2 * Copyright (c) 2003,2004,2009 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * lwkt_token - Implement soft token locks.
38 * Tokens are locks which serialize a thread only while the thread is
39 * running. If the thread blocks all tokens are released, then reacquired
40 * when the thread resumes.
42 * This implementation requires no critical sections or spin locks, but
43 * does use atomic_cmpset_ptr().
45 * Tokens may be recursively acquired by the same thread. However the
46 * caller must be sure to release such tokens in reverse order.
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/kernel.h>
52 #include <sys/rtprio.h>
53 #include <sys/queue.h>
54 #include <sys/sysctl.h>
56 #include <sys/kthread.h>
57 #include <machine/cpu.h>
60 #include <sys/spinlock.h>
62 #include <sys/thread2.h>
63 #include <sys/spinlock2.h>
64 #include <sys/mplock2.h>
67 #include <vm/vm_param.h>
68 #include <vm/vm_kern.h>
69 #include <vm/vm_object.h>
70 #include <vm/vm_page.h>
71 #include <vm/vm_map.h>
72 #include <vm/vm_pager.h>
73 #include <vm/vm_extern.h>
74 #include <vm/vm_zone.h>
76 #include <machine/stdarg.h>
77 #include <machine/smp.h>
79 #ifndef LWKT_NUM_POOL_TOKENS
80 #define LWKT_NUM_POOL_TOKENS 1024 /* power of 2 */
82 #define LWKT_MASK_POOL_TOKENS (LWKT_NUM_POOL_TOKENS - 1)
84 static lwkt_token pool_tokens[LWKT_NUM_POOL_TOKENS];
86 #define TOKEN_STRING "REF=%p TOK=%p TD=%p"
87 #define CONTENDED_STRING "REF=%p TOK=%p TD=%p (contention started)"
88 #define UNCONTENDED_STRING "REF=%p TOK=%p TD=%p (contention stopped)"
89 #if !defined(KTR_TOKENS)
90 #define KTR_TOKENS KTR_ALL
93 KTR_INFO_MASTER(tokens);
94 KTR_INFO(KTR_TOKENS, tokens, fail, 0, TOKEN_STRING, sizeof(void *) * 3);
95 KTR_INFO(KTR_TOKENS, tokens, succ, 1, TOKEN_STRING, sizeof(void *) * 3);
97 KTR_INFO(KTR_TOKENS, tokens, release, 2, TOKEN_STRING, sizeof(void *) * 3);
98 KTR_INFO(KTR_TOKENS, tokens, remote, 3, TOKEN_STRING, sizeof(void *) * 3);
99 KTR_INFO(KTR_TOKENS, tokens, reqremote, 4, TOKEN_STRING, sizeof(void *) * 3);
100 KTR_INFO(KTR_TOKENS, tokens, reqfail, 5, TOKEN_STRING, sizeof(void *) * 3);
101 KTR_INFO(KTR_TOKENS, tokens, drain, 6, TOKEN_STRING, sizeof(void *) * 3);
102 KTR_INFO(KTR_TOKENS, tokens, contention_start, 7, CONTENDED_STRING, sizeof(void *) * 3);
103 KTR_INFO(KTR_TOKENS, tokens, contention_stop, 7, UNCONTENDED_STRING, sizeof(void *) * 3);
106 #define logtoken(name, ref) \
107 KTR_LOG(tokens_ ## name, ref, ref->tr_tok, curthread)
110 * Cpu contention mask for directed wakeups.
112 cpumask_t cpu_contention_mask;
115 * Global tokens. These replace the MP lock for major subsystem locking.
116 * These tokens are initially used to lockup both global and individual
119 * Once individual structures get their own locks these tokens are used
120 * only to protect global lists & other variables and to interlock
121 * allocations and teardowns and such.
123 * The UP initializer causes token acquisition to also acquire the MP lock
124 * for maximum compatibility. The feature may be enabled and disabled at
125 * any time, the MP state is copied to the tokref when the token is acquired
126 * and will not race against sysctl changes.
