Linux Audio

Check our new training course

Loading...
Note: File does not exist in v5.4.
   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
 123enum {
 124	IO_CHECK_CQ_OVERFLOW_BIT,
 125	IO_CHECK_CQ_DROPPED_BIT,
 126};
 127
 128enum {
 129	IO_EVENTFD_OP_SIGNAL_BIT,
 130	IO_EVENTFD_OP_FREE_BIT,
 131};
 132
 133struct 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
 143static bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
 144					 struct task_struct *task,
 145					 bool cancel_all);
 146
 147static void io_dismantle_req(struct io_kiocb *req);
 148static void io_clean_op(struct io_kiocb *req);
 149static void io_queue_sqe(struct io_kiocb *req);
 150static void io_move_task_work_from_local(struct io_ring_ctx *ctx);
 151static void __io_submit_flush_completions(struct io_ring_ctx *ctx);
 152static __cold void io_fallback_tw(struct io_uring_task *tctx);
 153
 154static struct kmem_cache *req_cachep;
 155
 156struct sock *io_uring_get_socket(struct file *file)
 157{
 158#if defined(CONFIG_UNIX)
 159	if (io_is_uring_fops(file)) {
 160		struct io_ring_ctx *ctx = file->private_data;
 161
 162		return ctx->ring_sock->sk;
 163	}
 164#endif
 165	return NULL;
 166}
 167EXPORT_SYMBOL(io_uring_get_socket);
 168
 169static inline void io_submit_flush_completions(struct io_ring_ctx *ctx)
 170{
 171	if (!wq_list_empty(&ctx->submit_state.compl_reqs) ||
 172	    ctx->submit_state.cqes_count)
 173		__io_submit_flush_completions(ctx);
 174}
 175
 176static inline unsigned int __io_cqring_events(struct io_ring_ctx *ctx)
 177{
 178	return ctx->cached_cq_tail - READ_ONCE(ctx->rings->cq.head);
 179}
 180
 181static inline unsigned int __io_cqring_events_user(struct io_ring_ctx *ctx)
 182{
 183	return READ_ONCE(ctx->rings->cq.tail) - READ_ONCE(ctx->rings->cq.head);
 184}
 185
 186static bool io_match_linked(struct io_kiocb *head)
 187{
 188	struct io_kiocb *req;
 189
 190	io_for_each_link(req, head) {
 191		if (req->flags & REQ_F_INFLIGHT)
 192			return true;
 193	}
 194	return false;
 195}
 196
 197/*
 198 * As io_match_task() but protected against racing with linked timeouts.
 199 * User must not hold timeout_lock.
 200 */
 201bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
 202			bool cancel_all)
 203{
 204	bool matched;
 205
 206	if (task && head->task != task)
 207		return false;
 208	if (cancel_all)
 209		return true;
 210
 211	if (head->flags & REQ_F_LINK_TIMEOUT) {
 212		struct io_ring_ctx *ctx = head->ctx;
 213
 214		/* protect against races with linked timeouts */
 215		spin_lock_irq(&ctx->timeout_lock);
 216		matched = io_match_linked(head);
 217		spin_unlock_irq(&ctx->timeout_lock);
 218	} else {
 219		matched = io_match_linked(head);
 220	}
 221	return matched;
 222}
 223
 224static inline void req_fail_link_node(struct io_kiocb *req, int res)
 225{
 226	req_set_fail(req);
 227	io_req_set_res(req, res, 0);
 228}
 229
 230static inline void io_req_add_to_cache(struct io_kiocb *req, struct io_ring_ctx *ctx)
 231{
 232	wq_stack_add_head(&req->comp_list, &ctx->submit_state.free_list);
 233}
 234
 235static __cold void io_ring_ctx_ref_free(struct percpu_ref *ref)
 236{
 237	struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
 238
 239	complete(&ctx->ref_comp);
 240}
 241
 242static __cold void io_fallback_req_func(struct work_struct *work)
 243{
 244	struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx,
 245						fallback_work.work);
 246	struct llist_node *node = llist_del_all(&ctx->fallback_llist);
 247	struct io_kiocb *req, *tmp;
 248	bool locked = false;
 249
 250	percpu_ref_get(&ctx->refs);
 251	llist_for_each_entry_safe(req, tmp, node, io_task_work.node)
 252		req->io_task_work.func(req, &locked);
 253
 254	if (locked) {
 255		io_submit_flush_completions(ctx);
 256		mutex_unlock(&ctx->uring_lock);
 257	}
 258	percpu_ref_put(&ctx->refs);
 259}
 260
 261static int io_alloc_hash_table(struct io_hash_table *table, unsigned bits)
 262{
 263	unsigned hash_buckets = 1U << bits;
 264	size_t hash_size = hash_buckets * sizeof(table->hbs[0]);
 265
 266	table->hbs = kmalloc(hash_size, GFP_KERNEL);
 267	if (!table->hbs)
 268		return -ENOMEM;
 269
 270	table->hash_bits = bits;
 271	init_hash_table(table, hash_buckets);
 272	return 0;
 273}
 274
 275static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 276{
 277	struct io_ring_ctx *ctx;
 278	int hash_bits;
 279
 280	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
 281	if (!ctx)
 282		return NULL;
 283
 284	xa_init(&ctx->io_bl_xa);
 285
 286	/*
 287	 * Use 5 bits less than the max cq entries, that should give us around
 288	 * 32 entries per hash list if totally full and uniformly spread, but
 289	 * don't keep too many buckets to not overconsume memory.
 290	 */
 291	hash_bits = ilog2(p->cq_entries) - 5;
 292	hash_bits = clamp(hash_bits, 1, 8);
 293	if (io_alloc_hash_table(&ctx->cancel_table, hash_bits))
 294		goto err;
 295	if (io_alloc_hash_table(&ctx->cancel_table_locked, hash_bits))
 296		goto err;
 297
 298	ctx->dummy_ubuf = kzalloc(sizeof(*ctx->dummy_ubuf), GFP_KERNEL);
 299	if (!ctx->dummy_ubuf)
 300		goto err;
 301	/* set invalid range, so io_import_fixed() fails meeting it */
 302	ctx->dummy_ubuf->ubuf = -1UL;
 303
 304	if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
 305			    0, GFP_KERNEL))
 306		goto err;
 307
 308	ctx->flags = p->flags;
 309	init_waitqueue_head(&ctx->sqo_sq_wait);
 310	INIT_LIST_HEAD(&ctx->sqd_list);
 311	INIT_LIST_HEAD(&ctx->cq_overflow_list);
 312	INIT_LIST_HEAD(&ctx->io_buffers_cache);
 313	io_alloc_cache_init(&ctx->apoll_cache);
 314	io_alloc_cache_init(&ctx->netmsg_cache);
 315	init_completion(&ctx->ref_comp);
 316	xa_init_flags(&ctx->personalities, XA_FLAGS_ALLOC1);
 317	mutex_init(&ctx->uring_lock);
 318	init_waitqueue_head(&ctx->cq_wait);
 319	spin_lock_init(&ctx->completion_lock);
 320	spin_lock_init(&ctx->timeout_lock);
 321	INIT_WQ_LIST(&ctx->iopoll_list);
 322	INIT_LIST_HEAD(&ctx->io_buffers_pages);
 323	INIT_LIST_HEAD(&ctx->io_buffers_comp);
 324	INIT_LIST_HEAD(&ctx->defer_list);
 325	INIT_LIST_HEAD(&ctx->timeout_list);
 326	INIT_LIST_HEAD(&ctx->ltimeout_list);
 327	spin_lock_init(&ctx->rsrc_ref_lock);
 328	INIT_LIST_HEAD(&ctx->rsrc_ref_list);
 329	INIT_DELAYED_WORK(&ctx->rsrc_put_work, io_rsrc_put_work);
 330	init_task_work(&ctx->rsrc_put_tw, io_rsrc_put_tw);
 331	init_llist_head(&ctx->rsrc_put_llist);
 332	init_llist_head(&ctx->work_llist);
 333	INIT_LIST_HEAD(&ctx->tctx_list);
 334	ctx->submit_state.free_list.next = NULL;
 335	INIT_WQ_LIST(&ctx->locked_free_list);
 336	INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
 337	INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
 338	return ctx;
 339err:
 340	kfree(ctx->dummy_ubuf);
 341	kfree(ctx->cancel_table.hbs);
 342	kfree(ctx->cancel_table_locked.hbs);
 343	kfree(ctx->io_bl);
 344	xa_destroy(&ctx->io_bl_xa);
 345	kfree(ctx);
 346	return NULL;
 347}
 348
 349static void io_account_cq_overflow(struct io_ring_ctx *ctx)
 350{
 351	struct io_rings *r = ctx->rings;
 352
 353	WRITE_ONCE(r->cq_overflow, READ_ONCE(r->cq_overflow) + 1);
 354	ctx->cq_extra--;
 355}
 356
 357static bool req_need_defer(struct io_kiocb *req, u32 seq)
 358{
 359	if (unlikely(req->flags & REQ_F_IO_DRAIN)) {
 360		struct io_ring_ctx *ctx = req->ctx;
 361
 362		return seq + READ_ONCE(ctx->cq_extra) != ctx->cached_cq_tail;
 363	}
 364
 365	return false;
 366}
 367
 368static inline void io_req_track_inflight(struct io_kiocb *req)
 369{
 370	if (!(req->flags & REQ_F_INFLIGHT)) {
 371		req->flags |= REQ_F_INFLIGHT;
 372		atomic_inc(&req->task->io_uring->inflight_tracked);
 373	}
 374}
 375
 376static struct io_kiocb *__io_prep_linked_timeout(struct io_kiocb *req)
 377{
 378	if (WARN_ON_ONCE(!req->link))
 379		return NULL;
 380
 381	req->flags &= ~REQ_F_ARM_LTIMEOUT;
 382	req->flags |= REQ_F_LINK_TIMEOUT;
 383
 384	/* linked timeouts should have two refs once prep'ed */
 385	io_req_set_refcount(req);
 386	__io_req_set_refcount(req->link, 2);
 387	return req->link;
 388}
 389
 390static inline struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
 391{
 392	if (likely(!(req->flags & REQ_F_ARM_LTIMEOUT)))
 393		return NULL;
 394	return __io_prep_linked_timeout(req);
 395}
 396
 397static noinline void __io_arm_ltimeout(struct io_kiocb *req)
 398{
 399	io_queue_linked_timeout(__io_prep_linked_timeout(req));
 400}
 401
 402static inline void io_arm_ltimeout(struct io_kiocb *req)
 403{
 404	if (unlikely(req->flags & REQ_F_ARM_LTIMEOUT))
 405		__io_arm_ltimeout(req);
 406}
 407
 408static void io_prep_async_work(struct io_kiocb *req)
 409{
 410	const struct io_op_def *def = &io_op_defs[req->opcode];
 411	struct io_ring_ctx *ctx = req->ctx;
 412
 413	if (!(req->flags & REQ_F_CREDS)) {
 414		req->flags |= REQ_F_CREDS;
 415		req->creds = get_current_cred();
 416	}
 417
 418	req->work.list.next = NULL;
 419	req->work.flags = 0;
 420	req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
 421	if (req->flags & REQ_F_FORCE_ASYNC)
 422		req->work.flags |= IO_WQ_WORK_CONCURRENT;
 423
 424	if (req->file && !io_req_ffs_set(req))
 425		req->flags |= io_file_get_flags(req->file) << REQ_F_SUPPORT_NOWAIT_BIT;
 426
 427	if (req->flags & REQ_F_ISREG) {
 428		if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
 429			io_wq_hash_work(&req->work, file_inode(req->file));
 430	} else if (!req->file || !S_ISBLK(file_inode(req->file)->i_mode)) {
 431		if (def->unbound_nonreg_file)
 432			req->work.flags |= IO_WQ_WORK_UNBOUND;
 433	}
 434}
 435
 436static void io_prep_async_link(struct io_kiocb *req)
 437{
 438	struct io_kiocb *cur;
 439
 440	if (req->flags & REQ_F_LINK_TIMEOUT) {
 441		struct io_ring_ctx *ctx = req->ctx;
 442
 443		spin_lock_irq(&ctx->timeout_lock);
 444		io_for_each_link(cur, req)
 445			io_prep_async_work(cur);
 446		spin_unlock_irq(&ctx->timeout_lock);
 447	} else {
 448		io_for_each_link(cur, req)
 449			io_prep_async_work(cur);
 450	}
 451}
 452
 453void io_queue_iowq(struct io_kiocb *req, bool *dont_use)
 454{
 455	struct io_kiocb *link = io_prep_linked_timeout(req);
 456	struct io_uring_task *tctx = req->task->io_uring;
 457
 458	BUG_ON(!tctx);
 459	BUG_ON(!tctx->io_wq);
 460
 461	/* init ->work of the whole link before punting */
 462	io_prep_async_link(req);
 463
 464	/*
 465	 * Not expected to happen, but if we do have a bug where this _can_
 466	 * happen, catch it here and ensure the request is marked as
 467	 * canceled. That will make io-wq go through the usual work cancel
 468	 * procedure rather than attempt to run this request (or create a new
 469	 * worker for it).
 470	 */
 471	if (WARN_ON_ONCE(!same_thread_group(req->task, current)))
 472		req->work.flags |= IO_WQ_WORK_CANCEL;
 473
 474	trace_io_uring_queue_async_work(req, io_wq_is_hashed(&req->work));
 475	io_wq_enqueue(tctx->io_wq, &req->work);
 476	if (link)
 477		io_queue_linked_timeout(link);
 478}
 479
 480static __cold void io_queue_deferred(struct io_ring_ctx *ctx)
 481{
 482	while (!list_empty(&ctx->defer_list)) {
 483		struct io_defer_entry *de = list_first_entry(&ctx->defer_list,
 484						struct io_defer_entry, list);
 485
 486		if (req_need_defer(de->req, de->seq))
 487			break;
 488		list_del_init(&de->list);
 489		io_req_task_queue(de->req);
 490		kfree(de);
 491	}
 492}
 493
 494
 495static void io_eventfd_ops(struct rcu_head *rcu)
 496{
 497	struct io_ev_fd *ev_fd = container_of(rcu, struct io_ev_fd, rcu);
 498	int ops = atomic_xchg(&ev_fd->ops, 0);
 499
 500	if (ops & BIT(IO_EVENTFD_OP_SIGNAL_BIT))
 501		eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
 502
 503	/* IO_EVENTFD_OP_FREE_BIT may not be set here depending on callback
 504	 * ordering in a race but if references are 0 we know we have to free
 505	 * it regardless.
 506	 */
 507	if (atomic_dec_and_test(&ev_fd->refs)) {
 508		eventfd_ctx_put(ev_fd->cq_ev_fd);
 509		kfree(ev_fd);
 510	}
 511}
 512
 513static void io_eventfd_signal(struct io_ring_ctx *ctx)
 514{
 515	struct io_ev_fd *ev_fd = NULL;
 516
 517	rcu_read_lock();
 518	/*
 519	 * rcu_dereference ctx->io_ev_fd once and use it for both for checking
 520	 * and eventfd_signal
 521	 */
 522	ev_fd = rcu_dereference(ctx->io_ev_fd);
 523
 524	/*
 525	 * Check again if ev_fd exists incase an io_eventfd_unregister call
 526	 * completed between the NULL check of ctx->io_ev_fd at the start of
 527	 * the function and rcu_read_lock.
 528	 */
 529	if (unlikely(!ev_fd))
 530		goto out;
 531	if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED)
 532		goto out;
 533	if (ev_fd->eventfd_async && !io_wq_current_is_worker())
 534		goto out;
 535
 536	if (likely(eventfd_signal_allowed())) {
 537		eventfd_signal_mask(ev_fd->cq_ev_fd, 1, EPOLL_URING_WAKE);
 538	} else {
 539		atomic_inc(&ev_fd->refs);
 540		if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_SIGNAL_BIT), &ev_fd->ops))
 541			call_rcu_hurry(&ev_fd->rcu, io_eventfd_ops);
 542		else
 543			atomic_dec(&ev_fd->refs);
 544	}
 545
 546out:
 547	rcu_read_unlock();
 548}
 549
 550static void io_eventfd_flush_signal(struct io_ring_ctx *ctx)
 551{
 552	bool skip;
 553
 554	spin_lock(&ctx->completion_lock);
 555
 556	/*
 557	 * Eventfd should only get triggered when at least one event has been
 558	 * posted. Some applications rely on the eventfd notification count
 559	 * only changing IFF a new CQE has been added to the CQ ring. There's
 560	 * no depedency on 1:1 relationship between how many times this
 561	 * function is called (and hence the eventfd count) and number of CQEs
 562	 * posted to the CQ ring.
