xref: /dflybsd-src/sys/kern/kern_spinlock.c (revision 0d9899e72d0ab73653a47790ac79317a20a1601c)
1 /*
2  * Copyright (c) 2005 Jeffrey M. Hsu.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Jeffrey M. Hsu. and Matthew Dillon
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of The DragonFly Project nor the names of its
16  *    contributors may be used to endorse or promote products derived
17  *    from this software without specific, prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
21  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
22  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
23  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
24  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
25  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
26  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
27  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
28  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
29  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 /*
34  * The implementation is designed to avoid looping when compatible operations
35  * are executed.
36  *
37  * To acquire a spinlock we first increment counta.  Then we check if counta
38  * meets our requirements.  For an exclusive spinlock it must be 1, of a
39  * shared spinlock it must either be 1 or the SHARED_SPINLOCK bit must be set.
40  *
41  * Shared spinlock failure case: Decrement the count, loop until we can
42  * transition from 0 to SHARED_SPINLOCK|1, or until we find SHARED_SPINLOCK
43  * is set and increment the count.
44  *
45  * Exclusive spinlock failure case: While maintaining the count, clear the
46  * SHARED_SPINLOCK flag unconditionally.  Then use an atomic add to transfer
47  * the count from the low bits to the high bits of counta.  Then loop until
48  * all low bits are 0.  Once the low bits drop to 0 we can transfer the
49  * count back with an atomic_cmpset_int(), atomically, and return.
50  */
51 #include <sys/param.h>
52 #include <sys/systm.h>
53 #include <sys/types.h>
54 #include <sys/kernel.h>
55 #include <sys/sysctl.h>
56 #ifdef INVARIANTS
57 #include <sys/proc.h>
58 #endif
59 #include <sys/priv.h>
60 #include <machine/atomic.h>
61 #include <machine/cpu.h>
62 #include <machine/cpufunc.h>
63 #include <machine/specialreg.h>
64 #include <machine/clock.h>
65 #include <sys/spinlock.h>
66 #include <sys/spinlock2.h>
67 #include <sys/ktr.h>
68 
69 #ifdef _KERNEL_VIRTUAL
70 #include <pthread.h>
71 #endif
72 
73 struct spinlock pmap_spin = SPINLOCK_INITIALIZER(pmap_spin, "pmap_spin");
74 
75 struct indefinite_info {
76 	sysclock_t	base;
77 	int		secs;
78 	const char	*ident;
79 };
80 
81 /*
82  * Kernal Trace
83  */
84 #if !defined(KTR_SPIN_CONTENTION)
85 #define KTR_SPIN_CONTENTION	KTR_ALL
86 #endif
87 #define SPIN_STRING	"spin=%p type=%c"
88 #define SPIN_ARG_SIZE	(sizeof(void *) + sizeof(int))
89 
90 KTR_INFO_MASTER(spin);
91 #if 0
92 KTR_INFO(KTR_SPIN_CONTENTION, spin, beg, 0, SPIN_STRING, SPIN_ARG_SIZE);
93 KTR_INFO(KTR_SPIN_CONTENTION, spin, end, 1, SPIN_STRING, SPIN_ARG_SIZE);
94 #endif
95 
96 #define logspin(name, spin, type)			\
97 	KTR_LOG(spin_ ## name, spin, type)
98 
99 #ifdef INVARIANTS
100 static int spin_lock_test_mode;
101 #endif
102 
103 #ifdef DEBUG_LOCKS_LATENCY
104 
105 static long spinlocks_add_latency;
106 SYSCTL_LONG(_debug, OID_AUTO, spinlocks_add_latency, CTLFLAG_RW,
107     &spinlocks_add_latency, 0,
108     "Add spinlock latency");
109 
110 #endif
111 
112 static int spin_indefinite_check(struct spinlock *spin,
113 				  struct indefinite_info *info);
114 
115 /*
116  * We contested due to another exclusive lock holder.  We lose.
