perfmon.c revision 83366 
1/*
2 * Copyright 1996 Massachusetts Institute of Technology
3 *
4 * Permission to use, copy, modify, and distribute this software and
5 * its documentation for any purpose and without fee is hereby
6 * granted, provided that both the above copyright notice and this
7 * permission notice appear in all copies, that both the above
8 * copyright notice and this permission notice appear in all
9 * supporting documentation, and that the name of M.I.T. not be used
10 * in advertising or publicity pertaining to distribution of the
11 * software without specific, written prior permission.  M.I.T. makes
12 * no representations about the suitability of this software for any
13 * purpose.  It is provided "as is" without express or implied
14 * warranty.
15 *
16 * THIS SOFTWARE IS PROVIDED BY M.I.T. ``AS IS''.  M.I.T. DISCLAIMS
17 * ALL EXPRESS OR IMPLIED WARRANTIES WITH REGARD TO THIS SOFTWARE,
18 * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
19 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. IN NO EVENT
20 * SHALL M.I.T. BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
22 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF
23 * USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
24 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
25 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
26 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 *
29 * $FreeBSD: head/sys/i386/i386/perfmon.c 83366 2001-09-12 08:38:13Z julian $
30 */
31
32#include <sys/param.h>
33#include <sys/systm.h>
34#include <sys/conf.h>
35#include <sys/fcntl.h>
36
37#ifndef SMP
38#include <machine/cputypes.h>
39#endif
40#include <machine/clock.h>
41#include <machine/perfmon.h>
42
43static int perfmon_inuse;
44static int perfmon_cpuok;
45#ifndef SMP
46static int msr_ctl[NPMC];
47#endif
48static int msr_pmc[NPMC];
49static unsigned int ctl_shadow[NPMC];
50static quad_t pmc_shadow[NPMC];	/* used when ctr is stopped on P5 */
51static int (*writectl)(int);
52#ifndef SMP
53static int writectl5(int);
54static int writectl6(int);
55#endif
56
57static d_close_t perfmon_close;
58static d_open_t	perfmon_open;
59static d_ioctl_t perfmon_ioctl;
60
61#define CDEV_MAJOR 2	/* We're really a minor of mem.c */
62static struct cdevsw perfmon_cdevsw = {
63	/* open */      perfmon_open,
64	/* close */     perfmon_close,
65	/* read */      noread,
66	/* write */     nowrite,
67	/* ioctl */     perfmon_ioctl,
68	/* poll */      nopoll,
69	/* mmap */      nommap,
70	/* strategy */  nostrategy,
71	/* name */      "perfmon",
72	/* maj */       CDEV_MAJOR,
73	/* dump */      nodump,
74	/* psize */     nopsize,
75	/* flags */     0,
76};
77
78/*
79 * Must be called after cpu_class is set up.
80 */
81void
82perfmon_init(void)
83{
84#ifndef SMP
85	switch(cpu_class) {
86	case CPUCLASS_586:
87		perfmon_cpuok = 1;
88		msr_ctl[0] = 0x11;
89		msr_ctl[1] = 0x11;
90		msr_pmc[0] = 0x12;
91		msr_pmc[1] = 0x13;
92		writectl = writectl5;
93		break;
94	case CPUCLASS_686:
95		perfmon_cpuok = 1;