 563	 */
 564	skip = ctx->cached_cq_tail == ctx->evfd_last_cq_tail;
 565	ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
 566	spin_unlock(&ctx->completion_lock);
 567	if (skip)
 568		return;
 569
 570	io_eventfd_signal(ctx);
 571}
 572
 573void __io_commit_cqring_flush(struct io_ring_ctx *ctx)
 574{
 575	if (ctx->off_timeout_used)
 576		io_flush_timeouts(ctx);
 577	if (ctx->drain_active) {
 578		spin_lock(&ctx->completion_lock);
 579		io_queue_deferred(ctx);
 580		spin_unlock(&ctx->completion_lock);
 581	}
 582	if (ctx->has_evfd)
 583		io_eventfd_flush_signal(ctx);
 584}
 585
 586static inline void __io_cq_lock(struct io_ring_ctx *ctx)
 587	__acquires(ctx->completion_lock)
 588{
 589	if (!ctx->task_complete)
 590		spin_lock(&ctx->completion_lock);
 591}
 592
 593static inline void __io_cq_unlock(struct io_ring_ctx *ctx)
 594{
 595	if (!ctx->task_complete)
 596		spin_unlock(&ctx->completion_lock);
 597}
 598
 599static inline void io_cq_lock(struct io_ring_ctx *ctx)
 600	__acquires(ctx->completion_lock)
 601{
 602	spin_lock(&ctx->completion_lock);
 603}
 604
 605static inline void io_cq_unlock(struct io_ring_ctx *ctx)
 606	__releases(ctx->completion_lock)
 607{
 608	spin_unlock(&ctx->completion_lock);
 609}
 610
 611/* keep it inlined for io_submit_flush_completions() */
 612static inline void __io_cq_unlock_post(struct io_ring_ctx *ctx)
 613	__releases(ctx->completion_lock)
 614{
 615	io_commit_cqring(ctx);
 616	__io_cq_unlock(ctx);
 617	io_commit_cqring_flush(ctx);
 618	io_cqring_wake(ctx);
 619}
 620
 621void io_cq_unlock_post(struct io_ring_ctx *ctx)
 622	__releases(ctx->completion_lock)
 623{
 624	io_commit_cqring(ctx);
 625	spin_unlock(&ctx->completion_lock);
 626	io_commit_cqring_flush(ctx);
 627	io_cqring_wake(ctx);
 628}
 629
 630/* Returns true if there are no backlogged entries after the flush */
 631static void io_cqring_overflow_kill(struct io_ring_ctx *ctx)
 632{
 633	struct io_overflow_cqe *ocqe;
 634	LIST_HEAD(list);
 635
 636	io_cq_lock(ctx);
 637	list_splice_init(&ctx->cq_overflow_list, &list);
 638	clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
 639	io_cq_unlock(ctx);
 640
 641	while (!list_empty(&list)) {
 642		ocqe = list_first_entry(&list, struct io_overflow_cqe, list);
 643		list_del(&ocqe->list);
 644		kfree(ocqe);
 645	}
 646}
 647
 648/* Returns true if there are no backlogged entries after the flush */
 649static void __io_cqring_overflow_flush(struct io_ring_ctx *ctx)
 650{
 651	size_t cqe_size = sizeof(struct io_uring_cqe);
 652
 653	if (__io_cqring_events(ctx) == ctx->cq_entries)
 654		return;
 655
 656	if (ctx->flags & IORING_SETUP_CQE32)
 657		cqe_size <<= 1;
 658
 659	io_cq_lock(ctx);
 660	while (!list_empty(&ctx->cq_overflow_list)) {
 661		struct io_uring_cqe *cqe = io_get_cqe_overflow(ctx, true);
 662		struct io_overflow_cqe *ocqe;
 663
 664		if (!cqe)
 665			break;
 666		ocqe = list_first_entry(&ctx->cq_overflow_list,
 667					struct io_overflow_cqe, list);
 668		memcpy(cqe, &ocqe->cqe, cqe_size);
 669		list_del(&ocqe->list);
 670		kfree(ocqe);
 671	}
 672
 673	if (list_empty(&ctx->cq_overflow_list)) {
 674		clear_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
 675		atomic_andnot(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
 676	}
 677	io_cq_unlock_post(ctx);
 678}
 679
 680static void io_cqring_do_overflow_flush(struct io_ring_ctx *ctx)
 681{
 682	/* iopoll syncs against uring_lock, not completion_lock */
 683	if (ctx->flags & IORING_SETUP_IOPOLL)
 684		mutex_lock(&ctx->uring_lock);
 685	__io_cqring_overflow_flush(ctx);
 686	if (ctx->flags & IORING_SETUP_IOPOLL)
 687		mutex_unlock(&ctx->uring_lock);
 688}
 689
 690static void io_cqring_overflow_flush(struct io_ring_ctx *ctx)
 691{
 692	if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
 693		io_cqring_do_overflow_flush(ctx);
 694}
 695
 696void __io_put_task(struct task_struct *task, int nr)
 697{
 698	struct io_uring_task *tctx = task->io_uring;
 699
 700	percpu_counter_sub(&tctx->inflight, nr);
 701	if (unlikely(atomic_read(&tctx->in_idle)))
 702		wake_up(&tctx->wait);
 703	put_task_struct_many(task, nr);
 704}
 705
 706void io_task_refs_refill(struct io_uring_task *tctx)
 707{
 708	unsigned int refill = -tctx->cached_refs + IO_TCTX_REFS_CACHE_NR;
 709
 710	percpu_counter_add(&tctx->inflight, refill);
 711	refcount_add(refill, &current->usage);
 712	tctx->cached_refs += refill;
 713}
 714
 715static __cold void io_uring_drop_tctx_refs(struct task_struct *task)
 716{
 717	struct io_uring_task *tctx = task->io_uring;
 718	unsigned int refs = tctx->cached_refs;
 719
 720	if (refs) {
 721		tctx->cached_refs = 0;
 722		percpu_counter_sub(&tctx->inflight, refs);
 723		put_task_struct_many(task, refs);
 724	}
 725}
 726
 727static bool io_cqring_event_overflow(struct io_ring_ctx *ctx, u64 user_data,
 728				     s32 res, u32 cflags, u64 extra1, u64 extra2)
 729{
 730	struct io_overflow_cqe *ocqe;
 731	size_t ocq_size = sizeof(struct io_overflow_cqe);
 732	bool is_cqe32 = (ctx->flags & IORING_SETUP_CQE32);
 733
 734	lockdep_assert_held(&ctx->completion_lock);
 735
 736	if (is_cqe32)
 737		ocq_size += sizeof(struct io_uring_cqe);
 738
 739	ocqe = kmalloc(ocq_size, GFP_ATOMIC | __GFP_ACCOUNT);
 740	trace_io_uring_cqe_overflow(ctx, user_data, res, cflags, ocqe);
 741	if (!ocqe) {
 742		/*
 743		 * If we're in ring overflow flush mode, or in task cancel mode,
 744		 * or cannot allocate an overflow entry, then we need to drop it
 745		 * on the floor.
 746		 */
 747		io_account_cq_overflow(ctx);
 748		set_bit(IO_CHECK_CQ_DROPPED_BIT, &ctx->check_cq);
 749		return false;
 750	}
 751	if (list_empty(&ctx->cq_overflow_list)) {
 752		set_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq);
 753		atomic_or(IORING_SQ_CQ_OVERFLOW, &ctx->rings->sq_flags);
 754
 755	}
 756	ocqe->cqe.user_data = user_data;
 757	ocqe->cqe.res = res;
 758	ocqe->cqe.flags = cflags;
 759	if (is_cqe32) {
 760		ocqe->cqe.big_cqe[0] = extra1;
 761		ocqe->cqe.big_cqe[1] = extra2;
 762	}
 763	list_add_tail(&ocqe->list, &ctx->cq_overflow_list);
 764	return true;
 765}
 766
 767bool io_req_cqe_overflow(struct io_kiocb *req)
 768{
 769	if (!(req->flags & REQ_F_CQE32_INIT)) {
 770		req->extra1 = 0;
 771		req->extra2 = 0;
 772	}
 773	return io_cqring_event_overflow(req->ctx, req->cqe.user_data,
 774					req->cqe.res, req->cqe.flags,
 775					req->extra1, req->extra2);
 776}
 777
 778/*
 779 * writes to the cq entry need to come after reading head; the
 780 * control dependency is enough as we're using WRITE_ONCE to
 781 * fill the cq entry
 782 */
 783struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow)
 784{
 785	struct io_rings *rings = ctx->rings;
 786	unsigned int off = ctx->cached_cq_tail & (ctx->cq_entries - 1);
 787	unsigned int free, queued, len;
 788
 789	/*
 790	 * Posting into the CQ when there are pending overflowed CQEs may break
 791	 * ordering guarantees, which will affect links, F_MORE users and more.
 792	 * Force overflow the completion.
 793	 */
 794	if (!overflow && (ctx->check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT)))
 795		return NULL;
 796
 797	/* userspace may cheat modifying the tail, be safe and do min */
 798	queued = min(__io_cqring_events(ctx), ctx->cq_entries);
 799	free = ctx->cq_entries - queued;
 800	/* we need a contiguous range, limit based on the current array offset */
 801	len = min(free, ctx->cq_entries - off);
 802	if (!len)
 803		return NULL;
 804
 805	if (ctx->flags & IORING_SETUP_CQE32) {
 806		off <<= 1;
 807		len <<= 1;
 808	}
 809
 810	ctx->cqe_cached = &rings->cqes[off];
 811	ctx->cqe_sentinel = ctx->cqe_cached + len;
 812
 813	ctx->cached_cq_tail++;
 814	ctx->cqe_cached++;
 815	if (ctx->flags & IORING_SETUP_CQE32)
 816		ctx->cqe_cached++;
 817	return &rings->cqes[off];
 818}
 819
 820static bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res,
 821			      u32 cflags)
 822{
 823	struct io_uring_cqe *cqe;
 824
 825	ctx->cq_extra++;
 826
 827	/*
 828	 * If we can't get a cq entry, userspace overflowed the
 829	 * submission (by quite a lot). Increment the overflow count in
 830	 * the ring.
 831	 */
 832	cqe = io_get_cqe(ctx);
 833	if (likely(cqe)) {
 834		trace_io_uring_complete(ctx, NULL, user_data, res, cflags, 0, 0);
 835
 836		WRITE_ONCE(cqe->user_data, user_data);
 837		WRITE_ONCE(cqe->res, res);
 838		WRITE_ONCE(cqe->flags, cflags);
 839
 840		if (ctx->flags & IORING_SETUP_CQE32) {
 841			WRITE_ONCE(cqe->big_cqe[0], 0);
 842			WRITE_ONCE(cqe->big_cqe[1], 0);
 843		}
 844		return true;
 845	}
 846	return false;
 847}
 848
 849static void __io_flush_post_cqes(struct io_ring_ctx *ctx)
 850	__must_hold(&ctx->uring_lock)
 851{
 852	struct io_submit_state *state = &ctx->submit_state;
 853	unsigned int i;
 854
 855	lockdep_assert_held(&ctx->uring_lock);
 856	for (i = 0; i < state->cqes_count; i++) {
 857		struct io_uring_cqe *cqe = &state->cqes[i];
 858
 859		if (!io_fill_cqe_aux(ctx, cqe->user_data, cqe->res, cqe->flags)) {
 860			if (ctx->task_complete) {
 861				spin_lock(&ctx->completion_lock);
 862				io_cqring_event_overflow(ctx, cqe->user_data,
 863							cqe->res, cqe->flags, 0, 0);
 864				spin_unlock(&ctx->completion_lock);
 865			} else {
 866				io_cqring_event_overflow(ctx, cqe->user_data,
 867							cqe->res, cqe->flags, 0, 0);
 868			}
 869		}
 870	}
 871	state->cqes_count = 0;
 872}
 873
 874static bool __io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags,
 875			      bool allow_overflow)
 876{
 877	bool filled;
 878
 879	io_cq_lock(ctx);
 880	filled = io_fill_cqe_aux(ctx, user_data, res, cflags);
 881	if (!filled && allow_overflow)
 882		filled = io_cqring_event_overflow(ctx, user_data, res, cflags, 0, 0);
 883
 884	io_cq_unlock_post(ctx);
 885	return filled;
 886}
 887
 888bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags)
 889{
 890	return __io_post_aux_cqe(ctx, user_data, res, cflags, true);
 891}
 892
 893bool io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags,
 894		bool allow_overflow)
 895{
 896	struct io_uring_cqe *cqe;
 897	unsigned int length;
 898
 899	if (!defer)
 900		return __io_post_aux_cqe(ctx, user_data, res, cflags, allow_overflow);
 901
 902	length = ARRAY_SIZE(ctx->submit_state.cqes);
 903
 904	lockdep_assert_held(&ctx->uring_lock);
 905
 906	if (ctx->submit_state.cqes_count == length) {
 907		__io_cq_lock(ctx);
 908		__io_flush_post_cqes(ctx);
 909		/* no need to flush - flush is deferred */
 910		__io_cq_unlock_post(ctx);
 911	}
 912
 913	/* For defered completions this is not as strict as it is otherwise,
 914	 * however it's main job is to prevent unbounded posted completions,
 915	 * and in that it works just as well.
 916	 */
 917	if (!allow_overflow && test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq))
 918		return false;
 919
 920	cqe = &ctx->submit_state.cqes[ctx->submit_state.cqes_count++];
 921	cqe->user_data = user_data;
 922	cqe->res = res;
 923	cqe->flags = cflags;
 924	return true;
 925}
 926
 927static void __io_req_complete_post(struct io_kiocb *req)
 928{
 929	struct io_ring_ctx *ctx = req->ctx;
 930
 931	io_cq_lock(ctx);
 932	if (!(req->flags & REQ_F_CQE_SKIP))
 933		io_fill_cqe_req(ctx, req);
 934
 935	/*
 936	 * If we're the last reference to this request, add to our locked
 937	 * free_list cache.
 938	 */
 939	if (req_ref_put_and_test(req)) {
 940		if (req->flags & IO_REQ_LINK_FLAGS) {
 941			if (req->flags & IO_DISARM_MASK)
 942				io_disarm_next(req);
 943			if (req->link) {
 944				io_req_task_queue(req->link);
 945				req->link = NULL;
 946			}
 947		}
 948		io_req_put_rsrc(req);
 949		/*
 950		 * Selected buffer deallocation in io_clean_op() assumes that
 951		 * we don't hold ->completion_lock. Clean them here to avoid
 952		 * deadlocks.
 953		 */
 954		io_put_kbuf_comp(req);
 955		io_dismantle_req(req);
 956		io_put_task(req->task, 1);
 957		wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
 958		ctx->locked_free_nr++;
 959	}
 960	io_cq_unlock_post(ctx);
 961}
 962
 963void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags)
 964{
 965	if (req->ctx->task_complete && (issue_flags & IO_URING_F_IOWQ)) {
 966		req->io_task_work.func = io_req_task_complete;
 967		io_req_task_work_add(req);
 968	} else if (!(issue_flags & IO_URING_F_UNLOCKED) ||
 969		   !(req->ctx->flags & IORING_SETUP_IOPOLL)) {
 970		__io_req_complete_post(req);
 971	} else {
 972		struct io_ring_ctx *ctx = req->ctx;
 973
 974		mutex_lock(&ctx->uring_lock);
 975		__io_req_complete_post(req);
 976		mutex_unlock(&ctx->uring_lock);
 977	}
 978}
 979
 980void io_req_defer_failed(struct io_kiocb *req, s32 res)
 981	__must_hold(&ctx->uring_lock)
 982{
 983	const struct io_op_def *def = &io_op_defs[req->opcode];
 984
 985	lockdep_assert_held(&req->ctx->uring_lock);
 986
 987	req_set_fail(req);
 988	io_req_set_res(req, res, io_put_kbuf(req, IO_URING_F_UNLOCKED));
 989	if (def->fail)
 990		def->fail(req);
 991	io_req_complete_defer(req);
 992}
 993
 994/*
 995 * Don't initialise the fields below on every allocation, but do that in
 996 * advance and keep them valid across allocations.
 997 */
 998static void io_preinit_req(struct io_kiocb *req, struct io_ring_ctx *ctx)
 999{
1000	req->ctx = ctx;
1001	req->link = NULL;
1002	req->async_data = NULL;
1003	/* not necessary, but safer to zero */
1004	req->cqe.res = 0;
1005}
1006
1007static void io_flush_cached_locked_reqs(struct io_ring_ctx *ctx,
1008					struct io_submit_state *state)
1009{
1010	spin_lock(&ctx->completion_lock);
1011	wq_list_splice(&ctx->locked_free_list, &state->free_list);
1012	ctx->locked_free_nr = 0;
1013	spin_unlock(&ctx->completion_lock);
1014}
1015
1016/*
1017 * A request might get retired back into the request caches even before opcode
1018 * handlers and io_issue_sqe() are done with it, e.g. inline completion path.
1019 * Because of that, io_alloc_req() should be called only under ->uring_lock
1020 * and with extra caution to not get a request that is still worked on.
1021 */
1022__cold bool __io_alloc_req_refill(struct io_ring_ctx *ctx)
1023	__must_hold(&ctx->uring_lock)
1024{
1025	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1026	void *reqs[IO_REQ_ALLOC_BATCH];
1027	int ret, i;
1028
1029	/*
1030	 * If we have more than a batch's worth of requests in our IRQ side
1031	 * locked cache, grab the lock and move them over to our submission
1032	 * side cache.
1033	 */
1034	if (data_race(ctx->locked_free_nr) > IO_COMPL_BATCH) {
1035		io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
1036		if (!io_req_cache_empty(ctx))
1037			return true;
1038	}
1039
1040	ret = kmem_cache_alloc_bulk(req_cachep, gfp, ARRAY_SIZE(reqs), reqs);
1041
1042	/*
1043	 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1044	 * retry single alloc to be on the safe side.
1045	 */
1046	if (unlikely(ret <= 0)) {
1047		reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1048		if (!reqs[0])
1049			return false;
1050		ret = 1;
1051	}
1052
1053	percpu_ref_get_many(&ctx->refs, ret);
1054	for (i = 0; i < ret; i++) {
1055		struct io_kiocb *req = reqs[i];
1056
1057		io_preinit_req(req, ctx);
1058		io_req_add_to_cache(req, ctx);
1059	}
1060	return true;
1061}
1062
1063static inline void io_dismantle_req(struct io_kiocb *req)
1064{
1065	unsigned int flags = req->flags;
1066
1067	if (unlikely(flags & IO_REQ_CLEAN_FLAGS))
1068		io_clean_op(req);
1069	if (!(flags & REQ_F_FIXED_FILE))
1070		io_put_file(req->file);
1071}
1072
1073__cold void io_free_req(struct io_kiocb *req)
1074{
1075	struct io_ring_ctx *ctx = req->ctx;
1076
1077	io_req_put_rsrc(req);
1078	io_dismantle_req(req);
1079	io_put_task(req->task, 1);
1080
1081	spin_lock(&ctx->completion_lock);
1082	wq_list_add_head(&req->comp_list, &ctx->locked_free_list);
1083	ctx->locked_free_nr++;
1084	spin_unlock(&ctx->completion_lock);
1085}
1086
1087static void __io_req_find_next_prep(struct io_kiocb *req)
1088{
1089	struct io_ring_ctx *ctx = req->ctx;
1090
1091	spin_lock(&ctx->completion_lock);
1092	io_disarm_next(req);
1093	spin_unlock(&ctx->completion_lock);
1094}
1095
1096static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
1097{
1098	struct io_kiocb *nxt;
1099
1100	/*
1101	 * If LINK is set, we have dependent requests in this chain. If we
1102	 * didn't fail this request, queue the first one up, moving any other
1103	 * dependencies to the next request. In case of failure, fail the rest
1104	 * of the chain.
1105	 */
1106	if (unlikely(req->flags & IO_DISARM_MASK))
1107		__io_req_find_next_prep(req);
1108	nxt = req->link;
1109	req->link = NULL;
1110	return nxt;
1111}
1112
1113static void ctx_flush_and_put(struct io_ring_ctx *ctx, bool *locked)
1114{
1115	if (!ctx)
1116		return;
1117	if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1118		atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1119	if (*locked) {
1120		io_submit_flush_completions(ctx);
1121		mutex_unlock(&ctx->uring_lock);
1122		*locked = false;
1123	}
1124	percpu_ref_put(&ctx->refs);
1125}
1126
1127static unsigned int handle_tw_list(struct llist_node *node,
1128				   struct io_ring_ctx **ctx, bool *locked,
1129				   struct llist_node *last)
1130{
1131	unsigned int count = 0;
1132
1133	while (node != last) {
1134		struct llist_node *next = node->next;
1135		struct io_kiocb *req = container_of(node, struct io_kiocb,
1136						    io_task_work.node);
1137
1138		prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1139
1140		if (req->ctx != *ctx) {
1141			ctx_flush_and_put(*ctx, locked);
1142			*ctx = req->ctx;
1143			/* if not contended, grab and improve batching */
1144			*locked = mutex_trylock(&(*ctx)->uring_lock);
1145			percpu_ref_get(&(*ctx)->refs);
1146		}
1147		req->io_task_work.func(req, locked);
1148		node = next;
1149		count++;
1150	}
1151
1152	return count;
1153}
1154
1155/**
1156 * io_llist_xchg - swap all entries in a lock-less list
1157 * @head:	the head of lock-less list to delete all entries
1158 * @new:	new entry as the head of the list
1159 *
1160 * If list is empty, return NULL, otherwise, return the pointer to the first entry.
