Home | History | Annotate | Line # | Download | only in kern
sys_futex.c revision 1.11.6.1
      1 /*	$NetBSD: sys_futex.c,v 1.11.6.1 2021/08/01 22:42:38 thorpej Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 2018, 2019, 2020 The NetBSD Foundation, Inc.
      5  * All rights reserved.
      6  *
      7  * This code is derived from software contributed to The NetBSD Foundation
      8  * by Taylor R. Campbell and Jason R. Thorpe.
      9  *
     10  * Redistribution and use in source and binary forms, with or without
     11  * modification, are permitted provided that the following conditions
     12  * are met:
     13  * 1. Redistributions of source code must retain the above copyright
     14  *    notice, this list of conditions and the following disclaimer.
     15  * 2. Redistributions in binary form must reproduce the above copyright
     16  *    notice, this list of conditions and the following disclaimer in the
     17  *    documentation and/or other materials provided with the distribution.
     18  *
     19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     29  * POSSIBILITY OF SUCH DAMAGE.
     30  */
     31 
     32 #include <sys/cdefs.h>
     33 __KERNEL_RCSID(0, "$NetBSD: sys_futex.c,v 1.11.6.1 2021/08/01 22:42:38 thorpej Exp $");
     34 
     35 /*
     36  * Futexes
     37  *
     38  *	The futex system call coordinates notifying threads waiting for
     39  *	changes on a 32-bit word of memory.  The word can be managed by
     40  *	CPU atomic operations in userland, without system calls, as long
     41  *	as there is no contention.
     42  *
     43  *	The simplest use case demonstrating the utility is:
     44  *
     45  *		// 32-bit word of memory shared among threads or
     46  *		// processes in userland.  lock & 1 means owned;
     47  *		// lock & 2 means there are waiters waiting.
     48  *		volatile int lock = 0;
     49  *
     50  *		int v;
     51  *
     52  *		// Acquire a lock.
     53  *		do {
     54  *			v = lock;
     55  *			if (v & 1) {
     56  *				// Lock is held.  Set a bit to say that
     57  *				// there are waiters, and wait for lock
     58  *				// to change to anything other than v;
     59  *				// then retry.
     60  *				if (atomic_cas_uint(&lock, v, v | 2) != v)
     61  *					continue;
     62  *				futex(FUTEX_WAIT, &lock, v | 2, NULL, NULL, 0);
     63  *				continue;
     64  *			}
     65  *		} while (atomic_cas_uint(&lock, v, v & ~1) != v);
     66  *		membar_enter();
     67  *
     68  *		...
     69  *
     70  *		// Release the lock.  Optimistically assume there are
     71  *		// no waiters first until demonstrated otherwise.
     72  *		membar_exit();
     73  *		if (atomic_cas_uint(&lock, 1, 0) != 1) {
     74  *			// There may be waiters.
     75  *			v = atomic_swap_uint(&lock, 0);
     76  *			// If there are still waiters, wake one.
     77  *			if (v & 2)
     78  *				futex(FUTEX_WAKE, &lock, 1, NULL, NULL, 0);
     79  *		}
     80  *
     81  *	The goal is to avoid the futex system call unless there is
     82  *	contention; then if there is contention, to guarantee no missed
     83  *	wakeups.
     84  *
     85  *	For a simple implementation, futex(FUTEX_WAIT) could queue
     86  *	itself to be woken, double-check the lock word, and then sleep;
     87  *	spurious wakeups are generally a fact of life, so any
     88  *	FUTEX_WAKE could just wake every FUTEX_WAIT in the system.
     89  *
     90  *	If this were all there is to it, we could then increase
     91  *	parallelism by refining the approximation: partition the
     92  *	waiters into buckets by hashing the lock addresses to reduce
     93  *	the incidence of spurious wakeups.  But this is not all.
     94  *
     95  *	The futex(FUTEX_CMP_REQUEUE, &lock, n, &lock2, m, val)
     96  *	operation not only wakes n waiters on lock if lock == val, but
     97  *	also _transfers_ m additional waiters to lock2.  Unless wakeups
     98  *	on lock2 also trigger wakeups on lock, we cannot move waiters
     99  *	to lock2 if they merely share the same hash as waiters on lock.
    100  *	Thus, we can't approximately distribute waiters into queues by
    101  *	a hash function; we must distinguish futex queues exactly by
    102  *	lock address.
    103  *
    104  *	For now, we use a global red/black tree to index futexes.  This
    105  *	should be replaced by a lockless radix tree with a thread to
    106  *	free entries no longer in use once all lookups on all CPUs have
    107  *	completed.
    108  *
    109  *	Specifically, we maintain two maps:
    110  *
    111  *	futex_tab.va[vmspace, va] for private futexes
    112  *	futex_tab.oa[uvm_voaddr] for shared futexes
    113  *
    114  *	This implementation does not support priority inheritance.
    115  */
    116 
    117 #include <sys/param.h>
    118 #include <sys/types.h>
    119 #include <sys/atomic.h>
    120 #include <sys/condvar.h>
    121 #include <sys/futex.h>
    122 #include <sys/mutex.h>
    123 #include <sys/rbtree.h>
    124 #include <sys/queue.h>
    125 
    126 #include <sys/syscall.h>
    127 #include <sys/syscallargs.h>
    128 #include <sys/syscallvar.h>
    129 
    130 #include <uvm/uvm_extern.h>
    131 
    132 /*
    133  * Lock order:
    134  *
    135  *	futex_tab.lock
    136  *	futex::fx_qlock			ordered by kva of struct futex
    137  *	 -> futex_wait::fw_lock		only one at a time
    138  *	futex_wait::fw_lock		only one at a time
    139  *	 -> futex::fx_abortlock		only one at a time
    140  */
    141 
    142 /*
    143  * union futex_key
    144  *
    145  *	A futex is addressed either by a vmspace+va (private) or by
    146  *	a uvm_voaddr (shared).
    147  */
    148 union futex_key {
    149 	struct {
    150 		struct vmspace			*vmspace;
    151 		vaddr_t				va;
    152 	}			fk_private;
    153 	struct uvm_voaddr	fk_shared;
    154 };
    155 
    156 /*
    157  * struct futex
    158  *
    159  *	Kernel state for a futex located at a particular address in a
    160  *	particular virtual address space.
    161  *
    162  *	N.B. fx_refcnt is an unsigned long because we need to be able
    163  *	to operate on it atomically on all systems while at the same
    164  *	time rendering practically impossible the chance of it reaching
    165  *	its max value.  In practice, we're limited by the number of LWPs
    166  *	that can be present on the system at any given time, and the
    167  *	assumption is that limit will be good enough on a 32-bit platform.
    168  *	See futex_wake() for why overflow needs to be avoided.
    169  */
    170 struct futex {
    171 	union futex_key		fx_key;
    172 	unsigned long		fx_refcnt;
    173 	bool			fx_shared;
    174 	bool			fx_on_tree;
    175 	struct rb_node		fx_node;
    176 
    177 	kmutex_t			fx_qlock;
    178 	TAILQ_HEAD(, futex_wait)	fx_queue;
    179 
    180 	kmutex_t			fx_abortlock;
    181 	LIST_HEAD(, futex_wait)		fx_abortlist;
    182 	kcondvar_t			fx_abortcv;
    183 };
    184 
    185 /*
    186  * struct futex_wait
    187  *
    188  *	State for a thread to wait on a futex.  Threads wait on fw_cv
    189  *	for fw_bitset to be set to zero.  The thread may transition to
    190  *	a different futex queue at any time under the futex's lock.
    191  */
    192 struct futex_wait {
    193 	kmutex_t		fw_lock;
    194 	kcondvar_t		fw_cv;
    195 	struct futex		*fw_futex;
    196 	TAILQ_ENTRY(futex_wait)	fw_entry;	/* queue lock */
    197 	LIST_ENTRY(futex_wait)	fw_abort;	/* queue abortlock */
    198 	int			fw_bitset;
    199 	bool			fw_aborting;	/* fw_lock */
    200 };
    201 
    202 /*
    203  * futex_tab
    204  *
    205  *	Global trees of futexes by vmspace/va and VM object address.
    206  *
    207  *	XXX This obviously doesn't scale in parallel.  We could use a
    208  *	pserialize-safe data structure, but there may be a high cost to
    209  *	frequent deletion since we don't cache futexes after we're done
    210  *	with them.  We could use hashed locks.  But for now, just make
    211  *	sure userland can't DoS the serial performance, by using a
    212  *	balanced binary tree for lookup.
    213  *
    214  *	XXX We could use a per-process tree for the table indexed by
    215  *	virtual address to reduce contention between processes.
