xref: /illumos-gate/usr/src/uts/common/os/mutex.c (revision 48bbca81)
17c478bd9Sstevel@tonic-gate /*
27c478bd9Sstevel@tonic-gate  * CDDL HEADER START
37c478bd9Sstevel@tonic-gate  *
47c478bd9Sstevel@tonic-gate  * The contents of this file are subject to the terms of the
50efe5e54Sdv  * Common Development and Distribution License (the "License").
60efe5e54Sdv  * You may not use this file except in compliance with the License.
77c478bd9Sstevel@tonic-gate  *
87c478bd9Sstevel@tonic-gate  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
97c478bd9Sstevel@tonic-gate  * or http://www.opensolaris.org/os/licensing.
107c478bd9Sstevel@tonic-gate  * See the License for the specific language governing permissions
117c478bd9Sstevel@tonic-gate  * and limitations under the License.
127c478bd9Sstevel@tonic-gate  *
137c478bd9Sstevel@tonic-gate  * When distributing Covered Code, include this CDDL HEADER in each
147c478bd9Sstevel@tonic-gate  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
157c478bd9Sstevel@tonic-gate  * If applicable, add the following below this CDDL HEADER, with the
167c478bd9Sstevel@tonic-gate  * fields enclosed by brackets "[]" replaced with your own identifying
177c478bd9Sstevel@tonic-gate  * information: Portions Copyright [yyyy] [name of copyright owner]
187c478bd9Sstevel@tonic-gate  *
197c478bd9Sstevel@tonic-gate  * CDDL HEADER END
207c478bd9Sstevel@tonic-gate  */
217c478bd9Sstevel@tonic-gate /*
22753a6d45SSherry Moore  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
237c478bd9Sstevel@tonic-gate  * Use is subject to license terms.
24*48bbca81SDaniel Hoffman  * Copyright (c) 2016 by Delphix. All rights reserved.
257c478bd9Sstevel@tonic-gate  */
277c478bd9Sstevel@tonic-gate /*
287c478bd9Sstevel@tonic-gate  * Big Theory Statement for mutual exclusion locking primitives.
297c478bd9Sstevel@tonic-gate  *
307c478bd9Sstevel@tonic-gate  * A mutex serializes multiple threads so that only one thread
317c478bd9Sstevel@tonic-gate  * (the "owner" of the mutex) is active at a time.  See mutex(9F)
327c478bd9Sstevel@tonic-gate  * for a full description of the interfaces and programming model.
337c478bd9Sstevel@tonic-gate  * The rest of this comment describes the implementation.
347c478bd9Sstevel@tonic-gate  *
357c478bd9Sstevel@tonic-gate  * Mutexes come in two flavors: adaptive and spin.  mutex_init(9F)
367c478bd9Sstevel@tonic-gate  * determines the type based solely on the iblock cookie (PIL) argument.
377c478bd9Sstevel@tonic-gate  * PIL > LOCK_LEVEL implies a spin lock; everything else is adaptive.
387c478bd9Sstevel@tonic-gate  *
397c478bd9Sstevel@tonic-gate  * Spin mutexes block interrupts and spin until the lock becomes available.
407c478bd9Sstevel@tonic-gate  * A thread may not sleep, or call any function that might sleep, while
417c478bd9Sstevel@tonic-gate  * holding a spin mutex.  With few exceptions, spin mutexes should only
427c478bd9Sstevel@tonic-gate  * be used to synchronize with interrupt handlers.
437c478bd9Sstevel@tonic-gate  *
447c478bd9Sstevel@tonic-gate  * Adaptive mutexes (the default type) spin if the owner is running on
457c478bd9Sstevel@tonic-gate  * another CPU and block otherwise.  This policy is based on the assumption
467c478bd9Sstevel@tonic-gate  * that mutex hold times are typically short enough that the time spent
477c478bd9Sstevel@tonic-gate  * spinning is less than the time it takes to block.  If you need mutual
487c478bd9Sstevel@tonic-gate  * exclusion semantics with long hold times, consider an rwlock(9F) as
497c478bd9Sstevel@tonic-gate  * RW_WRITER.  Better still, reconsider the algorithm: if it requires
507c478bd9Sstevel@tonic-gate  * mutual exclusion for long periods of time, it's probably not scalable.
517c478bd9Sstevel@tonic-gate  *
527c478bd9Sstevel@tonic-gate  * Adaptive mutexes are overwhelmingly more common than spin mutexes,
537c478bd9Sstevel@tonic-gate  * so mutex_enter() assumes that the lock is adaptive.  We get away
547c478bd9Sstevel@tonic-gate  * with this by structuring mutexes so that an attempt to acquire a
557c478bd9Sstevel@tonic-gate  * spin mutex as adaptive always fails.  When mutex_enter() fails
567c478bd9Sstevel@tonic-gate  * it punts to mutex_vector_enter(), which does all the hard stuff.
577c478bd9Sstevel@tonic-gate  *
587c478bd9Sstevel@tonic-gate  * mutex_vector_enter() first checks the type.  If it's spin mutex,
597c478bd9Sstevel@tonic-gate  * we just call lock_set_spl() and return.  If it's an adaptive mutex,
607c478bd9Sstevel@tonic-gate  * we check to see what the owner is doing.  If the owner is running,
617c478bd9Sstevel@tonic-gate  * we spin until the lock becomes available; if not, we mark the lock
627c478bd9Sstevel@tonic-gate  * as having waiters and block.
637c478bd9Sstevel@tonic-gate  *
647c478bd9Sstevel@tonic-gate  * Blocking on a mutex is surprisingly delicate dance because, for speed,
657c478bd9Sstevel@tonic-gate  * mutex_exit() doesn't use an atomic instruction.  Thus we have to work
667c478bd9Sstevel@tonic-gate  * a little harder in the (rarely-executed) blocking path to make sure
677c478bd9Sstevel@tonic-gate  * we don't block on a mutex that's just been released -- otherwise we
687c478bd9Sstevel@tonic-gate  * might never be woken up.
697c478bd9Sstevel@tonic-gate  *
707c478bd9Sstevel@tonic-gate  * The logic for synchronizing mutex_vector_enter() with mutex_exit()
717c478bd9Sstevel@tonic-gate  * in the face of preemption and relaxed memory ordering is as follows:
727c478bd9Sstevel@tonic-gate  *
737c478bd9Sstevel@tonic-gate  * (1) Preemption in the middle of mutex_exit() must cause mutex_exit()
747c478bd9Sstevel@tonic-gate  *     to restart.  Each platform must enforce this by checking the
757c478bd9Sstevel@tonic-gate  *     interrupted PC in the interrupt handler (or on return from trap --
767c478bd9Sstevel@tonic-gate  *     whichever is more convenient for the platform).  If the PC
777c478bd9Sstevel@tonic-gate  *     lies within the critical region of mutex_exit(), the interrupt
787c478bd9Sstevel@tonic-gate  *     handler must reset the PC back to the beginning of mutex_exit().
