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)
85 static int token_debug = 0;
88 static lwkt_token pool_tokens[LWKT_NUM_POOL_TOKENS];
90 #define TOKEN_STRING "REF=%p TOK=%p TD=%p"
91 #define CONTENDED_STRING "REF=%p TOK=%p TD=%p (contention started)"
92 #define UNCONTENDED_STRING "REF=%p TOK=%p TD=%p (contention stopped)"
93 #if !defined(KTR_TOKENS)
94 #define KTR_TOKENS KTR_ALL
97 KTR_INFO_MASTER(tokens);
98 KTR_INFO(KTR_TOKENS, tokens, fail, 0, TOKEN_STRING, sizeof(void *) * 3);
99 KTR_INFO(KTR_TOKENS, tokens, succ, 1, TOKEN_STRING, sizeof(void *) * 3);
101 KTR_INFO(KTR_TOKENS, tokens, release, 2, TOKEN_STRING, sizeof(void *) * 3);
102 KTR_INFO(KTR_TOKENS, tokens, remote, 3, TOKEN_STRING, sizeof(void *) * 3);
103 KTR_INFO(KTR_TOKENS, tokens, reqremote, 4, TOKEN_STRING, sizeof(void *) * 3);
104 KTR_INFO(KTR_TOKENS, tokens, reqfail, 5, TOKEN_STRING, sizeof(void *) * 3);
105 KTR_INFO(KTR_TOKENS, tokens, drain, 6, TOKEN_STRING, sizeof(void *) * 3);
106 KTR_INFO(KTR_TOKENS, tokens, contention_start, 7, CONTENDED_STRING, sizeof(void *) * 3);
107 KTR_INFO(KTR_TOKENS, tokens, contention_stop, 7, UNCONTENDED_STRING, sizeof(void *) * 3);
110 #define logtoken(name, ref) \
111 KTR_LOG(tokens_ ## name, ref, ref->tr_tok, curthread)
114 SYSCTL_INT(_lwkt, OID_AUTO, token_debug, CTLFLAG_RW, &token_debug, 0, "");
118 * Global tokens. These replace the MP lock for major subsystem locking.
119 * These tokens are initially used to lockup both global and individual
122 * Once individual structures get their own locks these tokens are used
123 * only to protect global lists & other variables and to interlock
124 * allocations and teardowns and such.
126 * The UP initializer causes token acquisition to also acquire the MP lock
127 * for maximum compatibility. The feature may be enabled and disabled at
128 * any time, the MP state is copied to the tokref when the token is acquired
129 * and will not race against sysctl changes.
131 struct lwkt_token pmap_token = LWKT_TOKEN_UP_INITIALIZER;
132 struct lwkt_token dev_token = LWKT_TOKEN_UP_INITIALIZER;
133 struct lwkt_token vm_token = LWKT_TOKEN_UP_INITIALIZER;
134 struct lwkt_token kvm_token = LWKT_TOKEN_UP_INITIALIZER;
135 struct lwkt_token proc_token = LWKT_TOKEN_UP_INITIALIZER;
136 struct lwkt_token tty_token = LWKT_TOKEN_UP_INITIALIZER;
137 struct lwkt_token vnode_token = LWKT_TOKEN_UP_INITIALIZER;
139 SYSCTL_INT(_lwkt, OID_AUTO, pmap_mpsafe,
140 CTLFLAG_RW, &pmap_token.t_flags, 0, "");
141 SYSCTL_INT(_lwkt, OID_AUTO, dev_mpsafe,
142 CTLFLAG_RW, &dev_token.t_flags, 0, "");
143 SYSCTL_INT(_lwkt, OID_AUTO, vm_mpsafe,
144 CTLFLAG_RW, &vm_token.t_flags, 0, "");
145 SYSCTL_INT(_lwkt, OID_AUTO, kvm_mpsafe,
146 CTLFLAG_RW, &kvm_token.t_flags, 0, "");
147 SYSCTL_INT(_lwkt, OID_AUTO, proc_mpsafe,
148 CTLFLAG_RW, &proc_token.t_flags, 0, "");
149 SYSCTL_INT(_lwkt, OID_AUTO, tty_mpsafe,
150 CTLFLAG_RW, &tty_token.t_flags, 0, "");
151 SYSCTL_INT(_lwkt, OID_AUTO, vnode_mpsafe,
152 CTLFLAG_RW, &vnode_token.t_flags, 0, "");
155 * The collision count is bumped every time the LWKT scheduler fails
156 * to acquire needed tokens in addition to a normal lwkt_gettoken()
159 SYSCTL_LONG(_lwkt, OID_AUTO, pmap_collisions,
160 CTLFLAG_RW, &pmap_token.t_collisions, 0, "");
161 SYSCTL_LONG(_lwkt, OID_AUTO, dev_collisions,
162 CTLFLAG_RW, &dev_token.t_collisions, 0, "");
163 SYSCTL_LONG(_lwkt, OID_AUTO, vm_collisions,
164 CTLFLAG_RW, &vm_token.t_collisions, 0, "");
165 SYSCTL_LONG(_lwkt, OID_AUTO, kvm_collisions,
166 CTLFLAG_RW, &kvm_token.t_collisions, 0, "");
167 SYSCTL_LONG(_lwkt, OID_AUTO, proc_collisions,
168 CTLFLAG_RW, &proc_token.t_collisions, 0, "");
169 SYSCTL_LONG(_lwkt, OID_AUTO, tty_collisions,
170 CTLFLAG_RW, &tty_token.t_collisions, 0, "");
171 SYSCTL_LONG(_lwkt, OID_AUTO, vnode_collisions,
172 CTLFLAG_RW, &vnode_token.t_collisions, 0, "");
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.