    216  */
    217 static struct {
    218 	kmutex_t	lock;
    219 	struct rb_tree	va;
    220 	struct rb_tree	oa;
    221 } futex_tab __cacheline_aligned;
    222 
    223 static int
    224 compare_futex_key(void *cookie, const void *n, const void *k)
    225 {
    226 	const struct futex *fa = n;
    227 	const union futex_key *fka = &fa->fx_key;
    228 	const union futex_key *fkb = k;
    229 
    230 	if ((uintptr_t)fka->fk_private.vmspace <
    231 	    (uintptr_t)fkb->fk_private.vmspace)
    232 		return -1;
    233 	if ((uintptr_t)fka->fk_private.vmspace >
    234 	    (uintptr_t)fkb->fk_private.vmspace)
    235 		return +1;
    236 	if (fka->fk_private.va < fkb->fk_private.va)
    237 		return -1;
    238 	if (fka->fk_private.va > fkb->fk_private.va)
    239 		return -1;
    240 	return 0;
    241 }
    242 
    243 static int
    244 compare_futex(void *cookie, const void *na, const void *nb)
    245 {
    246 	const struct futex *fa = na;
    247 	const struct futex *fb = nb;
    248 
    249 	return compare_futex_key(cookie, fa, &fb->fx_key);
    250 }
    251 
    252 static const rb_tree_ops_t futex_rb_ops = {
    253 	.rbto_compare_nodes = compare_futex,
    254 	.rbto_compare_key = compare_futex_key,
    255 	.rbto_node_offset = offsetof(struct futex, fx_node),
    256 };
    257 
    258 static int
    259 compare_futex_shared_key(void *cookie, const void *n, const void *k)
    260 {
    261 	const struct futex *fa = n;
    262 	const union futex_key *fka = &fa->fx_key;
    263 	const union futex_key *fkb = k;
    264 
    265 	return uvm_voaddr_compare(&fka->fk_shared, &fkb->fk_shared);
    266 }
    267 
    268 static int
    269 compare_futex_shared(void *cookie, const void *na, const void *nb)
    270 {
    271 	const struct futex *fa = na;
    272 	const struct futex *fb = nb;
    273 
    274 	return compare_futex_shared_key(cookie, fa, &fb->fx_key);
    275 }
    276 
    277 static const rb_tree_ops_t futex_shared_rb_ops = {
    278 	.rbto_compare_nodes = compare_futex_shared,
    279 	.rbto_compare_key = compare_futex_shared_key,
    280 	.rbto_node_offset = offsetof(struct futex, fx_node),
    281 };
    282 
    283 static void	futex_wait_dequeue(struct futex_wait *, struct futex *);
    284 
    285 /*
    286  * futex_load(uaddr, kaddr)
    287  *
    288  *	Perform a single atomic load to read *uaddr, and return the
    289  *	result in *kaddr.  Return 0 on success, EFAULT if uaddr is not
    290  *	mapped.
    291  */
    292 static inline int
    293 futex_load(int *uaddr, int *kaddr)
    294 {
    295 	return ufetch_int((u_int *)uaddr, (u_int *)kaddr);
    296 }
    297 
    298 /*
    299  * futex_test(uaddr, expected)
    300  *
    301  *	True if *uaddr == expected.  False if *uaddr != expected, or if
    302  *	uaddr is not mapped.
    303  */
    304 static bool
    305 futex_test(int *uaddr, int expected)
    306 {
    307 	int val;
    308 	int error;
    309 
    310 	error = futex_load(uaddr, &val);
    311 	if (error)
    312 		return false;
    313 	return val == expected;
    314 }
    315 
    316 /*
    317  * futex_sys_init()
    318  *
    319  *	Initialize the futex subsystem.
    320  */
    321 void
    322 futex_sys_init(void)
    323 {
    324 
    325 	mutex_init(&futex_tab.lock, MUTEX_DEFAULT, IPL_NONE);
    326 	rb_tree_init(&futex_tab.va, &futex_rb_ops);
    327 	rb_tree_init(&futex_tab.oa, &futex_shared_rb_ops);
    328 }
    329 
    330 /*
    331  * futex_sys_fini()
    332  *
    333  *	Finalize the futex subsystem.
    334  */
    335 void
    336 futex_sys_fini(void)
    337 {
    338 
    339 	KASSERT(RB_TREE_MIN(&futex_tab.oa) == NULL);
    340 	KASSERT(RB_TREE_MIN(&futex_tab.va) == NULL);
    341 	mutex_destroy(&futex_tab.lock);
    342 }
    343 
    344 /*
    345  * futex_queue_init(f)
    346  *
    347  *	Initialize the futex queue.  Caller must call futex_queue_fini
    348  *	when done.
    349  *
    350  *	Never sleeps.
    351  */
    352 static void
    353 futex_queue_init(struct futex *f)
    354 {
    355 
    356 	mutex_init(&f->fx_qlock, MUTEX_DEFAULT, IPL_NONE);
    357 	mutex_init(&f->fx_abortlock, MUTEX_DEFAULT, IPL_NONE);
    358 	cv_init(&f->fx_abortcv, "fqabort");
    359 	LIST_INIT(&f->fx_abortlist);
    360 	TAILQ_INIT(&f->fx_queue);
    361 }
    362 
    363 /*
    364  * futex_queue_drain(f)
    365  *
    366  *	Wait for any aborting waiters in f; then empty the queue of
    367  *	any stragglers and wake them.  Caller must guarantee no new
    368  *	references to f.
    369  *
    370  *	May sleep.
    371  */
    372 static void
    373 futex_queue_drain(struct futex *f)
    374 {
    375 	struct futex_wait *fw, *fw_next;
    376 
    377 	mutex_enter(&f->fx_abortlock);
    378 	while (!LIST_EMPTY(&f->fx_abortlist))
    379 		cv_wait(&f->fx_abortcv, &f->fx_abortlock);
    380 	mutex_exit(&f->fx_abortlock);
    381 
    382 	mutex_enter(&f->fx_qlock);
    383 	TAILQ_FOREACH_SAFE(fw, &f->fx_queue, fw_entry, fw_next) {
    384 		mutex_enter(&fw->fw_lock);
    385 		futex_wait_dequeue(fw, f);
    386 		cv_broadcast(&fw->fw_cv);
    387 		mutex_exit(&fw->fw_lock);
    388 	}
    389 	mutex_exit(&f->fx_qlock);
    390 }
    391 
    392 /*
    393  * futex_queue_fini(fq)
    394  *
    395  *	Finalize the futex queue initialized by futex_queue_init.  Queue
    396  *	must be empty.  Caller must not use f again until a subsequent
    397  *	futex_queue_init.
    398  */
    399 static void
    400 futex_queue_fini(struct futex *f)
    401 {
    402 
    403 	KASSERT(TAILQ_EMPTY(&f->fx_queue));
    404 	KASSERT(LIST_EMPTY(&f->fx_abortlist));
    405 	mutex_destroy(&f->fx_qlock);
    406 	mutex_destroy(&f->fx_abortlock);
    407 	cv_destroy(&f->fx_abortcv);
    408 }
    409 
    410 /*
    411  * futex_key_init(key, vm, va, shared)
    412  *
    413  *	Initialize a futex key for lookup, etc.
    414  */
    415 static int
    416 futex_key_init(union futex_key *fk, struct vmspace *vm, vaddr_t va, bool shared)
    417 {
    418 	int error = 0;
    419 
    420 	if (__predict_false(shared)) {
    421 		if (!uvm_voaddr_acquire(&vm->vm_map, va, &fk->fk_shared))
    422 			error = EFAULT;
    423 	} else {
    424 		fk->fk_private.vmspace = vm;
    425 		fk->fk_private.va = va;
    426 	}
    427 
    428 	return error;
    429 }
    430 
    431 /*
    432  * futex_key_fini(key, shared)
    433  *
    434  *	Release a futex key.
    435  */
    436 static void
    437 futex_key_fini(union futex_key *fk, bool shared)
    438 {
    439 	if (__predict_false(shared))
    440 		uvm_voaddr_release(&fk->fk_shared);
    441 	memset(fk, 0, sizeof(*fk));
    442 }
    443 
    444 /*
    445  * futex_create(fk, shared)
    446  *
    447  *	Create a futex.  Initial reference count is 1, representing the
    448  *	caller.  Returns NULL on failure.  Always takes ownership of the
    449  *	key, either transferring it to the newly-created futex, or releasing
    450  *	the key if creation fails.
    451  *
    452  *	Never sleeps for memory, but may sleep to acquire a lock.
    453  */
    454 static struct futex *
    455 futex_create(union futex_key *fk, bool shared)
    456 {
    457 	struct futex *f;
    458 
    459 	f = kmem_alloc(sizeof(*f), KM_NOSLEEP);
    460 	if (f == NULL) {
    461 		futex_key_fini(fk, shared);
    462 		return NULL;
    463 	}
    464 	f->fx_key = *fk;
    465 	f->fx_refcnt = 1;
    466 	f->fx_shared = shared;
    467 	f->fx_on_tree = false;
    468 	futex_queue_init(f);
    469 
    470 	return f;
    471 }
    472 
    473 /*
    474  * futex_destroy(f)
    475  *
    476  *	Destroy a futex created with futex_create.  Reference count
    477  *	must be zero.
    478  *
    479  *	May sleep.
    480  */
    481 static void
    482 futex_destroy(struct futex *f)
    483 {
    484 
    485 	ASSERT_SLEEPABLE();
    486 
    487 	KASSERT(atomic_load_relaxed(&f->fx_refcnt) == 0);
    488 	KASSERT(!f->fx_on_tree);
    489 
    490 	/* Drain and destroy the private queue.  */
    491 	futex_queue_drain(f);
    492 	futex_queue_fini(f);
    493 
    494 	futex_key_fini(&f->fx_key, f->fx_shared);
    495 
    496 	kmem_free(f, sizeof(*f));
    497 }
    498 
    499 /*
    500  * futex_hold(f)
    501  *
    502  *	Attempt to acquire a reference to f.  Return 0 on success,
    503  *	ENFILE on too many references.
    504  *
    505  *	Never sleeps.
    506  */
    507 static int
    508 futex_hold(struct futex *f)
    509 {
    510 	unsigned long refcnt;
    511 
    512 	do {
    513 		refcnt = atomic_load_relaxed(&f->fx_refcnt);
    514 		if (refcnt == ULONG_MAX)
    515 			return ENFILE;
    516 	} while (atomic_cas_ulong(&f->fx_refcnt, refcnt, refcnt + 1) != refcnt);
    517 
    518 	return 0;
    519 }
    520 
    521 /*
    522  * futex_rele(f)
    523  *
    524  *	Release a reference to f acquired with futex_create or
    525  *	futex_hold.
