xref: /freebsd/sys/kern/kern_pmc.c (revision 685dc743)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2003-2008 Joseph Koshy
5  * Copyright (c) 2007 The FreeBSD Foundation
6  * All rights reserved.
7  *
8  * Portions of this software were developed by A. Joseph Koshy under
9  * sponsorship from the FreeBSD Foundation and Google, Inc.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #include <sys/cdefs.h>
34 #include "opt_hwpmc_hooks.h"
35 
36 #include <sys/param.h>
37 #include <sys/ctype.h>
38 #include <sys/domainset.h>
39 #include <sys/param.h>
40 #include <sys/malloc.h>
41 #include <sys/kernel.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/pmc.h>
45 #include <sys/pmckern.h>
46 #include <sys/smp.h>
47 #include <sys/sysctl.h>
48 #include <sys/systm.h>
49 
50 #include <vm/vm.h>
51 #include <vm/vm_extern.h>
52 #include <vm/vm_kern.h>
53 
54 #ifdef	HWPMC_HOOKS
55 FEATURE(hwpmc_hooks, "Kernel support for HW PMC");
56 #define	PMC_KERNEL_VERSION	PMC_VERSION
57 #else
58 #define	PMC_KERNEL_VERSION	0
59 #endif
60 
61 MALLOC_DECLARE(M_PMCHOOKS);
62 MALLOC_DEFINE(M_PMCHOOKS, "pmchooks", "Memory space for PMC hooks");
63 
64 /* memory pool */
65 MALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module");
66 
67 const int pmc_kernel_version = PMC_KERNEL_VERSION;
68 
69 /* Hook variable. */
70 int __read_mostly (*pmc_hook)(struct thread *td, int function, void *arg) = NULL;
71 
72 /* Interrupt handler */
73 int __read_mostly (*pmc_intr)(struct trapframe *tf) = NULL;
74 
75 DPCPU_DEFINE(uint8_t, pmc_sampled);
76 
77 /*
78  * A global count of SS mode PMCs.  When non-zero, this means that
79  * we have processes that are sampling the system as a whole.
80  */
81 volatile int pmc_ss_count;
82 
83 /*
84  * Since PMC(4) may not be loaded in the current kernel, the
85  * convention followed is that a non-NULL value of 'pmc_hook' implies
86  * the presence of this kernel module.
87  *
88  * This requires us to protect 'pmc_hook' with a
89  * shared (sx) lock -- thus making the process of calling into PMC(4)
90  * somewhat more expensive than a simple 'if' check and indirect call.
91  */
92 struct sx pmc_sx;
93 SX_SYSINIT(pmcsx, &pmc_sx, "pmc-sx");
94 
95 /*
96  * PMC Soft per cpu trapframe.
97  */
98 struct trapframe pmc_tf[MAXCPU];
99 
100 /*
101  * Per domain list of buffer headers
102  */
103 __read_mostly struct pmc_domain_buffer_header *pmc_dom_hdrs[MAXMEMDOM];
104 
105 /*
106  * PMC Soft use a global table to store registered events.
107  */
108 
109 SYSCTL_NODE(_kern, OID_AUTO, hwpmc, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
110     "HWPMC parameters");
111 
112 static int pmc_softevents = 16;
113 SYSCTL_INT(_kern_hwpmc, OID_AUTO, softevents, CTLFLAG_RDTUN,
114     &pmc_softevents, 0, "maximum number of soft events");
115 
116 int pmc_softs_count;
117 struct pmc_soft **pmc_softs;
118 
119 struct mtx pmc_softs_mtx;
120 MTX_SYSINIT(pmc_soft_mtx, &pmc_softs_mtx, "pmc-softs", MTX_SPIN);
121 
122 /*
123  * Helper functions.
124  */
125 
126 /*
127  * A note on the CPU numbering scheme used by the hwpmc(4) driver.
128  *
129  * CPUs are denoted using numbers in the range 0..[pmc_cpu_max()-1].
130  * CPUs could be numbered "sparsely" in this range; the predicate
131  * `pmc_cpu_is_present()' is used to test whether a given CPU is
132  * physically present.
133  *
134  * Further, a CPU that is physically present may be administratively
135  * disabled or otherwise unavailable for use by hwpmc(4).  The
136  * `pmc_cpu_is_active()' predicate tests for CPU usability.  An
137  * "active" CPU participates in thread scheduling and can field
138  * interrupts raised by PMC hardware.