198 _lwkt_tokref_init(lwkt_tokref_t ref, lwkt_token_t tok, thread_t td)
202 ref->tr_flags = tok->t_flags;
206 ref->tr_flags |= LWKT_TOKEN_MPSAFE;
210 * Obtain all the tokens required by the specified thread on the current
211 * cpu, return 0 on failure and non-zero on success. If a failure occurs
212 * any partially acquired tokens will be released prior to return.
214 * lwkt_getalltokens is called by the LWKT scheduler to acquire all
215 * tokens that the thread had acquired prior to going to sleep.
217 * The scheduler is responsible for maintaining the MP lock count, so
218 * we don't need to deal with tr_flags here. We also do not do any
219 * logging here. The logging done by lwkt_gettoken() is plenty good
220 * enough to get a feel for it.
222 * Called from a critical section.
225 lwkt_getalltokens(thread_t td)
232 * Acquire tokens in forward order, assign or validate tok->t_ref.
234 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
238 * Try to acquire the token if we do not already have
241 * NOTE: If atomic_cmpset_ptr() fails we have to
242 * loop and try again. It just means we
247 if (atomic_cmpset_ptr(&tok->t_ref, NULL, scan))
253 * Test if ref is already recursively held by this
254 * thread. We cannot safely dereference tok->t_ref
255 * (it might belong to another thread and is thus
256 * unstable), but we don't have to. We can simply
259 if (ref >= &td->td_toks_base && ref < td->td_toks_stop)
263 * Otherwise we failed to acquire all the tokens.
266 atomic_add_long(&tok->t_collisions, 1);
267 lwkt_relalltokens(td);
275 * Release all tokens owned by the specified thread on the current cpu.
277 * This code is really simple. Even in cases where we own all the tokens
278 * note that t_ref may not match the scan for recursively held tokens,
279 * or for the case where a lwkt_getalltokens() failed.
281 * The scheduler is responsible for maintaining the MP lock count, so
282 * we don't need to deal with tr_flags here.
284 * Called from a critical section.
287 lwkt_relalltokens(thread_t td)
292 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) {
294 if (tok->t_ref == scan)
300 * Token acquisition helper function. The caller must have already
301 * made nref visible by adjusting td_toks_stop and will be responsible
302 * for the disposition of nref on either success or failure.
304 * When acquiring tokens recursively we want tok->t_ref to point to
305 * the outer (first) acquisition so it gets cleared only on the last
310 _lwkt_trytokref2(lwkt_tokref_t nref, thread_t td)
315 KKASSERT(td->td_gd->gd_intr_nesting_level == 0);
318 * Make sure the compiler does not reorder prior instructions
319 * beyond this demark.
324 * Attempt to gain ownership
329 * Try to acquire the token if we do not already have
335 * NOTE: If atomic_cmpset_ptr() fails we have to
336 * loop and try again. It just means we
339 if (atomic_cmpset_ptr(&tok->t_ref, NULL, nref))
345 * Test if ref is already recursively held by this
346 * thread. We cannot safely dereference tok->t_ref
347 * (it might belong to another thread and is thus
348 * unstable), but we don't have to. We can simply
351 if (ref >= &td->td_toks_base && ref < td->td_toks_stop)
355 * Otherwise we failed.
