xref: /xnu-12377.41.6/bsd/kern/kern_aio.c (revision bbb1b6f9e71b8cdde6e5cd6f4841f207dee3d828)

/*
 * Copyright (c) 2003-2024 Apple Inc. All rights reserved.
 *
 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
 *
 * This file contains Original Code and/or Modifications of Original Code
 * as defined in and that are subject to the Apple Public Source License
 * Version 2.0 (the 'License'). You may not use this file except in
 * compliance with the License. The rights granted to you under the License
 * may not be used to create, or enable the creation or redistribution of,
 * unlawful or unlicensed copies of an Apple operating system, or to
 * circumvent, violate, or enable the circumvention or violation of, any
 * terms of an Apple operating system software license agreement.
 *
 * Please obtain a copy of the License at
 * http://www.opensource.apple.com/apsl/ and read it before using this file.
 *
 * The Original Code and all software distributed under the License are
 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
 * Please see the License for the specific language governing rights and
 * limitations under the License.
 *
 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
 */

/*
 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
 */

#include <sys/systm.h>
#include <sys/fcntl.h>
#include <sys/file_internal.h>
#include <sys/filedesc.h>
#include <sys/kdebug.h>
#include <sys/kernel.h>
#include <sys/vnode_internal.h>
#include <sys/kauth.h>
#include <sys/mount_internal.h>
#include <sys/param.h>
#include <sys/proc_internal.h>
#include <sys/sysctl.h>
#include <sys/unistd.h>
#include <sys/user.h>

#include <sys/aio_kern.h>
#include <sys/sysproto.h>

#include <machine/limits.h>

#include <mach/mach_types.h>
#include <kern/kern_types.h>
#include <kern/waitq.h>
#include <kern/zalloc.h>
#include <kern/task.h>
#include <kern/sched_prim.h>
#include <kern/ast.h>

#include <vm/vm_map_xnu.h>

#include <os/refcnt.h>

#include <kern/thread.h>
#include <kern/policy_internal.h>
#include <pthread/workqueue_internal.h>

#if 0
#undef KERNEL_DEBUG
#define KERNEL_DEBUG KERNEL_DEBUG_CONSTANT
#endif

#define AIO_work_queued                 1
#define AIO_worker_wake                 2
#define AIO_completion_sig              3
#define AIO_completion_kevent           4
#define AIO_completion_cleanup_wait     5
#define AIO_completion_cleanup_wake     6
#define AIO_completion_suspend_wake     7
#define AIO_cancel                      10
#define AIO_cancel_async_workq          11
#define AIO_cancel_sync_workq           12
#define AIO_cancel_activeq              13
#define AIO_cancel_doneq                14
#define AIO_fsync                       20
#define AIO_fsync_delay                 21
#define AIO_read                        30
#define AIO_write                       40
#define AIO_listio                      50
#define AIO_error                       60
#define AIO_error_val                   61
#define AIO_error_activeq               62
#define AIO_error_workq                 63
#define AIO_return                      70
#define AIO_return_val                  71
#define AIO_return_activeq              72
#define AIO_return_workq                73
#define AIO_exec                        80
#define AIO_exit                        90
#define AIO_exit_sleep                  91
#define AIO_close                       100
#define AIO_close_sleep                 101
#define AIO_suspend                     110
#define AIO_suspend_sleep               111
#define AIO_worker_thread               120
#define AIO_register_kevent             130
#define AIO_WQ_process_entry            140
#define AIO_WQ_aio_thread_create        141
#define AIO_WQ_aio_thread_terminate     142
#define AIO_WQ_aio_death_call           143
#define AIO_WQ_aio_thread_park          144
#define AIO_WQ_aio_select_req           145
#define AIO_WQ_aio_thread_create_failed 146
#define AIO_WQ_aio_thread_wakeup        147

static TUNABLE(uint32_t, bootarg_aio_new_workq, "aio_new_workq", 1);

__options_decl(aio_entry_flags_t, uint32_t, {
	AIO_READ        = 0x00000001, /* a read */
	AIO_WRITE       = 0x00000002, /* a write */
	AIO_FSYNC       = 0x00000004, /* aio_fsync with op = O_SYNC */
	AIO_DSYNC       = 0x00000008, /* aio_fsync with op = O_DSYNC (not supported yet) */
	AIO_LIO         = 0x00000010, /* lio_listio generated IO */
	AIO_LIO_WAIT    = 0x00000020, /* lio_listio is waiting on the leader */

	AIO_COMPLETED   = 0x00000100, /* request has completed */
	AIO_CANCELLED   = 0x00000200, /* request has been cancelled */

	/*
	 * These flags mean that this entry is blocking either:
	 * - close (AIO_CLOSE_WAIT)
	 * - exit or exec (AIO_EXIT_WAIT)
	 *
	 * These flags are mutually exclusive, and the AIO_EXIT_WAIT variant
	 * will also neuter notifications in do_aio_completion_and_unlock().
	 */
	AIO_CLOSE_WAIT  = 0x00004000,
	AIO_EXIT_WAIT   = 0x00008000,
});

/*! @struct aio_workq_entry
 *
 * @discussion
 * This represents a piece of aio/lio work.
 *
 * The ownership rules go as follows:
 *
 * - the "proc" owns one refcount on the entry (from creation), while it is
 *   enqueued on the aio_activeq and then the aio_doneq.
 *
 *   either aio_return() (user read the status) or _aio_exit() (the process
 *   died) will dequeue the entry and consume this ref.
 *
 * - the async workqueue owns one refcount once the work is submitted,
 *   which is consumed in do_aio_completion_and_unlock().
 *
 *   This ref protects the entry for the the end of
 *   do_aio_completion_and_unlock() (when signal delivery happens).
 *
 * - lio_listio() for batches picks one of the entries to be the "leader"
 *   of the batch. Each work item will have a refcount on its leader
 *   so that the accounting of the batch completion can be done on the leader
 *   (to be able to decrement lio_pending).
 *
 *   This ref is consumed in do_aio_completion_and_unlock() as well.
 *
 * - lastly, in lio_listio() when the LIO_WAIT behavior is requested,
 *   an extra ref is taken in this syscall as it needs to keep accessing
 *   the leader "lio_pending" field until it hits 0.
 */
struct aio_workq_entry {
	/* queue lock */
	TAILQ_ENTRY(aio_workq_entry)    aio_workq_link;

	/* Proc lock */
	TAILQ_ENTRY(aio_workq_entry)    aio_proc_link;  /* p_aio_activeq or p_aio_doneq */
	user_ssize_t                    returnval;      /* return value from read / write request */
	errno_t                         errorval;       /* error value from read / write request */
	os_refcnt_t                     aio_refcount;
	aio_entry_flags_t               flags;

	int                             lio_pending;    /* pending I/Os in lio group, only on leader */
	struct aio_workq_entry         *lio_leader;     /* pointer to the lio leader, can be self */

	/* Initialized and never changed, safe to access */
	struct proc                    *procp;          /* user proc that queued this request */
	user_addr_t                     uaiocbp;        /* pointer passed in from user land */
	struct user_aiocb               aiocb;          /* copy of aiocb from user land */
	struct vfs_context              context;        /* context which enqueued the request */

	/* Initialized, and possibly freed by aio_work_thread() or at free if cancelled */
	vm_map_t                        aio_map;        /* user land map we have a reference to */
};

/*
 * aio requests queue up on the aio_async_workq or lio_sync_workq (for
 * lio_listio LIO_WAIT).  Requests then move to the per process aio_activeq
 * (proc.aio_activeq) when one of our worker threads start the IO.
 * And finally, requests move to the per process aio_doneq (proc.aio_doneq)
 * when the IO request completes.  The request remains on aio_doneq until
 * user process calls aio_return or the process exits, either way that is our
 * trigger to release aio resources.
 */
typedef struct aio_workq   {
	TAILQ_HEAD(, aio_workq_entry)   aioq_entries;
	lck_spin_t                      aioq_lock;
	struct waitq                    aioq_waitq;
} *aio_workq_t;

#define AIO_NUM_WORK_QUEUES 1
struct aio_anchor_cb {
	os_atomic(int)          aio_total_count;        /* total extant entries */

	/* Hash table of queues here */
	int                     aio_num_workqs;
	struct aio_workq        aio_async_workqs[AIO_NUM_WORK_QUEUES];
};
typedef struct aio_anchor_cb aio_anchor_cb;


/* New per process workqueue */
#define WORKQUEUE_AIO_MAXTHREADS            16

TAILQ_HEAD(workq_aio_uthread_head, uthread);

typedef struct workq_aio_s {
	thread_call_t   wa_death_call;
	struct workq_aio_uthread_head wa_thrunlist;
	struct workq_aio_uthread_head wa_thidlelist;
	TAILQ_HEAD(, aio_workq_entry) wa_aioq_entries;
	proc_t wa_proc;
	workq_state_flags_t _Atomic wa_flags;
	uint16_t wa_nthreads;
	uint16_t wa_thidlecount;
	uint16_t wa_thdying_count;
} workq_aio_s, *workq_aio_t;

struct aio_workq_usec_var {
	uint32_t usecs;
	uint64_t abstime;
};

static int aio_workq_sysctl_handle_usecs SYSCTL_HANDLER_ARGS;

#define AIO_WORKQ_SYSCTL_USECS(var, init) \
	        static struct aio_workq_usec_var var = { .usecs = (init) }; \
	        SYSCTL_OID(_kern, OID_AUTO, var##_usecs, \
	                        CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, &(var), 0, \
	                        aio_workq_sysctl_handle_usecs, "I", "")

AIO_WORKQ_SYSCTL_USECS(aio_wq_reduce_pool_window, WQ_REDUCE_POOL_WINDOW_USECS);

#define WQ_AIO_TRACE(x, wq, a, b, c, d) \
	        ({ KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_AIO, (x)),\
	        proc_getpid((wq)->wa_proc), (a), (b), (c), (d)); })

#define WQ_AIO_TRACE_WQ(x, wq) \
	        ({ KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_AIO, (x)),\
	        proc_getpid((wq)->wa_proc),\
	        (uintptr_t)thread_tid(current_thread()),\
	        (wq)->wa_nthreads, (wq)->wa_thidlecount, (wq)->wa_thdying_count); })

/*
 * Notes on aio sleep / wake channels.
 * We currently pick a couple fields within the proc structure that will allow
 * us sleep channels that currently do not collide with any other kernel routines.
 * At this time, for binary compatibility reasons, we cannot create new proc fields.
 */
#define AIO_SUSPEND_SLEEP_CHAN  p_aio_activeq
#define AIO_CLEANUP_SLEEP_CHAN  p_aio_total_count

#define ASSERT_AIO_FROM_PROC(aiop, theproc)     \
	if ((aiop)->procp != (theproc)) {       \
	        panic("AIO on a proc list that does not belong to that proc."); \
	}

extern kern_return_t thread_terminate(thread_t);

/*
 *  LOCAL PROTOTYPES
 */
static void             aio_proc_lock(proc_t procp);
static void             aio_proc_lock_spin(proc_t procp);
static void             aio_proc_unlock(proc_t procp);
static lck_mtx_t       *aio_proc_mutex(proc_t procp);
static bool             aio_has_active_requests_for_process(proc_t procp);
static bool             aio_proc_has_active_requests_for_file(proc_t procp, int fd);
static boolean_t        is_already_queued(proc_t procp, user_addr_t aiocbp);

static aio_workq_t      aio_entry_workq(aio_workq_entry *entryp);
static void             aio_workq_remove_entry_locked(aio_workq_t queue, aio_workq_entry *entryp);
static void             aio_workq_add_entry_locked(aio_workq_t queue, aio_workq_entry *entryp);
static void             aio_entry_ref(aio_workq_entry *entryp);
static void             aio_entry_unref(aio_workq_entry *entryp);
static bool             aio_entry_try_workq_remove(proc_t p, aio_workq_entry *entryp);
static boolean_t        aio_delay_fsync_request(aio_workq_entry *entryp);
static void             aio_free_request(aio_workq_entry *entryp);

static void             aio_workq_init(aio_workq_t wq);
static void             aio_workq_lock_spin(aio_workq_t wq);
static void             aio_workq_unlock(aio_workq_t wq);
static lck_spin_t      *aio_workq_lock(aio_workq_t wq);

static void             aio_work_thread(void *arg, wait_result_t wr);
static aio_workq_entry *aio_get_some_work(void);

static int              aio_queue_async_request(proc_t procp, user_addr_t aiocbp, aio_entry_flags_t);
static int              aio_validate(proc_t, aio_workq_entry *entryp);

static int              do_aio_cancel_locked(proc_t p, int fd, user_addr_t aiocbp, aio_entry_flags_t);
static void             do_aio_completion_and_unlock(proc_t p, aio_workq_entry *entryp, aio_entry_flags_t reason);
static int              do_aio_fsync(aio_workq_entry *entryp);
static int              do_aio_read(aio_workq_entry *entryp);
static int              do_aio_write(aio_workq_entry *entryp);
static void             do_munge_aiocb_user32_to_user(struct user32_aiocb *my_aiocbp, struct user_aiocb *the_user_aiocbp);
static void             do_munge_aiocb_user64_to_user(struct user64_aiocb *my_aiocbp, struct user_aiocb *the_user_aiocbp);
static aio_workq_entry *aio_create_queue_entry(proc_t procp, user_addr_t aiocbp, aio_entry_flags_t);
static int              aio_copy_in_list(proc_t, user_addr_t, user_addr_t *, int);

static void             workq_aio_prepare(struct proc *p);
static bool             workq_aio_entry_add_locked(struct proc *p, aio_workq_entry *entryp);
static void             workq_aio_wakeup_thread(proc_t p);
static void             workq_aio_wakeup_thread_and_unlock(proc_t p);
static int              workq_aio_process_entry(aio_workq_entry *entryp);
static bool             workq_aio_entry_remove_locked(struct proc *p, aio_workq_entry *entryp);

static void             workq_aio_kill_old_threads_call(void *param0, void *param1 __unused);
static void             workq_aio_unpark_continue(void *parameter __unused, wait_result_t wr);

static void             workq_aio_mark_exiting(proc_t p);
static void             workq_aio_exit(proc_t p);

#define ASSERT_AIO_PROC_LOCK_OWNED(p)   LCK_MTX_ASSERT(aio_proc_mutex(p), LCK_MTX_ASSERT_OWNED)
#define ASSERT_AIO_WORKQ_LOCK_OWNED(q)  LCK_SPIN_ASSERT(aio_workq_lock(q), LCK_ASSERT_OWNED)

