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 * Global tokens. These replace the MP lock for major subsystem locking.
111 * These tokens are initially used to lockup both global and individual
114 * Once individual structures get their own locks these tokens are used
115 * only to protect global lists & other variables and to interlock
116 * allocations and teardowns and such.
118 * The UP initializer causes token acquisition to also acquire the MP lock
119 * for maximum compatibility. The feature may be enabled and disabled at
120 * any time, the MP state is copied to the tokref when the token is acquired
121 * and will not race against sysctl changes.
123 struct lwkt_token pmap_token = LWKT_TOKEN_UP_INITIALIZER(pmap_token);
124 struct lwkt_token dev_token = LWKT_TOKEN_UP_INITIALIZER(dev_token);
125 struct lwkt_token vm_token = LWKT_TOKEN_UP_INITIALIZER(vm_token);
126 struct lwkt_token vmspace_token = LWKT_TOKEN_UP_INITIALIZER(vmspace_token);
127 struct lwkt_token kvm_token = LWKT_TOKEN_UP_INITIALIZER(kvm_token);
128 struct lwkt_token proc_token = LWKT_TOKEN_UP_INITIALIZER(proc_token);
129 struct lwkt_token tty_token = LWKT_TOKEN_UP_INITIALIZER(tty_token);
130 struct lwkt_token vnode_token = LWKT_TOKEN_UP_INITIALIZER(vnode_token);
131 struct lwkt_token vmobj_token = LWKT_TOKEN_UP_INITIALIZER(vmobj_token);
133 SYSCTL_INT(_lwkt, OID_AUTO, pmap_mpsafe, CTLFLAG_RW,
134 &pmap_token.t_flags, 0, "Require MP lock for pmap_token");
135 SYSCTL_INT(_lwkt, OID_AUTO, dev_mpsafe, CTLFLAG_RW,
136 &dev_token.t_flags, 0, "Require MP lock for dev_token");
137 SYSCTL_INT(_lwkt, OID_AUTO, vm_mpsafe, CTLFLAG_RW,
138 &vm_token.t_flags, 0, "Require MP lock for vm_token");
139 SYSCTL_INT(_lwkt, OID_AUTO, vmspace_mpsafe, CTLFLAG_RW,
140 &vmspace_token.t_flags, 0, "Require MP lock for vmspace_token");
141 SYSCTL_INT(_lwkt, OID_AUTO, kvm_mpsafe, CTLFLAG_RW,
142 &kvm_token.t_flags, 0, "Require MP lock for kvm_token");
143 SYSCTL_INT(_lwkt, OID_AUTO, proc_mpsafe, CTLFLAG_RW,
144 &proc_token.t_flags, 0, "Require MP lock for proc_token");
145 SYSCTL_INT(_lwkt, OID_AUTO, tty_mpsafe, CTLFLAG_RW,
146 &tty_token.t_flags, 0, "Require MP lock for tty_token");
147 SYSCTL_INT(_lwkt, OID_AUTO, vnode_mpsafe, CTLFLAG_RW,
148 &vnode_token.t_flags, 0, "Require MP lock for vnode_token");
149 SYSCTL_INT(_lwkt, OID_AUTO, vmobj_mpsafe, CTLFLAG_RW,
150 &vmobj_token.t_flags, 0, "Require MP lock for vmobj_token");
153 * The collision count is bumped every time the LWKT scheduler fails
154 * to acquire needed tokens in addition to a normal lwkt_gettoken()
157 SYSCTL_LONG(_lwkt, OID_AUTO, pmap_collisions, CTLFLAG_RW,
158 &pmap_token.t_collisions, 0, "Collision counter of pmap_token");
159 SYSCTL_LONG(_lwkt, OID_AUTO, dev_collisions, CTLFLAG_RW,
160 &dev_token.t_collisions, 0, "Collision counter of dev_token");
161 SYSCTL_LONG(_lwkt, OID_AUTO, vm_collisions, CTLFLAG_RW,
162 &vm_token.t_collisions, 0, "Collision counter of vm_token");
163 SYSCTL_LONG(_lwkt, OID_AUTO, vmspace_collisions, CTLFLAG_RW,
164 &vmspace_token.t_collisions, 0, "Collision counter of vmspace_token");
165 SYSCTL_LONG(_lwkt, OID_AUTO, kvm_collisions, CTLFLAG_RW,
166 &kvm_token.t_collisions, 0, "Collision counter of kvm_token");
167 SYSCTL_LONG(_lwkt, OID_AUTO, proc_collisions, CTLFLAG_RW,
168 &proc_token.t_collisions, 0, "Collision counter of proc_token");
169 SYSCTL_LONG(_lwkt, OID_AUTO, tty_collisions, CTLFLAG_RW,
170 &tty_token.t_collisions, 0, "Collision counter of tty_token");
171 SYSCTL_LONG(_lwkt, OID_AUTO, vnode_collisions, CTLFLAG_RW,
172 &vnode_token.t_collisions, 0, "Collision counter of vnode_token");
175 * Return a pool token given an address
179 _lwkt_token_pool_lookup(void *ptr)
183 i = ((int)(intptr_t)ptr >> 2) ^ ((int)(intptr_t)ptr >> 12);
184 return(&pool_tokens[i & LWKT_MASK_POOL_TOKENS]);
188 * Initialize a tokref_t prior to making it visible in the thread's
191 * As an optimization we set the MPSAFE flag if the thread is already
192 * holding the MP lock. This bypasses unncessary calls to get_mplock() and
193 * rel_mplock() on tokens which are not normally MPSAFE when the thread
194 * is already holding the MP lock.
