Home | History | Annotate | Line # | Download | only in kern
sys_futex.c revision 1.6
      1 /*	$NetBSD: sys_futex.c,v 1.6 2020/04/28 16:22:25 riastradh 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 1.6 2020/04/28 16:22:25 riastradh 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/types.h>
    118 #include <sys/atomic.h>
    119 #include <sys/condvar.h>
    120 #include <sys/futex.h>
    121 #include <sys/mutex.h>
    122 #include <sys/rbtree.h>
    123 #include <sys/queue.h>
    124 
    125 #include <sys/syscall.h>
    126 #include <sys/syscallargs.h>
    127 #include <sys/syscallvar.h>
    128 
    129 #include <uvm/uvm_extern.h>
    130 
    131 /*
    132  * Lock order:
    133  *
    134  *	futex_tab.lock
    135  *	futex::fx_qlock			ordered by kva of struct futex
    136  *	 -> futex_wait::fw_lock		only one at a time
    137  *	futex_wait::fw_lock		only one at a time
    138  *	 -> futex::fx_abortlock		only one at a time
    139  */
    140 
    141 /*
    142  * union futex_key
    143  *
    144  *	A futex is addressed either by a vmspace+va (private) or by
    145  *	a uvm_voaddr (shared).
    146  */
    147 union futex_key {
    148 	struct {
    149 		struct vmspace			*vmspace;
    150 		vaddr_t				va;
    151 	}			fk_private;
    152 	struct uvm_voaddr	fk_shared;
    153 };
    154 
    155 /*
    156  * struct futex
    157  *
    158  *	Kernel state for a futex located at a particular address in a
    159  *	particular virtual address space.
    160  *
    161  *	N.B. fx_refcnt is an unsigned long because we need to be able
    162  *	to operate on it atomically on all systems while at the same
    163  *	time rendering practically impossible the chance of it reaching
    164  *	its max value.  In practice, we're limited by the number of LWPs
    165  *	that can be present on the system at any given time, and the
    166  *	assumption is that limit will be good enough on a 32-bit platform.
    167  *	See futex_wake() for why overflow needs to be avoided.
    168  */
    169 struct futex {
    170 	union futex_key		fx_key;
    171 	unsigned long		fx_refcnt;
    172 	bool			fx_shared;
    173 	bool			fx_on_tree;
    174 	struct rb_node		fx_node;
    175 
    176 	kmutex_t			fx_qlock;
    177 	TAILQ_HEAD(, futex_wait)	fx_queue;
    178 
    179 	kmutex_t			fx_abortlock;
    180 	LIST_HEAD(, futex_wait)		fx_abortlist;
    181 	kcondvar_t			fx_abortcv;
    182 };
    183 
    184 /*
    185  * struct futex_wait
    186  *
    187  *	State for a thread to wait on a futex.  Threads wait on fw_cv
    188  *	for fw_bitset to be set to zero.  The thread may transition to
    189  *	a different futex queue at any time under the futex's lock.
    190  */
    191 struct futex_wait {
    192 	kmutex_t		fw_lock;
    193 	kcondvar_t		fw_cv;
    194 	struct futex		*fw_futex;
    195 	TAILQ_ENTRY(futex_wait)	fw_entry;	/* queue lock */
    196 	LIST_ENTRY(futex_wait)	fw_abort;	/* queue abortlock */
    197 	int			fw_bitset;
    198 	bool			fw_aborting;	/* fw_lock */
    199 };
    200 
    201 /*
    202  * futex_tab
    203  *
    204  *	Global trees of futexes by vmspace/va and VM object address.
    205  *
    206  *	XXX This obviously doesn't scale in parallel.  We could use a
    207  *	pserialize-safe data structure, but there may be a high cost to
    208  *	frequent deletion since we don't cache futexes after we're done
    209  *	with them.  We could use hashed locks.  But for now, just make
    210  *	sure userland can't DoS the serial performance, by using a
    211  *	balanced binary tree for lookup.
    212  *
    213  *	XXX We could use a per-process tree for the table indexed by
    214  *	virtual address to reduce contention between processes.
    215  */
    216 static struct {
    217 	kmutex_t	lock;
    218 	struct rb_tree	va;
    219 	struct rb_tree	oa;
    220 } futex_tab __cacheline_aligned;
    221 
    222 static int
    223 compare_futex_key(void *cookie, const void *n, const void *k)
    224 {
    225 	const struct futex *fa = n;
    226 	const union futex_key *fka = &fa->fx_key;
    227 	const union futex_key *fkb = k;
    228 
    229 	if ((uintptr_t)fka->fk_private.vmspace <
    230 	    (uintptr_t)fkb->fk_private.vmspace)
    231 		return -1;
    232 	if ((uintptr_t)fka->fk_private.vmspace >
    233 	    (uintptr_t)fkb->fk_private.vmspace)
    234 		return +1;
    235 	if (fka->fk_private.va < fkb->fk_private.va)
    236 		return -1;
    237 	if (fka->fk_private.va > fkb->fk_private.va)
    238 		return -1;
    239 	return 0;
    240 }
    241 
    242 static int
    243 compare_futex(void *cookie, const void *na, const void *nb)
    244 {
    245 	const struct futex *fa = na;
    246 	const struct futex *fb = nb;
    247 
    248 	return compare_futex_key(cookie, fa, &fb->fx_key);
    249 }
    250 
    251 static const rb_tree_ops_t futex_rb_ops = {
    252 	.rbto_compare_nodes = compare_futex,
    253 	.rbto_compare_key = compare_futex_key,
    254 	.rbto_node_offset = offsetof(struct futex, fx_node),
    255 };
    256 
    257 static int
    258 compare_futex_shared_key(void *cookie, const void *n, const void *k)
    259 {
    260 	const struct futex *fa = n;
    261 	const union futex_key *fka = &fa->fx_key;
    262 	const union futex_key *fkb = k;
    263 
    264 	return uvm_voaddr_compare(&fka->fk_shared, &fkb->fk_shared);
    265 }
    266 
    267 static int
    268 compare_futex_shared(void *cookie, const void *na, const void *nb)
    269 {
    270 	const struct futex *fa = na;
    271 	const struct futex *fb = nb;
    272 
    273 	return compare_futex_shared_key(cookie, fa, &fb->fx_key);
    274 }
    275 
    276 static const rb_tree_ops_t futex_shared_rb_ops = {
    277 	.rbto_compare_nodes = compare_futex_shared,
    278 	.rbto_compare_key = compare_futex_shared_key,
    279 	.rbto_node_offset = offsetof(struct futex, fx_node),
    280 };
    281 
    282 static void	futex_wait_dequeue(struct futex_wait *, struct futex *);
    283 
    284 /*
    285  * futex_load(uaddr, kaddr)
    286  *
    287  *	Perform a single atomic load to read *uaddr, and return the
    288  *	result in *kaddr.  Return 0 on success, EFAULT if uaddr is not
    289  *	mapped.
    290  */
    291 static inline int
    292 futex_load(int *uaddr, int *kaddr)
    293 {
    294 	return ufetch_int((u_int *)uaddr, (u_int *)kaddr);
    295 }
    296 
    297 /*
    298  * futex_test(uaddr, expected)
    299  *
    300  *	True if *uaddr == expected.  False if *uaddr != expected, or if
    301  *	uaddr is not mapped.
    302  */
    303 static bool
    304 futex_test(int *uaddr, int expected)
    305 {
    306 	int val;
    307 	int error;
    308 
    309 	error = futex_load(uaddr, &val);
    310 	if (error)
    311 		return false;
    312 	return val == expected;
    313 }
    314 
    315 /*
    316  * futex_sys_init()
    317  *
    318  *	Initialize the futex subsystem.
    319  */
    320 void
    321 futex_sys_init(void)
    322 {
    323 
    324 	mutex_init(&futex_tab.lock, MUTEX_DEFAULT, IPL_NONE);
    325 	rb_tree_init(&futex_tab.va, &futex_rb_ops);
    326 	rb_tree_init(&futex_tab.oa, &futex_shared_rb_ops);
    327 }
    328 
    329 /*
    330  * futex_sys_fini()
    331  *
    332  *	Finalize the futex subsystem.
    333  */
    334 void
    335 futex_sys_fini(void)
    336 {
    337 
    338 	KASSERT(RB_TREE_MIN(&futex_tab.oa) == NULL);
    339 	KASSERT(RB_TREE_MIN(&futex_tab.va) == NULL);
    340 	mutex_destroy(&futex_tab.lock);
    341 }
    342 
    343 /*
    344  * futex_queue_init(f)
    345  *
    346  *	Initialize the futex queue.  Caller must call futex_queue_fini
    347  *	when done.
    348  *
    349  *	Never sleeps.
