xref: /freebsd/sys/dev/hwpmc/hwpmc_core.c (revision e52d92164754cbfff84767d4c6eb3cc93e8c21ae)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2008 Joseph Koshy
5  * All rights reserved.
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  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 /*
30  * Intel Core PMCs.
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include <sys/param.h>
37 #include <sys/bus.h>
38 #include <sys/pmc.h>
39 #include <sys/pmckern.h>
40 #include <sys/systm.h>
41 
42 #include <machine/intr_machdep.h>
43 #if (__FreeBSD_version >= 1100000)
44 #include <x86/apicvar.h>
45 #else
46 #include <machine/apicvar.h>
47 #endif
48 #include <machine/cpu.h>
49 #include <machine/cpufunc.h>
50 #include <machine/md_var.h>
51 #include <machine/specialreg.h>
52 
53 #define	CORE_CPUID_REQUEST		0xA
54 #define	CORE_CPUID_REQUEST_SIZE		0x4
55 #define	CORE_CPUID_EAX			0x0
56 #define	CORE_CPUID_EBX			0x1
57 #define	CORE_CPUID_ECX			0x2
58 #define	CORE_CPUID_EDX			0x3
59 
60 #define	IAF_PMC_CAPS			\
61 	(PMC_CAP_READ | PMC_CAP_WRITE | PMC_CAP_INTERRUPT | \
62 	 PMC_CAP_USER | PMC_CAP_SYSTEM)
63 #define	IAF_RI_TO_MSR(RI)		((RI) + (1 << 30))
64 
65 #define	IAP_PMC_CAPS (PMC_CAP_INTERRUPT | PMC_CAP_USER | PMC_CAP_SYSTEM | \
66     PMC_CAP_EDGE | PMC_CAP_THRESHOLD | PMC_CAP_READ | PMC_CAP_WRITE |	 \
67     PMC_CAP_INVERT | PMC_CAP_QUALIFIER | PMC_CAP_PRECISE)
68 
69 #define	EV_IS_NOTARCH		0
70 #define	EV_IS_ARCH_SUPP		1
71 #define	EV_IS_ARCH_NOTSUPP	-1
72 
73 /*
74  * "Architectural" events defined by Intel.  The values of these
75  * symbols correspond to positions in the bitmask returned by
76  * the CPUID.0AH instruction.
77  */
78 enum core_arch_events {
79 	CORE_AE_BRANCH_INSTRUCTION_RETIRED	= 5,
80 	CORE_AE_BRANCH_MISSES_RETIRED		= 6,
81 	CORE_AE_INSTRUCTION_RETIRED		= 1,
82 	CORE_AE_LLC_MISSES			= 4,
83 	CORE_AE_LLC_REFERENCE			= 3,
84 	CORE_AE_UNHALTED_REFERENCE_CYCLES	= 2,
85 	CORE_AE_UNHALTED_CORE_CYCLES		= 0
86 };
87 
88 static enum pmc_cputype	core_cputype;
89 
90 struct core_cpu {
91 	volatile uint32_t	pc_resync;
92 	volatile uint32_t	pc_iafctrl;	/* Fixed function control. */
93 	volatile uint64_t	pc_globalctrl;	/* Global control register. */
94 	struct pmc_hw		pc_corepmcs[];
95 };
96 
97 static struct core_cpu **core_pcpu;
98 
99 static uint32_t core_architectural_events;
100 static uint64_t core_pmcmask;
101 
102 static int core_iaf_ri;		/* relative index of fixed counters */
103 static int core_iaf_width;
104 static int core_iaf_npmc;
105 
106 static int core_iap_width;
107 static int core_iap_npmc;
108 static int core_iap_wroffset;
109 
110 static int
111 core_pcpu_noop(struct pmc_mdep *md, int cpu)
112 {
113 	(void) md;
114 	(void) cpu;
115 	return (0);
116 }
117 
118 static int
119 core_pcpu_init(struct pmc_mdep *md, int cpu)
120 {
121 	struct pmc_cpu *pc;
122 	struct core_cpu *cc;
123 	struct pmc_hw *phw;
124 	int core_ri, n, npmc;
125 
126 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
127 	    ("[iaf,%d] insane cpu number %d", __LINE__, cpu));
128 
129 	PMCDBG1(MDP,INI,1,"core-init cpu=%d", cpu);
130 
131 	core_ri = md->pmd_classdep[PMC_MDEP_CLASS_INDEX_IAP].pcd_ri;
132 	npmc = md->pmd_classdep[PMC_MDEP_CLASS_INDEX_IAP].pcd_num;
133 
134 	if (core_cputype != PMC_CPU_INTEL_CORE)
135 		npmc += md->pmd_classdep[PMC_MDEP_CLASS_INDEX_IAF].pcd_num;
136 
137 	cc = malloc(sizeof(struct core_cpu) + npmc * sizeof(struct pmc_hw),
138 	    M_PMC, M_WAITOK | M_ZERO);
139 
140 	core_pcpu[cpu] = cc;
141 	pc = pmc_pcpu[cpu];
142 
143 	KASSERT(pc != NULL && cc != NULL,
144 	    ("[core,%d] NULL per-cpu structures cpu=%d", __LINE__, cpu));
145 
146 	for (n = 0, phw = cc->pc_corepmcs; n < npmc; n++, phw++) {
147 		phw->phw_state 	  = PMC_PHW_FLAG_IS_ENABLED |
148 		    PMC_PHW_CPU_TO_STATE(cpu) |
149 		    PMC_PHW_INDEX_TO_STATE(n + core_ri);
150 		phw->phw_pmc	  = NULL;
151 		pc->pc_hwpmcs[n + core_ri]  = phw;
152 	}
153 
154 	return (0);
155 }
156 
157 static int
158 core_pcpu_fini(struct pmc_mdep *md, int cpu)
159 {
160 	int core_ri, n, npmc;
161 	struct pmc_cpu *pc;
162 	struct core_cpu *cc;
163 	uint64_t msr = 0;
164 
165 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
166 	    ("[core,%d] insane cpu number (%d)", __LINE__, cpu));
167 
168 	PMCDBG1(MDP,INI,1,"core-pcpu-fini cpu=%d", cpu);
169 
170 	if ((cc = core_pcpu[cpu]) == NULL)
171 		return (0);
172 
173 	core_pcpu[cpu] = NULL;
174 
175 	pc = pmc_pcpu[cpu];
176 
177 	KASSERT(pc != NULL, ("[core,%d] NULL per-cpu %d state", __LINE__,
178 		cpu));
179 
180 	npmc = md->pmd_classdep[PMC_MDEP_CLASS_INDEX_IAP].pcd_num;
181 	core_ri = md->pmd_classdep[PMC_MDEP_CLASS_INDEX_IAP].pcd_ri;
182 
183 	for (n = 0; n < npmc; n++) {
184 		msr = rdmsr(IAP_EVSEL0 + n) & ~IAP_EVSEL_MASK;
185 		wrmsr(IAP_EVSEL0 + n, msr);
186 	}
187 
188 	if (core_cputype != PMC_CPU_INTEL_CORE) {
189 		msr = rdmsr(IAF_CTRL) & ~IAF_CTRL_MASK;
190 		wrmsr(IAF_CTRL, msr);
191 		npmc += md->pmd_classdep[PMC_MDEP_CLASS_INDEX_IAF].pcd_num;
192 	}
193 
194 	for (n = 0; n < npmc; n++)
195 		pc->pc_hwpmcs[n + core_ri] = NULL;
196 
197 	free(cc, M_PMC);
198 
199 	return (0);
200 }
201 
202 /*
203  * Fixed function counters.