117  *
118  * We have to unwind the attempt and may acquire the spinlock
119  * anyway while doing so.
120  */
121 int
122 spin_trylock_contested(struct spinlock *spin)
123 {
124 	globaldata_t gd = mycpu;
125 
126 	/*
127 	 * Handle degenerate case, else fail.
128 	 */
129 	if (atomic_cmpset_int(&spin->counta, SPINLOCK_SHARED|0, 1))
130 		return TRUE;
131 	/*atomic_add_int(&spin->counta, -1);*/
132 	--gd->gd_spinlocks;
133 	--gd->gd_curthread->td_critcount;
134 	return (FALSE);
135 }
136 
137 /*
138  * The spin_lock() inline was unable to acquire the lock and calls this
139  * function with spin->counta already incremented.
140  *
141  * atomic_swap_int() is the absolute fastest spinlock instruction, at
142  * least on multi-socket systems.  All instructions seem to be about
143  * the same on single-socket multi-core systems.  However, atomic_swap_int()
144  * does not result in an even distribution of successful acquisitions.
145  *
146  * UNFORTUNATELY we cannot really use atomic_swap_int() when also implementing
147  * shared spin locks, so as we do a better job removing contention we've
148  * moved to atomic_cmpset_int() to be able handle multiple states.
149  *
150  * Another problem we have is that (at least on the 48-core opteron we test
151  * with) having all 48 cores contesting the same spin lock reduces
152  * performance to around 600,000 ops/sec, verses millions when fewer cores
153  * are going after the same lock.
154  *
155  * Backoff algorithms can create even worse starvation problems, and don't
156  * really improve performance when a lot of cores are contending.
157  *
158  * Our solution is to allow the data cache to lazy-update by reading it
159  * non-atomically and only attempting to acquire the lock if the lazy read
160  * looks good.  This effectively limits cache bus bandwidth.  A cpu_pause()
161  * (for intel/amd anyhow) is not strictly needed as cache bus resource use
162  * is governed by the lazy update.
163  *
164  * WARNING!!!!  Performance matters here, by a huge margin.
165  *
166  *	48-core test with pre-read / -j 48 no-modules kernel compile
167  *	with fanned-out inactive and active queues came in at 55 seconds.
168  *
169  *	48-core test with pre-read / -j 48 no-modules kernel compile
170  *	came in at 75 seconds.  Without pre-read it came in at 170 seconds.
171  *
172  *	4-core test with pre-read / -j 48 no-modules kernel compile
173  *	came in at 83 seconds.  Without pre-read it came in at 83 seconds
174  *	as well (no difference).
175  */
176 void
177 _spin_lock_contested(struct spinlock *spin, const char *ident, int value)
178 {
179 	struct indefinite_info info = { 0, 0, ident };
180 	int i;
181 
182 	/*
183 	 * Handle degenerate case.
184 	 */
185 	if (value == SPINLOCK_SHARED) {
186 		if (atomic_cmpset_int(&spin->counta, SPINLOCK_SHARED|0, 1))
187 			return;
188 	}
189 
190 	/*
191 	 * Transfer our count to the high bits, then loop until we can
192 	 * acquire the low counter (== 1).  No new shared lock can be
193 	 * acquired while we hold the EXCLWAIT bits.
194 	 *
195 	 * Force any existing shared locks to exclusive.  The shared unlock
196 	 * understands that this may occur.
197 	 */
198 	atomic_add_int(&spin->counta, SPINLOCK_EXCLWAIT - 1);
199 	if (value & SPINLOCK_SHARED)
200 		atomic_clear_int(&spin->counta, SPINLOCK_SHARED);
201 
202 #ifdef DEBUG_LOCKS_LATENCY
203 	long j;
204 	for (j = spinlocks_add_latency; j > 0; --j)
205 		cpu_ccfence();
206 #endif
207 	i = 0;
208 
209 	/*logspin(beg, spin, 'w');*/
210 	for (;;) {
211 		/*
212 		 * If the low bits are zero, try to acquire the exclusive lock
213 		 * by transfering our high bit counter to the low bits.