    350  */
    351 static void
    352 futex_queue_init(struct futex *f)
    353 {
    354 
    355 	mutex_init(&f->fx_qlock, MUTEX_DEFAULT, IPL_NONE);
    356 	mutex_init(&f->fx_abortlock, MUTEX_DEFAULT, IPL_NONE);
    357 	cv_init(&f->fx_abortcv, "fqabort");
    358 	LIST_INIT(&f->fx_abortlist);
    359 	TAILQ_INIT(&f->fx_queue);
    360 }
    361 
    362 /*
    363  * futex_queue_drain(f)
    364  *
    365  *	Wait for any aborting waiters in f; then empty the queue of
    366  *	any stragglers and wake them.  Caller must guarantee no new
    367  *	references to f.
    368  *
    369  *	May sleep.
    370  */
    371 static void
    372 futex_queue_drain(struct futex *f)
    373 {
    374 	struct futex_wait *fw, *fw_next;
    375 
    376 	mutex_enter(&f->fx_abortlock);
    377 	while (!LIST_EMPTY(&f->fx_abortlist))
    378 		cv_wait(&f->fx_abortcv, &f->fx_abortlock);
    379 	mutex_exit(&f->fx_abortlock);
    380 
    381 	mutex_enter(&f->fx_qlock);
    382 	TAILQ_FOREACH_SAFE(fw, &f->fx_queue, fw_entry, fw_next) {
    383 		mutex_enter(&fw->fw_lock);
    384 		futex_wait_dequeue(fw, f);
    385 		cv_broadcast(&fw->fw_cv);
    386 		mutex_exit(&fw->fw_lock);
    387 	}
    388 	mutex_exit(&f->fx_qlock);
    389 }
    390 
    391 /*
    392  * futex_queue_fini(fq)
    393  *
    394  *	Finalize the futex queue initialized by futex_queue_init.  Queue
    395  *	must be empty.  Caller must not use f again until a subsequent
    396  *	futex_queue_init.
    397  */
    398 static void
    399 futex_queue_fini(struct futex *f)
    400 {
    401 
    402 	KASSERT(TAILQ_EMPTY(&f->fx_queue));
    403 	KASSERT(LIST_EMPTY(&f->fx_abortlist));
    404 	mutex_destroy(&f->fx_qlock);
    405 	mutex_destroy(&f->fx_abortlock);
    406 	cv_destroy(&f->fx_abortcv);
    407 }
    408 
    409 /*
    410  * futex_key_init(key, vm, va, shared)
    411  *
    412  *	Initialize a futex key for lookup, etc.
    413  */
    414 static int
    415 futex_key_init(union futex_key *fk, struct vmspace *vm, vaddr_t va, bool shared)
    416 {
    417 	int error = 0;
    418 
    419 	if (__predict_false(shared)) {
    420 		if (!uvm_voaddr_acquire(&vm->vm_map, va, &fk->fk_shared))
    421 			error = EFAULT;
    422 	} else {
    423 		fk->fk_private.vmspace = vm;
    424 		fk->fk_private.va = va;
    425 	}
    426 
    427 	return error;
    428 }
    429 
    430 /*
    431  * futex_key_fini(key, shared)
    432  *
    433  *	Release a futex key.
    434  */
    435 static void
    436 futex_key_fini(union futex_key *fk, bool shared)
    437 {
    438 	if (__predict_false(shared))
    439 		uvm_voaddr_release(&fk->fk_shared);
    440 	memset(fk, 0, sizeof(*fk));
    441 }
    442 
    443 /*
    444  * futex_create(fk, shared)
    445  *
    446  *	Create a futex.  Initial reference count is 1, representing the
    447  *	caller.  Returns NULL on failure.  Always takes ownership of the
    448  *	key, either transferring it to the newly-created futex, or releasing
    449  *	the key if creation fails.
    450  *
    451  *	Never sleeps for memory, but may sleep to acquire a lock.
    452  */
    453 static struct futex *
    454 futex_create(union futex_key *fk, bool shared)
    455 {
    456 	struct futex *f;
    457 
    458 	f = kmem_alloc(sizeof(*f), KM_NOSLEEP);
    459 	if (f == NULL) {
    460 		futex_key_fini(fk, shared);
    461 		return NULL;
    462 	}
    463 	f->fx_key = *fk;
    464 	f->fx_refcnt = 1;
    465 	f->fx_shared = shared;
    466 	f->fx_on_tree = false;
    467 	futex_queue_init(f);
    468 
    469 	return f;
    470 }
    471 
    472 /*
    473  * futex_destroy(f)
    474  *
    475  *	Destroy a futex created with futex_create.  Reference count
    476  *	must be zero.
    477  *
    478  *	May sleep.
    479  */
    480 static void
    481 futex_destroy(struct futex *f)
    482 {
    483 
    484 	ASSERT_SLEEPABLE();
    485 
    486 	KASSERT(atomic_load_relaxed(&f->fx_refcnt) == 0);
    487 	KASSERT(!f->fx_on_tree);
    488 
    489 	/* Drain and destroy the private queue.  */
    490 	futex_queue_drain(f);
    491 	futex_queue_fini(f);
    492 
    493 	futex_key_fini(&f->fx_key, f->fx_shared);
    494 
    495 	kmem_free(f, sizeof(*f));
    496 }
    497 
    498 /*
    499  * futex_hold(f)
    500  *
    501  *	Attempt to acquire a reference to f.  Return 0 on success,
    502  *	ENFILE on too many references.
    503  *
    504  *	Never sleeps.
    505  */
    506 static int
    507 futex_hold(struct futex *f)
    508 {
    509 	unsigned long refcnt;
    510 
    511 	do {
    512 		refcnt = atomic_load_relaxed(&f->fx_refcnt);
    513 		if (refcnt == ULONG_MAX)
    514 			return ENFILE;
    515 	} while (atomic_cas_ulong(&f->fx_refcnt, refcnt, refcnt + 1) != refcnt);
    516 
    517 	return 0;
    518 }
    519 
    520 /*
    521  * futex_rele(f)
    522  *
    523  *	Release a reference to f acquired with futex_create or
    524  *	futex_hold.
    525  *
    526  *	May sleep to free f.
    527  */
    528 static void
    529 futex_rele(struct futex *f)
    530 {
    531 	unsigned long refcnt;
    532 
    533 	ASSERT_SLEEPABLE();
    534 
    535 	do {
    536 		refcnt = atomic_load_relaxed(&f->fx_refcnt);
    537 		if (refcnt == 1)
    538 			goto trylast;
    539 	} while (atomic_cas_ulong(&f->fx_refcnt, refcnt, refcnt - 1) != refcnt);
    540 	return;
    541 
    542 trylast:
    543 	mutex_enter(&futex_tab.lock);
    544 	if (atomic_dec_ulong_nv(&f->fx_refcnt) == 0) {
    545 		if (f->fx_on_tree) {
    546 			if (__predict_false(f->fx_shared))
    547 				rb_tree_remove_node(&futex_tab.oa, f);
    548 			else
    549 				rb_tree_remove_node(&futex_tab.va, f);
    550 			f->fx_on_tree = false;
    551 		}
    552 	} else {
    553 		/* References remain -- don't destroy it.  */
    554 		f = NULL;
    555 	}
    556 	mutex_exit(&futex_tab.lock);
    557 	if (f != NULL)
    558 		futex_destroy(f);
    559 }
    560 
    561 /*
    562  * futex_rele_not_last(f)
    563  *
    564  *	Release a reference to f acquired with futex_create or
    565  *	futex_hold.
    566  *
    567  *	This version asserts that we are not dropping the last
    568  *	reference to f.
    569  */
    570 static void
    571 futex_rele_not_last(struct futex *f)
    572 {
    573 	unsigned long refcnt;
    574 
    575 	do {
    576 		refcnt = atomic_load_relaxed(&f->fx_refcnt);
    577 		KASSERT(refcnt > 1);
    578 	} while (atomic_cas_ulong(&f->fx_refcnt, refcnt, refcnt - 1) != refcnt);
    579 }
    580 
    581 /*
    582  * futex_lookup_by_key(key, shared, &f)
    583  *
    584  *	Try to find an existing futex va reference in the specified key
    585  *	On success, return 0, set f to found futex or to NULL if not found,
    586  *	and increment f's reference count if found.
    587  *
    588  *	Return ENFILE if reference count too high.
    589  *
    590  *	Internal lookup routine shared by futex_lookup() and
    591  *	futex_lookup_create().
    592  */
    593 static int
    594 futex_lookup_by_key(union futex_key *fk, bool shared, struct futex **fp)
    595 {
    596 	struct futex *f;
    597 	int error = 0;
    598 
    599 	mutex_enter(&futex_tab.lock);
    600 	if (__predict_false(shared)) {
    601 		f = rb_tree_find_node(&futex_tab.oa, fk);
    602 	} else {
    603 		f = rb_tree_find_node(&futex_tab.va, fk);
    604 	}
    605 	if (f) {
    606 		error = futex_hold(f);
    607 		if (error)
    608 			f = NULL;
    609 	}
    610  	*fp = f;
    611 	mutex_exit(&futex_tab.lock);
    612 
    613 	return error;
    614 }
    615 
    616 /*
    617  * futex_insert(f, fp)
    618  *
    619  *	Try to insert the futex f into the tree by va.  If there
    620  *	already is a futex for its va, acquire a reference to it, and
    621  *	store it in *fp; otherwise store f in *fp.