204  */
205 
206 static pmc_value_t
207 iaf_perfctr_value_to_reload_count(pmc_value_t v)
208 {
209 
210 	/* If the PMC has overflowed, return a reload count of zero. */
211 	if ((v & (1ULL << (core_iaf_width - 1))) == 0)
212 		return (0);
213 	v &= (1ULL << core_iaf_width) - 1;
214 	return (1ULL << core_iaf_width) - v;
215 }
216 
217 static pmc_value_t
218 iaf_reload_count_to_perfctr_value(pmc_value_t rlc)
219 {
220 	return (1ULL << core_iaf_width) - rlc;
221 }
222 
223 static int
224 iaf_allocate_pmc(int cpu, int ri, struct pmc *pm,
225     const struct pmc_op_pmcallocate *a)
226 {
227 	enum pmc_event ev;
228 	uint32_t caps, flags, validflags;
229 
230 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
231 	    ("[core,%d] illegal CPU %d", __LINE__, cpu));
232 
233 	PMCDBG2(MDP,ALL,1, "iaf-allocate ri=%d reqcaps=0x%x", ri, pm->pm_caps);
234 
235 	if (ri < 0 || ri > core_iaf_npmc)
236 		return (EINVAL);
237 
238 	caps = a->pm_caps;
239 
240 	if (a->pm_class != PMC_CLASS_IAF ||
241 	    (caps & IAF_PMC_CAPS) != caps)
242 		return (EINVAL);
243 
244 	ev = pm->pm_event;
245 
246 
247 	if (ev == PMC_EV_IAF_INSTR_RETIRED_ANY && ri != 0)
248 		return (EINVAL);
249 	if (ev == PMC_EV_IAF_CPU_CLK_UNHALTED_CORE && ri != 1)
250 		return (EINVAL);
251 	if (ev == PMC_EV_IAF_CPU_CLK_UNHALTED_REF && ri != 2)
252 		return (EINVAL);
253 
254 	flags = a->pm_md.pm_iaf.pm_iaf_flags;
255 
256 	validflags = IAF_MASK;
257 
258 	if (caps & PMC_CAP_INTERRUPT)
259 		flags |= IAF_PMI;
260 	if (caps & PMC_CAP_SYSTEM)
261 		flags |= IAF_OS;
262 	if (caps & PMC_CAP_USER)
263 		flags |= IAF_USR;
264 	if ((caps & (PMC_CAP_USER | PMC_CAP_SYSTEM)) == 0)
265 		flags |= (IAF_OS | IAF_USR);
266 
267 	pm->pm_md.pm_iaf.pm_iaf_ctrl = (flags << (ri * 4));
268 
269 	PMCDBG1(MDP,ALL,2, "iaf-allocate config=0x%jx",
270 	    (uintmax_t) pm->pm_md.pm_iaf.pm_iaf_ctrl);
271 
272 	return (0);
273 }
274 
275 static int
276 iaf_config_pmc(int cpu, int ri, struct pmc *pm)
277 {
278 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
279 	    ("[core,%d] illegal CPU %d", __LINE__, cpu));
280 
281 	KASSERT(ri >= 0 && ri < core_iaf_npmc,
282 	    ("[core,%d] illegal row-index %d", __LINE__, ri));
283 
284 	PMCDBG3(MDP,CFG,1, "iaf-config cpu=%d ri=%d pm=%p", cpu, ri, pm);
285 
286 	KASSERT(core_pcpu[cpu] != NULL, ("[core,%d] null per-cpu %d", __LINE__,
287 	    cpu));
288 
289 	core_pcpu[cpu]->pc_corepmcs[ri + core_iaf_ri].phw_pmc = pm;
290 
291 	return (0);
292 }
293 
294 static int
295 iaf_describe(int cpu, int ri, struct pmc_info *pi, struct pmc **ppmc)
296 {
297 	int error;
298 	struct pmc_hw *phw;
299 	char iaf_name[PMC_NAME_MAX];
300 
301 	phw = &core_pcpu[cpu]->pc_corepmcs[ri + core_iaf_ri];
302 
303 	(void) snprintf(iaf_name, sizeof(iaf_name), "IAF-%d", ri);
304 	if ((error = copystr(iaf_name, pi->pm_name, PMC_NAME_MAX,
305 	    NULL)) != 0)
306 		return (error);
307 
308 	pi->pm_class = PMC_CLASS_IAF;
309 
310 	if (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) {
311 		pi->pm_enabled = TRUE;
312 		*ppmc          = phw->phw_pmc;
313 	} else {
314 		pi->pm_enabled = FALSE;
315 		*ppmc          = NULL;
316 	}
317 
318 	return (0);
319 }
320 
321 static int
322 iaf_get_config(int cpu, int ri, struct pmc **ppm)
323 {
324 	*ppm = core_pcpu[cpu]->pc_corepmcs[ri + core_iaf_ri].phw_pmc;
325 
326 	return (0);
327 }
328 
329 static int
330 iaf_get_msr(int ri, uint32_t *msr)
331 {
332 	KASSERT(ri >= 0 && ri < core_iaf_npmc,
333 	    ("[iaf,%d] ri %d out of range", __LINE__, ri));
334 
335 	*msr = IAF_RI_TO_MSR(ri);
336 
337 	return (0);
338 }
339 
340 static int
341 iaf_read_pmc(int cpu, int ri, pmc_value_t *v)
342 {
343 	struct pmc *pm;
344 	pmc_value_t tmp;
345 
346 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
347 	    ("[core,%d] illegal cpu value %d", __LINE__, cpu));
348 	KASSERT(ri >= 0 && ri < core_iaf_npmc,
349 	    ("[core,%d] illegal row-index %d", __LINE__, ri));
350 
351 	pm = core_pcpu[cpu]->pc_corepmcs[ri + core_iaf_ri].phw_pmc;
352 
353 	KASSERT(pm,
354 	    ("[core,%d] cpu %d ri %d(%d) pmc not configured", __LINE__, cpu,
355 		ri, ri + core_iaf_ri));
356 
357 	tmp = rdpmc(IAF_RI_TO_MSR(ri));
358 
359 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
360 		*v = iaf_perfctr_value_to_reload_count(tmp);
361 	else
362 		*v = tmp & ((1ULL << core_iaf_width) - 1);
363 
364 	PMCDBG4(MDP,REA,1, "iaf-read cpu=%d ri=%d msr=0x%x -> v=%jx", cpu, ri,
365 	    IAF_RI_TO_MSR(ri), *v);
