctx->cached_cq_tail ||
1460 			    wq_list_empty(&ctx->iopoll_list))
1461 				break;
1462 		}
1463 		ret = io_do_iopoll(ctx, !min);
1464 		if (ret < 0)
1465 			break;
1466 		nr_events += ret;
1467 		ret = 0;
1468 	} while (nr_events < min && !need_resched());
1469 
1470 	return ret;
1471 }
1472 
1473 void io_req_task_complete(struct io_kiocb *req, bool *locked)
1474 {
1475 	if (req->flags & (REQ_F_BUFFER_SELECTED|REQ_F_BUFFER_RING)) {
1476 		unsigned issue_flags = *locked ? 0 : IO_URING_F_UNLOCKED;
1477 
1478 		req->cqe.flags |= io_put_kbuf(req, issue_flags);
1479 	}
1480 
1481 	if (*locked)
1482 		io_req_complete_defer(req);
1483 	else
1484 		io_req_complete_post(req);
1485 }
1486 
1487 /*
1488  * After the iocb has been issued, it's safe to be found on the poll list.
1489  * Adding the kiocb to the list AFTER submission ensures that we don't
1490  * find it from a io_do_iopoll() thread before the issuer is done
1491  * accessing the kiocb cookie.
1492  */
1493 static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1494 {
1495 	struct io_ring_ctx *ctx = req->ctx;
1496 	const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1497 
1498 	/* workqueue context doesn't hold uring_lock, grab it now */
1499 	if (unlikely(needs_lock))
1500 		mutex_lock(&ctx->uring_lock);
1501 
1502 	/*
1503 	 * Track whether we have multiple files in our lists. This will impact
1504 	 * how we do polling eventually, not spinning if we're on potentially
1505 	 * different devices.
1506 	 */
1507 	if (wq_list_empty(&ctx->iopoll_list)) {
1508 		ctx->poll_multi_queue = false;
1509 	} else if (!ctx->poll_multi_queue) {
1510 		struct io_kiocb *list_req;
1511 
1512 		list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1513 					comp_list);
1514 		if (list_req->file != req->file)
1515 			ctx->poll_multi_queue = true;
1516 	}
1517 
1518 	/*
1519 	 * For fast devices, IO may have already completed. If it has, add
1520 	 * it to the front so we find it first.
1521 	 */
1522 	if (READ_ONCE(req->iopoll_completed))
1523 		wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1524 	else
1525 		wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1526 
1527 	if (unlikely(needs_lock)) {
1528 		/*
1529 		 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1530 		 * in sq thread task context or in io worker task context. If
1531 		 * current task context is sq thread, we don't need to check
1532 		 * whether should wake up sq thread.
1533 		 */
1534 		if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1535 		    wq_has_sleeper(&ctx->sq_data->wait))
1536 			wake_up(&ctx->sq_data->wait);
1537 
1538 		mutex_unlock(&ctx->uring_lock);
1539 	}
1540 }
1541 
1542 static bool io_bdev_nowait(struct block_device *bdev)
1543 {
1544 	return !bdev || bdev_nowait(bdev);
1545 }
1546 
1547 /*
1548  * If we tracked the file through the SCM inflight mechanism, we could support
1549  * any file. For now, just ensure that anything potentially problematic is done
1550  * inline.
1551  */
1552 static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1553 {
1554 	if (S_ISBLK(mode)) {
1555 		if (IS_ENABLED(CONFIG_BLOCK) &&
1556 		    io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1557 			return true;
1558 		return false;
1559 	}
1560 	if (S_ISSOCK(mode))
1561 		return true;
1562 	if (S_ISREG(mode)) {
1563 		if (IS_ENABLED(CONFIG_BLOCK) &&
1564 		    io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1565 		    !io_is_uring_fops(file))
1566 			return true;
1567 		return false;
1568 	}
1569 
1570 	/* any ->read/write should understand O_NONBLOCK */
1571 	if (file->f_flags & O_NONBLOCK)
1572 		return true;
1573 	return file->f_mode & FMODE_NOWAIT;
1574 }
1575 
1576 /*
1577  * If we tracked the file through the SCM inflight mechanism, we could support
1578  * any file. For now, just ensure that anything potentially problematic is done
1579  * inline.
1580  */
1581 unsigned int io_file_get_flags(struct file *file)
1582 {
1583 	umode_t mode = file_inode(file)->i_mode;
1584 	unsigned int res = 0;
1585 
1586 	if (S_ISREG(mode))
1587 		res |= FFS_ISREG;
1588 	if (__io_file_supports_nowait(file, mode))
1589 		res |= FFS_NOWAIT;
1590 	if (io_file_need_scm(file))
1591 		res |= FFS_SCM;
1592 	return res;
1593 }
1594 
1595 bool io_alloc_async_data(struct io_kiocb *req)
1596 {
1597 	WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1598 	req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1599 	if (req->async_data) {
1600 		req->flags |= REQ_F_ASYNC_DATA;
1601 		return false;
1602 	}
1603 	return true;
1604 }
1605 
1606 int io_req_prep_async(struct io_kiocb *req)
1607 {
1608 	const struct io_op_def *def = &io_op_defs[req->opcode];
1609 
1610 	/* assign early for deferred execution for non-fixed file */
1611 	if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1612 		req->file = io_file_get_normal(req, req->cqe.fd);
1613 	if (!def->prep_async)
1614 		return 0;
1615 	if (WARN_ON_ONCE(req_has_async_data(req)))
1616 		return -EFAULT;
1617 	if (!io_op_defs[req->opcode].manual_alloc) {
1618 		if (io_alloc_async_data(req))
1619 			return -EAGAIN;
1620 	}
1621 	return def->prep_async(req);
1622 }
1623 
1624 static u32 io_get_sequence(struct io_kiocb *req)
1625 {
1626 	u32 seq = req->ctx->cached_sq_head;
1627 	struct io_kiocb *cur;
1628 
1629 	/* need original cached_sq_head, but it was increased for each req */
1630 	io_for_each_link(cur, req)
1631 		seq--;
1632 	return seq;
1633 }
1634 
1635 static __cold void io_drain_req(struct io_kiocb *req)
1636 {
1637 	struct io_ring_ctx *ctx = req->ctx;
1638 	struct io_defer_entry *de;
1639 	int ret;
1640 	u32 seq = io_get_sequence(req);
1641 
1642 	/* Still need defer if there is pending req in defer list. */
1643 	spin_lock(&ctx->completion_lock);
1644 	if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1645 		spin_unlock(&ctx->completion_lock);
1646 queue:
1647 		ctx->drain_active = false;
1648 		io_req_task_queue(req);
1649 		return;
1650 	}
1651 	spin_unlock(&ctx->completion_lock);
1652 
1653 	ret = io_req_prep_async(req);
1654 	if (ret) {
1655 fail:
1656 		io_req_complete_failed(req, ret);
1657 		return;
1658 	}
1659 	io_prep_async_link(req);
1660 	de = kmalloc(sizeof(*de), GFP_KERNEL);
1661 	if (!de) {
1662 		ret = -ENOMEM;
1663 		goto fail;
1664 	}
1665 
1666 	spin_lock(&ctx->completion_lock);
1667 	if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1668 		spin_unlock(&ctx->completion_lock);
1669 		kfree(de);
1670 		goto queue;
1671 	}
1672 
1673 	trace_io_uring_defer(req);
1674 	de->req = req;
1675 	de->seq = seq;
1676 	list_add_tail(&de->list, &ctx->defer_list);
1677 	spin_unlock(&ctx->completion_lock);
1678 }
1679 
1680 static void io_clean_op(struct io_kiocb *req)
1681 {
1682 	if (req->flags & REQ_F_BUFFER_SELECTED) {
1683 		spin_lock(&req->ctx->completion_lock);
1684 		io_put_kbuf_comp(req);
1685 		spin_unlock(&req->ctx->completion_lock);
1686 	}
1687 
1688 	if (req->flags & REQ_F_NEED_CLEANUP) {
1689 		const struct io_op_def *def = &io_op_defs[req->opcode];
1690 
1691 		if (def->cleanup)
1692 			def->cleanup(req);
1693 	}
1694 	if ((req->flags & REQ_F_POLLED) && req->apoll) {
1695 		kfree(req->apoll->double_poll);
1696 		kfree(req->apoll);
1697 		req->apoll = NULL;
1698 	}
1699 	if (req->flags & REQ_F_INFLIGHT) {
1700 		struct io_uring_task *tctx = req->task->io_uring;
1701 
1702 		atomic_dec(&tctx->inflight_tracked);
1703 	}
1704 	if (req->flags & REQ_F_CREDS)
1705 		put_cred(req->creds);
1706 	if (req->flags & REQ_F_ASYNC_DATA) {
1707 		kfree(req->async_data);
1708 		req->async_data = NULL;
1709 	}
1710 	req->flags &= ~IO_REQ_CLEAN_FLAGS;
1711 }
1712 
1713 static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1714 {
1715 	if (req->file || !io_op_defs[req->opcode].needs_file)
1716 		return true;
1717 
1718 	if (req->flags & REQ_F_FIXED_FILE)
1719 		req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1720 	else
1721 		req->file = io_file_get_normal(req, req->cqe.fd);
1722 
1723 	return !!req->file;
1724 }
1725 
1726 static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1727 {
1728 	const struct io_op_def *def = &io_op_defs[req->opcode];
1729 	const struct cred *creds = NULL;
1730 	int ret;
1731 
1732 	if (unlikely(!io_assign_file(req, issue_flags)))
1733 		return -EBADF;
1734 
1735 	if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1736 		creds = override_creds(req->creds);
1737 
1738 	if (!def->audit_skip)
1739 		audit_uring_entry(req->opcode);
1740 
1741 	ret = def->issue(req, issue_flags);
1742 
1743 	if (!def->audit_skip)
1744 		audit_uring_exit(!ret, ret);
1745 
1746 	if (creds)
1747 		revert_creds(creds);
1748 
1749 	if (ret == IOU_OK) {
1750 		if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1751 			io_req_complete_defer(req);
1752 		else
1753 			io_req_complete_post(req);
1754 	} else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1755 		return ret;
1756 
1757 	/* If the op doesn't have a file, we're not polling for it */
1758 	if ((req->ctx->flags & IORING_SETUP_IOPOLL) && req->file)
1759 		io_iopoll_req_issued(req, issue_flags);
1760 
1761 	return 0;
1762 }
1763 
1764 int io_poll_issue(struct io_kiocb *req, bool *locked)
1765 {
1766 	io_tw_lock(req->ctx, locked);
1767 	if (unlikely(req->task->flags & PF_EXITING))
1768 		return -EFAULT;
1769 	return io_issue_sqe(req, IO_URING_F_NONBLOCK);
1770 }
1771 
1772 struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1773 {
1774 	struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1775 
1776 	req = io_put_req_find_next(req);
1777 	return req ? &req->work : NULL;
1778 }
1779 
1780 void io_wq_submit_work(struct io_wq_work *work)
1781 {
1782 	struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1783 	const struct io_op_def *def = &io_op_defs[req->opcode];
1784 	unsigned int issue_flags = IO_URING_F_UNLOCKED;
1785 	bool needs_poll = false;
1786 	int ret = 0, err = -ECANCELED;
1787 
1788 	/* one will be dropped by ->io_free_work() after returning to io-wq */
1789 	if (!(req->flags & REQ_F_REFCOUNT))
1790 		__io_req_set_refcount(req, 2);
1791 	else
1792 		req_ref_get(req);
1793 
1794 	io_arm_ltimeout(req);
1795 
1796 	/* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1797 	if (work->flags & IO_WQ_WORK_CANCEL) {
1798 fail:
1799 		io_req_task_queue_fail(req, err);
1800 		return;
1801 	}
1802 	if (!io_assign_file(req, issue_flags)) {
1803 		err = -EBADF;
1804 		work->flags |= IO_WQ_WORK_CANCEL;
1805 		goto fail;
1806 	}
1807 
1808 	if (req->flags & REQ_F_FORCE_ASYNC) {
1809 		bool opcode_poll = def->pollin || def->pollout;
1810 
1811 		if (opcode_poll && file_can_poll(req->file)) {
1812 			needs_poll = true;
1813 			issue_flags |= IO_URING_F_NONBLOCK;
1814 		}
1815 	}
1816 
1817 	do {
1818 		ret = io_issue_sqe(req, issue_flags);
1819 		if (ret != -EAGAIN)
1820 			break;
1821 		/*
1822 		 * We can get EAGAIN for iopolled IO even though we're
1823 		 * forcing a sync submission from here, since we can't
1824 		 * wait for request slots on the block side.
