xref: /freebsd/sys/dev/hwpmc/hwpmc_amd.c (revision d7d962ead0b6e5e8a39202d0590022082bf5bfb6)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 /* Support for the AMD K7 and later processors */
37 
38 #include <sys/param.h>
39 #include <sys/lock.h>
40 #include <sys/malloc.h>
41 #include <sys/mutex.h>
42 #include <sys/pcpu.h>
43 #include <sys/pmc.h>
44 #include <sys/pmckern.h>
45 #include <sys/smp.h>
46 #include <sys/systm.h>
47 
48 #include <machine/cpu.h>
49 #include <machine/cpufunc.h>
50 #include <machine/md_var.h>
51 #include <machine/specialreg.h>
52 
53 #ifdef	HWPMC_DEBUG
54 enum pmc_class	amd_pmc_class;
55 #endif
56 
57 #define	OVERFLOW_WAIT_COUNT	50
58 
59 DPCPU_DEFINE_STATIC(uint32_t, nmi_counter);
60 
61 /* AMD K7 & K8 PMCs */
62 struct amd_descr {
63 	struct pmc_descr pm_descr;  /* "base class" */
64 	uint32_t	pm_evsel;   /* address of EVSEL register */
65 	uint32_t	pm_perfctr; /* address of PERFCTR register */
66 };
67 
68 static  struct amd_descr amd_pmcdesc[AMD_NPMCS] =
69 {
70     {
71 	.pm_descr =
72 	{
73 		.pd_name  = "",
74 		.pd_class = -1,
75 		.pd_caps  = AMD_PMC_CAPS,
76 		.pd_width = 48
77 	},
78 	.pm_evsel   = AMD_PMC_EVSEL_0,
79 	.pm_perfctr = AMD_PMC_PERFCTR_0
80     },
81     {
82 	.pm_descr =
83 	{
84 		.pd_name  = "",
85 		.pd_class = -1,
86 		.pd_caps  = AMD_PMC_CAPS,
87 		.pd_width = 48
88 	},
89 	.pm_evsel   = AMD_PMC_EVSEL_1,
90 	.pm_perfctr = AMD_PMC_PERFCTR_1
91     },
92     {
93 	.pm_descr =
94 	{
95 		.pd_name  = "",
96 		.pd_class = -1,
97 		.pd_caps  = AMD_PMC_CAPS,
98 		.pd_width = 48
99 	},
100 	.pm_evsel   = AMD_PMC_EVSEL_2,
101 	.pm_perfctr = AMD_PMC_PERFCTR_2
102     },
103     {
104 	.pm_descr =
105 	{
106 		.pd_name  = "",
107 		.pd_class = -1,
108 		.pd_caps  = AMD_PMC_CAPS,
109 		.pd_width = 48
110 	},
111 	.pm_evsel   = AMD_PMC_EVSEL_3,
112 	.pm_perfctr = AMD_PMC_PERFCTR_3
113      },
114     {
115 	.pm_descr =
116 	{
117 		.pd_name  = "",
118 		.pd_class = -1,
119 		.pd_caps  = AMD_PMC_CAPS,
120 		.pd_width = 48
121 	},
122 	.pm_evsel   = AMD_PMC_EVSEL_4,
123 	.pm_perfctr = AMD_PMC_PERFCTR_4
124     },
125     {
126 	.pm_descr =
127 	{
128 		.pd_name  = "",
129 		.pd_class = -1,
130 		.pd_caps  = AMD_PMC_CAPS,
131 		.pd_width = 48
132 	},
133 	.pm_evsel   = AMD_PMC_EVSEL_5,
134 	.pm_perfctr = AMD_PMC_PERFCTR_5
135     },
136     {
137 	.pm_descr =
138 	{
139 		.pd_name  = "",
140 		.pd_class = -1,
141 		.pd_caps  = AMD_PMC_CAPS,
142 		.pd_width = 48
143 	},
144 	.pm_evsel   = AMD_PMC_EVSEL_EP_L3_0,