/*
 *  EXTERNAL PROTOTYPES
 */

/* in ...bsd/kern/sys_generic.c */
extern int dofileread(vfs_context_t ctx, struct fileproc *fp,
    user_addr_t bufp, user_size_t nbyte,
    off_t offset, int flags, user_ssize_t *retval);
extern int dofilewrite(vfs_context_t ctx, struct fileproc *fp,
    user_addr_t bufp, user_size_t nbyte, off_t offset,
    int flags, user_ssize_t *retval);

/*
 * aio external global variables.
 */
extern int aio_max_requests;                    /* AIO_MAX - configurable */
extern int aio_max_requests_per_process;        /* AIO_PROCESS_MAX - configurable */
extern int aio_worker_threads;                  /* AIO_THREAD_COUNT - configurable */


/*
 * aio static variables.
 */
static aio_anchor_cb aio_anchor = {
	.aio_num_workqs = AIO_NUM_WORK_QUEUES,
};
os_refgrp_decl(static, aio_refgrp, "aio", NULL);
static LCK_GRP_DECLARE(aio_proc_lock_grp, "aio_proc");
static LCK_GRP_DECLARE(aio_queue_lock_grp, "aio_queue");
static LCK_MTX_DECLARE(aio_proc_mtx, &aio_proc_lock_grp);

static struct klist aio_klist;
static LCK_GRP_DECLARE(aio_klist_lck_grp, "aio_klist");
static LCK_MTX_DECLARE(aio_klist_lock, &aio_klist_lck_grp);

static KALLOC_TYPE_DEFINE(aio_workq_zonep, aio_workq_entry, KT_DEFAULT);

/* Hash */
static aio_workq_t
aio_entry_workq(__unused aio_workq_entry *entryp)
{
	return &aio_anchor.aio_async_workqs[0];
}

static void
aio_workq_init(aio_workq_t wq)
{
	TAILQ_INIT(&wq->aioq_entries);
	lck_spin_init(&wq->aioq_lock, &aio_queue_lock_grp, LCK_ATTR_NULL);
	waitq_init(&wq->aioq_waitq, WQT_QUEUE, SYNC_POLICY_FIFO);
}


/*
 * Can be passed a queue which is locked spin.
 */
static void
aio_workq_remove_entry_locked(aio_workq_t queue, aio_workq_entry *entryp)
{
	ASSERT_AIO_WORKQ_LOCK_OWNED(queue);

	if (entryp->aio_workq_link.tqe_prev == NULL) {
		panic("Trying to remove an entry from a work queue, but it is not on a queue");
	}

	TAILQ_REMOVE(&queue->aioq_entries, entryp, aio_workq_link);
	entryp->aio_workq_link.tqe_prev = NULL; /* Not on a workq */
}

static void
aio_workq_add_entry_locked(aio_workq_t queue, aio_workq_entry *entryp)
{
	ASSERT_AIO_WORKQ_LOCK_OWNED(queue);

	if (bootarg_aio_new_workq) {
		panic("old workq implementation selected with bootarg set");
	}

	TAILQ_INSERT_TAIL(&queue->aioq_entries, entryp, aio_workq_link);
}

static void
aio_proc_lock(proc_t procp)
{
	lck_mtx_lock(aio_proc_mutex(procp));
}

static void
aio_proc_lock_spin(proc_t procp)
{
	lck_mtx_lock_spin(aio_proc_mutex(procp));
}

static bool
aio_has_any_work(void)
{
	return os_atomic_load(&aio_anchor.aio_total_count, relaxed) != 0;
}

static bool
aio_try_proc_insert_active_locked(proc_t procp, aio_workq_entry *entryp)
{
	int old, new;

	ASSERT_AIO_PROC_LOCK_OWNED(procp);

	if (procp->p_aio_total_count >= aio_max_requests_per_process) {
		return false;
	}

	if (is_already_queued(procp, entryp->uaiocbp)) {
		return false;
	}

	os_atomic_rmw_loop(&aio_anchor.aio_total_count, old, new, relaxed, {
		if (old >= aio_max_requests) {
		        os_atomic_rmw_loop_give_up(return false);
		}
		new = old + 1;
	});

	TAILQ_INSERT_TAIL(&procp->p_aio_activeq, entryp, aio_proc_link);
	procp->p_aio_total_count++;
	return true;
}

static void
aio_proc_move_done_locked(proc_t procp, aio_workq_entry *entryp)
{
	TAILQ_REMOVE(&procp->p_aio_activeq, entryp, aio_proc_link);
	TAILQ_INSERT_TAIL(&procp->p_aio_doneq, entryp, aio_proc_link);
}

static void
aio_proc_remove_done_locked(proc_t procp, aio_workq_entry *entryp)
{
	TAILQ_REMOVE(&procp->p_aio_doneq, entryp, aio_proc_link);
	entryp->aio_proc_link.tqe_prev = NULL;
	if (os_atomic_dec_orig(&aio_anchor.aio_total_count, relaxed) <= 0) {
		panic("Negative total AIO count!");
	}
	if (procp->p_aio_total_count-- <= 0) {
		panic("proc %p: p_aio_total_count accounting mismatch", procp);
	}
}

static void
aio_proc_unlock(proc_t procp)
{
	lck_mtx_unlock(aio_proc_mutex(procp));
}

static lck_mtx_t*
aio_proc_mutex(proc_t procp)
{
	return &procp->p_mlock;
}

static void
aio_entry_ref(aio_workq_entry *entryp)
{
	os_ref_retain(&entryp->aio_refcount);
}

static void
aio_entry_unref(aio_workq_entry *entryp)
{
	if (os_ref_release(&entryp->aio_refcount) == 0) {
		aio_free_request(entryp);
	}
}

static bool
aio_entry_try_workq_remove(proc_t p, aio_workq_entry *entryp)
{
	/* Can only be cancelled if it's still on a work queue */
	if (entryp->aio_workq_link.tqe_prev != NULL) {
		aio_workq_t queue;
		if (bootarg_aio_new_workq) {
			return workq_aio_entry_remove_locked(p, entryp);
		}

		/* Will have to check again under the lock */
		queue = aio_entry_workq(entryp);
		aio_workq_lock_spin(queue);
		if (entryp->aio_workq_link.tqe_prev != NULL) {
			aio_workq_remove_entry_locked(queue, entryp);
			aio_workq_unlock(queue);
			return true;
		} else {
			aio_workq_unlock(queue);
		}
	}

	return false;
}

static void
aio_workq_lock_spin(aio_workq_t wq)
{
	lck_spin_lock(aio_workq_lock(wq));
}

static void
aio_workq_unlock(aio_workq_t wq)
{
	lck_spin_unlock(aio_workq_lock(wq));
}

static lck_spin_t*
aio_workq_lock(aio_workq_t wq)
{
	return &wq->aioq_lock;
}

/*
 * aio_cancel - attempt to cancel one or more async IO requests currently
 * outstanding against file descriptor uap->fd.  If uap->aiocbp is not
 * NULL then only one specific IO is cancelled (if possible).  If uap->aiocbp
 * is NULL then all outstanding async IO request for the given file
 * descriptor are cancelled (if possible).
 */
int
aio_cancel(proc_t p, struct aio_cancel_args *uap, int *retval)
{
	struct user_aiocb my_aiocb;
	int               result;

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_cancel) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(p), uap->fd, uap->aiocbp, 0, 0);

	if (uap->fd) {
		vnode_t vp = NULLVP;
		const char *vname = NULL;

		result = vnode_getfromfd(vfs_context_current(), uap->fd, &vp);
		if (result != 0) {
			result = EBADF;
			goto ExitRoutine;
		}

		vname = vnode_getname(vp);
		/*
		 * The aio_cancel() system call will always	return AIO_NOTCANCELED for
		 * file	descriptor associated with raw disk device.
		 */
		if (vnode_ischr(vp) && vname && !strncmp(vname, "rdisk", 5)) {
			result = 0;
			*retval = AIO_NOTCANCELED;
		}

		if (vname) {
			vnode_putname(vname);
		}
		vnode_put(vp);

		if (result == 0 && *retval == AIO_NOTCANCELED) {
			goto ExitRoutine;
		}
	}

	/* quick check to see if there are any async IO requests queued up */
	if (!aio_has_any_work()) {
		result = 0;
		*retval = AIO_ALLDONE;
		goto ExitRoutine;
	}

	*retval = -1;
	if (uap->aiocbp != USER_ADDR_NULL) {
		if (proc_is64bit(p)) {
			struct user64_aiocb aiocb64;

			result = copyin(uap->aiocbp, &aiocb64, sizeof(aiocb64));
			if (result == 0) {
				do_munge_aiocb_user64_to_user(&aiocb64, &my_aiocb);
			}
		} else {
			struct user32_aiocb aiocb32;

			result = copyin(uap->aiocbp, &aiocb32, sizeof(aiocb32));
			if (result == 0) {
				do_munge_aiocb_user32_to_user(&aiocb32, &my_aiocb);
			}
		}

		if (result != 0) {
			result = EAGAIN;
			goto ExitRoutine;
		}

		/* NOTE - POSIX standard says a mismatch between the file */
		/* descriptor passed in and the file descriptor embedded in */
		/* the aiocb causes unspecified results.  We return EBADF in */
		/* that situation.  */
		if (uap->fd != my_aiocb.aio_fildes) {
			result = EBADF;
			goto ExitRoutine;
		}
	}

	aio_proc_lock(p);
	result = do_aio_cancel_locked(p, uap->fd, uap->aiocbp, 0);
	ASSERT_AIO_PROC_LOCK_OWNED(p);
	aio_proc_unlock(p);

	if (result != -1) {
		*retval = result;
		result = 0;
		goto ExitRoutine;
	}

	result = EBADF;

ExitRoutine:
	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_cancel) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(p), uap->fd, uap->aiocbp, result, 0);

	return result;
}


/*
 * _aio_close - internal function used to clean up async IO requests for
 * a file descriptor that is closing.
 * THIS MAY BLOCK.
 */
__private_extern__ void
_aio_close(proc_t p, int fd)
{
	int error;

	/* quick check to see if there are any async IO requests queued up */
	if (!aio_has_any_work()) {
		return;
	}

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_close) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(p), fd, 0, 0, 0);

	/* cancel all async IO requests on our todo queues for this file descriptor */
	aio_proc_lock(p);
	error = do_aio_cancel_locked(p, fd, USER_ADDR_NULL, AIO_CLOSE_WAIT);
	ASSERT_AIO_PROC_LOCK_OWNED(p);
	if (error == AIO_NOTCANCELED) {
		/*
		 * AIO_NOTCANCELED is returned when we find an aio request for this process
		 * and file descriptor on the active async IO queue.  Active requests cannot
		 * be cancelled so we must wait for them to complete.  We will get a special
		 * wake up call on our channel used to sleep for ALL active requests to
		 * complete.  This sleep channel (proc.AIO_CLEANUP_SLEEP_CHAN) is only used
		 * when we must wait for all active aio requests.
		 */

		KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_close_sleep) | DBG_FUNC_NONE,
		    VM_KERNEL_ADDRPERM(p), fd, 0, 0, 0);

		while (aio_proc_has_active_requests_for_file(p, fd)) {
			msleep(&p->AIO_CLEANUP_SLEEP_CHAN, aio_proc_mutex(p), PRIBIO, "aio_close", 0);
		}
	}

	aio_proc_unlock(p);

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_close) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(p), fd, 0, 0, 0);
}


/*
 * aio_error - return the error status associated with the async IO
 * request referred to by uap->aiocbp.  The error status is the errno
 * value that would be set by the corresponding IO request (read, wrtie,
 * fdatasync, or sync).
 */
int
aio_error(proc_t p, struct aio_error_args *uap, int *retval)
{
	aio_workq_entry *entryp;
	int              error;

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_error) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(p), uap->aiocbp, 0, 0, 0);

	/* see if there are any aios to check */
	if (!aio_has_any_work()) {
		return EINVAL;
	}

	aio_proc_lock(p);

	/* look for a match on our queue of async IO requests that have completed */
	TAILQ_FOREACH(entryp, &p->p_aio_doneq, aio_proc_link) {
		if (entryp->uaiocbp == uap->aiocbp) {
			ASSERT_AIO_FROM_PROC(entryp, p);

			*retval = entryp->errorval;
			error = 0;

			KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_error_val) | DBG_FUNC_NONE,
			    VM_KERNEL_ADDRPERM(p), uap->aiocbp, *retval, 0, 0);
			goto ExitRoutine;
		}
	}

	/* look for a match on our queue of active async IO requests */
	TAILQ_FOREACH(entryp, &p->p_aio_activeq, aio_proc_link) {
		if (entryp->uaiocbp == uap->aiocbp) {
			ASSERT_AIO_FROM_PROC(entryp, p);
			*retval = EINPROGRESS;
			error = 0;
			KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_error_activeq) | DBG_FUNC_NONE,
			    VM_KERNEL_ADDRPERM(p), uap->aiocbp, *retval, 0, 0);
			goto ExitRoutine;
		}
	}

	error = EINVAL;

ExitRoutine:
	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_error) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(p), uap->aiocbp, error, 0, 0);
	aio_proc_unlock(p);

	return error;
}


/*
 * aio_fsync - asynchronously force all IO operations associated
 * with the file indicated by the file descriptor (uap->aiocbp->aio_fildes) and
 * queued at the time of the call to the synchronized completion state.
 * NOTE - we do not support op O_DSYNC at this point since we do not support the
 * fdatasync() call.
 */
int
aio_fsync(proc_t p, struct aio_fsync_args *uap, int *retval)
{
	aio_entry_flags_t fsync_kind;
	int error;

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_fsync) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(p), uap->aiocbp, uap->op, 0, 0);

	*retval = 0;
	/* 0 := O_SYNC for binary backward compatibility with Panther */
	if (uap->op == O_SYNC || uap->op == 0) {
		fsync_kind = AIO_FSYNC;
	} else if (uap->op == O_DSYNC) {
		fsync_kind = AIO_DSYNC;
	} else {
		*retval = -1;
		error = EINVAL;
		goto ExitRoutine;
	}

	error = aio_queue_async_request(p, uap->aiocbp, fsync_kind);
	if (error != 0) {
		*retval = -1;
	}