366 
367 	return (0);
368 }
369 
370 static int
371 iaf_release_pmc(int cpu, int ri, struct pmc *pmc)
372 {
373 	PMCDBG3(MDP,REL,1, "iaf-release cpu=%d ri=%d pm=%p", cpu, ri, pmc);
374 
375 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
376 	    ("[core,%d] illegal CPU value %d", __LINE__, cpu));
377 	KASSERT(ri >= 0 && ri < core_iaf_npmc,
378 	    ("[core,%d] illegal row-index %d", __LINE__, ri));
379 
380 	KASSERT(core_pcpu[cpu]->pc_corepmcs[ri + core_iaf_ri].phw_pmc == NULL,
381 	    ("[core,%d] PHW pmc non-NULL", __LINE__));
382 
383 	return (0);
384 }
385 
386 static int
387 iaf_start_pmc(int cpu, int ri)
388 {
389 	struct pmc *pm;
390 	struct core_cpu *iafc;
391 	uint64_t msr = 0;
392 
393 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
394 	    ("[core,%d] illegal CPU value %d", __LINE__, cpu));
395 	KASSERT(ri >= 0 && ri < core_iaf_npmc,
396 	    ("[core,%d] illegal row-index %d", __LINE__, ri));
397 
398 	PMCDBG2(MDP,STA,1,"iaf-start cpu=%d ri=%d", cpu, ri);
399 
400 	iafc = core_pcpu[cpu];
401 	pm = iafc->pc_corepmcs[ri + core_iaf_ri].phw_pmc;
402 
403 	iafc->pc_iafctrl |= pm->pm_md.pm_iaf.pm_iaf_ctrl;
404 
405  	msr = rdmsr(IAF_CTRL) & ~IAF_CTRL_MASK;
406  	wrmsr(IAF_CTRL, msr | (iafc->pc_iafctrl & IAF_CTRL_MASK));
407 
408 	do {
409 		iafc->pc_resync = 0;
410 		iafc->pc_globalctrl |= (1ULL << (ri + IAF_OFFSET));
411  		msr = rdmsr(IA_GLOBAL_CTRL) & ~IAF_GLOBAL_CTRL_MASK;
412  		wrmsr(IA_GLOBAL_CTRL, msr | (iafc->pc_globalctrl &
413  					     IAF_GLOBAL_CTRL_MASK));
414 	} while (iafc->pc_resync != 0);
415 
416 	PMCDBG4(MDP,STA,1,"iafctrl=%x(%x) globalctrl=%jx(%jx)",
417 	    iafc->pc_iafctrl, (uint32_t) rdmsr(IAF_CTRL),
418 	    iafc->pc_globalctrl, rdmsr(IA_GLOBAL_CTRL));
419 
420 	return (0);
421 }
422 
423 static int
424 iaf_stop_pmc(int cpu, int ri)
425 {
426 	uint32_t fc;
427 	struct core_cpu *iafc;
428 	uint64_t msr = 0;
429 
430 	PMCDBG2(MDP,STO,1,"iaf-stop cpu=%d ri=%d", cpu, ri);
431 
432 	iafc = core_pcpu[cpu];
433 
434 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
435 	    ("[core,%d] illegal CPU value %d", __LINE__, cpu));
436 	KASSERT(ri >= 0 && ri < core_iaf_npmc,
437 	    ("[core,%d] illegal row-index %d", __LINE__, ri));
438 
439 	fc = (IAF_MASK << (ri * 4));
440 
441 	iafc->pc_iafctrl &= ~fc;
442 
443 	PMCDBG1(MDP,STO,1,"iaf-stop iafctrl=%x", iafc->pc_iafctrl);
444  	msr = rdmsr(IAF_CTRL) & ~IAF_CTRL_MASK;
445  	wrmsr(IAF_CTRL, msr | (iafc->pc_iafctrl & IAF_CTRL_MASK));
446 
447 	do {
448 		iafc->pc_resync = 0;
449 		iafc->pc_globalctrl &= ~(1ULL << (ri + IAF_OFFSET));
450  		msr = rdmsr(IA_GLOBAL_CTRL) & ~IAF_GLOBAL_CTRL_MASK;
451  		wrmsr(IA_GLOBAL_CTRL, msr | (iafc->pc_globalctrl &
452  					     IAF_GLOBAL_CTRL_MASK));
453 	} while (iafc->pc_resync != 0);
454 
455 	PMCDBG4(MDP,STO,1,"iafctrl=%x(%x) globalctrl=%jx(%jx)",
456 	    iafc->pc_iafctrl, (uint32_t) rdmsr(IAF_CTRL),
457 	    iafc->pc_globalctrl, rdmsr(IA_GLOBAL_CTRL));
458 
459 	return (0);
460 }
461 
462 static int
463 iaf_write_pmc(int cpu, int ri, pmc_value_t v)
464 {
465 	struct core_cpu *cc;
466 	struct pmc *pm;
467 	uint64_t msr;
468 
469 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
470 	    ("[core,%d] illegal cpu value %d", __LINE__, cpu));
471 	KASSERT(ri >= 0 && ri < core_iaf_npmc,
472 	    ("[core,%d] illegal row-index %d", __LINE__, ri));
473 
474 	cc = core_pcpu[cpu];
475 	pm = cc->pc_corepmcs[ri + core_iaf_ri].phw_pmc;
476 
477 	KASSERT(pm,
478 	    ("[core,%d] cpu %d ri %d pmc not configured", __LINE__, cpu, ri));
479 
480 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
481 		v = iaf_reload_count_to_perfctr_value(v);
482 
483 	/* Turn off fixed counters */
484 	msr = rdmsr(IAF_CTRL) & ~IAF_CTRL_MASK;
485 	wrmsr(IAF_CTRL, msr);
486 
487 	wrmsr(IAF_CTR0 + ri, v & ((1ULL << core_iaf_width) - 1));
488 
489 	/* Turn on fixed counters */
490 	msr = rdmsr(IAF_CTRL) & ~IAF_CTRL_MASK;
491 	wrmsr(IAF_CTRL, msr | (cc->pc_iafctrl & IAF_CTRL_MASK));
492 
493 	PMCDBG6(MDP,WRI,1, "iaf-write cpu=%d ri=%d msr=0x%x v=%jx iafctrl=%jx "
494 	    "pmc=%jx", cpu, ri, IAF_RI_TO_MSR(ri), v,
495 	    (uintmax_t) rdmsr(IAF_CTRL),
496 	    (uintmax_t) rdpmc(IAF_RI_TO_MSR(ri)));
497 
498 	return (0);
499 }
500 
501 
502 static void
503 iaf_initialize(struct pmc_mdep *md, int maxcpu, int npmc, int pmcwidth)
504 {
505 	struct pmc_classdep *pcd;
506 
507 	KASSERT(md != NULL, ("[iaf,%d] md is NULL", __LINE__));
508 
509 	PMCDBG0(MDP,INI,1, "iaf-initialize");