1161 * The order of entries returned is from the newest to the oldest added one.
1162 */
1163static inline struct llist_node *io_llist_xchg(struct llist_head *head,
1164					       struct llist_node *new)
1165{
1166	return xchg(&head->first, new);
1167}
1168
1169/**
1170 * io_llist_cmpxchg - possibly swap all entries in a lock-less list
1171 * @head:	the head of lock-less list to delete all entries
1172 * @old:	expected old value of the first entry of the list
1173 * @new:	new entry as the head of the list
1174 *
1175 * perform a cmpxchg on the first entry of the list.
1176 */
1177
1178static inline struct llist_node *io_llist_cmpxchg(struct llist_head *head,
1179						  struct llist_node *old,
1180						  struct llist_node *new)
1181{
1182	return cmpxchg(&head->first, old, new);
1183}
1184
1185void tctx_task_work(struct callback_head *cb)
1186{
1187	bool uring_locked = false;
1188	struct io_ring_ctx *ctx = NULL;
1189	struct io_uring_task *tctx = container_of(cb, struct io_uring_task,
1190						  task_work);
1191	struct llist_node fake = {};
1192	struct llist_node *node;
1193	unsigned int loops = 1;
1194	unsigned int count;
1195
1196	if (unlikely(current->flags & PF_EXITING)) {
1197		io_fallback_tw(tctx);
1198		return;
1199	}
1200
1201	node = io_llist_xchg(&tctx->task_list, &fake);
1202	count = handle_tw_list(node, &ctx, &uring_locked, NULL);
1203	node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1204	while (node != &fake) {
1205		loops++;
1206		node = io_llist_xchg(&tctx->task_list, &fake);
1207		count += handle_tw_list(node, &ctx, &uring_locked, &fake);
1208		node = io_llist_cmpxchg(&tctx->task_list, &fake, NULL);
1209	}
1210
1211	ctx_flush_and_put(ctx, &uring_locked);
1212
1213	/* relaxed read is enough as only the task itself sets ->in_idle */
1214	if (unlikely(atomic_read(&tctx->in_idle)))
1215		io_uring_drop_tctx_refs(current);
1216
1217	trace_io_uring_task_work_run(tctx, count, loops);
1218}
1219
1220static __cold void io_fallback_tw(struct io_uring_task *tctx)
1221{
1222	struct llist_node *node = llist_del_all(&tctx->task_list);
1223	struct io_kiocb *req;
1224
1225	while (node) {
1226		req = container_of(node, struct io_kiocb, io_task_work.node);
1227		node = node->next;
1228		if (llist_add(&req->io_task_work.node,
1229			      &req->ctx->fallback_llist))
1230			schedule_delayed_work(&req->ctx->fallback_work, 1);
1231	}
1232}
1233
1234static void io_req_local_work_add(struct io_kiocb *req)
1235{
1236	struct io_ring_ctx *ctx = req->ctx;
1237
1238	percpu_ref_get(&ctx->refs);
1239
1240	if (!llist_add(&req->io_task_work.node, &ctx->work_llist)) {
1241		percpu_ref_put(&ctx->refs);
1242		return;
1243	}
1244	/* need it for the following io_cqring_wake() */
1245	smp_mb__after_atomic();
1246
1247	if (unlikely(atomic_read(&req->task->io_uring->in_idle))) {
1248		io_move_task_work_from_local(ctx);
1249		percpu_ref_put(&ctx->refs);
1250		return;
1251	}
1252
1253	if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1254		atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1255
1256	if (ctx->has_evfd)
1257		io_eventfd_signal(ctx);
1258	__io_cqring_wake(ctx);
1259	percpu_ref_put(&ctx->refs);
1260}
1261
1262void __io_req_task_work_add(struct io_kiocb *req, bool allow_local)
1263{
1264	struct io_uring_task *tctx = req->task->io_uring;
1265	struct io_ring_ctx *ctx = req->ctx;
1266
1267	if (allow_local && ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
1268		io_req_local_work_add(req);
1269		return;
1270	}
1271
1272	/* task_work already pending, we're done */
1273	if (!llist_add(&req->io_task_work.node, &tctx->task_list))
1274		return;
1275
1276	if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1277		atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1278
1279	if (likely(!task_work_add(req->task, &tctx->task_work, ctx->notify_method)))
1280		return;
1281
1282	io_fallback_tw(tctx);
1283}
1284
1285static void __cold io_move_task_work_from_local(struct io_ring_ctx *ctx)
1286{
1287	struct llist_node *node;
1288
1289	node = llist_del_all(&ctx->work_llist);
1290	while (node) {
1291		struct io_kiocb *req = container_of(node, struct io_kiocb,
1292						    io_task_work.node);
1293
1294		node = node->next;
1295		__io_req_task_work_add(req, false);
1296	}
1297}
1298
1299int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked)
1300{
1301	struct llist_node *node;
1302	struct llist_node fake;
1303	struct llist_node *current_final = NULL;
1304	int ret;
1305	unsigned int loops = 1;
1306
1307	if (unlikely(ctx->submitter_task != current))
1308		return -EEXIST;
1309
1310	node = io_llist_xchg(&ctx->work_llist, &fake);
1311	ret = 0;
1312again:
1313	while (node != current_final) {
1314		struct llist_node *next = node->next;
1315		struct io_kiocb *req = container_of(node, struct io_kiocb,
1316						    io_task_work.node);
1317		prefetch(container_of(next, struct io_kiocb, io_task_work.node));
1318		req->io_task_work.func(req, locked);
1319		ret++;
1320		node = next;
1321	}
1322
1323	if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
1324		atomic_andnot(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
1325
1326	node = io_llist_cmpxchg(&ctx->work_llist, &fake, NULL);
1327	if (node != &fake) {
1328		loops++;
1329		current_final = &fake;
1330		node = io_llist_xchg(&ctx->work_llist, &fake);
1331		goto again;
1332	}
1333
1334	if (*locked)
1335		io_submit_flush_completions(ctx);
1336	trace_io_uring_local_work_run(ctx, ret, loops);
1337	return ret;
1338
1339}
1340
1341int io_run_local_work(struct io_ring_ctx *ctx)
1342{
1343	bool locked;
1344	int ret;
1345
1346	if (llist_empty(&ctx->work_llist))
1347		return 0;
1348
1349	__set_current_state(TASK_RUNNING);
1350	locked = mutex_trylock(&ctx->uring_lock);
1351	ret = __io_run_local_work(ctx, &locked);
1352	if (locked)
1353		mutex_unlock(&ctx->uring_lock);
1354
1355	return ret;
1356}
1357
1358static void io_req_task_cancel(struct io_kiocb *req, bool *locked)
1359{
1360	io_tw_lock(req->ctx, locked);
1361	io_req_defer_failed(req, req->cqe.res);
1362}
1363
1364void io_req_task_submit(struct io_kiocb *req, bool *locked)
1365{
1366	io_tw_lock(req->ctx, locked);
1367	/* req->task == current here, checking PF_EXITING is safe */
1368	if (likely(!(req->task->flags & PF_EXITING)))
1369		io_queue_sqe(req);
1370	else
1371		io_req_defer_failed(req, -EFAULT);
1372}
1373
1374void io_req_task_queue_fail(struct io_kiocb *req, int ret)
1375{
1376	io_req_set_res(req, ret, 0);
1377	req->io_task_work.func = io_req_task_cancel;
1378	io_req_task_work_add(req);
1379}
1380
1381void io_req_task_queue(struct io_kiocb *req)
1382{
1383	req->io_task_work.func = io_req_task_submit;
1384	io_req_task_work_add(req);
1385}
1386
1387void io_queue_next(struct io_kiocb *req)
1388{
1389	struct io_kiocb *nxt = io_req_find_next(req);
1390
1391	if (nxt)
1392		io_req_task_queue(nxt);
1393}
1394
1395void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node)
1396	__must_hold(&ctx->uring_lock)
1397{
1398	struct task_struct *task = NULL;
1399	int task_refs = 0;
1400
1401	do {
1402		struct io_kiocb *req = container_of(node, struct io_kiocb,
1403						    comp_list);
1404
1405		if (unlikely(req->flags & IO_REQ_CLEAN_SLOW_FLAGS)) {
1406			if (req->flags & REQ_F_REFCOUNT) {
1407				node = req->comp_list.next;
1408				if (!req_ref_put_and_test(req))
1409					continue;
1410			}
1411			if ((req->flags & REQ_F_POLLED) && req->apoll) {
1412				struct async_poll *apoll = req->apoll;
1413
1414				if (apoll->double_poll)
1415					kfree(apoll->double_poll);
1416				if (!io_alloc_cache_put(&ctx->apoll_cache, &apoll->cache))
1417					kfree(apoll);
1418				req->flags &= ~REQ_F_POLLED;
1419			}
1420			if (req->flags & IO_REQ_LINK_FLAGS)
1421				io_queue_next(req);
1422			if (unlikely(req->flags & IO_REQ_CLEAN_FLAGS))
1423				io_clean_op(req);
1424		}
1425		if (!(req->flags & REQ_F_FIXED_FILE))
1426			io_put_file(req->file);
1427
1428		io_req_put_rsrc_locked(req, ctx);
1429
1430		if (req->task != task) {
1431			if (task)
1432				io_put_task(task, task_refs);
1433			task = req->task;
1434			task_refs = 0;
1435		}
1436		task_refs++;
1437		node = req->comp_list.next;
1438		io_req_add_to_cache(req, ctx);
1439	} while (node);
1440
1441	if (task)
1442		io_put_task(task, task_refs);
1443}
1444
1445static void __io_submit_flush_completions(struct io_ring_ctx *ctx)
1446	__must_hold(&ctx->uring_lock)
1447{
1448	struct io_wq_work_node *node, *prev;
1449	struct io_submit_state *state = &ctx->submit_state;
1450
1451	__io_cq_lock(ctx);
1452	/* must come first to preserve CQE ordering in failure cases */
1453	if (state->cqes_count)
1454		__io_flush_post_cqes(ctx);
1455	wq_list_for_each(node, prev, &state->compl_reqs) {
1456		struct io_kiocb *req = container_of(node, struct io_kiocb,
1457					    comp_list);
1458
1459		if (!(req->flags & REQ_F_CQE_SKIP) &&
1460		    unlikely(!__io_fill_cqe_req(ctx, req))) {
1461			if (ctx->task_complete) {
1462				spin_lock(&ctx->completion_lock);
1463				io_req_cqe_overflow(req);
1464				spin_unlock(&ctx->completion_lock);
1465			} else {
1466				io_req_cqe_overflow(req);
1467			}
1468		}
1469	}
1470	__io_cq_unlock_post(ctx);
1471
1472	if (!wq_list_empty(&ctx->submit_state.compl_reqs)) {
1473		io_free_batch_list(ctx, state->compl_reqs.first);
1474		INIT_WQ_LIST(&state->compl_reqs);
1475	}
1476}
1477
1478/*
1479 * Drop reference to request, return next in chain (if there is one) if this
1480 * was the last reference to this request.
1481 */
1482static inline struct io_kiocb *io_put_req_find_next(struct io_kiocb *req)
1483{
1484	struct io_kiocb *nxt = NULL;
1485
1486	if (req_ref_put_and_test(req)) {
1487		if (unlikely(req->flags & IO_REQ_LINK_FLAGS))
1488			nxt = io_req_find_next(req);
1489		io_free_req(req);
1490	}
1491	return nxt;
1492}
1493
1494static unsigned io_cqring_events(struct io_ring_ctx *ctx)
1495{
1496	/* See comment at the top of this file */
1497	smp_rmb();
1498	return __io_cqring_events(ctx);
1499}
1500
1501/*
1502 * We can't just wait for polled events to come to us, we have to actively
1503 * find and complete them.
1504 */
1505static __cold void io_iopoll_try_reap_events(struct io_ring_ctx *ctx)
1506{
1507	if (!(ctx->flags & IORING_SETUP_IOPOLL))
1508		return;
1509
1510	mutex_lock(&ctx->uring_lock);
1511	while (!wq_list_empty(&ctx->iopoll_list)) {
1512		/* let it sleep and repeat later if can't complete a request */
1513		if (io_do_iopoll(ctx, true) == 0)
1514			break;
1515		/*
1516		 * Ensure we allow local-to-the-cpu processing to take place,
1517		 * in this case we need to ensure that we reap all events.
1518		 * Also let task_work, etc. to progress by releasing the mutex
1519		 */
1520		if (need_resched()) {
1521			mutex_unlock(&ctx->uring_lock);
1522			cond_resched();
1523			mutex_lock(&ctx->uring_lock);
1524		}
1525	}
1526	mutex_unlock(&ctx->uring_lock);
1527}
1528
1529static int io_iopoll_check(struct io_ring_ctx *ctx, long min)
1530{
1531	unsigned int nr_events = 0;
1532	int ret = 0;
1533	unsigned long check_cq;
1534
1535	if (!io_allowed_run_tw(ctx))
1536		return -EEXIST;
1537
1538	check_cq = READ_ONCE(ctx->check_cq);
1539	if (unlikely(check_cq)) {
1540		if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
1541			__io_cqring_overflow_flush(ctx);
1542		/*
1543		 * Similarly do not spin if we have not informed the user of any
1544		 * dropped CQE.
1545		 */
1546		if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
1547			return -EBADR;
1548	}
1549	/*
1550	 * Don't enter poll loop if we already have events pending.
1551	 * If we do, we can potentially be spinning for commands that
1552	 * already triggered a CQE (eg in error).
1553	 */
1554	if (io_cqring_events(ctx))
1555		return 0;
1556
1557	do {
1558		/*
1559		 * If a submit got punted to a workqueue, we can have the
1560		 * application entering polling for a command before it gets
1561		 * issued. That app will hold the uring_lock for the duration
1562		 * of the poll right here, so we need to take a breather every
1563		 * now and then to ensure that the issue has a chance to add
1564		 * the poll to the issued list. Otherwise we can spin here
1565		 * forever, while the workqueue is stuck trying to acquire the
1566		 * very same mutex.
1567		 */
1568		if (wq_list_empty(&ctx->iopoll_list) ||
1569		    io_task_work_pending(ctx)) {
1570			u32 tail = ctx->cached_cq_tail;
1571
1572			(void) io_run_local_work_locked(ctx);
1573
1574			if (task_work_pending(current) ||
1575			    wq_list_empty(&ctx->iopoll_list)) {
1576				mutex_unlock(&ctx->uring_lock);
1577				io_run_task_work();
1578				mutex_lock(&ctx->uring_lock);
1579			}
1580			/* some requests don't go through iopoll_list */
1581			if (tail != ctx->cached_cq_tail ||
1582			    wq_list_empty(&ctx->iopoll_list))
1583				break;
1584		}
1585		ret = io_do_iopoll(ctx, !min);
1586		if (ret < 0)
1587			break;
1588		nr_events += ret;
1589		ret = 0;
1590	} while (nr_events < min && !need_resched());
1591
1592	return ret;
1593}
1594
1595void io_req_task_complete(struct io_kiocb *req, bool *locked)
1596{
1597	if (*locked)
1598		io_req_complete_defer(req);
1599	else
1600		io_req_complete_post(req, IO_URING_F_UNLOCKED);
1601}
1602
1603/*
1604 * After the iocb has been issued, it's safe to be found on the poll list.
1605 * Adding the kiocb to the list AFTER submission ensures that we don't
1606 * find it from a io_do_iopoll() thread before the issuer is done
1607 * accessing the kiocb cookie.
1608 */
1609static void io_iopoll_req_issued(struct io_kiocb *req, unsigned int issue_flags)
1610{
1611	struct io_ring_ctx *ctx = req->ctx;
1612	const bool needs_lock = issue_flags & IO_URING_F_UNLOCKED;
1613
1614	/* workqueue context doesn't hold uring_lock, grab it now */
1615	if (unlikely(needs_lock))
1616		mutex_lock(&ctx->uring_lock);
1617
1618	/*
1619	 * Track whether we have multiple files in our lists. This will impact
1620	 * how we do polling eventually, not spinning if we're on potentially
1621	 * different devices.
1622	 */
1623	if (wq_list_empty(&ctx->iopoll_list)) {
1624		ctx->poll_multi_queue = false;
1625	} else if (!ctx->poll_multi_queue) {
1626		struct io_kiocb *list_req;
1627
1628		list_req = container_of(ctx->iopoll_list.first, struct io_kiocb,
1629					comp_list);
1630		if (list_req->file != req->file)
1631			ctx->poll_multi_queue = true;
1632	}
1633
1634	/*
1635	 * For fast devices, IO may have already completed. If it has, add
1636	 * it to the front so we find it first.
1637	 */
1638	if (READ_ONCE(req->iopoll_completed))
1639		wq_list_add_head(&req->comp_list, &ctx->iopoll_list);
1640	else
1641		wq_list_add_tail(&req->comp_list, &ctx->iopoll_list);
1642
1643	if (unlikely(needs_lock)) {
1644		/*
1645		 * If IORING_SETUP_SQPOLL is enabled, sqes are either handle
1646		 * in sq thread task context or in io worker task context. If
1647		 * current task context is sq thread, we don't need to check
1648		 * whether should wake up sq thread.
1649		 */
1650		if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1651		    wq_has_sleeper(&ctx->sq_data->wait))
1652			wake_up(&ctx->sq_data->wait);
1653
1654		mutex_unlock(&ctx->uring_lock);
1655	}
1656}
1657
1658static bool io_bdev_nowait(struct block_device *bdev)
1659{
1660	return !bdev || bdev_nowait(bdev);
1661}
1662
1663/*
1664 * If we tracked the file through the SCM inflight mechanism, we could support
1665 * any file. For now, just ensure that anything potentially problematic is done
1666 * inline.
1667 */
1668static bool __io_file_supports_nowait(struct file *file, umode_t mode)
1669{
1670	if (S_ISBLK(mode)) {
1671		if (IS_ENABLED(CONFIG_BLOCK) &&
1672		    io_bdev_nowait(I_BDEV(file->f_mapping->host)))
1673			return true;
1674		return false;
1675	}
1676	if (S_ISSOCK(mode))
1677		return true;
1678	if (S_ISREG(mode)) {
1679		if (IS_ENABLED(CONFIG_BLOCK) &&
1680		    io_bdev_nowait(file->f_inode->i_sb->s_bdev) &&
1681		    !io_is_uring_fops(file))
1682			return true;
1683		return false;
1684	}
1685
1686	/* any ->read/write should understand O_NONBLOCK */
1687	if (file->f_flags & O_NONBLOCK)
1688		return true;
1689	return file->f_mode & FMODE_NOWAIT;
1690}
1691
1692/*
1693 * If we tracked the file through the SCM inflight mechanism, we could support
1694 * any file. For now, just ensure that anything potentially problematic is done
1695 * inline.