797c478bd9Sstevel@tonic-gate  *     The critical region consists of all instructions up to, but not
807c478bd9Sstevel@tonic-gate  *     including, the store that clears the lock (which, of course,
817c478bd9Sstevel@tonic-gate  *     must never be executed twice.)
827c478bd9Sstevel@tonic-gate  *
837c478bd9Sstevel@tonic-gate  *     This ensures that the owner will always check for waiters after
847c478bd9Sstevel@tonic-gate  *     resuming from a previous preemption.
857c478bd9Sstevel@tonic-gate  *
867c478bd9Sstevel@tonic-gate  * (2) A thread resuming in mutex_exit() does (at least) the following:
877c478bd9Sstevel@tonic-gate  *
887c478bd9Sstevel@tonic-gate  *	when resuming:	set CPU_THREAD = owner
897c478bd9Sstevel@tonic-gate  *			membar #StoreLoad
907c478bd9Sstevel@tonic-gate  *
917c478bd9Sstevel@tonic-gate  *	in mutex_exit:	check waiters bit; do wakeup if set
927c478bd9Sstevel@tonic-gate  *			membar #LoadStore|#StoreStore
937c478bd9Sstevel@tonic-gate  *			clear owner
947c478bd9Sstevel@tonic-gate  *			(at this point, other threads may or may not grab
957c478bd9Sstevel@tonic-gate  *			the lock, and we may or may not reacquire it)
967c478bd9Sstevel@tonic-gate  *
977c478bd9Sstevel@tonic-gate  *	when blocking:	membar #StoreStore (due to disp_lock_enter())
987c478bd9Sstevel@tonic-gate  *			set CPU_THREAD = (possibly) someone else
997c478bd9Sstevel@tonic-gate  *
1007c478bd9Sstevel@tonic-gate  * (3) A thread blocking in mutex_vector_enter() does the following:
1017c478bd9Sstevel@tonic-gate  *
1027c478bd9Sstevel@tonic-gate  *			set waiters bit
1037c478bd9Sstevel@tonic-gate  *			membar #StoreLoad (via membar_enter())
104575a7426Spt  *			check CPU_THREAD for owner's t_cpu
105575a7426Spt  *				continue if owner running
1067c478bd9Sstevel@tonic-gate  *			membar #LoadLoad (via membar_consumer())
1077c478bd9Sstevel@tonic-gate  *			check owner and waiters bit; abort if either changed
1087c478bd9Sstevel@tonic-gate  *			block
1097c478bd9Sstevel@tonic-gate  *
1107c478bd9Sstevel@tonic-gate  * Thus the global memory orderings for (2) and (3) are as follows:
1117c478bd9Sstevel@tonic-gate  *
1127c478bd9Sstevel@tonic-gate  * (2M) mutex_exit() memory order:
1137c478bd9Sstevel@tonic-gate  *
1147c478bd9Sstevel@tonic-gate  *			STORE	CPU_THREAD = owner
1157c478bd9Sstevel@tonic-gate  *			LOAD	waiters bit
1167c478bd9Sstevel@tonic-gate  *			STORE	owner = NULL
1177c478bd9Sstevel@tonic-gate  *			STORE	CPU_THREAD = (possibly) someone else
1187c478bd9Sstevel@tonic-gate  *
1197c478bd9Sstevel@tonic-gate  * (3M) mutex_vector_enter() memory order:
1207c478bd9Sstevel@tonic-gate  *
1217c478bd9Sstevel@tonic-gate  *			STORE	waiters bit = 1
1227c478bd9Sstevel@tonic-gate  *			LOAD	CPU_THREAD for each CPU
1237c478bd9Sstevel@tonic-gate  *			LOAD	owner and waiters bit
1247c478bd9Sstevel@tonic-gate  *
1257c478bd9Sstevel@tonic-gate  * It has been verified by exhaustive simulation that all possible global
1267c478bd9Sstevel@tonic-gate  * memory orderings of (2M) interleaved with (3M) result in correct
1277c478bd9Sstevel@tonic-gate  * behavior.  Moreover, these ordering constraints are minimal: changing
1287c478bd9Sstevel@tonic-gate  * the ordering of anything in (2M) or (3M) breaks the algorithm, creating
1297c478bd9Sstevel@tonic-gate  * windows for missed wakeups.  Note: the possibility that other threads
1307c478bd9Sstevel@tonic-gate  * may grab the lock after the owner drops it can be factored out of the
1317c478bd9Sstevel@tonic-gate  * memory ordering analysis because mutex_vector_enter() won't block
1327c478bd9Sstevel@tonic-gate  * if the lock isn't still owned by the same thread.
1337c478bd9Sstevel@tonic-gate  *
1347c478bd9Sstevel@tonic-gate  * The only requirements of code outside the mutex implementation are
1357c478bd9Sstevel@tonic-gate  * (1) mutex_exit() preemption fixup in interrupt handlers or trap return,
136575a7426Spt  * (2) a membar #StoreLoad after setting CPU_THREAD in resume(),
137575a7426Spt  * (3) mutex_owner_running() preemption fixup in interrupt handlers
138575a7426Spt  * or trap returns.
1397c478bd9Sstevel@tonic-gate  * Note: idle threads cannot grab adaptive locks (since they cannot block),
1407c478bd9Sstevel@tonic-gate  * so the membar may be safely omitted when resuming an idle thread.
1417c478bd9Sstevel@tonic-gate  *
1427c478bd9Sstevel@tonic-gate  * When a mutex has waiters, mutex_vector_exit() has several options:
1437c478bd9Sstevel@tonic-gate  *
1447c478bd9Sstevel@tonic-gate  * (1) Choose a waiter and make that thread the owner before waking it;
1457c478bd9Sstevel@tonic-gate  *     this is known as "direct handoff" of ownership.
1467c478bd9Sstevel@tonic-gate  *
1477c478bd9Sstevel@tonic-gate  * (2) Drop the lock and wake one waiter.
1487c478bd9Sstevel@tonic-gate  *
1497c478bd9Sstevel@tonic-gate  * (3) Drop the lock, clear the waiters bit, and wake all waiters.