    526  *
    527  *	May sleep to free f.
    528  */
    529 static void
    530 futex_rele(struct futex *f)
    531 {
    532 	unsigned long refcnt;
    533 
    534 	ASSERT_SLEEPABLE();
    535 
    536 	do {
    537 		refcnt = atomic_load_relaxed(&f->fx_refcnt);
    538 		if (refcnt == 1)
    539 			goto trylast;
    540 	} while (atomic_cas_ulong(&f->fx_refcnt, refcnt, refcnt - 1) != refcnt);
    541 	return;
    542 
    543 trylast:
    544 	mutex_enter(&futex_tab.lock);
    545 	if (atomic_dec_ulong_nv(&f->fx_refcnt) == 0) {
    546 		if (f->fx_on_tree) {
    547 			if (__predict_false(f->fx_shared))
    548 				rb_tree_remove_node(&futex_tab.oa, f);
    549 			else
    550 				rb_tree_remove_node(&futex_tab.va, f);
    551 			f->fx_on_tree = false;
    552 		}
    553 	} else {
    554 		/* References remain -- don't destroy it.  */
    555 		f = NULL;
    556 	}
    557 	mutex_exit(&futex_tab.lock);
    558 	if (f != NULL)
    559 		futex_destroy(f);
    560 }
    561 
    562 /*
    563  * futex_rele_not_last(f)
    564  *
    565  *	Release a reference to f acquired with futex_create or
    566  *	futex_hold.
    567  *
    568  *	This version asserts that we are not dropping the last
    569  *	reference to f.
    570  */
    571 static void
    572 futex_rele_not_last(struct futex *f)
    573 {
    574 	unsigned long refcnt;
    575 
    576 	do {
    577 		refcnt = atomic_load_relaxed(&f->fx_refcnt);
    578 		KASSERT(refcnt > 1);
    579 	} while (atomic_cas_ulong(&f->fx_refcnt, refcnt, refcnt - 1) != refcnt);
    580 }
    581 
    582 /*
    583  * futex_lookup_by_key(key, shared, &f)
    584  *
    585  *	Try to find an existing futex va reference in the specified key
    586  *	On success, return 0, set f to found futex or to NULL if not found,
    587  *	and increment f's reference count if found.
    588  *
    589  *	Return ENFILE if reference count too high.
    590  *
    591  *	Internal lookup routine shared by futex_lookup() and
    592  *	futex_lookup_create().
    593  */
    594 static int
    595 futex_lookup_by_key(union futex_key *fk, bool shared, struct futex **fp)
    596 {
    597 	struct futex *f;
    598 	int error = 0;
    599 
    600 	mutex_enter(&futex_tab.lock);
    601 	if (__predict_false(shared)) {
    602 		f = rb_tree_find_node(&futex_tab.oa, fk);
    603 	} else {
    604 		f = rb_tree_find_node(&futex_tab.va, fk);
    605 	}
    606 	if (f) {
    607 		error = futex_hold(f);
    608 		if (error)
    609 			f = NULL;
    610 	}
    611  	*fp = f;
    612 	mutex_exit(&futex_tab.lock);
    613 
    614 	return error;
    615 }
    616 
    617 /*
    618  * futex_insert(f, fp)
    619  *
    620  *	Try to insert the futex f into the tree by va.  If there
    621  *	already is a futex for its va, acquire a reference to it, and
    622  *	store it in *fp; otherwise store f in *fp.
    623  *
    624  *	Return 0 on success, ENFILE if there already is a futex but its
    625  *	reference count is too high.
    626  */
    627 static int
    628 futex_insert(struct futex *f, struct futex **fp)
    629 {
    630 	struct futex *f0;
    631 	int error;
    632 
    633 	KASSERT(atomic_load_relaxed(&f->fx_refcnt) != 0);
    634 	KASSERT(!f->fx_on_tree);
    635 
    636 	mutex_enter(&futex_tab.lock);
    637 	if (__predict_false(f->fx_shared))
    638 		f0 = rb_tree_insert_node(&futex_tab.oa, f);
    639 	else
    640 		f0 = rb_tree_insert_node(&futex_tab.va, f);
    641 	if (f0 == f) {
    642 		f->fx_on_tree = true;
    643 		error = 0;
    644 	} else {
    645 		KASSERT(atomic_load_relaxed(&f0->fx_refcnt) != 0);
    646 		KASSERT(f0->fx_on_tree);
    647 		error = futex_hold(f0);
    648 		if (error)
    649 			goto out;
    650 	}
    651 	*fp = f0;
    652 out:	mutex_exit(&futex_tab.lock);
    653 
    654 	return error;
    655 }
    656 
    657 /*
    658  * futex_lookup(uaddr, shared, &f)
    659  *
    660  *	Find a futex at the userland pointer uaddr in the current
    661  *	process's VM space.  On success, return the futex in f and
    662  *	increment its reference count.
    663  *
    664  *	Caller must call futex_rele when done.
    665  */
    666 static int
    667 futex_lookup(int *uaddr, bool shared, struct futex **fp)
    668 {
    669 	union futex_key fk;
    670 	struct vmspace *vm = curproc->p_vmspace;
    671 	vaddr_t va = (vaddr_t)uaddr;
    672 	int error;
    673 
    674 	/*
    675 	 * Reject unaligned user pointers so we don't cross page
    676 	 * boundaries and so atomics will work.
    677 	 */
    678 	if ((va & 3) != 0)
    679 		return EINVAL;
    680 
    681 	/* Look it up. */
    682 	error = futex_key_init(&fk, vm, va, shared);
    683 	if (error)
    684 		return error;
    685 
    686 	error = futex_lookup_by_key(&fk, shared, fp);
    687 	futex_key_fini(&fk, shared);
    688 	if (error)
    689 		return error;
    690 
    691 	KASSERT(*fp == NULL || (*fp)->fx_shared == shared);
    692 	KASSERT(*fp == NULL || atomic_load_relaxed(&(*fp)->fx_refcnt) != 0);
    693 
    694 	/*
    695 	 * Success!  (Caller must still check whether we found
    696 	 * anything, but nothing went _wrong_ like trying to use
    697 	 * unmapped memory.)
    698 	 */
    699 	KASSERT(error == 0);
    700 
    701 	return error;
    702 }
    703 
    704 /*
    705  * futex_lookup_create(uaddr, shared, &f)
    706  *
    707  *	Find or create a futex at the userland pointer uaddr in the
    708  *	current process's VM space.  On success, return the futex in f
    709  *	and increment its reference count.
    710  *
    711  *	Caller must call futex_rele when done.
    712  */
    713 static int
    714 futex_lookup_create(int *uaddr, bool shared, struct futex **fp)
    715 {
    716 	union futex_key fk;
    717 	struct vmspace *vm = curproc->p_vmspace;
    718 	struct futex *f = NULL;
    719 	vaddr_t va = (vaddr_t)uaddr;
    720 	int error;
    721 
    722 	/*
    723 	 * Reject unaligned user pointers so we don't cross page
    724 	 * boundaries and so atomics will work.
    725 	 */
    726 	if ((va & 3) != 0)
    727 		return EINVAL;
    728 
    729 	error = futex_key_init(&fk, vm, va, shared);
    730 	if (error)
    731 		return error;
    732 
    733 	/*
    734 	 * Optimistically assume there already is one, and try to find
    735 	 * it.
    736 	 */
    737 	error = futex_lookup_by_key(&fk, shared, fp);
    738 	if (error || *fp != NULL) {
    739 		/*
    740 		 * We either found one, or there was an error.
    741 		 * In either case, we are done with the key.
    742 		 */
    743 		futex_key_fini(&fk, shared);
    744 		goto out;
    745 	}
    746 
    747 	/*
    748 	 * Create a futex recoard.  This tranfers ownership of the key
    749 	 * in all cases.
    750 	 */
    751 	f = futex_create(&fk, shared);
    752 	if (f == NULL) {
    753 		error = ENOMEM;
    754 		goto out;
    755 	}
    756 
    757 	/*
    758 	 * Insert our new futex, or use existing if someone else beat
    759 	 * us to it.
    760 	 */
    761 	error = futex_insert(f, fp);
    762 	if (error)
    763 		goto out;
    764 	if (*fp == f)
    765 		f = NULL;	/* don't release on exit */
    766 
    767 	/* Success!  */
    768 	KASSERT(error == 0);
    769 
    770 out:	if (f != NULL)
    771 		futex_rele(f);
    772 	KASSERT(error || *fp != NULL);
    773 	KASSERT(error || atomic_load_relaxed(&(*fp)->fx_refcnt) != 0);
    774 	return error;
    775 }
    776 
    777 /*
    778  * futex_wait_init(fw, bitset)
    779  *
    780  *	Initialize a record for a thread to wait on a futex matching
    781  *	the specified bit set.  Should be passed to futex_wait_enqueue
    782  *	before futex_wait, and should be passed to futex_wait_fini when
    783  *	done.
    784  */
    785 static void
    786 futex_wait_init(struct futex_wait *fw, int bitset)
    787 {
    788 
    789 	KASSERT(bitset);
    790 
    791 	mutex_init(&fw->fw_lock, MUTEX_DEFAULT, IPL_NONE);
    792 	cv_init(&fw->fw_cv, "futex");
    793 	fw->fw_futex = NULL;
    794 	fw->fw_bitset = bitset;
    795 	fw->fw_aborting = false;
    796 }
    797 
    798 /*
    799  * futex_wait_fini(fw)
    800  *
    801  *	Finalize a record for a futex waiter.  Must not be on any
    802  *	futex's queue.