1696 */
1697unsigned int io_file_get_flags(struct file *file)
1698{
1699	umode_t mode = file_inode(file)->i_mode;
1700	unsigned int res = 0;
1701
1702	if (S_ISREG(mode))
1703		res |= FFS_ISREG;
1704	if (__io_file_supports_nowait(file, mode))
1705		res |= FFS_NOWAIT;
1706	return res;
1707}
1708
1709bool io_alloc_async_data(struct io_kiocb *req)
1710{
1711	WARN_ON_ONCE(!io_op_defs[req->opcode].async_size);
1712	req->async_data = kmalloc(io_op_defs[req->opcode].async_size, GFP_KERNEL);
1713	if (req->async_data) {
1714		req->flags |= REQ_F_ASYNC_DATA;
1715		return false;
1716	}
1717	return true;
1718}
1719
1720int io_req_prep_async(struct io_kiocb *req)
1721{
1722	const struct io_op_def *def = &io_op_defs[req->opcode];
1723
1724	/* assign early for deferred execution for non-fixed file */
1725	if (def->needs_file && !(req->flags & REQ_F_FIXED_FILE))
1726		req->file = io_file_get_normal(req, req->cqe.fd);
1727	if (!def->prep_async)
1728		return 0;
1729	if (WARN_ON_ONCE(req_has_async_data(req)))
1730		return -EFAULT;
1731	if (!io_op_defs[req->opcode].manual_alloc) {
1732		if (io_alloc_async_data(req))
1733			return -EAGAIN;
1734	}
1735	return def->prep_async(req);
1736}
1737
1738static u32 io_get_sequence(struct io_kiocb *req)
1739{
1740	u32 seq = req->ctx->cached_sq_head;
1741	struct io_kiocb *cur;
1742
1743	/* need original cached_sq_head, but it was increased for each req */
1744	io_for_each_link(cur, req)
1745		seq--;
1746	return seq;
1747}
1748
1749static __cold void io_drain_req(struct io_kiocb *req)
1750	__must_hold(&ctx->uring_lock)
1751{
1752	struct io_ring_ctx *ctx = req->ctx;
1753	struct io_defer_entry *de;
1754	int ret;
1755	u32 seq = io_get_sequence(req);
1756
1757	/* Still need defer if there is pending req in defer list. */
1758	spin_lock(&ctx->completion_lock);
1759	if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) {
1760		spin_unlock(&ctx->completion_lock);
1761queue:
1762		ctx->drain_active = false;
1763		io_req_task_queue(req);
1764		return;
1765	}
1766	spin_unlock(&ctx->completion_lock);
1767
1768	io_prep_async_link(req);
1769	de = kmalloc(sizeof(*de), GFP_KERNEL);
1770	if (!de) {
1771		ret = -ENOMEM;
1772		io_req_defer_failed(req, ret);
1773		return;
1774	}
1775
1776	spin_lock(&ctx->completion_lock);
1777	if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) {
1778		spin_unlock(&ctx->completion_lock);
1779		kfree(de);
1780		goto queue;
1781	}
1782
1783	trace_io_uring_defer(req);
1784	de->req = req;
1785	de->seq = seq;
1786	list_add_tail(&de->list, &ctx->defer_list);
1787	spin_unlock(&ctx->completion_lock);
1788}
1789
1790static void io_clean_op(struct io_kiocb *req)
1791{
1792	if (req->flags & REQ_F_BUFFER_SELECTED) {
1793		spin_lock(&req->ctx->completion_lock);
1794		io_put_kbuf_comp(req);
1795		spin_unlock(&req->ctx->completion_lock);
1796	}
1797
1798	if (req->flags & REQ_F_NEED_CLEANUP) {
1799		const struct io_op_def *def = &io_op_defs[req->opcode];
1800
1801		if (def->cleanup)
1802			def->cleanup(req);
1803	}
1804	if ((req->flags & REQ_F_POLLED) && req->apoll) {
1805		kfree(req->apoll->double_poll);
1806		kfree(req->apoll);
1807		req->apoll = NULL;
1808	}
1809	if (req->flags & REQ_F_INFLIGHT) {
1810		struct io_uring_task *tctx = req->task->io_uring;
1811
1812		atomic_dec(&tctx->inflight_tracked);
1813	}
1814	if (req->flags & REQ_F_CREDS)
1815		put_cred(req->creds);
1816	if (req->flags & REQ_F_ASYNC_DATA) {
1817		kfree(req->async_data);
1818		req->async_data = NULL;
1819	}
1820	req->flags &= ~IO_REQ_CLEAN_FLAGS;
1821}
1822
1823static bool io_assign_file(struct io_kiocb *req, unsigned int issue_flags)
1824{
1825	if (req->file || !io_op_defs[req->opcode].needs_file)
1826		return true;
1827
1828	if (req->flags & REQ_F_FIXED_FILE)
1829		req->file = io_file_get_fixed(req, req->cqe.fd, issue_flags);
1830	else
1831		req->file = io_file_get_normal(req, req->cqe.fd);
1832
1833	return !!req->file;
1834}
1835
1836static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
1837{
1838	const struct io_op_def *def = &io_op_defs[req->opcode];
1839	const struct cred *creds = NULL;
1840	int ret;
1841
1842	if (unlikely(!io_assign_file(req, issue_flags)))
1843		return -EBADF;
1844
1845	if (unlikely((req->flags & REQ_F_CREDS) && req->creds != current_cred()))
1846		creds = override_creds(req->creds);
1847
1848	if (!def->audit_skip)
1849		audit_uring_entry(req->opcode);
1850
1851	ret = def->issue(req, issue_flags);
1852
1853	if (!def->audit_skip)
1854		audit_uring_exit(!ret, ret);
1855
1856	if (creds)
1857		revert_creds(creds);
1858
1859	if (ret == IOU_OK) {
1860		if (issue_flags & IO_URING_F_COMPLETE_DEFER)
1861			io_req_complete_defer(req);
1862		else
1863			io_req_complete_post(req, issue_flags);
1864	} else if (ret != IOU_ISSUE_SKIP_COMPLETE)
1865		return ret;
1866
1867	/* If the op doesn't have a file, we're not polling for it */
1868	if ((req->ctx->flags & IORING_SETUP_IOPOLL) && def->iopoll_queue)
1869		io_iopoll_req_issued(req, issue_flags);
1870
1871	return 0;
1872}
1873
1874int io_poll_issue(struct io_kiocb *req, bool *locked)
1875{
1876	io_tw_lock(req->ctx, locked);
1877	return io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_MULTISHOT|
1878				 IO_URING_F_COMPLETE_DEFER);
1879}
1880
1881struct io_wq_work *io_wq_free_work(struct io_wq_work *work)
1882{
1883	struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1884
1885	req = io_put_req_find_next(req);
1886	return req ? &req->work : NULL;
1887}
1888
1889void io_wq_submit_work(struct io_wq_work *work)
1890{
1891	struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1892	const struct io_op_def *def = &io_op_defs[req->opcode];
1893	unsigned int issue_flags = IO_URING_F_UNLOCKED | IO_URING_F_IOWQ;
1894	bool needs_poll = false;
1895	int ret = 0, err = -ECANCELED;
1896
1897	/* one will be dropped by ->io_wq_free_work() after returning to io-wq */
1898	if (!(req->flags & REQ_F_REFCOUNT))
1899		__io_req_set_refcount(req, 2);
1900	else
1901		req_ref_get(req);
1902
1903	io_arm_ltimeout(req);
1904
1905	/* either cancelled or io-wq is dying, so don't touch tctx->iowq */
1906	if (work->flags & IO_WQ_WORK_CANCEL) {
1907fail:
1908		io_req_task_queue_fail(req, err);
1909		return;
1910	}
1911	if (!io_assign_file(req, issue_flags)) {
1912		err = -EBADF;
1913		work->flags |= IO_WQ_WORK_CANCEL;
1914		goto fail;
1915	}
1916
1917	if (req->flags & REQ_F_FORCE_ASYNC) {
1918		bool opcode_poll = def->pollin || def->pollout;
1919
1920		if (opcode_poll && file_can_poll(req->file)) {
1921			needs_poll = true;
1922			issue_flags |= IO_URING_F_NONBLOCK;
1923		}
1924	}
1925
1926	do {
1927		ret = io_issue_sqe(req, issue_flags);
1928		if (ret != -EAGAIN)
1929			break;
1930		/*
1931		 * We can get EAGAIN for iopolled IO even though we're
1932		 * forcing a sync submission from here, since we can't
1933		 * wait for request slots on the block side.
1934		 */
1935		if (!needs_poll) {
1936			if (!(req->ctx->flags & IORING_SETUP_IOPOLL))
1937				break;
1938			cond_resched();
1939			continue;
1940		}
1941
1942		if (io_arm_poll_handler(req, issue_flags) == IO_APOLL_OK)
1943			return;
1944		/* aborted or ready, in either case retry blocking */
1945		needs_poll = false;
1946		issue_flags &= ~IO_URING_F_NONBLOCK;
1947	} while (1);
1948
1949	/* avoid locking problems by failing it from a clean context */
1950	if (ret < 0)
1951		io_req_task_queue_fail(req, ret);
1952}
1953
1954inline struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
1955				      unsigned int issue_flags)
1956{
1957	struct io_ring_ctx *ctx = req->ctx;
1958	struct file *file = NULL;
1959	unsigned long file_ptr;
1960
1961	io_ring_submit_lock(ctx, issue_flags);
1962
1963	if (unlikely((unsigned int)fd >= ctx->nr_user_files))
1964		goto out;
1965	fd = array_index_nospec(fd, ctx->nr_user_files);
1966	file_ptr = io_fixed_file_slot(&ctx->file_table, fd)->file_ptr;
1967	file = (struct file *) (file_ptr & FFS_MASK);
1968	file_ptr &= ~FFS_MASK;
1969	/* mask in overlapping REQ_F and FFS bits */
1970	req->flags |= (file_ptr << REQ_F_SUPPORT_NOWAIT_BIT);
1971	io_req_set_rsrc_node(req, ctx, 0);
1972out:
1973	io_ring_submit_unlock(ctx, issue_flags);
1974	return file;
1975}
1976
1977struct file *io_file_get_normal(struct io_kiocb *req, int fd)
1978{
1979	struct file *file = fget(fd);
1980
1981	trace_io_uring_file_get(req, fd);
1982
1983	/* we don't allow fixed io_uring files */
1984	if (file && io_is_uring_fops(file))
1985		io_req_track_inflight(req);
1986	return file;
1987}
1988
1989static void io_queue_async(struct io_kiocb *req, int ret)
1990	__must_hold(&req->ctx->uring_lock)
1991{
1992	struct io_kiocb *linked_timeout;
1993
1994	if (ret != -EAGAIN || (req->flags & REQ_F_NOWAIT)) {
1995		io_req_defer_failed(req, ret);
1996		return;
1997	}
1998
1999	linked_timeout = io_prep_linked_timeout(req);
2000
2001	switch (io_arm_poll_handler(req, 0)) {
2002	case IO_APOLL_READY:
2003		io_kbuf_recycle(req, 0);
2004		io_req_task_queue(req);
2005		break;
2006	case IO_APOLL_ABORTED:
2007		io_kbuf_recycle(req, 0);
2008		io_queue_iowq(req, NULL);
2009		break;
2010	case IO_APOLL_OK:
2011		break;
2012	}
2013
2014	if (linked_timeout)
2015		io_queue_linked_timeout(linked_timeout);
2016}
2017
2018static inline void io_queue_sqe(struct io_kiocb *req)
2019	__must_hold(&req->ctx->uring_lock)
2020{
2021	int ret;
2022
2023	ret = io_issue_sqe(req, IO_URING_F_NONBLOCK|IO_URING_F_COMPLETE_DEFER);
2024
2025	/*
2026	 * We async punt it if the file wasn't marked NOWAIT, or if the file
2027	 * doesn't support non-blocking read/write attempts
2028	 */
2029	if (likely(!ret))
2030		io_arm_ltimeout(req);
2031	else
2032		io_queue_async(req, ret);
2033}
2034
2035static void io_queue_sqe_fallback(struct io_kiocb *req)
2036	__must_hold(&req->ctx->uring_lock)
2037{
2038	if (unlikely(req->flags & REQ_F_FAIL)) {
2039		/*
2040		 * We don't submit, fail them all, for that replace hardlinks
2041		 * with normal links. Extra REQ_F_LINK is tolerated.
2042		 */
2043		req->flags &= ~REQ_F_HARDLINK;
2044		req->flags |= REQ_F_LINK;
2045		io_req_defer_failed(req, req->cqe.res);
2046	} else {
2047		int ret = io_req_prep_async(req);
2048
2049		if (unlikely(ret)) {
2050			io_req_defer_failed(req, ret);
2051			return;
2052		}
2053
2054		if (unlikely(req->ctx->drain_active))
2055			io_drain_req(req);
2056		else
2057			io_queue_iowq(req, NULL);
2058	}
2059}
2060
2061/*
2062 * Check SQE restrictions (opcode and flags).
2063 *
2064 * Returns 'true' if SQE is allowed, 'false' otherwise.
2065 */
2066static inline bool io_check_restriction(struct io_ring_ctx *ctx,
2067					struct io_kiocb *req,
2068					unsigned int sqe_flags)
2069{
2070	if (!test_bit(req->opcode, ctx->restrictions.sqe_op))
2071		return false;
2072
2073	if ((sqe_flags & ctx->restrictions.sqe_flags_required) !=
2074	    ctx->restrictions.sqe_flags_required)
2075		return false;
2076
2077	if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed |
2078			  ctx->restrictions.sqe_flags_required))
2079		return false;
2080
2081	return true;
2082}
2083
2084static void io_init_req_drain(struct io_kiocb *req)
2085{
2086	struct io_ring_ctx *ctx = req->ctx;
2087	struct io_kiocb *head = ctx->submit_state.link.head;
2088
2089	ctx->drain_active = true;
2090	if (head) {
2091		/*
2092		 * If we need to drain a request in the middle of a link, drain
2093		 * the head request and the next request/link after the current
2094		 * link. Considering sequential execution of links,
2095		 * REQ_F_IO_DRAIN will be maintained for every request of our
2096		 * link.
2097		 */
2098		head->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2099		ctx->drain_next = true;
2100	}
2101}
2102
2103static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
2104		       const struct io_uring_sqe *sqe)
2105	__must_hold(&ctx->uring_lock)
2106{
2107	const struct io_op_def *def;
2108	unsigned int sqe_flags;
2109	int personality;
2110	u8 opcode;
2111
2112	/* req is partially pre-initialised, see io_preinit_req() */
2113	req->opcode = opcode = READ_ONCE(sqe->opcode);
2114	/* same numerical values with corresponding REQ_F_*, safe to copy */
2115	req->flags = sqe_flags = READ_ONCE(sqe->flags);
2116	req->cqe.user_data = READ_ONCE(sqe->user_data);
2117	req->file = NULL;
2118	req->rsrc_node = NULL;
2119	req->task = current;
2120
2121	if (unlikely(opcode >= IORING_OP_LAST)) {
2122		req->opcode = 0;
2123		return -EINVAL;
2124	}
2125	def = &io_op_defs[opcode];
2126	if (unlikely(sqe_flags & ~SQE_COMMON_FLAGS)) {
2127		/* enforce forwards compatibility on users */
2128		if (sqe_flags & ~SQE_VALID_FLAGS)
2129			return -EINVAL;
2130		if (sqe_flags & IOSQE_BUFFER_SELECT) {
2131			if (!def->buffer_select)
2132				return -EOPNOTSUPP;
2133			req->buf_index = READ_ONCE(sqe->buf_group);
2134		}
2135		if (sqe_flags & IOSQE_CQE_SKIP_SUCCESS)
2136			ctx->drain_disabled = true;
2137		if (sqe_flags & IOSQE_IO_DRAIN) {
2138			if (ctx->drain_disabled)
2139				return -EOPNOTSUPP;
2140			io_init_req_drain(req);
2141		}
2142	}
2143	if (unlikely(ctx->restricted || ctx->drain_active || ctx->drain_next)) {
2144		if (ctx->restricted && !io_check_restriction(ctx, req, sqe_flags))
2145			return -EACCES;
2146		/* knock it to the slow queue path, will be drained there */
2147		if (ctx->drain_active)
2148			req->flags |= REQ_F_FORCE_ASYNC;
2149		/* if there is no link, we're at "next" request and need to drain */
2150		if (unlikely(ctx->drain_next) && !ctx->submit_state.link.head) {
2151			ctx->drain_next = false;
2152			ctx->drain_active = true;
2153			req->flags |= REQ_F_IO_DRAIN | REQ_F_FORCE_ASYNC;
2154		}
2155	}
2156
2157	if (!def->ioprio && sqe->ioprio)
2158		return -EINVAL;
2159	if (!def->iopoll && (ctx->flags & IORING_SETUP_IOPOLL))
2160		return -EINVAL;
2161
2162	if (def->needs_file) {
2163		struct io_submit_state *state = &ctx->submit_state;
2164
2165		req->cqe.fd = READ_ONCE(sqe->fd);
2166
2167		/*
2168		 * Plug now if we have more than 2 IO left after this, and the
2169		 * target is potentially a read/write to block based storage.
2170		 */
2171		if (state->need_plug && def->plug) {
2172			state->plug_started = true;
2173			state->need_plug = false;
2174			blk_start_plug_nr_ios(&state->plug, state->submit_nr);
2175		}
2176	}
2177
2178	personality = READ_ONCE(sqe->personality);
2179	if (personality) {
2180		int ret;
2181
2182		req->creds = xa_load(&ctx->personalities, personality);
2183		if (!req->creds)
2184			return -EINVAL;
2185		get_cred(req->creds);
2186		ret = security_uring_override_creds(req->creds);
2187		if (ret) {
2188			put_cred(req->creds);
2189			return ret;
2190		}
2191		req->flags |= REQ_F_CREDS;
2192	}
2193
2194	return def->prep(req, sqe);
2195}
2196
2197static __cold int io_submit_fail_init(const struct io_uring_sqe *sqe,
2198				      struct io_kiocb *req, int ret)
2199{
2200	struct io_ring_ctx *ctx = req->ctx;
2201	struct io_submit_link *link = &ctx->submit_state.link;
2202	struct io_kiocb *head = link->head;
2203
2204	trace_io_uring_req_failed(sqe, req, ret);
2205
2206	/*
2207	 * Avoid breaking links in the middle as it renders links with SQPOLL
2208	 * unusable. Instead of failing eagerly, continue assembling the link if
2209	 * applicable and mark the head with REQ_F_FAIL. The link flushing code
2210	 * should find the flag and handle the rest.