    622  *
    623  *	Return 0 on success, ENFILE if there already is a futex but its
    624  *	reference count is too high.
    625  */
    626 static int
    627 futex_insert(struct futex *f, struct futex **fp)
    628 {
    629 	struct futex *f0;
    630 	int error;
    631 
    632 	KASSERT(atomic_load_relaxed(&f->fx_refcnt) != 0);
    633 	KASSERT(!f->fx_on_tree);
    634 
    635 	mutex_enter(&futex_tab.lock);
    636 	if (__predict_false(f->fx_shared))
    637 		f0 = rb_tree_insert_node(&futex_tab.oa, f);
    638 	else
    639 		f0 = rb_tree_insert_node(&futex_tab.va, f);
    640 	if (f0 == f) {
    641 		f->fx_on_tree = true;
    642 		error = 0;
    643 	} else {
    644 		KASSERT(atomic_load_relaxed(&f0->fx_refcnt) != 0);
    645 		KASSERT(f0->fx_on_tree);
    646 		error = futex_hold(f0);
    647 		if (error)
    648 			goto out;
    649 	}
    650 	*fp = f0;
    651 out:	mutex_exit(&futex_tab.lock);
    652 
    653 	return error;
    654 }
    655 
    656 /*
    657  * futex_lookup(uaddr, shared, &f)
    658  *
    659  *	Find a futex at the userland pointer uaddr in the current
    660  *	process's VM space.  On success, return the futex in f and
    661  *	increment its reference count.
    662  *
    663  *	Caller must call futex_rele when done.
    664  */
    665 static int
    666 futex_lookup(int *uaddr, bool shared, struct futex **fp)
    667 {
    668 	union futex_key fk;
    669 	struct vmspace *vm = curproc->p_vmspace;
    670 	vaddr_t va = (vaddr_t)uaddr;
    671 	int error;
    672 
    673 	/*
    674 	 * Reject unaligned user pointers so we don't cross page
    675 	 * boundaries and so atomics will work.
    676 	 */
    677 	if ((va & 3) != 0)
    678 		return EINVAL;
    679 
    680 	/* Look it up. */
    681 	error = futex_key_init(&fk, vm, va, shared);
    682 	if (error)
    683 		return error;
    684 
    685 	error = futex_lookup_by_key(&fk, shared, fp);
    686 	futex_key_fini(&fk, shared);
    687 	if (error)
    688 		return error;
    689 
    690 	KASSERT(*fp == NULL || (*fp)->fx_shared == shared);
    691 	KASSERT(*fp == NULL || atomic_load_relaxed(&(*fp)->fx_refcnt) != 0);
    692 
    693 	/*
    694 	 * Success!  (Caller must still check whether we found
    695 	 * anything, but nothing went _wrong_ like trying to use
    696 	 * unmapped memory.)
    697 	 */
    698 	KASSERT(error == 0);
    699 
    700 	return error;
    701 }
    702 
    703 /*
    704  * futex_lookup_create(uaddr, shared, &f)
    705  *
    706  *	Find or create a futex at the userland pointer uaddr in the
    707  *	current process's VM space.  On success, return the futex in f
    708  *	and increment its reference count.
    709  *
    710  *	Caller must call futex_rele when done.
    711  */
    712 static int
    713 futex_lookup_create(int *uaddr, bool shared, struct futex **fp)
    714 {
    715 	union futex_key fk;
    716 	struct vmspace *vm = curproc->p_vmspace;
    717 	struct futex *f = NULL;
    718 	vaddr_t va = (vaddr_t)uaddr;
    719 	int error;
    720 
    721 	/*
    722 	 * Reject unaligned user pointers so we don't cross page
    723 	 * boundaries and so atomics will work.
    724 	 */
    725 	if ((va & 3) != 0)
    726 		return EINVAL;
    727 
    728 	error = futex_key_init(&fk, vm, va, shared);
    729 	if (error)
    730 		return error;
    731 
    732 	/*
    733 	 * Optimistically assume there already is one, and try to find
    734 	 * it.
    735 	 */
    736 	error = futex_lookup_by_key(&fk, shared, fp);
    737 	if (error || *fp != NULL) {
    738 		/*
    739 		 * We either found one, or there was an error.
    740 		 * In either case, we are done with the key.
    741 		 */
    742 		futex_key_fini(&fk, shared);
    743 		goto out;
    744 	}
    745 
    746 	/*
    747 	 * Create a futex recoard.  This tranfers ownership of the key
    748 	 * in all cases.
    749 	 */
    750 	f = futex_create(&fk, shared);
    751 	if (f == NULL) {
    752 		error = ENOMEM;
    753 		goto out;
    754 	}
    755 
    756 	/*
    757 	 * Insert our new futex, or use existing if someone else beat
    758 	 * us to it.
    759 	 */
    760 	error = futex_insert(f, fp);
    761 	if (error)
    762 		goto out;
    763 	if (*fp == f)
    764 		f = NULL;	/* don't release on exit */
    765 
    766 	/* Success!  */
    767 	KASSERT(error == 0);
    768 
    769 out:	if (f != NULL)
    770 		futex_rele(f);
    771 	KASSERT(error || *fp != NULL);
    772 	KASSERT(error || atomic_load_relaxed(&(*fp)->fx_refcnt) != 0);
    773 	return error;
    774 }
    775 
    776 /*
    777  * futex_wait_init(fw, bitset)
    778  *
    779  *	Initialize a record for a thread to wait on a futex matching
    780  *	the specified bit set.  Should be passed to futex_wait_enqueue
    781  *	before futex_wait, and should be passed to futex_wait_fini when
    782  *	done.
    783  */
    784 static void
    785 futex_wait_init(struct futex_wait *fw, int bitset)
    786 {
    787 
    788 	KASSERT(bitset);
    789 
    790 	mutex_init(&fw->fw_lock, MUTEX_DEFAULT, IPL_NONE);
    791 	cv_init(&fw->fw_cv, "futex");
    792 	fw->fw_futex = NULL;
    793 	fw->fw_bitset = bitset;
    794 	fw->fw_aborting = false;
    795 }
    796 
    797 /*
    798  * futex_wait_fini(fw)
    799  *
    800  *	Finalize a record for a futex waiter.  Must not be on any
    801  *	futex's queue.
    802  */
    803 static void
    804 futex_wait_fini(struct futex_wait *fw)
    805 {
    806 
    807 	KASSERT(fw->fw_futex == NULL);
    808 
    809 	cv_destroy(&fw->fw_cv);
    810 	mutex_destroy(&fw->fw_lock);
    811 }
    812 
    813 /*
    814  * futex_wait_enqueue(fw, f)
    815  *
    816  *	Put fw on the futex queue.  Must be done before futex_wait.
    817  *	Caller must hold fw's lock and f's lock, and fw must not be on
    818  *	any existing futex's waiter list.
    819  */
    820 static void
    821 futex_wait_enqueue(struct futex_wait *fw, struct futex *f)
    822 {
    823 
    824 	KASSERT(mutex_owned(&f->fx_qlock));
    825 	KASSERT(mutex_owned(&fw->fw_lock));
    826 	KASSERT(fw->fw_futex == NULL);
    827 	KASSERT(!fw->fw_aborting);
    828 
    829 	fw->fw_futex = f;
    830 	TAILQ_INSERT_TAIL(&f->fx_queue, fw, fw_entry);
    831 }
    832 
    833 /*
    834  * futex_wait_dequeue(fw, f)
    835  *
    836  *	Remove fw from the futex queue.  Precludes subsequent
    837  *	futex_wait until a futex_wait_enqueue.  Caller must hold fw's
    838  *	lock and f's lock, and fw must be on f.
    839  */
    840 static void
    841 futex_wait_dequeue(struct futex_wait *fw, struct futex *f)
    842 {
    843 
    844 	KASSERT(mutex_owned(&f->fx_qlock));
    845 	KASSERT(mutex_owned(&fw->fw_lock));
    846 	KASSERT(fw->fw_futex == f);
    847 
    848 	TAILQ_REMOVE(&f->fx_queue, fw, fw_entry);
    849 	fw->fw_futex = NULL;
    850 }
    851 
    852 /*
    853  * futex_wait_abort(fw)
    854  *
    855  *	Caller is no longer waiting for fw.  Remove it from any queue
    856  *	if it was on one.  Caller must hold fw->fw_lock.
    857  */
    858 static void
    859 futex_wait_abort(struct futex_wait *fw)
    860 {
    861 	struct futex *f;
    862 
    863 	KASSERT(mutex_owned(&fw->fw_lock));
    864 
    865 	/*
    866 	 * Grab the futex queue.  It can't go away as long as we hold
    867 	 * fw_lock.  However, we can't take the queue lock because
    868 	 * that's a lock order reversal.
    869 	 */
    870 	f = fw->fw_futex;
    871 
    872 	/* Put us on the abort list so that fq won't go away.  */
    873 	mutex_enter(&f->fx_abortlock);
    874 	LIST_INSERT_HEAD(&f->fx_abortlist, fw, fw_abort);
    875 	mutex_exit(&f->fx_abortlock);
    876 
    877 	/*
    878 	 * Mark fw as aborting so it won't lose wakeups and won't be
    879 	 * transferred to any other queue.