ExitRoutine:
	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_fsync) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(p), uap->aiocbp, error, 0, 0);

	return error;
}


/* aio_read - asynchronously read uap->aiocbp->aio_nbytes bytes from the
 * file descriptor (uap->aiocbp->aio_fildes) into the buffer
 * (uap->aiocbp->aio_buf).
 */
int
aio_read(proc_t p, struct aio_read_args *uap, int *retval)
{
	int error;

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_read) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(p), uap->aiocbp, 0, 0, 0);

	*retval = 0;

	error = aio_queue_async_request(p, uap->aiocbp, AIO_READ);
	if (error != 0) {
		*retval = -1;
	}

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_read) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(p), uap->aiocbp, error, 0, 0);

	return error;
}


/*
 * aio_return - return the return status associated with the async IO
 * request referred to by uap->aiocbp.  The return status is the value
 * that would be returned by corresponding IO request (read, write,
 * fdatasync, or sync).  This is where we release kernel resources
 * held for async IO call associated with the given aiocb pointer.
 */
int
aio_return(proc_t p, struct aio_return_args *uap, user_ssize_t *retval)
{
	aio_workq_entry *entryp;
	int              error = EINVAL;

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_return) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(p), uap->aiocbp, 0, 0, 0);

	/* See if there are any entries to check */
	if (!aio_has_any_work()) {
		goto ExitRoutine;
	}

	aio_proc_lock(p);
	*retval = 0;

	/* look for a match on our queue of async IO requests that have completed */
	TAILQ_FOREACH(entryp, &p->p_aio_doneq, aio_proc_link) {
		ASSERT_AIO_FROM_PROC(entryp, p);
		if (entryp->uaiocbp == uap->aiocbp) {
			/* Done and valid for aio_return(), pull it off the list */
			aio_proc_remove_done_locked(p, entryp);

			*retval = entryp->returnval;
			error = 0;
			aio_proc_unlock(p);

			aio_entry_unref(entryp);

			KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_return_val) | DBG_FUNC_NONE,
			    VM_KERNEL_ADDRPERM(p), uap->aiocbp, *retval, 0, 0);
			goto ExitRoutine;
		}
	}

	/* look for a match on our queue of active async IO requests */
	TAILQ_FOREACH(entryp, &p->p_aio_activeq, aio_proc_link) {
		ASSERT_AIO_FROM_PROC(entryp, p);
		if (entryp->uaiocbp == uap->aiocbp) {
			error = EINPROGRESS;
			KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_return_activeq) | DBG_FUNC_NONE,
			    VM_KERNEL_ADDRPERM(p), uap->aiocbp, *retval, 0, 0);
			break;
		}
	}

	aio_proc_unlock(p);

ExitRoutine:
	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_return) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(p), uap->aiocbp, error, 0, 0);

	return error;
}


/*
 * _aio_exec - internal function used to clean up async IO requests for
 * a process that is going away due to exec().  We cancel any async IOs
 * we can and wait for those already active.  We also disable signaling
 * for cancelled or active aio requests that complete.
 * This routine MAY block!
 */
__private_extern__ void
_aio_exec(proc_t p)
{
	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_exec) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(p), 0, 0, 0, 0);

	_aio_exit(p);

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_exec) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(p), 0, 0, 0, 0);
}


/*
 * _aio_exit - internal function used to clean up async IO requests for
 * a process that is terminating (via exit() or exec()).  We cancel any async IOs
 * we can and wait for those already active.  We also disable signaling
 * for cancelled or active aio requests that complete.  This routine MAY block!
 */
__private_extern__ void
_aio_exit(proc_t p)
{
	TAILQ_HEAD(, aio_workq_entry) tofree = TAILQ_HEAD_INITIALIZER(tofree);
	aio_workq_entry *entryp, *tmp;
	int              error;

	/* quick check to see if there are any async IO requests queued up */
	if (!aio_has_any_work()) {
		workq_aio_mark_exiting(p);
		workq_aio_exit(p);
		return;
	}

	workq_aio_mark_exiting(p);

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_exit) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(p), 0, 0, 0, 0);

	aio_proc_lock(p);

	/*
	 * cancel async IO requests on the todo work queue and wait for those
	 * already active to complete.
	 */
	error = do_aio_cancel_locked(p, -1, USER_ADDR_NULL, AIO_EXIT_WAIT);
	ASSERT_AIO_PROC_LOCK_OWNED(p);
	if (error == AIO_NOTCANCELED) {
		/*
		 * AIO_NOTCANCELED is returned when we find an aio request for this process
		 * on the active async IO queue.  Active requests cannot be cancelled so we
		 * must wait for them to complete.  We will get a special wake up call on
		 * our channel used to sleep for ALL active requests to complete.  This sleep
		 * channel (proc.AIO_CLEANUP_SLEEP_CHAN) is only used when we must wait for all
		 * active aio requests.
		 */

		KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_exit_sleep) | DBG_FUNC_NONE,
		    VM_KERNEL_ADDRPERM(p), 0, 0, 0, 0);

		while (aio_has_active_requests_for_process(p)) {
			msleep(&p->AIO_CLEANUP_SLEEP_CHAN, aio_proc_mutex(p), PRIBIO, "aio_exit", 0);
		}
	}

	assert(!aio_has_active_requests_for_process(p));

	/* release all aio resources used by this process */
	TAILQ_FOREACH_SAFE(entryp, &p->p_aio_doneq, aio_proc_link, tmp) {
		ASSERT_AIO_FROM_PROC(entryp, p);

		aio_proc_remove_done_locked(p, entryp);
		TAILQ_INSERT_TAIL(&tofree, entryp, aio_proc_link);
	}

	aio_proc_unlock(p);

	workq_aio_exit(p);

	/* free all the entries outside of the aio_proc_lock() */
	TAILQ_FOREACH_SAFE(entryp, &tofree, aio_proc_link, tmp) {
		entryp->aio_proc_link.tqe_prev = NULL;
		aio_entry_unref(entryp);
	}

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_exit) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(p), 0, 0, 0, 0);
}


static bool
should_cancel(aio_workq_entry *entryp, int fd, user_addr_t aiocbp,
    aio_entry_flags_t reason)
{
	if (reason & AIO_EXIT_WAIT) {
		/* caller is _aio_exit() */
		return true;
	}
	if (fd != entryp->aiocb.aio_fildes) {
		/* not the file we're looking for */
		return false;
	}
	/*
	 * aio_cancel() or _aio_close() cancel
	 * everything for a given fd when aiocbp is NULL
	 */
	return aiocbp == USER_ADDR_NULL || entryp->uaiocbp == aiocbp;
}

/*
 * do_aio_cancel_locked - cancel async IO requests (if possible).  We get called by
 * aio_cancel, close, and at exit.
 * There are three modes of operation: 1) cancel all async IOs for a process -
 * fd is 0 and aiocbp is NULL 2) cancel all async IOs for file descriptor - fd
 * is > 0 and aiocbp is NULL 3) cancel one async IO associated with the given
 * aiocbp.
 * Returns -1 if no matches were found, AIO_CANCELED when we cancelled all
 * target async IO requests, AIO_NOTCANCELED if we could not cancel all
 * target async IO requests, and AIO_ALLDONE if all target async IO requests
 * were already complete.
 * WARNING - do not deference aiocbp in this routine, it may point to user
 * land data that has not been copied in (when called from aio_cancel())
 *
 * Called with proc locked, and returns the same way.
 */
static int
do_aio_cancel_locked(proc_t p, int fd, user_addr_t aiocbp,
    aio_entry_flags_t reason)
{
	bool multiple_matches = (aiocbp == USER_ADDR_NULL);
	aio_workq_entry *entryp, *tmp;
	int result;

	ASSERT_AIO_PROC_LOCK_OWNED(p);

	/* look for a match on our queue of async todo work. */
again:
	result = -1;
	TAILQ_FOREACH_SAFE(entryp, &p->p_aio_activeq, aio_proc_link, tmp) {
		ASSERT_AIO_FROM_PROC(entryp, p);

		if (!should_cancel(entryp, fd, aiocbp, reason)) {
			continue;
		}

		if (reason) {
			/* mark the entry as blocking close or exit/exec */
			entryp->flags |= reason;
			if ((entryp->flags & AIO_EXIT_WAIT) && (entryp->flags & AIO_CLOSE_WAIT)) {
				panic("Close and exit flags set at the same time");
			}
		}

		/* Can only be cancelled if it's still on a work queue */
		if (aio_entry_try_workq_remove(p, entryp)) {
			entryp->errorval = ECANCELED;
			entryp->returnval = -1;

			/* Now it's officially cancelled.  Do the completion */
			KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_cancel_async_workq) | DBG_FUNC_NONE,
			    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
			    fd, 0, 0);
			do_aio_completion_and_unlock(p, entryp, AIO_CANCELLED);

			aio_proc_lock(p);

			if (multiple_matches) {
				/*
				 * Restart from the head of the proc active queue since it
				 * may have been changed while we were away doing completion
				 * processing.
				 *
				 * Note that if we found an uncancellable AIO before, we will
				 * either find it again or discover that it's been completed,
				 * so resetting the result will not cause us to return success
				 * despite outstanding AIOs.
				 */
				goto again;
			}

			return AIO_CANCELED;
		}

		/*
		 * It's been taken off the active queue already, i.e. is in flight.
		 * All we can do is ask for notification.
		 */
		KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_cancel_activeq) | DBG_FUNC_NONE,
		    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
		    fd, 0, 0);

		result = AIO_NOTCANCELED;
		if (!multiple_matches) {
			return result;
		}
	}

	/*
	 * if we didn't find any matches on the todo or active queues then look for a
	 * match on our queue of async IO requests that have completed and if found
	 * return AIO_ALLDONE result.
	 *
	 * Proc AIO lock is still held.
	 */
	if (result == -1) {
		TAILQ_FOREACH(entryp, &p->p_aio_doneq, aio_proc_link) {
			ASSERT_AIO_FROM_PROC(entryp, p);
			if (should_cancel(entryp, fd, aiocbp, reason)) {
				KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_cancel_doneq) | DBG_FUNC_NONE,
				    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
				    fd, 0, 0);

				result = AIO_ALLDONE;
				if (!multiple_matches) {
					return result;
				}
			}
		}
	}

	return result;
}


/*
 * aio_suspend - suspend the calling thread until at least one of the async
 * IO operations referenced by uap->aiocblist has completed, until a signal
 * interrupts the function, or uap->timeoutp time interval (optional) has
 * passed.
 * Returns 0 if one or more async IOs have completed else -1 and errno is
 * set appropriately - EAGAIN if timeout elapses or EINTR if an interrupt
 * woke us up.
 */
int
aio_suspend(proc_t p, struct aio_suspend_args *uap, int *retval)
{
	__pthread_testcancel(1);
	return aio_suspend_nocancel(p, (struct aio_suspend_nocancel_args *)uap, retval);
}


int
aio_suspend_nocancel(proc_t p, struct aio_suspend_nocancel_args *uap, int *retval)
{
	int                     error;
	int                     i;
	uint64_t                abstime;
	struct user_timespec    ts;
	aio_workq_entry        *entryp;
	user_addr_t            *aiocbpp;
	size_t                  aiocbpp_size;

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_suspend) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(p), uap->nent, 0, 0, 0);

	*retval = -1;
	abstime = 0;
	aiocbpp = NULL;

	if (!aio_has_any_work()) {
		error = EINVAL;
		goto ExitThisRoutine;
	}

	if (uap->nent < 1 || uap->nent > aio_max_requests_per_process ||
	    os_mul_overflow(sizeof(user_addr_t), uap->nent, &aiocbpp_size)) {
		error = EINVAL;
		goto ExitThisRoutine;
	}

	if (uap->timeoutp != USER_ADDR_NULL) {
		if (proc_is64bit(p)) {
			struct user64_timespec temp;
			error = copyin(uap->timeoutp, &temp, sizeof(temp));
			if (error == 0) {
				ts.tv_sec = (user_time_t)temp.tv_sec;
				ts.tv_nsec = (user_long_t)temp.tv_nsec;
			}
		} else {
			struct user32_timespec temp;
			error = copyin(uap->timeoutp, &temp, sizeof(temp));
			if (error == 0) {
				ts.tv_sec = temp.tv_sec;
				ts.tv_nsec = temp.tv_nsec;
			}
		}
		if (error != 0) {
			error = EAGAIN;
			goto ExitThisRoutine;
		}

		if (ts.tv_sec < 0 || ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000) {
			error = EINVAL;
			goto ExitThisRoutine;
		}

		nanoseconds_to_absolutetime((uint64_t)ts.tv_sec * NSEC_PER_SEC + ts.tv_nsec,
		    &abstime);
		clock_absolutetime_interval_to_deadline(abstime, &abstime);
	}

	aiocbpp = (user_addr_t *)kalloc_data(aiocbpp_size, Z_WAITOK);
	if (aiocbpp == NULL || aio_copy_in_list(p, uap->aiocblist, aiocbpp, uap->nent)) {
		error = EAGAIN;
		goto ExitThisRoutine;
	}

	/* check list of aio requests to see if any have completed */
check_for_our_aiocbp:
	aio_proc_lock_spin(p);
	for (i = 0; i < uap->nent; i++) {
		user_addr_t     aiocbp;

		/* NULL elements are legal so check for 'em */
		aiocbp = *(aiocbpp + i);
		if (aiocbp == USER_ADDR_NULL) {
			continue;
		}

		/* return immediately if any aio request in the list is done */
		TAILQ_FOREACH(entryp, &p->p_aio_doneq, aio_proc_link) {
			ASSERT_AIO_FROM_PROC(entryp, p);
			if (entryp->uaiocbp == aiocbp) {
				aio_proc_unlock(p);
				*retval = 0;
				error = 0;
				goto ExitThisRoutine;
			}
		}
	}

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_suspend_sleep) | DBG_FUNC_NONE,
	    VM_KERNEL_ADDRPERM(p), uap->nent, 0, 0, 0);