139  *
140  * On systems with hyperthreaded CPUs, multiple logical CPUs may share
141  * PMC hardware resources.  For such processors one logical CPU is
142  * denoted as the primary owner of the in-CPU PMC resources. The
143  * pmc_cpu_is_primary() predicate is used to distinguish this primary
144  * CPU from the others.
145  */
146 
147 int
pmc_cpu_is_active(int cpu)148 pmc_cpu_is_active(int cpu)
149 {
150 #ifdef	SMP
151 	return (pmc_cpu_is_present(cpu) &&
152 	    !CPU_ISSET(cpu, &hlt_cpus_mask));
153 #else
154 	return (1);
155 #endif
156 }
157 
158 /* Deprecated. */
159 int
pmc_cpu_is_disabled(int cpu)160 pmc_cpu_is_disabled(int cpu)
161 {
162 	return (!pmc_cpu_is_active(cpu));
163 }
164 
165 int
pmc_cpu_is_present(int cpu)166 pmc_cpu_is_present(int cpu)
167 {
168 #ifdef	SMP
169 	return (!CPU_ABSENT(cpu));
170 #else
171 	return (1);
172 #endif
173 }
174 
175 int
pmc_cpu_is_primary(int cpu)176 pmc_cpu_is_primary(int cpu)
177 {
178 #ifdef	SMP
179 	return (!CPU_ISSET(cpu, &logical_cpus_mask));
180 #else
181 	return (1);
182 #endif
183 }
184 
185 /*
186  * Return the maximum CPU number supported by the system.  The return
187  * value is used for scaling internal data structures and for runtime
188  * checks.
189  */
190 unsigned int
pmc_cpu_max(void)191 pmc_cpu_max(void)
192 {
193 #ifdef	SMP
194 	return (mp_maxid+1);
195 #else
196 	return (1);
197 #endif
198 }
199 
200 #ifdef	INVARIANTS
201 
202 /*
203  * Return the count of CPUs in the `active' state in the system.
204  */
205 int
pmc_cpu_max_active(void)206 pmc_cpu_max_active(void)
207 {
208 #ifdef	SMP
209 	/*
210 	 * When support for CPU hot-plugging is added to the kernel,
211 	 * this function would change to return the current number
212 	 * of "active" CPUs.
213 	 */
214 	return (mp_ncpus);
215 #else
216 	return (1);
217 #endif
218 }
219 
220 #endif
221 
222 /*
223  * Cleanup event name:
224  * - remove duplicate '_'
225  * - all uppercase
226  */
227 static void
pmc_soft_namecleanup(char * name)228 pmc_soft_namecleanup(char *name)
229 {
230 	char *p, *q;
231 
232 	p = q = name;
233 
234 	for ( ; *p == '_' ; p++)
235 		;
236 	for ( ; *p ; p++) {
237 		if (*p == '_' && (*(p + 1) == '_' || *(p + 1) == '\0'))
238 			continue;
239 		else
240 			*q++ = toupper(*p);
241 	}
242 	*q = '\0';
243 }
244 
245 void
pmc_soft_ev_register(struct pmc_soft * ps)246 pmc_soft_ev_register(struct pmc_soft *ps)
247 {
248 	static int warned = 0;
249 	int n;
250 
251 	ps->ps_running  = 0;
252 	ps->ps_ev.pm_ev_code = 0; /* invalid */
253 	pmc_soft_namecleanup(ps->ps_ev.pm_ev_name);
254 
255 	mtx_lock_spin(&pmc_softs_mtx);
256 
257 	if (pmc_softs_count >= pmc_softevents) {
258 		/*
259 		 * XXX Reusing events can enter a race condition where
260 		 * new allocated event will be used as an old one.