    803  */
    804 static void
    805 futex_wait_fini(struct futex_wait *fw)
    806 {
    807 
    808 	KASSERT(fw->fw_futex == NULL);
    809 
    810 	cv_destroy(&fw->fw_cv);
    811 	mutex_destroy(&fw->fw_lock);
    812 }
    813 
    814 /*
    815  * futex_wait_enqueue(fw, f)
    816  *
    817  *	Put fw on the futex queue.  Must be done before futex_wait.
    818  *	Caller must hold fw's lock and f's lock, and fw must not be on
    819  *	any existing futex's waiter list.
    820  */
    821 static void
    822 futex_wait_enqueue(struct futex_wait *fw, struct futex *f)
    823 {
    824 
    825 	KASSERT(mutex_owned(&f->fx_qlock));
    826 	KASSERT(mutex_owned(&fw->fw_lock));
    827 	KASSERT(fw->fw_futex == NULL);
    828 	KASSERT(!fw->fw_aborting);
    829 
    830 	fw->fw_futex = f;
    831 	TAILQ_INSERT_TAIL(&f->fx_queue, fw, fw_entry);
    832 }
    833 
    834 /*
    835  * futex_wait_dequeue(fw, f)
    836  *
    837  *	Remove fw from the futex queue.  Precludes subsequent
    838  *	futex_wait until a futex_wait_enqueue.  Caller must hold fw's
    839  *	lock and f's lock, and fw must be on f.
    840  */
    841 static void
    842 futex_wait_dequeue(struct futex_wait *fw, struct futex *f)
    843 {
    844 
    845 	KASSERT(mutex_owned(&f->fx_qlock));
    846 	KASSERT(mutex_owned(&fw->fw_lock));
    847 	KASSERT(fw->fw_futex == f);
    848 
    849 	TAILQ_REMOVE(&f->fx_queue, fw, fw_entry);
    850 	fw->fw_futex = NULL;
    851 }
    852 
    853 /*
    854  * futex_wait_abort(fw)
    855  *
    856  *	Caller is no longer waiting for fw.  Remove it from any queue
    857  *	if it was on one.  Caller must hold fw->fw_lock.
    858  */
    859 static void
    860 futex_wait_abort(struct futex_wait *fw)
    861 {
    862 	struct futex *f;
    863 
    864 	KASSERT(mutex_owned(&fw->fw_lock));
    865 
    866 	/*
    867 	 * Grab the futex queue.  It can't go away as long as we hold
    868 	 * fw_lock.  However, we can't take the queue lock because
    869 	 * that's a lock order reversal.
    870 	 */
    871 	f = fw->fw_futex;
    872 
    873 	/* Put us on the abort list so that fq won't go away.  */
    874 	mutex_enter(&f->fx_abortlock);
    875 	LIST_INSERT_HEAD(&f->fx_abortlist, fw, fw_abort);
    876 	mutex_exit(&f->fx_abortlock);
    877 
    878 	/*
    879 	 * Mark fw as aborting so it won't lose wakeups and won't be
    880 	 * transferred to any other queue.
    881 	 */
    882 	fw->fw_aborting = true;
    883 
    884 	/* f is now stable, so we can release fw_lock.  */
    885 	mutex_exit(&fw->fw_lock);
    886 
    887 	/* Now we can remove fw under the queue lock.  */
    888 	mutex_enter(&f->fx_qlock);
    889 	mutex_enter(&fw->fw_lock);
    890 	futex_wait_dequeue(fw, f);
    891 	mutex_exit(&fw->fw_lock);
    892 	mutex_exit(&f->fx_qlock);
    893 
    894 	/*
    895 	 * Finally, remove us from the abort list and notify anyone
    896 	 * waiting for the abort to complete if we were the last to go.
    897 	 */
    898 	mutex_enter(&f->fx_abortlock);
    899 	LIST_REMOVE(fw, fw_abort);
    900 	if (LIST_EMPTY(&f->fx_abortlist))
    901 		cv_broadcast(&f->fx_abortcv);
    902 	mutex_exit(&f->fx_abortlock);
    903 
    904 	/*
    905 	 * Release our reference to the futex now that we are not
    906 	 * waiting for it.
    907 	 */
    908 	futex_rele(f);
    909 
    910 	/*
    911 	 * Reacquire the fw lock as caller expects.  Verify that we're
    912 	 * aborting and no longer associated with a futex.
    913 	 */
    914 	mutex_enter(&fw->fw_lock);
    915 	KASSERT(fw->fw_aborting);
    916 	KASSERT(fw->fw_futex == NULL);
    917 }
    918 
    919 /*
    920  * futex_wait(fw, deadline, clkid)
    921  *
    922  *	fw must be a waiter on a futex's queue.  Wait until deadline on
    923  *	the clock clkid, or forever if deadline is NULL, for a futex
    924  *	wakeup.  Return 0 on explicit wakeup or destruction of futex,
    925  *	ETIMEDOUT on timeout, EINTR/ERESTART on signal.  Either way, fw
    926  *	will no longer be on a futex queue on return.
    927  */
    928 static int
    929 futex_wait(struct futex_wait *fw, const struct timespec *deadline,
    930     clockid_t clkid)
    931 {
    932 	int error = 0;
    933 
    934 	/* Test and wait under the wait lock.  */
    935 	mutex_enter(&fw->fw_lock);
    936 
    937 	for (;;) {
    938 		/* If we're done yet, stop and report success.  */
    939 		if (fw->fw_bitset == 0 || fw->fw_futex == NULL) {
    940 			error = 0;
    941 			break;
    942 		}
    943 
    944 		/* If anything went wrong in the last iteration, stop.  */
    945 		if (error)
    946 			break;
    947 
    948 		/* Not done yet.  Wait.  */
    949 		if (deadline) {
    950 			struct timespec ts;
    951 
    952 			/* Check our watch.  */
    953 			error = clock_gettime1(clkid, &ts);
    954 			if (error)
    955 				break;
    956 
    957 			/* If we're past the deadline, ETIMEDOUT.  */
    958 			if (timespeccmp(deadline, &ts, <=)) {
    959 				error = ETIMEDOUT;
    960 				break;
    961 			}
    962 
    963 			/* Count how much time is left.  */
    964 			timespecsub(deadline, &ts, &ts);
    965 
    966 			/* Wait for that much time, allowing signals.  */
    967 			error = cv_timedwait_sig(&fw->fw_cv, &fw->fw_lock,
    968 			    tstohz(&ts));
    969 		} else {
    970 			/* Wait indefinitely, allowing signals. */
    971 			error = cv_wait_sig(&fw->fw_cv, &fw->fw_lock);
    972 		}
    973 	}
    974 
    975 	/*
    976 	 * If we were woken up, the waker will have removed fw from the
    977 	 * queue.  But if anything went wrong, we must remove fw from
    978 	 * the queue ourselves.  While here, convert EWOULDBLOCK to
    979 	 * ETIMEDOUT.
    980 	 */
    981 	if (error) {
    982 		futex_wait_abort(fw);
    983 		if (error == EWOULDBLOCK)
    984 			error = ETIMEDOUT;
    985 	}
    986 
    987 	mutex_exit(&fw->fw_lock);
    988 
    989 	return error;
    990 }
    991 
    992 /*
    993  * futex_wake(f, nwake, f2, nrequeue, bitset)
    994  *
    995  *	Wake up to nwake waiters on f matching bitset; then, if f2 is
    996  *	provided, move up to nrequeue remaining waiters on f matching
    997  *	bitset to f2.  Return the number of waiters actually woken.
    998  *	Caller must hold the locks of f and f2, if provided.
    999  */
   1000 static unsigned
   1001 futex_wake(struct futex *f, unsigned nwake, struct futex *f2,
   1002     unsigned nrequeue, int bitset)
   1003 {
   1004 	struct futex_wait *fw, *fw_next;
   1005 	unsigned nwoken = 0;
   1006 	int hold_error __diagused;
   1007 
   1008 	KASSERT(mutex_owned(&f->fx_qlock));
   1009 	KASSERT(f2 == NULL || mutex_owned(&f2->fx_qlock));
   1010 
   1011 	/* Wake up to nwake waiters, and count the number woken.  */
   1012 	TAILQ_FOREACH_SAFE(fw, &f->fx_queue, fw_entry, fw_next) {
   1013 		if ((fw->fw_bitset & bitset) == 0)
   1014 			continue;
   1015 		if (nwake > 0) {
   1016 			mutex_enter(&fw->fw_lock);
   1017 			if (__predict_false(fw->fw_aborting)) {
   1018 				mutex_exit(&fw->fw_lock);
   1019 				continue;
   1020 			}
   1021 			futex_wait_dequeue(fw, f);
   1022 			fw->fw_bitset = 0;
   1023 			cv_broadcast(&fw->fw_cv);
   1024 			mutex_exit(&fw->fw_lock);
   1025 			nwake--;
   1026 			nwoken++;
   1027 			/*
   1028 			 * Drop the futex reference on behalf of the
   1029 			 * waiter.  We assert this is not the last
   1030 			 * reference on the futex (our caller should
   1031 			 * also have one).