	/*
	 * wait for an async IO to complete or a signal fires or timeout expires.
	 * we return EAGAIN (35) for timeout expiration and EINTR (4) when a signal
	 * interrupts us.  If an async IO completes before a signal fires or our
	 * timeout expires, we get a wakeup call from aio_work_thread().
	 */

	error = msleep1(&p->AIO_SUSPEND_SLEEP_CHAN, aio_proc_mutex(p),
	    PCATCH | PWAIT | PDROP, "aio_suspend", abstime);
	if (error == 0) {
		/*
		 * got our wakeup call from aio_work_thread().
		 * Since we can get a wakeup on this channel from another thread in the
		 * same process we head back up to make sure this is for the correct aiocbp.
		 * If it is the correct aiocbp we will return from where we do the check
		 * (see entryp->uaiocbp == aiocbp after check_for_our_aiocbp label)
		 * else we will fall out and just sleep again.
		 */
		goto check_for_our_aiocbp;
	} else if (error == EWOULDBLOCK) {
		/* our timeout expired */
		error = EAGAIN;
	} else {
		/* we were interrupted */
		error = EINTR;
	}

ExitThisRoutine:
	if (aiocbpp != NULL) {
		kfree_data(aiocbpp, aiocbpp_size);
	}

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_suspend) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(p), uap->nent, error, 0, 0);

	return error;
}


/* aio_write - asynchronously write uap->aiocbp->aio_nbytes bytes to the
 * file descriptor (uap->aiocbp->aio_fildes) from the buffer
 * (uap->aiocbp->aio_buf).
 */

int
aio_write(proc_t p, struct aio_write_args *uap, int *retval __unused)
{
	int error;

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_write) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(p), uap->aiocbp, 0, 0, 0);

	error = aio_queue_async_request(p, uap->aiocbp, AIO_WRITE);

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_write) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(p), uap->aiocbp, error, 0, 0);

	return error;
}


static int
aio_copy_in_list(proc_t procp, user_addr_t aiocblist, user_addr_t *aiocbpp,
    int nent)
{
	int result;

	/* copyin our aiocb pointers from list */
	result = copyin(aiocblist, aiocbpp,
	    proc_is64bit(procp) ? (nent * sizeof(user64_addr_t))
	    : (nent * sizeof(user32_addr_t)));
	if (result) {
		return result;
	}

	/*
	 * We depend on a list of user_addr_t's so we need to
	 * munge and expand when these pointers came from a
	 * 32-bit process
	 */
	if (!proc_is64bit(procp)) {
		/* copy from last to first to deal with overlap */
		user32_addr_t *my_ptrp = ((user32_addr_t *)aiocbpp) + (nent - 1);
		user_addr_t *my_addrp = aiocbpp + (nent - 1);

		for (int i = 0; i < nent; i++, my_ptrp--, my_addrp--) {
			*my_addrp = (user_addr_t) (*my_ptrp);
		}
	}

	return 0;
}


static int
aio_copy_in_sigev(proc_t procp, user_addr_t sigp, struct user_sigevent *sigev)
{
	int     result = 0;

	if (sigp == USER_ADDR_NULL) {
		goto out;
	}

	/*
	 * We need to munge aio_sigevent since it contains pointers.
	 * Since we do not know if sigev_value is an int or a ptr we do
	 * NOT cast the ptr to a user_addr_t.   This means if we send
	 * this info back to user space we need to remember sigev_value
	 * was not expanded for the 32-bit case.
	 *
	 * Notes:	 This does NOT affect us since we don't support
	 *		sigev_value yet in the aio context.
	 */
	if (proc_is64bit(procp)) {
#if __LP64__
		struct user64_sigevent sigevent64;

		result = copyin(sigp, &sigevent64, sizeof(sigevent64));
		if (result == 0) {
			sigev->sigev_notify = sigevent64.sigev_notify;
			sigev->sigev_signo = sigevent64.sigev_signo;
			sigev->sigev_value.size_equivalent.sival_int = sigevent64.sigev_value.size_equivalent.sival_int;
			sigev->sigev_notify_function = sigevent64.sigev_notify_function;
			sigev->sigev_notify_attributes = sigevent64.sigev_notify_attributes;
		}
#else
		panic("64bit process on 32bit kernel is not supported");
#endif
	} else {
		struct user32_sigevent sigevent32;

		result = copyin(sigp, &sigevent32, sizeof(sigevent32));
		if (result == 0) {
			sigev->sigev_notify = sigevent32.sigev_notify;
			sigev->sigev_signo = sigevent32.sigev_signo;
			sigev->sigev_value.size_equivalent.sival_int = sigevent32.sigev_value.sival_int;
			sigev->sigev_notify_function = CAST_USER_ADDR_T(sigevent32.sigev_notify_function);
			sigev->sigev_notify_attributes = CAST_USER_ADDR_T(sigevent32.sigev_notify_attributes);
		}
	}

	if (result != 0) {
		result = EAGAIN;
	}

out:
	return result;
}

/*
 * validate user_sigevent.  at this point we only support
 * sigev_notify equal to SIGEV_SIGNAL or SIGEV_NONE.  this means
 * sigev_value, sigev_notify_function, and sigev_notify_attributes
 * are ignored, since SIGEV_THREAD is unsupported.  This is consistent
 * with no [RTS] (RalTime Signal) option group support.
 */
static int
aio_sigev_validate(const struct user_sigevent *sigev)
{
	switch (sigev->sigev_notify) {
	case SIGEV_SIGNAL:
	{
		int signum;

		/* make sure we have a valid signal number */
		signum = sigev->sigev_signo;
		if (signum <= 0 || signum >= NSIG ||
		    signum == SIGKILL || signum == SIGSTOP) {
			return EINVAL;
		}
	}
	break;

	case SIGEV_NONE:
		break;

	case SIGEV_KEVENT:
		/*
		 * The sigev_signo should contain the descriptor of the kqueue.
		 * Validate that it contains some sane value.
		 */
		if (sigev->sigev_signo <= 0 || sigev->sigev_signo > maxfilesperproc) {
			return EINVAL;
		}
		break;

	case SIGEV_THREAD:
	/* Unsupported [RTS] */

	default:
		return EINVAL;
	}

	return 0;
}


/*
 * aio_try_enqueue_work_locked
 *
 * Queue up the entry on the aio asynchronous work queue in priority order
 * based on the relative priority of the request.  We calculate the relative
 * priority using the nice value of the caller and the value
 *
 * Parameters:	procp			Process queueing the I/O
 *		entryp			The work queue entry being queued
 *		leader			The work leader if any
 *
 * Returns:	Whether the enqueue was successful
 *
 * Notes:	This function is used for both lio_listio and aio
 *
 * XXX:		At some point, we may have to consider thread priority
 *		rather than process priority, but we don't maintain the
 *		adjusted priority for threads the POSIX way.
 *
 * Called with proc locked.
 */
static bool
aio_try_enqueue_work_locked(proc_t procp, aio_workq_entry *entryp,
    aio_workq_entry *leader)
{
	ASSERT_AIO_PROC_LOCK_OWNED(procp);

	/* Onto proc queue */
	if (!aio_try_proc_insert_active_locked(procp, entryp)) {
		return false;
	}

	if (leader) {
		aio_entry_ref(leader); /* consumed in do_aio_completion_and_unlock */
		leader->lio_pending++;
		entryp->lio_leader = leader;
	}

	/* And work queue */
	aio_entry_ref(entryp); /* consumed in do_aio_completion_and_unlock */
	if (bootarg_aio_new_workq) {
		if (!workq_aio_entry_add_locked(procp, entryp)) {
			(void)os_ref_release(&entryp->aio_refcount);
			return false;
		}
	} else {
		aio_workq_t queue = aio_entry_workq(entryp);
		aio_workq_lock_spin(queue);
		aio_workq_add_entry_locked(queue, entryp);
		waitq_wakeup64_one(&queue->aioq_waitq, CAST_EVENT64_T(queue),
		    THREAD_AWAKENED, WAITQ_WAKEUP_DEFAULT);
		aio_workq_unlock(queue);
	}

	KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_AIO, AIO_work_queued) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(procp), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
	    entryp->flags, entryp->aiocb.aio_fildes, 0);
	KERNEL_DEBUG_CONSTANT(BSDDBG_CODE(DBG_BSD_AIO, AIO_work_queued) | DBG_FUNC_END,
	    entryp->aiocb.aio_offset, 0, entryp->aiocb.aio_nbytes, 0, 0);
	return true;
}

/*
 * EV_FLAG0/1 are filter specific flags.
 * Repurpose EV_FLAG0 to indicate the kevent is registered from kernel.
 */
#define EV_KERNEL    EV_FLAG0

/* Register a kevent for AIO completion notification. */
static int
aio_register_kevent(proc_t procp, aio_workq_entry *entryp)
{
	struct kevent_qos_s kev;
	struct fileproc *fp = NULL;
	kqueue_t kqu;
	int kqfd = entryp->aiocb.aio_sigevent.sigev_signo;
	int error;

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_register_kevent) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(procp), VM_KERNEL_ADDRPERM(entryp),
	    VM_KERNEL_ADDRPERM(entryp->uaiocbp), kqfd, 0);

	error = fp_get_ftype(procp, kqfd, DTYPE_KQUEUE, EBADF, &fp);
	if (error) {
		goto exit;
	}

	kqu.kq = (struct kqueue *)fp_get_data(fp);

	memset(&kev, 0, sizeof(kev));
	kev.ident = (uintptr_t)entryp->uaiocbp;
	kev.filter = EVFILT_AIO;
	/*
	 * Set the EV_FLAG0 to indicate the event is registered from the kernel.
	 * This flag later is checked in filt_aioattach() and to determine if
	 * a kevent is registered from kernel or user-space.
	 */
	kev.flags = EV_ADD | EV_ENABLE | EV_CLEAR | EV_ONESHOT | EV_KERNEL;
	kev.udata = entryp->aiocb.aio_sigevent.sigev_value.sival_ptr;
	kev.data = (intptr_t)entryp;

	error = kevent_register(kqu.kq, &kev, NULL);
	assert((error & FILTER_REGISTER_WAIT) == 0);

exit:
	if (fp) {
		fp_drop(procp, kqfd, fp, 0);
	}

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_register_kevent) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(procp), VM_KERNEL_ADDRPERM(entryp), error, 0, 0);

	return error;
}

/*
 * lio_listio - initiate a list of IO requests.  We process the list of
 * aiocbs either synchronously (mode == LIO_WAIT) or asynchronously
 * (mode == LIO_NOWAIT).
 *
 * The caller gets error and return status for each aiocb in the list
 * via aio_error and aio_return.  We must keep completed requests until
 * released by the aio_return call.
 */
int
lio_listio(proc_t p, struct lio_listio_args *uap, int *retval __unused)
{
	aio_workq_entry         *entries[AIO_LISTIO_MAX] = { };
	user_addr_t              aiocbpp[AIO_LISTIO_MAX];
	struct user_sigevent     aiosigev = { };
	int                      result = 0;
	int                      lio_count = 0;

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_listio) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(p), uap->nent, uap->mode, 0, 0);

	if (!(uap->mode == LIO_NOWAIT || uap->mode == LIO_WAIT)) {
		result = EINVAL;
		goto ExitRoutine;
	}

	if (uap->nent < 1 || uap->nent > AIO_LISTIO_MAX) {
		result = EINVAL;
		goto ExitRoutine;
	}

	/*
	 * Use sigevent passed in to lio_listio for each of our calls, but
	 * only do completion notification after the last request completes.
	 */
	if (uap->sigp != USER_ADDR_NULL) {
		result = aio_copy_in_sigev(p, uap->sigp, &aiosigev);
		if (result) {
			goto ExitRoutine;
		}
		result = aio_sigev_validate(&aiosigev);
		if (result) {
			goto ExitRoutine;
		}
	}

	if (aio_copy_in_list(p, uap->aiocblist, aiocbpp, uap->nent)) {
		result = EAGAIN;
		goto ExitRoutine;
	}

	/*
	 * allocate/parse all entries
	 */
	for (int i = 0; i < uap->nent; i++) {
		aio_workq_entry *entryp;

		/* NULL elements are legal so check for 'em */
		if (aiocbpp[i] == USER_ADDR_NULL) {
			continue;
		}

		entryp = aio_create_queue_entry(p, aiocbpp[i], AIO_LIO);
		if (entryp == NULL) {
			result = EAGAIN;
			goto ExitRoutine;
		}

		/*
		 * This refcount is cleaned up on exit if the entry
		 * isn't submitted
		 */
		entries[lio_count++] = entryp;
		if ((uap->mode == LIO_NOWAIT) &&
		    (entryp->aiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)) {
			/* Set signal hander, if any */
			entryp->aiocb.aio_sigevent = aiosigev;
		}
	}

	if (lio_count == 0) {
		/* There's nothing to submit */
		goto ExitRoutine;
	}

	/*
	 * Past this point we're commited and will not bail out
	 *
	 * - keep a reference on the leader for LIO_WAIT
	 * - perform the submissions and optionally wait
	 */

	aio_workq_entry *leader = entries[0];
	if (uap->mode == LIO_WAIT) {
		aio_entry_ref(leader); /* consumed below */
	}

	aio_proc_lock(p);

	for (int i = 0; i < lio_count; i++) {
		if (aio_try_enqueue_work_locked(p, entries[i], leader)) {
			workq_aio_wakeup_thread(p); /* this may drop and reacquire the proc lock */
			/*
			 * For SIGEV_KEVENT, every AIO in the list would get its own kevent
			 * notification upon completion as opposed to SIGEV_SIGNAL which a
			 * single notification is deliverd when all AIOs have completed.
			 */
			if ((uap->mode == LIO_NOWAIT) &&
			    (entries[i]->aiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT)) {
				aio_register_kevent(p, entries[i]);
			}
			entries[i] = NULL; /* the entry was submitted */
		} else {
			result = EAGAIN;
		}
	}

	if (uap->mode == LIO_WAIT && result == 0) {
		leader->flags |= AIO_LIO_WAIT;

		while (leader->lio_pending) {
			/* If we were interrupted, fail out (even if all finished) */
			if (msleep(leader, aio_proc_mutex(p),
			    PCATCH | PRIBIO | PSPIN, "lio_listio", 0) != 0) {
				result = EINTR;
				break;
			}
		}

		leader->flags &= ~AIO_LIO_WAIT;
	}

	aio_proc_unlock(p);

	if (uap->mode == LIO_WAIT) {
		aio_entry_unref(leader);
	}

ExitRoutine:
	/* Consume unsubmitted entries */
	for (int i = 0; i < lio_count; i++) {
		if (entries[i]) {
			aio_entry_unref(entries[i]);
		}
	}

