xref: /freebsd/sys/dev/hwpmc/hwpmc_amd.c (revision a259b98fa211ed87bfee58c575de4e2de94ee0fa)
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 AUTHOR 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 AUTHOR 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 /* Support for the AMD K8 and later processors */
34 
35 #include <sys/param.h>
36 #include <sys/lock.h>
37 #include <sys/malloc.h>
38 #include <sys/mutex.h>
39 #include <sys/pcpu.h>
40 #include <sys/pmc.h>
41 #include <sys/pmckern.h>
42 #include <sys/smp.h>
43 #include <sys/sysctl.h>
44 #include <sys/systm.h>
45 
46 #define	EXTERR_CATEGORY	EXTERR_CAT_HWPMC_AMD
47 #include <sys/exterrvar.h>
48 
49 #include <machine/cpu.h>
50 #include <machine/cpufunc.h>
51 #include <machine/md_var.h>
52 #include <machine/specialreg.h>
53 
54 #define	OVERFLOW_WAIT_COUNT	50
55 
56 DPCPU_DEFINE_STATIC(uint32_t, nmi_counter);
57 
58 /* AMD K8 PMCs */
59 struct amd_descr {
60 	struct pmc_descr pm_descr;   /* "base class" */
61 	uint32_t	pm_evsel;    /* address of EVSEL register */
62 	uint32_t	pm_perfctr;  /* address of PERFCTR register */
63 	enum sub_class	pm_subclass; /* register subclass */
64 };
65 
66 static int amd_npmcs;
67 static int amd_core_npmcs, amd_l3_npmcs, amd_df_npmcs, amd_umc_npmcs;
68 static struct amd_descr amd_pmcdesc[AMD_NPMCS_MAX];
69 struct amd_event_code_map {
70 	enum pmc_event	pe_ev;	 /* enum value */
71 	uint16_t	pe_code; /* encoded event mask */
72 	uint8_t		pe_mask; /* bits allowed in unit mask */
73 };
74 
75 const struct amd_event_code_map amd_event_codes[] = {
76 	{ PMC_EV_K8_FP_DISPATCHED_FPU_OPS,		0x00, 0x3F },
77 	{ PMC_EV_K8_FP_CYCLES_WITH_NO_FPU_OPS_RETIRED,	0x01, 0x00 },
78 	{ PMC_EV_K8_FP_DISPATCHED_FPU_FAST_FLAG_OPS,	0x02, 0x00 },
79 
80 	{ PMC_EV_K8_LS_SEGMENT_REGISTER_LOAD, 		0x20, 0x7F },
81 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SELF_MODIFYING_CODE,
82 	  						0x21, 0x00 },
83 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x22, 0x00 },
84 	{ PMC_EV_K8_LS_BUFFER2_FULL,			0x23, 0x00 },
85 	{ PMC_EV_K8_LS_LOCKED_OPERATION,		0x24, 0x07 },
86 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_LATE_CANCEL,	0x25, 0x00 },
87 	{ PMC_EV_K8_LS_RETIRED_CFLUSH_INSTRUCTIONS,	0x26, 0x00 },
88 	{ PMC_EV_K8_LS_RETIRED_CPUID_INSTRUCTIONS,	0x27, 0x00 },
89 
90 	{ PMC_EV_K8_DC_ACCESS,				0x40, 0x00 },
91 	{ PMC_EV_K8_DC_MISS,				0x41, 0x00 },
92 	{ PMC_EV_K8_DC_REFILL_FROM_L2,			0x42, 0x1F },
93 	{ PMC_EV_K8_DC_REFILL_FROM_SYSTEM,		0x43, 0x1F },
94 	{ PMC_EV_K8_DC_COPYBACK,			0x44, 0x1F },
95 	{ PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_HIT,	0x45, 0x00 },
96 	{ PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_MISS,	0x46, 0x00 },
97 	{ PMC_EV_K8_DC_MISALIGNED_DATA_REFERENCE,	0x47, 0x00 },
98 	{ PMC_EV_K8_DC_MICROARCHITECTURAL_LATE_CANCEL,	0x48, 0x00 },
99 	{ PMC_EV_K8_DC_MICROARCHITECTURAL_EARLY_CANCEL, 0x49, 0x00 },
100 	{ PMC_EV_K8_DC_ONE_BIT_ECC_ERROR,		0x4A, 0x03 },
101 	{ PMC_EV_K8_DC_DISPATCHED_PREFETCH_INSTRUCTIONS, 0x4B, 0x07 },
102 	{ PMC_EV_K8_DC_DCACHE_ACCESSES_BY_LOCKS,	0x4C, 0x03 },
103 
104 	{ PMC_EV_K8_BU_CPU_CLK_UNHALTED,		0x76, 0x00 },
105 	{ PMC_EV_K8_BU_INTERNAL_L2_REQUEST,		0x7D, 0x1F },
106 	{ PMC_EV_K8_BU_FILL_REQUEST_L2_MISS,		0x7E, 0x07 },
107 	{ PMC_EV_K8_BU_FILL_INTO_L2,			0x7F, 0x03 },
108 
109 	{ PMC_EV_K8_IC_FETCH,				0x80, 0x00 },
110 	{ PMC_EV_K8_IC_MISS,				0x81, 0x00 },
111 	{ PMC_EV_K8_IC_REFILL_FROM_L2,			0x82, 0x00 },
112 	{ PMC_EV_K8_IC_REFILL_FROM_SYSTEM,		0x83, 0x00 },
113 	{ PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_HIT,	0x84, 0x00 },
114 	{ PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_MISS,	0x85, 0x00 },
115 	{ PMC_EV_K8_IC_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x86, 0x00 },
116 	{ PMC_EV_K8_IC_INSTRUCTION_FETCH_STALL,		0x87, 0x00 },
117 	{ PMC_EV_K8_IC_RETURN_STACK_HIT,		0x88, 0x00 },
118 	{ PMC_EV_K8_IC_RETURN_STACK_OVERFLOW,		0x89, 0x00 },
119 
120 	{ PMC_EV_K8_FR_RETIRED_X86_INSTRUCTIONS,	0xC0, 0x00 },
121 	{ PMC_EV_K8_FR_RETIRED_UOPS,			0xC1, 0x00 },
122 	{ PMC_EV_K8_FR_RETIRED_BRANCHES,		0xC2, 0x00 },
123 	{ PMC_EV_K8_FR_RETIRED_BRANCHES_MISPREDICTED,	0xC3, 0x00 },
124 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES,		0xC4, 0x00 },
125 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED, 0xC5, 0x00 },
126 	{ PMC_EV_K8_FR_RETIRED_FAR_CONTROL_TRANSFERS,	0xC6, 0x00 },
127 	{ PMC_EV_K8_FR_RETIRED_RESYNCS,			0xC7, 0x00 },
128 	{ PMC_EV_K8_FR_RETIRED_NEAR_RETURNS,		0xC8, 0x00 },
129 	{ PMC_EV_K8_FR_RETIRED_NEAR_RETURNS_MISPREDICTED, 0xC9, 0x00 },
130 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED_BY_ADDR_MISCOMPARE,
131 							0xCA, 0x00 },
132 	{ PMC_EV_K8_FR_RETIRED_FPU_INSTRUCTIONS,	0xCB, 0x0F },
133 	{ PMC_EV_K8_FR_RETIRED_FASTPATH_DOUBLE_OP_INSTRUCTIONS,
134 							0xCC, 0x07 },