   1032 			 */
   1033 			futex_rele_not_last(f);
   1034 		} else {
   1035 			break;
   1036 		}
   1037 	}
   1038 
   1039 	if (f2) {
   1040 		/* Move up to nrequeue waiters from f's queue to f2's queue. */
   1041 		TAILQ_FOREACH_SAFE(fw, &f->fx_queue, fw_entry, fw_next) {
   1042 			if ((fw->fw_bitset & bitset) == 0)
   1043 				continue;
   1044 			if (nrequeue > 0) {
   1045 				mutex_enter(&fw->fw_lock);
   1046 				if (__predict_false(fw->fw_aborting)) {
   1047 					mutex_exit(&fw->fw_lock);
   1048 					continue;
   1049 				}
   1050 				futex_wait_dequeue(fw, f);
   1051 				futex_wait_enqueue(fw, f2);
   1052 				mutex_exit(&fw->fw_lock);
   1053 				nrequeue--;
   1054 				/*
   1055 				 * Transfer the reference from f to f2.
   1056 				 * As above, we assert that we are not
   1057 				 * dropping the last reference to f here.
   1058 				 *
   1059 				 * XXX futex_hold() could theoretically
   1060 				 * XXX fail here.
   1061 				 */
   1062 				futex_rele_not_last(f);
   1063 				hold_error = futex_hold(f2);
   1064 				KASSERT(hold_error == 0);
   1065 			} else {
   1066 				break;
   1067 			}
   1068 		}
   1069 	} else {
   1070 		KASSERT(nrequeue == 0);
   1071 	}
   1072 
   1073 	/* Return the number of waiters woken.  */
   1074 	return nwoken;
   1075 }
   1076 
   1077 /*
   1078  * futex_queue_lock(f)
   1079  *
   1080  *	Acquire the queue lock of f.  Pair with futex_queue_unlock.  Do
   1081  *	not use if caller needs to acquire two locks; use
   1082  *	futex_queue_lock2 instead.
   1083  */
   1084 static void
   1085 futex_queue_lock(struct futex *f)
   1086 {
   1087 	mutex_enter(&f->fx_qlock);
   1088 }
   1089 
   1090 /*
   1091  * futex_queue_unlock(f)
   1092  *
   1093  *	Release the queue lock of f.
   1094  */
   1095 static void
   1096 futex_queue_unlock(struct futex *f)
   1097 {
   1098 	mutex_exit(&f->fx_qlock);
   1099 }
   1100 
   1101 /*
   1102  * futex_queue_lock2(f, f2)
   1103  *
   1104  *	Acquire the queue locks of both f and f2, which may be null, or
   1105  *	which may have the same underlying queue.  If they are
   1106  *	distinct, an arbitrary total order is chosen on the locks.
   1107  *
   1108  *	Callers should only ever acquire multiple queue locks
   1109  *	simultaneously using futex_queue_lock2.
   1110  */
   1111 static void
   1112 futex_queue_lock2(struct futex *f, struct futex *f2)
   1113 {
   1114 
   1115 	/*
   1116 	 * If both are null, do nothing; if one is null and the other
   1117 	 * is not, lock the other and be done with it.
   1118 	 */
   1119 	if (f == NULL && f2 == NULL) {
   1120 		return;
   1121 	} else if (f == NULL) {
   1122 		mutex_enter(&f2->fx_qlock);
   1123 		return;
   1124 	} else if (f2 == NULL) {
   1125 		mutex_enter(&f->fx_qlock);
   1126 		return;
   1127 	}
   1128 
   1129 	/* If both futexes are the same, acquire only one. */
   1130 	if (f == f2) {
   1131 		mutex_enter(&f->fx_qlock);
   1132 		return;
   1133 	}
   1134 
   1135 	/* Otherwise, use the ordering on the kva of the futex pointer.  */
   1136 	if ((uintptr_t)f < (uintptr_t)f2) {
   1137 		mutex_enter(&f->fx_qlock);
   1138 		mutex_enter(&f2->fx_qlock);
   1139 	} else {
   1140 		mutex_enter(&f2->fx_qlock);
   1141 		mutex_enter(&f->fx_qlock);
   1142 	}
   1143 }
   1144 
   1145 /*
   1146  * futex_queue_unlock2(f, f2)
   1147  *
   1148  *	Release the queue locks of both f and f2, which may be null, or
   1149  *	which may have the same underlying queue.
   1150  */
   1151 static void
   1152 futex_queue_unlock2(struct futex *f, struct futex *f2)
   1153 {
   1154 
   1155 	/*
   1156 	 * If both are null, do nothing; if one is null and the other
   1157 	 * is not, unlock the other and be done with it.
   1158 	 */
   1159 	if (f == NULL && f2 == NULL) {
   1160 		return;
   1161 	} else if (f == NULL) {
   1162 		mutex_exit(&f2->fx_qlock);
   1163 		return;
   1164 	} else if (f2 == NULL) {
   1165 		mutex_exit(&f->fx_qlock);
   1166 		return;
   1167 	}
   1168 
   1169 	/* If both futexes are the same, release only one. */
   1170 	if (f == f2) {
   1171 		mutex_exit(&f->fx_qlock);
   1172 		return;
   1173 	}
   1174 
   1175 	/* Otherwise, use the ordering on the kva of the futex pointer.  */
   1176 	if ((uintptr_t)f < (uintptr_t)f2) {
   1177 		mutex_exit(&f2->fx_qlock);
   1178 		mutex_exit(&f->fx_qlock);
   1179 	} else {
   1180 		mutex_exit(&f->fx_qlock);
   1181 		mutex_exit(&f2->fx_qlock);
   1182 	}
   1183 }
   1184 
   1185 /*
   1186  * futex_func_wait(uaddr, val, val3, timeout, clkid, clkflags, retval)
   1187  *
   1188  *	Implement futex(FUTEX_WAIT).
   1189  */
   1190 static int
   1191 futex_func_wait(bool shared, int *uaddr, int val, int val3,
   1192     const struct timespec *timeout, clockid_t clkid, int clkflags,
   1193     register_t *retval)
   1194 {
   1195 	struct futex *f;
   1196 	struct futex_wait wait, *fw = &wait;
   1197 	struct timespec ts;
   1198 	const struct timespec *deadline;
   1199 	int error;
   1200 
   1201 	/*
   1202 	 * If there's nothing to wait for, and nobody will ever wake
   1203 	 * us, then don't set anything up to wait -- just stop here.
   1204 	 */
   1205 	if (val3 == 0)
   1206 		return EINVAL;
   1207 
   1208 	/* Optimistically test before anything else.  */
   1209 	if (!futex_test(uaddr, val))
   1210 		return EAGAIN;
   1211 
   1212 	/* Determine a deadline on the specified clock.  */
   1213 	if (timeout == NULL || (clkflags & TIMER_ABSTIME) == TIMER_ABSTIME) {
   1214 		deadline = timeout;
   1215 	} else {
   1216 		error = clock_gettime1(clkid, &ts);
   1217 		if (error)
   1218 			return error;
   1219 		timespecadd(&ts, timeout, &ts);
   1220 		deadline = &ts;
   1221 	}
   1222 
   1223 	/* Get the futex, creating it if necessary.  */
   1224 	error = futex_lookup_create(uaddr, shared, &f);
   1225 	if (error)
   1226 		return error;
   1227 	KASSERT(f);
   1228 
   1229 	/* Get ready to wait.  */
   1230 	futex_wait_init(fw, val3);
   1231 
   1232 	/*
   1233 	 * Under the queue lock, check the value again: if it has
   1234 	 * already changed, EAGAIN; otherwise enqueue the waiter.
   1235 	 * Since FUTEX_WAKE will use the same lock and be done after
   1236 	 * modifying the value, the order in which we check and enqueue
   1237 	 * is immaterial.
   1238 	 */
   1239 	futex_queue_lock(f);
   1240 	if (!futex_test(uaddr, val)) {
   1241 		futex_queue_unlock(f);
   1242 		error = EAGAIN;
   1243 		goto out;
   1244 	}
   1245 	mutex_enter(&fw->fw_lock);
   1246 	futex_wait_enqueue(fw, f);
   1247 	mutex_exit(&fw->fw_lock);
   1248 	futex_queue_unlock(f);
   1249 
   1250 	/*
   1251 	 * We cannot drop our reference to the futex here, because
   1252 	 * we might be enqueued on a different one when we are awakened.
   1253 	 * The references will be managed on our behalf in the requeue
   1254 	 * and wake cases.
   1255 	 */
   1256 	f = NULL;
   1257 
   1258 	/* Wait. */
   1259 	error = futex_wait(fw, deadline, clkid);
   1260 	if (error)
   1261 		goto out;
   1262 
   1263 	/* Return 0 on success, error on failure. */
   1264 	*retval = 0;
   1265 
   1266 out:	if (f != NULL)
   1267 		futex_rele(f);
   1268 	futex_wait_fini(fw);
   1269 	return error;
   1270 }
   1271 
   1272 /*
   1273  * futex_func_wake(uaddr, val, val3, retval)
   1274  *
   1275  *	Implement futex(FUTEX_WAKE) and futex(FUTEX_WAKE_BITSET).
   1276  */
   1277 static int
   1278 futex_func_wake(bool shared, int *uaddr, int val, int val3, register_t *retval)
   1279 {
   1280 	struct futex *f;
   1281 	unsigned int nwoken = 0;
   1282 	int error = 0;
   1283 
   1284 	/* Reject negative number of wakeups.  */
   1285 	if (val < 0) {
   1286 		error = EINVAL;
   1287 		goto out;
   1288 	}
   1289 
   1290 	/* Look up the futex, if any.  */
   1291 	error = futex_lookup(uaddr, shared, &f);
   1292 	if (error)
   1293 		goto out;
   1294 
   1295 	/* If there's no futex, there are no waiters to wake.  */
   1296 	if (f == NULL)
   1297 		goto out;
   1298 
   1299 	/*
   1300 	 * Under f's queue lock, wake the waiters and remember the
   1301 	 * number woken.