214 		 *
215 		 * NOTE: Reading spin->counta prior to the swap is extremely
216 		 *	 important on multi-chip/many-core boxes.  On 48-core
217 		 *	 this one change improves fully concurrent all-cores
218 		 *	 compiles by 100% or better.
219 		 *
220 		 *	 I can't emphasize enough how important the pre-read
221 		 *	 is in preventing hw cache bus armageddon on
222 		 *	 multi-chip systems.  And on single-chip/multi-core
223 		 *	 systems it just doesn't hurt.
224 		 */
225 		uint32_t ovalue = spin->counta;
226 		cpu_ccfence();
227 		if ((ovalue & (SPINLOCK_EXCLWAIT - 1)) == 0 &&
228 		    atomic_cmpset_int(&spin->counta, ovalue,
229 				      (ovalue - SPINLOCK_EXCLWAIT) | 1)) {
230 			break;
231 		}
232 		if ((++i & 0x7F) == 0x7F) {
233 			mycpu->gd_cnt.v_lock_name[0] = 'X';
234 			strncpy(mycpu->gd_cnt.v_lock_name + 1,
235 				ident,
236 				sizeof(mycpu->gd_cnt.v_lock_name) - 2);
237 			++mycpu->gd_cnt.v_lock_colls;
238 			if (spin_indefinite_check(spin, &info))
239 				break;
240 		}
241 #ifdef _KERNEL_VIRTUAL
242 		pthread_yield();
243 #endif
244 	}
245 	/*logspin(end, spin, 'w');*/
246 }
247 
248 /*
249  * Shared spinlock attempt was contested.
250  *
251  * The caller has not modified counta.
252  */
253 void
254 _spin_lock_shared_contested(struct spinlock *spin, const char *ident, int value)
255 {
256 	struct indefinite_info info = { 0, 0, ident };
257 	int i;
258 
259 #ifdef DEBUG_LOCKS_LATENCY
260 	long j;
261 	for (j = spinlocks_add_latency; j > 0; --j)
262 		cpu_ccfence();
263 #endif
264 	i = 0;
265 
266 	/*logspin(beg, spin, 'w');*/
267 	for (;;) {
268 		/*
269 		 * Loop until we can acquire the shared spinlock.  Note that
270 		 * the low bits can be zero while the high EXCLWAIT bits are
271 		 * non-zero.  In this situation exclusive requesters have
272 		 * priority (otherwise shared users on multiple cpus can hog
273 		 * the spinlnock).
274 		 *
275 		 * NOTE: Reading spin->counta prior to the swap is extremely
276 		 *	 important on multi-chip/many-core boxes.  On 48-core
277 		 *	 this one change improves fully concurrent all-cores
278 		 *	 compiles by 100% or better.
279 		 *
280 		 *	 I can't emphasize enough how important the pre-read
281 		 *	 is in preventing hw cache bus armageddon on
282 		 *	 multi-chip systems.  And on single-chip/multi-core
283 		 *	 systems it just doesn't hurt.