135 	{ PMC_EV_K8_FR_INTERRUPTS_MASKED_CYCLES,	0xCD, 0x00 },
136 	{ PMC_EV_K8_FR_INTERRUPTS_MASKED_WHILE_PENDING_CYCLES, 0xCE, 0x00 },
137 	{ PMC_EV_K8_FR_TAKEN_HARDWARE_INTERRUPTS,	0xCF, 0x00 },
138 
139 	{ PMC_EV_K8_FR_DECODER_EMPTY,			0xD0, 0x00 },
140 	{ PMC_EV_K8_FR_DISPATCH_STALLS,			0xD1, 0x00 },
141 	{ PMC_EV_K8_FR_DISPATCH_STALL_FROM_BRANCH_ABORT_TO_RETIRE,
142 							0xD2, 0x00 },
143 	{ PMC_EV_K8_FR_DISPATCH_STALL_FOR_SERIALIZATION, 0xD3, 0x00 },
144 	{ PMC_EV_K8_FR_DISPATCH_STALL_FOR_SEGMENT_LOAD,	0xD4, 0x00 },
145 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_REORDER_BUFFER_IS_FULL,
146 							0xD5, 0x00 },
147 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_RESERVATION_STATIONS_ARE_FULL,
148 							0xD6, 0x00 },
149 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FPU_IS_FULL,	0xD7, 0x00 },
150 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_LS_IS_FULL,	0xD8, 0x00 },
151 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_WAITING_FOR_ALL_TO_BE_QUIET,
152 							0xD9, 0x00 },
153 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FAR_XFER_OR_RESYNC_BRANCH_PENDING,
154 							0xDA, 0x00 },
155 	{ PMC_EV_K8_FR_FPU_EXCEPTIONS,			0xDB, 0x0F },
156 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR0,	0xDC, 0x00 },
157 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR1,	0xDD, 0x00 },
158 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR2,	0xDE, 0x00 },
159 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR3,	0xDF, 0x00 },
160 
161 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_ACCESS_EVENT, 0xE0, 0x7 },
162 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_TABLE_OVERFLOW, 0xE1, 0x00 },
163 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_DRAM_COMMAND_SLOTS_MISSED,
164 							0xE2, 0x00 },
165 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_TURNAROUND,	0xE3, 0x07 },
166 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_BYPASS_SATURATION, 0xE4, 0x0F },
167 	{ PMC_EV_K8_NB_SIZED_COMMANDS,			0xEB, 0x7F },
168 	{ PMC_EV_K8_NB_PROBE_RESULT,			0xEC, 0x0F },
169 	{ PMC_EV_K8_NB_HT_BUS0_BANDWIDTH,		0xF6, 0x0F },
170 	{ PMC_EV_K8_NB_HT_BUS1_BANDWIDTH,		0xF7, 0x0F },
171 	{ PMC_EV_K8_NB_HT_BUS2_BANDWIDTH,		0xF8, 0x0F }
172 
173 };
174 
175 const int amd_event_codes_size = nitems(amd_event_codes);
176 
177 /*
178  * Per-processor information
179  */
180 struct amd_cpu {
181 	struct pmc_hw	pc_amdpmcs[AMD_NPMCS_MAX];
182 };
183 static struct amd_cpu **amd_pcpu;
184 
185 /* Populated by amd_init_policy(); PRECISERETIRE is OR-ed in per-allocation. */
186 static uint64_t amd_core_allowed_mask;
187 static uint64_t amd_l3_allowed_mask;
188 static uint64_t amd_df_allowed_mask;
189 static uint64_t amd_umc_allowed_mask;
190 
191 static uint64_t amd_core_extra_mask;
192 static uint64_t amd_l3_extra_mask;
193 static uint64_t amd_df_extra_mask;
194 static uint64_t amd_umc_extra_mask;
195 
196 SYSCTL_DECL(_kern_hwpmc);
197 
198 SYSCTL_U64(_kern_hwpmc, OID_AUTO, amd_core_extra_mask, CTLFLAG_RDTUN,
199     &amd_core_extra_mask, 0,
200     "Extra allowed bits in AMD core PMU PERFEVTSEL (override; default 0)");
201 
202 SYSCTL_U64(_kern_hwpmc, OID_AUTO, amd_l3_extra_mask, CTLFLAG_RDTUN,
203     &amd_l3_extra_mask, 0,
204     "Extra allowed bits in AMD L3 PMU control (override; default 0)");
205 
206 SYSCTL_U64(_kern_hwpmc, OID_AUTO, amd_df_extra_mask, CTLFLAG_RDTUN,
207     &amd_df_extra_mask, 0,
208     "Extra allowed bits in AMD DF PMU control (override; default 0)");
209 
210 SYSCTL_U64(_kern_hwpmc, OID_AUTO, amd_umc_extra_mask, CTLFLAG_RDTUN,
211     &amd_umc_extra_mask, 0,
212     "Extra allowed bits in AMD UMC PMU control (override; default 0)");
213 
214 static void
215 amd_init_policy(void)
216 {
217 	int family;
218 
219 	family = CPUID_TO_FAMILY(cpu_id);
220 
221 	amd_core_allowed_mask = AMD_VALID_BITS;
222 
223 	amd_l3_allowed_mask = (family <= 0x17) ?
224 	    AMD_PMC_L3_FAMILY17_MASK : AMD_PMC_L3_FAMILY19_MASK;
225 
226 	amd_df_allowed_mask = (family <= 0x19) ?
227 	    AMD_PMC_DF_FAMILY17_MASK : AMD_PMC_DF_FAMILY1A_MASK;
228 
229 	amd_umc_allowed_mask = AMD_PMC_UMC_MASK;
230 }
231 
232 static uint64_t
233 amd_config_mask(enum sub_class subclass, uint64_t caps)
234 {
235 
236 	switch (subclass) {
237 	case PMC_AMD_SUB_CLASS_CORE:
238 		return (amd_core_allowed_mask | amd_core_extra_mask |
239 		    (((caps & PMC_CAP_PRECISE) != 0) ?
240 		    AMD_PMC_PRECISERETIRE : 0));
241 	case PMC_AMD_SUB_CLASS_L3_CACHE:
242 		return (amd_l3_allowed_mask | amd_l3_extra_mask);
243 	case PMC_AMD_SUB_CLASS_DATA_FABRIC:
244 		return (amd_df_allowed_mask | amd_df_extra_mask);
245 	case PMC_AMD_SUB_CLASS_UMC:
246 		return (amd_umc_allowed_mask | amd_umc_extra_mask);
247 	default:
248 		return (0);
249 	}
250 }
251 
252 /*
253  * Read a PMC value from the MSR.