1825 		 */
1826 		if (!needs_poll) {
1827 			if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1828 				break;
1829 			cond_resched();
1830 			continue;
1831 		}
1832 
1833 		if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1834 			return;
1835 		/* aborted or ready, in either case retry blocking */
1836 		needs_poll = false;
1837 		issue_flags &= ~IO_URING_F_NONBLOCK;
1838 	} while (1);
1839 
1840 	/* avoid locking problems by failing it from a clean context */
1841 	if (ret < 0)
1842 		io_req_task_queue_fail(req, ret);
1843 }
1844 
1845 inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1846 				      unsigned int issue_flags)
1847 {
1848 	struct io_ring_ctx *ctx = req->ctx;
1849 	struct file *file = NULL;
1850 	unsigned long file_ptr;
1851 
1852 	io_ring_submit_lock(ctx, issue_flags);
1853 
1854 	if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1855 		goto out;
1856 	fd = array_index_nospec(fd, ctx->nr_user_files);
1857 	file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1858 	file = (struct file *) (file_ptr & FFS_MASK);
1859 	file_ptr &= ~FFS_MASK;
1860 	/* mask in overlapping REQ_F and FFS bits */
1861 	req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1862 	io_req_set_rsrc_node(req, ctx, 0);
1863 	WARN_ON_ONCE(file && !test_bit(fd, ctx->file_table.bitmap));
1864 out:
1865 	io_ring_submit_unlock(ctx, issue_flags);
1866 	return file;
1867 }
1868 
1869 struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1870 {
1871 	struct file *file = fget(fd);
1872 
1873 	trace_io_uring_file_get(req, fd);
1874 
1875 	/* we don't allow fixed io_uring files */
1876 	if (file && io_is_uring_fops(file))
1877 		io_req_track_inflight(req);
1878 	return file;
1879 }
1880 
1881 static void io_queue_async(struct io_kiocb *req, int ret)
1882 	__must_hold(&req->ctx->uring_lock)
1883 {
1884 	struct io_kiocb *linked_timeout;
1885 
1886 	if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1887 		io_req_complete_failed(req, ret);
1888 		return;
1889 	}
1890 
1891 	linked_timeout = io_prep_linked_timeout(req);
1892 
1893 	switch (io_arm_poll_handler(req, 0)) {
1894 	case IO_APOLL_READY:
1895 		io_kbuf_recycle(req, 0);
1896 		io_req_task_queue(req);
1897 		break;
1898 	case IO_APOLL_ABORTED:
1899 		io_kbuf_recycle(req, 0);
1900 		io_queue_iowq(req, NULL);
1901 		break;
1902 	case IO_APOLL_OK:
1903 		break;
1904 	}
1905 
1906 	if (linked_timeout)
1907 		io_queue_linked_timeout(linked_timeout);
1908 }
1909 
1910 static inline void io_queue_sqe(struct io_kiocb *req)
1911 	__must_hold(&req->ctx->uring_lock)
1912 {
1913 	int ret;
1914 
1915 	ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
1916 
1917 	/*
1918 	 * We async punt it if the file wasn't marked NOWAIT, or if the file
1919 	 * doesn't support non-blocking read/write attempts
1920 	 */
1921 	if (likely(!ret))
1922 		io_arm_ltimeout(req);
1923 	else
1924 		io_queue_async(req, ret);
1925 }
1926 
1927 static void io_queue_sqe_fallback(struct io_kiocb *req)
1928 	__must_hold(&req->ctx->uring_lock)
1929 {
1930 	if (unlikely(req->flags & REQ_F_FAIL)) {
1931 		/*
1932 		 * We don't submit, fail them all, for that replace hardlinks
1933 		 * with normal links. Extra REQ_F_LINK is tolerated.
1934 		 */
1935 		req->flags &= ~REQ_F_HARDLINK;
1936 		req->flags |= REQ_F_LINK;
1937 		io_req_complete_failed(req, req->cqe.res);
1938 	} else if (unlikely(req->ctx->drain_active)) {
1939 		io_drain_req(req);
1940 	} else {
1941 		int ret = io_req_prep_async(req);
1942 
1943 		if (unlikely(ret))
1944 			io_req_complete_failed(req, ret);
1945 		else
1946 			io_queue_iowq(req, NULL);
1947 	}
1948 }
1949 
1950 /*
1951  * Check SQE restrictions (opcode and flags).
1952  *
1953  * Returns 'true' if SQE is allowed, 'false' otherwise.
1954  */
1955 static inline bool io_check_restriction(struct io_ring_ctx *ctx,
1956 					struct io_kiocb *req,
1957 					unsigned int sqe_flags)
1958 {
1959 	if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
1960 		return false;
1961 
1962 	if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
1963 	    ctx->restrictions.sqe_flags_required)
1964 		return false;
1965 
1966 	if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
1967 			  ctx->restrictions.sqe_flags_required))
1968 		return false;
1969 
1970 	return true;
1971 }
1972 
1973 static void io_init_req_drain(struct io_kiocb *req)
1974 {
1975 	struct io_ring_ctx *ctx = req->ctx;
1976 	struct io_kiocb *head = ctx->submit_state.link.head;
1977 
1978 	ctx->drain_active = true;
1979 	if (head) {
1980 		/*
1981 		 * If we need to drain a request in the middle of a link, drain
1982 		 * the head request and the next request/link after the current
1983 		 * link. Considering sequential execution of links,
1984 		 * REQ_F_IO_DRAIN will be maintained for every request of our
1985 		 * link.
1986 		 */
1987 		head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
1988 		ctx->drain_next = true;
1989 	}
1990 }
1991 
1992 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
1993 		       const struct io_uring_sqe *sqe)
1994 	__must_hold(&ctx->uring_lock)
1995 {
1996 	const struct io_op_def *def;
1997 	unsigned int sqe_flags;
1998 	int personality;
1999 	u8 opcode;
2000 
2001 	/* req is partially pre-initialised, see io_preinit_req() */
2002 	req->opcode = opcode = READ_ONCE(sqe->opcode);
2003 	/* same numerical values with corresponding REQ_F_*, safe to copy */
2004 	req->flags = sqe_flags = READ_ONCE(sqe->flags);
2005 	req->cqe.user_data = READ_ONCE(sqe->user_data);
2006 	req->file = NULL;
2007 	req->rsrc_node = NULL;
2008 	req->task = current;
2009 
2010 	if (unlikely(opcode >= IORING_OP_LAST)) {
2011 		req->opcode = 0;
2012 		return -EINVAL;
2013 	}
2014 	def = &io_op_defs[opcode];
2015 	if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2016 		/* enforce forwards compatibility on users */
2017 		if (sqe_flags & ~SQE_VALID_FLAGS)
2018 			return -EINVAL;
2019 		if (sqe_flags & IOSQE_BUFFER_SELECT) {
2020 			if (!def->buffer_select)
2021 				return -EOPNOTSUPP;
2022 			req->buf_index = READ_ONCE(sqe->buf_group);
2023 		}
2024 		if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2025 			ctx->drain_disabled = true;
2026 		if (sqe_flags & IOSQE_IO_DRAIN) {
2027 			if (ctx->drain_disabled)
2028 				return -EOPNOTSUPP;
2029 			io_init_req_drain(req);
2030 		}
2031 	}
2032 	if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2033 		if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2034 			return -EACCES;
2035 		/* knock it to the slow queue path, will be drained there */
2036 		if (ctx->drain_active)
2037 			req->flags |= REQ_F_FORCE_ASYNC;
2038 		/* if there is no link, we're at "next" request and need to drain */
2039 		if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2040 			ctx->drain_next = false;
2041 			ctx->drain_active = true;
2042 			req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2043 		}
2044 	}
2045 
2046 	if (!def->ioprio && sqe->ioprio)
2047 		return -EINVAL;
2048 	if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2049 		return -EINVAL;
2050 
2051 	if (def->needs_file) {
2052 		struct io_submit_state *state = &ctx->submit_state;
2053 
2054 		req->cqe.fd = READ_ONCE(sqe->fd);
2055 
2056 		/*
2057 		 * Plug now if we have more than 2 IO left after this, and the
2058 		 * target is potentially a read/write to block based storage.
2059 		 */
2060 		if (state->need_plug && def->plug) {
2061 			state->plug_started = true;
2062 			state->need_plug = false;
2063 			blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2064 		}
2065 	}
2066 
2067 	personality = READ_ONCE(sqe->personality);
2068 	if (personality) {
2069 		int ret;
2070 
2071 		req->creds = xa_load(&ctx->personalities, personality);
2072 		if (!req->creds)
2073 			return -EINVAL;
2074 		get_cred(req->creds);
2075 		ret = security_uring_override_creds(req->creds);
2076 		if (ret) {
2077 			put_cred(req->creds);
2078 			return ret;
2079 		}
2080 		req->flags |= REQ_F_CREDS;
2081 	}
2082 
2083 	return def->prep(req, sqe);
2084 }
2085 
2086 static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2087 				      struct io_kiocb *req, int ret)
2088 {
2089 	struct io_ring_ctx *ctx = req->ctx;
2090 	struct io_submit_link *link = &ctx->submit_state.link;
2091 	struct io_kiocb *head = link->head;
2092 
2093 	trace_io_uring_req_failed(sqe, req, ret);
2094 
2095 	/*
2096 	 * Avoid breaking links in the middle as it renders links with SQPOLL
2097 	 * unusable. Instead of failing eagerly, continue assembling the link if
2098 	 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2099 	 * should find the flag and handle the rest.
2100 	 */
2101 	req_fail_link_node(req, ret);
2102 	if (head && !(head->flags & REQ_F_FAIL))
2103 		req_fail_link_node(head, -ECANCELED);
2104 
2105 	if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2106 		if (head) {
2107 			link->last->link = req;
2108 			link->head = NULL;
2109 			req = head;
2110 		}
2111 		io_queue_sqe_fallback(req);
2112 		return ret;
2113 	}
2114 
2115 	if (head)
2116 		link->last->link = req;
2117 	else
2118 		link->head = req;
2119 	link->last = req;
2120 	return 0;
2121 }
2122 
2123 static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2124 			 const struct io_uring_sqe *sqe)
2125 	__must_hold(&ctx->uring_lock)
2126 {
2127 	struct io_submit_link *link = &ctx->submit_state.link;
2128 	int ret;
2129 
2130 	ret = io_init_req(ctx, req, sqe);
2131 	if (unlikely(ret))
2132 		return io_submit_fail_init(sqe, req, ret);
2133 
2134 	/* don't need @sqe from now on */
2135 	trace_io_uring_submit_sqe(req, true);
2136 
2137 	/*
2138 	 * If we already have a head request, queue this one for async
2139 	 * submittal once the head completes. If we don't have a head but
2140 	 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2141 	 * submitted sync once the chain is complete. If none of those
2142 	 * conditions are true (normal request), then just queue it.