510 
511 	pcd = &md->pmd_classdep[PMC_MDEP_CLASS_INDEX_IAF];
512 
513 	pcd->pcd_caps	= IAF_PMC_CAPS;
514 	pcd->pcd_class	= PMC_CLASS_IAF;
515 	pcd->pcd_num	= npmc;
516 	pcd->pcd_ri	= md->pmd_npmc;
517 	pcd->pcd_width	= pmcwidth;
518 
519 	pcd->pcd_allocate_pmc	= iaf_allocate_pmc;
520 	pcd->pcd_config_pmc	= iaf_config_pmc;
521 	pcd->pcd_describe	= iaf_describe;
522 	pcd->pcd_get_config	= iaf_get_config;
523 	pcd->pcd_get_msr	= iaf_get_msr;
524 	pcd->pcd_pcpu_fini	= core_pcpu_noop;
525 	pcd->pcd_pcpu_init	= core_pcpu_noop;
526 	pcd->pcd_read_pmc	= iaf_read_pmc;
527 	pcd->pcd_release_pmc	= iaf_release_pmc;
528 	pcd->pcd_start_pmc	= iaf_start_pmc;
529 	pcd->pcd_stop_pmc	= iaf_stop_pmc;
530 	pcd->pcd_write_pmc	= iaf_write_pmc;
531 
532 	md->pmd_npmc	       += npmc;
533 }
534 
535 /*
536  * Intel programmable PMCs.
537  */
538 
539 /* Sub fields of UMASK that this event supports. */
540 #define	IAP_M_CORE		(1 << 0) /* Core specificity */
541 #define	IAP_M_AGENT		(1 << 1) /* Agent specificity */
542 #define	IAP_M_PREFETCH		(1 << 2) /* Prefetch */
543 #define	IAP_M_MESI		(1 << 3) /* MESI */
544 #define	IAP_M_SNOOPRESPONSE	(1 << 4) /* Snoop response */
545 #define	IAP_M_SNOOPTYPE		(1 << 5) /* Snoop type */
546 #define	IAP_M_TRANSITION	(1 << 6) /* Transition */
547 
548 #define	IAP_F_CORE		(0x3 << 14) /* Core specificity */
549 #define	IAP_F_AGENT		(0x1 << 13) /* Agent specificity */
550 #define	IAP_F_PREFETCH		(0x3 << 12) /* Prefetch */
551 #define	IAP_F_MESI		(0xF <<  8) /* MESI */
552 #define	IAP_F_SNOOPRESPONSE	(0xB <<  8) /* Snoop response */
553 #define	IAP_F_SNOOPTYPE		(0x3 <<  8) /* Snoop type */
554 #define	IAP_F_TRANSITION	(0x1 << 12) /* Transition */
555 
556 #define	IAP_PREFETCH_RESERVED	(0x2 << 12)
557 #define	IAP_CORE_THIS		(0x1 << 14)
558 #define	IAP_CORE_ALL		(0x3 << 14)
559 #define	IAP_F_CMASK		0xFF000000
560 
561 static pmc_value_t
562 iap_perfctr_value_to_reload_count(pmc_value_t v)
563 {
564 
565 	/* If the PMC has overflowed, return a reload count of zero. */
566 	if ((v & (1ULL << (core_iap_width - 1))) == 0)
567 		return (0);
568 	v &= (1ULL << core_iap_width) - 1;
569 	return (1ULL << core_iap_width) - v;
570 }
571 
572 static pmc_value_t
573 iap_reload_count_to_perfctr_value(pmc_value_t rlc)
574 {
575 	return (1ULL << core_iap_width) - rlc;
576 }
577 
578 static int
579 iap_pmc_has_overflowed(int ri)
580 {
581 	uint64_t v;
582 
583 	/*
584 	 * We treat a Core (i.e., Intel architecture v1) PMC as has
585 	 * having overflowed if its MSB is zero.
586 	 */
587 	v = rdpmc(ri);
588 	return ((v & (1ULL << (core_iap_width - 1))) == 0);
589 }
590 
591 static int
592 iap_event_corei7_ok_on_counter(uint8_t evsel, int ri)
593 {
594 	uint32_t mask;
595 
596 	switch (evsel) {
597 		/*
598 		 * Events valid only on counter 0, 1.
599 		 */
600 		case 0x40:
601 		case 0x41:
602 		case 0x42:
603 		case 0x43:
604 		case 0x51:
605 		case 0x63:
606 			mask = 0x3;
607 		break;
608 
609 		default:
610 		mask = ~0;	/* Any row index is ok. */
611 	}
612 
613 	return (mask & (1 << ri));
614 }
615 
616 static int
617 iap_event_westmere_ok_on_counter(uint8_t evsel, int ri)
618 {
619 	uint32_t mask;
620 
621 	switch (evsel) {
622 		/*
623 		 * Events valid only on counter 0.
624 		 */
625 		case 0x60:
626 		case 0xB3:
627 		mask = 0x1;
628 		break;
629 
630 		/*
631 		 * Events valid only on counter 0, 1.
632 		 */
633 		case 0x4C:
634 		case 0x4E:
635 		case 0x51:
636 		case 0x63:
637 		mask = 0x3;
638 		break;
639 
640 	default:
641 		mask = ~0;	/* Any row index is ok. */
642 	}
643 
644 	return (mask & (1 << ri));
645 }
646 
647 static int
648 iap_event_sb_sbx_ib_ibx_ok_on_counter(uint8_t evsel, int ri)
649 {
650 	uint32_t mask;
651 
652 	switch (evsel) {
653 		/* Events valid only on counter 0. */
654     case 0xB7:
655 		mask = 0x1;
656 		break;
657 		/* Events valid only on counter 1. */
658 	case 0xC0:
659 		mask = 0x2;
660 		break;
661 		/* Events valid only on counter 2. */
662 	case 0x48:
663 	case 0xA2:
664 	case 0xA3:
665 		mask = 0x4;
666 		break;
667 		/* Events valid only on counter 3. */
668 	case 0xBB:
669 	case 0xCD:
670 		mask = 0x8;
671 		break;
672 	default:
673 		mask = ~0;	/* Any row index is ok. */
674 	}
675 
676 	return (mask & (1 << ri));
677 }
678 
679 static int
680 iap_event_ok_on_counter(uint8_t evsel, int ri)
681 {
682 	uint32_t mask;
683 
684 	switch (evsel) {
685 		/*
686 		 * Events valid only on counter 0.