2211	 */
2212	req_fail_link_node(req, ret);
2213	if (head && !(head->flags & REQ_F_FAIL))
2214		req_fail_link_node(head, -ECANCELED);
2215
2216	if (!(req->flags & IO_REQ_LINK_FLAGS)) {
2217		if (head) {
2218			link->last->link = req;
2219			link->head = NULL;
2220			req = head;
2221		}
2222		io_queue_sqe_fallback(req);
2223		return ret;
2224	}
2225
2226	if (head)
2227		link->last->link = req;
2228	else
2229		link->head = req;
2230	link->last = req;
2231	return 0;
2232}
2233
2234static inline int io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2235			 const struct io_uring_sqe *sqe)
2236	__must_hold(&ctx->uring_lock)
2237{
2238	struct io_submit_link *link = &ctx->submit_state.link;
2239	int ret;
2240
2241	ret = io_init_req(ctx, req, sqe);
2242	if (unlikely(ret))
2243		return io_submit_fail_init(sqe, req, ret);
2244
2245	/* don't need @sqe from now on */
2246	trace_io_uring_submit_sqe(req, true);
2247
2248	/*
2249	 * If we already have a head request, queue this one for async
2250	 * submittal once the head completes. If we don't have a head but
2251	 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2252	 * submitted sync once the chain is complete. If none of those
2253	 * conditions are true (normal request), then just queue it.
2254	 */
2255	if (unlikely(link->head)) {
2256		ret = io_req_prep_async(req);
2257		if (unlikely(ret))
2258			return io_submit_fail_init(sqe, req, ret);
2259
2260		trace_io_uring_link(req, link->head);
2261		link->last->link = req;
2262		link->last = req;
2263
2264		if (req->flags & IO_REQ_LINK_FLAGS)
2265			return 0;
2266		/* last request of the link, flush it */
2267		req = link->head;
2268		link->head = NULL;
2269		if (req->flags & (REQ_F_FORCE_ASYNC | REQ_F_FAIL))
2270			goto fallback;
2271
2272	} else if (unlikely(req->flags & (IO_REQ_LINK_FLAGS |
2273					  REQ_F_FORCE_ASYNC | REQ_F_FAIL))) {
2274		if (req->flags & IO_REQ_LINK_FLAGS) {
2275			link->head = req;
2276			link->last = req;
2277		} else {
2278fallback:
2279			io_queue_sqe_fallback(req);
2280		}
2281		return 0;
2282	}
2283
2284	io_queue_sqe(req);
2285	return 0;
2286}
2287
2288/*
2289 * Batched submission is done, ensure local IO is flushed out.
2290 */
2291static void io_submit_state_end(struct io_ring_ctx *ctx)
2292{
2293	struct io_submit_state *state = &ctx->submit_state;
2294
2295	if (unlikely(state->link.head))
2296		io_queue_sqe_fallback(state->link.head);
2297	/* flush only after queuing links as they can generate completions */
2298	io_submit_flush_completions(ctx);
2299	if (state->plug_started)
2300		blk_finish_plug(&state->plug);
2301}
2302
2303/*
2304 * Start submission side cache.
2305 */
2306static void io_submit_state_start(struct io_submit_state *state,
2307				  unsigned int max_ios)
2308{
2309	state->plug_started = false;
2310	state->need_plug = max_ios > 2;
2311	state->submit_nr = max_ios;
2312	/* set only head, no need to init link_last in advance */
2313	state->link.head = NULL;
2314}
2315
2316static void io_commit_sqring(struct io_ring_ctx *ctx)
2317{
2318	struct io_rings *rings = ctx->rings;
2319
2320	/*
2321	 * Ensure any loads from the SQEs are done at this point,
2322	 * since once we write the new head, the application could
2323	 * write new data to them.
2324	 */
2325	smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2326}
2327
2328/*
2329 * Fetch an sqe, if one is available. Note this returns a pointer to memory
2330 * that is mapped by userspace. This means that care needs to be taken to
2331 * ensure that reads are stable, as we cannot rely on userspace always
2332 * being a good citizen. If members of the sqe are validated and then later
2333 * used, it's important that those reads are done through READ_ONCE() to
2334 * prevent a re-load down the line.
2335 */
2336static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
2337{
2338	unsigned head, mask = ctx->sq_entries - 1;
2339	unsigned sq_idx = ctx->cached_sq_head++ & mask;
2340
2341	/*
2342	 * The cached sq head (or cq tail) serves two purposes:
2343	 *
2344	 * 1) allows us to batch the cost of updating the user visible
2345	 *    head updates.
2346	 * 2) allows the kernel side to track the head on its own, even
2347	 *    though the application is the one updating it.
2348	 */
2349	head = READ_ONCE(ctx->sq_array[sq_idx]);
2350	if (likely(head < ctx->sq_entries)) {
2351		/* double index for 128-byte SQEs, twice as long */
2352		if (ctx->flags & IORING_SETUP_SQE128)
2353			head <<= 1;
2354		return &ctx->sq_sqes[head];
2355	}
2356
2357	/* drop invalid entries */
2358	ctx->cq_extra--;
2359	WRITE_ONCE(ctx->rings->sq_dropped,
2360		   READ_ONCE(ctx->rings->sq_dropped) + 1);
2361	return NULL;
2362}
2363
2364int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
2365	__must_hold(&ctx->uring_lock)
2366{
2367	unsigned int entries = io_sqring_entries(ctx);
2368	unsigned int left;
2369	int ret;
2370
2371	if (unlikely(!entries))
2372		return 0;
2373	/* make sure SQ entry isn't read before tail */
2374	ret = left = min3(nr, ctx->sq_entries, entries);
2375	io_get_task_refs(left);
2376	io_submit_state_start(&ctx->submit_state, left);
2377
2378	do {
2379		const struct io_uring_sqe *sqe;
2380		struct io_kiocb *req;
2381
2382		if (unlikely(!io_alloc_req_refill(ctx)))
2383			break;
2384		req = io_alloc_req(ctx);
2385		sqe = io_get_sqe(ctx);
2386		if (unlikely(!sqe)) {
2387			io_req_add_to_cache(req, ctx);
2388			break;
2389		}
2390
2391		/*
2392		 * Continue submitting even for sqe failure if the
2393		 * ring was setup with IORING_SETUP_SUBMIT_ALL
2394		 */
2395		if (unlikely(io_submit_sqe(ctx, req, sqe)) &&
2396		    !(ctx->flags & IORING_SETUP_SUBMIT_ALL)) {
2397			left--;
2398			break;
2399		}
2400	} while (--left);
2401
2402	if (unlikely(left)) {
2403		ret -= left;
2404		/* try again if it submitted nothing and can't allocate a req */
2405		if (!ret && io_req_cache_empty(ctx))
2406			ret = -EAGAIN;
2407		current->io_uring->cached_refs += left;
2408	}
2409
2410	io_submit_state_end(ctx);
2411	 /* Commit SQ ring head once we've consumed and submitted all SQEs */
2412	io_commit_sqring(ctx);
2413	return ret;
2414}
2415
2416struct io_wait_queue {
2417	struct wait_queue_entry wq;
2418	struct io_ring_ctx *ctx;
2419	unsigned cq_tail;
2420	unsigned nr_timeouts;
2421};
2422
2423static inline bool io_has_work(struct io_ring_ctx *ctx)
2424{
2425	return test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq) ||
2426	       ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
2427		!llist_empty(&ctx->work_llist));
2428}
2429
2430static inline bool io_should_wake(struct io_wait_queue *iowq)
2431{
2432	struct io_ring_ctx *ctx = iowq->ctx;
2433	int dist = READ_ONCE(ctx->rings->cq.tail) - (int) iowq->cq_tail;
2434
2435	/*
2436	 * Wake up if we have enough events, or if a timeout occurred since we
2437	 * started waiting. For timeouts, we always want to return to userspace,
2438	 * regardless of event count.
2439	 */
2440	return dist >= 0 || atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2441}
2442
2443static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2444			    int wake_flags, void *key)
2445{
2446	struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2447							wq);
2448	struct io_ring_ctx *ctx = iowq->ctx;
2449
2450	/*
2451	 * Cannot safely flush overflowed CQEs from here, ensure we wake up
2452	 * the task, and the next invocation will do it.
2453	 */
2454	if (io_should_wake(iowq) || io_has_work(ctx))
2455		return autoremove_wake_function(curr, mode, wake_flags, key);
2456	return -1;
2457}
2458
2459int io_run_task_work_sig(struct io_ring_ctx *ctx)
2460{
2461	if (io_run_task_work_ctx(ctx) > 0)
2462		return 1;
2463	if (task_sigpending(current))
2464		return -EINTR;
2465	return 0;
2466}
2467
2468/* when returns >0, the caller should retry */
2469static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
2470					  struct io_wait_queue *iowq,
2471					  ktime_t *timeout)
2472{
2473	int ret;
2474	unsigned long check_cq;
2475
2476	/* make sure we run task_work before checking for signals */
2477	ret = io_run_task_work_sig(ctx);
2478	if (ret || io_should_wake(iowq))
2479		return ret;
2480
2481	check_cq = READ_ONCE(ctx->check_cq);
2482	if (unlikely(check_cq)) {
2483		/* let the caller flush overflows, retry */
2484		if (check_cq & BIT(IO_CHECK_CQ_OVERFLOW_BIT))
2485			return 1;
2486		if (check_cq & BIT(IO_CHECK_CQ_DROPPED_BIT))
2487			return -EBADR;
2488	}
2489	if (!schedule_hrtimeout(timeout, HRTIMER_MODE_ABS))
2490		return -ETIME;
2491
2492	/*
2493	 * Run task_work after scheduling. If we got woken because of
2494	 * task_work being processed, run it now rather than let the caller
2495	 * do another wait loop.
2496	 */
2497	ret = io_run_task_work_sig(ctx);
2498	return ret < 0 ? ret : 1;
2499}
2500
2501/*
2502 * Wait until events become available, if we don't already have some. The
2503 * application must reap them itself, as they reside on the shared cq ring.
2504 */
2505static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2506			  const sigset_t __user *sig, size_t sigsz,
2507			  struct __kernel_timespec __user *uts)
2508{
2509	struct io_wait_queue iowq;
2510	struct io_rings *rings = ctx->rings;
2511	ktime_t timeout = KTIME_MAX;
2512	int ret;
2513
2514	if (!io_allowed_run_tw(ctx))
2515		return -EEXIST;
2516
2517	do {
2518		/* always run at least 1 task work to process local work */
2519		ret = io_run_task_work_ctx(ctx);
2520		if (ret < 0)
2521			return ret;
2522		io_cqring_overflow_flush(ctx);
2523
2524		/* if user messes with these they will just get an early return */
2525		if (__io_cqring_events_user(ctx) >= min_events)
2526			return 0;
2527	} while (ret > 0);
2528
2529	if (sig) {
2530#ifdef CONFIG_COMPAT
2531		if (in_compat_syscall())
2532			ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2533						      sigsz);
2534		else
2535#endif
2536			ret = set_user_sigmask(sig, sigsz);
2537
2538		if (ret)
2539			return ret;
2540	}
2541
2542	if (uts) {
2543		struct timespec64 ts;
2544
2545		if (get_timespec64(&ts, uts))
2546			return -EFAULT;
2547		timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
2548	}
2549
2550	init_waitqueue_func_entry(&iowq.wq, io_wake_function);
2551	iowq.wq.private = current;
2552	INIT_LIST_HEAD(&iowq.wq.entry);
2553	iowq.ctx = ctx;
2554	iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2555	iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
2556
2557	trace_io_uring_cqring_wait(ctx, min_events);
2558	do {
2559		if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2560			finish_wait(&ctx->cq_wait, &iowq.wq);
2561			io_cqring_do_overflow_flush(ctx);
2562		}
2563		prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
2564						TASK_INTERRUPTIBLE);
2565		ret = io_cqring_wait_schedule(ctx, &iowq, &timeout);
2566		if (__io_cqring_events_user(ctx) >= min_events)
2567			break;
2568		cond_resched();
2569	} while (ret > 0);
2570
2571	finish_wait(&ctx->cq_wait, &iowq.wq);
2572	restore_saved_sigmask_unless(ret == -EINTR);
2573
2574	return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2575}
2576
2577static void io_mem_free(void *ptr)
2578{
2579	struct page *page;
2580
2581	if (!ptr)
2582		return;
2583
2584	page = virt_to_head_page(ptr);
2585	if (put_page_testzero(page))
2586		free_compound_page(page);
2587}
2588
2589static void *io_mem_alloc(size_t size)
2590{
2591	gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP;
2592
2593	return (void *) __get_free_pages(gfp, get_order(size));
2594}
2595
2596static unsigned long rings_size(struct io_ring_ctx *ctx, unsigned int sq_entries,
2597				unsigned int cq_entries, size_t *sq_offset)
2598{
2599	struct io_rings *rings;
2600	size_t off, sq_array_size;
2601
2602	off = struct_size(rings, cqes, cq_entries);
2603	if (off == SIZE_MAX)
2604		return SIZE_MAX;
2605	if (ctx->flags & IORING_SETUP_CQE32) {
2606		if (check_shl_overflow(off, 1, &off))
2607			return SIZE_MAX;
2608	}
2609
2610#ifdef CONFIG_SMP
2611	off = ALIGN(off, SMP_CACHE_BYTES);
2612	if (off == 0)
2613		return SIZE_MAX;
2614#endif
2615
2616	if (sq_offset)
2617		*sq_offset = off;
2618
2619	sq_array_size = array_size(sizeof(u32), sq_entries);
2620	if (sq_array_size == SIZE_MAX)
2621		return SIZE_MAX;
2622
2623	if (check_add_overflow(off, sq_array_size, &off))
2624		return SIZE_MAX;
2625
2626	return off;
2627}
2628
2629static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg,
2630			       unsigned int eventfd_async)
2631{
2632	struct io_ev_fd *ev_fd;
2633	__s32 __user *fds = arg;
2634	int fd;
2635
2636	ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2637					lockdep_is_held(&ctx->uring_lock));
2638	if (ev_fd)
2639		return -EBUSY;
2640
2641	if (copy_from_user(&fd, fds, sizeof(*fds)))
2642		return -EFAULT;
2643
2644	ev_fd = kmalloc(sizeof(*ev_fd), GFP_KERNEL);
2645	if (!ev_fd)
2646		return -ENOMEM;
2647
2648	ev_fd->cq_ev_fd = eventfd_ctx_fdget(fd);
2649	if (IS_ERR(ev_fd->cq_ev_fd)) {
2650		int ret = PTR_ERR(ev_fd->cq_ev_fd);
2651		kfree(ev_fd);
2652		return ret;
2653	}
2654
2655	spin_lock(&ctx->completion_lock);
2656	ctx->evfd_last_cq_tail = ctx->cached_cq_tail;
2657	spin_unlock(&ctx->completion_lock);
2658
2659	ev_fd->eventfd_async = eventfd_async;
2660	ctx->has_evfd = true;
2661	rcu_assign_pointer(ctx->io_ev_fd, ev_fd);
2662	atomic_set(&ev_fd->refs, 1);
2663	atomic_set(&ev_fd->ops, 0);
2664	return 0;
2665}
2666
2667static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2668{
2669	struct io_ev_fd *ev_fd;
2670
2671	ev_fd = rcu_dereference_protected(ctx->io_ev_fd,
2672					lockdep_is_held(&ctx->uring_lock));
2673	if (ev_fd) {
2674		ctx->has_evfd = false;
2675		rcu_assign_pointer(ctx->io_ev_fd, NULL);
2676		if (!atomic_fetch_or(BIT(IO_EVENTFD_OP_FREE_BIT), &ev_fd->ops))
2677			call_rcu(&ev_fd->rcu, io_eventfd_ops);
2678		return 0;
2679	}
2680
2681	return -ENXIO;
2682}
2683
2684static void io_req_caches_free(struct io_ring_ctx *ctx)
2685{
2686	int nr = 0;
2687
2688	mutex_lock(&ctx->uring_lock);
2689	io_flush_cached_locked_reqs(ctx, &ctx->submit_state);
2690
2691	while (!io_req_cache_empty(ctx)) {
2692		struct io_kiocb *req = io_alloc_req(ctx);
2693
2694		kmem_cache_free(req_cachep, req);
2695		nr++;
2696	}
2697	if (nr)
2698		percpu_ref_put_many(&ctx->refs, nr);
2699	mutex_unlock(&ctx->uring_lock);
2700}
2701
2702static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
2703{
2704	io_sq_thread_finish(ctx);
2705	io_rsrc_refs_drop(ctx);
2706	/* __io_rsrc_put_work() may need uring_lock to progress, wait w/o it */
2707	io_wait_rsrc_data(ctx->buf_data);
2708	io_wait_rsrc_data(ctx->file_data);
2709
2710	mutex_lock(&ctx->uring_lock);
2711	if (ctx->buf_data)
2712		__io_sqe_buffers_unregister(ctx);
2713	if (ctx->file_data)
2714		__io_sqe_files_unregister(ctx);
2715	io_cqring_overflow_kill(ctx);
2716	io_eventfd_unregister(ctx);
2717	io_alloc_cache_free(&ctx->apoll_cache, io_apoll_cache_free);
2718	io_alloc_cache_free(&ctx->netmsg_cache, io_netmsg_cache_free);
2719	mutex_unlock(&ctx->uring_lock);
2720	io_destroy_buffers(ctx);
2721	if (ctx->sq_creds)
2722		put_cred(ctx->sq_creds);
2723	if (ctx->submitter_task)
2724		put_task_struct(ctx->submitter_task);
2725
2726	/* there are no registered resources left, nobody uses it */
2727	if (ctx->rsrc_node)
2728		io_rsrc_node_destroy(ctx->rsrc_node);
2729	if (ctx->rsrc_backup_node)
2730		io_rsrc_node_destroy(ctx->rsrc_backup_node);
2731	flush_delayed_work(&ctx->rsrc_put_work);
2732	flush_delayed_work(&ctx->fallback_work);
2733
2734	WARN_ON_ONCE(!list_empty(&ctx->rsrc_ref_list));
2735	WARN_ON_ONCE(!llist_empty(&ctx->rsrc_put_llist));
2736
2737#if defined(CONFIG_UNIX)
2738	if (ctx->ring_sock) {
2739		ctx->ring_sock->file = NULL; /* so that iput() is called */
2740		sock_release(ctx->ring_sock);
2741	}
2742#endif
2743	WARN_ON_ONCE(!list_empty(&ctx->ltimeout_list));
2744
2745	if (ctx->mm_account) {
2746		mmdrop(ctx->mm_account);
2747		ctx->mm_account = NULL;
2748	}
2749	io_mem_free(ctx->rings);
2750	io_mem_free(ctx->sq_sqes);
2751
2752	percpu_ref_exit(&ctx->refs);
2753	free_uid(ctx->user);
2754	io_req_caches_free(ctx);
2755	if (ctx->hash_map)
2756		io_wq_put_hash(ctx->hash_map);
2757	kfree(ctx->cancel_table.hbs);
2758	kfree(ctx->cancel_table_locked.hbs);
2759	kfree(ctx->dummy_ubuf);
2760	kfree(ctx->io_bl);
2761	xa_destroy(&ctx->io_bl_xa);
2762	kfree(ctx);
2763}
2764
2765static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2766{
2767	struct io_ring_ctx *ctx = file->private_data;
2768	__poll_t mask = 0;
2769
2770	poll_wait(file, &ctx->cq_wait, wait);
2771	/*
2772	 * synchronizes with barrier from wq_has_sleeper call in
2773	 * io_commit_cqring
2774	 */
2775	smp_rmb();
2776	if (!io_sqring_full(ctx))
2777		mask |= EPOLLOUT | EPOLLWRNORM;
2778
2779	/*
2780	 * Don't flush cqring overflow list here, just do a simple check.