1507c478bd9Sstevel@tonic-gate  *
1517c478bd9Sstevel@tonic-gate  * In many ways (1) is the cleanest solution, but if a lock is moderately
1527c478bd9Sstevel@tonic-gate  * contended it defeats the adaptive spin logic.  If we make some other
153*48bbca81SDaniel Hoffman  * thread the owner, but it's not ONPROC yet, then all other threads on
1547c478bd9Sstevel@tonic-gate  * other cpus that try to get the lock will conclude that the owner is
1557c478bd9Sstevel@tonic-gate  * blocked, so they'll block too.  And so on -- it escalates quickly,
1567c478bd9Sstevel@tonic-gate  * with every thread taking the blocking path rather than the spin path.
1577c478bd9Sstevel@tonic-gate  * Thus, direct handoff is *not* a good idea for adaptive mutexes.
1587c478bd9Sstevel@tonic-gate  *
1597c478bd9Sstevel@tonic-gate  * Option (2) is the next most natural-seeming option, but it has several
1607c478bd9Sstevel@tonic-gate  * annoying properties.  If there's more than one waiter, we must preserve
1617c478bd9Sstevel@tonic-gate  * the waiters bit on an unheld lock.  On cas-capable platforms, where
1627c478bd9Sstevel@tonic-gate  * the waiters bit is part of the lock word, this means that both 0x0
1637c478bd9Sstevel@tonic-gate  * and 0x1 represent unheld locks, so we have to cas against *both*.
1647c478bd9Sstevel@tonic-gate  * Priority inheritance also gets more complicated, because a lock can
1657c478bd9Sstevel@tonic-gate  * have waiters but no owner to whom priority can be willed.  So while
1667c478bd9Sstevel@tonic-gate  * it is possible to make option (2) work, it's surprisingly vile.
1677c478bd9Sstevel@tonic-gate  *
1687c478bd9Sstevel@tonic-gate  * Option (3), the least-intuitive at first glance, is what we actually do.
1697c478bd9Sstevel@tonic-gate  * It has the advantage that because you always wake all waiters, you
1707c478bd9Sstevel@tonic-gate  * never have to preserve the waiters bit.  Waking all waiters seems like
1717c478bd9Sstevel@tonic-gate  * begging for a thundering herd problem, but consider: under option (2),
1727c478bd9Sstevel@tonic-gate  * every thread that grabs and drops the lock will wake one waiter -- so
1737c478bd9Sstevel@tonic-gate  * if the lock is fairly active, all waiters will be awakened very quickly
1747c478bd9Sstevel@tonic-gate  * anyway.  Moreover, this is how adaptive locks are *supposed* to work.
1757c478bd9Sstevel@tonic-gate  * The blocking case is rare; the more common case (by 3-4 orders of
1767c478bd9Sstevel@tonic-gate  * magnitude) is that one or more threads spin waiting to get the lock.
1777c478bd9Sstevel@tonic-gate  * Only direct handoff can prevent the thundering herd problem, but as
1787c478bd9Sstevel@tonic-gate  * mentioned earlier, that would tend to defeat the adaptive spin logic.
1797c478bd9Sstevel@tonic-gate  * In practice, option (3) works well because the blocking case is rare.
1807c478bd9Sstevel@tonic-gate  */
1827c478bd9Sstevel@tonic-gate /*
1837c478bd9Sstevel@tonic-gate  * delayed lock retry with exponential delay for spin locks
1847c478bd9Sstevel@tonic-gate  *
1857c478bd9Sstevel@tonic-gate  * It is noted above that for both the spin locks and the adaptive locks,
1867c478bd9Sstevel@tonic-gate  * spinning is the dominate mode of operation.  So long as there is only
1877c478bd9Sstevel@tonic-gate  * one thread waiting on a lock, the naive spin loop works very well in
1887c478bd9Sstevel@tonic-gate  * cache based architectures.  The lock data structure is pulled into the
1897c478bd9Sstevel@tonic-gate  * cache of the processor with the waiting/spinning thread and no further
1907c478bd9Sstevel@tonic-gate  * memory traffic is generated until the lock is released.  Unfortunately,
1917c478bd9Sstevel@tonic-gate  * once two or more threads are waiting on a lock, the naive spin has
1927c478bd9Sstevel@tonic-gate  * the property of generating maximum memory traffic from each spinning
1937c478bd9Sstevel@tonic-gate  * thread as the spinning threads contend for the lock data structure.
1947c478bd9Sstevel@tonic-gate  *
1957c478bd9Sstevel@tonic-gate  * By executing a delay loop before retrying a lock, a waiting thread
1967c478bd9Sstevel@tonic-gate  * can reduce its memory traffic by a large factor, depending on the
1977c478bd9Sstevel@tonic-gate  * size of the delay loop.  A large delay loop greatly reduced the memory
1987c478bd9Sstevel@tonic-gate  * traffic, but has the drawback of having a period of time when
1997c478bd9Sstevel@tonic-gate  * no thread is attempting to gain the lock even though several threads
2007c478bd9Sstevel@tonic-gate  * might be waiting.  A small delay loop has the drawback of not
2017c478bd9Sstevel@tonic-gate  * much reduction in memory traffic, but reduces the potential idle time.
2027c478bd9Sstevel@tonic-gate  * The theory of the exponential delay code is to start with a short
2037c478bd9Sstevel@tonic-gate  * delay loop and double the waiting time on each iteration, up to
204575a7426Spt  * a preselected maximum.
2057c478bd9Sstevel@tonic-gate  */
2077c478bd9Sstevel@tonic-gate #include <sys/param.h>
2087c478bd9Sstevel@tonic-gate #include <sys/time.h>
2097c478bd9Sstevel@tonic-gate #include <sys/cpuvar.h>
2107c478bd9Sstevel@tonic-gate #include <sys/thread.h>
2117c478bd9Sstevel@tonic-gate #include <sys/debug.h>
2127c478bd9Sstevel@tonic-gate #include <sys/cmn_err.h>
2137c478bd9Sstevel@tonic-gate #include <sys/sobject.h>
2147c478bd9Sstevel@tonic-gate #include <sys/turnstile.h>
2157c478bd9Sstevel@tonic-gate #include <sys/systm.h>
2167c478bd9Sstevel@tonic-gate #include <sys/mutex_impl.h>
2177c478bd9Sstevel@tonic-gate #include <sys/spl.h>
2187c478bd9Sstevel@tonic-gate #include <sys/lockstat.h>
2197c478bd9Sstevel@tonic-gate #include <sys/atomic.h>
2207c478bd9Sstevel@tonic-gate #include <sys/cpu.h>
2217c478bd9Sstevel@tonic-gate #include <sys/stack.h>
222843e1988Sjohnlev #include <sys/archsystm.h>
223575a7426Spt #include <sys/machsystm.h>
224575a7426Spt #include <sys/x_call.h>
2267c478bd9Sstevel@tonic-gate /*
2277c478bd9Sstevel@tonic-gate  * The sobj_ops vector exports a set of functions needed when a thread
2287c478bd9Sstevel@tonic-gate  * is asleep on a synchronization object of this type.