2143 	 */
2144 	if (unlikely(link->head)) {
2145 		ret = io_req_prep_async(req);
2146 		if (unlikely(ret))
2147 			return io_submit_fail_init(sqe, req, ret);
2148 
2149 		trace_io_uring_link(req, link->head);
2150 		link->last->link = req;
2151 		link->last = req;
2152 
2153 		if (req->flags & IO_REQ_LINK_FLAGS)
2154 			return 0;
2155 		/* last request of the link, flush it */
2156 		req = link->head;
2157 		link->head = NULL;
2158 		if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2159 			goto fallback;
2160 
2161 	} else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2162 					  REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2163 		if (req->flags & IO_REQ_LINK_FLAGS) {
2164 			link->head = req;
2165 			link->last = req;
2166 		} else {
2167 fallback:
2168 			io_queue_sqe_fallback(req);
2169 		}
2170 		return 0;
2171 	}
2172 
2173 	io_queue_sqe(req);
2174 	return 0;
2175 }
2176 
2177 /*
2178  * Batched submission is done, ensure local IO is flushed out.
2179  */
2180 static void io_submit_state_end(struct io_ring_ctx *ctx)
2181 {
2182 	struct io_submit_state *state = &ctx->submit_state;
2183 
2184 	if (unlikely(state->link.head))
2185 		io_queue_sqe_fallback(state->link.head);
2186 	/* flush only after queuing links as they can generate completions */
2187 	io_submit_flush_completions(ctx);
2188 	if (state->plug_started)
2189 		blk_finish_plug(&state->plug);
2190 }
2191 
2192 /*
2193  * Start submission side cache.
2194  */
2195 static void io_submit_state_start(struct io_submit_state *state,
2196 				  unsigned int max_ios)
2197 {
2198 	state->plug_started = false;
2199 	state->need_plug = max_ios > 2;
2200 	state->submit_nr = max_ios;
2201 	/* set only head, no need to init link_last in advance */
2202 	state->link.head = NULL;
2203 }
2204 
2205 static void io_commit_sqring(struct io_ring_ctx *ctx)
2206 {
2207 	struct io_rings *rings = ctx->rings;
2208 
2209 	/*
2210 	 * Ensure any loads from the SQEs are done at this point,
2211 	 * since once we write the new head, the application could
2212 	 * write new data to them.
2213 	 */
2214 	smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2215 }
2216 
2217 /*
2218  * Fetch an sqe, if one is available. Note this returns a pointer to memory
2219  * that is mapped by userspace. This means that care needs to be taken to
2220  * ensure that reads are stable, as we cannot rely on userspace always
2221  * being a good citizen. If members of the sqe are validated and then later
2222  * used, it's important that those reads are done through READ_ONCE() to
2223  * prevent a re-load down the line.
2224  */
2225 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2226 {
2227 	unsigned head, mask = ctx->sq_entries - 1;
2228 	unsigned sq_idx = ctx->cached_sq_head++ & mask;
2229 
2230 	/*
2231 	 * The cached sq head (or cq tail) serves two purposes:
2232 	 *
2233 	 * 1) allows us to batch the cost of updating the user visible
2234 	 *    head updates.
2235 	 * 2) allows the kernel side to track the head on its own, even
2236 	 *    though the application is the one updating it.
2237 	 */
2238 	head = READ_ONCE(ctx->sq_array[sq_idx]);
2239 	if (likely(head < ctx->sq_entries)) {
2240 		/* double index for 128-byte SQEs, twice as long */
2241 		if (ctx->flags & IORING_SETUP_SQE128)
2242 			head <<= 1;
2243 		return &ctx->sq_sqes[head];
2244 	}
2245 
2246 	/* drop invalid entries */
2247 	ctx->cq_extra--;
2248 	WRITE_ONCE(ctx->rings->sq_dropped,
2249 		   READ_ONCE(ctx->rings->sq_dropped) + 1);
2250 	return NULL;
2251 }
2252 
2253 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2254 	__must_hold(&ctx->uring_lock)
2255 {
2256 	unsigned int entries = io_sqring_entries(ctx);
2257 	unsigned int left;
2258 	int ret;
2259 
2260 	if (unlikely(!entries))
2261 		return 0;
2262 	/* make sure SQ entry isn't read before tail */
2263 	ret = left = min3(nr, ctx->sq_entries, entries);
2264 	io_get_task_refs(left);
2265 	io_submit_state_start(&ctx->submit_state, left);
2266 
2267 	do {
2268 		const struct io_uring_sqe *sqe;
2269 		struct io_kiocb *req;
2270 
2271 		if (unlikely(!io_alloc_req_refill(ctx)))
2272 			break;
2273 		req = io_alloc_req(ctx);
2274 		sqe = io_get_sqe(ctx);
2275 		if (unlikely(!sqe)) {
2276 			io_req_add_to_cache(req, ctx);
2277 			break;
2278 		}
2279 
2280 		/*
2281 		 * Continue submitting even for sqe failure if the
2282 		 * ring was setup with IORING_SETUP_SUBMIT_ALL
2283 		 */
2284 		if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2285 		    !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2286 			left--;
2287 			break;
2288 		}
2289 	} while (--left);
2290 
2291 	if (unlikely(left)) {
2292 		ret -= left;
2293 		/* try again if it submitted nothing and can't allocate a req */
2294 		if (!ret && io_req_cache_empty(ctx))
2295 			ret = -EAGAIN;
2296 		current->io_uring->cached_refs += left;
2297 	}
2298 
2299 	io_submit_state_end(ctx);
2300 	 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2301 	io_commit_sqring(ctx);
2302 	return ret;
2303 }
2304 
2305 struct io_wait_queue {
2306 	struct wait_queue_entry wq;
2307 	struct io_ring_ctx *ctx;
2308 	unsigned cq_tail;
2309 	unsigned nr_timeouts;
2310 };
2311 
2312 static inline bool io_has_work(struct io_ring_ctx *ctx)
2313 {
2314 	return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2315 	       ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
2316 		!llist_empty(&ctx->work_llist));
2317 }
2318 
2319 static inline bool io_should_wake(struct io_wait_queue *iowq)
2320 {
2321 	struct io_ring_ctx *ctx = iowq->ctx;
2322 	int dist = ctx->cached_cq_tail - (int) iowq->cq_tail;
2323 
2324 	/*
2325 	 * Wake up if we have enough events, or if a timeout occurred since we
2326 	 * started waiting. For timeouts, we always want to return to userspace,
2327 	 * regardless of event count.
2328 	 */
2329 	return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2330 }
2331 
2332 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2333 			    int wake_flags, void *key)
2334 {
2335 	struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2336 							wq);
2337 	struct io_ring_ctx *ctx = iowq->ctx;
2338 
2339 	/*
2340 	 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2341 	 * the task, and the next invocation will do it.
2342 	 */
2343 	if (io_should_wake(iowq) || io_has_work(ctx))
2344 		return autoremove_wake_function(curr, mode, wake_flags, key);
2345 	return -1;
2346 }
2347 
2348 int io_run_task_work_sig(struct io_ring_ctx *ctx)
2349 {
2350 	if (io_run_task_work_ctx(ctx) > 0)
2351 		return 1;
2352 	if (task_sigpending(current))
2353 		return -EINTR;
2354 	return 0;
2355 }
2356 
2357 /* when returns >0, the caller should retry */
2358 static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2359 					  struct io_wait_queue *iowq,
2360 					  ktime_t timeout)
2361 {
2362 	int ret;
2363 	unsigned long check_cq;
2364 
2365 	/* make sure we run task_work before checking for signals */
2366 	ret = io_run_task_work_sig(ctx);
2367 	if (ret || io_should_wake(iowq))
2368 		return ret;
2369 
2370 	check_cq = READ_ONCE(ctx->check_cq);
2371 	if (unlikely(check_cq)) {
2372 		/* let the caller flush overflows, retry */
2373 		if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2374 			return 1;
2375 		if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2376 			return -EBADR;
2377 	}
2378 	if (!schedule_hrtimeout(&timeout, HRTIMER_MODE_ABS))
2379 		return -ETIME;
2380 	return 1;
2381 }
2382 
2383 /*
2384  * Wait until events become available, if we don't already have some. The
2385  * application must reap them itself, as they reside on the shared cq ring.