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_listio) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(p), result, 0, 0, 0);

	return result;
}


/*
 * aio worker thread.  this is where all the real work gets done.
 * we get a wake up call on sleep channel &aio_anchor.aio_async_workq
 * after new work is queued up.
 */
__attribute__((noreturn))
static void
aio_work_thread(void *arg __unused, wait_result_t wr __unused)
{
	aio_workq_entry         *entryp;
	int                     error;
	vm_map_switch_context_t switch_ctx;
	struct uthread          *uthreadp = NULL;
	proc_t                  p = NULL;

	for (;;) {
		/*
		 * returns with the entry ref'ed.
		 * sleeps until work is available.
		 */
		entryp = aio_get_some_work();
		p = entryp->procp;

		KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_worker_thread) | DBG_FUNC_START,
		    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
		    entryp->flags, 0, 0);

		/*
		 * Assume the target's address space identity for the duration
		 * of the IO.  Note: don't need to have the entryp locked,
		 * because the proc and map don't change until it's freed.
		 */
		uthreadp = (struct uthread *) current_uthread();
		assert(get_task_map(proc_task(current_proc())) != entryp->aio_map);
		assert(uthreadp->uu_aio_task == NULL);

		/*
		 * workq entries at this stage cause _aio_exec() and _aio_exit() to
		 * block until we hit `do_aio_completion_and_unlock()` below,
		 * which means that it is safe to dereference p->task without
		 * holding a lock or taking references.
		 */
		uthreadp->uu_aio_task = proc_task(p);
		switch_ctx = vm_map_switch_to(entryp->aio_map);

		if ((entryp->flags & AIO_READ) != 0) {
			error = do_aio_read(entryp);
		} else if ((entryp->flags & AIO_WRITE) != 0) {
			error = do_aio_write(entryp);
		} else if ((entryp->flags & (AIO_FSYNC | AIO_DSYNC)) != 0) {
			error = do_aio_fsync(entryp);
		} else {
			error = EINVAL;
		}

		/* Restore old map */
		vm_map_switch_back(switch_ctx);
		uthreadp->uu_aio_task = NULL;

		/* liberate unused map */
		vm_map_deallocate(entryp->aio_map);
		entryp->aio_map = VM_MAP_NULL;

		KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_worker_thread) | DBG_FUNC_END,
		    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
		    entryp->errorval, entryp->returnval, 0);

		/* we're done with the IO request so pop it off the active queue and */
		/* push it on the done queue */
		aio_proc_lock(p);
		entryp->errorval = error;
		do_aio_completion_and_unlock(p, entryp, AIO_COMPLETED);
	}
}


/*
 * aio_get_some_work - get the next async IO request that is ready to be executed.
 * aio_fsync complicates matters a bit since we cannot do the fsync until all async
 * IO requests at the time the aio_fsync call came in have completed.
 * NOTE - AIO_LOCK must be held by caller
 */
static aio_workq_entry *
aio_get_some_work(void)
{
	aio_workq_entry *entryp = NULL;
	aio_workq_t      queue = NULL;

	/* Just one queue for the moment.  In the future there will be many. */
	queue = &aio_anchor.aio_async_workqs[0];
	aio_workq_lock_spin(queue);

	/*
	 * Hold the queue lock.
	 *
	 * pop some work off the work queue and add to our active queue
	 * Always start with the queue lock held.
	 */
	while ((entryp = TAILQ_FIRST(&queue->aioq_entries))) {
		/*
		 * Pull of of work queue.  Once it's off, it can't be cancelled,
		 * so we can take our ref once we drop the queue lock.
		 */

		aio_workq_remove_entry_locked(queue, entryp);

		aio_workq_unlock(queue);

		/*
		 * Check if it's an fsync that must be delayed.  No need to lock the entry;
		 * that flag would have been set at initialization.
		 */
		if ((entryp->flags & AIO_FSYNC) != 0) {
			/*
			 * Check for unfinished operations on the same file
			 * in this proc's queue.
			 */
			aio_proc_lock_spin(entryp->procp);
			if (aio_delay_fsync_request(entryp)) {
				/* It needs to be delayed.  Put it back on the end of the work queue */
				KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_fsync_delay) | DBG_FUNC_NONE,
				    VM_KERNEL_ADDRPERM(entryp->procp),
				    VM_KERNEL_ADDRPERM(entryp->uaiocbp), 0, 0, 0);

				aio_proc_unlock(entryp->procp);

				aio_workq_lock_spin(queue);
				aio_workq_add_entry_locked(queue, entryp);
				continue;
			}
			aio_proc_unlock(entryp->procp);
		}

		return entryp;
	}

	/* We will wake up when someone enqueues something */
	waitq_assert_wait64(&queue->aioq_waitq, CAST_EVENT64_T(queue), THREAD_UNINT, 0);
	aio_workq_unlock(queue);
	thread_block(aio_work_thread);

	__builtin_unreachable();
}

/*
 * aio_delay_fsync_request - look to see if this aio_fsync request should be delayed.
 * A big, simple hammer: only send it off if it's the most recently filed IO which has
 * not been completed.
 */
static boolean_t
aio_delay_fsync_request(aio_workq_entry *entryp)
{
	if (proc_in_teardown(entryp->procp)) {
		/*
		 * we can't delay FSYNCS when in teardown as it will confuse _aio_exit,
		 * if it was dequeued, then we must now commit to it
		 */
		return FALSE;
	}

	if (entryp == TAILQ_FIRST(&entryp->procp->p_aio_activeq)) {
		return FALSE;
	}

	return TRUE;
}

static aio_workq_entry *
aio_create_queue_entry(proc_t procp, user_addr_t aiocbp, aio_entry_flags_t flags)
{
	aio_workq_entry *entryp;

	entryp = zalloc_flags(aio_workq_zonep, Z_WAITOK | Z_ZERO);
	entryp->procp = procp;
	entryp->uaiocbp = aiocbp;
	entryp->flags = flags;
	/* consumed in aio_return or _aio_exit */
	os_ref_init(&entryp->aio_refcount, &aio_refgrp);

	if (proc_is64bit(procp)) {
		struct user64_aiocb aiocb64;

		if (copyin(aiocbp, &aiocb64, sizeof(aiocb64)) != 0) {
			goto error_exit;
		}
		do_munge_aiocb_user64_to_user(&aiocb64, &entryp->aiocb);
	} else {
		struct user32_aiocb aiocb32;

		if (copyin(aiocbp, &aiocb32, sizeof(aiocb32)) != 0) {
			goto error_exit;
		}
		do_munge_aiocb_user32_to_user(&aiocb32, &entryp->aiocb);
	}

	/* do some more validation on the aiocb and embedded file descriptor */
	if (aio_validate(procp, entryp) != 0) {
		goto error_exit;
	}

	/* get a reference on the current_thread, which is passed in vfs_context. */
	entryp->context = *vfs_context_current();
	thread_reference(entryp->context.vc_thread);
	kauth_cred_ref(entryp->context.vc_ucred);

	if (bootarg_aio_new_workq) {
		entryp->aio_map = VM_MAP_NULL;
		workq_aio_prepare(procp);
	} else {
		/* get a reference to the user land map in order to keep it around */
		entryp->aio_map = get_task_map(proc_task(procp));
		vm_map_reference(entryp->aio_map);
	}
	return entryp;

error_exit:
	zfree(aio_workq_zonep, entryp);
	return NULL;
}


/*
 * aio_queue_async_request - queue up an async IO request on our work queue then
 * wake up one of our worker threads to do the actual work.  We get a reference
 * to our caller's user land map in order to keep it around while we are
 * processing the request.
 */
static int
aio_queue_async_request(proc_t procp, user_addr_t aiocbp,
    aio_entry_flags_t flags)
{
	aio_workq_entry *entryp;
	int              result;

	entryp = aio_create_queue_entry(procp, aiocbp, flags);
	if (entryp == NULL) {
		result = EAGAIN;
		goto error_noalloc;
	}

	aio_proc_lock(procp);
	if (!aio_try_enqueue_work_locked(procp, entryp, NULL)) {
		result = EAGAIN;
		goto error_exit;
	}

	if ((entryp->aiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) &&
	    (aio_register_kevent(procp, entryp) != 0)) {
		result = EAGAIN;
		goto error_exit;
	}
	workq_aio_wakeup_thread_and_unlock(procp);
	return 0;

error_exit:
	/*
	 * This entry has not been queued up so no worries about
	 * unlocked state and aio_map
	 */
	aio_proc_unlock(procp);
	aio_free_request(entryp);
error_noalloc:
	return result;
}


/*
 * aio_free_request - remove our reference on the user land map and
 * free the work queue entry resources.  The entry is off all lists
 * and has zero refcount, so no one can have a pointer to it.
 */
static void
aio_free_request(aio_workq_entry *entryp)
{
	if (entryp->aio_proc_link.tqe_prev || entryp->aio_workq_link.tqe_prev) {
		panic("aio_workq_entry %p being freed while still enqueued", entryp);
	}

	/* remove our reference to the user land map. */
	if (VM_MAP_NULL != entryp->aio_map) {
		vm_map_deallocate(entryp->aio_map);
	}

	/* remove our reference to thread which enqueued the request */
	if (entryp->context.vc_thread) {
		thread_deallocate(entryp->context.vc_thread);
	}
	kauth_cred_unref(&entryp->context.vc_ucred);

	zfree(aio_workq_zonep, entryp);
}


/*
 * aio_validate
 *
 * validate the aiocb passed in by one of the aio syscalls.
 */
static int
aio_validate(proc_t p, aio_workq_entry *entryp)
{
	struct fileproc *fp;
	int              flag;
	int              result;

	result = 0;

	if ((entryp->flags & AIO_LIO) != 0) {
		if (entryp->aiocb.aio_lio_opcode == LIO_READ) {
			entryp->flags |= AIO_READ;
		} else if (entryp->aiocb.aio_lio_opcode == LIO_WRITE) {
			entryp->flags |= AIO_WRITE;
		} else if (entryp->aiocb.aio_lio_opcode == LIO_NOP) {
			return 0;
		} else {
			return EINVAL;
		}
	}

	flag = FREAD;
	if ((entryp->flags & (AIO_WRITE | AIO_FSYNC | AIO_DSYNC)) != 0) {
		flag = FWRITE;
	}

	if ((entryp->flags & (AIO_READ | AIO_WRITE)) != 0) {
		if (entryp->aiocb.aio_nbytes > INT_MAX ||
		    entryp->aiocb.aio_buf == USER_ADDR_NULL ||
		    entryp->aiocb.aio_offset < 0) {
			return EINVAL;
		}
	}

	result = aio_sigev_validate(&entryp->aiocb.aio_sigevent);
	if (result) {
		return result;
	}

	/* validate the file descriptor and that the file was opened
	 * for the appropriate read / write access.
	 */
	proc_fdlock(p);

	fp = fp_get_noref_locked(p, entryp->aiocb.aio_fildes);
	if (fp == NULL) {
		result = EBADF;
	} else if ((fp->fp_glob->fg_flag & flag) == 0) {
		/* we don't have read or write access */
		result = EBADF;
	} else if (FILEGLOB_DTYPE(fp->fp_glob) != DTYPE_VNODE) {
		/* this is not a file */
		result = ESPIPE;
	} else {
		fp->fp_flags |= FP_AIOISSUED;
	}

	proc_fdunlock(p);

	return result;
}

/*
 * do_aio_completion_and_unlock.  Handle async IO completion.
 */
static void
do_aio_completion_and_unlock(proc_t p, aio_workq_entry *entryp,
    aio_entry_flags_t reason)
{
	aio_workq_entry *leader = entryp->lio_leader;
	int              lio_pending = 0;
	bool             do_signal, do_kevent;

	ASSERT_AIO_PROC_LOCK_OWNED(p);

	aio_proc_move_done_locked(p, entryp);
	entryp->flags |= reason;

	if (leader) {
		lio_pending = --leader->lio_pending;
		if (lio_pending < 0) {
			panic("lio_pending accounting mistake");
		}
		if (lio_pending == 0 && (leader->flags & AIO_LIO_WAIT)) {
			wakeup(leader);
		}
		entryp->lio_leader = NULL; /* no dangling pointers please */
	}

	/*
	 * need to handle case where a process is trying to exit, exec, or
	 * close and is currently waiting for active aio requests to complete.
	 * If AIO_CLEANUP_WAIT is set then we need to look to see if there are any
	 * other requests in the active queue for this process.  If there are
	 * none then wakeup using the AIO_CLEANUP_SLEEP_CHAN tsleep channel.
	 * If there are some still active then do nothing - we only want to
	 * wakeup when all active aio requests for the process are complete.
	 */
	do_signal = do_kevent = false;
	if (__improbable(entryp->flags & AIO_EXIT_WAIT)) {
		KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_cleanup_wait) | DBG_FUNC_NONE,
		    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
		    0, 0, 0);

		if (!aio_has_active_requests_for_process(p)) {
			/*
			 * no active aio requests for this process, continue exiting.  In this
			 * case, there should be no one else waiting on the proc in AIO...
			 */
			wakeup_one((caddr_t)&p->AIO_CLEANUP_SLEEP_CHAN);

			KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_cleanup_wake) | DBG_FUNC_NONE,
			    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
			    0, 0, 0);
		}
	} else if (entryp->aiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
		/*
		 * If this was the last request in the group, or not part of
		 * a group, and that a signal is desired, send one.
		 */
		do_signal = (lio_pending == 0);
	} else if (entryp->aiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
		/*
		 * For SIGEV_KEVENT, every AIO (even it is part of a group) would get
		 * a kevent notification.
		 */
		do_kevent = true;
	}

	if (__improbable(entryp->flags & AIO_CLOSE_WAIT)) {
		KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_cleanup_wait) | DBG_FUNC_NONE,
		    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
		    0, 0, 0);

		if (!aio_proc_has_active_requests_for_file(p, entryp->aiocb.aio_fildes)) {
			/* Can't wakeup_one(); multiple closes might be in progress. */
			wakeup(&p->AIO_CLEANUP_SLEEP_CHAN);

			KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_cleanup_wake) | DBG_FUNC_NONE,
			    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
			    0, 0, 0);
		}
	}

	aio_proc_unlock(p);

	if (do_signal) {
		KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_sig) | DBG_FUNC_NONE,
		    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
		    entryp->aiocb.aio_sigevent.sigev_signo, 0, 0);

		psignal(p, entryp->aiocb.aio_sigevent.sigev_signo);
	} else if (do_kevent) {
		KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_kevent) | DBG_FUNC_NONE,
		    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
		    entryp->aiocb.aio_sigevent.sigev_signo, 0, 0);