2297c478bd9Sstevel@tonic-gate  */
2307c478bd9Sstevel@tonic-gate static sobj_ops_t mutex_sobj_ops = {
2317c478bd9Sstevel@tonic-gate 	SOBJ_MUTEX, mutex_owner, turnstile_stay_asleep, turnstile_change_pri
2327c478bd9Sstevel@tonic-gate };
2347c478bd9Sstevel@tonic-gate /*
2357c478bd9Sstevel@tonic-gate  * If the system panics on a mutex, save the address of the offending
2367c478bd9Sstevel@tonic-gate  * mutex in panic_mutex_addr, and save the contents in panic_mutex.
2377c478bd9Sstevel@tonic-gate  */
2387c478bd9Sstevel@tonic-gate static mutex_impl_t panic_mutex;
2397c478bd9Sstevel@tonic-gate static mutex_impl_t *panic_mutex_addr;
2417c478bd9Sstevel@tonic-gate static void
mutex_panic(char * msg,mutex_impl_t * lp)2427c478bd9Sstevel@tonic-gate mutex_panic(char *msg, mutex_impl_t *lp)
2437c478bd9Sstevel@tonic-gate {
2447c478bd9Sstevel@tonic-gate 	if (panicstr)
2457c478bd9Sstevel@tonic-gate 		return;
24775d94465SJosef 'Jeff' Sipek 	if (atomic_cas_ptr(&panic_mutex_addr, NULL, lp) == NULL)
2487c478bd9Sstevel@tonic-gate 		panic_mutex = *lp;
2507c478bd9Sstevel@tonic-gate 	panic("%s, lp=%p owner=%p thread=%p",
2518793b36bSNick Todd 	    msg, (void *)lp, (void *)MUTEX_OWNER(&panic_mutex),
2528793b36bSNick Todd 	    (void *)curthread);
2537c478bd9Sstevel@tonic-gate }
255575a7426Spt /* "tunables" for per-platform backoff constants. */
256575a7426Spt uint_t mutex_backoff_cap = 0;
257575a7426Spt ushort_t mutex_backoff_base = MUTEX_BACKOFF_BASE;
258575a7426Spt ushort_t mutex_cap_factor = MUTEX_CAP_FACTOR;
259575a7426Spt uchar_t mutex_backoff_shift = MUTEX_BACKOFF_SHIFT;
261575a7426Spt void
mutex_sync(void)262575a7426Spt mutex_sync(void)
263575a7426Spt {
264575a7426Spt 	MUTEX_SYNC();
265575a7426Spt }
267575a7426Spt /* calculate the backoff interval */
268374ae87fSsvemuri uint_t
default_lock_backoff(uint_t backoff)269575a7426Spt default_lock_backoff(uint_t backoff)
270575a7426Spt {
271575a7426Spt 	uint_t cap;		/* backoff cap calculated */
273575a7426Spt 	if (backoff == 0) {
274575a7426Spt 		backoff = mutex_backoff_base;
275575a7426Spt 		/* first call just sets the base */
276575a7426Spt 		return (backoff);
277575a7426Spt 	}
279575a7426Spt 	/* set cap */
280575a7426Spt 	if (mutex_backoff_cap == 0) {
281575a7426Spt 		/*
282575a7426Spt 		 * For a contended lock, in the worst case a load + cas may
283575a7426Spt 		 * be queued  at the controller for each contending CPU.
284575a7426Spt 		 * Therefore, to avoid queueing, the accesses for all CPUS must
285575a7426Spt 		 * be spread out in time over an interval of (ncpu *
286575a7426Spt 		 * cap-factor).  Maximum backoff is set to this value, and
287575a7426Spt 		 * actual backoff is a random number from 0 to the current max.
288575a7426Spt 		 */
289575a7426Spt 		cap = ncpus_online * mutex_cap_factor;
290575a7426Spt 	} else {
291575a7426Spt 		cap = mutex_backoff_cap;
292575a7426Spt 	}
294575a7426Spt 	/* calculate new backoff value */
295575a7426Spt 	backoff <<= mutex_backoff_shift;	/* increase backoff */
296575a7426Spt 	if (backoff > cap) {
297575a7426Spt 		if (cap < mutex_backoff_base)
298575a7426Spt 			backoff = mutex_backoff_base;
299575a7426Spt 		else
300575a7426Spt 			backoff = cap;
301575a7426Spt 	}
303575a7426Spt 	return (backoff);
304575a7426Spt }
306575a7426Spt /*
307575a7426Spt  * default delay function for mutexes.
308575a7426Spt  */
309374ae87fSsvemuri void
default_lock_delay(uint_t backoff)310575a7426Spt default_lock_delay(uint_t backoff)
311575a7426Spt {
312575a7426Spt 	ulong_t rnd;		/* random factor */
313575a7426Spt 	uint_t cur_backoff;	/* calculated backoff */
314575a7426Spt 	uint_t backctr;
316575a7426Spt 	/*
317575a7426Spt 	 * Modify backoff by a random amount to avoid lockstep, and to
318575a7426Spt 	 * make it probable that some thread gets a small backoff, and
319575a7426Spt 	 * re-checks quickly
320575a7426Spt 	 */
321575a7426Spt 	rnd = (((long)curthread >> PTR24_LSB) ^ (long)MUTEX_GETTICK());
322575a7426Spt 	cur_backoff = (uint_t)(rnd % (backoff - mutex_backoff_base + 1)) +
323575a7426Spt 	    mutex_backoff_base;
325575a7426Spt 	/*
326575a7426Spt 	 * Delay before trying
327575a7426Spt 	 * to touch the mutex data structure.
328575a7426Spt 	 */
329575a7426Spt 	for (backctr = cur_backoff; backctr; backctr--) {
330575a7426Spt 		MUTEX_DELAY();
331575a7426Spt 	};
332575a7426Spt }
334575a7426Spt uint_t (*mutex_lock_backoff)(uint_t) = default_lock_backoff;
335575a7426Spt void (*mutex_lock_delay)(uint_t) = default_lock_delay;
336575a7426Spt void (*mutex_delay)(void) = mutex_delay_default;
3387c478bd9Sstevel@tonic-gate /*
3397c478bd9Sstevel@tonic-gate  * mutex_vector_enter() is called from the assembly mutex_enter() routine
3407c478bd9Sstevel@tonic-gate  * if the lock is held or is not of type MUTEX_ADAPTIVE.