    880 	 */
    881 	fw->fw_aborting = true;
    882 
    883 	/* f is now stable, so we can release fw_lock.  */
    884 	mutex_exit(&fw->fw_lock);
    885 
    886 	/* Now we can remove fw under the queue lock.  */
    887 	mutex_enter(&f->fx_qlock);
    888 	mutex_enter(&fw->fw_lock);
    889 	futex_wait_dequeue(fw, f);
    890 	mutex_exit(&fw->fw_lock);
    891 	mutex_exit(&f->fx_qlock);
    892 
    893 	/*
    894 	 * Finally, remove us from the abort list and notify anyone
    895 	 * waiting for the abort to complete if we were the last to go.
    896 	 */
    897 	mutex_enter(&f->fx_abortlock);
    898 	LIST_REMOVE(fw, fw_abort);
    899 	if (LIST_EMPTY(&f->fx_abortlist))
    900 		cv_broadcast(&f->fx_abortcv);
    901 	mutex_exit(&f->fx_abortlock);
    902 
    903 	/*
    904 	 * Release our reference to the futex now that we are not
    905 	 * waiting for it.
    906 	 */
    907 	futex_rele(f);
    908 
    909 	/*
    910 	 * Reacquire the fw lock as caller expects.  Verify that we're
    911 	 * aborting and no longer associated with a futex.
    912 	 */
    913 	mutex_enter(&fw->fw_lock);
    914 	KASSERT(fw->fw_aborting);
    915 	KASSERT(fw->fw_futex == NULL);
    916 }
    917 
    918 /*
    919  * futex_wait(fw, deadline, clkid)
    920  *
    921  *	fw must be a waiter on a futex's queue.  Wait until deadline on
    922  *	the clock clkid, or forever if deadline is NULL, for a futex
    923  *	wakeup.  Return 0 on explicit wakeup or destruction of futex,
    924  *	ETIMEDOUT on timeout, EINTR/ERESTART on signal.  Either way, fw
    925  *	will no longer be on a futex queue on return.
    926  */
    927 static int
    928 futex_wait(struct futex_wait *fw, const struct timespec *deadline,
    929     clockid_t clkid)
    930 {
    931 	int error = 0;
    932 
    933 	/* Test and wait under the wait lock.  */
    934 	mutex_enter(&fw->fw_lock);
    935 
    936 	for (;;) {
    937 		/* If we're done yet, stop and report success.  */
    938 		if (fw->fw_bitset == 0 || fw->fw_futex == NULL) {
    939 			error = 0;
    940 			break;
    941 		}
    942 
    943 		/* If anything went wrong in the last iteration, stop.  */
    944 		if (error)
    945 			break;
    946 
    947 		/* Not done yet.  Wait.  */
    948 		if (deadline) {
    949 			struct timespec ts;
    950 
    951 			/* Check our watch.  */
    952 			error = clock_gettime1(clkid, &ts);
    953 			if (error)
    954 				break;
    955 
    956 			/* If we're past the deadline, ETIMEDOUT.  */
    957 			if (timespeccmp(deadline, &ts, <=)) {
    958 				error = ETIMEDOUT;
    959 				break;
    960 			}
    961 
    962 			/* Count how much time is left.  */
    963 			timespecsub(deadline, &ts, &ts);
    964 
    965 			/* Wait for that much time, allowing signals.  */
    966 			error = cv_timedwait_sig(&fw->fw_cv, &fw->fw_lock,
    967 			    tstohz(&ts));
    968 		} else {
    969 			/* Wait indefinitely, allowing signals. */
    970 			error = cv_wait_sig(&fw->fw_cv, &fw->fw_lock);
    971 		}
    972 	}
    973 
    974 	/*
    975 	 * If we were woken up, the waker will have removed fw from the
    976 	 * queue.  But if anything went wrong, we must remove fw from
    977 	 * the queue ourselves.  While here, convert EWOULDBLOCK to
    978 	 * ETIMEDOUT.
    979 	 */
    980 	if (error) {
    981 		futex_wait_abort(fw);
    982 		if (error == EWOULDBLOCK)
    983 			error = ETIMEDOUT;
    984 	}
    985 
    986 	mutex_exit(&fw->fw_lock);
    987 
    988 	return error;
    989 }
    990 
    991 /*
    992  * futex_wake(f, nwake, f2, nrequeue, bitset)
    993  *
    994  *	Wake up to nwake waiters on f matching bitset; then, if f2 is
    995  *	provided, move up to nrequeue remaining waiters on f matching
    996  *	bitset to f2.  Return the number of waiters actually woken.
    997  *	Caller must hold the locks of f and f2, if provided.
    998  */
    999 static unsigned
   1000 futex_wake(struct futex *f, unsigned nwake, struct futex *f2,
   1001     unsigned nrequeue, int bitset)
   1002 {
   1003 	struct futex_wait *fw, *fw_next;
   1004 	unsigned nwoken = 0;
   1005 	int hold_error __diagused;
   1006 
   1007 	KASSERT(mutex_owned(&f->fx_qlock));
   1008 	KASSERT(f2 == NULL || mutex_owned(&f2->fx_qlock));
   1009 
   1010 	/* Wake up to nwake waiters, and count the number woken.  */
   1011 	TAILQ_FOREACH_SAFE(fw, &f->fx_queue, fw_entry, fw_next) {
   1012 		if ((fw->fw_bitset & bitset) == 0)
   1013 			continue;
   1014 		if (nwake > 0) {
   1015 			mutex_enter(&fw->fw_lock);
   1016 			if (__predict_false(fw->fw_aborting)) {
   1017 				mutex_exit(&fw->fw_lock);
   1018 				continue;
   1019 			}
   1020 			futex_wait_dequeue(fw, f);
   1021 			fw->fw_bitset = 0;
   1022 			cv_broadcast(&fw->fw_cv);
   1023 			mutex_exit(&fw->fw_lock);
   1024 			nwake--;
   1025 			nwoken++;
   1026 			/*
   1027 			 * Drop the futex reference on behalf of the
   1028 			 * waiter.  We assert this is not the last
   1029 			 * reference on the futex (our caller should
   1030 			 * also have one).
   1031 			 */
   1032 			futex_rele_not_last(f);
   1033 		} else {
   1034 			break;
   1035 		}
   1036 	}
   1037 
   1038 	if (f2) {
   1039 		/* Move up to nrequeue waiters from f's queue to f2's queue. */
   1040 		TAILQ_FOREACH_SAFE(fw, &f->fx_queue, fw_entry, fw_next) {
   1041 			if ((fw->fw_bitset & bitset) == 0)
   1042 				continue;
   1043 			if (nrequeue > 0) {
   1044 				mutex_enter(&fw->fw_lock);
   1045 				if (__predict_false(fw->fw_aborting)) {
   1046 					mutex_exit(&fw->fw_lock);
   1047 					continue;
   1048 				}
   1049 				futex_wait_dequeue(fw, f);
   1050 				futex_wait_enqueue(fw, f2);
   1051 				mutex_exit(&fw->fw_lock);
   1052 				nrequeue--;
   1053 				/*
   1054 				 * Transfer the reference from f to f2.
   1055 				 * As above, we assert that we are not
   1056 				 * dropping the last reference to f here.
   1057 				 *
   1058 				 * XXX futex_hold() could theoretically
   1059 				 * XXX fail here.
   1060 				 */
   1061 				futex_rele_not_last(f);
   1062 				hold_error = futex_hold(f2);
   1063 				KASSERT(hold_error == 0);
   1064 			} else {
   1065 				break;
   1066 			}
   1067 		}
   1068 	} else {
   1069 		KASSERT(nrequeue == 0);
   1070 	}
   1071 
   1072 	/* Return the number of waiters woken.  */
   1073 	return nwoken;
   1074 }
   1075 
   1076 /*
   1077  * futex_queue_lock(f)
   1078  *
   1079  *	Acquire the queue lock of f.  Pair with futex_queue_unlock.  Do
   1080  *	not use if caller needs to acquire two locks; use
   1081  *	futex_queue_lock2 instead.
   1082  */
   1083 static void
   1084 futex_queue_lock(struct futex *f)
   1085 {
   1086 	mutex_enter(&f->fx_qlock);
   1087 }
   1088 
   1089 /*
   1090  * futex_queue_unlock(f)
   1091  *
   1092  *	Release the queue lock of f.
   1093  */
   1094 static void
   1095 futex_queue_unlock(struct futex *f)
   1096 {
   1097 	mutex_exit(&f->fx_qlock);
   1098 }
   1099 
   1100 /*
   1101  * futex_queue_lock2(f, f2)
   1102  *
   1103  *	Acquire the queue locks of both f and f2, which may be null, or
   1104  *	which may have the same underlying queue.  If they are
   1105  *	distinct, an arbitrary total order is chosen on the locks.