254  */
255 static int
256 amd_read_pmc(int cpu, int ri, struct pmc *pm, pmc_value_t *v)
257 {
258 	const struct amd_descr *pd;
259 	pmc_value_t tmp;
260 	enum pmc_mode mode;
261 
262 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
263 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
264 	KASSERT(ri >= 0 && ri < amd_npmcs,
265 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
266 	KASSERT(amd_pcpu[cpu],
267 	    ("[amd,%d] null per-cpu, cpu %d", __LINE__, cpu));
268 
269 	pd = &amd_pmcdesc[ri];
270 	mode = PMC_TO_MODE(pm);
271 
272 	PMCDBG2(MDP, REA, 1, "amd-read id=%d class=%d", ri,
273 	    pd->pm_descr.pd_class);
274 
275 	tmp = rdmsr(pd->pm_perfctr); /* RDMSR serializes */
276 	PMCDBG2(MDP, REA, 2, "amd-read (pre-munge) id=%d -> %jd", ri, tmp);
277 	if (PMC_IS_SAMPLING_MODE(mode)) {
278 		/*
279 		 * Clamp value to 0 if the counter just overflowed,
280 		 * otherwise the returned reload count would wrap to a
281 		 * huge value.
282 		 */
283 		if ((tmp & (1ULL << 47)) == 0)
284 			tmp = 0;
285 		else {
286 			/* Sign extend 48 bit value to 64 bits. */
287 			tmp = (pmc_value_t) ((int64_t)(tmp << 16) >> 16);
288 			tmp = AMD_PERFCTR_VALUE_TO_RELOAD_COUNT(tmp);
289 		}
290 	}
291 	*v = tmp;
292 
293 	PMCDBG2(MDP, REA, 2, "amd-read (post-munge) id=%d -> %jd", ri, *v);
294 
295 	return (0);
296 }
297 
298 /*
299  * Write a PMC MSR.
300  */
301 static int
302 amd_write_pmc(int cpu, int ri, struct pmc *pm, pmc_value_t v)
303 {
304 	const struct amd_descr *pd;
305 	enum pmc_mode mode;
306 
307 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
308 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
309 	KASSERT(ri >= 0 && ri < amd_npmcs,
310 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
311 
312 	pd = &amd_pmcdesc[ri];
313 	mode = PMC_TO_MODE(pm);
314 
315 	/* use 2's complement of the count for sampling mode PMCs */
316 	if (PMC_IS_SAMPLING_MODE(mode))
317 		v = AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v);
318 
319 	PMCDBG3(MDP, WRI, 1, "amd-write cpu=%d ri=%d v=%jx", cpu, ri, v);
320 
321 	/* write the PMC value */
322 	wrmsr(pd->pm_perfctr, v);
323 	return (0);
324 }
325 
326 /*
327  * Configure hardware PMC according to the configuration recorded in 'pm'.
328  */
329 static int
330 amd_config_pmc(int cpu, int ri, struct pmc *pm)
331 {
332 	struct pmc_hw *phw;
333 
334 	PMCDBG3(MDP, CFG, 1, "cpu=%d ri=%d pm=%p", cpu, ri, pm);
335 
336 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
337 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
338 	KASSERT(ri >= 0 && ri < amd_npmcs,
339 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
340 
341 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
342 
343 	KASSERT(pm == NULL || phw->phw_pmc == NULL,
344 	    ("[amd,%d] pm=%p phw->pm=%p hwpmc not unconfigured",
345 		__LINE__, pm, phw->phw_pmc));
346 
347 	phw->phw_pmc = pm;
348 	return (0);
349 }
350 
351 /*
352  * Retrieve a configured PMC pointer from hardware state.
353  */
354 static int
355 amd_get_config(int cpu, int ri, struct pmc **ppm)
356 {
357 	*ppm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc;
358 	return (0);
359 }
360 
361 /*
362  * Machine-dependent actions taken during the context switch in of a
363  * thread.
364  */
365 static int
366 amd_switch_in(struct pmc_cpu *pc __pmcdbg_used, struct pmc_process *pp)
367 {
368 	PMCDBG3(MDP, SWI, 1, "pc=%p pp=%p enable-msr=%d", pc, pp,
369 	    (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0);
370 
371 	/* enable the RDPMC instruction if needed */
372 	if (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS)
373 		load_cr4(rcr4() | CR4_PCE);
374 
375 	return (0);
376 }
377 
378 /*
379  * Machine-dependent actions taken during the context switch out of a
380  * thread.
381  */
382 static int
383 amd_switch_out(struct pmc_cpu *pc __pmcdbg_used,
384     struct pmc_process *pp __pmcdbg_used)
385 {
386 	PMCDBG3(MDP, SWO, 1, "pc=%p pp=%p enable-msr=%d", pc, pp, pp ?
387 	    (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) == 1 : 0);
388 
389 	/* always turn off the RDPMC instruction */
390 	load_cr4(rcr4() & ~CR4_PCE);
391 
392 	return (0);
393 }
394 
395 /*
396  * Check if a given PMC allocation is feasible.
397  */
398 static int
399 amd_allocate_pmc(int cpu __unused, int ri, struct pmc *pm,
400     const struct pmc_op_pmcallocate *a)
401 {
402 	const struct pmc_descr *pd;
403 	uint64_t allowed_unitmask, caps, config, unitmask;
404 	enum pmc_event pe;
405 	int i;
406 
407 	KASSERT(ri >= 0 && ri < amd_npmcs,
408 	    ("[amd,%d] illegal row index %d", __LINE__, ri));
409 
410 	pd = &amd_pmcdesc[ri].pm_descr;
411 
412 	/* check class match */
413 	if (pd->pd_class != a->pm_class)
414 		return (EINVAL);
415 
416 	caps = pm->pm_caps;
417 
418 	PMCDBG2(MDP, ALL, 1,"amd-allocate ri=%d caps=0x%x", ri, caps);
419 
420 	/* Validate sub-class. */
421 	if (amd_pmcdesc[ri].pm_subclass != a->pm_md.pm_amd.pm_amd_sub_class)
422 		return (EINVAL);
423 
424 	if (((caps & PMC_CAP_PRECISE) != 0) &&
425 	    ((pd->pd_caps & PMC_CAP_PRECISE) == 0))
426 		return (EINVAL);
427 
428 	/* PMC_F_EV_PMU: config comes from pmu-events tables. */
429 	if ((a->pm_flags & PMC_F_EV_PMU) != 0) {
430 		config = a->pm_md.pm_amd.pm_amd_config;
431 		if ((config & ~amd_config_mask(amd_pmcdesc[ri].pm_subclass,
432 		    caps)) != 0)
433 			return (EXTERROR(EINVAL,
434 			    "AMD PMU config has unsupported bits %#jx",
435 			    (uintmax_t)(config & ~amd_config_mask(
436 			    amd_pmcdesc[ri].pm_subclass, caps))));
437 		pm->pm_md.pm_amd.pm_amd_evsel = config;
438 		PMCDBG2(MDP, ALL, 2, "amd-allocate ri=%d -> config=0x%jx",
439 		    ri, (uintmax_t)config);
440 		return (0);
441 	}
442 
443 	/*
444 	 * Everything below this is for supporting older processors.