		/* We only support one event type for either completed/cancelled AIO. */
		lck_mtx_lock(&aio_klist_lock);
		KNOTE(&aio_klist, 1);
		lck_mtx_unlock(&aio_klist_lock);
	}

	/*
	 * A thread in aio_suspend() wants to known about completed IOs.  If it checked
	 * the done list before we moved our AIO there, then it already asserted its wait,
	 * and we can wake it up without holding the lock.  If it checked the list after
	 * we did our move, then it already has seen the AIO that we moved.  Herego, we
	 * can do our wakeup without holding the lock.
	 */
	wakeup(&p->AIO_SUSPEND_SLEEP_CHAN);
	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_completion_suspend_wake) | DBG_FUNC_NONE,
	    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp), 0, 0, 0);

	aio_entry_unref(entryp); /* see aio_try_enqueue_work_locked */
	if (leader) {
		aio_entry_unref(leader); /* see lio_listio */
	}
}


/*
 * do_aio_read
 */
static int
do_aio_read(aio_workq_entry *entryp)
{
	struct proc     *p = entryp->procp;
	struct fileproc *fp;
	int error;

	if ((error = fp_lookup(p, entryp->aiocb.aio_fildes, &fp, 0))) {
		return error;
	}

	if (fp->fp_glob->fg_flag & FREAD) {
		error = dofileread(&entryp->context, fp,
		    entryp->aiocb.aio_buf,
		    entryp->aiocb.aio_nbytes,
		    entryp->aiocb.aio_offset, FOF_OFFSET,
		    &entryp->returnval);
	} else {
		error = EBADF;
	}

	fp_drop(p, entryp->aiocb.aio_fildes, fp, 0);
	return error;
}


/*
 * do_aio_write
 */
static int
do_aio_write(aio_workq_entry *entryp)
{
	struct proc     *p = entryp->procp;
	struct fileproc *fp;
	int error;

	if ((error = fp_lookup(p, entryp->aiocb.aio_fildes, &fp, 0))) {
		return error;
	}

	if (fp->fp_glob->fg_flag & FWRITE) {
		int flags = 0;

		if ((fp->fp_glob->fg_flag & O_APPEND) == 0) {
			flags |= FOF_OFFSET;
		}

		/* NB: tell dofilewrite the offset, and to use the proc cred */
		error = dofilewrite(&entryp->context,
		    fp,
		    entryp->aiocb.aio_buf,
		    entryp->aiocb.aio_nbytes,
		    entryp->aiocb.aio_offset,
		    flags,
		    &entryp->returnval);
	} else {
		error = EBADF;
	}

	fp_drop(p, entryp->aiocb.aio_fildes, fp, 0);
	return error;
}


/*
 * aio_has_active_requests_for_process - return whether the process has active
 * requests pending.
 */
static bool
aio_has_active_requests_for_process(proc_t procp)
{
	return !TAILQ_EMPTY(&procp->p_aio_activeq);
}

/*
 * Called with the proc locked.
 */
static bool
aio_proc_has_active_requests_for_file(proc_t procp, int fd)
{
	aio_workq_entry *entryp;

	TAILQ_FOREACH(entryp, &procp->p_aio_activeq, aio_proc_link) {
		if (entryp->aiocb.aio_fildes == fd) {
			return true;
		}
	}

	return false;
}


/*
 * do_aio_fsync
 */
static int
do_aio_fsync(aio_workq_entry *entryp)
{
	struct proc            *p = entryp->procp;
	struct vnode           *vp;
	struct fileproc        *fp;
	int                     sync_flag;
	int                     error;

	/*
	 * We are never called unless either AIO_FSYNC or AIO_DSYNC are set.
	 *
	 * If AIO_DSYNC is set, we can tell the lower layers that it is OK
	 * to mark for update the metadata not strictly necessary for data
	 * retrieval, rather than forcing it to disk.
	 *
	 * If AIO_FSYNC is set, we have to also wait for metadata not really
	 * necessary to data retrival are committed to stable storage (e.g.
	 * atime, mtime, ctime, etc.).
	 *
	 * Metadata necessary for data retrieval ust be committed to stable
	 * storage in either case (file length, etc.).
	 */
	if (entryp->flags & AIO_FSYNC) {
		sync_flag = MNT_WAIT;
	} else {
		sync_flag = MNT_DWAIT;
	}

	error = fp_get_ftype(p, entryp->aiocb.aio_fildes, DTYPE_VNODE, ENOTSUP, &fp);
	if (error != 0) {
		entryp->returnval = -1;
		return error;
	}
	vp = fp_get_data(fp);

	if ((error = vnode_getwithref(vp)) == 0) {
		error = VNOP_FSYNC(vp, sync_flag, &entryp->context);

		(void)vnode_put(vp);
	} else {
		entryp->returnval = -1;
	}

	fp_drop(p, entryp->aiocb.aio_fildes, fp, 0);
	return error;
}


/*
 * is_already_queued - runs through our queues to see if the given
 * aiocbp / process is there.  Returns TRUE if there is a match
 * on any of our aio queues.
 *
 * Called with proc aio lock held (can be held spin)
 */
static boolean_t
is_already_queued(proc_t procp, user_addr_t aiocbp)
{
	aio_workq_entry *entryp;
	boolean_t        result;

	result = FALSE;

	/* look for matches on our queue of async IO requests that have completed */
	TAILQ_FOREACH(entryp, &procp->p_aio_doneq, aio_proc_link) {
		if (aiocbp == entryp->uaiocbp) {
			result = TRUE;
			goto ExitThisRoutine;
		}
	}

	/* look for matches on our queue of active async IO requests */
	TAILQ_FOREACH(entryp, &procp->p_aio_activeq, aio_proc_link) {
		if (aiocbp == entryp->uaiocbp) {
			result = TRUE;
			goto ExitThisRoutine;
		}
	}

ExitThisRoutine:
	return result;
}


/*
 * aio initialization
 */
__private_extern__ void
aio_init(void)
{
	for (int i = 0; i < AIO_NUM_WORK_QUEUES; i++) {
		aio_workq_init(&aio_anchor.aio_async_workqs[i]);
	}

	if (bootarg_aio_new_workq) {
		printf("New aio workqueue implementation selected\n");
	} else {
		_aio_create_worker_threads(aio_worker_threads);
	}

	klist_init(&aio_klist);

	clock_interval_to_absolutetime_interval(aio_wq_reduce_pool_window.usecs,
	    NSEC_PER_USEC, &aio_wq_reduce_pool_window.abstime);
}


/*
 * aio worker threads created here.
 */
__private_extern__ void
_aio_create_worker_threads(int num)
{
	int i;

	/* create some worker threads to handle the async IO requests */
	for (i = 0; i < num; i++) {
		thread_t                myThread;

		if (KERN_SUCCESS != kernel_thread_start(aio_work_thread, NULL, &myThread)) {
			printf("%s - failed to create a work thread \n", __FUNCTION__);
		} else {
			thread_deallocate(myThread);
		}
	}
}

/*
 * Return the current activation utask
 */
task_t
get_aiotask(void)
{
	return current_uthread()->uu_aio_task;
}


/*
 * In the case of an aiocb from a
 * 32-bit process we need to expand some longs and pointers to the correct
 * sizes in order to let downstream code always work on the same type of
 * aiocb (in our case that is a user_aiocb)
 */
static void
do_munge_aiocb_user32_to_user(struct user32_aiocb *my_aiocbp, struct user_aiocb *the_user_aiocbp)
{
	the_user_aiocbp->aio_fildes = my_aiocbp->aio_fildes;
	the_user_aiocbp->aio_offset = my_aiocbp->aio_offset;
	the_user_aiocbp->aio_buf = CAST_USER_ADDR_T(my_aiocbp->aio_buf);
	the_user_aiocbp->aio_nbytes = my_aiocbp->aio_nbytes;
	the_user_aiocbp->aio_reqprio = my_aiocbp->aio_reqprio;
	the_user_aiocbp->aio_lio_opcode = my_aiocbp->aio_lio_opcode;

	/* special case here.  since we do not know if sigev_value is an */
	/* int or a ptr we do NOT cast the ptr to a user_addr_t.   This  */
	/* means if we send this info back to user space we need to remember */
	/* sigev_value was not expanded for the 32-bit case.  */
	/* NOTE - this does NOT affect us since we don't support sigev_value */
	/* yet in the aio context.  */
	//LP64
	the_user_aiocbp->aio_sigevent.sigev_notify = my_aiocbp->aio_sigevent.sigev_notify;
	the_user_aiocbp->aio_sigevent.sigev_signo = my_aiocbp->aio_sigevent.sigev_signo;
	the_user_aiocbp->aio_sigevent.sigev_value.sival_ptr =
	    my_aiocbp->aio_sigevent.sigev_value.sival_ptr;
	the_user_aiocbp->aio_sigevent.sigev_notify_function =
	    CAST_USER_ADDR_T(my_aiocbp->aio_sigevent.sigev_notify_function);
	the_user_aiocbp->aio_sigevent.sigev_notify_attributes =
	    CAST_USER_ADDR_T(my_aiocbp->aio_sigevent.sigev_notify_attributes);
}

/* Similar for 64-bit user process, so that we don't need to satisfy
 * the alignment constraints of the original user64_aiocb
 */
#if !__LP64__
__dead2
#endif
static void
do_munge_aiocb_user64_to_user(struct user64_aiocb *my_aiocbp, struct user_aiocb *the_user_aiocbp)
{
#if __LP64__
	the_user_aiocbp->aio_fildes = my_aiocbp->aio_fildes;
	the_user_aiocbp->aio_offset = my_aiocbp->aio_offset;
	the_user_aiocbp->aio_buf = my_aiocbp->aio_buf;
	the_user_aiocbp->aio_nbytes = my_aiocbp->aio_nbytes;
	the_user_aiocbp->aio_reqprio = my_aiocbp->aio_reqprio;
	the_user_aiocbp->aio_lio_opcode = my_aiocbp->aio_lio_opcode;

	the_user_aiocbp->aio_sigevent.sigev_notify = my_aiocbp->aio_sigevent.sigev_notify;
	the_user_aiocbp->aio_sigevent.sigev_signo = my_aiocbp->aio_sigevent.sigev_signo;
	the_user_aiocbp->aio_sigevent.sigev_value.sival_ptr =
	    my_aiocbp->aio_sigevent.sigev_value.sival_ptr;
	the_user_aiocbp->aio_sigevent.sigev_notify_function =
	    my_aiocbp->aio_sigevent.sigev_notify_function;
	the_user_aiocbp->aio_sigevent.sigev_notify_attributes =
	    my_aiocbp->aio_sigevent.sigev_notify_attributes;
#else
#pragma unused(my_aiocbp, the_user_aiocbp)
	panic("64bit process on 32bit kernel is not supported");
#endif
}


static int
filt_aioattach(struct knote *kn, struct kevent_qos_s *kev)
{
	aio_workq_entry *entryp = (aio_workq_entry *)kev->data;

	/* Don't allow kevent registration from the user-space. */
	if ((kev->flags & EV_KERNEL) == 0) {
		return EPERM;
	}

	kev->flags &= ~EV_KERNEL;
	/* Clear the 'kn_fflags' state afte the knote has been processed. */
	kn->kn_flags |= EV_CLEAR;

	/* Associate the knote with the AIO work. */
	knote_kn_hook_set_raw(kn, (void *)entryp);

	lck_mtx_lock(&aio_klist_lock);
	KNOTE_ATTACH(&aio_klist, kn);
	lck_mtx_unlock(&aio_klist_lock);

	return 0;
}

static void
filt_aiodetach(struct knote *kn)
{
	lck_mtx_lock(&aio_klist_lock);
	KNOTE_DETACH(&aio_klist, kn);
	lck_mtx_unlock(&aio_klist_lock);
}

/*
 * The 'f_event' is called with 'aio_klist_lock' held when KNOTE() was called
 * in do_aio_completion_and_unlock().
 */
static int
filt_aioevent(struct knote *kn, long hint)
{
	aio_workq_entry *entryp;
	int activate = 0;

	/*
	 * The 'f_event' and 'f_process' can run concurrently so it is possible
	 * the aio_workq_entry has been detached from the knote when the
	 * filt_aioprocess() was called earlier. In this case, we will skip
	 * activating the event.
	 */
	entryp = knote_kn_hook_get_raw(kn);
	if (__improbable(entryp == NULL)) {
		goto out;
	}

	/* We can only activate the filter if the AIO work has completed. */
	if (entryp->flags & AIO_COMPLETED) {
		kn->kn_fflags |= hint;
		activate = FILTER_ACTIVE;
	}

out:
	return activate;
}

static int
filt_aiotouch(struct knote *kn, struct kevent_qos_s *kev)
{
	panic("%s: kn %p kev %p (NOT EXPECTED TO BE CALLED!!)", __func__, kn, kev);
}

static int
filt_aioprocess(struct knote *kn, struct kevent_qos_s *kev)
{
	aio_workq_entry *entryp;
	proc_t p;
	int res = 0;

	entryp = knote_kn_hook_get_raw(kn);
	assert(entryp);
	p = entryp->procp;

	lck_mtx_lock(&aio_klist_lock);
	if (kn->kn_fflags) {
		/* Propagate the error status and return value back to the user. */
		kn->kn_ext[0] = entryp->errorval;
		kn->kn_ext[1] = entryp->returnval;
		knote_fill_kevent(kn, kev, 0);
		knote_kn_hook_set_raw(kn, NULL);

		aio_proc_lock(p);
		aio_proc_remove_done_locked(p, entryp);
		aio_proc_unlock(p);
		aio_entry_unref(entryp);

		res = FILTER_ACTIVE;
	}
	lck_mtx_unlock(&aio_klist_lock);

	return res;
}

SECURITY_READ_ONLY_EARLY(struct filterops) aio_filtops = {
	.f_isfd = 0,
	.f_attach = filt_aioattach,
	.f_detach = filt_aiodetach,
	.f_event = filt_aioevent,
	.f_touch = filt_aiotouch,
	.f_process = filt_aioprocess,
};

#pragma mark per process aio workqueue

/*
 * The per process workq threads call this function to handle the aio request. The threads
 * belong to the same process so we don't need to change the vm maps as we would for kernel
 * threads.
 */
static int
workq_aio_process_entry(aio_workq_entry *entryp)
{
	proc_t p = entryp->procp;
	int error;

	assert(current_proc() == p && current_thread() != vfs_context_thread(&entryp->context));