2386  */
2387 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2388 			  const sigset_t __user *sig, size_t sigsz,
2389 			  struct __kernel_timespec __user *uts)
2390 {
2391 	struct io_wait_queue iowq;
2392 	struct io_rings *rings = ctx->rings;
2393 	ktime_t timeout = KTIME_MAX;
2394 	int ret;
2395 
2396 	if (!io_allowed_run_tw(ctx))
2397 		return -EEXIST;
2398 
2399 	do {
2400 		/* always run at least 1 task work to process local work */
2401 		ret = io_run_task_work_ctx(ctx);
2402 		if (ret < 0)
2403 			return ret;
2404 		io_cqring_overflow_flush(ctx);
2405 
2406 		if (io_cqring_events(ctx) >= min_events)
2407 			return 0;
2408 	} while (ret > 0);
2409 
2410 	if (sig) {
2411 #ifdef CONFIG_COMPAT
2412 		if (in_compat_syscall())
2413 			ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2414 						      sigsz);
2415 		else
2416 #endif
2417 			ret = set_user_sigmask(sig, sigsz);
2418 
2419 		if (ret)
2420 			return ret;
2421 	}
2422 
2423 	if (uts) {
2424 		struct timespec64 ts;
2425 
2426 		if (get_timespec64(&ts, uts))
2427 			return -EFAULT;
2428 		timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2429 	}
2430 
2431 	init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2432 	iowq.wq.private = current;
2433 	INIT_LIST_HEAD(&iowq.wq.entry);
2434 	iowq.ctx = ctx;
2435 	iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2436 	iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2437 
2438 	trace_io_uring_cqring_wait(ctx, min_events);
2439 	do {
2440 		/* if we can't even flush overflow, don't wait for more */
2441 		if (!io_cqring_overflow_flush(ctx)) {
2442 			ret = -EBUSY;
2443 			break;
2444 		}
2445 		prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2446 						TASK_INTERRUPTIBLE);
2447 		ret = io_cqring_wait_schedule(ctx, &iowq, timeout);
2448 		cond_resched();
2449 	} while (ret > 0);
2450 
2451 	finish_wait(&ctx->cq_wait, &iowq.wq);
2452 	restore_saved_sigmask_unless(ret == -EINTR);
2453 
2454 	return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2455 }
2456 
2457 static void io_mem_free(void *ptr)
2458 {
2459 	struct page *page;
2460 
2461 	if (!ptr)
2462 		return;
2463 
2464 	page = virt_to_head_page(ptr);
2465 	if (put_page_testzero(page))
2466 		free_compound_page(page);
2467 }
2468 
2469 static void *io_mem_alloc(size_t size)
2470 {
2471 	gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2472 
2473 	return (void *) __get_free_pages(gfp, get_order(size));
2474 }
2475 
2476 static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2477 				unsigned int cq_entries, size_t *sq_offset)
2478 {
2479 	struct io_rings *rings;
2480 	size_t off, sq_array_size;
2481 
2482 	off = struct_size(rings, cqes, cq_entries);
2483 	if (off == SIZE_MAX)
2484 		return SIZE_MAX;
2485 	if (ctx->flags & IORING_SETUP_CQE32) {
2486 		if (check_shl_overflow(off, 1, &off))
2487 			return SIZE_MAX;
2488 	}
2489 
2490 #ifdef CONFIG_SMP
2491 	off = ALIGN(off, SMP_CACHE_BYTES);
2492 	if (off == 0)
2493 		return SIZE_MAX;
2494 #endif
2495 
2496 	if (sq_offset)
2497 		*sq_offset = off;
2498 
2499 	sq_array_size = array_size(sizeof(u32), sq_entries);
2500 	if (sq_array_size == SIZE_MAX)
2501 		return SIZE_MAX;
2502 
2503 	if (check_add_overflow(off, sq_array_size, &off))
2504 		return SIZE_MAX;
2505 
2506 	return off;
2507 }
2508 
2509 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2510 			       unsigned int eventfd_async)
2511 {
2512 	struct io_ev_fd *ev_fd;
2513 	__s32 __user *fds = arg;
2514 	int fd;
2515 
2516 	ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2517 					lockdep_is_held(&ctx->uring_lock));
2518 	if (ev_fd)
2519 		return -EBUSY;
2520 
2521 	if (copy_from_user(&fd, fds, sizeof(*fds)))
2522 		return -EFAULT;
2523 
2524 	ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2525 	if (!ev_fd)
2526 		return -ENOMEM;
2527 
2528 	ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2529 	if (IS_ERR(ev_fd->cq_ev_fd)) {
2530 		int ret = PTR_ERR(ev_fd->cq_ev_fd);
2531 		kfree(ev_fd);
2532 		return ret;
2533 	}
2534 
2535 	spin_lock(&ctx->completion_lock);
2536 	ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2537 	spin_unlock(&ctx->completion_lock);
2538 
2539 	ev_fd->eventfd_async = eventfd_async;
2540 	ctx->has_evfd = true;
2541 	rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2542 	atomic_set(&ev_fd->refs, 1);
2543 	atomic_set(&ev_fd->ops, 0);
2544 	return 0;
2545 }
2546 
2547 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2548 {
2549 	struct io_ev_fd *ev_fd;
2550 
2551 	ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2552 					lockdep_is_held(&ctx->uring_lock));
2553 	if (ev_fd) {
2554 		ctx->has_evfd = false;
2555 		rcu_assign_pointer(ctx->io_ev_fd, NULL);
2556 		if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2557 			call_rcu(&ev_fd->rcu, io_eventfd_ops);
2558 		return 0;
2559 	}
2560 
2561 	return -ENXIO;
2562 }
2563 
2564 static void io_req_caches_free(struct io_ring_ctx *ctx)
2565 {
2566 	struct io_submit_state *state = &ctx->submit_state;
2567 	int nr = 0;
2568 
2569 	mutex_lock(&ctx->uring_lock);
2570 	io_flush_cached_locked_reqs(ctx, state);
2571 
2572 	while (!io_req_cache_empty(ctx)) {
2573 		struct io_wq_work_node *node;
2574 		struct io_kiocb *req;
2575 
2576 		node = wq_stack_extract(&state->free_list);
2577 		req = container_of(node, struct io_kiocb, comp_list);
2578 		kmem_cache_free(req_cachep, req);
2579 		nr++;
2580 	}
2581 	if (nr)
2582 		percpu_ref_put_many(&ctx->refs, nr);
2583 	mutex_unlock(&ctx->uring_lock);
2584 }
2585 
2586 static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2587 {
2588 	io_sq_thread_finish(ctx);
2589 	io_rsrc_refs_drop(ctx);
2590 	/* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2591 	io_wait_rsrc_data(ctx->buf_data);
2592 	io_wait_rsrc_data(ctx->file_data);
2593 
2594 	mutex_lock(&ctx->uring_lock);
2595 	if (ctx->buf_data)
2596 		__io_sqe_buffers_unregister(ctx);
2597 	if (ctx->file_data)
2598 		__io_sqe_files_unregister(ctx);
2599 	if (ctx->rings)
2600 		__io_cqring_overflow_flush(ctx, true);
2601 	io_eventfd_unregister(ctx);
2602 	io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2603 	io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2604 	mutex_unlock(&ctx->uring_lock);
2605 	io_destroy_buffers(ctx);
2606 	if (ctx->sq_creds)
2607 		put_cred(ctx->sq_creds);
2608 	if (ctx->submitter_task)
2609 		put_task_struct(ctx->submitter_task);
2610 
2611 	/* there are no registered resources left, nobody uses it */
2612 	if (ctx->rsrc_node)
2613 		io_rsrc_node_destroy(ctx->rsrc_node);
2614 	if (ctx->rsrc_backup_node)
2615 		io_rsrc_node_destroy(ctx->rsrc_backup_node);
2616 	flush_delayed_work(&ctx->rsrc_put_work);
2617 	flush_delayed_work(&ctx->fallback_work);
2618 
2619 	WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2620 	WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2621 
2622 #if defined(CONFIG_UNIX)
2623 	if (ctx->ring_sock) {
2624 		ctx->ring_sock->file = NULL; /* so that iput() is called */
2625 		sock_release(ctx->ring_sock);
2626 	}
2627 #endif
2628 	WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2629 
2630 	if (ctx->mm_account) {
2631 		mmdrop(ctx->mm_account);
2632 		ctx->mm_account = NULL;
2633 	}
2634 	io_mem_free(ctx->rings);
2635 	io_mem_free(ctx->sq_sqes);
2636 
2637 	percpu_ref_exit(&ctx->refs);
2638 	free_uid(ctx->user);
2639 	io_req_caches_free(ctx);
2640 	if (ctx->hash_map)
2641 		io_wq_put_hash(ctx->hash_map);
2642 	kfree(ctx->cancel_table.hbs);
2643 	kfree(ctx->cancel_table_locked.hbs);
2644 	kfree(ctx->dummy_ubuf);
2645 	kfree(ctx->io_bl);
2646 	xa_destroy(&ctx->io_bl_xa);
2647 	kfree(ctx);
2648 }
2649 
2650 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2651 {
2652 	struct io_ring_ctx *ctx = file->private_data;
2653 	__poll_t mask = 0;
2654 
2655 	poll_wait(file, &ctx->cq_wait, wait);
2656 	/*
2657 	 * synchronizes with barrier from wq_has_sleeper call in
2658 	 * io_commit_cqring
2659 	 */
2660 	smp_rmb();
2661 	if (!io_sqring_full(ctx))
2662 		mask |= EPOLLOUT | EPOLLWRNORM;
2663 
2664 	/*
2665 	 * Don't flush cqring overflow list here, just do a simple check.
2666 	 * Otherwise there could possible be ABBA deadlock:
2667 	 *      CPU0                    CPU1
2668 	 *      ----                    ----
2669 	 * lock(&ctx->uring_lock);
2670 	 *                              lock(&ep->mtx);
2671 	 *                              lock(&ctx->uring_lock);
2672 	 * lock(&ep->mtx);
2673 	 *
2674 	 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2675 	 * pushs them to do the flush.
2676 	 */
2677 
2678 	if (io_cqring_events(ctx) || io_has_work(ctx))
2679 		mask |= EPOLLIN | EPOLLRDNORM;
2680 
2681 	return mask;
2682 }
2683 
2684 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2685 {
2686 	const struct cred *creds;
2687 
2688 	creds = xa_erase(&ctx->personalities, id);
2689 	if (creds) {
2690 		put_cred(creds);
2691 		return 0;
2692 	}
2693 
2694 	return -EINVAL;
2695 }
2696 
2697 struct io_tctx_exit {
2698 	struct callback_head		task_work;
2699 	struct completion		completion;
2700 	struct io_ring_ctx		*ctx;
2701 };
2702 
2703 static __cold void io_tctx_exit_cb(struct callback_head *cb)
2704 {
2705 	struct io_uring_task *tctx = current->io_uring;
2706 	struct io_tctx_exit *work;
2707 
2708 	work = container_of(cb, struct io_tctx_exit, task_work);
2709 	/*
2710 	 * When @in_idle, we're in cancellation and it's racy to remove the
2711 	 * node. It'll be removed by the end of cancellation, just ignore it.
2712 	 */
2713 	if (!atomic_read(&tctx->in_idle))
2714 		io_uring_del_tctx_node((unsigned long)work->ctx);
2715 	complete(&work->completion);
2716 }
2717 
2718 static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2719 {
2720 	struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2721 
2722 	return req->ctx == data;
2723 }
2724 
2725 static __cold void io_ring_exit_work(struct work_struct *work)
2726 {
2727 	struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2728 	unsigned long timeout = jiffies + HZ * 60 * 5;
2729 	unsigned long interval = HZ / 20;
2730 	struct io_tctx_exit exit;
2731 	struct io_tctx_node *node;
2732 	int ret;
2733 
2734 	/*
2735 	 * If we're doing polled IO and end up having requests being
2736 	 * submitted async (out-of-line), then completions can come in while
2737 	 * we're waiting for refs to drop. We need to reap these manually,
2738 	 * as nobody else will be looking for them.
2739 	 */
2740 	do {
2741 		if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2742 			io_move_task_work_from_local(ctx);
2743 
2744 		while (io_uring_try_cancel_requests(ctx, NULL, true))
2745 			cond_resched();
2746 
2747 		if (ctx->sq_data) {
2748 			struct io_sq_data *sqd = ctx->sq_data;
2749 			struct task_struct *tsk;
2750 
2751 			io_sq_thread_park(sqd);
2752 			tsk = sqd->thread;
2753 			if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2754 				io_wq_cancel_cb(tsk->io_uring->io_wq,
2755 						io_cancel_ctx_cb, ctx, true);
2756 			io_sq_thread_unpark(sqd);
2757 		}
2758 
2759 		io_req_caches_free(ctx);
2760 
2761 		if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2762 			/* there is little hope left, don't run it too often */
2763 			interval = HZ * 60;
2764 		}
2765 	} while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2766 
2767 	init_completion(&exit.completion);
2768 	init_task_work(&exit.task_work, io_tctx_exit_cb);
2769 	exit.ctx = ctx;
2770 	/*
2771 	 * Some may use context even when all refs and requests have been put,
2772 	 * and they are free to do so while still holding uring_lock or
2773 	 * completion_lock, see io_req_task_submit(). Apart from other work,
2774 	 * this lock/unlock section also waits them to finish.
2775 	 */
2776 	mutex_lock(&ctx->uring_lock);
2777 	while (!list_empty(&ctx->tctx_list)) {
2778 		WARN_ON_ONCE(time_after(jiffies, timeout));
2779 
2780 		node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2781 					ctx_node);
2782 		/* don't spin on a single task if cancellation failed */
2783 		list_rotate_left(&ctx->tctx_list);
2784 		ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2785 		if (WARN_ON_ONCE(ret))
2786 			continue;
2787 
2788 		mutex_unlock(&ctx->uring_lock);
2789 		wait_for_completion(&exit.completion);
2790 		mutex_lock(&ctx->uring_lock);
2791 	}
2792 	mutex_unlock(&ctx->uring_lock);
2793 	spin_lock(&ctx->completion_lock);
2794 	spin_unlock(&ctx->completion_lock);
2795 
2796 	io_ring_ctx_free(ctx);
2797 }
2798 
2799 static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2800 {
2801 	unsigned long index;
2802 	struct creds *creds;
2803 
2804 	mutex_lock(&ctx->uring_lock);
2805 	percpu_ref_kill(&ctx->refs);
2806 	if (ctx->rings)
2807 		__io_cqring_overflow_flush(ctx, true);
2808 	xa_for_each(&ctx->personalities, index, creds)
2809 		io_unregister_personality(ctx, index);
2810 	if (ctx->rings)
2811 		io_poll_remove_all(ctx, NULL, true);
2812 	mutex_unlock(&ctx->uring_lock);
2813 
2814 	/* failed during ring init, it couldn't have issued any requests */
2815 	if (ctx->rings) {
2816 		io_kill_timeouts(ctx, NULL, true);
2817 		/* if we failed setting up the ctx, we might not have any rings */
2818 		io_iopoll_try_reap_events(ctx);
2819 		/* drop cached put refs after potentially doing completions */
2820 		if (current->io_uring)
2821 			io_uring_drop_tctx_refs(current);
2822 	}
2823 
2824 	INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2825 	/*
2826 	 * Use system_unbound_wq to avoid spawning tons of event kworkers
2827 	 * if we're exiting a ton of rings at the same time. It just adds
2828 	 * noise and overhead, there's no discernable change in runtime
2829 	 * over using system_wq.