2781	 * Otherwise there could possible be ABBA deadlock:
2782	 *      CPU0                    CPU1
2783	 *      ----                    ----
2784	 * lock(&ctx->uring_lock);
2785	 *                              lock(&ep->mtx);
2786	 *                              lock(&ctx->uring_lock);
2787	 * lock(&ep->mtx);
2788	 *
2789	 * Users may get EPOLLIN meanwhile seeing nothing in cqring, this
2790	 * pushes them to do the flush.
2791	 */
2792
2793	if (io_cqring_events(ctx) || io_has_work(ctx))
2794		mask |= EPOLLIN | EPOLLRDNORM;
2795
2796	return mask;
2797}
2798
2799static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
2800{
2801	const struct cred *creds;
2802
2803	creds = xa_erase(&ctx->personalities, id);
2804	if (creds) {
2805		put_cred(creds);
2806		return 0;
2807	}
2808
2809	return -EINVAL;
2810}
2811
2812struct io_tctx_exit {
2813	struct callback_head		task_work;
2814	struct completion		completion;
2815	struct io_ring_ctx		*ctx;
2816};
2817
2818static __cold void io_tctx_exit_cb(struct callback_head *cb)
2819{
2820	struct io_uring_task *tctx = current->io_uring;
2821	struct io_tctx_exit *work;
2822
2823	work = container_of(cb, struct io_tctx_exit, task_work);
2824	/*
2825	 * When @in_idle, we're in cancellation and it's racy to remove the
2826	 * node. It'll be removed by the end of cancellation, just ignore it.
2827	 * tctx can be NULL if the queueing of this task_work raced with
2828	 * work cancelation off the exec path.
2829	 */
2830	if (tctx && !atomic_read(&tctx->in_idle))
2831		io_uring_del_tctx_node((unsigned long)work->ctx);
2832	complete(&work->completion);
2833}
2834
2835static __cold bool io_cancel_ctx_cb(struct io_wq_work *work, void *data)
2836{
2837	struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2838
2839	return req->ctx == data;
2840}
2841
2842static __cold void io_ring_exit_work(struct work_struct *work)
2843{
2844	struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, exit_work);
2845	unsigned long timeout = jiffies + HZ * 60 * 5;
2846	unsigned long interval = HZ / 20;
2847	struct io_tctx_exit exit;
2848	struct io_tctx_node *node;
2849	int ret;
2850
2851	/*
2852	 * If we're doing polled IO and end up having requests being
2853	 * submitted async (out-of-line), then completions can come in while
2854	 * we're waiting for refs to drop. We need to reap these manually,
2855	 * as nobody else will be looking for them.
2856	 */
2857	do {
2858		if (test_bit(IO_CHECK_CQ_OVERFLOW_BIT, &ctx->check_cq)) {
2859			mutex_lock(&ctx->uring_lock);
2860			io_cqring_overflow_kill(ctx);
2861			mutex_unlock(&ctx->uring_lock);
2862		}
2863
2864		if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
2865			io_move_task_work_from_local(ctx);
2866
2867		while (io_uring_try_cancel_requests(ctx, NULL, true))
2868			cond_resched();
2869
2870		if (ctx->sq_data) {
2871			struct io_sq_data *sqd = ctx->sq_data;
2872			struct task_struct *tsk;
2873
2874			io_sq_thread_park(sqd);
2875			tsk = sqd->thread;
2876			if (tsk && tsk->io_uring && tsk->io_uring->io_wq)
2877				io_wq_cancel_cb(tsk->io_uring->io_wq,
2878						io_cancel_ctx_cb, ctx, true);
2879			io_sq_thread_unpark(sqd);
2880		}
2881
2882		io_req_caches_free(ctx);
2883
2884		if (WARN_ON_ONCE(time_after(jiffies, timeout))) {
2885			/* there is little hope left, don't run it too often */
2886			interval = HZ * 60;
2887		}
2888	} while (!wait_for_completion_timeout(&ctx->ref_comp, interval));
2889
2890	init_completion(&exit.completion);
2891	init_task_work(&exit.task_work, io_tctx_exit_cb);
2892	exit.ctx = ctx;
2893	/*
2894	 * Some may use context even when all refs and requests have been put,
2895	 * and they are free to do so while still holding uring_lock or
2896	 * completion_lock, see io_req_task_submit(). Apart from other work,
2897	 * this lock/unlock section also waits them to finish.
2898	 */
2899	mutex_lock(&ctx->uring_lock);
2900	while (!list_empty(&ctx->tctx_list)) {
2901		WARN_ON_ONCE(time_after(jiffies, timeout));
2902
2903		node = list_first_entry(&ctx->tctx_list, struct io_tctx_node,
2904					ctx_node);
2905		/* don't spin on a single task if cancellation failed */
2906		list_rotate_left(&ctx->tctx_list);
2907		ret = task_work_add(node->task, &exit.task_work, TWA_SIGNAL);
2908		if (WARN_ON_ONCE(ret))
2909			continue;
2910
2911		mutex_unlock(&ctx->uring_lock);
2912		wait_for_completion(&exit.completion);
2913		mutex_lock(&ctx->uring_lock);
2914	}
2915	mutex_unlock(&ctx->uring_lock);
2916	spin_lock(&ctx->completion_lock);
2917	spin_unlock(&ctx->completion_lock);
2918
2919	io_ring_ctx_free(ctx);
2920}
2921
2922static __cold void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2923{
2924	unsigned long index;
2925	struct creds *creds;
2926
2927	mutex_lock(&ctx->uring_lock);
2928	percpu_ref_kill(&ctx->refs);
2929	xa_for_each(&ctx->personalities, index, creds)
2930		io_unregister_personality(ctx, index);
2931	if (ctx->rings)
2932		io_poll_remove_all(ctx, NULL, true);
2933	mutex_unlock(&ctx->uring_lock);
2934
2935	/*
2936	 * If we failed setting up the ctx, we might not have any rings
2937	 * and therefore did not submit any requests
2938	 */
2939	if (ctx->rings)
2940		io_kill_timeouts(ctx, NULL, true);
2941
2942	INIT_WORK(&ctx->exit_work, io_ring_exit_work);
2943	/*
2944	 * Use system_unbound_wq to avoid spawning tons of event kworkers
2945	 * if we're exiting a ton of rings at the same time. It just adds
2946	 * noise and overhead, there's no discernable change in runtime
2947	 * over using system_wq.
2948	 */
2949	queue_work(system_unbound_wq, &ctx->exit_work);
2950}
2951
2952static int io_uring_release(struct inode *inode, struct file *file)
2953{
2954	struct io_ring_ctx *ctx = file->private_data;
2955
2956	file->private_data = NULL;
2957	io_ring_ctx_wait_and_kill(ctx);
2958	return 0;
2959}
2960
2961struct io_task_cancel {
2962	struct task_struct *task;
2963	bool all;
2964};
2965
2966static bool io_cancel_task_cb(struct io_wq_work *work, void *data)
2967{
2968	struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2969	struct io_task_cancel *cancel = data;
2970
2971	return io_match_task_safe(req, cancel->task, cancel->all);
2972}
2973
2974static __cold bool io_cancel_defer_files(struct io_ring_ctx *ctx,
2975					 struct task_struct *task,
2976					 bool cancel_all)
2977{
2978	struct io_defer_entry *de;
2979	LIST_HEAD(list);
2980
2981	spin_lock(&ctx->completion_lock);
2982	list_for_each_entry_reverse(de, &ctx->defer_list, list) {
2983		if (io_match_task_safe(de->req, task, cancel_all)) {
2984			list_cut_position(&list, &ctx->defer_list, &de->list);
2985			break;
2986		}
2987	}
2988	spin_unlock(&ctx->completion_lock);
2989	if (list_empty(&list))
2990		return false;
2991
2992	while (!list_empty(&list)) {
2993		de = list_first_entry(&list, struct io_defer_entry, list);
2994		list_del_init(&de->list);
2995		io_req_task_queue_fail(de->req, -ECANCELED);
2996		kfree(de);
2997	}
2998	return true;
2999}
3000
3001static __cold bool io_uring_try_cancel_iowq(struct io_ring_ctx *ctx)
3002{
3003	struct io_tctx_node *node;
3004	enum io_wq_cancel cret;
3005	bool ret = false;
3006
3007	mutex_lock(&ctx->uring_lock);
3008	list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3009		struct io_uring_task *tctx = node->task->io_uring;
3010
3011		/*
3012		 * io_wq will stay alive while we hold uring_lock, because it's
3013		 * killed after ctx nodes, which requires to take the lock.
3014		 */
3015		if (!tctx || !tctx->io_wq)
3016			continue;
3017		cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_ctx_cb, ctx, true);
3018		ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3019	}
3020	mutex_unlock(&ctx->uring_lock);
3021
3022	return ret;
3023}
3024
3025static __cold bool io_uring_try_cancel_requests(struct io_ring_ctx *ctx,
3026						struct task_struct *task,
3027						bool cancel_all)
3028{
3029	struct io_task_cancel cancel = { .task = task, .all = cancel_all, };
3030	struct io_uring_task *tctx = task ? task->io_uring : NULL;
3031	enum io_wq_cancel cret;
3032	bool ret = false;
3033
3034	/* failed during ring init, it couldn't have issued any requests */
3035	if (!ctx->rings)
3036		return false;
3037
3038	if (!task) {
3039		ret |= io_uring_try_cancel_iowq(ctx);
3040	} else if (tctx && tctx->io_wq) {
3041		/*
3042		 * Cancels requests of all rings, not only @ctx, but
3043		 * it's fine as the task is in exit/exec.
3044		 */
3045		cret = io_wq_cancel_cb(tctx->io_wq, io_cancel_task_cb,
3046				       &cancel, true);
3047		ret |= (cret != IO_WQ_CANCEL_NOTFOUND);
3048	}
3049
3050	/* SQPOLL thread does its own polling */
3051	if ((!(ctx->flags & IORING_SETUP_SQPOLL) && cancel_all) ||
3052	    (ctx->sq_data && ctx->sq_data->thread == current)) {
3053		while (!wq_list_empty(&ctx->iopoll_list)) {
3054			io_iopoll_try_reap_events(ctx);
3055			ret = true;
3056		}
3057	}
3058
3059	if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3060		ret |= io_run_local_work(ctx) > 0;
3061	ret |= io_cancel_defer_files(ctx, task, cancel_all);
3062	mutex_lock(&ctx->uring_lock);
3063	ret |= io_poll_remove_all(ctx, task, cancel_all);
3064	mutex_unlock(&ctx->uring_lock);
3065	ret |= io_kill_timeouts(ctx, task, cancel_all);
3066	if (task)
3067		ret |= io_run_task_work() > 0;
3068	return ret;
3069}
3070
3071static s64 tctx_inflight(struct io_uring_task *tctx, bool tracked)
3072{
3073	if (tracked)
3074		return atomic_read(&tctx->inflight_tracked);
3075	return percpu_counter_sum(&tctx->inflight);
3076}
3077
3078/*
3079 * Find any io_uring ctx that this task has registered or done IO on, and cancel
3080 * requests. @sqd should be not-null IFF it's an SQPOLL thread cancellation.
3081 */
3082__cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd)
3083{
3084	struct io_uring_task *tctx = current->io_uring;
3085	struct io_ring_ctx *ctx;
3086	s64 inflight;
3087	DEFINE_WAIT(wait);
3088
3089	WARN_ON_ONCE(sqd && sqd->thread != current);
3090
3091	if (!current->io_uring)
3092		return;
3093	if (tctx->io_wq)
3094		io_wq_exit_start(tctx->io_wq);
3095
3096	atomic_inc(&tctx->in_idle);
3097	do {
3098		bool loop = false;
3099
3100		io_uring_drop_tctx_refs(current);
3101		/* read completions before cancelations */
3102		inflight = tctx_inflight(tctx, !cancel_all);
3103		if (!inflight)
3104			break;
3105
3106		if (!sqd) {
3107			struct io_tctx_node *node;
3108			unsigned long index;
3109
3110			xa_for_each(&tctx->xa, index, node) {
3111				/* sqpoll task will cancel all its requests */
3112				if (node->ctx->sq_data)
3113					continue;
3114				loop |= io_uring_try_cancel_requests(node->ctx,
3115							current, cancel_all);
3116			}
3117		} else {
3118			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
3119				loop |= io_uring_try_cancel_requests(ctx,
3120								     current,
3121								     cancel_all);
3122		}
3123
3124		if (loop) {
3125			cond_resched();
3126			continue;
3127		}
3128
3129		prepare_to_wait(&tctx->wait, &wait, TASK_INTERRUPTIBLE);
3130		io_run_task_work();
3131		io_uring_drop_tctx_refs(current);
3132
3133		/*
3134		 * If we've seen completions, retry without waiting. This
3135		 * avoids a race where a completion comes in before we did
3136		 * prepare_to_wait().
3137		 */
3138		if (inflight == tctx_inflight(tctx, !cancel_all))
3139			schedule();
3140		finish_wait(&tctx->wait, &wait);
3141	} while (1);
3142
3143	io_uring_clean_tctx(tctx);
3144	if (cancel_all) {
3145		/*
3146		 * We shouldn't run task_works after cancel, so just leave
3147		 * ->in_idle set for normal exit.
3148		 */
3149		atomic_dec(&tctx->in_idle);
3150		/* for exec all current's requests should be gone, kill tctx */
3151		__io_uring_free(current);
3152	}
3153}
3154
3155void __io_uring_cancel(bool cancel_all)
3156{
3157	io_uring_cancel_generic(cancel_all, NULL);
3158}
3159
3160static void *io_uring_validate_mmap_request(struct file *file,
3161					    loff_t pgoff, size_t sz)
3162{
3163	struct io_ring_ctx *ctx = file->private_data;
3164	loff_t offset = pgoff << PAGE_SHIFT;
3165	struct page *page;
3166	void *ptr;
3167
3168	switch (offset) {
3169	case IORING_OFF_SQ_RING:
3170	case IORING_OFF_CQ_RING:
3171		ptr = ctx->rings;
3172		break;
3173	case IORING_OFF_SQES:
3174		ptr = ctx->sq_sqes;
3175		break;
3176	default:
3177		return ERR_PTR(-EINVAL);
3178	}
3179
3180	page = virt_to_head_page(ptr);
3181	if (sz > page_size(page))
3182		return ERR_PTR(-EINVAL);
3183
3184	return ptr;
3185}
3186
3187#ifdef CONFIG_MMU
3188
3189static __cold int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3190{
3191	size_t sz = vma->vm_end - vma->vm_start;
3192	unsigned long pfn;
3193	void *ptr;
3194
3195	ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
3196	if (IS_ERR(ptr))
3197		return PTR_ERR(ptr);
3198
3199	pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3200	return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3201}
3202
3203#else /* !CONFIG_MMU */
3204
3205static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3206{
3207	return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
3208}
3209
3210static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
3211{
3212	return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
3213}
3214
3215static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
3216	unsigned long addr, unsigned long len,
3217	unsigned long pgoff, unsigned long flags)
3218{
3219	void *ptr;
3220
3221	ptr = io_uring_validate_mmap_request(file, pgoff, len);
3222	if (IS_ERR(ptr))
3223		return PTR_ERR(ptr);
3224
3225	return (unsigned long) ptr;
3226}
3227
3228#endif /* !CONFIG_MMU */
3229
3230static int io_validate_ext_arg(unsigned flags, const void __user *argp, size_t argsz)
3231{
3232	if (flags & IORING_ENTER_EXT_ARG) {
3233		struct io_uring_getevents_arg arg;
3234
3235		if (argsz != sizeof(arg))
3236			return -EINVAL;
3237		if (copy_from_user(&arg, argp, sizeof(arg)))
3238			return -EFAULT;
3239	}
3240	return 0;
3241}
3242
3243static int io_get_ext_arg(unsigned flags, const void __user *argp, size_t *argsz,
3244			  struct __kernel_timespec __user **ts,
3245			  const sigset_t __user **sig)
3246{
3247	struct io_uring_getevents_arg arg;
3248
3249	/*
3250	 * If EXT_ARG isn't set, then we have no timespec and the argp pointer
3251	 * is just a pointer to the sigset_t.
3252	 */
3253	if (!(flags & IORING_ENTER_EXT_ARG)) {
3254		*sig = (const sigset_t __user *) argp;
3255		*ts = NULL;
3256		return 0;
3257	}
3258
3259	/*
3260	 * EXT_ARG is set - ensure we agree on the size of it and copy in our
3261	 * timespec and sigset_t pointers if good.
3262	 */
3263	if (*argsz != sizeof(arg))
3264		return -EINVAL;
3265	if (copy_from_user(&arg, argp, sizeof(arg)))
3266		return -EFAULT;
3267	if (arg.pad)
3268		return -EINVAL;
3269	*sig = u64_to_user_ptr(arg.sigmask);
3270	*argsz = arg.sigmask_sz;
3271	*ts = u64_to_user_ptr(arg.ts);
3272	return 0;
3273}
3274
3275SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3276		u32, min_complete, u32, flags, const void __user *, argp,
3277		size_t, argsz)
3278{
3279	struct io_ring_ctx *ctx;
3280	struct fd f;
3281	long ret;
3282
3283	if (unlikely(flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP |
3284			       IORING_ENTER_SQ_WAIT | IORING_ENTER_EXT_ARG |
3285			       IORING_ENTER_REGISTERED_RING)))
3286		return -EINVAL;
3287
3288	/*
3289	 * Ring fd has been registered via IORING_REGISTER_RING_FDS, we
3290	 * need only dereference our task private array to find it.