145 	.pm_perfctr = AMD_PMC_PERFCTR_EP_L3_0
146     },
147     {
148 	.pm_descr =
149 	{
150 		.pd_name  = "",
151 		.pd_class = -1,
152 		.pd_caps  = AMD_PMC_CAPS,
153 		.pd_width = 48
154 	},
155 	.pm_evsel   = AMD_PMC_EVSEL_EP_L3_1,
156 	.pm_perfctr = AMD_PMC_PERFCTR_EP_L3_1
157     },
158     {
159 	.pm_descr =
160 	{
161 		.pd_name  = "",
162 		.pd_class = -1,
163 		.pd_caps  = AMD_PMC_CAPS,
164 		.pd_width = 48
165 	},
166 	.pm_evsel   = AMD_PMC_EVSEL_EP_L3_2,
167 	.pm_perfctr = AMD_PMC_PERFCTR_EP_L3_2
168     },
169     {
170 	.pm_descr =
171 	{
172 		.pd_name  = "",
173 		.pd_class = -1,
174 		.pd_caps  = AMD_PMC_CAPS,
175 		.pd_width = 48
176 	},
177 	.pm_evsel   = AMD_PMC_EVSEL_EP_L3_3,
178 	.pm_perfctr = AMD_PMC_PERFCTR_EP_L3_3
179     },
180     {
181 	.pm_descr =
182 	{
183 		.pd_name  = "",
184 		.pd_class = -1,
185 		.pd_caps  = AMD_PMC_CAPS,
186 		.pd_width = 48
187 	},
188 	.pm_evsel   = AMD_PMC_EVSEL_EP_L3_4,
189 	.pm_perfctr = AMD_PMC_PERFCTR_EP_L3_4
190     },
191     {
192 	.pm_descr =
193 	{
194 		.pd_name  = "",
195 		.pd_class = -1,
196 		.pd_caps  = AMD_PMC_CAPS,
197 		.pd_width = 48
198 	},
199 	.pm_evsel   = AMD_PMC_EVSEL_EP_L3_5,
200 	.pm_perfctr = AMD_PMC_PERFCTR_EP_L3_5
201     },
202     {
203 	.pm_descr =
204 	{
205 		.pd_name  = "",
206 		.pd_class = -1,
207 		.pd_caps  = AMD_PMC_CAPS,
208 		.pd_width = 48
209 	},
210 	.pm_evsel   = AMD_PMC_EVSEL_EP_DF_0,
211 	.pm_perfctr = AMD_PMC_PERFCTR_EP_DF_0
212     },
213     {
214 	.pm_descr =
215 	{
216 		.pd_name  = "",
217 		.pd_class = -1,
218 		.pd_caps  = AMD_PMC_CAPS,
219 		.pd_width = 48
220 	},
221 	.pm_evsel   = AMD_PMC_EVSEL_EP_DF_1,
222 	.pm_perfctr = AMD_PMC_PERFCTR_EP_DF_1
223     },
224     {
225 	.pm_descr =
226 	{
227 		.pd_name  = "",
228 		.pd_class = -1,
229 		.pd_caps  = AMD_PMC_CAPS,
230 		.pd_width = 48
231 	},
232 	.pm_evsel   = AMD_PMC_EVSEL_EP_DF_2,
233 	.pm_perfctr = AMD_PMC_PERFCTR_EP_DF_2
234     },
235     {
236 	.pm_descr =
237 	{
238 		.pd_name  = "",
239 		.pd_class = -1,
240 		.pd_caps  = AMD_PMC_CAPS,
241 		.pd_width = 48
242 	},
243 	.pm_evsel   = AMD_PMC_EVSEL_EP_DF_3,
244 	.pm_perfctr = AMD_PMC_PERFCTR_EP_DF_3
245      }
246 };
247 
248 struct amd_event_code_map {
249 	enum pmc_event	pe_ev;	 /* enum value */
250 	uint16_t	pe_code; /* encoded event mask */
251 	uint8_t		pe_mask; /* bits allowed in unit mask */
252 };
253 
254 const struct amd_event_code_map amd_event_codes[] = {
255 #if	defined(__i386__)	/* 32 bit Athlon (K7) only */
256 	{ PMC_EV_K7_DC_ACCESSES, 		0x40, 0 },