3417c478bd9Sstevel@tonic-gate  */
3427c478bd9Sstevel@tonic-gate void
mutex_vector_enter(mutex_impl_t * lp)3437c478bd9Sstevel@tonic-gate mutex_vector_enter(mutex_impl_t *lp)
3447c478bd9Sstevel@tonic-gate {
3457c478bd9Sstevel@tonic-gate 	kthread_id_t	owner;
346575a7426Spt 	kthread_id_t	lastowner = MUTEX_NO_OWNER; /* track owner changes */
3477c478bd9Sstevel@tonic-gate 	hrtime_t	sleep_time = 0;	/* how long we slept */
3489d68b18eSck 	hrtime_t	spin_time = 0;	/* how long we spun */
349575a7426Spt 	cpu_t 		*cpup;
3507c478bd9Sstevel@tonic-gate 	turnstile_t	*ts;
3517c478bd9Sstevel@tonic-gate 	volatile mutex_impl_t *vlp = (volatile mutex_impl_t *)lp;
352575a7426Spt 	uint_t		backoff = 0;	/* current backoff */
353575a7426Spt 	int		changecnt = 0;	/* count of owner changes */
3557c478bd9Sstevel@tonic-gate 	ASSERT_STACK_ALIGNED();
3577c478bd9Sstevel@tonic-gate 	if (MUTEX_TYPE_SPIN(lp)) {
3587c478bd9Sstevel@tonic-gate 		lock_set_spl(&lp->m_spin.m_spinlock, lp->m_spin.m_minspl,
3597c478bd9Sstevel@tonic-gate 		    &lp->m_spin.m_oldspl);
3607c478bd9Sstevel@tonic-gate 		return;
3617c478bd9Sstevel@tonic-gate 	}
3637c478bd9Sstevel@tonic-gate 	if (!MUTEX_TYPE_ADAPTIVE(lp)) {
3647c478bd9Sstevel@tonic-gate 		mutex_panic("mutex_enter: bad mutex", lp);
3657c478bd9Sstevel@tonic-gate 		return;
3667c478bd9Sstevel@tonic-gate 	}
3687c478bd9Sstevel@tonic-gate 	/*
3697c478bd9Sstevel@tonic-gate 	 * Adaptive mutexes must not be acquired from above LOCK_LEVEL.
3707c478bd9Sstevel@tonic-gate 	 * We can migrate after loading CPU but before checking CPU_ON_INTR,
3717c478bd9Sstevel@tonic-gate 	 * so we must verify by disabling preemption and loading CPU again.
3727c478bd9Sstevel@tonic-gate 	 */
3737c478bd9Sstevel@tonic-gate 	cpup = CPU;
3747c478bd9Sstevel@tonic-gate 	if (CPU_ON_INTR(cpup) && !panicstr) {
3757c478bd9Sstevel@tonic-gate 		kpreempt_disable();
3767c478bd9Sstevel@tonic-gate 		if (CPU_ON_INTR(CPU))
3777c478bd9Sstevel@tonic-gate 			mutex_panic("mutex_enter: adaptive at high PIL", lp);
3787c478bd9Sstevel@tonic-gate 		kpreempt_enable();
3797c478bd9Sstevel@tonic-gate 	}
3817c478bd9Sstevel@tonic-gate 	CPU_STATS_ADDQ(cpup, sys, mutex_adenters, 1);
3839d68b18eSck 	spin_time = LOCKSTAT_START_TIME(LS_MUTEX_ENTER_SPIN);
385575a7426Spt 	backoff = mutex_lock_backoff(0);	/* set base backoff */
3867c478bd9Sstevel@tonic-gate 	for (;;) {
387575a7426Spt 		mutex_lock_delay(backoff); /* backoff delay */
3897c478bd9Sstevel@tonic-gate 		if (panicstr)
3907c478bd9Sstevel@tonic-gate 			return;
3927c478bd9Sstevel@tonic-gate 		if ((owner = MUTEX_OWNER(vlp)) == NULL) {
393575a7426Spt 			if (mutex_adaptive_tryenter(lp)) {
3947c478bd9Sstevel@tonic-gate 				break;
395575a7426Spt 			}
396575a7426Spt 			/* increase backoff only on failed attempt. */
397575a7426Spt 			backoff = mutex_lock_backoff(backoff);
398575a7426Spt 			changecnt++;
3997c478bd9Sstevel@tonic-gate 			continue;
400575a7426Spt 		} else if (lastowner != owner) {
401575a7426Spt 			lastowner = owner;
402575a7426Spt 			backoff = mutex_lock_backoff(backoff);
403575a7426Spt 			changecnt++;
404575a7426Spt 		}
406575a7426Spt 		if (changecnt >= ncpus_online) {
407575a7426Spt 			backoff = mutex_lock_backoff(0);
408575a7426Spt 			changecnt = 0;
4097c478bd9Sstevel@tonic-gate 		}
4117c478bd9Sstevel@tonic-gate 		if (owner == curthread)
4127c478bd9Sstevel@tonic-gate 			mutex_panic("recursive mutex_enter", lp);
4147c478bd9Sstevel@tonic-gate 		/*
4157c478bd9Sstevel@tonic-gate 		 * If lock is held but owner is not yet set, spin.
4167c478bd9Sstevel@tonic-gate 		 * (Only relevant for platforms that don't have cas.)
4177c478bd9Sstevel@tonic-gate 		 */
4187c478bd9Sstevel@tonic-gate 		if (owner == MUTEX_NO_OWNER)
4197c478bd9Sstevel@tonic-gate 			continue;
421575a7426Spt 		if (mutex_owner_running(lp) != NULL)  {
422575a7426Spt 			continue;
423575a7426Spt 		}
4257c478bd9Sstevel@tonic-gate 		/*
4267c478bd9Sstevel@tonic-gate 		 * The owner appears not to be running, so block.
4277c478bd9Sstevel@tonic-gate 		 * See the Big Theory Statement for memory ordering issues.
4287c478bd9Sstevel@tonic-gate 		 */
4297c478bd9Sstevel@tonic-gate 		ts = turnstile_lookup(lp);
4307c478bd9Sstevel@tonic-gate 		MUTEX_SET_WAITERS(lp);
4317c478bd9Sstevel@tonic-gate 		membar_enter();
4337c478bd9Sstevel@tonic-gate 		/*
4347c478bd9Sstevel@tonic-gate 		 * Recheck whether owner is running after waiters bit hits
4357c478bd9Sstevel@tonic-gate 		 * global visibility (above).  If owner is running, spin.
4367c478bd9Sstevel@tonic-gate 		 */
437575a7426Spt 		if (mutex_owner_running(lp) != NULL) {
438575a7426Spt 			turnstile_exit(lp);
439575a7426Spt 			continue;
440575a7426Spt 		}
4417c478bd9Sstevel@tonic-gate 		membar_consumer();
4437c478bd9Sstevel@tonic-gate 		/*
4447c478bd9Sstevel@tonic-gate 		 * If owner and waiters bit are unchanged, block.
