xref: /titanic_41/usr/src/uts/common/os/dtrace_subr.c (revision 8d62b2231107e557adb17ea75579d9c49511457b)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #include <sys/dtrace.h>
28 #include <sys/cmn_err.h>
29 #include <sys/tnf.h>
30 #include <sys/atomic.h>
31 #include <sys/prsystm.h>
32 #include <sys/modctl.h>
33 #include <sys/aio_impl.h>
34 
35 #ifdef __sparc
36 #include <sys/privregs.h>
37 #endif
38 
39 void (*dtrace_cpu_init)(processorid_t);
40 void (*dtrace_modload)(struct modctl *);
41 void (*dtrace_modunload)(struct modctl *);
42 void (*dtrace_helpers_cleanup)(void);
43 void (*dtrace_helpers_fork)(proc_t *, proc_t *);
44 void (*dtrace_cpustart_init)(void);
45 void (*dtrace_cpustart_fini)(void);
46 void (*dtrace_cpc_fire)(uint64_t);
47 void (*dtrace_closef)(void);
48 
49 void (*dtrace_debugger_init)(void);
50 void (*dtrace_debugger_fini)(void);
51 
52 dtrace_vtime_state_t dtrace_vtime_active = 0;
53 dtrace_cacheid_t dtrace_predcache_id = DTRACE_CACHEIDNONE + 1;
54 
55 /*
56  * dtrace_cpc_in_use usage statement: this global variable is used by the cpc
57  * hardware overflow interrupt handler and the kernel cpc framework to check
58  * whether or not the DTrace cpc provider is currently in use. The variable is
59  * set before counters are enabled with the first enabling and cleared when
60  * the last enabling is disabled. Its value at any given time indicates the
61  * number of active dcpc based enablings. The global 'kcpc_cpuctx_lock' rwlock
62  * is held during initial setting to protect races between kcpc_open() and the
63  * first enabling. The locking provided by the DTrace subsystem, the kernel
64  * cpc framework and the cpu management framework protect consumers from race
65  * conditions on enabling and disabling probes.
66  */
67 uint32_t dtrace_cpc_in_use = 0;
68 
69 typedef struct dtrace_hrestime {
70 	lock_t		dthr_lock;		/* lock for this element */
71 	timestruc_t	dthr_hrestime;		/* hrestime value */
72 	int64_t		dthr_adj;		/* hrestime_adj value */
73 	hrtime_t	dthr_hrtime;		/* hrtime value */
74 } dtrace_hrestime_t;
75 
76 static dtrace_hrestime_t dtrace_hrestime[2];
77 
78 /*
79  * Making available adjustable high-resolution time in DTrace is regrettably
80  * more complicated than one might think it should be.  The problem is that
81  * the variables related to adjusted high-resolution time (hrestime,
82  * hrestime_adj and friends) are adjusted under hres_lock -- and this lock may
83  * be held when we enter probe context.  One might think that we could address
84  * this by having a single snapshot copy that is stored under a different lock
85  * from hres_tick(), using the snapshot iff hres_lock is locked in probe
86  * context.  Unfortunately, this too won't work:  because hres_lock is grabbed
87  * in more than just hres_tick() context, we could enter probe context
88  * concurrently on two different CPUs with both locks (hres_lock and the
89  * snapshot lock) held.  As this implies, the fundamental problem is that we
90  * need to have access to a snapshot of these variables that we _know_ will
91  * not be locked in probe context.  To effect this, we have two snapshots
92  * protected by two different locks, and we mandate that these snapshots are
93  * recorded in succession by a single thread calling dtrace_hres_tick().  (We
94  * assure this by calling it out of the same CY_HIGH_LEVEL cyclic that calls
95  * hres_tick().)  A single thread can't be in two places at once:  one of the
96  * snapshot locks is guaranteed to be unheld at all times.  The
97  * dtrace_gethrestime() algorithm is thus to check first one snapshot and then
98  * the other to find the unlocked snapshot.