128 struct lwkt_token mp_token = LWKT_TOKEN_MP_INITIALIZER(mp_token);
129 struct lwkt_token pmap_token = LWKT_TOKEN_UP_INITIALIZER(pmap_token);
130 struct lwkt_token dev_token = LWKT_TOKEN_UP_INITIALIZER(dev_token);
131 struct lwkt_token vm_token = LWKT_TOKEN_UP_INITIALIZER(vm_token);
132 struct lwkt_token vmspace_token = LWKT_TOKEN_UP_INITIALIZER(vmspace_token);
133 struct lwkt_token kvm_token = LWKT_TOKEN_UP_INITIALIZER(kvm_token);
134 struct lwkt_token proc_token = LWKT_TOKEN_UP_INITIALIZER(proc_token);
135 struct lwkt_token tty_token = LWKT_TOKEN_UP_INITIALIZER(tty_token);
136 struct lwkt_token vnode_token = LWKT_TOKEN_UP_INITIALIZER(vnode_token);
137 struct lwkt_token vmobj_token = LWKT_TOKEN_UP_INITIALIZER(vmobj_token);
139 SYSCTL_INT(_lwkt, OID_AUTO, pmap_mpsafe, CTLFLAG_RW,
140 &pmap_token.t_flags, 0, "Require MP lock for pmap_token");
141 SYSCTL_INT(_lwkt, OID_AUTO, dev_mpsafe, CTLFLAG_RW,
142 &dev_token.t_flags, 0, "Require MP lock for dev_token");
143 SYSCTL_INT(_lwkt, OID_AUTO, vm_mpsafe, CTLFLAG_RW,
144 &vm_token.t_flags, 0, "Require MP lock for vm_token");
145 SYSCTL_INT(_lwkt, OID_AUTO, vmspace_mpsafe, CTLFLAG_RW,
146 &vmspace_token.t_flags, 0, "Require MP lock for vmspace_token");
147 SYSCTL_INT(_lwkt, OID_AUTO, kvm_mpsafe, CTLFLAG_RW,
148 &kvm_token.t_flags, 0, "Require MP lock for kvm_token");
149 SYSCTL_INT(_lwkt, OID_AUTO, proc_mpsafe, CTLFLAG_RW,
150 &proc_token.t_flags, 0, "Require MP lock for proc_token");
151 SYSCTL_INT(_lwkt, OID_AUTO, tty_mpsafe, CTLFLAG_RW,
152 &tty_token.t_flags, 0, "Require MP lock for tty_token");
153 SYSCTL_INT(_lwkt, OID_AUTO, vnode_mpsafe, CTLFLAG_RW,
154 &vnode_token.t_flags, 0, "Require MP lock for vnode_token");
155 SYSCTL_INT(_lwkt, OID_AUTO, vmobj_mpsafe, CTLFLAG_RW,
156 &vmobj_token.t_flags, 0, "Require MP lock for vmobj_token");
159 * The collision count is bumped every time the LWKT scheduler fails
160 * to acquire needed tokens in addition to a normal lwkt_gettoken()
163 SYSCTL_LONG(_lwkt, OID_AUTO, mp_collisions, CTLFLAG_RW,
164 &mp_token.t_collisions, 0, "Collision counter of mp_token");
165 SYSCTL_LONG(_lwkt, OID_AUTO, pmap_collisions, CTLFLAG_RW,
166 &pmap_token.t_collisions, 0, "Collision counter of pmap_token");
167 SYSCTL_LONG(_lwkt, OID_AUTO, dev_collisions, CTLFLAG_RW,
168 &dev_token.t_collisions, 0, "Collision counter of dev_token");
169 SYSCTL_LONG(_lwkt, OID_AUTO, vm_collisions, CTLFLAG_RW,
170 &vm_token.t_collisions, 0, "Collision counter of vm_token");
171 SYSCTL_LONG(_lwkt, OID_AUTO, vmspace_collisions, CTLFLAG_RW,
172 &vmspace_token.t_collisions, 0, "Collision counter of vmspace_token");
173 SYSCTL_LONG(_lwkt, OID_AUTO, kvm_collisions, CTLFLAG_RW,
174 &kvm_token.t_collisions, 0, "Collision counter of kvm_token");
175 SYSCTL_LONG(_lwkt, OID_AUTO, proc_collisions, CTLFLAG_RW,
176 &proc_token.t_collisions, 0, "Collision counter of proc_token");
177 SYSCTL_LONG(_lwkt, OID_AUTO, tty_collisions, CTLFLAG_RW,
178 &tty_token.t_collisions, 0, "Collision counter of tty_token");
179 SYSCTL_LONG(_lwkt, OID_AUTO, vnode_collisions, CTLFLAG_RW,
180 &vnode_token.t_collisions, 0, "Collision counter of vnode_token");
184 * Acquire the initial mplock
186 * (low level boot only)
189 cpu_get_initial_mplock(void)
191 KKASSERT(mp_token.t_ref == NULL);