   1302 	 */
   1303 	futex_queue_lock(f);
   1304 	nwoken = futex_wake(f, val, NULL, 0, val3);
   1305 	futex_queue_unlock(f);
   1306 
   1307 	/* Release the futex.  */
   1308 	futex_rele(f);
   1309 
   1310 out:
   1311 	/* Return the number of waiters woken.  */
   1312 	*retval = nwoken;
   1313 
   1314 	/* Success!  */
   1315 	return error;
   1316 }
   1317 
   1318 /*
   1319  * futex_func_requeue(op, uaddr, val, uaddr2, val2, val3, retval)
   1320  *
   1321  *	Implement futex(FUTEX_REQUEUE) and futex(FUTEX_CMP_REQUEUE).
   1322  */
   1323 static int
   1324 futex_func_requeue(bool shared, int op, int *uaddr, int val, int *uaddr2,
   1325     int val2, int val3, register_t *retval)
   1326 {
   1327 	struct futex *f = NULL, *f2 = NULL;
   1328 	unsigned nwoken = 0;	/* default to zero woken on early return */
   1329 	int error;
   1330 
   1331 	/* Reject negative number of wakeups or requeues. */
   1332 	if (val < 0 || val2 < 0) {
   1333 		error = EINVAL;
   1334 		goto out;
   1335 	}
   1336 
   1337 	/* Look up the source futex, if any. */
   1338 	error = futex_lookup(uaddr, shared, &f);
   1339 	if (error)
   1340 		goto out;
   1341 
   1342 	/* If there is none, nothing to do. */
   1343 	if (f == NULL)
   1344 		goto out;
   1345 
   1346 	/*
   1347 	 * We may need to create the destination futex because it's
   1348 	 * entirely possible it does not currently have any waiters.
   1349 	 */
   1350 	error = futex_lookup_create(uaddr2, shared, &f2);
   1351 	if (error)
   1352 		goto out;
   1353 
   1354 	/*
   1355 	 * Under the futexes' queue locks, check the value; if
   1356 	 * unchanged from val3, wake the waiters.
   1357 	 */
   1358 	futex_queue_lock2(f, f2);
   1359 	if (op == FUTEX_CMP_REQUEUE && !futex_test(uaddr, val3)) {
   1360 		error = EAGAIN;
   1361 	} else {
   1362 		error = 0;
   1363 		nwoken = futex_wake(f, val, f2, val2, FUTEX_BITSET_MATCH_ANY);
   1364 	}
   1365 	futex_queue_unlock2(f, f2);
   1366 
   1367 out:
   1368 	/* Return the number of waiters woken.  */
   1369 	*retval = nwoken;
   1370 
   1371 	/* Release the futexes if we got them.  */
   1372 	if (f2)
   1373 		futex_rele(f2);
   1374 	if (f)
   1375 		futex_rele(f);
   1376 	return error;
   1377 }
   1378 
   1379 /*
   1380  * futex_validate_op_cmp(val3)
   1381  *
   1382  *	Validate an op/cmp argument for FUTEX_WAKE_OP.
   1383  */
   1384 static int
   1385 futex_validate_op_cmp(int val3)
   1386 {
   1387 	int op = __SHIFTOUT(val3, FUTEX_OP_OP_MASK);
   1388 	int cmp = __SHIFTOUT(val3, FUTEX_OP_CMP_MASK);
   1389 
   1390 	if (op & FUTEX_OP_OPARG_SHIFT) {
   1391 		int oparg = __SHIFTOUT(val3, FUTEX_OP_OPARG_MASK);
   1392 		if (oparg < 0)
   1393 			return EINVAL;
   1394 		if (oparg >= 32)
   1395 			return EINVAL;
   1396 		op &= ~FUTEX_OP_OPARG_SHIFT;
   1397 	}
   1398 
   1399 	switch (op) {
   1400 	case FUTEX_OP_SET:
   1401 	case FUTEX_OP_ADD:
   1402 	case FUTEX_OP_OR:
   1403 	case FUTEX_OP_ANDN:
   1404 	case FUTEX_OP_XOR:
   1405 		break;
   1406 	default:
   1407 		return EINVAL;
   1408 	}
   1409 
   1410 	switch (cmp) {
   1411 	case FUTEX_OP_CMP_EQ:
   1412 	case FUTEX_OP_CMP_NE:
   1413 	case FUTEX_OP_CMP_LT:
   1414 	case FUTEX_OP_CMP_LE:
   1415 	case FUTEX_OP_CMP_GT:
   1416 	case FUTEX_OP_CMP_GE:
   1417 		break;
   1418 	default:
   1419 		return EINVAL;
   1420 	}
   1421 
   1422 	return 0;
   1423 }
   1424 
   1425 /*
   1426  * futex_compute_op(oldval, val3)
   1427  *
   1428  *	Apply a FUTEX_WAIT_OP operation to oldval.
   1429  */
   1430 static int
   1431 futex_compute_op(int oldval, int val3)
   1432 {
   1433 	int op = __SHIFTOUT(val3, FUTEX_OP_OP_MASK);
   1434 	int oparg = __SHIFTOUT(val3, FUTEX_OP_OPARG_MASK);
   1435 
   1436 	if (op & FUTEX_OP_OPARG_SHIFT) {
   1437 		KASSERT(oparg >= 0);
   1438 		KASSERT(oparg < 32);
   1439 		oparg = 1u << oparg;
   1440 		op &= ~FUTEX_OP_OPARG_SHIFT;
   1441 	}
   1442 
   1443 	switch (op) {
   1444 	case FUTEX_OP_SET:
   1445 		return oparg;
   1446 
   1447 	case FUTEX_OP_ADD:
   1448 		/*
   1449 		 * Avoid signed arithmetic overflow by doing
   1450 		 * arithmetic unsigned and converting back to signed
   1451 		 * at the end.
   1452 		 */
   1453 		return (int)((unsigned)oldval + (unsigned)oparg);
   1454 
   1455 	case FUTEX_OP_OR:
   1456 		return oldval | oparg;
   1457 
   1458 	case FUTEX_OP_ANDN:
   1459 		return oldval & ~oparg;
   1460 
   1461 	case FUTEX_OP_XOR:
   1462 		return oldval ^ oparg;
   1463 
   1464 	default:
   1465 		panic("invalid futex op");
   1466 	}
   1467 }
   1468 
   1469 /*
   1470  * futex_compute_cmp(oldval, val3)
   1471  *
   1472  *	Apply a FUTEX_WAIT_OP comparison to oldval.
   1473  */
   1474 static bool
   1475 futex_compute_cmp(int oldval, int val3)
   1476 {
   1477 	int cmp = __SHIFTOUT(val3, FUTEX_OP_CMP_MASK);
   1478 	int cmparg = __SHIFTOUT(val3, FUTEX_OP_CMPARG_MASK);
   1479 
   1480 	switch (cmp) {
   1481 	case FUTEX_OP_CMP_EQ:
   1482 		return (oldval == cmparg);
   1483 
   1484 	case FUTEX_OP_CMP_NE:
   1485 		return (oldval != cmparg);
   1486 
   1487 	case FUTEX_OP_CMP_LT:
   1488 		return (oldval < cmparg);
   1489 
   1490 	case FUTEX_OP_CMP_LE:
   1491 		return (oldval <= cmparg);
   1492 
   1493 	case FUTEX_OP_CMP_GT:
   1494 		return (oldval > cmparg);
   1495 
   1496 	case FUTEX_OP_CMP_GE:
   1497 		return (oldval >= cmparg);
   1498 
   1499 	default:
   1500 		panic("invalid futex cmp operation");
   1501 	}
   1502 }
   1503 
   1504 /*
   1505  * futex_func_wake_op(uaddr, val, uaddr2, val2, val3, retval)
   1506  *
   1507  *	Implement futex(FUTEX_WAKE_OP).
   1508  */
   1509 static int
   1510 futex_func_wake_op(bool shared, int *uaddr, int val, int *uaddr2, int val2,
   1511     int val3, register_t *retval)
   1512 {
   1513 	struct futex *f = NULL, *f2 = NULL;
   1514 	int oldval, newval, actual;
   1515 	unsigned nwoken = 0;
   1516 	int error;
   1517 
   1518 	/* Reject negative number of wakeups.  */
   1519 	if (val < 0 || val2 < 0) {
   1520 		error = EINVAL;
   1521 		goto out;
   1522 	}
   1523 
   1524 	/* Reject invalid operations before we start doing things.  */
   1525 	if ((error = futex_validate_op_cmp(val3)) != 0)
   1526 		goto out;
   1527 
   1528 	/* Look up the first futex, if any.  */
   1529 	error = futex_lookup(uaddr, shared, &f);
   1530 	if (error)
   1531 		goto out;
   1532 
   1533 	/* Look up the second futex, if any.  */
   1534 	error = futex_lookup(uaddr2, shared, &f2);
   1535 	if (error)
   1536 		goto out;
   1537 
   1538 	/*
   1539 	 * Under the queue locks:
   1540 	 *
   1541 	 * 1. Read/modify/write: *uaddr2 op= oparg.
   1542 	 * 2. Unconditionally wake uaddr.
   1543 	 * 3. Conditionally wake uaddr2, if it previously matched val2.