   1106  *
   1107  *	Callers should only ever acquire multiple queue locks
   1108  *	simultaneously using futex_queue_lock2.
   1109  */
   1110 static void
   1111 futex_queue_lock2(struct futex *f, struct futex *f2)
   1112 {
   1113 
   1114 	/*
   1115 	 * If both are null, do nothing; if one is null and the other
   1116 	 * is not, lock the other and be done with it.
   1117 	 */
   1118 	if (f == NULL && f2 == NULL) {
   1119 		return;
   1120 	} else if (f == NULL) {
   1121 		mutex_enter(&f2->fx_qlock);
   1122 		return;
   1123 	} else if (f2 == NULL) {
   1124 		mutex_enter(&f->fx_qlock);
   1125 		return;
   1126 	}
   1127 
   1128 	/* If both futexes are the same, acquire only one. */
   1129 	if (f == f2) {
   1130 		mutex_enter(&f->fx_qlock);
   1131 		return;
   1132 	}
   1133 
   1134 	/* Otherwise, use the ordering on the kva of the futex pointer.  */
   1135 	if ((uintptr_t)f < (uintptr_t)f2) {
   1136 		mutex_enter(&f->fx_qlock);
   1137 		mutex_enter(&f2->fx_qlock);
   1138 	} else {
   1139 		mutex_enter(&f2->fx_qlock);
   1140 		mutex_enter(&f->fx_qlock);
   1141 	}
   1142 }
   1143 
   1144 /*
   1145  * futex_queue_unlock2(f, f2)
   1146  *
   1147  *	Release the queue locks of both f and f2, which may be null, or
   1148  *	which may have the same underlying queue.
   1149  */
   1150 static void
   1151 futex_queue_unlock2(struct futex *f, struct futex *f2)
   1152 {
   1153 
   1154 	/*
   1155 	 * If both are null, do nothing; if one is null and the other
   1156 	 * is not, unlock the other and be done with it.
   1157 	 */
   1158 	if (f == NULL && f2 == NULL) {
   1159 		return;
   1160 	} else if (f == NULL) {
   1161 		mutex_exit(&f2->fx_qlock);
   1162 		return;
   1163 	} else if (f2 == NULL) {
   1164 		mutex_exit(&f->fx_qlock);
   1165 		return;
   1166 	}
   1167 
   1168 	/* If both futexes are the same, release only one. */
   1169 	if (f == f2) {
   1170 		mutex_exit(&f->fx_qlock);
   1171 		return;
   1172 	}
   1173 
   1174 	/* Otherwise, use the ordering on the kva of the futex pointer.  */
   1175 	if ((uintptr_t)f < (uintptr_t)f2) {
   1176 		mutex_exit(&f2->fx_qlock);
   1177 		mutex_exit(&f->fx_qlock);
   1178 	} else {
   1179 		mutex_exit(&f->fx_qlock);
   1180 		mutex_exit(&f2->fx_qlock);
   1181 	}
   1182 }
   1183 
   1184 /*
   1185  * futex_func_wait(uaddr, val, val3, timeout, clkid, clkflags, retval)
   1186  *
   1187  *	Implement futex(FUTEX_WAIT).
   1188  */
   1189 static int
   1190 futex_func_wait(bool shared, int *uaddr, int val, int val3,
   1191     const struct timespec *timeout, clockid_t clkid, int clkflags,
   1192     register_t *retval)
   1193 {
   1194 	struct futex *f;
   1195 	struct futex_wait wait, *fw = &wait;
   1196 	struct timespec ts;
   1197 	const struct timespec *deadline;
   1198 	int error;
   1199 
   1200 	/*
   1201 	 * If there's nothing to wait for, and nobody will ever wake
   1202 	 * us, then don't set anything up to wait -- just stop here.
   1203 	 */
   1204 	if (val3 == 0)
   1205 		return 0;
   1206 
   1207 	/* Optimistically test before anything else.  */
   1208 	if (!futex_test(uaddr, val))
   1209 		return EAGAIN;
   1210 
   1211 	/* Determine a deadline on the specified clock.  */
   1212 	if (timeout == NULL || (clkflags & TIMER_ABSTIME) == TIMER_ABSTIME) {
   1213 		deadline = timeout;
   1214 	} else {
   1215 		error = clock_gettime1(clkid, &ts);
   1216 		if (error)
   1217 			return error;
   1218 		timespecadd(&ts, timeout, &ts);
   1219 		deadline = &ts;
   1220 	}
   1221 
   1222 	/* Get the futex, creating it if necessary.  */
   1223 	error = futex_lookup_create(uaddr, shared, &f);
   1224 	if (error)
   1225 		return error;
   1226 	KASSERT(f);
   1227 
   1228 	/* Get ready to wait.  */
   1229 	futex_wait_init(fw, val3);
   1230 
   1231 	/*
   1232 	 * Under the queue lock, check the value again: if it has
   1233 	 * already changed, EAGAIN; otherwise enqueue the waiter.
   1234 	 * Since FUTEX_WAKE will use the same lock and be done after
   1235 	 * modifying the value, the order in which we check and enqueue
   1236 	 * is immaterial.
   1237 	 */
   1238 	futex_queue_lock(f);
   1239 	if (!futex_test(uaddr, val)) {
   1240 		futex_queue_unlock(f);
   1241 		error = EAGAIN;
   1242 		goto out;
   1243 	}
   1244 	mutex_enter(&fw->fw_lock);
   1245 	futex_wait_enqueue(fw, f);
   1246 	mutex_exit(&fw->fw_lock);
   1247 	futex_queue_unlock(f);
   1248 
   1249 	/*
   1250 	 * We cannot drop our reference to the futex here, because
   1251 	 * we might be enqueued on a different one when we are awakened.
   1252 	 * The references will be managed on our behalf in the requeue
   1253 	 * and wake cases.
   1254 	 */
   1255 	f = NULL;
   1256 
   1257 	/* Wait. */
   1258 	error = futex_wait(fw, deadline, clkid);
   1259 	if (error)
   1260 		goto out;
   1261 
   1262 	/* Return 0 on success, error on failure. */
   1263 	*retval = 0;
   1264 
   1265 out:	if (f != NULL)
   1266 		futex_rele(f);
   1267 	futex_wait_fini(fw);
   1268 	return error;
   1269 }
   1270 
   1271 /*
   1272  * futex_func_wake(uaddr, val, val3, retval)
   1273  *
   1274  *	Implement futex(FUTEX_WAKE) and futex(FUTEX_WAKE_BITSET).
   1275  */
   1276 static int
   1277 futex_func_wake(bool shared, int *uaddr, int val, int val3, register_t *retval)
   1278 {
   1279 	struct futex *f;
   1280 	unsigned int nwoken = 0;
   1281 	int error = 0;
   1282 
   1283 	/* Reject negative number of wakeups.  */
   1284 	if (val < 0) {
   1285 		error = EINVAL;
   1286 		goto out;
   1287 	}
   1288 
   1289 	/* Look up the futex, if any.  */
   1290 	error = futex_lookup(uaddr, shared, &f);
   1291 	if (error)
   1292 		goto out;
   1293 
   1294 	/* If there's no futex, there are no waiters to wake.  */
   1295 	if (f == NULL)
   1296 		goto out;
   1297 
   1298 	/*
   1299 	 * Under f's queue lock, wake the waiters and remember the
   1300 	 * number woken.
   1301 	 */
   1302 	futex_queue_lock(f);
   1303 	nwoken = futex_wake(f, val, NULL, 0, val3);
   1304 	futex_queue_unlock(f);
   1305 
   1306 	/* Release the futex.  */
   1307 	futex_rele(f);
   1308 
   1309 out:
   1310 	/* Return the number of waiters woken.  */
   1311 	*retval = nwoken;
   1312 
   1313 	/* Success!  */
   1314 	return error;
   1315 }
   1316 
   1317 /*
   1318  * futex_func_requeue(op, uaddr, val, uaddr2, val2, val3, retval)
   1319  *
   1320  *	Implement futex(FUTEX_REQUEUE) and futex(FUTEX_CMP_REQUEUE).
   1321  */
   1322 static int
   1323 futex_func_requeue(bool shared, int op, int *uaddr, int val, int *uaddr2,
   1324     int val2, int val3, register_t *retval)
   1325 {
   1326 	struct futex *f = NULL, *f2 = NULL;
   1327 	unsigned nwoken = 0;	/* default to zero woken on early return */
   1328 	int error;
   1329 
   1330 	/* Reject negative number of wakeups or requeues. */
   1331 	if (val < 0 || val2 < 0) {
   1332 		error = EINVAL;
   1333 		goto out;
   1334 	}
   1335 
   1336 	/* Look up the source futex, if any. */
   1337 	error = futex_lookup(uaddr, shared, &f);
   1338 	if (error)
   1339 		goto out;
   1340 
   1341 	/* If there is none, nothing to do. */
   1342 	if (f == NULL)
   1343 		goto out;
   1344 
   1345 	/*
   1346 	 * We may need to create the destination futex because it's
   1347 	 * entirely possible it does not currently have any waiters.