192 if (lwkt_trytoken(&mp_token) == FALSE)
193 panic("cpu_get_initial_mplock");
198 * Return a pool token given an address
202 _lwkt_token_pool_lookup(void *ptr)
206 i = ((int)(intptr_t)ptr >> 2) ^ ((int)(intptr_t)ptr >> 12);
207 return(&pool_tokens[i & LWKT_MASK_POOL_TOKENS]);
211 * Initialize a tokref_t prior to making it visible in the thread's
214 * As an optimization we set the MPSAFE flag if the thread is already
215 * holding the mp_token. This bypasses unncessary calls to get_mplock() and
216 * rel_mplock() on tokens which are not normally MPSAFE when the thread
217 * is already holding the MP lock.
221 _lwkt_tok_flags(lwkt_token_t tok, thread_t td)
226 * tok->t_flags can change out from under us, make sure we have
229 flags = tok->t_flags;
232 if ((flags & LWKT_TOKEN_MPSAFE) == 0 &&
233 _lwkt_token_held(&mp_token, td)) {
234 return (flags | LWKT_TOKEN_MPSAFE);
239 return (flags | LWKT_TOKEN_MPSAFE);
245 _lwkt_tokref_init(lwkt_tokref_t ref, lwkt_token_t tok, thread_t td,
250 ref->tr_flags = flags;
254 * Obtain all the tokens required by the specified thread on the current
255 * cpu, return 0 on failure and non-zero on success. If a failure occurs
256 * any partially acquired tokens will be released prior to return.
258 * lwkt_getalltokens is called by the LWKT scheduler to acquire all
259 * tokens that the thread had acquired prior to going to sleep.
261 * The scheduler is responsible for maintaining the MP lock count, so
262 * we don't need to deal with tr_flags here. We also do not do any
263 * logging here. The logging done by lwkt_gettoken() is plenty good
264 * enough to get a feel for it.
266 * Called from a critical section.
269 lwkt_getalltokens(thread_t td)
276 * Acquire tokens in forward order, assign or validate tok->t_ref.
278 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
282 * Try to acquire the token if we do not already have
285 * NOTE: If atomic_cmpset_ptr() fails we have to
286 * loop and try again. It just means we
291 if (atomic_cmpset_ptr(&tok->t_ref, NULL, scan))
297 * Test if ref is already recursively held by this
298 * thread. We cannot safely dereference tok->t_ref
299 * (it might belong to another thread and is thus
300 * unstable), but we don't have to. We can simply
303 if (ref >= &td->td_toks_base && ref < td->td_toks_stop)
307 * Otherwise we failed to acquire all the tokens.
310 td->td_wmesg = tok->t_desc;
311 atomic_add_long(&tok->t_collisions, 1);
312 lwkt_relalltokens(td);
320 * Release all tokens owned by the specified thread on the current cpu.
322 * This code is really simple. Even in cases where we own all the tokens
323 * note that t_ref may not match the scan for recursively held tokens,
324 * or for the case where a lwkt_getalltokens() failed.
326 * The scheduler is responsible for maintaining the MP lock count, so
327 * we don't need to deal with tr_flags here.
329 * Called from a critical section.
332 lwkt_relalltokens(thread_t td)
337 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
339 if (tok->t_ref == scan)
345 * Token acquisition helper function. The caller must have already
346 * made nref visible by adjusting td_toks_stop and will be responsible
347 * for the disposition of nref on either success or failure.