196 * WARNING: The inherited td_xpcount does not count here because a switch
197 * could schedule the preempted thread and blow away the inherited
202 _lwkt_tokref_init(lwkt_tokref_t ref, lwkt_token_t tok, thread_t td)
206 ref->tr_flags = tok->t_flags;
210 ref->tr_flags |= LWKT_TOKEN_MPSAFE;
214 * Obtain all the tokens required by the specified thread on the current
215 * cpu, return 0 on failure and non-zero on success. If a failure occurs
216 * any partially acquired tokens will be released prior to return.
218 * lwkt_getalltokens is called by the LWKT scheduler to acquire all
219 * tokens that the thread had acquired prior to going to sleep.
221 * The scheduler is responsible for maintaining the MP lock count, so
222 * we don't need to deal with tr_flags here. We also do not do any
223 * logging here. The logging done by lwkt_gettoken() is plenty good
224 * enough to get a feel for it.
226 * Called from a critical section.
229 lwkt_getalltokens(thread_t td, const char **msgp, const void **addrp)
236 * Acquire tokens in forward order, assign or validate tok->t_ref.
238 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
242 * Try to acquire the token if we do not already have
245 * NOTE: If atomic_cmpset_ptr() fails we have to
246 * loop and try again. It just means we
251 if (atomic_cmpset_ptr(&tok->t_ref, NULL, scan))
257 * Test if ref is already recursively held by this
258 * thread. We cannot safely dereference tok->t_ref
259 * (it might belong to another thread and is thus
260 * unstable), but we don't have to. We can simply
263 if (ref >= &td->td_toks_base && ref < td->td_toks_stop)
267 * Otherwise we failed to acquire all the tokens.
271 *addrp = scan->tr_stallpc;
272 atomic_add_long(&tok->t_collisions, 1);
273 lwkt_relalltokens(td);
281 * Release all tokens owned by the specified thread on the current cpu.
283 * This code is really simple. Even in cases where we own all the tokens
284 * note that t_ref may not match the scan for recursively held tokens,
285 * or for the case where a lwkt_getalltokens() failed.
287 * The scheduler is responsible for maintaining the MP lock count, so
288 * we don't need to deal with tr_flags here.
290 * Called from a critical section.
293 lwkt_relalltokens(thread_t td)
298 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
300 if (tok->t_ref == scan)
306 * Token acquisition helper function. The caller must have already
307 * made nref visible by adjusting td_toks_stop and will be responsible
308 * for the disposition of nref on either success or failure.
310 * When acquiring tokens recursively we want tok->t_ref to point to
311 * the outer (first) acquisition so it gets cleared only on the last
316 _lwkt_trytokref2(lwkt_tokref_t nref, thread_t td, int blocking)
322 * Make sure the compiler does not reorder prior instructions
323 * beyond this demark.
328 * Attempt to gain ownership
333 * Try to acquire the token if we do not already have
334 * it. This is not allowed if we are in a hard code
335 * section (because it 'might' have blocked).
339 KASSERT((blocking == 0 ||
340 td->td_gd->gd_intr_nesting_level == 0 ||
341 panic_cpu_gd == mycpu),
342 ("Attempt to acquire token %p not already "
343 "held in hard code section", tok));
346 * NOTE: If atomic_cmpset_ptr() fails we have to
347 * loop and try again. It just means we
350 if (atomic_cmpset_ptr(&tok->t_ref, NULL, nref))
356 * Test if ref is already recursively held by this
357 * thread. We cannot safely dereference tok->t_ref
358 * (it might belong to another thread and is thus
359 * unstable), but we don't have to. We can simply
362 * It is ok to acquire a token that is already held
363 * by the current thread when in a hard code section.
365 if (ref >= &td->td_toks_base && ref < td->td_toks_stop)
369 * Otherwise we failed, and it is not ok to attempt to
370 * acquire a token in a hard code section.
372 KASSERT((blocking == 0 ||
373 td->td_gd->gd_intr_nesting_level == 0),
374 ("Attempt to acquire token %p not already "
375 "held in hard code section", tok));
382 * Acquire a serializing token. This routine does not block.
386 _lwkt_trytokref(lwkt_tokref_t ref, thread_t td)
388 if ((ref->tr_flags & LWKT_TOKEN_MPSAFE) == 0) {
389 if (try_mplock() == 0) {
394 if (_lwkt_trytokref2(ref, td, 0) == FALSE) {
396 * Cleanup, deactivate the failed token.
398 if ((ref->tr_flags & LWKT_TOKEN_MPSAFE) == 0)
407 * Acquire a serializing token. This routine can block.