687 		 */
688 	case 0x10:
689 	case 0x14:
690 	case 0x18:
691 	case 0xB3:
692 	case 0xC1:
693 	case 0xCB:
694 		mask = (1 << 0);
695 		break;
696 
697 		/*
698 		 * Events valid only on counter 1.
699 		 */
700 	case 0x11:
701 	case 0x12:
702 	case 0x13:
703 		mask = (1 << 1);
704 		break;
705 
706 	default:
707 		mask = ~0;	/* Any row index is ok. */
708 	}
709 
710 	return (mask & (1 << ri));
711 }
712 
713 static int
714 iap_allocate_pmc(int cpu, int ri, struct pmc *pm,
715     const struct pmc_op_pmcallocate *a)
716 {
717 	enum pmc_event map;
718 	uint8_t ev;
719 	uint32_t caps;
720 	const struct pmc_md_iap_op_pmcallocate *iap;
721 
722 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
723 	    ("[core,%d] illegal CPU %d", __LINE__, cpu));
724 	KASSERT(ri >= 0 && ri < core_iap_npmc,
725 	    ("[core,%d] illegal row-index value %d", __LINE__, ri));
726 
727 	/* check requested capabilities */
728 	caps = a->pm_caps;
729 	if ((IAP_PMC_CAPS & caps) != caps)
730 		return (EPERM);
731 	map = 0;	/* XXX: silent GCC warning */
732 	iap = &a->pm_md.pm_iap;
733 	ev = IAP_EVSEL_GET(iap->pm_iap_config);
734 
735 	switch (core_cputype) {
736 	case PMC_CPU_INTEL_COREI7:
737 	case PMC_CPU_INTEL_NEHALEM_EX:
738 		if (iap_event_corei7_ok_on_counter(ev, ri) == 0)
739 			return (EINVAL);
740 		break;
741 	case PMC_CPU_INTEL_SKYLAKE:
742 	case PMC_CPU_INTEL_SKYLAKE_XEON:
743 	case PMC_CPU_INTEL_BROADWELL:
744 	case PMC_CPU_INTEL_BROADWELL_XEON:
745 	case PMC_CPU_INTEL_SANDYBRIDGE:
746 	case PMC_CPU_INTEL_SANDYBRIDGE_XEON:
747 	case PMC_CPU_INTEL_IVYBRIDGE:
748 	case PMC_CPU_INTEL_IVYBRIDGE_XEON:
749 	case PMC_CPU_INTEL_HASWELL:
750 	case PMC_CPU_INTEL_HASWELL_XEON:
751 		if (iap_event_sb_sbx_ib_ibx_ok_on_counter(ev, ri) == 0)
752 			return (EINVAL);
753 		break;
754 	case PMC_CPU_INTEL_WESTMERE:
755 	case PMC_CPU_INTEL_WESTMERE_EX:
756 		if (iap_event_westmere_ok_on_counter(ev, ri) == 0)
757 			return (EINVAL);
758 		break;
759 	default:
760 		if (iap_event_ok_on_counter(ev, ri) == 0)
761 			return (EINVAL);
762 	}
763 
764 	pm->pm_md.pm_iap.pm_iap_evsel = iap->pm_iap_config;
765 	return (0);
766 }
767 
768 static int
769 iap_config_pmc(int cpu, int ri, struct pmc *pm)
770 {
771 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
772 	    ("[core,%d] illegal CPU %d", __LINE__, cpu));
773 
774 	KASSERT(ri >= 0 && ri < core_iap_npmc,
775 	    ("[core,%d] illegal row-index %d", __LINE__, ri));
776 
777 	PMCDBG3(MDP,CFG,1, "iap-config cpu=%d ri=%d pm=%p", cpu, ri, pm);
778 
779 	KASSERT(core_pcpu[cpu] != NULL, ("[core,%d] null per-cpu %d", __LINE__,
780 	    cpu));
781 
782 	core_pcpu[cpu]->pc_corepmcs[ri].phw_pmc = pm;
783 
784 	return (0);
785 }
786 
787 static int
788 iap_describe(int cpu, int ri, struct pmc_info *pi, struct pmc **ppmc)
789 {
790 	int error;
791 	struct pmc_hw *phw;
792 	char iap_name[PMC_NAME_MAX];
793 
794 	phw = &core_pcpu[cpu]->pc_corepmcs[ri];
795 
796 	(void) snprintf(iap_name, sizeof(iap_name), "IAP-%d", ri);
797 	if ((error = copystr(iap_name, pi->pm_name, PMC_NAME_MAX,
798 	    NULL)) != 0)
799 		return (error);
800 
801 	pi->pm_class = PMC_CLASS_IAP;
802 
803 	if (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) {
804 		pi->pm_enabled = TRUE;
805 		*ppmc          = phw->phw_pmc;
806 	} else {
807 		pi->pm_enabled = FALSE;
808 		*ppmc          = NULL;
809 	}
810 
811 	return (0);
812 }
813 
814 static int
815 iap_get_config(int cpu, int ri, struct pmc **ppm)
816 {
817 	*ppm = core_pcpu[cpu]->pc_corepmcs[ri].phw_pmc;
818 
819 	return (0);
820 }
821 
822 static int
823 iap_get_msr(int ri, uint32_t *msr)
824 {
825 	KASSERT(ri >= 0 && ri < core_iap_npmc,
826 	    ("[iap,%d] ri %d out of range", __LINE__, ri));
827 
828 	*msr = ri;
829 
830 	return (0);
831 }
832 
833 static int
834 iap_read_pmc(int cpu, int ri, pmc_value_t *v)
835 {
836 	struct pmc *pm;
837 	pmc_value_t tmp;
838 
839 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
840 	    ("[core,%d] illegal cpu value %d", __LINE__, cpu));
841 	KASSERT(ri >= 0 && ri < core_iap_npmc,
842 	    ("[core,%d] illegal row-index %d", __LINE__, ri));
843 
844 	pm = core_pcpu[cpu]->pc_corepmcs[ri].phw_pmc;
845 
846 	KASSERT(pm,
847 	    ("[core,%d] cpu %d ri %d pmc not configured", __LINE__, cpu,
848 		ri));
849 