261 		 */
262 		for (n = 0; n < pmc_softevents; n++)
263 			if (pmc_softs[n] == NULL)
264 				break;
265 		if (n == pmc_softevents) {
266 			mtx_unlock_spin(&pmc_softs_mtx);
267 			if (!warned) {
268 				printf("hwpmc: too many soft events, "
269 				    "increase kern.hwpmc.softevents tunable\n");
270 				warned = 1;
271 			}
272 			return;
273 		}
274 
275 		ps->ps_ev.pm_ev_code = PMC_EV_SOFT_FIRST + n;
276 		pmc_softs[n] = ps;
277 	} else {
278 		ps->ps_ev.pm_ev_code = PMC_EV_SOFT_FIRST + pmc_softs_count;
279 		pmc_softs[pmc_softs_count++] = ps;
280 	}
281 
282 	mtx_unlock_spin(&pmc_softs_mtx);
283 }
284 
285 void
pmc_soft_ev_deregister(struct pmc_soft * ps)286 pmc_soft_ev_deregister(struct pmc_soft *ps)
287 {
288 
289 	KASSERT(ps != NULL, ("pmc_soft_deregister: called with NULL"));
290 
291 	mtx_lock_spin(&pmc_softs_mtx);
292 
293 	if (ps->ps_ev.pm_ev_code != 0 &&
294 	    (ps->ps_ev.pm_ev_code - PMC_EV_SOFT_FIRST) < pmc_softevents) {
295 		KASSERT((int)ps->ps_ev.pm_ev_code >= PMC_EV_SOFT_FIRST &&
296 		    (int)ps->ps_ev.pm_ev_code <= PMC_EV_SOFT_LAST,
297 		    ("pmc_soft_deregister: invalid event value"));
298 		pmc_softs[ps->ps_ev.pm_ev_code - PMC_EV_SOFT_FIRST] = NULL;
299 	}
300 
301 	mtx_unlock_spin(&pmc_softs_mtx);
302 }
303 
304 struct pmc_soft *
pmc_soft_ev_acquire(enum pmc_event ev)305 pmc_soft_ev_acquire(enum pmc_event ev)
306 {
307 	struct pmc_soft *ps;
308 
309 	if (ev == 0 || (ev - PMC_EV_SOFT_FIRST) >= pmc_softevents)
310 		return NULL;
311 
312 	KASSERT((int)ev >= PMC_EV_SOFT_FIRST &&
313 	    (int)ev <= PMC_EV_SOFT_LAST,
314 	    ("event out of range"));
315 
316 	mtx_lock_spin(&pmc_softs_mtx);
317 
318 	ps = pmc_softs[ev - PMC_EV_SOFT_FIRST];
319 	if (ps == NULL)
320 		mtx_unlock_spin(&pmc_softs_mtx);
321 
322 	return ps;
323 }
324 
325 void
pmc_soft_ev_release(struct pmc_soft * ps)326 pmc_soft_ev_release(struct pmc_soft *ps)
327 {
328 
329 	mtx_unlock_spin(&pmc_softs_mtx);
330 }
331 
332 /*
333  *  Initialise hwpmc.
334  */
335 static void
init_hwpmc(void * dummy __unused)336 init_hwpmc(void *dummy __unused)
337 {
338 	int domain, cpu;
339 
340 	if (pmc_softevents <= 0 ||
341 	    pmc_softevents > PMC_EV_DYN_COUNT) {
342 		(void) printf("hwpmc: tunable \"softevents\"=%d out of "
343 		    "range.\n", pmc_softevents);
344 		pmc_softevents = PMC_EV_DYN_COUNT;
345 	}
346 	pmc_softs = malloc(pmc_softevents * sizeof(*pmc_softs), M_PMCHOOKS,
347 	    M_WAITOK | M_ZERO);
348 
349 	for (domain = 0; domain < vm_ndomains; domain++) {
350 		pmc_dom_hdrs[domain] = malloc_domainset(
351 		    sizeof(struct pmc_domain_buffer_header), M_PMC,
352 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
353 		mtx_init(&pmc_dom_hdrs[domain]->pdbh_mtx, "pmc_bufferlist_mtx", "pmc-leaf", MTX_SPIN);
354 		TAILQ_INIT(&pmc_dom_hdrs[domain]->pdbh_head);
355 	}
356 	CPU_FOREACH(cpu) {
357 		domain = pcpu_find(cpu)->pc_domain;
358 		KASSERT(pmc_dom_hdrs[domain] != NULL, ("no mem allocated for domain: %d", domain));
359 		pmc_dom_hdrs[domain]->pdbh_ncpus++;
360 	}
361 
362 }
363 
364 SYSINIT(hwpmc, SI_SUB_KDTRACE, SI_ORDER_FIRST, init_hwpmc, NULL);
365