   1544 	 */
   1545 	futex_queue_lock2(f, f2);
   1546 	do {
   1547 		error = futex_load(uaddr2, &oldval);
   1548 		if (error)
   1549 			goto out_unlock;
   1550 		newval = futex_compute_op(oldval, val3);
   1551 		error = ucas_int(uaddr2, oldval, newval, &actual);
   1552 		if (error)
   1553 			goto out_unlock;
   1554 	} while (actual != oldval);
   1555 	nwoken = (f ? futex_wake(f, val, NULL, 0, FUTEX_BITSET_MATCH_ANY) : 0);
   1556 	if (f2 && futex_compute_cmp(oldval, val3))
   1557 		nwoken += futex_wake(f2, val2, NULL, 0,
   1558 		    FUTEX_BITSET_MATCH_ANY);
   1559 
   1560 	/* Success! */
   1561 	error = 0;
   1562 out_unlock:
   1563 	futex_queue_unlock2(f, f2);
   1564 
   1565 out:
   1566 	/* Return the number of waiters woken. */
   1567 	*retval = nwoken;
   1568 
   1569 	/* Release the futexes, if we got them. */
   1570 	if (f2)
   1571 		futex_rele(f2);
   1572 	if (f)
   1573 		futex_rele(f);
   1574 	return error;
   1575 }
   1576 
   1577 /*
   1578  * do_futex(uaddr, op, val, timeout, uaddr2, val2, val3)
   1579  *
   1580  *	Implement the futex system call with all the parameters
   1581  *	parsed out.
   1582  */
   1583 int
   1584 do_futex(int *uaddr, int op, int val, const struct timespec *timeout,
   1585     int *uaddr2, int val2, int val3, register_t *retval)
   1586 {
   1587 	const bool shared = (op & FUTEX_PRIVATE_FLAG) ? false : true;
   1588 	const clockid_t clkid = (op & FUTEX_CLOCK_REALTIME) ? CLOCK_REALTIME
   1589 							    : CLOCK_MONOTONIC;
   1590 
   1591 	op &= FUTEX_CMD_MASK;
   1592 
   1593 	switch (op) {
   1594 	case FUTEX_WAIT:
   1595 		return futex_func_wait(shared, uaddr, val,
   1596 		    FUTEX_BITSET_MATCH_ANY, timeout, clkid, TIMER_RELTIME,
   1597 		    retval);
   1598 
   1599 	case FUTEX_WAKE:
   1600 		val3 = FUTEX_BITSET_MATCH_ANY;
   1601 		/* FALLTHROUGH */
   1602 	case FUTEX_WAKE_BITSET:
   1603 		return futex_func_wake(shared, uaddr, val, val3, retval);
   1604 
   1605 	case FUTEX_REQUEUE:
   1606 	case FUTEX_CMP_REQUEUE:
   1607 		return futex_func_requeue(shared, op, uaddr, val, uaddr2,
   1608 		    val2, val3, retval);
   1609 
   1610 	case FUTEX_WAIT_BITSET:
   1611 		return futex_func_wait(shared, uaddr, val, val3, timeout,
   1612 		    clkid, TIMER_ABSTIME, retval);
   1613 
   1614 	case FUTEX_WAKE_OP:
   1615 		return futex_func_wake_op(shared, uaddr, val, uaddr2, val2,
   1616 		    val3, retval);
   1617 
   1618 	case FUTEX_FD:
   1619 	default:
   1620 		return ENOSYS;
   1621 	}
   1622 }
   1623 
   1624 /*
   1625  * sys___futex(l, uap, retval)
   1626  *
   1627  *	__futex(2) system call: generic futex operations.
   1628  */
   1629 int
   1630 sys___futex(struct lwp *l, const struct sys___futex_args *uap,
   1631     register_t *retval)
   1632 {
   1633 	/* {
   1634 		syscallarg(int *) uaddr;
   1635 		syscallarg(int) op;
   1636 		syscallarg(int) val;
   1637 		syscallarg(const struct timespec *) timeout;
   1638 		syscallarg(int *) uaddr2;
   1639 		syscallarg(int) val2;
   1640 		syscallarg(int) val3;
   1641 	} */
   1642 	struct timespec ts, *tsp;
   1643 	int error;
   1644 
   1645 	/*
   1646 	 * Copy in the timeout argument, if specified.
   1647 	 */
   1648 	if (SCARG(uap, timeout)) {
   1649 		error = copyin(SCARG(uap, timeout), &ts, sizeof(ts));
   1650 		if (error)
   1651 			return error;
   1652 		tsp = &ts;
   1653 	} else {
   1654 		tsp = NULL;
   1655 	}
   1656 
   1657 	return do_futex(SCARG(uap, uaddr), SCARG(uap, op), SCARG(uap, val),
   1658 	    tsp, SCARG(uap, uaddr2), SCARG(uap, val2), SCARG(uap, val3),
   1659 	    retval);
   1660 }
   1661 
   1662 /*
   1663  * sys___futex_set_robust_list(l, uap, retval)
   1664  *
   1665  *	__futex_set_robust_list(2) system call for robust futexes.
   1666  */
   1667 int
   1668 sys___futex_set_robust_list(struct lwp *l,
   1669     const struct sys___futex_set_robust_list_args *uap, register_t *retval)
   1670 {
   1671 	/* {
   1672 		syscallarg(void *) head;
   1673 		syscallarg(size_t) len;
   1674 	} */
   1675 	void *head = SCARG(uap, head);
   1676 
   1677 	if (SCARG(uap, len) != _FUTEX_ROBUST_HEAD_SIZE)
   1678 		return EINVAL;
   1679 	if ((uintptr_t)head % sizeof(u_long))
   1680 		return EINVAL;
   1681 
   1682 	l->l_robust_head = (uintptr_t)head;
   1683 
   1684 	return 0;
   1685 }
   1686 
   1687 /*
   1688  * sys___futex_get_robust_list(l, uap, retval)
   1689  *
   1690  *	__futex_get_robust_list(2) system call for robust futexes.
   1691  */
   1692 int
   1693 sys___futex_get_robust_list(struct lwp *l,
   1694     const struct sys___futex_get_robust_list_args *uap, register_t *retval)
   1695 {
   1696 	/* {
   1697 		syscallarg(lwpid_t) lwpid;
   1698 		syscallarg(void **) headp;
   1699 		syscallarg(size_t *) lenp;
   1700 	} */
   1701 	void *head;
   1702 	const size_t len = _FUTEX_ROBUST_HEAD_SIZE;
   1703 	int error;
   1704 
   1705 	error = futex_robust_head_lookup(l, SCARG(uap, lwpid), &head);
   1706 	if (error)
   1707 		return error;
   1708 
   1709 	/* Copy out the head pointer and the head structure length. */
   1710 	error = copyout(&head, SCARG(uap, headp), sizeof(head));
   1711 	if (__predict_true(error == 0)) {
   1712 		error = copyout(&len, SCARG(uap, lenp), sizeof(len));
   1713 	}
   1714 
   1715 	return error;
   1716 }
   1717 
   1718 /*
   1719  * release_futex(uva, tid)
   1720  *
   1721  *	Try to release the robust futex at uva in the current process
   1722  *	on lwp exit.  If anything goes wrong, silently fail.  It is the
   1723  *	userland program's obligation to arrange correct behaviour.
   1724  */
   1725 static void
   1726 release_futex(uintptr_t const uptr, lwpid_t const tid, bool const is_pi,
   1727     bool const is_pending)
   1728 {
   1729 	int *uaddr;
   1730 	struct futex *f;
   1731 	int oldval, newval, actual;
   1732 	int error;
   1733 
   1734 	/* If it's misaligned, tough.  */
   1735 	if (__predict_false(uptr & 3))
   1736 		return;
   1737 	uaddr = (int *)uptr;
   1738 
   1739 	error = futex_load(uaddr, &oldval);
   1740 	if (__predict_false(error))
   1741 		return;
   1742 
   1743 	/*
   1744 	 * There are two race conditions we need to handle here:
   1745 	 *
   1746 	 * 1. User space cleared the futex word but died before
   1747 	 *    being able to issue the wakeup.  No wakeups will
   1748 	 *    ever be issued, oops!
   1749 	 *
   1750 	 * 2. Awakened waiter died before being able to acquire
   1751 	 *    the futex in user space.  Any other waiters are
   1752 	 *    now stuck, oops!
   1753 	 *
   1754 	 * In both of these cases, the futex word will be 0 (because
   1755 	 * it's updated before the wake is issued).  The best we can
   1756 	 * do is detect this situation if it's the pending futex and
   1757 	 * issue a wake without modifying the futex word.
   1758 	 *
   1759 	 * XXX eventual PI handling?
   1760 	 */
   1761 	if (__predict_false(is_pending && (oldval & ~FUTEX_WAITERS) == 0)) {
   1762 		register_t retval;
   1763 		(void) futex_func_wake(/*shared*/true, uaddr, 1,
   1764 		    FUTEX_BITSET_MATCH_ANY, &retval);
   1765 		return;
   1766 	}
   1767 
   1768 	/* Optimistically test whether we need to do anything at all.  */
   1769 	if ((oldval & FUTEX_TID_MASK) != tid)
   1770 		return;
   1771 
   1772 	/*
   1773 	 * We need to handle the case where this thread owned the futex,
   1774 	 * but it was uncontended.  In this case, there won't be any
   1775 	 * kernel state to look up.  All we can do is mark the futex
   1776 	 * as a zombie to be mopped up the next time another thread
   1777 	 * attempts to acquire it.
   1778 	 *
   1779 	 * N.B. It's important to ensure to set FUTEX_OWNER_DIED in
   1780 	 * this loop, even if waiters appear while we're are doing
   1781 	 * so.  This is beause FUTEX_WAITERS is set by user space
   1782 	 * before calling __futex() to wait, and the futex needs
   1783 	 * to be marked as a zombie when the new waiter gets into
   1784 	 * the kernel.