850 	tmp = rdpmc(ri);
851 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
852 		*v = iap_perfctr_value_to_reload_count(tmp);
853 	else
854 		*v = tmp & ((1ULL << core_iap_width) - 1);
855 
856 	PMCDBG4(MDP,REA,1, "iap-read cpu=%d ri=%d msr=0x%x -> v=%jx", cpu, ri,
857 	    IAP_PMC0 + ri, *v);
858 
859 	return (0);
860 }
861 
862 static int
863 iap_release_pmc(int cpu, int ri, struct pmc *pm)
864 {
865 	(void) pm;
866 
867 	PMCDBG3(MDP,REL,1, "iap-release cpu=%d ri=%d pm=%p", cpu, ri,
868 	    pm);
869 
870 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
871 	    ("[core,%d] illegal CPU value %d", __LINE__, cpu));
872 	KASSERT(ri >= 0 && ri < core_iap_npmc,
873 	    ("[core,%d] illegal row-index %d", __LINE__, ri));
874 
875 	KASSERT(core_pcpu[cpu]->pc_corepmcs[ri].phw_pmc
876 	    == NULL, ("[core,%d] PHW pmc non-NULL", __LINE__));
877 
878 	return (0);
879 }
880 
881 static int
882 iap_start_pmc(int cpu, int ri)
883 {
884 	struct pmc *pm;
885 	uint32_t evsel;
886 	struct core_cpu *cc;
887 
888 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
889 	    ("[core,%d] illegal CPU value %d", __LINE__, cpu));
890 	KASSERT(ri >= 0 && ri < core_iap_npmc,
891 	    ("[core,%d] illegal row-index %d", __LINE__, ri));
892 
893 	cc = core_pcpu[cpu];
894 	pm = cc->pc_corepmcs[ri].phw_pmc;
895 
896 	KASSERT(pm,
897 	    ("[core,%d] starting cpu%d,ri%d with no pmc configured",
898 		__LINE__, cpu, ri));
899 
900 	PMCDBG2(MDP,STA,1, "iap-start cpu=%d ri=%d", cpu, ri);
901 
902 	evsel = pm->pm_md.pm_iap.pm_iap_evsel;
903 
904 	PMCDBG4(MDP,STA,2, "iap-start/2 cpu=%d ri=%d evselmsr=0x%x evsel=0x%x",
905 	    cpu, ri, IAP_EVSEL0 + ri, evsel);
906 
907 	/* Event specific configuration. */
908 
909 	switch (IAP_EVSEL_GET(evsel)) {
910 	case 0xB7:
911 		wrmsr(IA_OFFCORE_RSP0, pm->pm_md.pm_iap.pm_iap_rsp);
912 		break;
913 	case 0xBB:
914 		wrmsr(IA_OFFCORE_RSP1, pm->pm_md.pm_iap.pm_iap_rsp);
915 		break;
916 	default:
917 		break;
918 	}
919 
920 	wrmsr(IAP_EVSEL0 + ri, evsel | IAP_EN);
921 
922 	if (core_cputype == PMC_CPU_INTEL_CORE)
923 		return (0);
924 
925 	do {
926 		cc->pc_resync = 0;
927 		cc->pc_globalctrl |= (1ULL << ri);
928 		wrmsr(IA_GLOBAL_CTRL, cc->pc_globalctrl);
929 	} while (cc->pc_resync != 0);
930 
931 	return (0);
932 }
933 
934 static int
935 iap_stop_pmc(int cpu, int ri)
936 {
937 	struct pmc *pm;
938 	struct core_cpu *cc;
939 	uint64_t msr;
940 
941 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
942 	    ("[core,%d] illegal cpu value %d", __LINE__, cpu));
943 	KASSERT(ri >= 0 && ri < core_iap_npmc,
944 	    ("[core,%d] illegal row index %d", __LINE__, ri));
945 
946 	cc = core_pcpu[cpu];
947 	pm = cc->pc_corepmcs[ri].phw_pmc;
948 
949 	KASSERT(pm,
950 	    ("[core,%d] cpu%d ri%d no configured PMC to stop", __LINE__,
951 		cpu, ri));
952 
953 	PMCDBG2(MDP,STO,1, "iap-stop cpu=%d ri=%d", cpu, ri);
954 
955 	msr = rdmsr(IAP_EVSEL0 + ri) & ~IAP_EVSEL_MASK;
956 	wrmsr(IAP_EVSEL0 + ri, msr);	/* stop hw */
957 
958 	if (core_cputype == PMC_CPU_INTEL_CORE)
959 		return (0);
960 
961 	msr = 0;
962 	do {
963 		cc->pc_resync = 0;
964 		cc->pc_globalctrl &= ~(1ULL << ri);
965 		msr = rdmsr(IA_GLOBAL_CTRL) & ~IA_GLOBAL_CTRL_MASK;
966 		wrmsr(IA_GLOBAL_CTRL, cc->pc_globalctrl);
967 	} while (cc->pc_resync != 0);
968 
969 	return (0);
970 }
971 
972 static int
973 iap_write_pmc(int cpu, int ri, pmc_value_t v)
974 {
975 	struct pmc *pm;
976 	struct core_cpu *cc;
977 
978 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
979 	    ("[core,%d] illegal cpu value %d", __LINE__, cpu));
980 	KASSERT(ri >= 0 && ri < core_iap_npmc,
981 	    ("[core,%d] illegal row index %d", __LINE__, ri));
982 
983 	cc = core_pcpu[cpu];
984 	pm = cc->pc_corepmcs[ri].phw_pmc;
985 
986 	KASSERT(pm,
987 	    ("[core,%d] cpu%d ri%d no configured PMC to stop", __LINE__,
988 		cpu, ri));
989 
990 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
991 		v = iap_reload_count_to_perfctr_value(v);
992 
993 	v &= (1ULL << core_iap_width) - 1;
994 
995 	PMCDBG4(MDP,WRI,1, "iap-write cpu=%d ri=%d msr=0x%x v=%jx", cpu, ri,
996 	    IAP_PMC0 + ri, v);
997 
998 	/*
999 	 * Write the new value to the counter (or it's alias).  The
1000 	 * counter will be in a stopped state when the pcd_write()
1001 	 * entry point is called.