445 	 */
446 	pe = a->pm_ev;
447 
448 	/* map ev to the correct event mask code */
449 	config = allowed_unitmask = 0;
450 	for (i = 0; i < amd_event_codes_size; i++) {
451 		if (amd_event_codes[i].pe_ev == pe) {
452 			config =
453 			    AMD_PMC_TO_EVENTMASK(amd_event_codes[i].pe_code);
454 			allowed_unitmask =
455 			    AMD_PMC_TO_UNITMASK(amd_event_codes[i].pe_mask);
456 			break;
457 		}
458 	}
459 	if (i == amd_event_codes_size)
460 		return (EXTERROR(EINVAL,
461 		    "AMD legacy event %ju is not supported",
462 		    (uintmax_t)pe));
463 
464 	unitmask = a->pm_md.pm_amd.pm_amd_config & AMD_PMC_UNITMASK;
465 	if ((unitmask & ~allowed_unitmask) != 0) /* disallow reserved bits */
466 		return (EXTERROR(EINVAL,
467 		    "AMD unitmask %#jx exceeds allowed mask %#jx",
468 		    (uintmax_t)unitmask, (uintmax_t)allowed_unitmask));
469 
470 	if (unitmask && (caps & PMC_CAP_QUALIFIER) != 0)
471 		config |= unitmask;
472 
473 	if ((caps & PMC_CAP_THRESHOLD) != 0)
474 		config |= a->pm_md.pm_amd.pm_amd_config & AMD_PMC_COUNTERMASK;
475 
476 	/* Set at least one of the 'usr' or 'os' caps. */
477 	if ((caps & PMC_CAP_USER) != 0)
478 		config |= AMD_PMC_USR;
479 	if ((caps & PMC_CAP_SYSTEM) != 0)
480 		config |= AMD_PMC_OS;
481 	if ((caps & (PMC_CAP_USER | PMC_CAP_SYSTEM)) == 0)
482 		config |= (AMD_PMC_USR|AMD_PMC_OS);
483 
484 	if ((caps & PMC_CAP_EDGE) != 0)
485 		config |= AMD_PMC_EDGE;
486 	if ((caps & PMC_CAP_INVERT) != 0)
487 		config |= AMD_PMC_INVERT;
488 	if ((caps & PMC_CAP_INTERRUPT) != 0)
489 		config |= AMD_PMC_INT;
490 
491 	pm->pm_md.pm_amd.pm_amd_evsel = config; /* save config value */
492 
493 	PMCDBG2(MDP, ALL, 2, "amd-allocate ri=%d -> config=0x%x", ri, config);
494 
495 	return (0);
496 }
497 
498 /*
499  * Release machine dependent state associated with a PMC.  This is a
500  * no-op on this architecture.
501  */
502 static int
503 amd_release_pmc(int cpu, int ri, struct pmc *pmc __unused)
504 {
505 	struct pmc_hw *phw __diagused;
506 
507 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
508 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
509 	KASSERT(ri >= 0 && ri < amd_npmcs,
510 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
511 
512 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
513 
514 	KASSERT(phw->phw_pmc == NULL,
515 	    ("[amd,%d] PHW pmc %p non-NULL", __LINE__, phw->phw_pmc));
516 
517 	return (0);
518 }
519 
520 /*
521  * Start a PMC.
522  */
523 static int
524 amd_start_pmc(int cpu __diagused, int ri, struct pmc *pm)
525 {
526 	const struct amd_descr *pd;
527 	uint64_t config;
528 
529 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
530 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
531 	KASSERT(ri >= 0 && ri < amd_npmcs,
532 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
533 
534 	pd = &amd_pmcdesc[ri];
535 
536 	PMCDBG2(MDP, STA, 1, "amd-start cpu=%d ri=%d", cpu, ri);
537 
538 	/*
539 	 * Asserts triggered by DF/UMC counters because all DF/UMC MSRs are
540 	 * shared.  While userspace now honors the per-node flags, we should
541 	 * enforce this in the kernel.
542 	 */
543 	/* turn on the PMC ENABLE bit */
544 	if (pd->pm_subclass == PMC_AMD_SUB_CLASS_UMC) {
545 		KASSERT(AMD_PMC_UMC_IS_STOPPED(pd->pm_evsel),
546 		    ("[amd,%d] pmc%d,cpu%d: Starting active PMC \"%s\"",
547 		    __LINE__, ri, cpu, pd->pm_descr.pd_name));
548 
549 		config = pm->pm_md.pm_amd.pm_amd_evsel | AMD_PMC_UMC_ENABLE;
550 	} else {
551 		KASSERT(AMD_PMC_IS_STOPPED(pd->pm_evsel),
552 		    ("[amd,%d] pmc%d,cpu%d: Starting active PMC \"%s\"",
553 		    __LINE__, ri, cpu, pd->pm_descr.pd_name));
554 
555 		config = pm->pm_md.pm_amd.pm_amd_evsel | AMD_PMC_ENABLE;
556 	}
557 
558 	PMCDBG1(MDP, STA, 2, "amd-start config=0x%x", config);
559 
560 	wrmsr(pd->pm_evsel, config);
561 	return (0);
562 }
563 
564 /*
565  * Stop a PMC.
566  */
567 static int
568 amd_stop_pmc(int cpu __diagused, int ri, struct pmc *pm)
569 {
570 	const struct amd_descr *pd;
571 	uint64_t config;
572 	int i;
573 
574 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
575 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
576 	KASSERT(ri >= 0 && ri < amd_npmcs,
577 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
578 
579 	pd = &amd_pmcdesc[ri];
580 
581 	PMCDBG1(MDP, STO, 1, "amd-stop ri=%d", ri);
582 
583 	/* turn off the PMC ENABLE bit */
584 	if (pd->pm_subclass == PMC_AMD_SUB_CLASS_UMC) {
585 		KASSERT(!AMD_PMC_UMC_IS_STOPPED(pd->pm_evsel),
586 		    ("[amd,%d] PMC%d, CPU%d \"%s\" already stopped",
587 		    __LINE__, ri, cpu, pd->pm_descr.pd_name));
588 
589 		config = pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_UMC_ENABLE;
590 	} else {
591 		KASSERT(!AMD_PMC_IS_STOPPED(pd->pm_evsel),
592 		    ("[amd,%d] PMC%d, CPU%d \"%s\" already stopped",
593 		    __LINE__, ri, cpu, pd->pm_descr.pd_name));
594 
595 		config = pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE;
596 	}
597 	wrmsr(pd->pm_evsel, config);
598 
599 	/*
600 	 * Due to NMI latency on newer AMD processors
601 	 * NMI interrupts are ignored, which leads to
602 	 * panic or messages based on kernel configuration
603 	 */
604 
605 	/* Wait for the count to be reset */
606 	for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
607 		if (rdmsr(pd->pm_perfctr) & (1 << (pd->pm_descr.pd_width - 1)))
608 			break;
609 
610 		DELAY(1);
611 	}
612 
613 	return (0);
614 }
615 
616 /*
617  * Interrupt handler.  This function needs to return '1' if the
618  * interrupt was this CPU's PMCs or '0' otherwise.  It is not allowed
619  * to sleep or do anything a 'fast' interrupt handler is not allowed
620  * to do.