2830 	 */
2831 	queue_work(system_unbound_wq, &ctx->exit_work);
2832 }
2833 
2834 static int io_uring_release(struct inode *inode, struct file *file)
2835 {
2836 	struct io_ring_ctx *ctx = file->private_data;
2837 
2838 	file->private_data = NULL;
2839 	io_ring_ctx_wait_and_kill(ctx);
2840 	return 0;
2841 }
2842 
2843 struct io_task_cancel {
2844 	struct task_struct *task;
2845 	bool all;
2846 };
2847 
2848 static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2849 {
2850 	struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2851 	struct io_task_cancel *cancel = data;
2852 
2853 	return io_match_task_safe(req, cancel->task, cancel->all);
2854 }
2855 
2856 static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2857 					 struct task_struct *task,
2858 					 bool cancel_all)
2859 {
2860 	struct io_defer_entry *de;
2861 	LIST_HEAD(list);
2862 
2863 	spin_lock(&ctx->completion_lock);
2864 	list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2865 		if (io_match_task_safe(de->req, task, cancel_all)) {
2866 			list_cut_position(&list, &ctx->defer_list, &de->list);
2867 			break;
2868 		}
2869 	}
2870 	spin_unlock(&ctx->completion_lock);
2871 	if (list_empty(&list))
2872 		return false;
2873 
2874 	while (!list_empty(&list)) {
2875 		de = list_first_entry(&list, struct io_defer_entry, list);
2876 		list_del_init(&de->list);
2877 		io_req_complete_failed(de->req, -ECANCELED);
2878 		kfree(de);
2879 	}
2880 	return true;
2881 }
2882 
2883 static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
2884 {
2885 	struct io_tctx_node *node;
2886 	enum io_wq_cancel cret;
2887 	bool ret = false;
2888 
2889 	mutex_lock(&ctx->uring_lock);
2890 	list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
2891 		struct io_uring_task *tctx = node->task->io_uring;
2892 
2893 		/*
2894 		 * io_wq will stay alive while we hold uring_lock, because it's
2895 		 * killed after ctx nodes, which requires to take the lock.
2896 		 */
2897 		if (!tctx || !tctx->io_wq)
2898 			continue;
2899 		cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
2900 		ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2901 	}
2902 	mutex_unlock(&ctx->uring_lock);
2903 
2904 	return ret;
2905 }
2906 
2907 static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
2908 						struct task_struct *task,
2909 						bool cancel_all)
2910 {
2911 	struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
2912 	struct io_uring_task *tctx = task ? task->io_uring : NULL;
2913 	enum io_wq_cancel cret;
2914 	bool ret = false;
2915 
2916 	/* failed during ring init, it couldn't have issued any requests */
2917 	if (!ctx->rings)
2918 		return false;
2919 
2920 	if (!task) {
2921 		ret |= io_uring_try_cancel_iowq(ctx);
2922 	} else if (tctx && tctx->io_wq) {
2923 		/*
2924 		 * Cancels requests of all rings, not only @ctx, but
2925 		 * it's fine as the task is in exit/exec.
2926 		 */
2927 		cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
2928 				       &cancel, true);
2929 		ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
2930 	}
2931 
2932 	/* SQPOLL thread does its own polling */
2933 	if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
2934 	    (ctx->sq_data && ctx->sq_data->thread == current)) {
2935 		while (!wq_list_empty(&ctx->iopoll_list)) {
2936 			io_iopoll_try_reap_events(ctx);
2937 			ret = true;
2938 		}
2939 	}
2940 
2941 	if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2942 		ret |= io_run_local_work(ctx) > 0;
2943 	ret |= io_cancel_defer_files(ctx, task, cancel_all);
2944 	mutex_lock(&ctx->uring_lock);
2945 	ret |= io_poll_remove_all(ctx, task, cancel_all);
2946 	mutex_unlock(&ctx->uring_lock);
2947 	ret |= io_kill_timeouts(ctx, task, cancel_all);
2948 	if (task)
2949 		ret |= io_run_task_work() > 0;
2950 	return ret;
2951 }
2952 
2953 static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
2954 {
2955 	if (tracked)
2956 		return atomic_read(&tctx->inflight_tracked);
2957 	return percpu_counter_sum(&tctx->inflight);
2958 }
2959 
2960 /*
2961  * Find any io_uring ctx that this task has registered or done IO on, and cancel
2962  * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
2963  */
2964 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
2965 {
2966 	struct io_uring_task *tctx = current->io_uring;
2967 	struct io_ring_ctx *ctx;
2968 	s64 inflight;
2969 	DEFINE_WAIT(wait);
2970 
2971 	WARN_ON_ONCE(sqd && sqd->thread != current);
2972 
2973 	if (!current->io_uring)
2974 		return;
2975 	if (tctx->io_wq)
2976 		io_wq_exit_start(tctx->io_wq);
2977 
2978 	atomic_inc(&tctx->in_idle);
2979 	do {
2980 		bool loop = false;
2981 
2982 		io_uring_drop_tctx_refs(current);
2983 		/* read completions before cancelations */
2984 		inflight = tctx_inflight(tctx, !cancel_all);
2985 		if (!inflight)
2986 			break;
2987 
2988 		if (!sqd) {
2989 			struct io_tctx_node *node;
2990 			unsigned long index;
2991 
2992 			xa_for_each(&tctx->xa, index, node) {
2993 				/* sqpoll task will cancel all its requests */
2994 				if (node->ctx->sq_data)
2995 					continue;
2996 				loop |= io_uring_try_cancel_requests(node->ctx,
2997 							current, cancel_all);
2998 			}
2999 		} else {
3000 			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3001 				loop |= io_uring_try_cancel_requests(ctx,
3002 								     current,
3003 								     cancel_all);
3004 		}
3005 
3006 		if (loop) {
3007 			cond_resched();
3008 			continue;
3009 		}
3010 
3011 		prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3012 		io_run_task_work();
3013 		io_uring_drop_tctx_refs(current);
3014 
3015 		/*
3016 		 * If we've seen completions, retry without waiting. This
3017 		 * avoids a race where a completion comes in before we did
3018 		 * prepare_to_wait().
3019 		 */
3020 		if (inflight == tctx_inflight(tctx, !cancel_all))
3021 			schedule();
3022 		finish_wait(&tctx->wait, &wait);
3023 	} while (1);
3024 
3025 	io_uring_clean_tctx(tctx);
3026 	if (cancel_all) {
3027 		/*
3028 		 * We shouldn't run task_works after cancel, so just leave
3029 		 * ->in_idle set for normal exit.
3030 		 */
3031 		atomic_dec(&tctx->in_idle);
3032 		/* for exec all current's requests should be gone, kill tctx */
3033 		__io_uring_free(current);
3034 	}
3035 }
3036 
3037 void __io_uring_cancel(bool cancel_all)
3038 {
3039 	io_uring_cancel_generic(cancel_all, NULL);
3040 }
3041 
3042 static void *io_uring_validate_mmap_request(struct file *file,
3043 					    loff_t pgoff, size_t sz)
3044 {
3045 	struct io_ring_ctx *ctx = file->private_data;
3046 	loff_t offset = pgoff << PAGE_SHIFT;
3047 	struct page *page;
3048 	void *ptr;
3049 
3050 	switch (offset) {
3051 	case IORING_OFF_SQ_RING:
3052 	case IORING_OFF_CQ_RING:
3053 		ptr = ctx->rings;
3054 		break;
3055 	case IORING_OFF_SQES:
3056 		ptr = ctx->sq_sqes;
3057 		break;
3058 	default:
3059 		return ERR_PTR(-EINVAL);
3060 	}
3061 
3062 	page = virt_to_head_page(ptr);
3063 	if (sz > page_size(page))
3064 		return ERR_PTR(-EINVAL);
3065 
3066 	return ptr;
3067 }
3068 
3069 #ifdef CONFIG_MMU
3070 
3071 static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3072 {
3073 	size_t sz = vma->vm_end - vma->vm_start;
3074 	unsigned long pfn;
3075 	void *ptr;
3076 
3077 	ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3078 	if (IS_ERR(ptr))
3079 		return PTR_ERR(ptr);
3080 
3081 	pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3082 	return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3083 }
3084 
3085 #else /* !CONFIG_MMU */
3086 
3087 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3088 {
3089 	return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
3090 }
3091 
3092 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3093 {
3094 	return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3095 }
3096 
3097 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3098 	unsigned long addr, unsigned long len,
3099 	unsigned long pgoff, unsigned long flags)
3100 {
3101 	void *ptr;
3102 
3103 	ptr = io_uring_validate_mmap_request(file, pgoff, len);
3104 	if (IS_ERR(ptr))
3105 		return PTR_ERR(ptr);
3106 
3107 	return (unsigned long) ptr;
3108 }
3109 
3110 #endif /* !CONFIG_MMU */
3111 
3112 static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3113 {
3114 	if (flags & IORING_ENTER_EXT_ARG) {
3115 		struct io_uring_getevents_arg arg;
3116 
3117 		if (argsz != sizeof(arg))
3118 			return -EINVAL;
3119 		if (copy_from_user(&arg, argp, sizeof(arg)))
3120 			return -EFAULT;
3121 	}
3122 	return 0;
3123 }
3124 
3125 static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3126 			  struct __kernel_timespec __user **ts,
3127 			  const sigset_t __user **sig)
3128 {
3129 	struct io_uring_getevents_arg arg;
3130 
3131 	/*
3132 	 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3133 	 * is just a pointer to the sigset_t.
3134 	 */
3135 	if (!(flags & IORING_ENTER_EXT_ARG)) {
3136 		*sig = (const sigset_t __user *) argp;
3137 		*ts = NULL;
3138 		return 0;
3139 	}
3140 
3141 	/*
3142 	 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3143 	 * timespec and sigset_t pointers if good.
3144 	 */
3145 	if (*argsz != sizeof(arg))
3146 		return -EINVAL;
3147 	if (copy_from_user(&arg, argp, sizeof(arg)))
3148 		return -EFAULT;
3149 	if (arg.pad)
3150 		return -EINVAL;
3151 	*sig = u64_to_user_ptr(arg.sigmask);
3152 	*argsz = arg.sigmask_sz;
3153 	*ts = u64_to_user_ptr(arg.ts);
3154 	return 0;
3155 }
3156 
3157 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3158 		u32, min_complete, u32, flags, const void __user *, argp,
3159 		size_t, argsz)
3160 {
3161 	struct io_ring_ctx *ctx;
3162 	struct fd f;
3163 	long ret;
3164 
3165 	if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3166 			       IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3167 			       IORING_ENTER_REGISTERED_RING)))
3168 		return -EINVAL;
3169 
3170 	/*
3171 	 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3172 	 * need only dereference our task private array to find it.