1002 	 */
1003 	wrmsr(core_iap_wroffset + IAP_PMC0 + ri, v);
1004 	return (0);
1005 }
1006 
1007 
1008 static void
1009 iap_initialize(struct pmc_mdep *md, int maxcpu, int npmc, int pmcwidth,
1010     int flags)
1011 {
1012 	struct pmc_classdep *pcd;
1013 
1014 	KASSERT(md != NULL, ("[iap,%d] md is NULL", __LINE__));
1015 
1016 	PMCDBG0(MDP,INI,1, "iap-initialize");
1017 
1018 	/* Remember the set of architectural events supported. */
1019 	core_architectural_events = ~flags;
1020 
1021 	pcd = &md->pmd_classdep[PMC_MDEP_CLASS_INDEX_IAP];
1022 
1023 	pcd->pcd_caps	= IAP_PMC_CAPS;
1024 	pcd->pcd_class	= PMC_CLASS_IAP;
1025 	pcd->pcd_num	= npmc;
1026 	pcd->pcd_ri	= md->pmd_npmc;
1027 	pcd->pcd_width	= pmcwidth;
1028 
1029 	pcd->pcd_allocate_pmc	= iap_allocate_pmc;
1030 	pcd->pcd_config_pmc	= iap_config_pmc;
1031 	pcd->pcd_describe	= iap_describe;
1032 	pcd->pcd_get_config	= iap_get_config;
1033 	pcd->pcd_get_msr	= iap_get_msr;
1034 	pcd->pcd_pcpu_fini	= core_pcpu_fini;
1035 	pcd->pcd_pcpu_init	= core_pcpu_init;
1036 	pcd->pcd_read_pmc	= iap_read_pmc;
1037 	pcd->pcd_release_pmc	= iap_release_pmc;
1038 	pcd->pcd_start_pmc	= iap_start_pmc;
1039 	pcd->pcd_stop_pmc	= iap_stop_pmc;
1040 	pcd->pcd_write_pmc	= iap_write_pmc;
1041 
1042 	md->pmd_npmc	       += npmc;
1043 }
1044 
1045 static int
1046 core_intr(int cpu, struct trapframe *tf)
1047 {
1048 	pmc_value_t v;
1049 	struct pmc *pm;
1050 	struct core_cpu *cc;
1051 	int error, found_interrupt, ri;
1052 	uint64_t msr;
1053 
1054 	PMCDBG3(MDP,INT, 1, "cpu=%d tf=0x%p um=%d", cpu, (void *) tf,
1055 	    TRAPF_USERMODE(tf));
1056 
1057 	found_interrupt = 0;
1058 	cc = core_pcpu[cpu];
1059 
1060 	for (ri = 0; ri < core_iap_npmc; ri++) {
1061 
1062 		if ((pm = cc->pc_corepmcs[ri].phw_pmc) == NULL ||
1063 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1064 			continue;
1065 
1066 		if (!iap_pmc_has_overflowed(ri))
1067 			continue;
1068 
1069 		found_interrupt = 1;
1070 
1071 		if (pm->pm_state != PMC_STATE_RUNNING)
1072 			continue;
1073 
1074 		error = pmc_process_interrupt(cpu, PMC_HR, pm, tf,
1075 		    TRAPF_USERMODE(tf));
1076 
1077 		v = pm->pm_sc.pm_reloadcount;
1078 		v = iap_reload_count_to_perfctr_value(v);
1079 
1080 		/*
1081 		 * Stop the counter, reload it but only restart it if
1082 		 * the PMC is not stalled.
1083 		 */
1084 		msr = rdmsr(IAP_EVSEL0 + ri) & ~IAP_EVSEL_MASK;
1085 		wrmsr(IAP_EVSEL0 + ri, msr);
1086 		wrmsr(core_iap_wroffset + IAP_PMC0 + ri, v);
1087 
1088 		if (error)
1089 			continue;
1090 
1091 		wrmsr(IAP_EVSEL0 + ri, msr | (pm->pm_md.pm_iap.pm_iap_evsel |
1092 					      IAP_EN));
1093 	}
1094 
1095 	if (found_interrupt)
1096 		lapic_reenable_pmc();
1097 
1098 	if (found_interrupt)
1099 		counter_u64_add(pmc_stats.pm_intr_processed, 1);
1100 	else
1101 		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
1102 
1103 	return (found_interrupt);
1104 }
1105 
1106 static int
1107 core2_intr(int cpu, struct trapframe *tf)
1108 {
1109 	int error, found_interrupt, n;
1110 	uint64_t flag, intrstatus, intrenable, msr;
1111 	struct pmc *pm;
1112 	struct core_cpu *cc;
1113 	pmc_value_t v;
1114 
1115 	PMCDBG3(MDP,INT, 1, "cpu=%d tf=0x%p um=%d", cpu, (void *) tf,
1116 	    TRAPF_USERMODE(tf));
1117 
1118 	/*
1119 	 * The IA_GLOBAL_STATUS (MSR 0x38E) register indicates which
1120 	 * PMCs have a pending PMI interrupt.  We take a 'snapshot' of
1121 	 * the current set of interrupting PMCs and process these
1122 	 * after stopping them.
1123 	 */
1124 	intrstatus = rdmsr(IA_GLOBAL_STATUS);
1125 	intrenable = intrstatus & core_pmcmask;
1126 
1127 	PMCDBG2(MDP,INT, 1, "cpu=%d intrstatus=%jx", cpu,
1128 	    (uintmax_t) intrstatus);
1129 
1130 	found_interrupt = 0;
1131 	cc = core_pcpu[cpu];
1132 
1133 	KASSERT(cc != NULL, ("[core,%d] null pcpu", __LINE__));
1134 
1135 	cc->pc_globalctrl &= ~intrenable;
1136 	cc->pc_resync = 1;	/* MSRs now potentially out of sync. */
1137 
1138 	/*
1139 	 * Stop PMCs and clear overflow status bits.
1140 	 */
1141 	msr = rdmsr(IA_GLOBAL_CTRL) & ~IA_GLOBAL_CTRL_MASK;
1142 	wrmsr(IA_GLOBAL_CTRL, msr);
1143 	wrmsr(IA_GLOBAL_OVF_CTRL, intrenable |
1144 	    IA_GLOBAL_STATUS_FLAG_OVFBUF |
1145 	    IA_GLOBAL_STATUS_FLAG_CONDCHG);
1146 
1147 	/*
1148 	 * Look for interrupts from fixed function PMCs.
1149 	 */
1150 	for (n = 0, flag = (1ULL << IAF_OFFSET); n < core_iaf_npmc;
1151 	     n++, flag <<= 1) {
1152 
1153 		if ((intrstatus & flag) == 0)
1154 			continue;
1155 
1156 		found_interrupt = 1;
1157 
1158 		pm = cc->pc_corepmcs[n + core_iaf_ri].phw_pmc;
1159 		if (pm == NULL || pm->pm_state != PMC_STATE_RUNNING ||
1160 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1161 			continue;
1162 
1163 		error = pmc_process_interrupt(cpu, PMC_HR, pm, tf,
1164 		    TRAPF_USERMODE(tf));
1165 
1166 		if (error)
1167 			intrenable &= ~flag;
1168 
1169 		v = iaf_reload_count_to_perfctr_value(pm->pm_sc.pm_reloadcount);
1170 
1171 		/* Reload sampling count. */
1172 		wrmsr(IAF_CTR0 + n, v);
1173 
1174 		PMCDBG4(MDP,INT, 1, "iaf-intr cpu=%d error=%d v=%jx(%jx)", cpu,
1175 		    error, (uintmax_t) v, (uintmax_t) rdpmc(IAF_RI_TO_MSR(n)));
1176 	}
1177 
1178 	/*
1179 	 * Process interrupts from the programmable counters.