621  */
622 static int
623 amd_intr(struct trapframe *tf)
624 {
625 	struct amd_cpu *pac;
626 	struct pmc *pm;
627 	pmc_value_t v;
628 	uint64_t config, evsel, perfctr;
629 	uint32_t active = 0, count = 0;
630 	int i, error, retval, cpu;
631 
632 	cpu = curcpu;
633 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
634 	    ("[amd,%d] out of range CPU %d", __LINE__, cpu));
635 
636 	PMCDBG3(MDP, INT, 1, "cpu=%d tf=%p um=%d", cpu, tf, TRAPF_USERMODE(tf));
637 
638 	retval = 0;
639 
640 	pac = amd_pcpu[cpu];
641 
642 	retval = pmc_ibs_intr(tf);
643 	if (retval)
644 		goto done;
645 
646 	/*
647 	 * look for all PMCs that have interrupted:
648 	 * - look for a running, sampling PMC which has overflowed
649 	 *   and which has a valid 'struct pmc' association
650 	 *
651 	 * If found, we call a helper to process the interrupt.
652 	 *
653 	 * PMCs interrupting at the same time are collapsed into
654 	 * a single interrupt. Check all the valid pmcs for
655 	 * overflow.
656 	 */
657 	for (i = 0; i < amd_npmcs; i++) {
658 		if (amd_pmcdesc[i].pm_subclass != PMC_AMD_SUB_CLASS_CORE)
659 		    break;
660 
661 		if ((pm = pac->pc_amdpmcs[i].phw_pmc) == NULL ||
662 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) {
663 			continue;
664 		}
665 
666 		/* Consider pmc with valid handle as active */
667 		active++;
668 
669 		if (!AMD_PMC_HAS_OVERFLOWED(i))
670 			continue;
671 
672 		retval = 1;	/* Found an interrupting PMC. */
673 
674 		if (pm->pm_state != PMC_STATE_RUNNING)
675 			continue;
676 
677 		/* Stop the PMC, reload count. */
678 		evsel   = amd_pmcdesc[i].pm_evsel;
679 		perfctr = amd_pmcdesc[i].pm_perfctr;
680 		v       = pm->pm_sc.pm_reloadcount;
681 		config  = rdmsr(evsel);
682 
683 
684 		if (amd_pmcdesc[i].pm_subclass == PMC_AMD_SUB_CLASS_UMC) {
685 			KASSERT((config & ~AMD_PMC_UMC_ENABLE) ==
686 			    (pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_UMC_ENABLE),
687 			    ("[amd,%d] config mismatch reg=0x%jx pm=0x%jx", __LINE__,
688 			    (uintmax_t)config, (uintmax_t)pm->pm_md.pm_amd.pm_amd_evsel));
689 
690 			wrmsr(evsel, config & ~AMD_PMC_UMC_ENABLE);
691 		} else {
692 			KASSERT((config & ~AMD_PMC_ENABLE) ==
693 			    (pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE),
694 			    ("[amd,%d] config mismatch reg=0x%jx pm=0x%jx", __LINE__,
695 			    (uintmax_t)config, (uintmax_t)pm->pm_md.pm_amd.pm_amd_evsel));
696 
697 			wrmsr(evsel, config & ~AMD_PMC_ENABLE);
698 		}
699 		wrmsr(perfctr, AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v));
700 
701 		/* Restart the counter if logging succeeded. */
702 		error = pmc_process_interrupt(PMC_HR, pm, tf);
703 		if (error == 0)
704 			wrmsr(evsel, config);
705 	}
706 
707 	/*
708 	 * Due to NMI latency, there can be a scenario in which
709 	 * multiple pmcs gets serviced in an earlier NMI and we
710 	 * do not find an overflow in the subsequent NMI.
711 	 *
712 	 * For such cases we keep a per-cpu count of active NMIs
713 	 * and compare it with min(active pmcs, 2) to determine
714 	 * if this NMI was for a pmc overflow which was serviced
715 	 * in an earlier request or should be ignored.
716 	 */
717 	if (retval) {
718 		DPCPU_SET(nmi_counter, min(2, active));
719 	} else {
720 		if ((count = DPCPU_GET(nmi_counter))) {
721 			retval = 1;
722 			DPCPU_SET(nmi_counter, --count);
723 		}
724 	}
725 
726 done:
727 	if (retval)
728 		counter_u64_add(pmc_stats.pm_intr_processed, 1);
729 	else
730 		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
731 
732 	PMCDBG1(MDP, INT, 2, "retval=%d", retval);
733 	return (retval);
734 }
735 
736 /*
737  * Describe a PMC.
738  */
739 static int
740 amd_describe(int cpu, int ri, struct pmc_info *pi, struct pmc **ppmc)
741 {
742 	const struct amd_descr *pd;
743 	struct pmc_hw *phw;
744 
745 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
746 	    ("[amd,%d] illegal CPU %d", __LINE__, cpu));
747 	KASSERT(ri >= 0 && ri < amd_npmcs,
748 	    ("[amd,%d] row-index %d out of range", __LINE__, ri));
749 
750 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
751 	pd  = &amd_pmcdesc[ri];
752 
753 	strlcpy(pi->pm_name, pd->pm_descr.pd_name, sizeof(pi->pm_name));
754 	pi->pm_class = pd->pm_descr.pd_class;
755 
756 	if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) != 0) {
757 		pi->pm_enabled = true;
758 		*ppmc          = phw->phw_pmc;
759 	} else {
760 		pi->pm_enabled = false;
761 		*ppmc          = NULL;
762 	}
763 
764 	return (0);
765 }
766 
767 /*
768  * Return the MSR address of the given PMC.
769  */
770 static int
771 amd_get_msr(int ri, uint32_t *msr)
772 {
773 	int df_idx;
774 
775 	KASSERT(ri >= 0 && ri < amd_npmcs,
776 	    ("[amd,%d] ri %d out of range", __LINE__, ri));
777 
778 	/*
779 	 * Map counter row index to RDPMC ECX value.
780 	 *
781 	 * AMD BKDG 24594 rev 3.37, page 440,
782 	 * "RDPMC Read Performance-Monitoring Counter":
783 	 *   ECX 0-5:   Core counters 0-5
784 	 *   ECX 6-9:   DF/Northbridge counters 0-3
785 	 *   ECX 10-15: L3 Cache counters 0-5
786 	 *   ECX 16-27: DF/Northbridge counters 4-15
787 	 *
788 	 * AMD PPR 57930-A0 section 2.1.9,
789 	 * "Register Sharing" for DF counter details.