3173 	 */
3174 	if (flags & IORING_ENTER_REGISTERED_RING) {
3175 		struct io_uring_task *tctx = current->io_uring;
3176 
3177 		if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3178 			return -EINVAL;
3179 		fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3180 		f.file = tctx->registered_rings[fd];
3181 		f.flags = 0;
3182 		if (unlikely(!f.file))
3183 			return -EBADF;
3184 	} else {
3185 		f = fdget(fd);
3186 		if (unlikely(!f.file))
3187 			return -EBADF;
3188 		ret = -EOPNOTSUPP;
3189 		if (unlikely(!io_is_uring_fops(f.file)))
3190 			goto out;
3191 	}
3192 
3193 	ctx = f.file->private_data;
3194 	ret = -EBADFD;
3195 	if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3196 		goto out;
3197 
3198 	/*
3199 	 * For SQ polling, the thread will do all submissions and completions.
3200 	 * Just return the requested submit count, and wake the thread if
3201 	 * we were asked to.
3202 	 */
3203 	ret = 0;
3204 	if (ctx->flags & IORING_SETUP_SQPOLL) {
3205 		io_cqring_overflow_flush(ctx);
3206 
3207 		if (unlikely(ctx->sq_data->thread == NULL)) {
3208 			ret = -EOWNERDEAD;
3209 			goto out;
3210 		}
3211 		if (flags & IORING_ENTER_SQ_WAKEUP)
3212 			wake_up(&ctx->sq_data->wait);
3213 		if (flags & IORING_ENTER_SQ_WAIT) {
3214 			ret = io_sqpoll_wait_sq(ctx);
3215 			if (ret)
3216 				goto out;
3217 		}
3218 		ret = to_submit;
3219 	} else if (to_submit) {
3220 		ret = io_uring_add_tctx_node(ctx);
3221 		if (unlikely(ret))
3222 			goto out;
3223 
3224 		mutex_lock(&ctx->uring_lock);
3225 		ret = io_submit_sqes(ctx, to_submit);
3226 		if (ret != to_submit) {
3227 			mutex_unlock(&ctx->uring_lock);
3228 			goto out;
3229 		}
3230 		if (flags & IORING_ENTER_GETEVENTS) {
3231 			if (ctx->syscall_iopoll)
3232 				goto iopoll_locked;
3233 			/*
3234 			 * Ignore errors, we'll soon call io_cqring_wait() and
3235 			 * it should handle ownership problems if any.
3236 			 */
3237 			if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3238 				(void)io_run_local_work_locked(ctx);
3239 		}
3240 		mutex_unlock(&ctx->uring_lock);
3241 	}
3242 
3243 	if (flags & IORING_ENTER_GETEVENTS) {
3244 		int ret2;
3245 
3246 		if (ctx->syscall_iopoll) {
3247 			/*
3248 			 * We disallow the app entering submit/complete with
3249 			 * polling, but we still need to lock the ring to
3250 			 * prevent racing with polled issue that got punted to
3251 			 * a workqueue.
3252 			 */
3253 			mutex_lock(&ctx->uring_lock);
3254 iopoll_locked:
3255 			ret2 = io_validate_ext_arg(flags, argp, argsz);
3256 			if (likely(!ret2)) {
3257 				min_complete = min(min_complete,
3258 						   ctx->cq_entries);
3259 				ret2 = io_iopoll_check(ctx, min_complete);
3260 			}
3261 			mutex_unlock(&ctx->uring_lock);
3262 		} else {
3263 			const sigset_t __user *sig;
3264 			struct __kernel_timespec __user *ts;
3265 
3266 			ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3267 			if (likely(!ret2)) {
3268 				min_complete = min(min_complete,
3269 						   ctx->cq_entries);
3270 				ret2 = io_cqring_wait(ctx, min_complete, sig,
3271 						      argsz, ts);
3272 			}
3273 		}
3274 
3275 		if (!ret) {
3276 			ret = ret2;
3277 
3278 			/*
3279 			 * EBADR indicates that one or more CQE were dropped.
3280 			 * Once the user has been informed we can clear the bit
3281 			 * as they are obviously ok with those drops.
3282 			 */
3283 			if (unlikely(ret2 == -EBADR))
3284 				clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3285 					  &ctx->check_cq);
3286 		}
3287 	}
3288 out:
3289 	fdput(f);
3290 	return ret;
3291 }
3292 
3293 static const struct file_operations io_uring_fops = {
3294 	.release	= io_uring_release,
3295 	.mmap		= io_uring_mmap,
3296 #ifndef CONFIG_MMU
3297 	.get_unmapped_area = io_uring_nommu_get_unmapped_area,
3298 	.mmap_capabilities = io_uring_nommu_mmap_capabilities,
3299 #endif
3300 	.poll		= io_uring_poll,
3301 #ifdef CONFIG_PROC_FS
3302 	.show_fdinfo	= io_uring_show_fdinfo,
3303 #endif
3304 };
3305 
3306 bool io_is_uring_fops(struct file *file)
3307 {
3308 	return file->f_op == &io_uring_fops;
3309 }
3310 
3311 static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3312 					 struct io_uring_params *p)
3313 {
3314 	struct io_rings *rings;
3315 	size_t size, sq_array_offset;
3316 
3317 	/* make sure these are sane, as we already accounted them */
3318 	ctx->sq_entries = p->sq_entries;
3319 	ctx->cq_entries = p->cq_entries;
3320 
3321 	size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3322 	if (size == SIZE_MAX)
3323 		return -EOVERFLOW;
3324 
3325 	rings = io_mem_alloc(size);
3326 	if (!rings)
3327 		return -ENOMEM;
3328 
3329 	ctx->rings = rings;
3330 	ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3331 	rings->sq_ring_mask = p->sq_entries - 1;
3332 	rings->cq_ring_mask = p->cq_entries - 1;
3333 	rings->sq_ring_entries = p->sq_entries;
3334 	rings->cq_ring_entries = p->cq_entries;
3335 
3336 	if (p->flags & IORING_SETUP_SQE128)
3337 		size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3338 	else
3339 		size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3340 	if (size == SIZE_MAX) {
3341 		io_mem_free(ctx->rings);
3342 		ctx->rings = NULL;
3343 		return -EOVERFLOW;
3344 	}
3345 
3346 	ctx->sq_sqes = io_mem_alloc(size);
3347 	if (!ctx->sq_sqes) {
3348 		io_mem_free(ctx->rings);
3349 		ctx->rings = NULL;
3350 		return -ENOMEM;
3351 	}
3352 
3353 	return 0;
3354 }
3355 
3356 static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3357 {
3358 	int ret, fd;
3359 
3360 	fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3361 	if (fd < 0)
3362 		return fd;
3363 
3364 	ret = __io_uring_add_tctx_node(ctx);
3365 	if (ret) {
3366 		put_unused_fd(fd);
3367 		return ret;
3368 	}
3369 	fd_install(fd, file);
3370 	return fd;
3371 }
3372 
3373 /*
3374  * Allocate an anonymous fd, this is what constitutes the application
3375  * visible backing of an io_uring instance. The application mmaps this
3376  * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3377  * we have to tie this fd to a socket for file garbage collection purposes.
3378  */
3379 static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3380 {
3381 	struct file *file;
3382 #if defined(CONFIG_UNIX)
3383 	int ret;
3384 
3385 	ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3386 				&ctx->ring_sock);
3387 	if (ret)
3388 		return ERR_PTR(ret);
3389 #endif
3390 
3391 	file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3392 					 O_RDWR | O_CLOEXEC, NULL);
3393 #if defined(CONFIG_UNIX)
3394 	if (IS_ERR(file)) {
3395 		sock_release(ctx->ring_sock);
3396 		ctx->ring_sock = NULL;
3397 	} else {
3398 		ctx->ring_sock->file = file;
3399 	}
3400 #endif
3401 	return file;
3402 }
3403 
3404 static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3405 				  struct io_uring_params __user *params)
3406 {
3407 	struct io_ring_ctx *ctx;
3408 	struct file *file;
3409 	int ret;
3410 
3411 	if (!entries)
3412 		return -EINVAL;
3413 	if (entries > IORING_MAX_ENTRIES) {
3414 		if (!(p->flags & IORING_SETUP_CLAMP))
3415 			return -EINVAL;
3416 		entries = IORING_MAX_ENTRIES;
3417 	}
3418 
3419 	/*
3420 	 * Use twice as many entries for the CQ ring. It's possible for the
3421 	 * application to drive a higher depth than the size of the SQ ring,
3422 	 * since the sqes are only used at submission time. This allows for
3423 	 * some flexibility in overcommitting a bit. If the application has
3424 	 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3425 	 * of CQ ring entries manually.
3426 	 */
3427 	p->sq_entries = roundup_pow_of_two(entries);
3428 	if (p->flags & IORING_SETUP_CQSIZE) {
3429 		/*
3430 		 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3431 		 * to a power-of-two, if it isn't already. We do NOT impose
3432 		 * any cq vs sq ring sizing.
3433 		 */
3434 		if (!p->cq_entries)
3435 			return -EINVAL;
3436 		if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3437 			if (!(p->flags & IORING_SETUP_CLAMP))
3438 				return -EINVAL;
3439 			p->cq_entries = IORING_MAX_CQ_ENTRIES;
3440 		}
3441 		p->cq_entries = roundup_pow_of_two(p->cq_entries);
3442 		if (p->cq_entries < p->sq_entries)
3443 			return -EINVAL;
3444 	} else {
3445 		p->cq_entries = 2 * p->sq_entries;
3446 	}
3447 
3448 	ctx = io_ring_ctx_alloc(p);
3449 	if (!ctx)
3450 		return -ENOMEM;
3451 
3452 	/*
3453 	 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3454 	 * space applications don't need to do io completion events
3455 	 * polling again, they can rely on io_sq_thread to do polling
3456 	 * work, which can reduce cpu usage and uring_lock contention.
3457 	 */
3458 	if (ctx->flags & IORING_SETUP_IOPOLL &&
3459 	    !(ctx->flags & IORING_SETUP_SQPOLL))
3460 		ctx->syscall_iopoll = 1;
3461 
3462 	ctx->compat = in_compat_syscall();
3463 	if (!capable(CAP_IPC_LOCK))
3464 		ctx->user = get_uid(current_user());
3465 
3466 	/*
3467 	 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3468 	 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3469 	 */
3470 	ret = -EINVAL;
3471 	if (ctx->flags & IORING_SETUP_SQPOLL) {
3472 		/* IPI related flags don't make sense with SQPOLL */
3473 		if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3474 				  IORING_SETUP_TASKRUN_FLAG |
3475 				  IORING_SETUP_DEFER_TASKRUN))
3476 			goto err;
3477 		ctx->notify_method = TWA_SIGNAL_NO_IPI;
3478 	} else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3479 		ctx->notify_method = TWA_SIGNAL_NO_IPI;
3480 	} else {
3481 		if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3482 		    !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3483 			goto err;
3484 		ctx->notify_method = TWA_SIGNAL;
3485 	}
3486 
3487 	/*
3488 	 * For DEFER_TASKRUN we require the completion task to be the same as the
3489 	 * submission task. This implies that there is only one submitter, so enforce
3490 	 * that.
3491 	 */
3492 	if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3493 	    !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3494 		goto err;
3495 	}
3496 
3497 	/*
3498 	 * This is just grabbed for accounting purposes. When a process exits,
3499 	 * the mm is exited and dropped before the files, hence we need to hang
3500 	 * on to this mm purely for the purposes of being able to unaccount
3501 	 * memory (locked/pinned vm). It's not used for anything else.