1180 	 */
1181 	for (n = 0, flag = 1; n < core_iap_npmc; n++, flag <<= 1) {
1182 		if ((intrstatus & flag) == 0)
1183 			continue;
1184 
1185 		found_interrupt = 1;
1186 
1187 		pm = cc->pc_corepmcs[n].phw_pmc;
1188 		if (pm == NULL || pm->pm_state != PMC_STATE_RUNNING ||
1189 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1190 			continue;
1191 
1192 		error = pmc_process_interrupt(cpu, PMC_HR, pm, tf,
1193 		    TRAPF_USERMODE(tf));
1194 		if (error)
1195 			intrenable &= ~flag;
1196 
1197 		v = iap_reload_count_to_perfctr_value(pm->pm_sc.pm_reloadcount);
1198 
1199 		PMCDBG3(MDP,INT, 1, "iap-intr cpu=%d error=%d v=%jx", cpu, error,
1200 		    (uintmax_t) v);
1201 
1202 		/* Reload sampling count. */
1203 		wrmsr(core_iap_wroffset + IAP_PMC0 + n, v);
1204 	}
1205 
1206 	/*
1207 	 * Reenable all non-stalled PMCs.
1208 	 */
1209 	PMCDBG2(MDP,INT, 1, "cpu=%d intrenable=%jx", cpu,
1210 	    (uintmax_t) intrenable);
1211 
1212 	cc->pc_globalctrl |= intrenable;
1213 
1214 	wrmsr(IA_GLOBAL_CTRL, cc->pc_globalctrl & IA_GLOBAL_CTRL_MASK);
1215 
1216 	PMCDBG5(MDP,INT, 1, "cpu=%d fixedctrl=%jx globalctrl=%jx status=%jx "
1217 	    "ovf=%jx", cpu, (uintmax_t) rdmsr(IAF_CTRL),
1218 	    (uintmax_t) rdmsr(IA_GLOBAL_CTRL),
1219 	    (uintmax_t) rdmsr(IA_GLOBAL_STATUS),
1220 	    (uintmax_t) rdmsr(IA_GLOBAL_OVF_CTRL));
1221 
1222 	if (found_interrupt)
1223 		lapic_reenable_pmc();
1224 
1225 	if (found_interrupt)
1226 		counter_u64_add(pmc_stats.pm_intr_processed, 1);
1227 	else
1228 		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
1229 
1230 	return (found_interrupt);
1231 }
1232 
1233 int
1234 pmc_core_initialize(struct pmc_mdep *md, int maxcpu, int version_override)
1235 {
1236 	int cpuid[CORE_CPUID_REQUEST_SIZE];
1237 	int ipa_version, flags, nflags;
1238 
1239 	do_cpuid(CORE_CPUID_REQUEST, cpuid);
1240 
1241 	ipa_version = (version_override > 0) ? version_override :
1242 	    cpuid[CORE_CPUID_EAX] & 0xFF;
1243 	core_cputype = md->pmd_cputype;
1244 
1245 	PMCDBG3(MDP,INI,1,"core-init cputype=%d ncpu=%d ipa-version=%d",
1246 	    core_cputype, maxcpu, ipa_version);
1247 
1248 	if (ipa_version < 1 || ipa_version > 4 ||
1249 	    (core_cputype != PMC_CPU_INTEL_CORE && ipa_version == 1)) {
1250 		/* Unknown PMC architecture. */
1251 		printf("hwpc_core: unknown PMC architecture: %d\n",
1252 		    ipa_version);
1253 		return (EPROGMISMATCH);
1254 	}
1255 
1256 	core_iap_wroffset = 0;
1257 	if (cpu_feature2 & CPUID2_PDCM) {
1258 		if (rdmsr(IA32_PERF_CAPABILITIES) & PERFCAP_FW_WRITE) {
1259 			PMCDBG0(MDP, INI, 1,
1260 			    "core-init full-width write supported");
1261 			core_iap_wroffset = IAP_A_PMC0 - IAP_PMC0;
1262 		} else
1263 			PMCDBG0(MDP, INI, 1,
1264 			    "core-init full-width write NOT supported");
1265 	} else
1266 		PMCDBG0(MDP, INI, 1, "core-init pdcm not supported");
1267 
1268 	core_pmcmask = 0;
1269 
1270 	/*
1271 	 * Initialize programmable counters.
1272 	 */
1273 	core_iap_npmc = (cpuid[CORE_CPUID_EAX] >> 8) & 0xFF;
1274 	core_iap_width = (cpuid[CORE_CPUID_EAX] >> 16) & 0xFF;
1275 
1276 	core_pmcmask |= ((1ULL << core_iap_npmc) - 1);
1277 
1278 	nflags = (cpuid[CORE_CPUID_EAX] >> 24) & 0xFF;
1279 	flags = cpuid[CORE_CPUID_EBX] & ((1 << nflags) - 1);
1280 
1281 	iap_initialize(md, maxcpu, core_iap_npmc, core_iap_width, flags);
1282 
1283 	/*
1284 	 * Initialize fixed function counters, if present.
1285 	 */
1286 	if (core_cputype != PMC_CPU_INTEL_CORE) {
1287 		core_iaf_ri = core_iap_npmc;
1288 		core_iaf_npmc = cpuid[CORE_CPUID_EDX] & 0x1F;
1289 		core_iaf_width = (cpuid[CORE_CPUID_EDX] >> 5) & 0xFF;
1290 
1291 		iaf_initialize(md, maxcpu, core_iaf_npmc, core_iaf_width);
1292 		core_pmcmask |= ((1ULL << core_iaf_npmc) - 1) << IAF_OFFSET;
1293 	}
1294 
1295 	PMCDBG2(MDP,INI,1,"core-init pmcmask=0x%jx iafri=%d", core_pmcmask,
1296 	    core_iaf_ri);
1297 
1298 	core_pcpu = malloc(sizeof(*core_pcpu) * maxcpu, M_PMC,
1299 	    M_ZERO | M_WAITOK);
1300 
1301 	/*
1302 	 * Choose the appropriate interrupt handler.
1303 	 */
1304 	if (ipa_version == 1)
1305 		md->pmd_intr = core_intr;
1306 	else
1307 		md->pmd_intr = core2_intr;
1308 
1309 	md->pmd_pcpu_fini = NULL;
1310 	md->pmd_pcpu_init = NULL;
1311 
1312 	return (0);
1313 }
1314 
1315 void
1316 pmc_core_finalize(struct pmc_mdep *md)
1317 {
1318 	PMCDBG0(MDP,INI,1, "core-finalize");
1319 
1320 	free(core_pcpu, M_PMC);
1321 	core_pcpu = NULL;
1322 }
1323