3291	 */
3292	if (flags & IORING_ENTER_REGISTERED_RING) {
3293		struct io_uring_task *tctx = current->io_uring;
3294
3295		if (unlikely(!tctx || fd >= IO_RINGFD_REG_MAX))
3296			return -EINVAL;
3297		fd = array_index_nospec(fd, IO_RINGFD_REG_MAX);
3298		f.file = tctx->registered_rings[fd];
3299		f.flags = 0;
3300		if (unlikely(!f.file))
3301			return -EBADF;
3302	} else {
3303		f = fdget(fd);
3304		if (unlikely(!f.file))
3305			return -EBADF;
3306		ret = -EOPNOTSUPP;
3307		if (unlikely(!io_is_uring_fops(f.file)))
3308			goto out;
3309	}
3310
3311	ctx = f.file->private_data;
3312	ret = -EBADFD;
3313	if (unlikely(ctx->flags & IORING_SETUP_R_DISABLED))
3314		goto out;
3315
3316	/*
3317	 * For SQ polling, the thread will do all submissions and completions.
3318	 * Just return the requested submit count, and wake the thread if
3319	 * we were asked to.
3320	 */
3321	ret = 0;
3322	if (ctx->flags & IORING_SETUP_SQPOLL) {
3323		io_cqring_overflow_flush(ctx);
3324
3325		if (unlikely(ctx->sq_data->thread == NULL)) {
3326			ret = -EOWNERDEAD;
3327			goto out;
3328		}
3329		if (flags & IORING_ENTER_SQ_WAKEUP)
3330			wake_up(&ctx->sq_data->wait);
3331		if (flags & IORING_ENTER_SQ_WAIT) {
3332			ret = io_sqpoll_wait_sq(ctx);
3333			if (ret)
3334				goto out;
3335		}
3336		ret = to_submit;
3337	} else if (to_submit) {
3338		ret = io_uring_add_tctx_node(ctx);
3339		if (unlikely(ret))
3340			goto out;
3341
3342		mutex_lock(&ctx->uring_lock);
3343		ret = io_submit_sqes(ctx, to_submit);
3344		if (ret != to_submit) {
3345			mutex_unlock(&ctx->uring_lock);
3346			goto out;
3347		}
3348		if (flags & IORING_ENTER_GETEVENTS) {
3349			if (ctx->syscall_iopoll)
3350				goto iopoll_locked;
3351			/*
3352			 * Ignore errors, we'll soon call io_cqring_wait() and
3353			 * it should handle ownership problems if any.
3354			 */
3355			if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
3356				(void)io_run_local_work_locked(ctx);
3357		}
3358		mutex_unlock(&ctx->uring_lock);
3359	}
3360
3361	if (flags & IORING_ENTER_GETEVENTS) {
3362		int ret2;
3363
3364		if (ctx->syscall_iopoll) {
3365			/*
3366			 * We disallow the app entering submit/complete with
3367			 * polling, but we still need to lock the ring to
3368			 * prevent racing with polled issue that got punted to
3369			 * a workqueue.
3370			 */
3371			mutex_lock(&ctx->uring_lock);
3372iopoll_locked:
3373			ret2 = io_validate_ext_arg(flags, argp, argsz);
3374			if (likely(!ret2)) {
3375				min_complete = min(min_complete,
3376						   ctx->cq_entries);
3377				ret2 = io_iopoll_check(ctx, min_complete);
3378			}
3379			mutex_unlock(&ctx->uring_lock);
3380		} else {
3381			const sigset_t __user *sig;
3382			struct __kernel_timespec __user *ts;
3383
3384			ret2 = io_get_ext_arg(flags, argp, &argsz, &ts, &sig);
3385			if (likely(!ret2)) {
3386				min_complete = min(min_complete,
3387						   ctx->cq_entries);
3388				ret2 = io_cqring_wait(ctx, min_complete, sig,
3389						      argsz, ts);
3390			}
3391		}
3392
3393		if (!ret) {
3394			ret = ret2;
3395
3396			/*
3397			 * EBADR indicates that one or more CQE were dropped.
3398			 * Once the user has been informed we can clear the bit
3399			 * as they are obviously ok with those drops.
3400			 */
3401			if (unlikely(ret2 == -EBADR))
3402				clear_bit(IO_CHECK_CQ_DROPPED_BIT,
3403					  &ctx->check_cq);
3404		}
3405	}
3406out:
3407	fdput(f);
3408	return ret;
3409}
3410
3411static const struct file_operations io_uring_fops = {
3412	.release	= io_uring_release,
3413	.mmap		= io_uring_mmap,
3414#ifndef CONFIG_MMU
3415	.get_unmapped_area = io_uring_nommu_get_unmapped_area,
3416	.mmap_capabilities = io_uring_nommu_mmap_capabilities,
3417#endif
3418	.poll		= io_uring_poll,
3419#ifdef CONFIG_PROC_FS
3420	.show_fdinfo	= io_uring_show_fdinfo,
3421#endif
3422};
3423
3424bool io_is_uring_fops(struct file *file)
3425{
3426	return file->f_op == &io_uring_fops;
3427}
3428
3429static __cold int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3430					 struct io_uring_params *p)
3431{
3432	struct io_rings *rings;
3433	size_t size, sq_array_offset;
3434
3435	/* make sure these are sane, as we already accounted them */
3436	ctx->sq_entries = p->sq_entries;
3437	ctx->cq_entries = p->cq_entries;
3438
3439	size = rings_size(ctx, p->sq_entries, p->cq_entries, &sq_array_offset);
3440	if (size == SIZE_MAX)
3441		return -EOVERFLOW;
3442
3443	rings = io_mem_alloc(size);
3444	if (!rings)
3445		return -ENOMEM;
3446
3447	ctx->rings = rings;
3448	ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3449	rings->sq_ring_mask = p->sq_entries - 1;
3450	rings->cq_ring_mask = p->cq_entries - 1;
3451	rings->sq_ring_entries = p->sq_entries;
3452	rings->cq_ring_entries = p->cq_entries;
3453
3454	if (p->flags & IORING_SETUP_SQE128)
3455		size = array_size(2 * sizeof(struct io_uring_sqe), p->sq_entries);
3456	else
3457		size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3458	if (size == SIZE_MAX) {
3459		io_mem_free(ctx->rings);
3460		ctx->rings = NULL;
3461		return -EOVERFLOW;
3462	}
3463
3464	ctx->sq_sqes = io_mem_alloc(size);
3465	if (!ctx->sq_sqes) {
3466		io_mem_free(ctx->rings);
3467		ctx->rings = NULL;
3468		return -ENOMEM;
3469	}
3470
3471	return 0;
3472}
3473
3474static int io_uring_install_fd(struct io_ring_ctx *ctx, struct file *file)
3475{
3476	int ret, fd;
3477
3478	fd = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3479	if (fd < 0)
3480		return fd;
3481
3482	ret = __io_uring_add_tctx_node(ctx);
3483	if (ret) {
3484		put_unused_fd(fd);
3485		return ret;
3486	}
3487	fd_install(fd, file);
3488	return fd;
3489}
3490
3491/*
3492 * Allocate an anonymous fd, this is what constitutes the application
3493 * visible backing of an io_uring instance. The application mmaps this
3494 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3495 * we have to tie this fd to a socket for file garbage collection purposes.
3496 */
3497static struct file *io_uring_get_file(struct io_ring_ctx *ctx)
3498{
3499	struct file *file;
3500#if defined(CONFIG_UNIX)
3501	int ret;
3502
3503	ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3504				&ctx->ring_sock);
3505	if (ret)
3506		return ERR_PTR(ret);
3507#endif
3508
3509	file = anon_inode_getfile_secure("[io_uring]", &io_uring_fops, ctx,
3510					 O_RDWR | O_CLOEXEC, NULL);
3511#if defined(CONFIG_UNIX)
3512	if (IS_ERR(file)) {
3513		sock_release(ctx->ring_sock);
3514		ctx->ring_sock = NULL;
3515	} else {
3516		ctx->ring_sock->file = file;
3517	}
3518#endif
3519	return file;
3520}
3521
3522static __cold int io_uring_create(unsigned entries, struct io_uring_params *p,
3523				  struct io_uring_params __user *params)
3524{
3525	struct io_ring_ctx *ctx;
3526	struct file *file;
3527	int ret;
3528
3529	if (!entries)
3530		return -EINVAL;
3531	if (entries > IORING_MAX_ENTRIES) {
3532		if (!(p->flags & IORING_SETUP_CLAMP))
3533			return -EINVAL;
3534		entries = IORING_MAX_ENTRIES;
3535	}
3536
3537	/*
3538	 * Use twice as many entries for the CQ ring. It's possible for the
3539	 * application to drive a higher depth than the size of the SQ ring,
3540	 * since the sqes are only used at submission time. This allows for
3541	 * some flexibility in overcommitting a bit. If the application has
3542	 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
3543	 * of CQ ring entries manually.
3544	 */
3545	p->sq_entries = roundup_pow_of_two(entries);
3546	if (p->flags & IORING_SETUP_CQSIZE) {
3547		/*
3548		 * If IORING_SETUP_CQSIZE is set, we do the same roundup
3549		 * to a power-of-two, if it isn't already. We do NOT impose
3550		 * any cq vs sq ring sizing.
3551		 */
3552		if (!p->cq_entries)
3553			return -EINVAL;
3554		if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
3555			if (!(p->flags & IORING_SETUP_CLAMP))
3556				return -EINVAL;
3557			p->cq_entries = IORING_MAX_CQ_ENTRIES;
3558		}
3559		p->cq_entries = roundup_pow_of_two(p->cq_entries);
3560		if (p->cq_entries < p->sq_entries)
3561			return -EINVAL;
3562	} else {
3563		p->cq_entries = 2 * p->sq_entries;
3564	}
3565
3566	ctx = io_ring_ctx_alloc(p);
3567	if (!ctx)
3568		return -ENOMEM;
3569
3570	if ((ctx->flags & IORING_SETUP_DEFER_TASKRUN) &&
3571	    !(ctx->flags & IORING_SETUP_IOPOLL) &&
3572	    !(ctx->flags & IORING_SETUP_SQPOLL))
3573		ctx->task_complete = true;
3574
3575	/*
3576	 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
3577	 * space applications don't need to do io completion events
3578	 * polling again, they can rely on io_sq_thread to do polling
3579	 * work, which can reduce cpu usage and uring_lock contention.
3580	 */
3581	if (ctx->flags & IORING_SETUP_IOPOLL &&
3582	    !(ctx->flags & IORING_SETUP_SQPOLL))
3583		ctx->syscall_iopoll = 1;
3584
3585	ctx->compat = in_compat_syscall();
3586	if (!capable(CAP_IPC_LOCK))
3587		ctx->user = get_uid(current_user());
3588
3589	/*
3590	 * For SQPOLL, we just need a wakeup, always. For !SQPOLL, if
3591	 * COOP_TASKRUN is set, then IPIs are never needed by the app.
3592	 */
3593	ret = -EINVAL;
3594	if (ctx->flags & IORING_SETUP_SQPOLL) {
3595		/* IPI related flags don't make sense with SQPOLL */
3596		if (ctx->flags & (IORING_SETUP_COOP_TASKRUN |
3597				  IORING_SETUP_TASKRUN_FLAG |
3598				  IORING_SETUP_DEFER_TASKRUN))
3599			goto err;
3600		ctx->notify_method = TWA_SIGNAL_NO_IPI;
3601	} else if (ctx->flags & IORING_SETUP_COOP_TASKRUN) {
3602		ctx->notify_method = TWA_SIGNAL_NO_IPI;
3603	} else {
3604		if (ctx->flags & IORING_SETUP_TASKRUN_FLAG &&
3605		    !(ctx->flags & IORING_SETUP_DEFER_TASKRUN))
3606			goto err;
3607		ctx->notify_method = TWA_SIGNAL;
3608	}
3609
3610	/*
3611	 * For DEFER_TASKRUN we require the completion task to be the same as the
3612	 * submission task. This implies that there is only one submitter, so enforce
3613	 * that.
3614	 */
3615	if (ctx->flags & IORING_SETUP_DEFER_TASKRUN &&
3616	    !(ctx->flags & IORING_SETUP_SINGLE_ISSUER)) {
3617		goto err;
3618	}
3619
3620	/*
3621	 * This is just grabbed for accounting purposes. When a process exits,
3622	 * the mm is exited and dropped before the files, hence we need to hang
3623	 * on to this mm purely for the purposes of being able to unaccount
3624	 * memory (locked/pinned vm). It's not used for anything else.
3625	 */
3626	mmgrab(current->mm);
3627	ctx->mm_account = current->mm;
3628
3629	ret = io_allocate_scq_urings(ctx, p);
3630	if (ret)
3631		goto err;
3632
3633	ret = io_sq_offload_create(ctx, p);
3634	if (ret)
3635		goto err;
3636	/* always set a rsrc node */
3637	ret = io_rsrc_node_switch_start(ctx);
3638	if (ret)
3639		goto err;
3640	io_rsrc_node_switch(ctx, NULL);
3641
3642	memset(&p->sq_off, 0, sizeof(p->sq_off));
3643	p->sq_off.head = offsetof(struct io_rings, sq.head);
3644	p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3645	p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3646	p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3647	p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3648	p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3649	p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3650
3651	memset(&p->cq_off, 0, sizeof(p->cq_off));
3652	p->cq_off.head = offsetof(struct io_rings, cq.head);
3653	p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3654	p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3655	p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3656	p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3657	p->cq_off.cqes = offsetof(struct io_rings, cqes);
3658	p->cq_off.flags = offsetof(struct io_rings, cq_flags);
3659
3660	p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
3661			IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
3662			IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL |
3663			IORING_FEAT_POLL_32BITS | IORING_FEAT_SQPOLL_NONFIXED |
3664			IORING_FEAT_EXT_ARG | IORING_FEAT_NATIVE_WORKERS |
3665			IORING_FEAT_RSRC_TAGS | IORING_FEAT_CQE_SKIP |
3666			IORING_FEAT_LINKED_FILE;
3667
3668	if (copy_to_user(params, p, sizeof(*p))) {
3669		ret = -EFAULT;
3670		goto err;
3671	}
3672
3673	if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
3674	    && !(ctx->flags & IORING_SETUP_R_DISABLED))
3675		WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3676
3677	file = io_uring_get_file(ctx);
3678	if (IS_ERR(file)) {
3679		ret = PTR_ERR(file);
3680		goto err;
3681	}
3682
3683	/*
3684	 * Install ring fd as the very last thing, so we don't risk someone
3685	 * having closed it before we finish setup
3686	 */
3687	ret = io_uring_install_fd(ctx, file);
3688	if (ret < 0) {
3689		/* fput will clean it up */
3690		fput(file);
3691		return ret;
3692	}
3693
3694	trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
3695	return ret;
3696err:
3697	io_ring_ctx_wait_and_kill(ctx);
3698	return ret;
3699}
3700
3701/*
3702 * Sets up an aio uring context, and returns the fd. Applications asks for a
3703 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3704 * params structure passed in.