99  */
100 void
101 dtrace_hres_tick(void)
102 {
103 	int i;
104 	ushort_t spl;
105 
106 	for (i = 0; i < 2; i++) {
107 		dtrace_hrestime_t tmp;
108 
109 		spl = hr_clock_lock();
110 		tmp.dthr_hrestime = hrestime;
111 		tmp.dthr_adj = hrestime_adj;
112 		tmp.dthr_hrtime = dtrace_gethrtime();
113 		hr_clock_unlock(spl);
114 
115 		lock_set(&dtrace_hrestime[i].dthr_lock);
116 		dtrace_hrestime[i].dthr_hrestime = tmp.dthr_hrestime;
117 		dtrace_hrestime[i].dthr_adj = tmp.dthr_adj;
118 		dtrace_hrestime[i].dthr_hrtime = tmp.dthr_hrtime;
119 		dtrace_membar_producer();
120 
121 		/*
122 		 * To allow for lock-free examination of this lock, we use
123 		 * the same trick that is used hres_lock; for more details,
124 		 * see the description of this technique in sun4u/sys/clock.h.
125 		 */
126 		dtrace_hrestime[i].dthr_lock++;
127 	}
128 }
129 
130 hrtime_t
131 dtrace_gethrestime(void)
132 {
133 	dtrace_hrestime_t snap;
134 	hrtime_t now;
135 	int i = 0, adj, nslt;
136 
137 	for (;;) {
138 		snap.dthr_lock = dtrace_hrestime[i].dthr_lock;
139 		dtrace_membar_consumer();
140 		snap.dthr_hrestime = dtrace_hrestime[i].dthr_hrestime;
141 		snap.dthr_hrtime = dtrace_hrestime[i].dthr_hrtime;
142 		snap.dthr_adj = dtrace_hrestime[i].dthr_adj;
143 		dtrace_membar_consumer();
144 
145 		if ((snap.dthr_lock & ~1) == dtrace_hrestime[i].dthr_lock)
146 			break;
147 
148 		/*
149 		 * If we're here, the lock was either locked, or it
150 		 * transitioned while we were taking the snapshot.  Either
151 		 * way, we're going to try the other dtrace_hrestime element;
152 		 * we know that it isn't possible for both to be locked
153 		 * simultaneously, so we will ultimately get a good snapshot.
154 		 */
155 		i ^= 1;
156 	}
157 
158 	/*
159 	 * We have a good snapshot.  Now perform any necessary adjustments.
160 	 */
161 	nslt = dtrace_gethrtime() - snap.dthr_hrtime;
162 	ASSERT(nslt >= 0);
163 
164 	now = ((hrtime_t)snap.dthr_hrestime.tv_sec * (hrtime_t)NANOSEC) +
165 	    snap.dthr_hrestime.tv_nsec;
166 
167 	if (snap.dthr_adj != 0) {
168 		if (snap.dthr_adj > 0) {
169 			adj = (nslt >> adj_shift);
170 			if (adj > snap.dthr_adj)
171 				adj = (int)snap.dthr_adj;
172 		} else {
173 			adj = -(nslt >> adj_shift);
174 			if (adj < snap.dthr_adj)
175 				adj = (int)snap.dthr_adj;
176 		}
177 		now += adj;
178 	}
179 
180 	return (now);
181 }
182 
183 void
184 dtrace_vtime_enable(void)
185 {
186 	dtrace_vtime_state_t state, nstate;
187 
188 	do {
189 		state = dtrace_vtime_active;
190 
191 		switch (state) {
192 		case DTRACE_VTIME_INACTIVE:
193 			nstate = DTRACE_VTIME_ACTIVE;
194 			break;
195 
196 		case DTRACE_VTIME_INACTIVE_TNF:
197 			nstate = DTRACE_VTIME_ACTIVE_TNF;
198 			break;
199 
200 		case DTRACE_VTIME_ACTIVE:
201 		case DTRACE_VTIME_ACTIVE_TNF:
202 			panic("DTrace virtual time already enabled");
203 			/*NOTREACHED*/
204 		}
205 
206 	} while	(atomic_cas_32((uint32_t *)&dtrace_vtime_active,
207 	    state, nstate) != state);
208 }
209 
210 void
211 dtrace_vtime_disable(void)
212 {
213 	dtrace_vtime_state_t state, nstate;
214 
215 	do {
216 		state = dtrace_vtime_active;
217 
218 		switch (state) {