4457c478bd9Sstevel@tonic-gate 		 */
4467c478bd9Sstevel@tonic-gate 		if (MUTEX_OWNER(vlp) == owner && MUTEX_HAS_WAITERS(vlp)) {
4477c478bd9Sstevel@tonic-gate 			sleep_time -= gethrtime();
4487c478bd9Sstevel@tonic-gate 			(void) turnstile_block(ts, TS_WRITER_Q, lp,
4497c478bd9Sstevel@tonic-gate 			    &mutex_sobj_ops, NULL, NULL);
4507c478bd9Sstevel@tonic-gate 			sleep_time += gethrtime();
451575a7426Spt 			/* reset backoff after turnstile */
452575a7426Spt 			backoff = mutex_lock_backoff(0);
4537c478bd9Sstevel@tonic-gate 		} else {
4547c478bd9Sstevel@tonic-gate 			turnstile_exit(lp);
4557c478bd9Sstevel@tonic-gate 		}
4567c478bd9Sstevel@tonic-gate 	}
4587c478bd9Sstevel@tonic-gate 	ASSERT(MUTEX_OWNER(lp) == curthread);
4600efe5e54Sdv 	if (sleep_time != 0) {
4610efe5e54Sdv 		/*
4620efe5e54Sdv 		 * Note, sleep time is the sum of all the sleeping we
4630efe5e54Sdv 		 * did.
4640efe5e54Sdv 		 */
4657c478bd9Sstevel@tonic-gate 		LOCKSTAT_RECORD(LS_MUTEX_ENTER_BLOCK, lp, sleep_time);
4667c478bd9Sstevel@tonic-gate 	}
4689d68b18eSck 	/* record spin time, don't count sleep time */
4699d68b18eSck 	if (spin_time != 0) {
4719d68b18eSck 		    spin_time + sleep_time);
472575a7426Spt 	}
4747c478bd9Sstevel@tonic-gate 	LOCKSTAT_RECORD0(LS_MUTEX_ENTER_ACQUIRE, lp);
4757c478bd9Sstevel@tonic-gate }
4777c478bd9Sstevel@tonic-gate /*
4787c478bd9Sstevel@tonic-gate  * mutex_vector_tryenter() is called from the assembly mutex_tryenter()
4797c478bd9Sstevel@tonic-gate  * routine if the lock is held or is not of type MUTEX_ADAPTIVE.
4807c478bd9Sstevel@tonic-gate  */
4817c478bd9Sstevel@tonic-gate int
mutex_vector_tryenter(mutex_impl_t * lp)4827c478bd9Sstevel@tonic-gate mutex_vector_tryenter(mutex_impl_t *lp)
4837c478bd9Sstevel@tonic-gate {
4847c478bd9Sstevel@tonic-gate 	int s;
4867c478bd9Sstevel@tonic-gate 	if (MUTEX_TYPE_ADAPTIVE(lp))
4877c478bd9Sstevel@tonic-gate 		return (0);		/* we already tried in assembly */
4897c478bd9Sstevel@tonic-gate 	if (!MUTEX_TYPE_SPIN(lp)) {
4907c478bd9Sstevel@tonic-gate 		mutex_panic("mutex_tryenter: bad mutex", lp);
4917c478bd9Sstevel@tonic-gate 		return (0);
4927c478bd9Sstevel@tonic-gate 	}
4947c478bd9Sstevel@tonic-gate 	s = splr(lp->m_spin.m_minspl);
4957c478bd9Sstevel@tonic-gate 	if (lock_try(&lp->m_spin.m_spinlock)) {
4967c478bd9Sstevel@tonic-gate 		lp->m_spin.m_oldspl = (ushort_t)s;
4977c478bd9Sstevel@tonic-gate 		return (1);
4987c478bd9Sstevel@tonic-gate 	}
4997c478bd9Sstevel@tonic-gate 	splx(s);
5007c478bd9Sstevel@tonic-gate 	return (0);
5017c478bd9Sstevel@tonic-gate }
5037c478bd9Sstevel@tonic-gate /*
5047c478bd9Sstevel@tonic-gate  * mutex_vector_exit() is called from mutex_exit() if the lock is not
5057c478bd9Sstevel@tonic-gate  * adaptive, has waiters, or is not owned by the current thread (panic).
5067c478bd9Sstevel@tonic-gate  */
5077c478bd9Sstevel@tonic-gate void
mutex_vector_exit(mutex_impl_t * lp)5087c478bd9Sstevel@tonic-gate mutex_vector_exit(mutex_impl_t *lp)
5097c478bd9Sstevel@tonic-gate {
5107c478bd9Sstevel@tonic-gate 	turnstile_t *ts;
5127c478bd9Sstevel@tonic-gate 	if (MUTEX_TYPE_SPIN(lp)) {
5137c478bd9Sstevel@tonic-gate 		lock_clear_splx(&lp->m_spin.m_spinlock, lp->m_spin.m_oldspl);
5147c478bd9Sstevel@tonic-gate 		return;
5157c478bd9Sstevel@tonic-gate 	}
5177c478bd9Sstevel@tonic-gate 	if (MUTEX_OWNER(lp) != curthread) {
5187c478bd9Sstevel@tonic-gate 		mutex_panic("mutex_exit: not owner", lp);
5197c478bd9Sstevel@tonic-gate 		return;
5207c478bd9Sstevel@tonic-gate 	}
5227c478bd9Sstevel@tonic-gate 	ts = turnstile_lookup(lp);
5237c478bd9Sstevel@tonic-gate 	MUTEX_CLEAR_LOCK_AND_WAITERS(lp);
5247c478bd9Sstevel@tonic-gate 	if (ts == NULL)
5257c478bd9Sstevel@tonic-gate 		turnstile_exit(lp);
5267c478bd9Sstevel@tonic-gate 	else
5277c478bd9Sstevel@tonic-gate 		turnstile_wakeup(ts, TS_WRITER_Q, ts->ts_waiters, NULL);
5287c478bd9Sstevel@tonic-gate 	LOCKSTAT_RECORD0(LS_MUTEX_EXIT_RELEASE, lp);
5297c478bd9Sstevel@tonic-gate }
5317c478bd9Sstevel@tonic-gate int
mutex_owned(const kmutex_t * mp)532b5fca8f8Stomee mutex_owned(const kmutex_t *mp)
5337c478bd9Sstevel@tonic-gate {
534b5fca8f8Stomee 	const mutex_impl_t *lp = (const mutex_impl_t *)mp;
53619397407SSherry Moore 	if (panicstr || quiesce_active)
5377c478bd9Sstevel@tonic-gate 		return (1);
5397c478bd9Sstevel@tonic-gate 	if (MUTEX_TYPE_ADAPTIVE(lp))
5407c478bd9Sstevel@tonic-gate 		return (MUTEX_OWNER(lp) == curthread);
5417c478bd9Sstevel@tonic-gate 	return (LOCK_HELD(&lp->m_spin.m_spinlock));
5427c478bd9Sstevel@tonic-gate }
5447c478bd9Sstevel@tonic-gate kthread_t *
mutex_owner(const kmutex_t * mp)545b5fca8f8Stomee mutex_owner(const kmutex_t *mp)
5467c478bd9Sstevel@tonic-gate {
547b5fca8f8Stomee 	const mutex_impl_t *lp = (const mutex_impl_t *)mp;
5487c478bd9Sstevel@tonic-gate 	kthread_id_t t;
5507c478bd9Sstevel@tonic-gate 	if (MUTEX_TYPE_ADAPTIVE(lp) && (t = MUTEX_OWNER(lp)) != MUTEX_NO_OWNER)
5517c478bd9Sstevel@tonic-gate 		return (t);
5527c478bd9Sstevel@tonic-gate 	return (NULL);
5537c478bd9Sstevel@tonic-gate }
5557c478bd9Sstevel@tonic-gate /*
5567c478bd9Sstevel@tonic-gate  * The iblock cookie 'ibc' is the spl level associated with the lock;
5577c478bd9Sstevel@tonic-gate  * this alone determines whether the lock will be ADAPTIVE or SPIN.