96		msr_ctl[0] = 0x186;
97		msr_ctl[1] = 0x187;
98		msr_pmc[0] = 0xc1;
99		msr_pmc[1] = 0xc2;
100		writectl = writectl6;
101		break;
102
103	default:
104		perfmon_cpuok = 0;
105		break;
106	}
107#endif /* SMP */
108	make_dev(&perfmon_cdevsw, 32, UID_ROOT, GID_KMEM, 0640, "perfmon");
109}
110
111int
112perfmon_avail(void)
113{
114	return perfmon_cpuok;
115}
116
117int
118perfmon_setup(int pmc, unsigned int control)
119{
120	critical_t	savecrit;
121
122	if (pmc < 0 || pmc >= NPMC)
123		return EINVAL;
124
125	perfmon_inuse |= (1 << pmc);
126	control &= ~(PMCF_SYS_FLAGS << 16);
127	savecrit = critical_enter();
128	ctl_shadow[pmc] = control;
129	writectl(pmc);
130	wrmsr(msr_pmc[pmc], pmc_shadow[pmc] = 0);
131	critical_exit(savecrit);
132	return 0;
133}
134
135int
136perfmon_get(int pmc, unsigned int *control)
137{
138	if (pmc < 0 || pmc >= NPMC)
139		return EINVAL;
140
141	if (perfmon_inuse & (1 << pmc)) {
142		*control = ctl_shadow[pmc];
143		return 0;
144	}
145	return EBUSY;		/* XXX reversed sense */
146}
147
148int
149perfmon_fini(int pmc)
150{
151	if (pmc < 0 || pmc >= NPMC)
152		return EINVAL;
153
154	if (perfmon_inuse & (1 << pmc)) {
155		perfmon_stop(pmc);
156		ctl_shadow[pmc] = 0;
157		perfmon_inuse &= ~(1 << pmc);
158		return 0;
159	}
160	return EBUSY;		/* XXX reversed sense */
161}
162
163int
164perfmon_start(int pmc)
165{
166	critical_t	savecrit;
167
168	if (pmc < 0 || pmc >= NPMC)
169		return EINVAL;
170
171	if (perfmon_inuse & (1 << pmc)) {
172		savecrit = critical_enter();
173		ctl_shadow[pmc] |= (PMCF_EN << 16);
174		wrmsr(msr_pmc[pmc], pmc_shadow[pmc]);
175		writectl(pmc);
176		critical_exit(savecrit);
177		return 0;
178	}
179	return EBUSY;
180}
181
182int
183perfmon_stop(int pmc)
184{
185	critical_t	savecrit;
186
187	if (pmc < 0 || pmc >= NPMC)
188		return EINVAL;
189
190	if (perfmon_inuse & (1 << pmc)) {
191		savecrit = critical_enter();
192		pmc_shadow[pmc] = rdmsr(msr_pmc[pmc]) & 0xffffffffffULL;
193		ctl_shadow[pmc] &= ~(PMCF_EN << 16);
194		writectl(pmc);
195		critical_exit(savecrit);
196		return 0;
197	}
198	return EBUSY;
199}
200
201int
202perfmon_read(int pmc, quad_t *val)
203{
204	if (pmc < 0 || pmc >= NPMC)
205		return EINVAL;
206
207	if (perfmon_inuse & (1 << pmc)) {
208		if (ctl_shadow[pmc] & (PMCF_EN << 16))
209			*val = rdmsr(msr_pmc[pmc]) & 0xffffffffffULL;
210		else
211			*val = pmc_shadow[pmc];
212		return 0;
213	}
214
215	return EBUSY;
216}
217
218int
219perfmon_reset(int pmc)
220{
221	if (pmc < 0 || pmc >= NPMC)
222		return EINVAL;
223
224	if (perfmon_inuse & (1 << pmc)) {
225		wrmsr(msr_pmc[pmc], pmc_shadow[pmc] = 0);
226		return 0;
227	}
228	return EBUSY;
229}
230
231#ifndef SMP
232/*
233 * Unfortunately, the performance-monitoring registers are laid out
234 * differently in the P5 and P6.  We keep everything in P6 format
235 * internally (except for the event code), and convert to P5
236 * format as needed on those CPUs.  The writectl function pointer
237 * is set up to point to one of these functions by perfmon_init().