3502 	 */
3503 	mmgrab(current->mm);
3504 	ctx->mm_account = current->mm;
3505 
3506 	ret = io_allocate_scq_urings(ctx, p);
3507 	if (ret)
3508 		goto err;
3509 
3510 	ret = io_sq_offload_create(ctx, p);
3511 	if (ret)
3512 		goto err;
3513 	/* always set a rsrc node */
3514 	ret = io_rsrc_node_switch_start(ctx);
3515 	if (ret)
3516 		goto err;
3517 	io_rsrc_node_switch(ctx, NULL);
3518 
3519 	memset(&p->sq_off, 0, sizeof(p->sq_off));
3520 	p->sq_off.head = offsetof(struct io_rings, sq.head);
3521 	p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3522 	p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3523 	p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3524 	p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3525 	p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3526 	p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3527 
3528 	memset(&p->cq_off, 0, sizeof(p->cq_off));
3529 	p->cq_off.head = offsetof(struct io_rings, cq.head);
3530 	p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3531 	p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3532 	p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3533 	p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3534 	p->cq_off.cqes = offsetof(struct io_rings, cqes);
3535 	p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3536 
3537 	p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3538 			IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3539 			IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3540 			IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3541 			IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3542 			IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3543 			IORING_FEAT_LINKED_FILE;
3544 
3545 	if (copy_to_user(params, p, sizeof(*p))) {
3546 		ret = -EFAULT;
3547 		goto err;
3548 	}
3549 
3550 	if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3551 	    && !(ctx->flags & IORING_SETUP_R_DISABLED))
3552 		ctx->submitter_task = get_task_struct(current);
3553 
3554 	file = io_uring_get_file(ctx);
3555 	if (IS_ERR(file)) {
3556 		ret = PTR_ERR(file);
3557 		goto err;
3558 	}
3559 
3560 	/*
3561 	 * Install ring fd as the very last thing, so we don't risk someone
3562 	 * having closed it before we finish setup
3563 	 */
3564 	ret = io_uring_install_fd(ctx, file);
3565 	if (ret < 0) {
3566 		/* fput will clean it up */
3567 		fput(file);
3568 		return ret;
3569 	}
3570 
3571 	trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3572 	return ret;
3573 err:
3574 	io_ring_ctx_wait_and_kill(ctx);
3575 	return ret;
3576 }
3577 
3578 /*
3579  * Sets up an aio uring context, and returns the fd. Applications asks for a
3580  * ring size, we return the actual sq/cq ring sizes (among other things) in the
3581  * params structure passed in.
3582  */
3583 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3584 {
3585 	struct io_uring_params p;
3586 	int i;
3587 
3588 	if (copy_from_user(&p, params, sizeof(p)))
3589 		return -EFAULT;
3590 	for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3591 		if (p.resv[i])
3592 			return -EINVAL;
3593 	}
3594 
3595 	if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3596 			IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3597 			IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3598 			IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3599 			IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3600 			IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3601 			IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3602 		return -EINVAL;
3603 
3604 	return io_uring_create(entries, &p, params);
3605 }
3606 
3607 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3608 		struct io_uring_params __user *, params)
3609 {
3610 	return io_uring_setup(entries, params);
3611 }
3612 
3613 static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3614 			   unsigned nr_args)
3615 {
3616 	struct io_uring_probe *p;
3617 	size_t size;
3618 	int i, ret;
3619 
3620 	size = struct_size(p, ops, nr_args);
3621 	if (size == SIZE_MAX)
3622 		return -EOVERFLOW;
3623 	p = kzalloc(size, GFP_KERNEL);
3624 	if (!p)
3625 		return -ENOMEM;
3626 
3627 	ret = -EFAULT;
3628 	if (copy_from_user(p, arg, size))
3629 		goto out;
3630 	ret = -EINVAL;
3631 	if (memchr_inv(p, 0, size))
3632 		goto out;
3633 
3634 	p->last_op = IORING_OP_LAST - 1;
3635 	if (nr_args > IORING_OP_LAST)
3636 		nr_args = IORING_OP_LAST;
3637 
3638 	for (i = 0; i < nr_args; i++) {
3639 		p->ops[i].op = i;
3640 		if (!io_op_defs[i].not_supported)
3641 			p->ops[i].flags = IO_URING_OP_SUPPORTED;
3642 	}
3643 	p->ops_len = i;
3644 
3645 	ret = 0;
3646 	if (copy_to_user(arg, p, size))
3647 		ret = -EFAULT;
3648 out:
3649 	kfree(p);
3650 	return ret;
3651 }
3652 
3653 static int io_register_personality(struct io_ring_ctx *ctx)
3654 {
3655 	const struct cred *creds;
3656 	u32 id;
3657 	int ret;
3658 
3659 	creds = get_current_cred();
3660 
3661 	ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3662 			XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3663 	if (ret < 0) {
3664 		put_cred(creds);
3665 		return ret;
3666 	}
3667 	return id;
3668 }
3669 
3670 static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3671 					   void __user *arg, unsigned int nr_args)
3672 {
3673 	struct io_uring_restriction *res;
3674 	size_t size;
3675 	int i, ret;
3676 
3677 	/* Restrictions allowed only if rings started disabled */
3678 	if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3679 		return -EBADFD;
3680 
3681 	/* We allow only a single restrictions registration */
3682 	if (ctx->restrictions.registered)
3683 		return -EBUSY;
3684 
3685 	if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3686 		return -EINVAL;
3687 
3688 	size = array_size(nr_args, sizeof(*res));
3689 	if (size == SIZE_MAX)
3690 		return -EOVERFLOW;
3691 
3692 	res = memdup_user(arg, size);
3693 	if (IS_ERR(res))
3694 		return PTR_ERR(res);
3695 
3696 	ret = 0;
3697 
3698 	for (i = 0; i < nr_args; i++) {
3699 		switch (res[i].opcode) {
3700 		case IORING_RESTRICTION_REGISTER_OP:
3701 			if (res[i].register_op >= IORING_REGISTER_LAST) {
3702 				ret = -EINVAL;
3703 				goto out;
3704 			}
3705 
3706 			__set_bit(res[i].register_op,
3707 				  ctx->restrictions.register_op);
3708 			break;
3709 		case IORING_RESTRICTION_SQE_OP:
3710 			if (res[i].sqe_op >= IORING_OP_LAST) {
3711 				ret = -EINVAL;
3712 				goto out;
3713 			}
3714 
3715 			__set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3716 			break;
3717 		case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3718 			ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3719 			break;
3720 		case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3721 			ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3722 			break;
3723 		default:
3724 			ret = -EINVAL;
3725 			goto out;
3726 		}
3727 	}
3728 
3729 out:
3730 	/* Reset all restrictions if an error happened */
3731 	if (ret != 0)
3732 		memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3733 	else
3734 		ctx->restrictions.registered = true;
3735 
3736 	kfree(res);
3737 	return ret;
3738 }
3739 
3740 static int io_register_enable_rings(struct io_ring_ctx *ctx)
3741 {
3742 	if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3743 		return -EBADFD;
3744 
3745 	if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task)
3746 		ctx->submitter_task = get_task_struct(current);
3747 
3748 	if (ctx->restrictions.registered)
3749 		ctx->restricted = 1;
3750 
3751 	ctx->flags &= ~IORING_SETUP_R_DISABLED;
3752 	if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3753 		wake_up(&ctx->sq_data->wait);
3754 	return 0;
3755 }
3756 
3757 static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3758 				       void __user *arg, unsigned len)
3759 {
3760 	struct io_uring_task *tctx = current->io_uring;
3761 	cpumask_var_t new_mask;
3762 	int ret;
3763 
3764 	if (!tctx || !tctx->io_wq)
3765 		return -EINVAL;
3766 
3767 	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3768 		return -ENOMEM;
3769 
3770 	cpumask_clear(new_mask);
3771 	if (len > cpumask_size())
3772 		len = cpumask_size();
3773 
3774 	if (in_compat_syscall()) {
3775 		ret = compat_get_bitmap(cpumask_bits(new_mask),
3776 					(const compat_ulong_t __user *)arg,
3777 					len * 8 /* CHAR_BIT */);
3778 	} else {
3779 		ret = copy_from_user(new_mask, arg, len);
3780 	}
3781 
3782 	if (ret) {
3783 		free_cpumask_var(new_mask);
3784 		return -EFAULT;
3785 	}
3786 
3787 	ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3788 	free_cpumask_var(new_mask);
3789 	return ret;
3790 }
3791 
3792 static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3793 {
3794 	struct io_uring_task *tctx = current->io_uring;
3795 
3796 	if (!tctx || !tctx->io_wq)
3797 		return -EINVAL;
3798 
3799 	return io_wq_cpu_affinity(tctx->io_wq, NULL);
3800 }
3801 
3802 static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3803 					       void __user *arg)
3804 	__must_hold(&ctx->uring_lock)
3805 {
3806 	struct io_tctx_node *node;
3807 	struct io_uring_task *tctx = NULL;
3808 	struct io_sq_data *sqd = NULL;
3809 	__u32 new_count[2];
3810 	int i, ret;
3811 
3812 	if (copy_from_user(new_count, arg, sizeof(new_count)))
3813 		return -EFAULT;
3814 	for (i = 0; i < ARRAY_SIZE(new_count); i++)
3815 		if (new_count[i] > INT_MAX)
3816 			return -EINVAL;
3817 
3818 	if (ctx->flags & IORING_SETUP_SQPOLL) {
3819 		sqd = ctx->sq_data;
3820 		if (sqd) {
3821 			/*
3822 			 * Observe the correct sqd->lock -> ctx->uring_lock
3823 			 * ordering. Fine to drop uring_lock here, we hold
3824 			 * a ref to the ctx.
3825 			 */
3826 			refcount_inc(&sqd->refs);
3827 			mutex_unlock(&ctx->uring_lock);
3828 			mutex_lock(&sqd->lock);
3829 			mutex_lock(&ctx->uring_lock);
3830 			if (sqd->thread)
3831 				tctx = sqd->thread->io_uring;
3832 		}
3833 	} else {
3834 		tctx = current->io_uring;
3835 	}
3836 
3837 	BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3838 
3839 	for (i = 0; i < ARRAY_SIZE(new_count); i++)
3840 		if (new_count[i])
3841 			ctx->iowq_limits[i] = new_count[i];
3842 	ctx->iowq_limits_set = true;
3843 
3844 	if (tctx && tctx->io_wq) {
3845 		ret = io_wq_max_workers(tctx->io_wq, new_count);
3846 		if (ret)
3847 			goto err;
3848 	} else {
3849 		memset(new_count, 0, sizeof(new_count));
3850 	}
3851 
3852 	if (sqd) {
3853 		mutex_unlock(&sqd->lock);
3854 		io_put_sq_data(sqd);
3855 	}
3856 
3857 	if (copy_to_user(arg, new_count, sizeof(new_count)))
3858 		return -EFAULT;
3859 
3860 	/* that's it for SQPOLL, only the SQPOLL task creates requests */
3861 	if (sqd)
3862 		return 0;
3863 
3864 	/* now propagate the restriction to all registered users */
3865 	list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3866 		struct io_uring_task *tctx = node->task->io_uring;
3867 
3868 		if (WARN_ON_ONCE(!tctx->io_wq))
3869 			continue;
3870 
3871 		for (i = 0; i < ARRAY_SIZE(new_count); i++)
3872 			new_count[i] = ctx->iowq_limits[i];
3873 		/* ignore errors, it always returns zero anyway */
3874 		(void)io_wq_max_workers(tctx->io_wq, new_count);
3875 	}
3876 	return 0;
3877 err:
3878 	if (sqd) {
3879 		mutex_unlock(&sqd->lock);
3880 		io_put_sq_data(sqd);
3881 	}
3882 	return ret;
3883 }
3884 
3885 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3886 			       void __user *arg, unsigned nr_args)
3887 	__releases(ctx->uring_lock)
3888 	__acquires(ctx->uring_lock)
3889 {
3890 	int ret;
3891 
3892 	/*
3893 	 * We don't quiesce the refs for register anymore and so it can't be
3894 	 * dying as we're holding a file ref here.