790 	 */
791 	if (ri < amd_core_npmcs) {
792 		/* ECX 0-5: Core counters */
793 		*msr = ri;
794 	} else if (ri < amd_core_npmcs + amd_l3_npmcs) {
795 		/* ECX 10-15: L3 Cache counters */
796 		*msr = 10 + (ri - amd_core_npmcs);
797 	} else {
798 		/* ECX 6-9: DF counters 0-3
799 		 * ECX 16-27: DF counters 4-15 */
800 		df_idx = ri - amd_core_npmcs - amd_l3_npmcs;
801 		if (df_idx < 4)
802 			*msr = 6 + df_idx;
803 		else if (df_idx < 16)
804 			*msr = 16 + (df_idx - 4);
805 		else
806 			return (EINVAL);
807 	}
808 	return (0);
809 }
810 
811 /*
812  * Return the capabilities of the given PMC.
813  */
814 static int
815 amd_get_caps(int ri, uint32_t *caps)
816 {
817 	KASSERT(ri >= 0 && ri < amd_npmcs,
818 	    ("[amd,%d] ri %d out of range", __LINE__, ri));
819 
820 	*caps = amd_pmcdesc[ri].pm_descr.pd_caps;
821 
822 	return (0);
823 }
824 
825 /*
826  * Processor-dependent initialization.
827  */
828 static int
829 amd_pcpu_init(struct pmc_mdep *md, int cpu)
830 {
831 	struct amd_cpu *pac;
832 	struct pmc_cpu *pc;
833 	struct pmc_hw  *phw;
834 	int first_ri, n;
835 
836 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
837 	    ("[amd,%d] insane cpu number %d", __LINE__, cpu));
838 
839 	PMCDBG1(MDP, INI, 1, "amd-init cpu=%d", cpu);
840 
841 	amd_pcpu[cpu] = pac = malloc(sizeof(struct amd_cpu), M_PMC,
842 	    M_WAITOK | M_ZERO);
843 
844 	/*
845 	 * Set the content of the hardware descriptors to a known
846 	 * state and initialize pointers in the MI per-cpu descriptor.
847 	 */
848 	pc = pmc_pcpu[cpu];
849 	first_ri = md->pmd_classdep[PMC_MDEP_CLASS_INDEX_K8].pcd_ri;
850 
851 	KASSERT(pc != NULL, ("[amd,%d] NULL per-cpu pointer", __LINE__));
852 
853 	for (n = 0, phw = pac->pc_amdpmcs; n < amd_npmcs; n++, phw++) {
854 		phw->phw_state = PMC_PHW_FLAG_IS_ENABLED |
855 		    PMC_PHW_CPU_TO_STATE(cpu) | PMC_PHW_INDEX_TO_STATE(n);
856 		phw->phw_pmc = NULL;
857 		pc->pc_hwpmcs[n + first_ri] = phw;
858 	}
859 
860 	return (0);
861 }
862 
863 /*
864  * Processor-dependent cleanup prior to the KLD being unloaded.
865  */
866 static int
867 amd_pcpu_fini(struct pmc_mdep *md, int cpu)
868 {
869 	struct amd_cpu *pac;
870 	struct pmc_cpu *pc;
871 	int first_ri, i;
872 
873 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
874 	    ("[amd,%d] insane cpu number (%d)", __LINE__, cpu));
875 
876 	PMCDBG1(MDP, INI, 1, "amd-cleanup cpu=%d", cpu);
877 
878 	/*
879 	 * Next, free up allocated space.
880 	 */
881 	if ((pac = amd_pcpu[cpu]) == NULL)
882 		return (0);
883 
884 	amd_pcpu[cpu] = NULL;
885 
886 #ifdef	HWPMC_DEBUG
887 	for (i = 0; i < AMD_NPMCS_K8; i++) {
888 		KASSERT(pac->pc_amdpmcs[i].phw_pmc == NULL,
889 		    ("[amd,%d] CPU%d/PMC%d in use", __LINE__, cpu, i));
890 		KASSERT(AMD_PMC_IS_STOPPED(AMD_PMC_EVSEL_0 + i),
891 		    ("[amd,%d] CPU%d/PMC%d not stopped", __LINE__, cpu, i));
892 	}
893 #endif
894 
895 	pc = pmc_pcpu[cpu];
896 	KASSERT(pc != NULL, ("[amd,%d] NULL per-cpu state", __LINE__));
897 
898 	first_ri = md->pmd_classdep[PMC_MDEP_CLASS_INDEX_K8].pcd_ri;
899 
900 	/*
901 	 * Reset pointers in the MI 'per-cpu' state.
902 	 */
903 	for (i = 0; i < amd_npmcs; i++)
904 		pc->pc_hwpmcs[i + first_ri] = NULL;
905 
906 	free(pac, M_PMC);
907 	return (0);
908 }
909 
910 /*
911  * Check that the PMC hardware is safe to use.  First, we check that the PMCs
912  * are not in use by firmware or another module.  Second, if none of the PMC
913  * feature flags are set, we check that the event selector is working, because
914  * virtual machines have no way to communicate the absence of PMCs.
915  */
916 static int
917 amd_hwcheck(void)
918 {
919 	uint64_t reg;
920 	int error, i;
921 
922 	/*
923 	 * Some PC vendors enable the core counters in firmware to track
924 	 * performance.  The best guess is that this is being used to control
925 	 * power management from within the SMM mode.  We shouldn't just take
926 	 * over the PMCs in this case.  The user should try disabling any
927 	 * performance monitoring or power management functions in the BIOS to
928 	 * safely make use of the counters.
929 	 */
930 	for (i = 0; i < amd_core_npmcs; i++) {
931 		error = rdmsr_safe(amd_pmcdesc[i].pm_evsel, &reg);
932 		if (error != 0) {
933 			printf("hwpmc: AMD evsel %d rdmsr failed!\n", i);
934 			return (-1);
935 		}
936 
937 		if ((reg & AMD_PMC_ENABLE) != 0) {
938 			printf("hwpmc: PMCs maybe in use by firmware!\n");
939 			printf("hwpmc: Disable the PMC use in the BIOS before loading\n");
940 			return (-1);
941 		}
942 	}
943 
944 	/*
945 	 * Unfortunately, there is no way to communicate that the original four
946 	 * core counters are disabled through CPUIDs alone.  We attempt to
947 	 * write and read back the MSR to validate that it is working.
948 	 *
949 	 * Referenced the BIOS and Kernel Developer Guide for AMD Athlon 64 and
950 	 * AMD Opteron Processors 26094 Rev. 3.24 January, 2005 to ensure these
951 	 * fields are valid.