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_WQ_process_entry) | DBG_FUNC_START,
	    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
	    entryp->flags, 0, 0);

	if ((entryp->flags & AIO_READ) != 0) {
		error = do_aio_read(entryp);
	} else if ((entryp->flags & AIO_WRITE) != 0) {
		error = do_aio_write(entryp);
	} else if ((entryp->flags & (AIO_FSYNC | AIO_DSYNC)) != 0) {
		if ((entryp->flags & AIO_FSYNC) != 0) {
			/*
			 * Check for unfinished operations on the same file
			 * in this proc's queue.
			 */
			aio_proc_lock_spin(p);
			if (aio_delay_fsync_request(entryp)) {
				/* It needs to be delayed.  Put it back on the end of the work queue */
				KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_fsync_delay) | DBG_FUNC_NONE,
				    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
				    0, 0, 0);

				/* The references on this entry havn't been consumed */
				if (!workq_aio_entry_add_locked(p, entryp)) {
					entryp->errorval = ECANCELED;
					entryp->returnval = -1;

					/* Now it's officially cancelled.  Do the completion */
					KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_cancel_async_workq) | DBG_FUNC_NONE,
					    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
					    entryp->aiocb.aio_fildes, 0, 0);

					do_aio_completion_and_unlock(p, entryp, AIO_CANCELLED);
				} else {
					workq_aio_wakeup_thread_and_unlock(p);
				}
				return 0;
			}
			aio_proc_unlock(entryp->procp);
		}
		error = do_aio_fsync(entryp);
	} else {
		error = EINVAL;
	}

	KERNEL_DEBUG(BSDDBG_CODE(DBG_BSD_AIO, AIO_WQ_process_entry) | DBG_FUNC_END,
	    VM_KERNEL_ADDRPERM(p), VM_KERNEL_ADDRPERM(entryp->uaiocbp),
	    entryp->errorval, entryp->returnval, 0);

	/* we're done with the IO request so pop it off the active queue and */
	/* push it on the done queue */
	aio_proc_lock(p);
	entryp->errorval = error;
	do_aio_completion_and_unlock(p, entryp, AIO_COMPLETED);
	return 0;
}

/*
 * The functions below implement a workqueue for aio which is taken from the
 * workqueue implementation for libdispatch/pthreads. They are stripped down versions
 * of the corresponding functions for libdispatch/pthreads.
 */

static int
aio_workq_sysctl_handle_usecs SYSCTL_HANDLER_ARGS
{
#pragma unused(arg2)
	struct aio_workq_usec_var *v = arg1;
	int error = sysctl_handle_int(oidp, &v->usecs, 0, req);
	if (error || !req->newptr) {
		return error;
	}
	clock_interval_to_absolutetime_interval(v->usecs, NSEC_PER_USEC,
	    &v->abstime);
	return 0;
}

#pragma mark wq_flags

#define AIO_WQ_DEAD 0x1000

static inline uint32_t
_wa_flags(workq_aio_t wq_aio)
{
	return os_atomic_load(&wq_aio->wa_flags, relaxed);
}

static inline bool
_wq_exiting(workq_aio_t wq_aio)
{
	return _wa_flags(wq_aio) & WQ_EXITING;
}

static inline bool
_wq_dead(workq_aio_t wq_aio)
{
	return _wa_flags(wq_aio) & AIO_WQ_DEAD;
}

#define AIO_WQPTR_IS_INITING_VALUE ((workq_aio_t)~(uintptr_t)0)

static workq_aio_t
proc_get_aio_wqptr_fast(struct proc *p)
{
	return os_atomic_load(&p->p_aio_wqptr, relaxed);
}

static workq_aio_t
proc_get_aio_wqptr(struct proc *p)
{
	workq_aio_t wq_aio = proc_get_aio_wqptr_fast(p);
	return wq_aio == AIO_WQPTR_IS_INITING_VALUE ? NULL : wq_aio;
}

static void
proc_set_aio_wqptr(struct proc *p, workq_aio_t wq_aio)
{
	wq_aio = os_atomic_xchg(&p->p_aio_wqptr, wq_aio, release);
	if (wq_aio == AIO_WQPTR_IS_INITING_VALUE) {
		proc_lock(p);
		thread_wakeup(&p->p_aio_wqptr);
		proc_unlock(p);
	}
}

static bool
proc_init_aio_wqptr_or_wait(struct proc *p)
{
	workq_aio_t wq_aio;

	proc_lock(p);
	wq_aio = os_atomic_load(&p->p_aio_wqptr, relaxed);

	if (wq_aio == NULL) {
		os_atomic_store(&p->p_aio_wqptr, AIO_WQPTR_IS_INITING_VALUE, relaxed);
		proc_unlock(p);
		return true;
	}

	if (wq_aio == AIO_WQPTR_IS_INITING_VALUE) {
		assert_wait(&p->p_aio_wqptr, THREAD_UNINT);
		proc_unlock(p);
		thread_block(THREAD_CONTINUE_NULL);
	} else {
		proc_unlock(p);
	}
	return false;
}

static inline event_t
workq_aio_parked_wait_event(struct uthread *uth)
{
	return (event_t)&uth->uu_workq_stackaddr;
}

static inline void
workq_aio_thread_wakeup(struct uthread *uth)
{
	thread_wakeup_thread(workq_aio_parked_wait_event(uth), get_machthread(uth));
}

/*
 * Routine:	workq_aio_mark_exiting
 *
 * Function:	Mark the work queue such that new threads will not be added to the
 *		work queue after we return.
 *
 * Conditions:	Called against the current process.
 */
static void
workq_aio_mark_exiting(proc_t p)
{
	workq_aio_t wq_aio = proc_get_aio_wqptr(p);
	uint32_t wq_flags;

	if (!wq_aio) {
		return;
	}

	wq_flags = os_atomic_or_orig(&wq_aio->wa_flags, WQ_EXITING, relaxed);
	if (__improbable(wq_flags & WQ_EXITING)) {
		panic("workq_aio_mark_exiting_locked called twice");
	}

	/*
	 * Opportunistically try to cancel thread calls that are likely in flight.
	 * workq_aio_exit() will do the proper cleanup.
	 */
	if (wq_flags & WQ_DEATH_CALL_SCHEDULED) {
		thread_call_cancel(wq_aio->wa_death_call);
	}
}

static void
workq_aio_exit(proc_t p)
{
	workq_aio_t wq_aio;

	wq_aio = os_atomic_xchg(&p->p_aio_wqptr, NULL, release);

	if (!wq_aio) {
		return;
	}

	/*
	 * Thread calls are always scheduled by the proc itself or under the
	 * workqueue spinlock if WQ_EXITING is not yet set.
	 *
	 * Either way, when this runs, the proc has no threads left beside
	 * the one running this very code, so we know no thread call can be
	 * dispatched anymore.
	 */

	thread_call_cancel_wait(wq_aio->wa_death_call);
	thread_call_free(wq_aio->wa_death_call);

	/*
	 * Clean up workqueue data structures for threads that exited and
	 * didn't get a chance to clean up after themselves.
	 *
	 * idle/new threads should have been interrupted and died on their own
	 */

	assert(TAILQ_EMPTY(&wq_aio->wa_aioq_entries));
	assert(TAILQ_EMPTY(&wq_aio->wa_thrunlist));

	if (wq_aio->wa_nthreads) {
		os_atomic_or(&wq_aio->wa_flags, AIO_WQ_DEAD, relaxed);
		aio_proc_lock_spin(p);
		if (wq_aio->wa_nthreads) {
			struct uthread *uth;

			TAILQ_FOREACH(uth, &wq_aio->wa_thidlelist, uu_workq_entry) {
				if (uth->uu_workq_flags & UT_WORKQ_DYING) {
					workq_aio_thread_wakeup(uth);
					continue;
				}
				wq_aio->wa_thdying_count++;
				uth->uu_workq_flags |= UT_WORKQ_DYING;
				workq_aio_thread_wakeup(uth);
			}
			while (wq_aio->wa_nthreads) {
				msleep(&wq_aio->wa_nthreads, aio_proc_mutex(p), PRIBIO | PSPIN, "aio_workq_exit", 0);
			}
		}
		aio_proc_unlock(p);
	}

	assertf(TAILQ_EMPTY(&wq_aio->wa_thidlelist),
	    "wa_thidlecount = %d, wa_thdying_count = %d",
	    wq_aio->wa_thidlecount, wq_aio->wa_thdying_count);
	assertf(wq_aio->wa_thidlecount == 0,
	    "wa_thidlecount = %d, wa_thdying_count = %d",
	    wq_aio->wa_thidlecount, wq_aio->wa_thdying_count);
	assertf(wq_aio->wa_thdying_count == 0,
	    "wa_thdying_count = %d", wq_aio->wa_thdying_count);

	kfree_type(workq_aio_s, wq_aio);
}

static int
workq_aio_open(struct proc *p)
{
	workq_aio_t wq_aio;
	int error = 0;

	if (proc_get_aio_wqptr(p) == NULL) {
		if (proc_init_aio_wqptr_or_wait(p) == FALSE) {
			assert(proc_get_aio_wqptr(p) != NULL);
			goto out;
		}

		wq_aio = kalloc_type(workq_aio_s, Z_WAITOK | Z_ZERO);

		wq_aio->wa_proc = p;

		TAILQ_INIT(&wq_aio->wa_thidlelist);
		TAILQ_INIT(&wq_aio->wa_thrunlist);
		TAILQ_INIT(&wq_aio->wa_aioq_entries);

		wq_aio->wa_death_call = thread_call_allocate_with_options(
			workq_aio_kill_old_threads_call, wq_aio,
			THREAD_CALL_PRIORITY_USER, THREAD_CALL_OPTIONS_ONCE);

		proc_set_aio_wqptr(p, wq_aio);
	}
out:
	return error;
}

#pragma mark aio workqueue idle thread accounting

static inline struct uthread *
workq_oldest_killable_idle_aio_thread(workq_aio_t wq_aio)
{
	return TAILQ_LAST(&wq_aio->wa_thidlelist, workq_aio_uthread_head);
}

static inline uint64_t
workq_kill_delay_for_idle_aio_thread()
{
	return aio_wq_reduce_pool_window.abstime;
}

static inline bool
workq_should_kill_idle_aio_thread(struct uthread *uth, uint64_t now)
{
	uint64_t delay = workq_kill_delay_for_idle_aio_thread();
	return now - uth->uu_save.uus_workq_park_data.idle_stamp > delay;
}

static void
workq_aio_death_call_schedule(workq_aio_t wq_aio, uint64_t deadline)
{
	uint32_t wa_flags = os_atomic_load(&wq_aio->wa_flags, relaxed);

	if (wa_flags & (WQ_EXITING | WQ_DEATH_CALL_SCHEDULED)) {
		return;
	}
	os_atomic_or(&wq_aio->wa_flags, WQ_DEATH_CALL_SCHEDULED, relaxed);

	/*
	 * <rdar://problem/13139182> Due to how long term timers work, the leeway
	 * can't be too short, so use 500ms which is long enough that we will not
	 * wake up the CPU for killing threads, but short enough that it doesn't
	 * fall into long-term timer list shenanigans.
	 */
	thread_call_enter_delayed_with_leeway(wq_aio->wa_death_call, NULL, deadline,
	    aio_wq_reduce_pool_window.abstime / 10,
	    THREAD_CALL_DELAY_LEEWAY | THREAD_CALL_DELAY_USER_BACKGROUND);
}

/*
 * `decrement` is set to the number of threads that are no longer dying:
 * - because they have been resuscitated just in time (workq_pop_idle_thread)
 * - or have been killed (workq_thread_terminate).
 */
static void
workq_aio_death_policy_evaluate(workq_aio_t wq_aio, uint16_t decrement)
{
	struct uthread *uth;

	assert(wq_aio->wa_thdying_count >= decrement);
#if 0
	if (decrement) {
		printf("VV_DEBUG_AIO : %s:%d : pid = %d, ctid = %d, after decrement thdying_count = %d\n",
		    __FUNCTION__, __LINE__, proc_pid(current_proc()), thread_get_ctid(thr),
		    wq_aio->wa_thdying_count - decrement);
	}
#endif

	if ((wq_aio->wa_thdying_count -= decrement) > 0) {
		return;
	}

	if (wq_aio->wa_thidlecount <= 1) {
		return;
	}

	if (((uth = workq_oldest_killable_idle_aio_thread(wq_aio)) == NULL)) {
		return;
	}

	uint64_t now = mach_absolute_time();
	uint64_t delay = workq_kill_delay_for_idle_aio_thread();

	if (now - uth->uu_save.uus_workq_park_data.idle_stamp > delay) {
		if (!(uth->uu_workq_flags & UT_WORKQ_DYING)) {
			wq_aio->wa_thdying_count++;
			uth->uu_workq_flags |= UT_WORKQ_DYING;
		}
		workq_aio_thread_wakeup(uth);
		return;
	}

	workq_aio_death_call_schedule(wq_aio,
	    uth->uu_save.uus_workq_park_data.idle_stamp + delay);
}

static void
workq_aio_kill_old_threads_call(void *param0, void *param1 __unused)
{
	workq_aio_t wq_aio = param0;

	aio_proc_lock_spin(wq_aio->wa_proc);
	WQ_AIO_TRACE_WQ(AIO_WQ_aio_death_call | DBG_FUNC_START, wq_aio);
	os_atomic_andnot(&wq_aio->wa_flags, WQ_DEATH_CALL_SCHEDULED, relaxed);
	workq_aio_death_policy_evaluate(wq_aio, 0);
	WQ_AIO_TRACE_WQ(AIO_WQ_aio_death_call | DBG_FUNC_END, wq_aio);
	aio_proc_unlock(wq_aio->wa_proc);;
}

#define WORKQ_UNPARK_FOR_DEATH_WAS_IDLE 0x1
#define WQ_SETUP_NONE  0

__attribute__((noreturn, noinline))
static void
workq_aio_unpark_for_death_and_unlock(proc_t p, workq_aio_t wq_aio,
    struct uthread *uth, uint32_t death_flags, __unused uint32_t setup_flags)
{
	if (death_flags & WORKQ_UNPARK_FOR_DEATH_WAS_IDLE) {
		wq_aio->wa_thidlecount--;
		TAILQ_REMOVE(&wq_aio->wa_thidlelist, uth, uu_workq_entry);
	}

	if (uth->uu_workq_flags & UT_WORKQ_DYING) {
		wq_aio->wa_thdying_count--;
	}
	assert(wq_aio->wa_nthreads > 0);
	wq_aio->wa_nthreads--;