219 		case DTRACE_VTIME_ACTIVE:
220 			nstate = DTRACE_VTIME_INACTIVE;
221 			break;
222 
223 		case DTRACE_VTIME_ACTIVE_TNF:
224 			nstate = DTRACE_VTIME_INACTIVE_TNF;
225 			break;
226 
227 		case DTRACE_VTIME_INACTIVE:
228 		case DTRACE_VTIME_INACTIVE_TNF:
229 			panic("DTrace virtual time already disabled");
230 			/*NOTREACHED*/
231 		}
232 
233 	} while	(atomic_cas_32((uint32_t *)&dtrace_vtime_active,
234 	    state, nstate) != state);
235 }
236 
237 void
238 dtrace_vtime_enable_tnf(void)
239 {
240 	dtrace_vtime_state_t state, nstate;
241 
242 	do {
243 		state = dtrace_vtime_active;
244 
245 		switch (state) {
246 		case DTRACE_VTIME_ACTIVE:
247 			nstate = DTRACE_VTIME_ACTIVE_TNF;
248 			break;
249 
250 		case DTRACE_VTIME_INACTIVE:
251 			nstate = DTRACE_VTIME_INACTIVE_TNF;
252 			break;
253 
254 		case DTRACE_VTIME_ACTIVE_TNF:
255 		case DTRACE_VTIME_INACTIVE_TNF:
256 			panic("TNF already active");
257 			/*NOTREACHED*/
258 		}
259 
260 	} while	(atomic_cas_32((uint32_t *)&dtrace_vtime_active,
261 	    state, nstate) != state);
262 }
263 
264 void
265 dtrace_vtime_disable_tnf(void)
266 {
267 	dtrace_vtime_state_t state, nstate;
268 
269 	do {
270 		state = dtrace_vtime_active;
271 
272 		switch (state) {
273 		case DTRACE_VTIME_ACTIVE_TNF:
274 			nstate = DTRACE_VTIME_ACTIVE;
275 			break;
276 
277 		case DTRACE_VTIME_INACTIVE_TNF:
278 			nstate = DTRACE_VTIME_INACTIVE;
279 			break;
280 
281 		case DTRACE_VTIME_ACTIVE:
282 		case DTRACE_VTIME_INACTIVE:
283 			panic("TNF already inactive");
284 			/*NOTREACHED*/
285 		}
286 
287 	} while	(atomic_cas_32((uint32_t *)&dtrace_vtime_active,
288 	    state, nstate) != state);
289 }
290 
291 void
292 dtrace_vtime_switch(kthread_t *next)
293 {
294 	dtrace_icookie_t cookie;
295 	hrtime_t ts;
296 
297 	if (tnf_tracing_active) {
298 		tnf_thread_switch(next);
299 
300 		if (dtrace_vtime_active == DTRACE_VTIME_INACTIVE_TNF)
301 			return;
302 	}
303 
304 	cookie = dtrace_interrupt_disable();
305 	ts = dtrace_gethrtime();
306 
307 	if (curthread->t_dtrace_start != 0) {
308 		curthread->t_dtrace_vtime += ts - curthread->t_dtrace_start;
309 		curthread->t_dtrace_start = 0;
310 	}
311 
312 	next->t_dtrace_start = ts;
313 
314 	dtrace_interrupt_enable(cookie);
315 }
316 
317 void (*dtrace_fasttrap_fork_ptr)(proc_t *, proc_t *);
318 void (*dtrace_fasttrap_exec_ptr)(proc_t *);
319 void (*dtrace_fasttrap_exit_ptr)(proc_t *);
320 
321 /*
322  * This function is called by cfork() in the event that it appears that
323  * there may be dtrace tracepoints active in the parent process's address
324  * space. This first confirms the existence of dtrace tracepoints in the
325  * parent process and calls into the fasttrap module to remove the
326  * corresponding tracepoints from the child. By knowing that there are
327  * existing tracepoints, and ensuring they can't be removed, we can rely
328  * on the fasttrap module remaining loaded.
329  */
330 void
331 dtrace_fasttrap_fork(proc_t *p, proc_t *cp)
332 {
333 	ASSERT(p->p_proc_flag & P_PR_LOCK);
334 	ASSERT(p->p_dtrace_count > 0);
335 	ASSERT(dtrace_fasttrap_fork_ptr != NULL);
336 
337 	dtrace_fasttrap_fork_ptr(p, cp);
338 }
339