952 	 */
953 	if ((amd_feature2 & AMDID2_PCXC) == 0) {
954 		error = wrmsr_safe(AMD_PMC_EVSEL_0, AMD_PMC_OS | AMD_PMC_USR);
955 		if (error != 0) {
956 			printf("hwpmc: AMD evsel 0 wrmsr failed!\n");
957 			return (-1);
958 		}
959 
960 		error = rdmsr_safe(AMD_PMC_EVSEL_0, &reg);
961 		if (error != 0) {
962 			printf("hwpmc: AMD evsel 0 rdmsr failed!\n");
963 			return (-1);
964 		}
965 
966 		if (reg == 0) {
967 			printf("hwpmc: AMD evsel returned invalid value! "
968 			    "You may be in a VM without PMC support.\n");
969 			return (-1);
970 		}
971 
972 		wrmsr(AMD_PMC_EVSEL_0, 0);
973 	}
974 
975 	return (0);
976 }
977 
978 /*
979  * Initialize ourselves.
980  */
981 struct pmc_mdep *
982 pmc_amd_initialize(void)
983 {
984 	u_int regs[4];
985 	struct amd_descr *d;
986 	struct pmc_classdep *pcd;
987 	struct pmc_mdep *pmc_mdep;
988 	enum pmc_cputype cputype;
989 	int ncpus, nclasses, i;
990 	int family, model, stepping;
991 	int error;
992 	int pmcs_per_umc;
993 
994 	/*
995 	 * The presence of hardware performance counters on the AMD
996 	 * Athlon, Duron or later processors, is _not_ indicated by
997 	 * any of the processor feature flags set by the 'CPUID'
998 	 * instruction, so we only check the 'instruction family'
999 	 * field returned by CPUID for instruction family >= 6.
1000 	 */
1001 
1002 	family = CPUID_TO_FAMILY(cpu_id);
1003 	model = CPUID_TO_MODEL(cpu_id);
1004 	stepping = CPUID_TO_STEPPING(cpu_id);
1005 
1006 	if (family == 0x18)
1007 		snprintf(pmc_cpuid, sizeof(pmc_cpuid), "HygonGenuine-%d-%02X-%X",
1008 		    family, model, stepping);
1009 	else
1010 		snprintf(pmc_cpuid, sizeof(pmc_cpuid), "AuthenticAMD-%d-%02X-%X",
1011 		    family, model, stepping);
1012 
1013 	switch (cpu_id & 0xF00) {
1014 	case 0xF00:		/* Athlon64/Opteron processor */
1015 		cputype = PMC_CPU_AMD_K8;
1016 		break;
1017 	default:
1018 		printf("pmc: Unknown AMD CPU %x %d-%d.\n", cpu_id, family,
1019 		    model);
1020 		return (NULL);
1021 	}
1022 
1023 	/*
1024 	 * From PPR for AMD Family 1Ah, a new cpuid leaf specifies the maximum
1025 	 * number of PMCs of each type.  If we do not have that leaf, we use
1026 	 * the prior default values that are only valid if we have the feature
1027 	 * bit enabled in CPU.
1028 	 */
1029 	if ((amd_feature2 & AMDID2_PCXC) != 0) {
1030 		amd_core_npmcs = AMD_PMC_CORE_DEFAULT;
1031 	} else {
1032 		amd_core_npmcs = AMD_NPMCS_K8;
1033 	}
1034 	amd_l3_npmcs = AMD_PMC_L3_DEFAULT;
1035 	amd_df_npmcs = AMD_PMC_DF_DEFAULT;
1036 	amd_umc_npmcs = 0;
1037 	pmcs_per_umc = 0;
1038 
1039 	if (cpu_exthigh >= CPUID_EXTPERFMON) {
1040 		do_cpuid(CPUID_EXTPERFMON, regs);
1041 		if (regs[1] != 0) {
1042 			amd_core_npmcs = EXTPERFMON_CORE_PMCS(regs[1]);
1043 			amd_df_npmcs = EXTPERFMON_DF_PMCS(regs[1]);
1044 			amd_umc_npmcs = EXTPERFMON_UMC_PMCS(regs[1]);
1045 			pmcs_per_umc = amd_umc_npmcs / popcntq(regs[2]);
1046 		}
1047 	}
1048 
1049 	/* Enable the newer core counters */
1050 	for (i = 0; i < amd_core_npmcs; i++) {
1051 		d = &amd_pmcdesc[i];
1052 		snprintf(d->pm_descr.pd_name, PMC_NAME_MAX,
1053 		    "K8-%d", i);
1054 		d->pm_descr.pd_class = PMC_CLASS_K8;
1055 		d->pm_descr.pd_caps = AMD_PMC_CAPS;
1056 		/*
1057 		 * Zen 5 can precisely count retire events.
1058 		 *
1059 		 * Refer to PPR Vol 1 for AMD Family 1Ah Model 02h C1 57238
1060 		 * Rev. 0.24 September 29, 2024.
1061 		 */
1062 		if ((family >= 0x1a) && (i == 2))
1063 			d->pm_descr.pd_caps |= PMC_CAP_PRECISE;
1064 		d->pm_descr.pd_width = 48;
1065 		if ((amd_feature2 & AMDID2_PCXC) != 0) {
1066 			d->pm_evsel = AMD_PMC_CORE_BASE + 2 * i;
1067 			d->pm_perfctr = AMD_PMC_CORE_BASE + 2 * i + 1;
1068 		} else {
1069 			d->pm_evsel = AMD_PMC_EVSEL_0 + i;
1070 			d->pm_perfctr = AMD_PMC_PERFCTR_0 + i;
1071 		}
1072 		d->pm_subclass = PMC_AMD_SUB_CLASS_CORE;
1073 	}
1074 	amd_npmcs = amd_core_npmcs;
1075 
1076 	if ((amd_feature2 & AMDID2_PTSCEL2I) != 0) {
1077 		/* Enable the LLC/L3 counters */
1078 		for (i = 0; i < amd_l3_npmcs; i++) {
1079 			d = &amd_pmcdesc[amd_npmcs + i];
1080 			snprintf(d->pm_descr.pd_name, PMC_NAME_MAX,
1081 			    "K8-L3-%d", i);
1082 			d->pm_descr.pd_class = PMC_CLASS_K8;
1083 			d->pm_descr.pd_caps = AMD_PMC_L3_CAPS;
1084 			d->pm_descr.pd_width = 48;
1085 			d->pm_evsel = AMD_PMC_L3_BASE + 2 * i;
1086 			d->pm_perfctr = AMD_PMC_L3_BASE + 2 * i + 1;
1087 			d->pm_subclass = PMC_AMD_SUB_CLASS_L3_CACHE;
1088 		}
1089 		amd_npmcs += amd_l3_npmcs;
1090 	}
1091 
1092 	if ((amd_feature2 & AMDID2_PNXC) != 0) {
1093 		/* Enable the data fabric counters */
1094 		for (i = 0; i < amd_df_npmcs; i++) {
1095 			d = &amd_pmcdesc[amd_npmcs + i];
1096 			snprintf(d->pm_descr.pd_name, PMC_NAME_MAX,
1097 			    "K8-DF-%d", i);
1098 			d->pm_descr.pd_class = PMC_CLASS_K8;
1099 			d->pm_descr.pd_caps = AMD_PMC_DF_CAPS;
1100 			d->pm_descr.pd_width = 48;
1101 			d->pm_evsel = AMD_PMC_DF_BASE + 2 * i;
1102 			d->pm_perfctr = AMD_PMC_DF_BASE + 2 * i + 1;
1103 			d->pm_subclass = PMC_AMD_SUB_CLASS_DATA_FABRIC;
1104 		}
1105 		amd_npmcs += amd_df_npmcs;
1106 	}
1107 
1108 	for (i = 0; i < amd_umc_npmcs; i++) {
1109 		d = &amd_pmcdesc[amd_npmcs + i];
1110 		snprintf(d->pm_descr.pd_name, PMC_NAME_MAX,
1111 		    "K8-UMC%d-%d", i / pmcs_per_umc, i);
1112 		d->pm_descr.pd_class = PMC_CLASS_K8;
1113 		d->pm_descr.pd_caps = AMD_PMC_UMC_CAPS;
1114 		d->pm_descr.pd_width = 48;
1115 		d->pm_evsel = AMD_PMC_UMC_BASE + 2 * i;
1116 		d->pm_perfctr = AMD_PMC_UMC_BASE + 2 * i + 1;
1117 		d->pm_subclass = PMC_AMD_SUB_CLASS_UMC;
1118 	}
1119 	amd_npmcs += amd_umc_npmcs;
1120 
1121 	/*
1122 	 * Sanity check that the hardware is safe to use.  Do not read or write
1123 	 * any of the PMC MSRs until after this check passes.