349 * When acquiring tokens recursively we want tok->t_ref to point to
350 * the outer (first) acquisition so it gets cleared only on the last
355 _lwkt_trytokref2(lwkt_tokref_t nref, thread_t td, int blocking)
361 * Make sure the compiler does not reorder prior instructions
362 * beyond this demark.
367 * Attempt to gain ownership
372 * Try to acquire the token if we do not already have
373 * it. This is not allowed if we are in a hard code
374 * section (because it 'might' have blocked).
378 KASSERT((blocking == 0 ||
379 td->td_gd->gd_intr_nesting_level == 0 ||
380 panic_cpu_gd == mycpu),
381 ("Attempt to acquire token %p not already "
382 "held in hard code section", tok));
385 * NOTE: If atomic_cmpset_ptr() fails we have to
386 * loop and try again. It just means we
389 if (atomic_cmpset_ptr(&tok->t_ref, NULL, nref))
395 * Test if ref is already recursively held by this
396 * thread. We cannot safely dereference tok->t_ref
397 * (it might belong to another thread and is thus
398 * unstable), but we don't have to. We can simply
401 * It is ok to acquire a token that is already held
402 * by the current thread when in a hard code section.
404 if (ref >= &td->td_toks_base && ref < td->td_toks_stop)
408 * Otherwise we failed, and it is not ok to attempt to
409 * acquire a token in a hard code section.
411 KASSERT((blocking == 0 ||
412 td->td_gd->gd_intr_nesting_level == 0),
413 ("Attempt to acquire token %p not already "
414 "held in hard code section", tok));
421 * Get a serializing token. This routine can block.
424 lwkt_gettoken(lwkt_token_t tok)
426 thread_t td = curthread;
430 flags = _lwkt_tok_flags(tok, td);
431 if ((flags & LWKT_TOKEN_MPSAFE) == 0)
434 ref = td->td_toks_stop;
435 KKASSERT(ref < &td->td_toks_end);
438 _lwkt_tokref_init(ref, tok, td, flags);
440 if (_lwkt_trytokref2(ref, td, 1) == FALSE) {
442 * Give up running if we can't acquire the token right now.
444 * Since the tokref is already active the scheduler now
445 * takes care of acquisition, so we need only call
448 * Since we failed this was not a recursive token so upon
449 * return tr_tok->t_ref should be assigned to this specific
452 atomic_add_long(&ref->tr_tok->t_collisions, 1);
456 KKASSERT(ref->tr_tok->t_ref == ref);
461 lwkt_gettoken_hard(lwkt_token_t tok)
463 thread_t td = curthread;
467 flags = _lwkt_tok_flags(tok, td);
468 if ((flags & LWKT_TOKEN_MPSAFE) == 0)
471 ref = td->td_toks_stop;
472 KKASSERT(ref < &td->td_toks_end);
475 _lwkt_tokref_init(ref, tok, td, flags);
477 if (_lwkt_trytokref2(ref, td, 1) == FALSE) {
479 * Give up running if we can't acquire the token right now.
481 * Since the tokref is already active the scheduler now
482 * takes care of acquisition, so we need only call
485 * Since we failed this was not a recursive token so upon
486 * return tr_tok->t_ref should be assigned to this specific
489 atomic_add_long(&ref->tr_tok->t_collisions, 1);
493 KKASSERT(ref->tr_tok->t_ref == ref);
495 crit_enter_hard_gd(td->td_gd);
499 lwkt_getpooltoken(void *ptr)
501 thread_t td = curthread;
506 tok = _lwkt_token_pool_lookup(ptr);
507 flags = _lwkt_tok_flags(tok, td);
508 if ((flags & LWKT_TOKEN_MPSAFE) == 0)
511 ref = td->td_toks_stop;
512 KKASSERT(ref < &td->td_toks_end);
515 _lwkt_tokref_init(ref, tok, td, flags);
517 if (_lwkt_trytokref2(ref, td, 1) == FALSE) {
519 * Give up running if we can't acquire the token right now.
521 * Since the tokref is already active the scheduler now
522 * takes care of acquisition, so we need only call
525 * Since we failed this was not a recursive token so upon
526 * return tr_tok->t_ref should be assigned to this specific
529 atomic_add_long(&ref->tr_tok->t_collisions, 1);
533 KKASSERT(ref->tr_tok->t_ref == ref);
539 * Attempt to acquire a token, return TRUE on success, FALSE on failure.