238 */
239int
240writectl6(int pmc)
241{
242	if (pmc > 0 && !(ctl_shadow[pmc] & (PMCF_EN << 16))) {
243		wrmsr(msr_ctl[pmc], 0);
244	} else {
245		wrmsr(msr_ctl[pmc], ctl_shadow[pmc]);
246	}
247	return 0;
248}
249
250#define	P5FLAG_P	0x200
251#define	P5FLAG_E	0x100
252#define	P5FLAG_USR	0x80
253#define	P5FLAG_OS	0x40
254
255int
256writectl5(int pmc)
257{
258	quad_t newval = 0;
259
260	if (ctl_shadow[1] & (PMCF_EN << 16)) {
261		if (ctl_shadow[1] & (PMCF_USR << 16))
262			newval |= P5FLAG_USR << 16;
263		if (ctl_shadow[1] & (PMCF_OS << 16))
264			newval |= P5FLAG_OS << 16;
265		if (!(ctl_shadow[1] & (PMCF_E << 16)))
266			newval |= P5FLAG_E << 16;
267		newval |= (ctl_shadow[1] & 0x3f) << 16;
268	}
269	if (ctl_shadow[0] & (PMCF_EN << 16)) {
270		if (ctl_shadow[0] & (PMCF_USR << 16))
271			newval |= P5FLAG_USR;
272		if (ctl_shadow[0] & (PMCF_OS << 16))
273			newval |= P5FLAG_OS;
274		if (!(ctl_shadow[0] & (PMCF_E << 16)))
275			newval |= P5FLAG_E;
276		newval |= ctl_shadow[0] & 0x3f;
277	}
278
279	wrmsr(msr_ctl[0], newval);
280	return 0;		/* XXX should check for unimplemented bits */
281}
282#endif /* !SMP */
283
284/*
285 * Now the user-mode interface, called from a subdevice of mem.c.
286 */
287static int writer;
288static int writerpmc;
289
290static int
291perfmon_open(dev_t dev, int flags, int fmt, struct thread *td)
292{
293	if (!perfmon_cpuok)
294		return ENXIO;
295
296	if (flags & FWRITE) {
297		if (writer) {
298			return EBUSY;
299		} else {
300			writer = 1;
301			writerpmc = 0;
302		}
303	}
304	return 0;
305}
306
307static int
308perfmon_close(dev_t dev, int flags, int fmt, struct thread *td)
309{
310	if (flags & FWRITE) {
311		int i;
312
313		for (i = 0; i < NPMC; i++) {
314			if (writerpmc & (1 << i))
315				perfmon_fini(i);
316		}
317		writer = 0;
318	}
319	return 0;
320}
321
322static int
323perfmon_ioctl(dev_t dev, u_long cmd, caddr_t param, int flags, struct thread *td)
324{
325	struct pmc *pmc;
326	struct pmc_data *pmcd;
327	struct pmc_tstamp *pmct;
328	int *ip;
329	int rv;
330
331	switch(cmd) {
332	case PMIOSETUP:
333		if (!(flags & FWRITE))
334			return EPERM;
335		pmc = (struct pmc *)param;
336
337		rv = perfmon_setup(pmc->pmc_num, pmc->pmc_val);
338		if (!rv) {
339			writerpmc |= (1 << pmc->pmc_num);
340		}
341		break;
342
343	case PMIOGET:
344		pmc = (struct pmc *)param;
345		rv = perfmon_get(pmc->pmc_num, &pmc->pmc_val);
346		break;
347
348	case PMIOSTART:
349		if (!(flags & FWRITE))
350			return EPERM;
351
352		ip = (int *)param;
353		rv = perfmon_start(*ip);
354		break;
355
356	case PMIOSTOP:
357		if (!(flags & FWRITE))
358			return EPERM;
359
360		ip = (int *)param;
361		rv = perfmon_stop(*ip);
362		break;
363
364	case PMIORESET:
365		if (!(flags & FWRITE))
366			return EPERM;
367
368		ip = (int *)param;
369		rv = perfmon_reset(*ip);
370		break;
371
372	case PMIOREAD:
373		pmcd = (struct pmc_data *)param;
374		rv = perfmon_read(pmcd->pmcd_num, &pmcd->pmcd_value);
375		break;
376
377	case PMIOTSTAMP:
378		if (!tsc_freq) {
379			rv = ENOTTY;
380			break;
381		}
382		pmct = (struct pmc_tstamp *)param;
383		/* XXX interface loses precision. */
384		pmct->pmct_rate = tsc_freq / 1000000;
385		pmct->pmct_value = rdtsc();
386		rv = 0;
387		break;
388	default:
389		rv = ENOTTY;
390	}
391
392	return rv;
393}
394