   1348 	 */
   1349 	error = futex_lookup_create(uaddr2, shared, &f2);
   1350 	if (error)
   1351 		goto out;
   1352 
   1353 	/*
   1354 	 * Under the futexes' queue locks, check the value; if
   1355 	 * unchanged from val3, wake the waiters.
   1356 	 */
   1357 	futex_queue_lock2(f, f2);
   1358 	if (op == FUTEX_CMP_REQUEUE && !futex_test(uaddr, val3)) {
   1359 		error = EAGAIN;
   1360 	} else {
   1361 		error = 0;
   1362 		nwoken = futex_wake(f, val, f2, val2, FUTEX_BITSET_MATCH_ANY);
   1363 	}
   1364 	futex_queue_unlock2(f, f2);
   1365 
   1366 out:
   1367 	/* Return the number of waiters woken.  */
   1368 	*retval = nwoken;
   1369 
   1370 	/* Release the futexes if we got them.  */
   1371 	if (f2)
   1372 		futex_rele(f2);
   1373 	if (f)
   1374 		futex_rele(f);
   1375 	return error;
   1376 }
   1377 
   1378 /*
   1379  * futex_validate_op_cmp(val3)
   1380  *
   1381  *	Validate an op/cmp argument for FUTEX_WAKE_OP.
   1382  */
   1383 static int
   1384 futex_validate_op_cmp(int val3)
   1385 {
   1386 	int op = __SHIFTOUT(val3, FUTEX_OP_OP_MASK);
   1387 	int cmp = __SHIFTOUT(val3, FUTEX_OP_CMP_MASK);
   1388 
   1389 	if (op & FUTEX_OP_OPARG_SHIFT) {
   1390 		int oparg = __SHIFTOUT(val3, FUTEX_OP_OPARG_MASK);
   1391 		if (oparg < 0)
   1392 			return EINVAL;
   1393 		if (oparg >= 32)
   1394 			return EINVAL;
   1395 		op &= ~FUTEX_OP_OPARG_SHIFT;
   1396 	}
   1397 
   1398 	switch (op) {
   1399 	case FUTEX_OP_SET:
   1400 	case FUTEX_OP_ADD:
   1401 	case FUTEX_OP_OR:
   1402 	case FUTEX_OP_ANDN:
   1403 	case FUTEX_OP_XOR:
   1404 		break;
   1405 	default:
   1406 		return EINVAL;
   1407 	}
   1408 
   1409 	switch (cmp) {
   1410 	case FUTEX_OP_CMP_EQ:
   1411 	case FUTEX_OP_CMP_NE:
   1412 	case FUTEX_OP_CMP_LT:
   1413 	case FUTEX_OP_CMP_LE:
   1414 	case FUTEX_OP_CMP_GT:
   1415 	case FUTEX_OP_CMP_GE:
   1416 		break;
   1417 	default:
   1418 		return EINVAL;
   1419 	}
   1420 
   1421 	return 0;
   1422 }
   1423 
   1424 /*
   1425  * futex_compute_op(oldval, val3)
   1426  *
   1427  *	Apply a FUTEX_WAIT_OP operation to oldval.
   1428  */
   1429 static int
   1430 futex_compute_op(int oldval, int val3)
   1431 {
   1432 	int op = __SHIFTOUT(val3, FUTEX_OP_OP_MASK);
   1433 	int oparg = __SHIFTOUT(val3, FUTEX_OP_OPARG_MASK);
   1434 
   1435 	if (op & FUTEX_OP_OPARG_SHIFT) {
   1436 		KASSERT(oparg >= 0);
   1437 		KASSERT(oparg < 32);
   1438 		oparg = 1u << oparg;
   1439 		op &= ~FUTEX_OP_OPARG_SHIFT;
   1440 	}
   1441 
   1442 	switch (op) {
   1443 	case FUTEX_OP_SET:
   1444 		return oparg;
   1445 
   1446 	case FUTEX_OP_ADD:
   1447 		/*
   1448 		 * Avoid signed arithmetic overflow by doing
   1449 		 * arithmetic unsigned and converting back to signed
   1450 		 * at the end.
   1451 		 */
   1452 		return (int)((unsigned)oldval + (unsigned)oparg);
   1453 
   1454 	case FUTEX_OP_OR:
   1455 		return oldval | oparg;
   1456 
   1457 	case FUTEX_OP_ANDN:
   1458 		return oldval & ~oparg;
   1459 
   1460 	case FUTEX_OP_XOR:
   1461 		return oldval ^ oparg;
   1462 
   1463 	default:
   1464 		panic("invalid futex op");
   1465 	}
   1466 }
   1467 
   1468 /*
   1469  * futex_compute_cmp(oldval, val3)
   1470  *
   1471  *	Apply a FUTEX_WAIT_OP comparison to oldval.
   1472  */
   1473 static bool
   1474 futex_compute_cmp(int oldval, int val3)
   1475 {
   1476 	int cmp = __SHIFTOUT(val3, FUTEX_OP_CMP_MASK);
   1477 	int cmparg = __SHIFTOUT(val3, FUTEX_OP_CMPARG_MASK);
   1478 
   1479 	switch (cmp) {
   1480 	case FUTEX_OP_CMP_EQ:
   1481 		return (oldval == cmparg);
   1482 
   1483 	case FUTEX_OP_CMP_NE:
   1484 		return (oldval != cmparg);
   1485 
   1486 	case FUTEX_OP_CMP_LT:
   1487 		return (oldval < cmparg);
   1488 
   1489 	case FUTEX_OP_CMP_LE:
   1490 		return (oldval <= cmparg);
   1491 
   1492 	case FUTEX_OP_CMP_GT:
   1493 		return (oldval > cmparg);
   1494 
   1495 	case FUTEX_OP_CMP_GE:
   1496 		return (oldval >= cmparg);
   1497 
   1498 	default:
   1499 		panic("invalid futex cmp operation");
   1500 	}
   1501 }
   1502 
   1503 /*
   1504  * futex_func_wake_op(uaddr, val, uaddr2, val2, val3, retval)
   1505  *
   1506  *	Implement futex(FUTEX_WAKE_OP).
   1507  */
   1508 static int
   1509 futex_func_wake_op(bool shared, int *uaddr, int val, int *uaddr2, int val2,
   1510     int val3, register_t *retval)
   1511 {
   1512 	struct futex *f = NULL, *f2 = NULL;
   1513 	int oldval, newval, actual;
   1514 	unsigned nwoken = 0;
   1515 	int error;
   1516 
   1517 	/* Reject negative number of wakeups.  */
   1518 	if (val < 0 || val2 < 0) {
   1519 		error = EINVAL;
   1520 		goto out;
   1521 	}
   1522 
   1523 	/* Reject invalid operations before we start doing things.  */
   1524 	if ((error = futex_validate_op_cmp(val3)) != 0)
   1525 		goto out;
   1526 
   1527 	/* Look up the first futex, if any.  */
   1528 	error = futex_lookup(uaddr, shared, &f);
   1529 	if (error)
   1530 		goto out;
   1531 
   1532 	/* Look up the second futex, if any.  */
   1533 	error = futex_lookup(uaddr2, shared, &f2);
   1534 	if (error)
   1535 		goto out;
   1536 
   1537 	/*
   1538 	 * Under the queue locks:
   1539 	 *
   1540 	 * 1. Read/modify/write: *uaddr2 op= oparg.
   1541 	 * 2. Unconditionally wake uaddr.
   1542 	 * 3. Conditionally wake uaddr2, if it previously matched val2.
   1543 	 */
   1544 	futex_queue_lock2(f, f2);
   1545 	do {
   1546 		error = futex_load(uaddr2, &oldval);
   1547 		if (error)
   1548 			goto out_unlock;
   1549 		newval = futex_compute_op(oldval, val3);
   1550 		error = ucas_int(uaddr2, oldval, newval, &actual);
   1551 		if (error)
   1552 			goto out_unlock;
   1553 	} while (actual != oldval);
   1554 	nwoken = (f ? futex_wake(f, val, NULL, 0, FUTEX_BITSET_MATCH_ANY) : 0);
   1555 	if (f2 && futex_compute_cmp(oldval, val3))
   1556 		nwoken += futex_wake(f2, val2, NULL, 0,
   1557 		    FUTEX_BITSET_MATCH_ANY);
   1558 
   1559 	/* Success! */
   1560 	error = 0;
   1561 out_unlock:
   1562 	futex_queue_unlock2(f, f2);
   1563 
   1564 out:
   1565 	/* Return the number of waiters woken. */
   1566 	*retval = nwoken;
   1567 
   1568 	/* Release the futexes, if we got them. */
   1569 	if (f2)
   1570 		futex_rele(f2);
   1571 	if (f)
   1572 		futex_rele(f);
   1573 	return error;
   1574 }
   1575 
   1576 /*
   1577  * do_futex(uaddr, op, val, timeout, uaddr2, val2, val3)
   1578  *
   1579  *	Implement the futex system call with all the parameters
   1580  *	parsed out.