5587c478bd9Sstevel@tonic-gate  *
5597c478bd9Sstevel@tonic-gate  * Adaptive mutexes created in zeroed memory do not need to call
5607c478bd9Sstevel@tonic-gate  * mutex_init() as their allocation in this fashion guarantees
5617c478bd9Sstevel@tonic-gate  * their initialization.
5627c478bd9Sstevel@tonic-gate  *   eg adaptive mutexes created as static within the BSS or allocated
5637c478bd9Sstevel@tonic-gate  *      by kmem_zalloc().
5647c478bd9Sstevel@tonic-gate  */
5657c478bd9Sstevel@tonic-gate /* ARGSUSED */
5667c478bd9Sstevel@tonic-gate void
mutex_init(kmutex_t * mp,char * name,kmutex_type_t type,void * ibc)5677c478bd9Sstevel@tonic-gate mutex_init(kmutex_t *mp, char *name, kmutex_type_t type, void *ibc)
5687c478bd9Sstevel@tonic-gate {
5697c478bd9Sstevel@tonic-gate 	mutex_impl_t *lp = (mutex_impl_t *)mp;
5717c478bd9Sstevel@tonic-gate 	ASSERT(ibc < (void *)KERNELBASE);	/* see 1215173 */
5737c478bd9Sstevel@tonic-gate 	if ((intptr_t)ibc > ipltospl(LOCK_LEVEL) && ibc < (void *)KERNELBASE) {
5747c478bd9Sstevel@tonic-gate 		ASSERT(type != MUTEX_ADAPTIVE && type != MUTEX_DEFAULT);
5757c478bd9Sstevel@tonic-gate 		MUTEX_SET_TYPE(lp, MUTEX_SPIN);
5767c478bd9Sstevel@tonic-gate 		LOCK_INIT_CLEAR(&lp->m_spin.m_spinlock);
5777c478bd9Sstevel@tonic-gate 		LOCK_INIT_HELD(&lp->m_spin.m_dummylock);
5787c478bd9Sstevel@tonic-gate 		lp->m_spin.m_minspl = (int)(intptr_t)ibc;
5797c478bd9Sstevel@tonic-gate 	} else {
5807f30f491Sck #ifdef MUTEX_ALIGN
5817f30f491Sck 		static int misalign_cnt = 0;
5837f30f491Sck 		if (((uintptr_t)lp & (uintptr_t)(MUTEX_ALIGN - 1)) &&
5847f30f491Sck 		    (misalign_cnt < MUTEX_ALIGN_WARNINGS)) {
5857f30f491Sck 			/*
5867f30f491Sck 			 * The mutex is not aligned and may cross a cache line.
5877f30f491Sck 			 * This is not supported and may cause a panic.
5887f30f491Sck 			 * Show a warning that the mutex is not aligned
5897f30f491Sck 			 * and attempt to identify the origin.
5907f30f491Sck 			 * Unaligned mutexes are not (supposed to be)
5917f30f491Sck 			 * possible on SPARC.
5927f30f491Sck 			 */
5937f30f491Sck 			char *funcname;
5947f30f491Sck 			ulong_t offset = 0;
5967f30f491Sck 			funcname = modgetsymname((uintptr_t)caller(), &offset);
5977f30f491Sck 			cmn_err(CE_WARN, "mutex_init: %p is not %d byte "
5987f30f491Sck 			    "aligned; caller %s+%lx in module %s. "
5997f30f491Sck 			    "This is unsupported and may cause a panic. "
6007f30f491Sck 			    "Please report this to the kernel module supplier.",
6011f7e2746Sck 			    (void *)lp, MUTEX_ALIGN,
6027f30f491Sck 			    funcname ? funcname : "unknown", offset,
6037f30f491Sck 			    mod_containing_pc(caller()));
6047f30f491Sck 			misalign_cnt++;
6057f30f491Sck 			if (misalign_cnt >= MUTEX_ALIGN_WARNINGS) {
6067f30f491Sck 				cmn_err(CE_WARN, "mutex_init: further unaligned"
6077f30f491Sck 				    " mutex warnings will be suppressed.");
6087f30f491Sck 			}
6097f30f491Sck 		}
6107f30f491Sck #endif	/* MUTEX_ALIGN */
6117c478bd9Sstevel@tonic-gate 		ASSERT(type != MUTEX_SPIN);
6137c478bd9Sstevel@tonic-gate 		MUTEX_SET_TYPE(lp, MUTEX_ADAPTIVE);
6147c478bd9Sstevel@tonic-gate 		MUTEX_CLEAR_LOCK_AND_WAITERS(lp);
6157c478bd9Sstevel@tonic-gate 	}
6167c478bd9Sstevel@tonic-gate }
6187c478bd9Sstevel@tonic-gate void
mutex_destroy(kmutex_t * mp)6197c478bd9Sstevel@tonic-gate mutex_destroy(kmutex_t *mp)
6207c478bd9Sstevel@tonic-gate {
6217c478bd9Sstevel@tonic-gate 	mutex_impl_t *lp = (mutex_impl_t *)mp;
6237c478bd9Sstevel@tonic-gate 	if (lp->m_owner == 0 && !MUTEX_HAS_WAITERS(lp)) {
6247c478bd9Sstevel@tonic-gate 		MUTEX_DESTROY(lp);
6257c478bd9Sstevel@tonic-gate 	} else if (MUTEX_TYPE_SPIN(lp)) {
6267c478bd9Sstevel@tonic-gate 		LOCKSTAT_RECORD0(LS_MUTEX_DESTROY_RELEASE, lp);
6277c478bd9Sstevel@tonic-gate 		MUTEX_DESTROY(lp);
6287c478bd9Sstevel@tonic-gate 	} else if (MUTEX_TYPE_ADAPTIVE(lp)) {
6297c478bd9Sstevel@tonic-gate 		LOCKSTAT_RECORD0(LS_MUTEX_DESTROY_RELEASE, lp);
6307c478bd9Sstevel@tonic-gate 		if (MUTEX_OWNER(lp) != curthread)
6317c478bd9Sstevel@tonic-gate 			mutex_panic("mutex_destroy: not owner", lp);
6327c478bd9Sstevel@tonic-gate 		if (MUTEX_HAS_WAITERS(lp)) {
6337c478bd9Sstevel@tonic-gate 			turnstile_t *ts = turnstile_lookup(lp);
6347c478bd9Sstevel@tonic-gate 			turnstile_exit(lp);
6357c478bd9Sstevel@tonic-gate 			if (ts != NULL)
6367c478bd9Sstevel@tonic-gate 				mutex_panic("mutex_destroy: has waiters", lp);
6377c478bd9Sstevel@tonic-gate 		}
6387c478bd9Sstevel@tonic-gate 		MUTEX_DESTROY(lp);
6397c478bd9Sstevel@tonic-gate 	} else {
6407c478bd9Sstevel@tonic-gate 		mutex_panic("mutex_destroy: bad mutex", lp);
6417c478bd9Sstevel@tonic-gate 	}
6427c478bd9Sstevel@tonic-gate }
6447c478bd9Sstevel@tonic-gate /*
6457c478bd9Sstevel@tonic-gate  * Simple C support for the cases where spin locks miss on the first try.