542 lwkt_trytoken(lwkt_token_t tok)
544 thread_t td = curthread;
548 flags = _lwkt_tok_flags(tok, td);
549 if ((flags & LWKT_TOKEN_MPSAFE) == 0) {
550 if (try_mplock() == 0)
554 ref = td->td_toks_stop;
555 KKASSERT(ref < &td->td_toks_end);
558 _lwkt_tokref_init(ref, tok, td, flags);
560 if (_lwkt_trytokref2(ref, td, 0) == FALSE) {
562 * Cleanup, deactivate the failed token.
564 if ((ref->tr_flags & LWKT_TOKEN_MPSAFE) == 0) {
579 * Release a serializing token.
581 * WARNING! All tokens must be released in reverse order. This will be
585 lwkt_reltoken(lwkt_token_t tok)
587 thread_t td = curthread;
591 * Remove ref from thread token list and assert that it matches
592 * the token passed in. Tokens must be released in reverse order.
594 ref = td->td_toks_stop - 1;
595 KKASSERT(ref >= &td->td_toks_base && ref->tr_tok == tok);
598 * Only clear the token if it matches ref. If ref was a recursively
599 * acquired token it may not match. Then adjust td_toks_stop.
601 * Some comparisons must be run prior to adjusting td_toks_stop
602 * to avoid racing against a fast interrupt/ ipi which tries to
605 * We must also be absolutely sure that the compiler does not
606 * reorder the clearing of t_ref and the adjustment of td_toks_stop,
607 * or reorder the adjustment of td_toks_stop against the conditional.
609 * NOTE: The mplock is a token also so sequencing is a bit complex.
611 if (tok->t_ref == ref)
614 if ((ref->tr_flags & LWKT_TOKEN_MPSAFE) == 0) {
616 td->td_toks_stop = ref;
621 td->td_toks_stop = ref;
624 KKASSERT(tok->t_ref != ref);
628 lwkt_reltoken_hard(lwkt_token_t tok)
635 * It is faster for users of lwkt_getpooltoken() to use the returned
636 * token and just call lwkt_reltoken(), but for convenience we provide
637 * this function which looks the token up based on the ident.
640 lwkt_relpooltoken(void *ptr)
642 lwkt_token_t tok = _lwkt_token_pool_lookup(ptr);
647 * Return a count of the number of token refs the thread has to the
648 * specified token, whether it currently owns the token or not.
651 lwkt_cnttoken(lwkt_token_t tok, thread_t td)
656 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
657 if (scan->tr_tok == tok)
665 * Pool tokens are used to provide a type-stable serializing token
666 * pointer that does not race against disappearing data structures.
668 * This routine is called in early boot just after we setup the BSP's
669 * globaldata structure.
672 lwkt_token_pool_init(void)
676 for (i = 0; i < LWKT_NUM_POOL_TOKENS; ++i)
677 lwkt_token_init(&pool_tokens[i], 1, "pool");
681 lwkt_token_pool_lookup(void *ptr)
683 return (_lwkt_token_pool_lookup(ptr));
687 * Initialize a token. If mpsafe is 0, the MP lock is acquired before
688 * acquiring the token and released after releasing the token.
691 lwkt_token_init(lwkt_token_t tok, int mpsafe, const char *desc)
694 tok->t_flags = mpsafe ? LWKT_TOKEN_MPSAFE : 0;
695 tok->t_collisions = 0;
700 lwkt_token_uninit(lwkt_token_t tok)
707 lwkt_token_is_stale(lwkt_tokref_t ref)
709 lwkt_token_t tok = ref->tr_tok;
711 KKASSERT(tok->t_owner == curthread && ref->tr_state == 1 &&
714 /* Token is not stale */
715 if (tok->t_lastowner == tok->t_owner)
719 * The token is stale. Reset to not stale so that the next call to
720 * lwkt_token_is_stale will return "not stale" unless the token
721 * was acquired in-between by another thread.
723 tok->t_lastowner = tok->t_owner;