   1581  */
   1582 int
   1583 do_futex(int *uaddr, int op, int val, const struct timespec *timeout,
   1584     int *uaddr2, int val2, int val3, register_t *retval)
   1585 {
   1586 	const bool shared = (op & FUTEX_PRIVATE_FLAG) ? false : true;
   1587 	const clockid_t clkid = (op & FUTEX_CLOCK_REALTIME) ? CLOCK_REALTIME
   1588 							    : CLOCK_MONOTONIC;
   1589 
   1590 	op &= FUTEX_CMD_MASK;
   1591 
   1592 	switch (op) {
   1593 	case FUTEX_WAIT:
   1594 		return futex_func_wait(shared, uaddr, val,
   1595 		    FUTEX_BITSET_MATCH_ANY, timeout, clkid, TIMER_RELTIME,
   1596 		    retval);
   1597 
   1598 	case FUTEX_WAKE:
   1599 		val3 = FUTEX_BITSET_MATCH_ANY;
   1600 		/* FALLTHROUGH */
   1601 	case FUTEX_WAKE_BITSET:
   1602 		return futex_func_wake(shared, uaddr, val, val3, retval);
   1603 
   1604 	case FUTEX_REQUEUE:
   1605 	case FUTEX_CMP_REQUEUE:
   1606 		return futex_func_requeue(shared, op, uaddr, val, uaddr2,
   1607 		    val2, val3, retval);
   1608 
   1609 	case FUTEX_WAIT_BITSET:
   1610 		return futex_func_wait(shared, uaddr, val, val3, timeout,
   1611 		    clkid, TIMER_ABSTIME, retval);
   1612 
   1613 	case FUTEX_WAKE_OP:
   1614 		return futex_func_wake_op(shared, uaddr, val, uaddr2, val2,
   1615 		    val3, retval);
   1616 
   1617 	case FUTEX_FD:
   1618 	default:
   1619 		return ENOSYS;
   1620 	}
   1621 }
   1622 
   1623 /*
   1624  * sys___futex(l, uap, retval)
   1625  *
   1626  *	__futex(2) system call: generic futex operations.
   1627  */
   1628 int
   1629 sys___futex(struct lwp *l, const struct sys___futex_args *uap,
   1630     register_t *retval)
   1631 {
   1632 	/* {
   1633 		syscallarg(int *) uaddr;
   1634 		syscallarg(int) op;
   1635 		syscallarg(int) val;
   1636 		syscallarg(const struct timespec *) timeout;
   1637 		syscallarg(int *) uaddr2;
   1638 		syscallarg(int) val2;
   1639 		syscallarg(int) val3;
   1640 	} */
   1641 	struct timespec ts, *tsp;
   1642 	int error;
   1643 
   1644 	/*
   1645 	 * Copy in the timeout argument, if specified.
   1646 	 */
   1647 	if (SCARG(uap, timeout)) {
   1648 		error = copyin(SCARG(uap, timeout), &ts, sizeof(ts));
   1649 		if (error)
   1650 			return error;
   1651 		tsp = &ts;
   1652 	} else {
   1653 		tsp = NULL;
   1654 	}
   1655 
   1656 	return do_futex(SCARG(uap, uaddr), SCARG(uap, op), SCARG(uap, val),
   1657 	    tsp, SCARG(uap, uaddr2), SCARG(uap, val2), SCARG(uap, val3),
   1658 	    retval);
   1659 }
   1660 
   1661 /*
   1662  * sys___futex_set_robust_list(l, uap, retval)
   1663  *
   1664  *	__futex_set_robust_list(2) system call for robust futexes.
   1665  */
   1666 int
   1667 sys___futex_set_robust_list(struct lwp *l,
   1668     const struct sys___futex_set_robust_list_args *uap, register_t *retval)
   1669 {
   1670 	/* {
   1671 		syscallarg(void *) head;
   1672 		syscallarg(size_t) len;
   1673 	} */
   1674 	void *head = SCARG(uap, head);
   1675 
   1676 	if (SCARG(uap, len) != _FUTEX_ROBUST_HEAD_SIZE)
   1677 		return EINVAL;
   1678 	if ((uintptr_t)head % sizeof(u_long))
   1679 		return EINVAL;
   1680 
   1681 	l->l_robust_head = (uintptr_t)head;
   1682 
   1683 	return 0;
   1684 }
   1685 
   1686 /*
   1687  * sys___futex_get_robust_list(l, uap, retval)
   1688  *
   1689  *	__futex_get_robust_list(2) system call for robust futexes.
   1690  */
   1691 int
   1692 sys___futex_get_robust_list(struct lwp *l,
   1693     const struct sys___futex_get_robust_list_args *uap, register_t *retval)
   1694 {
   1695 	/* {
   1696 		syscallarg(lwpid_t) lwpid;
   1697 		syscallarg(void **) headp;
   1698 		syscallarg(size_t *) lenp;
   1699 	} */
   1700 	void *head;
   1701 	const size_t len = _FUTEX_ROBUST_HEAD_SIZE;
   1702 	int error;
   1703 
   1704 	error = futex_robust_head_lookup(l, SCARG(uap, lwpid), &head);
   1705 	if (error)
   1706 		return error;
   1707 
   1708 	/* Copy out the head pointer and the head structure length. */
   1709 	error = copyout(&head, SCARG(uap, headp), sizeof(head));
   1710 	if (__predict_true(error == 0)) {
   1711 		error = copyout(&len, SCARG(uap, lenp), sizeof(len));
   1712 	}
   1713 
   1714 	return error;
   1715 }
   1716 
   1717 /*
   1718  * release_futex(uva, tid)
   1719  *
   1720  *	Try to release the robust futex at uva in the current process
   1721  *	on lwp exit.  If anything goes wrong, silently fail.  It is the
   1722  *	userland program's obligation to arrange correct behaviour.
   1723  */
   1724 static void
   1725 release_futex(uintptr_t const uptr, lwpid_t const tid, bool const is_pi,
   1726     bool const is_pending)
   1727 {
   1728 	int *uaddr;
   1729 	struct futex *f;
   1730 	int oldval, newval, actual;
   1731 	int error;
   1732 
   1733 	/* If it's misaligned, tough.  */
   1734 	if (__predict_false(uptr & 3))
   1735 		return;
   1736 	uaddr = (int *)uptr;
   1737 
   1738 	error = futex_load(uaddr, &oldval);
   1739 	if (__predict_false(error))
   1740 		return;
   1741 
   1742 	/*
   1743 	 * There are two race conditions we need to handle here:
   1744 	 *
   1745 	 * 1. User space cleared the futex word but died before
   1746 	 *    being able to issue the wakeup.  No wakeups will
   1747 	 *    ever be issued, oops!
   1748 	 *
   1749 	 * 2. Awakened waiter died before being able to acquire
   1750 	 *    the futex in user space.  Any other waiters are
   1751 	 *    now stuck, oops!
   1752 	 *
   1753 	 * In both of these cases, the futex word will be 0 (because
   1754 	 * it's updated before the wake is issued).  The best we can
   1755 	 * do is detect this situation if it's the pending futex and
   1756 	 * issue a wake without modifying the futex word.
   1757 	 *
   1758 	 * XXX eventual PI handling?
   1759 	 */
   1760 	if (__predict_false(is_pending && (oldval & ~FUTEX_WAITERS) == 0)) {
   1761 		register_t retval;
   1762 		(void) futex_func_wake(/*shared*/true, uaddr, 1,
   1763 		    FUTEX_BITSET_MATCH_ANY, &retval);
   1764 		return;
   1765 	}
   1766 
   1767 	/* Optimistically test whether we need to do anything at all.  */
   1768 	if ((oldval & FUTEX_TID_MASK) != tid)
   1769 		return;
   1770 
   1771 	/*
   1772 	 * We need to handle the case where this thread owned the futex,
   1773 	 * but it was uncontended.  In this case, there won't be any
   1774 	 * kernel state to look up.  All we can do is mark the futex
   1775 	 * as a zombie to be mopped up the next time another thread
   1776 	 * attempts to acquire it.
   1777 	 *
   1778 	 * N.B. It's important to ensure to set FUTEX_OWNER_DIED in
   1779 	 * this loop, even if waiters appear while we're are doing
   1780 	 * so.  This is beause FUTEX_WAITERS is set by user space
   1781 	 * before calling __futex() to wait, and the futex needs
   1782 	 * to be marked as a zombie when the new waiter gets into
   1783 	 * the kernel.
   1784 	 */
   1785 	if ((oldval & FUTEX_WAITERS) == 0) {
   1786 		do {
   1787 			error = futex_load(uaddr, &oldval);
   1788 			if (error)
   1789 				return;
   1790 			if ((oldval & FUTEX_TID_MASK) != tid)
   1791 				return;
   1792 			newval = oldval | FUTEX_OWNER_DIED;
   1793 			error = ucas_int(uaddr, oldval, newval, &actual);
   1794 			if (error)
   1795 				return;
   1796 		} while (actual != oldval);
   1797 
   1798 		/*
   1799 		 * If where is still no indication of waiters, then there is
   1800 		 * no more work for us to do.