1124 	 */
1125 	if (amd_hwcheck() < 0) {
1126 		return (NULL);
1127 	}
1128 
1129 	/*
1130 	 * Allocate space for pointers to PMC HW descriptors and for
1131 	 * the MDEP structure used by MI code.
1132 	 */
1133 	amd_pcpu = malloc(sizeof(struct amd_cpu *) * pmc_cpu_max(), M_PMC,
1134 	    M_WAITOK | M_ZERO);
1135 
1136 	/*
1137 	 * These processors have two or three classes of PMCs: the TSC,
1138 	 * programmable PMCs, and AMD IBS.  One extra class slot is reserved
1139 	 * for the optional RAPL energy counters.
1140 	 */
1141 	if ((amd_feature2 & AMDID2_IBS) != 0) {
1142 		nclasses = 4;
1143 	} else {
1144 		nclasses = 3;
1145 	}
1146 
1147 	/*
1148 	 * Detect support for MPERF and APERF MSRs. tsc_perf_stat is set by the
1149 	 * kernel's generic TSC initialization (start_TSC(), called at boot via
1150 	 * cpu_startup() -> startrtclock()), not by hwpmc's TSC PMC class. It is
1151 	 * set only after confirming both MSRs actually increment (some emulators
1152 	 * expose the CPUID bit without real MSR support).
1153 	 */
1154 	if ((cpu_power_ecx & CPUID_PERF_STAT) && (tsc_perf_stat == 1)) {
1155 		nclasses++;
1156 	}
1157 
1158 	pmc_mdep = pmc_mdep_alloc(nclasses);
1159 
1160 	ncpus = pmc_cpu_max();
1161 
1162 	/* Initialize the TSC. */
1163 	error = pmc_tsc_initialize(pmc_mdep, ncpus);
1164 	if (error != 0)
1165 		goto error;
1166 
1167 	/* Initialize AMD K8 PMC handling. */
1168 	pcd = &pmc_mdep->pmd_classdep[PMC_MDEP_CLASS_INDEX_K8];
1169 
1170 	pcd->pcd_caps		= AMD_PMC_CAPS;
1171 	pcd->pcd_class		= PMC_CLASS_K8;
1172 	pcd->pcd_num		= amd_npmcs;
1173 	pcd->pcd_ri		= pmc_mdep->pmd_npmc;
1174 	pcd->pcd_width		= 48;
1175 
1176 	pcd->pcd_allocate_pmc	= amd_allocate_pmc;
1177 	pcd->pcd_config_pmc	= amd_config_pmc;
1178 	pcd->pcd_describe	= amd_describe;
1179 	pcd->pcd_get_config	= amd_get_config;
1180 	pcd->pcd_get_msr	= amd_get_msr;
1181 	pcd->pcd_pcpu_fini	= amd_pcpu_fini;
1182 	pcd->pcd_pcpu_init	= amd_pcpu_init;
1183 	pcd->pcd_read_pmc	= amd_read_pmc;
1184 	pcd->pcd_release_pmc	= amd_release_pmc;
1185 	pcd->pcd_start_pmc	= amd_start_pmc;
1186 	pcd->pcd_stop_pmc	= amd_stop_pmc;
1187 	pcd->pcd_write_pmc	= amd_write_pmc;
1188 	pcd->pcd_get_caps	= amd_get_caps;
1189 
1190 	pmc_mdep->pmd_cputype	= cputype;
1191 	pmc_mdep->pmd_intr	= amd_intr;
1192 	pmc_mdep->pmd_switch_in	= amd_switch_in;
1193 	pmc_mdep->pmd_switch_out = amd_switch_out;
1194 
1195 	pmc_mdep->pmd_npmc	+= amd_npmcs;
1196 
1197 	amd_init_policy();
1198 
1199 	PMCDBG0(MDP, INI, 0, "amd-initialize");
1200 
1201 	if ((amd_feature2 & AMDID2_IBS) != 0) {
1202 		error = pmc_ibs_initialize(pmc_mdep, ncpus);
1203 		if (error != 0)
1204 			goto error;
1205 	}
1206 
1207 	/* Initialize PERF class. */
1208 	pmc_perf_initialize(pmc_mdep, ncpus, nclasses - 1);
1209 
1210 	/* RAPL takes the reserved last slot; drop it if the probe fails. */
1211 	error = pmc_rapl_initialize(pmc_mdep, ncpus, pmc_mdep->pmd_nclass - 1);
1212 	if (error != 0)
1213 		pmc_mdep->pmd_nclass--;
1214 
1215 	return (pmc_mdep);
1216 
1217 error:
1218 	free(pmc_mdep, M_PMC);
1219 	return (NULL);
1220 }
1221 
1222 /*
1223  * Finalization code for AMD CPUs.
1224  */
1225 void
1226 pmc_amd_finalize(struct pmc_mdep *md)
1227 {
1228 	PMCDBG0(MDP, INI, 1, "amd-finalize");
1229 
1230 	/* Safe even if the RAPL class was skipped at initialize time. */
1231 	pmc_rapl_finalize(md);
1232 
1233 	pmc_tsc_finalize(md);
1234 
1235 	pmc_perf_finalize(md);
1236 
1237 	for (int i = 0; i < pmc_cpu_max(); i++)
1238 		KASSERT(amd_pcpu[i] == NULL,
1239 		    ("[amd,%d] non-null pcpu cpu %d", __LINE__, i));
1240 
1241 	free(amd_pcpu, M_PMC);
1242 	amd_pcpu = NULL;
1243 }
1244