411 _lwkt_gettokref(lwkt_tokref_t ref, thread_t td, const void **stkframe)
413 if ((ref->tr_flags & LWKT_TOKEN_MPSAFE) == 0)
415 if (_lwkt_trytokref2(ref, td, 1) == FALSE) {
417 * Give up running if we can't acquire the token right now.
419 * Since the tokref is already active the scheduler now
420 * takes care of acquisition, so we need only call
423 * Since we failed this was not a recursive token so upon
424 * return tr_tok->t_ref should be assigned to this specific
427 ref->tr_stallpc = stkframe[-1];
428 atomic_add_long(&ref->tr_tok->t_collisions, 1);
432 KKASSERT(ref->tr_tok->t_ref == ref);
437 lwkt_gettoken(lwkt_token_t tok)
439 thread_t td = curthread;
442 ref = td->td_toks_stop;
443 KKASSERT(ref < &td->td_toks_end);
446 _lwkt_tokref_init(ref, tok, td);
447 _lwkt_gettokref(ref, td, (const void **)&tok);
451 lwkt_gettoken_hard(lwkt_token_t tok)
453 thread_t td = curthread;
456 ref = td->td_toks_stop;
457 KKASSERT(ref < &td->td_toks_end);
460 _lwkt_tokref_init(ref, tok, td);
461 _lwkt_gettokref(ref, td, (const void **)&tok);
462 crit_enter_hard_gd(td->td_gd);
466 lwkt_getpooltoken(void *ptr)
468 thread_t td = curthread;
472 ref = td->td_toks_stop;
473 KKASSERT(ref < &td->td_toks_end);
476 tok = _lwkt_token_pool_lookup(ptr);
477 _lwkt_tokref_init(ref, tok, td);
478 _lwkt_gettokref(ref, td, (const void **)&ptr);
483 * Attempt to acquire a token, return TRUE on success, FALSE on failure.
486 lwkt_trytoken(lwkt_token_t tok)
488 thread_t td = curthread;
491 ref = td->td_toks_stop;
492 KKASSERT(ref < &td->td_toks_end);
495 _lwkt_tokref_init(ref, tok, td);
496 return(_lwkt_trytokref(ref, td));
500 * Release a serializing token.
502 * WARNING! All tokens must be released in reverse order. This will be
506 lwkt_reltoken(lwkt_token_t tok)
508 thread_t td = curthread;
512 * Remove ref from thread token list and assert that it matches
513 * the token passed in. Tokens must be released in reverse order.
515 ref = td->td_toks_stop - 1;
516 KKASSERT(ref >= &td->td_toks_base && ref->tr_tok == tok);
519 * Only clear the token if it matches ref. If ref was a recursively
520 * acquired token it may not match.
522 * If the token was not MPSAFE release the MP lock.
524 * NOTE: We have to do this before adjust td_toks_stop, otherwise
525 * a fast interrupt can come along and reuse our ref while
526 * tok is still attached to it.
528 if (tok->t_ref == ref)
531 if ((ref->tr_flags & LWKT_TOKEN_MPSAFE) == 0)
535 * Finally adjust td_toks_stop, be very sure that the compiler
536 * does not reorder the clearing of tok->t_ref with the
537 * decrementing of td->td_toks_stop.
540 td->td_toks_stop = ref;
541 KKASSERT(tok->t_ref != ref);
545 lwkt_reltoken_hard(lwkt_token_t tok)
552 * It is faster for users of lwkt_getpooltoken() to use the returned
553 * token and just call lwkt_reltoken(), but for convenience we provide
554 * this function which looks the token up based on the ident.
557 lwkt_relpooltoken(void *ptr)
559 lwkt_token_t tok = _lwkt_token_pool_lookup(ptr);
565 * Pool tokens are used to provide a type-stable serializing token
566 * pointer that does not race against disappearing data structures.
568 * This routine is called in early boot just after we setup the BSP's
569 * globaldata structure.
572 lwkt_token_pool_init(void)
576 for (i = 0; i < LWKT_NUM_POOL_TOKENS; ++i)
577 lwkt_token_init(&pool_tokens[i], 1, "pool");
581 lwkt_token_pool_lookup(void *ptr)
583 return (_lwkt_token_pool_lookup(ptr));
587 * Initialize a token. If mpsafe is 0, the MP lock is acquired before
588 * acquiring the token and released after releasing the token.
591 lwkt_token_init(lwkt_token_t tok, int mpsafe, const char *desc)
594 tok->t_flags = mpsafe ? LWKT_TOKEN_MPSAFE : 0;
595 tok->t_collisions = 0;
600 lwkt_token_uninit(lwkt_token_t tok)
607 lwkt_token_is_stale(lwkt_tokref_t ref)
609 lwkt_token_t tok = ref->tr_tok;
611 KKASSERT(tok->t_owner == curthread && ref->tr_state == 1 &&
614 /* Token is not stale */
615 if (tok->t_lastowner == tok->t_owner)
619 * The token is stale. Reset to not stale so that the next call to
620 * lwkt_token_is_stale will return "not stale" unless the token
621 * was acquired in-between by another thread.
623 tok->t_lastowner = tok->t_owner;