3705 */
3706static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3707{
3708	struct io_uring_params p;
3709	int i;
3710
3711	if (copy_from_user(&p, params, sizeof(p)))
3712		return -EFAULT;
3713	for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3714		if (p.resv[i])
3715			return -EINVAL;
3716	}
3717
3718	if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3719			IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
3720			IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ |
3721			IORING_SETUP_R_DISABLED | IORING_SETUP_SUBMIT_ALL |
3722			IORING_SETUP_COOP_TASKRUN | IORING_SETUP_TASKRUN_FLAG |
3723			IORING_SETUP_SQE128 | IORING_SETUP_CQE32 |
3724			IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN))
3725		return -EINVAL;
3726
3727	return io_uring_create(entries, &p, params);
3728}
3729
3730SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3731		struct io_uring_params __user *, params)
3732{
3733	return io_uring_setup(entries, params);
3734}
3735
3736static __cold int io_probe(struct io_ring_ctx *ctx, void __user *arg,
3737			   unsigned nr_args)
3738{
3739	struct io_uring_probe *p;
3740	size_t size;
3741	int i, ret;
3742
3743	size = struct_size(p, ops, nr_args);
3744	if (size == SIZE_MAX)
3745		return -EOVERFLOW;
3746	p = kzalloc(size, GFP_KERNEL);
3747	if (!p)
3748		return -ENOMEM;
3749
3750	ret = -EFAULT;
3751	if (copy_from_user(p, arg, size))
3752		goto out;
3753	ret = -EINVAL;
3754	if (memchr_inv(p, 0, size))
3755		goto out;
3756
3757	p->last_op = IORING_OP_LAST - 1;
3758	if (nr_args > IORING_OP_LAST)
3759		nr_args = IORING_OP_LAST;
3760
3761	for (i = 0; i < nr_args; i++) {
3762		p->ops[i].op = i;
3763		if (!io_op_defs[i].not_supported)
3764			p->ops[i].flags = IO_URING_OP_SUPPORTED;
3765	}
3766	p->ops_len = i;
3767
3768	ret = 0;
3769	if (copy_to_user(arg, p, size))
3770		ret = -EFAULT;
3771out:
3772	kfree(p);
3773	return ret;
3774}
3775
3776static int io_register_personality(struct io_ring_ctx *ctx)
3777{
3778	const struct cred *creds;
3779	u32 id;
3780	int ret;
3781
3782	creds = get_current_cred();
3783
3784	ret = xa_alloc_cyclic(&ctx->personalities, &id, (void *)creds,
3785			XA_LIMIT(0, USHRT_MAX), &ctx->pers_next, GFP_KERNEL);
3786	if (ret < 0) {
3787		put_cred(creds);
3788		return ret;
3789	}
3790	return id;
3791}
3792
3793static __cold int io_register_restrictions(struct io_ring_ctx *ctx,
3794					   void __user *arg, unsigned int nr_args)
3795{
3796	struct io_uring_restriction *res;
3797	size_t size;
3798	int i, ret;
3799
3800	/* Restrictions allowed only if rings started disabled */
3801	if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3802		return -EBADFD;
3803
3804	/* We allow only a single restrictions registration */
3805	if (ctx->restrictions.registered)
3806		return -EBUSY;
3807
3808	if (!arg || nr_args > IORING_MAX_RESTRICTIONS)
3809		return -EINVAL;
3810
3811	size = array_size(nr_args, sizeof(*res));
3812	if (size == SIZE_MAX)
3813		return -EOVERFLOW;
3814
3815	res = memdup_user(arg, size);
3816	if (IS_ERR(res))
3817		return PTR_ERR(res);
3818
3819	ret = 0;
3820
3821	for (i = 0; i < nr_args; i++) {
3822		switch (res[i].opcode) {
3823		case IORING_RESTRICTION_REGISTER_OP:
3824			if (res[i].register_op >= IORING_REGISTER_LAST) {
3825				ret = -EINVAL;
3826				goto out;
3827			}
3828
3829			__set_bit(res[i].register_op,
3830				  ctx->restrictions.register_op);
3831			break;
3832		case IORING_RESTRICTION_SQE_OP:
3833			if (res[i].sqe_op >= IORING_OP_LAST) {
3834				ret = -EINVAL;
3835				goto out;
3836			}
3837
3838			__set_bit(res[i].sqe_op, ctx->restrictions.sqe_op);
3839			break;
3840		case IORING_RESTRICTION_SQE_FLAGS_ALLOWED:
3841			ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags;
3842			break;
3843		case IORING_RESTRICTION_SQE_FLAGS_REQUIRED:
3844			ctx->restrictions.sqe_flags_required = res[i].sqe_flags;
3845			break;
3846		default:
3847			ret = -EINVAL;
3848			goto out;
3849		}
3850	}
3851
3852out:
3853	/* Reset all restrictions if an error happened */
3854	if (ret != 0)
3855		memset(&ctx->restrictions, 0, sizeof(ctx->restrictions));
3856	else
3857		ctx->restrictions.registered = true;
3858
3859	kfree(res);
3860	return ret;
3861}
3862
3863static int io_register_enable_rings(struct io_ring_ctx *ctx)
3864{
3865	if (!(ctx->flags & IORING_SETUP_R_DISABLED))
3866		return -EBADFD;
3867
3868	if (ctx->flags & IORING_SETUP_SINGLE_ISSUER && !ctx->submitter_task)
3869		WRITE_ONCE(ctx->submitter_task, get_task_struct(current));
3870
3871	if (ctx->restrictions.registered)
3872		ctx->restricted = 1;
3873
3874	ctx->flags &= ~IORING_SETUP_R_DISABLED;
3875	if (ctx->sq_data && wq_has_sleeper(&ctx->sq_data->wait))
3876		wake_up(&ctx->sq_data->wait);
3877	return 0;
3878}
3879
3880static __cold int io_register_iowq_aff(struct io_ring_ctx *ctx,
3881				       void __user *arg, unsigned len)
3882{
3883	struct io_uring_task *tctx = current->io_uring;
3884	cpumask_var_t new_mask;
3885	int ret;
3886
3887	if (!tctx || !tctx->io_wq)
3888		return -EINVAL;
3889
3890	if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3891		return -ENOMEM;
3892
3893	cpumask_clear(new_mask);
3894	if (len > cpumask_size())
3895		len = cpumask_size();
3896
3897	if (in_compat_syscall()) {
3898		ret = compat_get_bitmap(cpumask_bits(new_mask),
3899					(const compat_ulong_t __user *)arg,
3900					len * 8 /* CHAR_BIT */);
3901	} else {
3902		ret = copy_from_user(new_mask, arg, len);
3903	}
3904
3905	if (ret) {
3906		free_cpumask_var(new_mask);
3907		return -EFAULT;
3908	}
3909
3910	ret = io_wq_cpu_affinity(tctx->io_wq, new_mask);
3911	free_cpumask_var(new_mask);
3912	return ret;
3913}
3914
3915static __cold int io_unregister_iowq_aff(struct io_ring_ctx *ctx)
3916{
3917	struct io_uring_task *tctx = current->io_uring;
3918
3919	if (!tctx || !tctx->io_wq)
3920		return -EINVAL;
3921
3922	return io_wq_cpu_affinity(tctx->io_wq, NULL);
3923}
3924
3925static __cold int io_register_iowq_max_workers(struct io_ring_ctx *ctx,
3926					       void __user *arg)
3927	__must_hold(&ctx->uring_lock)
3928{
3929	struct io_tctx_node *node;
3930	struct io_uring_task *tctx = NULL;
3931	struct io_sq_data *sqd = NULL;
3932	__u32 new_count[2];
3933	int i, ret;
3934
3935	if (copy_from_user(new_count, arg, sizeof(new_count)))
3936		return -EFAULT;
3937	for (i = 0; i < ARRAY_SIZE(new_count); i++)
3938		if (new_count[i] > INT_MAX)
3939			return -EINVAL;
3940
3941	if (ctx->flags & IORING_SETUP_SQPOLL) {
3942		sqd = ctx->sq_data;
3943		if (sqd) {
3944			/*
3945			 * Observe the correct sqd->lock -> ctx->uring_lock
3946			 * ordering. Fine to drop uring_lock here, we hold
3947			 * a ref to the ctx.
3948			 */
3949			refcount_inc(&sqd->refs);
3950			mutex_unlock(&ctx->uring_lock);
3951			mutex_lock(&sqd->lock);
3952			mutex_lock(&ctx->uring_lock);
3953			if (sqd->thread)
3954				tctx = sqd->thread->io_uring;
3955		}
3956	} else {
3957		tctx = current->io_uring;
3958	}
3959
3960	BUILD_BUG_ON(sizeof(new_count) != sizeof(ctx->iowq_limits));
3961
3962	for (i = 0; i < ARRAY_SIZE(new_count); i++)
3963		if (new_count[i])
3964			ctx->iowq_limits[i] = new_count[i];
3965	ctx->iowq_limits_set = true;
3966
3967	if (tctx && tctx->io_wq) {
3968		ret = io_wq_max_workers(tctx->io_wq, new_count);
3969		if (ret)
3970			goto err;
3971	} else {
3972		memset(new_count, 0, sizeof(new_count));
3973	}
3974
3975	if (sqd) {
3976		mutex_unlock(&sqd->lock);
3977		io_put_sq_data(sqd);
3978	}
3979
3980	if (copy_to_user(arg, new_count, sizeof(new_count)))
3981		return -EFAULT;
3982
3983	/* that's it for SQPOLL, only the SQPOLL task creates requests */
3984	if (sqd)
3985		return 0;
3986
3987	/* now propagate the restriction to all registered users */
3988	list_for_each_entry(node, &ctx->tctx_list, ctx_node) {
3989		struct io_uring_task *tctx = node->task->io_uring;
3990
3991		if (WARN_ON_ONCE(!tctx->io_wq))
3992			continue;
3993
3994		for (i = 0; i < ARRAY_SIZE(new_count); i++)
3995			new_count[i] = ctx->iowq_limits[i];
3996		/* ignore errors, it always returns zero anyway */
3997		(void)io_wq_max_workers(tctx->io_wq, new_count);
3998	}
3999	return 0;
4000err:
4001	if (sqd) {
4002		mutex_unlock(&sqd->lock);
4003		io_put_sq_data(sqd);
4004	}
4005	return ret;
4006}
4007
4008static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4009			       void __user *arg, unsigned nr_args)
4010	__releases(ctx->uring_lock)
4011	__acquires(ctx->uring_lock)
4012{
4013	int ret;
4014
4015	/*
4016	 * We don't quiesce the refs for register anymore and so it can't be
4017	 * dying as we're holding a file ref here.
4018	 */
4019	if (WARN_ON_ONCE(percpu_ref_is_dying(&ctx->refs)))
4020		return -ENXIO;
4021
4022	if (ctx->submitter_task && ctx->submitter_task != current)
4023		return -EEXIST;
4024
4025	if (ctx->restricted) {
4026		opcode = array_index_nospec(opcode, IORING_REGISTER_LAST);
4027		if (!test_bit(opcode, ctx->restrictions.register_op))
4028			return -EACCES;
4029	}
4030
4031	switch (opcode) {
4032	case IORING_REGISTER_BUFFERS:
4033		ret = -EFAULT;
4034		if (!arg)
4035			break;
4036		ret = io_sqe_buffers_register(ctx, arg, nr_args, NULL);
4037		break;
4038	case IORING_UNREGISTER_BUFFERS:
4039		ret = -EINVAL;
4040		if (arg || nr_args)
4041			break;
4042		ret = io_sqe_buffers_unregister(ctx);
4043		break;
4044	case IORING_REGISTER_FILES:
4045		ret = -EFAULT;
4046		if (!arg)
4047			break;
4048		ret = io_sqe_files_register(ctx, arg, nr_args, NULL);
4049		break;
4050	case IORING_UNREGISTER_FILES:
4051		ret = -EINVAL;
4052		if (arg || nr_args)
4053			break;
4054		ret = io_sqe_files_unregister(ctx);
4055		break;
4056	case IORING_REGISTER_FILES_UPDATE:
4057		ret = io_register_files_update(ctx, arg, nr_args);
4058		break;
4059	case IORING_REGISTER_EVENTFD:
4060		ret = -EINVAL;
4061		if (nr_args != 1)
4062			break;
4063		ret = io_eventfd_register(ctx, arg, 0);
4064		break;
4065	case IORING_REGISTER_EVENTFD_ASYNC:
4066		ret = -EINVAL;
4067		if (nr_args != 1)
4068			break;
4069		ret = io_eventfd_register(ctx, arg, 1);
4070		break;
4071	case IORING_UNREGISTER_EVENTFD:
4072		ret = -EINVAL;
4073		if (arg || nr_args)
4074			break;
4075		ret = io_eventfd_unregister(ctx);
4076		break;
4077	case IORING_REGISTER_PROBE:
4078		ret = -EINVAL;
4079		if (!arg || nr_args > 256)
4080			break;
4081		ret = io_probe(ctx, arg, nr_args);
4082		break;
4083	case IORING_REGISTER_PERSONALITY:
4084		ret = -EINVAL;
4085		if (arg || nr_args)
4086			break;
4087		ret = io_register_personality(ctx);
4088		break;
4089	case IORING_UNREGISTER_PERSONALITY:
4090		ret = -EINVAL;
4091		if (arg)
4092			break;
4093		ret = io_unregister_personality(ctx, nr_args);
4094		break;
4095	case IORING_REGISTER_ENABLE_RINGS:
4096		ret = -EINVAL;
4097		if (arg || nr_args)
4098			break;
4099		ret = io_register_enable_rings(ctx);
4100		break;
4101	case IORING_REGISTER_RESTRICTIONS:
4102		ret = io_register_restrictions(ctx, arg, nr_args);
4103		break;
4104	case IORING_REGISTER_FILES2:
4105		ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_FILE);
4106		break;
4107	case IORING_REGISTER_FILES_UPDATE2:
4108		ret = io_register_rsrc_update(ctx, arg, nr_args,
4109					      IORING_RSRC_FILE);
4110		break;
4111	case IORING_REGISTER_BUFFERS2:
4112		ret = io_register_rsrc(ctx, arg, nr_args, IORING_RSRC_BUFFER);
4113		break;
4114	case IORING_REGISTER_BUFFERS_UPDATE:
4115		ret = io_register_rsrc_update(ctx, arg, nr_args,
4116					      IORING_RSRC_BUFFER);
4117		break;
4118	case IORING_REGISTER_IOWQ_AFF:
4119		ret = -EINVAL;
4120		if (!arg || !nr_args)
4121			break;
4122		ret = io_register_iowq_aff(ctx, arg, nr_args);
4123		break;
4124	case IORING_UNREGISTER_IOWQ_AFF:
4125		ret = -EINVAL;
4126		if (arg || nr_args)
4127			break;
4128		ret = io_unregister_iowq_aff(ctx);
4129		break;
4130	case IORING_REGISTER_IOWQ_MAX_WORKERS:
4131		ret = -EINVAL;
4132		if (!arg || nr_args != 2)
4133			break;
4134		ret = io_register_iowq_max_workers(ctx, arg);
4135		break;
4136	case IORING_REGISTER_RING_FDS:
4137		ret = io_ringfd_register(ctx, arg, nr_args);
4138		break;
4139	case IORING_UNREGISTER_RING_FDS:
4140		ret = io_ringfd_unregister(ctx, arg, nr_args);
4141		break;
4142	case IORING_REGISTER_PBUF_RING:
4143		ret = -EINVAL;
4144		if (!arg || nr_args != 1)
4145			break;
4146		ret = io_register_pbuf_ring(ctx, arg);
4147		break;
4148	case IORING_UNREGISTER_PBUF_RING:
4149		ret = -EINVAL;
4150		if (!arg || nr_args != 1)
4151			break;
4152		ret = io_unregister_pbuf_ring(ctx, arg);
4153		break;
4154	case IORING_REGISTER_SYNC_CANCEL:
4155		ret = -EINVAL;
4156		if (!arg || nr_args != 1)
4157			break;
4158		ret = io_sync_cancel(ctx, arg);
4159		break;
4160	case IORING_REGISTER_FILE_ALLOC_RANGE:
4161		ret = -EINVAL;
4162		if (!arg || nr_args)
4163			break;
4164		ret = io_register_file_alloc_range(ctx, arg);
4165		break;
4166	default:
4167		ret = -EINVAL;
4168		break;
4169	}
4170
4171	return ret;
4172}
4173
4174SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
4175		void __user *, arg, unsigned int, nr_args)
4176{
4177	struct io_ring_ctx *ctx;
4178	long ret = -EBADF;
4179	struct fd f;
4180
4181	if (opcode >= IORING_REGISTER_LAST)
4182		return -EINVAL;
4183
4184	f = fdget(fd);
4185	if (!f.file)
4186		return -EBADF;
4187
4188	ret = -EOPNOTSUPP;
4189	if (!io_is_uring_fops(f.file))
4190		goto out_fput;
4191
4192	ctx = f.file->private_data;
4193
4194	mutex_lock(&ctx->uring_lock);
4195	ret = __io_uring_register(ctx, opcode, arg, nr_args);
4196	mutex_unlock(&ctx->uring_lock);
4197	trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, ret);
4198out_fput:
4199	fdput(f);
4200	return ret;
4201}
4202
4203static int __init io_uring_init(void)
4204{
4205#define __BUILD_BUG_VERIFY_OFFSET_SIZE(stype, eoffset, esize, ename) do { \
4206	BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
4207	BUILD_BUG_ON(sizeof_field(stype, ename) != esize); \
4208} while (0)
4209
4210#define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
4211	__BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, sizeof(etype), ename)
4212#define BUILD_BUG_SQE_ELEM_SIZE(eoffset, esize, ename) \
4213	__BUILD_BUG_VERIFY_OFFSET_SIZE(struct io_uring_sqe, eoffset, esize, ename)
4214	BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
4215	BUILD_BUG_SQE_ELEM(0,  __u8,   opcode);
4216	BUILD_BUG_SQE_ELEM(1,  __u8,   flags);
4217	BUILD_BUG_SQE_ELEM(2,  __u16,  ioprio);
4218	BUILD_BUG_SQE_ELEM(4,  __s32,  fd);
4219	BUILD_BUG_SQE_ELEM(8,  __u64,  off);
4220	BUILD_BUG_SQE_ELEM(8,  __u64,  addr2);
4221	BUILD_BUG_SQE_ELEM(8,  __u32,  cmd_op);
4222	BUILD_BUG_SQE_ELEM(12, __u32, __pad1);
4223	BUILD_BUG_SQE_ELEM(16, __u64,  addr);
4224	BUILD_BUG_SQE_ELEM(16, __u64,  splice_off_in);
4225	BUILD_BUG_SQE_ELEM(24, __u32,  len);
4226	BUILD_BUG_SQE_ELEM(28,     __kernel_rwf_t, rw_flags);
4227	BUILD_BUG_SQE_ELEM(28, /* compat */   int, rw_flags);
4228	BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
4229	BUILD_BUG_SQE_ELEM(28, __u32,  fsync_flags);
4230	BUILD_BUG_SQE_ELEM(28, /* compat */ __u16,  poll_events);
4231	BUILD_BUG_SQE_ELEM(28, __u32,  poll32_events);
4232	BUILD_BUG_SQE_ELEM(28, __u32,  sync_range_flags);
4233	BUILD_BUG_SQE_ELEM(28, __u32,  msg_flags);
4234	BUILD_BUG_SQE_ELEM(28, __u32,  timeout_flags);
4235	BUILD_BUG_SQE_ELEM(28, __u32,  accept_flags);
4236	BUILD_BUG_SQE_ELEM(28, __u32,  cancel_flags);
4237	BUILD_BUG_SQE_ELEM(28, __u32,  open_flags);
4238	BUILD_BUG_SQE_ELEM(28, __u32,  statx_flags);
4239	BUILD_BUG_SQE_ELEM(28, __u32,  fadvise_advice);
4240	BUILD_BUG_SQE_ELEM(28, __u32,  splice_flags);
4241	BUILD_BUG_SQE_ELEM(28, __u32,  rename_flags);
4242	BUILD_BUG_SQE_ELEM(28, __u32,  unlink_flags);
4243	BUILD_BUG_SQE_ELEM(28, __u32,  hardlink_flags);
4244	BUILD_BUG_SQE_ELEM(28, __u32,  xattr_flags);
4245	BUILD_BUG_SQE_ELEM(28, __u32,  msg_ring_flags);
4246	BUILD_BUG_SQE_ELEM(32, __u64,  user_data);
4247	BUILD_BUG_SQE_ELEM(40, __u16,  buf_index);
4248	BUILD_BUG_SQE_ELEM(40, __u16,  buf_group);
4249	BUILD_BUG_SQE_ELEM(42, __u16,  personality);
4250	BUILD_BUG_SQE_ELEM(44, __s32,  splice_fd_in);
4251	BUILD_BUG_SQE_ELEM(44, __u32,  file_index);
4252	BUILD_BUG_SQE_ELEM(44, __u16,  addr_len);
4253	BUILD_BUG_SQE_ELEM(46, __u16,  __pad3[0]);
4254	BUILD_BUG_SQE_ELEM(48, __u64,  addr3);
4255	BUILD_BUG_SQE_ELEM_SIZE(48, 0, cmd);
4256	BUILD_BUG_SQE_ELEM(56, __u64,  __pad2);
4257
4258	BUILD_BUG_ON(sizeof(struct io_uring_files_update) !=
4259		     sizeof(struct io_uring_rsrc_update));
4260	BUILD_BUG_ON(sizeof(struct io_uring_rsrc_update) >
4261		     sizeof(struct io_uring_rsrc_update2));
4262
4263	/* ->buf_index is u16 */
4264	BUILD_BUG_ON(offsetof(struct io_uring_buf_ring, bufs) != 0);
4265	BUILD_BUG_ON(offsetof(struct io_uring_buf, resv) !=
4266		     offsetof(struct io_uring_buf_ring, tail));
4267
4268	/* should fit into one byte */
4269	BUILD_BUG_ON(SQE_VALID_FLAGS >= (1 << 8));
4270	BUILD_BUG_ON(SQE_COMMON_FLAGS >= (1 << 8));
4271	BUILD_BUG_ON((SQE_VALID_FLAGS | SQE_COMMON_FLAGS) != SQE_VALID_FLAGS);
4272
4273	BUILD_BUG_ON(__REQ_F_LAST_BIT > 8 * sizeof(int));
4274
4275	BUILD_BUG_ON(sizeof(atomic_t) != sizeof(u32));
4276
4277	io_uring_optable_init();
4278
4279	req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
4280				SLAB_ACCOUNT);
4281	return 0;
4282};
4283__initcall(io_uring_init);