	WQ_AIO_TRACE_WQ(AIO_WQ_aio_thread_terminate | DBG_FUNC_NONE, wq_aio);

	if (_wq_dead(wq_aio) && (wq_aio->wa_nthreads == 0)) {
		wakeup(&wq_aio->wa_nthreads);
	}

	aio_proc_unlock(p);

	thread_t th = get_machthread(uth);
	assert(th == current_thread());

	thread_deallocate(th);
	thread_terminate(th);
	thread_exception_return();
	__builtin_unreachable();
}

static void
workq_push_idle_aio_thread(proc_t p, workq_aio_t wq_aio, struct uthread *uth,
    uint32_t setup_flags)
{
	uint64_t now = mach_absolute_time();

	uth->uu_workq_flags &= ~(UT_WORKQ_RUNNING);
	TAILQ_REMOVE(&wq_aio->wa_thrunlist, uth, uu_workq_entry);

	uth->uu_save.uus_workq_park_data.idle_stamp = now;

	struct uthread *oldest = workq_oldest_killable_idle_aio_thread(wq_aio);
	uint16_t cur_idle = wq_aio->wa_thidlecount;

	if (_wq_exiting(wq_aio) || (wq_aio->wa_thdying_count == 0 && oldest &&
	    workq_should_kill_idle_aio_thread(oldest, now))) {
		/*
		 * Immediately kill threads if we have too may of them.
		 *
		 * And swap "place" with the oldest one we'd have woken up.
		 * This is a relatively desperate situation where we really
		 * need to kill threads quickly and it's best to kill
		 * the one that's currently on core than context switching.
		 */
		if (oldest) {
			oldest->uu_save.uus_workq_park_data.idle_stamp = now;
			TAILQ_REMOVE(&wq_aio->wa_thidlelist, oldest, uu_workq_entry);
			TAILQ_INSERT_HEAD(&wq_aio->wa_thidlelist, oldest, uu_workq_entry);
		}

		if (!(uth->uu_workq_flags & UT_WORKQ_DYING)) {
			wq_aio->wa_thdying_count++;
			uth->uu_workq_flags |= UT_WORKQ_DYING;
		}
		workq_aio_unpark_for_death_and_unlock(p, wq_aio, uth, 0, setup_flags);
		__builtin_unreachable();
	}

	struct uthread *tail = TAILQ_LAST(&wq_aio->wa_thidlelist, workq_aio_uthread_head);

	cur_idle += 1;
	wq_aio->wa_thidlecount = cur_idle;
	uth->uu_save.uus_workq_park_data.has_stack = false;
	TAILQ_INSERT_HEAD(&wq_aio->wa_thidlelist, uth, uu_workq_entry);

	if (!tail) {
		uint64_t delay = workq_kill_delay_for_idle_aio_thread();
		workq_aio_death_call_schedule(wq_aio, now + delay);
	}
}

/*
 * We have no work to do, park ourselves on the idle list.
 *
 * Consumes the workqueue lock and does not return.
 */
__attribute__((noreturn, noinline))
static void
workq_aio_park_and_unlock(proc_t p, workq_aio_t wq_aio, struct uthread *uth,
    uint32_t setup_flags)
{
	assert(uth == current_uthread());
	assert(uth->uu_kqr_bound == NULL);

	workq_push_idle_aio_thread(p, wq_aio, uth, setup_flags); // may not return

	if (uth->uu_workq_flags & UT_WORKQ_DYING) {
		workq_aio_unpark_for_death_and_unlock(p, wq_aio, uth,
		    WORKQ_UNPARK_FOR_DEATH_WAS_IDLE, setup_flags);
		__builtin_unreachable();
	}

	WQ_AIO_TRACE_WQ(AIO_WQ_aio_thread_park | DBG_FUNC_NONE, wq_aio);

	thread_set_pending_block_hint(get_machthread(uth), kThreadWaitParkedWorkQueue);
	/* XXX this should probably be THREAD_UNINT */
	assert_wait(workq_aio_parked_wait_event(uth), THREAD_INTERRUPTIBLE);
	aio_proc_unlock(p);
	thread_block(workq_aio_unpark_continue);
	__builtin_unreachable();
}

#define WORKQ_POLICY_INIT(qos) \
	         (struct uu_workq_policy){ .qos_req = (qos), .qos_bucket = (qos) }

/*
 * This function is always called with the workq lock.
 */
static void
workq_aio_thread_reset_pri(struct uthread *uth, thread_t src_th)
{
	thread_t th = get_machthread(uth);
	thread_qos_t qos = (thread_qos_t)proc_get_effective_thread_policy(src_th, TASK_POLICY_QOS);
	int priority = 31;
	int policy = POLICY_TIMESHARE;

	uth->uu_workq_pri = WORKQ_POLICY_INIT(qos);
	thread_set_workq_pri(th, qos, priority, policy);
}

static inline void
workq_aio_thread_set_type(struct uthread *uth, uint16_t flags)
{
	uth->uu_workq_flags &= ~(UT_WORKQ_OVERCOMMIT | UT_WORKQ_COOPERATIVE);
	uth->uu_workq_flags |= flags;
}

__attribute__((noreturn, noinline))
static void
workq_aio_unpark_select_req_or_park_and_unlock(proc_t p, workq_aio_t wq_aio,
    struct uthread *uth, uint32_t setup_flags)
{
	aio_workq_entry *entryp;
	thread_t last_thread = NULL;

	WQ_AIO_TRACE_WQ(AIO_WQ_aio_select_req | DBG_FUNC_START, wq_aio);
	thread_freeze_base_pri(get_machthread(uth));
	workq_aio_thread_set_type(uth, 0);
	while ((entryp = TAILQ_FIRST(&wq_aio->wa_aioq_entries))) {
		if (__improbable(_wq_exiting(wq_aio))) {
			break;
		}

		TAILQ_REMOVE(&wq_aio->wa_aioq_entries, entryp, aio_workq_link);
		entryp->aio_workq_link.tqe_prev = NULL; /* Not on a workq */

		aio_proc_unlock(p);

		thread_t thr = vfs_context_thread(&entryp->context);
		if (last_thread != thr) {
			workq_aio_thread_reset_pri(uth, thr);
			last_thread = thr;
		}

		/* this frees references to workq entry */
		workq_aio_process_entry(entryp);

		ast_check_async_thread();

		aio_proc_lock_spin(p);
	}
	WQ_AIO_TRACE_WQ(AIO_WQ_aio_select_req | DBG_FUNC_END, wq_aio);
	thread_unfreeze_base_pri(get_machthread(uth));
	workq_aio_park_and_unlock(p, wq_aio, uth, setup_flags);
}

/*
 * parked idle thread wakes up
 */
__attribute__((noreturn, noinline))
static void
workq_aio_unpark_continue(void *parameter __unused, wait_result_t wr)
{
	thread_t th = current_thread();
	struct uthread *uth = get_bsdthread_info(th);
	proc_t p = current_proc();
	workq_aio_t wq_aio = proc_get_aio_wqptr_fast(p);

	aio_proc_lock_spin(p);

	if (__probable(uth->uu_workq_flags & UT_WORKQ_RUNNING)) {
		workq_aio_unpark_select_req_or_park_and_unlock(p, wq_aio, uth, WQ_SETUP_NONE);
		__builtin_unreachable();
	}

	if (__probable(wr == THREAD_AWAKENED)) {
		/*
		 * We were set running, but for the purposes of dying.
		 */
		assert(uth->uu_workq_flags & UT_WORKQ_DYING);
		assert((uth->uu_workq_flags & UT_WORKQ_NEW) == 0);
	} else {
		/*
		 * workaround for <rdar://problem/38647347>,
		 * in case we do hit userspace, make sure calling
		 * workq_thread_terminate() does the right thing here,
		 * and if we never call it, that workq_exit() will too because it sees
		 * this thread on the runlist.
		 */
		assert(wr == THREAD_INTERRUPTED);

		if (!(uth->uu_workq_flags & UT_WORKQ_DYING)) {
			wq_aio->wa_thdying_count++;
			uth->uu_workq_flags |= UT_WORKQ_DYING;
		}
	}

	workq_aio_unpark_for_death_and_unlock(p, wq_aio, uth,
	    WORKQ_UNPARK_FOR_DEATH_WAS_IDLE, WQ_SETUP_NONE);

	__builtin_unreachable();
}

/*
 * Called by thread_create_workq_aio_waiting() during thread initialization, before
 * assert_wait, before the thread has been started.
 */
event_t
aio_workq_thread_init_and_wq_lock(task_t task, thread_t th)
{
	struct uthread *uth = get_bsdthread_info(th);

	uth->uu_workq_flags = UT_WORKQ_NEW;
	uth->uu_workq_pri = WORKQ_POLICY_INIT(THREAD_QOS_LEGACY);
	uth->uu_workq_thport = MACH_PORT_NULL;
	uth->uu_workq_stackaddr = 0;
	uth->uu_workq_pthread_kill_allowed = 0;

	thread_set_tag(th, THREAD_TAG_AIO_WORKQUEUE);
	thread_reset_workq_qos(th, THREAD_QOS_LEGACY);

	aio_proc_lock(get_bsdtask_info(task));
	return workq_aio_parked_wait_event(uth);
}

/**
 * Try to add a new workqueue thread for aio.
 *
 * - called with workq lock held
 * - dropped and retaken around thread creation
 * - return with workq lock held
 * - aio threads do not call into pthread functions to setup or destroy stacks.
 */
static kern_return_t
workq_aio_add_new_thread(proc_t p, workq_aio_t wq_aio)
{
	kern_return_t kret;
	thread_t th;

	wq_aio->wa_nthreads++;

	aio_proc_unlock(p);

	kret = thread_create_aio_workq_waiting(proc_task(p),
	    workq_aio_unpark_continue,
	    &th);

	if (kret != KERN_SUCCESS) {
		WQ_AIO_TRACE(AIO_WQ_aio_thread_create_failed | DBG_FUNC_NONE, wq_aio,
		    kret, 0, 0, 0);
		goto out;
	}

	/*
	 * thread_create_aio_workq_waiting() will return with the wq lock held
	 * on success, because it calls workq_thread_init_and_wq_lock().
	 */
	struct uthread *uth = get_bsdthread_info(th);
	TAILQ_INSERT_TAIL(&wq_aio->wa_thidlelist, uth, uu_workq_entry);
	wq_aio->wa_thidlecount++;
	uth->uu_workq_flags &= ~UT_WORKQ_NEW;
	WQ_AIO_TRACE_WQ(AIO_WQ_aio_thread_create | DBG_FUNC_NONE, wq_aio);
	return kret;

out:
	aio_proc_lock(p);
	/*
	 * Do not redrive here if we went under wq_max_threads again,
	 * it is the responsibility of the callers of this function
	 * to do so when it fails.
	 */
	wq_aio->wa_nthreads--;
	return kret;
}

static void
workq_aio_wakeup_thread_internal(proc_t p, bool unlock)
{
	workq_aio_t wq_aio = proc_get_aio_wqptr(p);
	bool needs_wakeup = false;
	struct uthread *uth = NULL;

	if (!wq_aio) {
		goto out;
	}

	uth = TAILQ_FIRST(&wq_aio->wa_thidlelist);
	while (!uth && (wq_aio->wa_nthreads < WORKQUEUE_AIO_MAXTHREADS) &&
	    !(thread_get_tag(current_thread()) & THREAD_TAG_AIO_WORKQUEUE)) {
		if (workq_aio_add_new_thread(p, wq_aio) != KERN_SUCCESS) {
			break;
		}
		uth = TAILQ_FIRST(&wq_aio->wa_thidlelist);
	}

	if (!uth) {
		goto out;
	}

	TAILQ_REMOVE(&wq_aio->wa_thidlelist, uth, uu_workq_entry);
	wq_aio->wa_thidlecount--;

	TAILQ_INSERT_TAIL(&wq_aio->wa_thrunlist, uth, uu_workq_entry);
	assert((uth->uu_workq_flags & UT_WORKQ_RUNNING) == 0);
	uth->uu_workq_flags |= UT_WORKQ_RUNNING;

	WQ_AIO_TRACE_WQ(AIO_WQ_aio_thread_wakeup | DBG_FUNC_NONE, wq_aio);

	if (__improbable(uth->uu_workq_flags & UT_WORKQ_DYING)) {
		uth->uu_workq_flags ^= UT_WORKQ_DYING;
		workq_aio_death_policy_evaluate(wq_aio, 1);
		needs_wakeup = false;
	} else {
		needs_wakeup = true;
	}
out:
	if (unlock) {
		aio_proc_unlock(p);
	}

	if (uth && needs_wakeup) {
		workq_aio_thread_wakeup(uth);
	}
}

static void
workq_aio_wakeup_thread_and_unlock(proc_t p)
{
	return workq_aio_wakeup_thread_internal(p, true);
}

static void
workq_aio_wakeup_thread(proc_t p)
{
	return workq_aio_wakeup_thread_internal(p, false);
}

void
workq_aio_prepare(struct proc *p)
{
	workq_aio_t wq_aio = proc_get_aio_wqptr(p);

	if (__improbable(!wq_aio && !proc_in_teardown(p))) {
		workq_aio_open(p);
	}
}

bool
workq_aio_entry_add_locked(struct proc *p, aio_workq_entry *entryp)
{
	workq_aio_t wq_aio = proc_get_aio_wqptr(p);
	bool ret = false;

	ASSERT_AIO_PROC_LOCK_OWNED(p);

	if (!proc_in_teardown(p) && wq_aio && !_wq_exiting(wq_aio)) {
		TAILQ_INSERT_TAIL(&wq_aio->wa_aioq_entries, entryp, aio_workq_link);
		ret = true;
	}

	return ret;
}

bool
workq_aio_entry_remove_locked(struct proc *p, aio_workq_entry *entryp)
{
	workq_aio_t  wq_aio = proc_get_aio_wqptr(p);

	ASSERT_AIO_PROC_LOCK_OWNED(p);

	if (entryp->aio_workq_link.tqe_prev == NULL) {
		panic("Trying to remove an entry from a work queue, but it is not on a queue");
	}

	TAILQ_REMOVE(&wq_aio->wa_aioq_entries, entryp, aio_workq_link);
	entryp->aio_workq_link.tqe_prev = NULL; /* Not on a workq */

	return true;
}