3895 	 */
3896 	if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
3897 		return -ENXIO;
3898 
3899 	if (ctx->submitter_task && ctx->submitter_task != current)
3900 		return -EEXIST;
3901 
3902 	if (ctx->restricted) {
3903 		if (opcode >= IORING_REGISTER_LAST)
3904 			return -EINVAL;
3905 		opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
3906 		if (!test_bit(opcode, ctx->restrictions.register_op))
3907 			return -EACCES;
3908 	}
3909 
3910 	switch (opcode) {
3911 	case IORING_REGISTER_BUFFERS:
3912 		ret = -EFAULT;
3913 		if (!arg)
3914 			break;
3915 		ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
3916 		break;
3917 	case IORING_UNREGISTER_BUFFERS:
3918 		ret = -EINVAL;
3919 		if (arg || nr_args)
3920 			break;
3921 		ret = io_sqe_buffers_unregister(ctx);
3922 		break;
3923 	case IORING_REGISTER_FILES:
3924 		ret = -EFAULT;
3925 		if (!arg)
3926 			break;
3927 		ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
3928 		break;
3929 	case IORING_UNREGISTER_FILES:
3930 		ret = -EINVAL;
3931 		if (arg || nr_args)
3932 			break;
3933 		ret = io_sqe_files_unregister(ctx);
3934 		break;
3935 	case IORING_REGISTER_FILES_UPDATE:
3936 		ret = io_register_files_update(ctx, arg, nr_args);
3937 		break;
3938 	case IORING_REGISTER_EVENTFD:
3939 		ret = -EINVAL;
3940 		if (nr_args != 1)
3941 			break;
3942 		ret = io_eventfd_register(ctx, arg, 0);
3943 		break;
3944 	case IORING_REGISTER_EVENTFD_ASYNC:
3945 		ret = -EINVAL;
3946 		if (nr_args != 1)
3947 			break;
3948 		ret = io_eventfd_register(ctx, arg, 1);
3949 		break;
3950 	case IORING_UNREGISTER_EVENTFD:
3951 		ret = -EINVAL;
3952 		if (arg || nr_args)
3953 			break;
3954 		ret = io_eventfd_unregister(ctx);
3955 		break;
3956 	case IORING_REGISTER_PROBE:
3957 		ret = -EINVAL;
3958 		if (!arg || nr_args > 256)
3959 			break;
3960 		ret = io_probe(ctx, arg, nr_args);
3961 		break;
3962 	case IORING_REGISTER_PERSONALITY:
3963 		ret = -EINVAL;
3964 		if (arg || nr_args)
3965 			break;
3966 		ret = io_register_personality(ctx);
3967 		break;
3968 	case IORING_UNREGISTER_PERSONALITY:
3969 		ret = -EINVAL;
3970 		if (arg)
3971 			break;
3972 		ret = io_unregister_personality(ctx, nr_args);
3973 		break;
3974 	case IORING_REGISTER_ENABLE_RINGS:
3975 		ret = -EINVAL;
3976 		if (arg || nr_args)
3977 			break;
3978 		ret = io_register_enable_rings(ctx);
3979 		break;
3980 	case IORING_REGISTER_RESTRICTIONS:
3981 		ret = io_register_restrictions(ctx, arg, nr_args);
3982 		break;
3983 	case IORING_REGISTER_FILES2:
3984 		ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
3985 		break;
3986 	case IORING_REGISTER_FILES_UPDATE2:
3987 		ret = io_register_rsrc_update(ctx, arg, nr_args,
3988 					      IORING_RSRC_FILE);
3989 		break;
3990 	case IORING_REGISTER_BUFFERS2:
3991 		ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
3992 		break;
3993 	case IORING_REGISTER_BUFFERS_UPDATE:
3994 		ret = io_register_rsrc_update(ctx, arg, nr_args,
3995 					      IORING_RSRC_BUFFER);
3996 		break;
3997 	case IORING_REGISTER_IOWQ_AFF:
3998 		ret = -EINVAL;
3999 		if (!arg || !nr_args)
4000 			break;
4001 		ret = io_register_iowq_aff(ctx, arg, nr_args);
4002 		break;
4003 	case IORING_UNREGISTER_IOWQ_AFF:
4004 		ret = -EINVAL;
4005 		if (arg || nr_args)
4006 			break;
4007 		ret = io_unregister_iowq_aff(ctx);
4008 		break;
4009 	case IORING_REGISTER_IOWQ_MAX_WORKERS:
4010 		ret = -EINVAL;
4011 		if (!arg || nr_args != 2)
4012 			break;
4013 		ret = io_register_iowq_max_workers(ctx, arg);
4014 		break;
4015 	case IORING_REGISTER_RING_FDS:
4016 		ret = io_ringfd_register(ctx, arg, nr_args);
4017 		break;
4018 	case IORING_UNREGISTER_RING_FDS:
4019 		ret = io_ringfd_unregister(ctx, arg, nr_args);
4020 		break;
4021 	case IORING_REGISTER_PBUF_RING:
4022 		ret = -EINVAL;
4023 		if (!arg || nr_args != 1)
4024 			break;
4025 		ret = io_register_pbuf_ring(ctx, arg);
4026 		break;
4027 	case IORING_UNREGISTER_PBUF_RING:
4028 		ret = -EINVAL;
4029 		if (!arg || nr_args != 1)
4030 			break;
4031 		ret = io_unregister_pbuf_ring(ctx, arg);
4032 		break;
4033 	case IORING_REGISTER_SYNC_CANCEL:
4034 		ret = -EINVAL;
4035 		if (!arg || nr_args != 1)
4036 			break;
4037 		ret = io_sync_cancel(ctx, arg);
4038 		break;
4039 	case IORING_REGISTER_FILE_ALLOC_RANGE:
4040 		ret = -EINVAL;
4041 		if (!arg || nr_args)
4042 			break;
4043 		ret = io_register_file_alloc_range(ctx, arg);
4044 		break;
4045 	default:
4046 		ret = -EINVAL;
4047 		break;
4048 	}
4049 
4050 	return ret;
4051 }
4052 
4053 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4054 		void __user *, arg, unsigned int, nr_args)
4055 {
4056 	struct io_ring_ctx *ctx;
4057 	long ret = -EBADF;
4058 	struct fd f;
4059 
4060 	f = fdget(fd);
4061 	if (!f.file)
4062 		return -EBADF;
4063 
4064 	ret = -EOPNOTSUPP;
4065 	if (!io_is_uring_fops(f.file))
4066 		goto out_fput;
4067 
4068 	ctx = f.file->private_data;
4069 
4070 	io_run_task_work_ctx(ctx);
4071 
4072 	mutex_lock(&ctx->uring_lock);
4073 	ret = __io_uring_register(ctx, opcode, arg, nr_args);
4074 	mutex_unlock(&ctx->uring_lock);
4075 	trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4076 out_fput:
4077 	fdput(f);
4078 	return ret;
4079 }
4080 
4081 static int __init io_uring_init(void)
4082 {
4083 #define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4084 	BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4085 	BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4086 } while (0)
4087 
4088 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4089 	__BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4090 #define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4091 	__BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4092 	BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4093 	BUILD_BUG_SQE_ELEM(0,  __u8,   opcode);
4094 	BUILD_BUG_SQE_ELEM(1,  __u8,   flags);
4095 	BUILD_BUG_SQE_ELEM(2,  __u16,  ioprio);
4096 	BUILD_BUG_SQE_ELEM(4,  __s32,  fd);
4097 	BUILD_BUG_SQE_ELEM(8,  __u64,  off);
4098 	BUILD_BUG_SQE_ELEM(8,  __u64,  addr2);
4099 	BUILD_BUG_SQE_ELEM(8,  __u32,  cmd_op);
4100 	BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4101 	BUILD_BUG_SQE_ELEM(16, __u64,  addr);
4102 	BUILD_BUG_SQE_ELEM(16, __u64,  splice_off_in);
4103 	BUILD_BUG_SQE_ELEM(24, __u32,  len);
4104 	BUILD_BUG_SQE_ELEM(28,     __kernel_rwf_t, rw_flags);
4105 	BUILD_BUG_SQE_ELEM(28, /* compat */   int, rw_flags);
4106 	BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4107 	BUILD_BUG_SQE_ELEM(28, __u32,  fsync_flags);
4108 	BUILD_BUG_SQE_ELEM(28, /* compat */ __u16,  poll_events);
4109 	BUILD_BUG_SQE_ELEM(28, __u32,  poll32_events);
4110 	BUILD_BUG_SQE_ELEM(28, __u32,  sync_range_flags);
4111 	BUILD_BUG_SQE_ELEM(28, __u32,  msg_flags);
4112 	BUILD_BUG_SQE_ELEM(28, __u32,  timeout_flags);
4113 	BUILD_BUG_SQE_ELEM(28, __u32,  accept_flags);
4114 	BUILD_BUG_SQE_ELEM(28, __u32,  cancel_flags);
4115 	BUILD_BUG_SQE_ELEM(28, __u32,  open_flags);
4116 	BUILD_BUG_SQE_ELEM(28, __u32,  statx_flags);
4117 	BUILD_BUG_SQE_ELEM(28, __u32,  fadvise_advice);
4118 	BUILD_BUG_SQE_ELEM(28, __u32,  splice_flags);
4119 	BUILD_BUG_SQE_ELEM(28, __u32,  rename_flags);
4120 	BUILD_BUG_SQE_ELEM(28, __u32,  unlink_flags);
4121 	BUILD_BUG_SQE_ELEM(28, __u32,  hardlink_flags);
4122 	BUILD_BUG_SQE_ELEM(28, __u32,  xattr_flags);
4123 	BUILD_BUG_SQE_ELEM(28, __u32,  msg_ring_flags);
4124 	BUILD_BUG_SQE_ELEM(32, __u64,  user_data);
4125 	BUILD_BUG_SQE_ELEM(40, __u16,  buf_index);
4126 	BUILD_BUG_SQE_ELEM(40, __u16,  buf_group);
4127 	BUILD_BUG_SQE_ELEM(42, __u16,  personality);
4128 	BUILD_BUG_SQE_ELEM(44, __s32,  splice_fd_in);
4129 	BUILD_BUG_SQE_ELEM(44, __u32,  file_index);
4130 	BUILD_BUG_SQE_ELEM(44, __u16,  addr_len);
4131 	BUILD_BUG_SQE_ELEM(46, __u16,  __pad3[0]);
4132 	BUILD_BUG_SQE_ELEM(48, __u64,  addr3);
4133 	BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4134 	BUILD_BUG_SQE_ELEM(56, __u64,  __pad2);
4135 
4136 	BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4137 		     sizeof(struct io_uring_rsrc_update));
4138 	BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4139 		     sizeof(struct io_uring_rsrc_update2));
4140 
4141 	/* ->buf_index is u16 */
4142 	BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4143 	BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4144 		     offsetof(struct io_uring_buf_ring, tail));
4145 
4146 	/* should fit into one byte */
4147 	BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4148 	BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4149 	BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4150 
4151 	BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4152 
4153 	BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4154 
4155 	io_uring_optable_init();
4156 
4157 	req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4158 				SLAB_ACCOUNT);
4159 	return 0;
4160 };
4161 __initcall(io_uring_init);
4162