   1785 	 */
   1786 	if ((oldval & FUTEX_WAITERS) == 0) {
   1787 		do {
   1788 			error = futex_load(uaddr, &oldval);
   1789 			if (error)
   1790 				return;
   1791 			if ((oldval & FUTEX_TID_MASK) != tid)
   1792 				return;
   1793 			newval = oldval | FUTEX_OWNER_DIED;
   1794 			error = ucas_int(uaddr, oldval, newval, &actual);
   1795 			if (error)
   1796 				return;
   1797 		} while (actual != oldval);
   1798 
   1799 		/*
   1800 		 * If where is still no indication of waiters, then there is
   1801 		 * no more work for us to do.
   1802 		 */
   1803 		if ((oldval & FUTEX_WAITERS) == 0)
   1804 			return;
   1805 	}
   1806 
   1807 	/*
   1808 	 * Look for a shared futex since we have no positive indication
   1809 	 * it is private.  If we can't, tough.
   1810 	 */
   1811 	error = futex_lookup(uaddr, /*shared*/true, &f);
   1812 	if (error)
   1813 		return;
   1814 
   1815 	/*
   1816 	 * If there's no kernel state for this futex, there's nothing to
   1817 	 * release.
   1818 	 */
   1819 	if (f == NULL)
   1820 		return;
   1821 
   1822 	/* Work under the futex queue lock.  */
   1823 	futex_queue_lock(f);
   1824 
   1825 	/*
   1826 	 * Fetch the word: if the tid doesn't match ours, skip;
   1827 	 * otherwise, set the owner-died bit, atomically.
   1828 	 */
   1829 	do {
   1830 		error = futex_load(uaddr, &oldval);
   1831 		if (error)
   1832 			goto out;
   1833 		if ((oldval & FUTEX_TID_MASK) != tid)
   1834 			goto out;
   1835 		newval = oldval | FUTEX_OWNER_DIED;
   1836 		error = ucas_int(uaddr, oldval, newval, &actual);
   1837 		if (error)
   1838 			goto out;
   1839 	} while (actual != oldval);
   1840 
   1841 	/*
   1842 	 * If there may be waiters, try to wake one.  If anything goes
   1843 	 * wrong, tough.
   1844 	 *
   1845 	 * XXX eventual PI handling?
   1846 	 */
   1847 	if (oldval & FUTEX_WAITERS)
   1848 		(void)futex_wake(f, 1, NULL, 0, FUTEX_BITSET_MATCH_ANY);
   1849 
   1850 	/* Unlock the queue and release the futex.  */
   1851 out:	futex_queue_unlock(f);
   1852 	futex_rele(f);
   1853 }
   1854 
   1855 /*
   1856  * futex_robust_head_lookup(l, lwpid)
   1857  *
   1858  *	Helper function to look up a robust head by LWP ID.
   1859  */
   1860 int
   1861 futex_robust_head_lookup(struct lwp *l, lwpid_t lwpid, void **headp)
   1862 {
   1863 	struct proc *p = l->l_proc;
   1864 
   1865 	/* Find the other lwp, if requested; otherwise use our robust head.  */
   1866 	if (lwpid) {
   1867 		mutex_enter(p->p_lock);
   1868 		l = lwp_find(p, lwpid);
   1869 		if (l == NULL) {
   1870 			mutex_exit(p->p_lock);
   1871 			return ESRCH;
   1872 		}
   1873 		*headp = (void *)l->l_robust_head;
   1874 		mutex_exit(p->p_lock);
   1875 	} else {
   1876 		*headp = (void *)l->l_robust_head;
   1877 	}
   1878 	return 0;
   1879 }
   1880 
   1881 /*
   1882  * futex_fetch_robust_head(uaddr)
   1883  *
   1884  *	Helper routine to fetch the futex robust list head that
   1885  *	handles 32-bit binaries running on 64-bit kernels.
   1886  */
   1887 static int
   1888 futex_fetch_robust_head(uintptr_t uaddr, u_long *rhead)
   1889 {
   1890 #ifdef _LP64
   1891 	if (curproc->p_flag & PK_32) {
   1892 		uint32_t rhead32[_FUTEX_ROBUST_HEAD_NWORDS];
   1893 		int error;
   1894 
   1895 		error = copyin((void *)uaddr, rhead32, sizeof(rhead32));
   1896 		if (__predict_true(error == 0)) {
   1897 			for (int i = 0; i < _FUTEX_ROBUST_HEAD_NWORDS; i++) {
   1898 				if (i == _FUTEX_ROBUST_HEAD_OFFSET) {
   1899 					/*
   1900 					 * Make sure the offset is sign-
   1901 					 * extended.
   1902 					 */
   1903 					rhead[i] = (int32_t)rhead32[i];
   1904 				} else {
   1905 					rhead[i] = rhead32[i];
   1906 				}
   1907 			}
   1908 		}
   1909 		return error;
   1910 	}
   1911 #endif /* _L64 */
   1912 
   1913 	return copyin((void *)uaddr, rhead,
   1914 	    sizeof(*rhead) * _FUTEX_ROBUST_HEAD_NWORDS);
   1915 }
   1916 
   1917 /*
   1918  * futex_decode_robust_word(word)
   1919  *
   1920  *	Decode a robust futex list word into the entry and entry
   1921  *	properties.
   1922  */
   1923 static inline void
   1924 futex_decode_robust_word(uintptr_t const word, uintptr_t * const entry,
   1925     bool * const is_pi)
   1926 {
   1927 	*is_pi = (word & _FUTEX_ROBUST_ENTRY_PI) ? true : false;
   1928 	*entry = word & ~_FUTEX_ROBUST_ENTRY_PI;
   1929 }
   1930 
   1931 /*
   1932  * futex_fetch_robust_entry(uaddr)
   1933  *
   1934  *	Helper routine to fetch and decode a robust futex entry
   1935  *	that handles 32-bit binaries running on 64-bit kernels.
   1936  */
   1937 static int
   1938 futex_fetch_robust_entry(uintptr_t const uaddr, uintptr_t * const valp,
   1939     bool * const is_pi)
   1940 {
   1941 	uintptr_t val = 0;
   1942 	int error = 0;
   1943 
   1944 #ifdef _LP64
   1945 	if (curproc->p_flag & PK_32) {
   1946 		uint32_t val32;
   1947 
   1948 		error = ufetch_32((uint32_t *)uaddr, &val32);
   1949 		if (__predict_true(error == 0))
   1950 			val = val32;
   1951 	} else
   1952 #endif /* _LP64 */
   1953 		error = ufetch_long((u_long *)uaddr, (u_long *)&val);
   1954 	if (__predict_false(error))
   1955 		return error;
   1956 
   1957 	futex_decode_robust_word(val, valp, is_pi);
   1958 	return 0;
   1959 }
   1960 
   1961 /*
   1962  * futex_release_all_lwp(l, tid)
   1963  *
   1964  *	Release all l's robust futexes.  If anything looks funny in
   1965  *	the process, give up -- it's userland's responsibility to dot
   1966  *	the i's and cross the t's.
   1967  */
   1968 void
   1969 futex_release_all_lwp(struct lwp * const l, lwpid_t const tid)
   1970 {
   1971 	u_long rhead[_FUTEX_ROBUST_HEAD_NWORDS];
   1972 	int limit = 1000000;
   1973 	int error;
   1974 
   1975 	/* If there's no robust list there's nothing to do. */
   1976 	if (l->l_robust_head == 0)
   1977 		return;
   1978 
   1979 	/* Read the final snapshot of the robust list head. */
   1980 	error = futex_fetch_robust_head(l->l_robust_head, rhead);
   1981 	if (error) {
   1982 		printf("WARNING: pid %jd (%s) lwp %jd tid %jd:"
   1983 		    " unmapped robust futex list head\n",
   1984 		    (uintmax_t)l->l_proc->p_pid, l->l_proc->p_comm,
   1985 		    (uintmax_t)l->l_lid, (uintmax_t)tid);
   1986 		return;
   1987 	}
   1988 
   1989 	const long offset = (long)rhead[_FUTEX_ROBUST_HEAD_OFFSET];
   1990 
   1991 	uintptr_t next, pending;
   1992 	bool is_pi, pending_is_pi;
   1993 
   1994 	futex_decode_robust_word(rhead[_FUTEX_ROBUST_HEAD_LIST],
   1995 	    &next, &is_pi);
   1996 	futex_decode_robust_word(rhead[_FUTEX_ROBUST_HEAD_PENDING],
   1997 	    &pending, &pending_is_pi);
   1998 
   1999 	/*
   2000 	 * Walk down the list of locked futexes and release them, up
   2001 	 * to one million of them before we give up.
   2002 	 */
   2003 
   2004 	while (next != l->l_robust_head && limit-- > 0) {
   2005 		/* pending handled below. */
   2006 		if (next != pending)
   2007 			release_futex(next + offset, tid, is_pi, false);
   2008 		error = futex_fetch_robust_entry(next, &next, &is_pi);
   2009 		if (error)
   2010 			break;
   2011 		preempt_point();
   2012 	}
   2013 	if (limit <= 0) {
   2014 		printf("WARNING: pid %jd (%s) lwp %jd tid %jd:"
   2015 		    " exhausted robust futex limit\n",
   2016 		    (uintmax_t)l->l_proc->p_pid, l->l_proc->p_comm,
   2017 		    (uintmax_t)l->l_lid, (uintmax_t)tid);
   2018 	}
   2019 
   2020 	/* If there's a pending futex, it may need to be released too. */
   2021 	if (pending != 0) {
   2022 		release_futex(pending + offset, tid, pending_is_pi, true);
   2023 	}
   2024 }
   2025