   1801 		 */
   1802 		if ((oldval & FUTEX_WAITERS) == 0)
   1803 			return;
   1804 	}
   1805 
   1806 	/*
   1807 	 * Look for a shared futex since we have no positive indication
   1808 	 * it is private.  If we can't, tough.
   1809 	 */
   1810 	error = futex_lookup(uaddr, /*shared*/true, &f);
   1811 	if (error)
   1812 		return;
   1813 
   1814 	/*
   1815 	 * If there's no kernel state for this futex, there's nothing to
   1816 	 * release.
   1817 	 */
   1818 	if (f == NULL)
   1819 		return;
   1820 
   1821 	/* Work under the futex queue lock.  */
   1822 	futex_queue_lock(f);
   1823 
   1824 	/*
   1825 	 * Fetch the word: if the tid doesn't match ours, skip;
   1826 	 * otherwise, set the owner-died bit, atomically.
   1827 	 */
   1828 	do {
   1829 		error = futex_load(uaddr, &oldval);
   1830 		if (error)
   1831 			goto out;
   1832 		if ((oldval & FUTEX_TID_MASK) != tid)
   1833 			goto out;
   1834 		newval = oldval | FUTEX_OWNER_DIED;
   1835 		error = ucas_int(uaddr, oldval, newval, &actual);
   1836 		if (error)
   1837 			goto out;
   1838 	} while (actual != oldval);
   1839 
   1840 	/*
   1841 	 * If there may be waiters, try to wake one.  If anything goes
   1842 	 * wrong, tough.
   1843 	 *
   1844 	 * XXX eventual PI handling?
   1845 	 */
   1846 	if (oldval & FUTEX_WAITERS)
   1847 		(void)futex_wake(f, 1, NULL, 0, FUTEX_BITSET_MATCH_ANY);
   1848 
   1849 	/* Unlock the queue and release the futex.  */
   1850 out:	futex_queue_unlock(f);
   1851 	futex_rele(f);
   1852 }
   1853 
   1854 /*
   1855  * futex_robust_head_lookup(l, lwpid)
   1856  *
   1857  *	Helper function to look up a robust head by LWP ID.
   1858  */
   1859 int
   1860 futex_robust_head_lookup(struct lwp *l, lwpid_t lwpid, void **headp)
   1861 {
   1862 	struct proc *p = l->l_proc;
   1863 
   1864 	/* Find the other lwp, if requested; otherwise use our robust head.  */
   1865 	if (lwpid) {
   1866 		mutex_enter(p->p_lock);
   1867 		l = lwp_find(p, lwpid);
   1868 		if (l == NULL) {
   1869 			mutex_exit(p->p_lock);
   1870 			return ESRCH;
   1871 		}
   1872 		*headp = (void *)l->l_robust_head;
   1873 		mutex_exit(p->p_lock);
   1874 	} else {
   1875 		*headp = (void *)l->l_robust_head;
   1876 	}
   1877 	return 0;
   1878 }
   1879 
   1880 /*
   1881  * futex_fetch_robust_head(uaddr)
   1882  *
   1883  *	Helper routine to fetch the futex robust list head that
   1884  *	handles 32-bit binaries running on 64-bit kernels.
   1885  */
   1886 static int
   1887 futex_fetch_robust_head(uintptr_t uaddr, u_long *rhead)
   1888 {
   1889 #ifdef _LP64
   1890 	if (curproc->p_flag & PK_32) {
   1891 		uint32_t rhead32[_FUTEX_ROBUST_HEAD_NWORDS];
   1892 		int error;
   1893 
   1894 		error = copyin((void *)uaddr, rhead32, sizeof(rhead32));
   1895 		if (__predict_true(error == 0)) {
   1896 			for (int i = 0; i < _FUTEX_ROBUST_HEAD_NWORDS; i++) {
   1897 				if (i == _FUTEX_ROBUST_HEAD_OFFSET) {
   1898 					/*
   1899 					 * Make sure the offset is sign-
   1900 					 * extended.
   1901 					 */
   1902 					rhead[i] = (int32_t)rhead32[i];
   1903 				} else {
   1904 					rhead[i] = rhead32[i];
   1905 				}
   1906 			}
   1907 		}
   1908 		return error;
   1909 	}
   1910 #endif /* _L64 */
   1911 
   1912 	return copyin((void *)uaddr, rhead,
   1913 	    sizeof(*rhead) * _FUTEX_ROBUST_HEAD_NWORDS);
   1914 }
   1915 
   1916 /*
   1917  * futex_decode_robust_word(word)
   1918  *
   1919  *	Decode a robust futex list word into the entry and entry
   1920  *	properties.
   1921  */
   1922 static inline void
   1923 futex_decode_robust_word(uintptr_t const word, uintptr_t * const entry,
   1924     bool * const is_pi)
   1925 {
   1926 	*is_pi = (word & _FUTEX_ROBUST_ENTRY_PI) ? true : false;
   1927 	*entry = word & ~_FUTEX_ROBUST_ENTRY_PI;
   1928 }
   1929 
   1930 /*
   1931  * futex_fetch_robust_entry(uaddr)
   1932  *
   1933  *	Helper routine to fetch and decode a robust futex entry
   1934  *	that handles 32-bit binaries running on 64-bit kernels.
   1935  */
   1936 static int
   1937 futex_fetch_robust_entry(uintptr_t const uaddr, uintptr_t * const valp,
   1938     bool * const is_pi)
   1939 {
   1940 	uintptr_t val = 0;
   1941 	int error = 0;
   1942 
   1943 #ifdef _LP64
   1944 	if (curproc->p_flag & PK_32) {
   1945 		uint32_t val32;
   1946 
   1947 		error = ufetch_32((uint32_t *)uaddr, &val32);
   1948 		if (__predict_true(error == 0))
   1949 			val = val32;
   1950 	} else
   1951 #endif /* _LP64 */
   1952 		error = ufetch_long((u_long *)uaddr, (u_long *)&val);
   1953 	if (__predict_false(error))
   1954 		return error;
   1955 
   1956 	futex_decode_robust_word(val, valp, is_pi);
   1957 	return 0;
   1958 }
   1959 
   1960 /*
   1961  * futex_release_all_lwp(l, tid)
   1962  *
   1963  *	Release all l's robust futexes.  If anything looks funny in
   1964  *	the process, give up -- it's userland's responsibility to dot
   1965  *	the i's and cross the t's.
   1966  */
   1967 void
   1968 futex_release_all_lwp(struct lwp * const l, lwpid_t const tid)
   1969 {
   1970 	u_long rhead[_FUTEX_ROBUST_HEAD_NWORDS];
   1971 	int limit = 1000000;
   1972 	int error;
   1973 
   1974 	/* If there's no robust list there's nothing to do. */
   1975 	if (l->l_robust_head == 0)
   1976 		return;
   1977 
   1978 	/* Read the final snapshot of the robust list head. */
   1979 	error = futex_fetch_robust_head(l->l_robust_head, rhead);
   1980 	if (error) {
   1981 		printf("WARNING: pid %jd (%s) lwp %jd tid %jd:"
   1982 		    " unmapped robust futex list head\n",
   1983 		    (uintmax_t)l->l_proc->p_pid, l->l_proc->p_comm,
   1984 		    (uintmax_t)l->l_lid, (uintmax_t)tid);
   1985 		return;
   1986 	}
   1987 
   1988 	const long offset = (long)rhead[_FUTEX_ROBUST_HEAD_OFFSET];
   1989 
   1990 	uintptr_t next, pending;
   1991 	bool is_pi, pending_is_pi;
   1992 
   1993 	futex_decode_robust_word(rhead[_FUTEX_ROBUST_HEAD_LIST],
   1994 	    &next, &is_pi);
   1995 	futex_decode_robust_word(rhead[_FUTEX_ROBUST_HEAD_PENDING],
   1996 	    &pending, &pending_is_pi);
   1997 
   1998 	/*
   1999 	 * Walk down the list of locked futexes and release them, up
   2000 	 * to one million of them before we give up.
   2001 	 */
   2002 
   2003 	while (next != l->l_robust_head && limit-- > 0) {
   2004 		/* pending handled below. */
   2005 		if (next != pending)
   2006 			release_futex(next + offset, tid, is_pi, false);
   2007 		error = futex_fetch_robust_entry(next, &next, &is_pi);
   2008 		if (error)
   2009 			break;
   2010 		preempt_point();
   2011 	}
   2012 	if (limit <= 0) {
   2013 		printf("WARNING: pid %jd (%s) lwp %jd tid %jd:"
   2014 		    " exhausted robust futex limit\n",
   2015 		    (uintmax_t)l->l_proc->p_pid, l->l_proc->p_comm,
   2016 		    (uintmax_t)l->l_lid, (uintmax_t)tid);
   2017 	}
   2018 
   2019 	/* If there's a pending futex, it may need to be released too. */
   2020 	if (pending != 0) {
   2021 		release_futex(pending + offset, tid, pending_is_pi, true);
   2022 	}
   2023 }
   2024