6467c478bd9Sstevel@tonic-gate  */
6477c478bd9Sstevel@tonic-gate void
lock_set_spin(lock_t * lp)6487c478bd9Sstevel@tonic-gate lock_set_spin(lock_t *lp)
6497c478bd9Sstevel@tonic-gate {
650575a7426Spt 	int loop_count = 0;
651575a7426Spt 	uint_t backoff = 0;	/* current backoff */
6529d68b18eSck 	hrtime_t spin_time = 0;	/* how long we spun */
6547c478bd9Sstevel@tonic-gate 	if (panicstr)
6557c478bd9Sstevel@tonic-gate 		return;
6577c478bd9Sstevel@tonic-gate 	if (ncpus == 1)
6588793b36bSNick Todd 		panic("lock_set: %p lock held and only one CPU", (void *)lp);
6609d68b18eSck 	spin_time = LOCKSTAT_START_TIME(LS_LOCK_SET_SPIN);
6627c478bd9Sstevel@tonic-gate 	while (LOCK_HELD(lp) || !lock_spin_try(lp)) {
6637c478bd9Sstevel@tonic-gate 		if (panicstr)
6647c478bd9Sstevel@tonic-gate 			return;
665575a7426Spt 		loop_count++;
667575a7426Spt 		if (ncpus_online == loop_count) {
668575a7426Spt 			backoff = mutex_lock_backoff(0);
669575a7426Spt 			loop_count = 0;
670575a7426Spt 		} else {
671575a7426Spt 			backoff = mutex_lock_backoff(backoff);
6727c478bd9Sstevel@tonic-gate 		}
673575a7426Spt 		mutex_lock_delay(backoff);
6747c478bd9Sstevel@tonic-gate 	}
6769d68b18eSck 	LOCKSTAT_RECORD_TIME(LS_LOCK_SET_SPIN, lp, spin_time);
6787c478bd9Sstevel@tonic-gate 	LOCKSTAT_RECORD0(LS_LOCK_SET_ACQUIRE, lp);
6797c478bd9Sstevel@tonic-gate }
6817c478bd9Sstevel@tonic-gate void
lock_set_spl_spin(lock_t * lp,int new_pil,ushort_t * old_pil_addr,int old_pil)6827c478bd9Sstevel@tonic-gate lock_set_spl_spin(lock_t *lp, int new_pil, ushort_t *old_pil_addr, int old_pil)
6837c478bd9Sstevel@tonic-gate {
684575a7426Spt 	int loop_count = 0;
685575a7426Spt 	uint_t backoff = 0;	/* current backoff */
6869d68b18eSck 	hrtime_t spin_time = 0;	/* how long we spun */
6887c478bd9Sstevel@tonic-gate 	if (panicstr)
6897c478bd9Sstevel@tonic-gate 		return;
6917c478bd9Sstevel@tonic-gate 	if (ncpus == 1)
6928793b36bSNick Todd 		panic("lock_set_spl: %p lock held and only one CPU",
6938793b36bSNick Todd 		    (void *)lp);
6957c478bd9Sstevel@tonic-gate 	ASSERT(new_pil > LOCK_LEVEL);
6979d68b18eSck 	spin_time = LOCKSTAT_START_TIME(LS_LOCK_SET_SPL_SPIN);
6997c478bd9Sstevel@tonic-gate 	do {
7007c478bd9Sstevel@tonic-gate 		splx(old_pil);
7017c478bd9Sstevel@tonic-gate 		while (LOCK_HELD(lp)) {
702575a7426Spt 			loop_count++;
7047c478bd9Sstevel@tonic-gate 			if (panicstr) {
7057c478bd9Sstevel@tonic-gate 				*old_pil_addr = (ushort_t)splr(new_pil);
7067c478bd9Sstevel@tonic-gate 				return;
7077c478bd9Sstevel@tonic-gate 			}
708575a7426Spt 			if (ncpus_online == loop_count) {
709575a7426Spt 				backoff = mutex_lock_backoff(0);
710575a7426Spt 				loop_count = 0;
711e603b7d4Spm 			} else {
712575a7426Spt 				backoff = mutex_lock_backoff(backoff);
7137c478bd9Sstevel@tonic-gate 			}
714575a7426Spt 			mutex_lock_delay(backoff);
7157c478bd9Sstevel@tonic-gate 		}
7167c478bd9Sstevel@tonic-gate 		old_pil = splr(new_pil);
7177c478bd9Sstevel@tonic-gate 	} while (!lock_spin_try(lp));
7197c478bd9Sstevel@tonic-gate 	*old_pil_addr = (ushort_t)old_pil;
7219d68b18eSck 	LOCKSTAT_RECORD_TIME(LS_LOCK_SET_SPL_SPIN, lp, spin_time);
7247c478bd9Sstevel@tonic-gate }