284 		 */
285 		uint32_t ovalue = spin->counta;
286 
287 		cpu_ccfence();
288 		if (ovalue == 0) {
289 			if (atomic_cmpset_int(&spin->counta, 0,
290 					      SPINLOCK_SHARED | 1))
291 				break;
292 		} else if (ovalue & SPINLOCK_SHARED) {
293 			if (atomic_cmpset_int(&spin->counta, ovalue,
294 					      ovalue + 1))
295 				break;
296 		}
297 		if ((++i & 0x7F) == 0x7F) {
298 			mycpu->gd_cnt.v_lock_name[0] = 'S';
299 			strncpy(mycpu->gd_cnt.v_lock_name + 1,
300 				ident,
301 				sizeof(mycpu->gd_cnt.v_lock_name) - 2);
302 			++mycpu->gd_cnt.v_lock_colls;
303 			if (spin_indefinite_check(spin, &info))
304 				break;
305 		}
306 #ifdef _KERNEL_VIRTUAL
307 		pthread_yield();
308 #endif
309 	}
310 	/*logspin(end, spin, 'w');*/
311 }
312 
313 static
314 int
315 spin_indefinite_check(struct spinlock *spin, struct indefinite_info *info)
316 {
317 	sysclock_t count;
318 
319 	cpu_spinlock_contested();
320 
321 	count = sys_cputimer->count();
322 	if (info->secs == 0) {
323 		info->base = count;
324 		++info->secs;
325 	} else if (count - info->base > sys_cputimer->freq) {
326 		kprintf("spin_lock: %s(%p), indefinite wait (%d secs)!\n",
327 			info->ident, spin, info->secs);
328 		info->base = count;
329 		++info->secs;
330 		if (panicstr)
331 			return (TRUE);
332 #if defined(INVARIANTS)
333 		if (spin_lock_test_mode) {
334 			print_backtrace(-1);
335 			return (TRUE);
336 		}
337 #endif
338 #if defined(INVARIANTS)
339 		if (info->secs == 11)
340 			print_backtrace(-1);
341 #endif
342 		if (info->secs == 60)
343 			panic("spin_lock: %s(%p), indefinite wait!",
344 			      info->ident, spin);
345 	}
346 	return (FALSE);
347 }
348 
349 /*
350  * If INVARIANTS is enabled various spinlock timing tests can be run
351  * by setting debug.spin_lock_test:
352  *
353  *	1	Test the indefinite wait code
354  *	2	Time the best-case exclusive lock overhead (spin_test_count)
355  *	3	Time the best-case shared lock overhead (spin_test_count)
356  */
357 
358 #ifdef INVARIANTS
359 
360 static int spin_test_count = 10000000;
361 SYSCTL_INT(_debug, OID_AUTO, spin_test_count, CTLFLAG_RW, &spin_test_count, 0,
362     "Number of iterations to use for spinlock wait code test");
363 
364 static int
365 sysctl_spin_lock_test(SYSCTL_HANDLER_ARGS)
366 {
367         struct spinlock spin;
368 	int error;
369 	int value = 0;
370 	int i;
371 
372 	if ((error = priv_check(curthread, PRIV_ROOT)) != 0)
373 		return (error);
374 	if ((error = SYSCTL_IN(req, &value, sizeof(value))) != 0)
375 		return (error);
376 
377 	/*
378 	 * Indefinite wait test
379 	 */
380 	if (value == 1) {
381 		spin_init(&spin, "sysctllock");
382 		spin_lock(&spin);	/* force an indefinite wait */
383 		spin_lock_test_mode = 1;
384 		spin_lock(&spin);
385 		spin_unlock(&spin);	/* Clean up the spinlock count */
386 		spin_unlock(&spin);
387 		spin_lock_test_mode = 0;
388 	}
389 
390 	/*
391 	 * Time best-case exclusive spinlocks
392 	 */
393 	if (value == 2) {
394 		globaldata_t gd = mycpu;
395 
396 		spin_init(&spin, "sysctllocktest");
397 		for (i = spin_test_count; i > 0; --i) {
398 		    _spin_lock_quick(gd, &spin, "test");
399 		    spin_unlock_quick(gd, &spin);
400 		}
401 	}
402 
403         return (0);
404 }
405 
406 SYSCTL_PROC(_debug, KERN_PROC_ALL, spin_lock_test, CTLFLAG_RW|CTLTYPE_INT,
407         0, 0, sysctl_spin_lock_test, "I", "Test spinlock wait code");
408 
409 #endif	/* INVARIANTS */
410