362 * Acquire a serializing token. This routine does not block.
366 _lwkt_trytokref(lwkt_tokref_t ref, thread_t td)
368 if ((ref->tr_flags & LWKT_TOKEN_MPSAFE) == 0) {
369 if (try_mplock() == 0)
372 if (_lwkt_trytokref2(ref, td) == FALSE) {
374 * Cleanup, deactivate the failed token.
377 if ((ref->tr_flags & LWKT_TOKEN_MPSAFE) == 0)
385 * Acquire a serializing token. This routine can block.
389 _lwkt_gettokref(lwkt_tokref_t ref, thread_t td)
391 if ((ref->tr_flags & LWKT_TOKEN_MPSAFE) == 0)
393 if (_lwkt_trytokref2(ref, td) == FALSE) {
395 * Give up running if we can't acquire the token right now.
397 * Since the tokref is already active the scheduler now
398 * takes care of acquisition, so we need only call
401 * Since we failed this was not a recursive token so upon
402 * return tr_tok->t_ref should be assigned to this specific
405 atomic_add_long(&ref->tr_tok->t_collisions, 1);
409 KKASSERT(ref->tr_tok->t_ref == ref);
414 lwkt_gettoken(lwkt_token_t tok)
416 thread_t td = curthread;
419 ref = td->td_toks_stop;
420 KKASSERT(ref < &td->td_toks_end);
421 _lwkt_tokref_init(ref, tok, td);
423 _lwkt_gettokref(ref, td);
427 lwkt_getpooltoken(void *ptr)
429 thread_t td = curthread;
433 ref = td->td_toks_stop;
434 KKASSERT(ref < &td->td_toks_end);
435 tok = _lwkt_token_pool_lookup(ptr);
436 _lwkt_tokref_init(ref, tok, td);
438 _lwkt_gettokref(ref, td);
443 lwkt_trytoken(lwkt_token_t tok)
445 thread_t td = curthread;
448 ref = td->td_toks_stop;
449 KKASSERT(ref < &td->td_toks_end);
450 _lwkt_tokref_init(ref, tok, td);
452 return(_lwkt_trytokref(ref, td));
456 * Release a serializing token.
458 * WARNING! All tokens must be released in reverse order. This will be
462 lwkt_reltoken(lwkt_token_t tok)
464 thread_t td = curthread;
468 * Remove ref from thread token list and assert that it matches
469 * the token passed in. Tokens must be released in reverse order.
471 ref = td->td_toks_stop - 1;
472 KKASSERT(ref >= &td->td_toks_base && ref->tr_tok == tok);
473 td->td_toks_stop = ref;
476 * If the token was not MPSAFE release the MP lock.
478 if ((ref->tr_flags & LWKT_TOKEN_MPSAFE) == 0)
482 * Make sure the compiler does not reorder the clearing of
488 * Only clear the token if it matches ref. If ref was a recursively
489 * acquired token it may not match.
491 if (tok->t_ref == ref)
496 * Pool tokens are used to provide a type-stable serializing token
497 * pointer that does not race against disappearing data structures.
499 * This routine is called in early boot just after we setup the BSP's
500 * globaldata structure.
503 lwkt_token_pool_init(void)
507 for (i = 0; i < LWKT_NUM_POOL_TOKENS; ++i)
508 lwkt_token_init(&pool_tokens[i], 1);
512 lwkt_token_pool_lookup(void *ptr)
514 return (_lwkt_token_pool_lookup(ptr));
518 * Initialize the owner and release-to cpu to the current cpu
519 * and reset the generation count.
522 lwkt_token_init(lwkt_token_t tok, int mpsafe)
525 tok->t_flags = mpsafe ? LWKT_TOKEN_MPSAFE : 0;
529 lwkt_token_uninit(lwkt_token_t tok)
536 lwkt_token_is_stale(lwkt_tokref_t ref)
538 lwkt_token_t tok = ref->tr_tok;
540 KKASSERT(tok->t_owner == curthread && ref->tr_state == 1 &&
543 /* Token is not stale */
544 if (tok->t_lastowner == tok->t_owner)
548 * The token is stale. Reset to not stale so that the next call to
549 * lwkt_token_is_stale will return "not stale" unless the token
550 * was acquired in-between by another thread.
552 tok->t_lastowner = tok->t_owner;