257 	{ PMC_EV_K7_DC_MISSES,			0x41, 0 },
258 	{ PMC_EV_K7_DC_REFILLS_FROM_L2,		0x42, AMD_PMC_UNITMASK_MOESI },
259 	{ PMC_EV_K7_DC_REFILLS_FROM_SYSTEM,	0x43, AMD_PMC_UNITMASK_MOESI },
260 	{ PMC_EV_K7_DC_WRITEBACKS,		0x44, AMD_PMC_UNITMASK_MOESI },
261 	{ PMC_EV_K7_L1_DTLB_MISS_AND_L2_DTLB_HITS, 0x45, 0 },
262 	{ PMC_EV_K7_L1_AND_L2_DTLB_MISSES,	0x46, 0 },
263 	{ PMC_EV_K7_MISALIGNED_REFERENCES,	0x47, 0 },
264 
265 	{ PMC_EV_K7_IC_FETCHES,			0x80, 0 },
266 	{ PMC_EV_K7_IC_MISSES,			0x81, 0 },
267 
268 	{ PMC_EV_K7_L1_ITLB_MISSES,		0x84, 0 },
269 	{ PMC_EV_K7_L1_L2_ITLB_MISSES,		0x85, 0 },
270 
271 	{ PMC_EV_K7_RETIRED_INSTRUCTIONS,	0xC0, 0 },
272 	{ PMC_EV_K7_RETIRED_OPS,		0xC1, 0 },
273 	{ PMC_EV_K7_RETIRED_BRANCHES,		0xC2, 0 },
274 	{ PMC_EV_K7_RETIRED_BRANCHES_MISPREDICTED, 0xC3, 0 },
275 	{ PMC_EV_K7_RETIRED_TAKEN_BRANCHES, 	0xC4, 0 },
276 	{ PMC_EV_K7_RETIRED_TAKEN_BRANCHES_MISPREDICTED, 0xC5, 0 },
277 	{ PMC_EV_K7_RETIRED_FAR_CONTROL_TRANSFERS, 0xC6, 0 },
278 	{ PMC_EV_K7_RETIRED_RESYNC_BRANCHES,	0xC7, 0 },
279 	{ PMC_EV_K7_INTERRUPTS_MASKED_CYCLES,	0xCD, 0 },
280 	{ PMC_EV_K7_INTERRUPTS_MASKED_WHILE_PENDING_CYCLES, 0xCE, 0 },
281 	{ PMC_EV_K7_HARDWARE_INTERRUPTS,	0xCF, 0 },
282 #endif
283 
284 	{ PMC_EV_K8_FP_DISPATCHED_FPU_OPS,		0x00, 0x3F },
285 	{ PMC_EV_K8_FP_CYCLES_WITH_NO_FPU_OPS_RETIRED,	0x01, 0x00 },
286 	{ PMC_EV_K8_FP_DISPATCHED_FPU_FAST_FLAG_OPS,	0x02, 0x00 },
287 
288 	{ PMC_EV_K8_LS_SEGMENT_REGISTER_LOAD, 		0x20, 0x7F },
289 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SELF_MODIFYING_CODE,
290 	  						0x21, 0x00 },
291 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x22, 0x00 },
292 	{ PMC_EV_K8_LS_BUFFER2_FULL,			0x23, 0x00 },
293 	{ PMC_EV_K8_LS_LOCKED_OPERATION,		0x24, 0x07 },
294 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_LATE_CANCEL,	0x25, 0x00 },
295 	{ PMC_EV_K8_LS_RETIRED_CFLUSH_INSTRUCTIONS,	0x26, 0x00 },
296 	{ PMC_EV_K8_LS_RETIRED_CPUID_INSTRUCTIONS,	0x27, 0x00 },
297 
298 	{ PMC_EV_K8_DC_ACCESS,				0x40, 0x00 },
299 	{ PMC_EV_K8_DC_MISS,				0x41, 0x00 },
300 	{ PMC_EV_K8_DC_REFILL_FROM_L2,			0x42, 0x1F },
301 	{ PMC_EV_K8_DC_REFILL_FROM_SYSTEM,		0x43, 0x1F },
302 	{ PMC_EV_K8_DC_COPYBACK,			0x44, 0x1F },
303 	{ PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_HIT,	0x45, 0x00 },
304 	{ PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_MISS,	0x46, 0x00 },
305 	{ PMC_EV_K8_DC_MISALIGNED_DATA_REFERENCE,	0x47, 0x00 },
306 	{ PMC_EV_K8_DC_MICROARCHITECTURAL_LATE_CANCEL,	0x48, 0x00 },
307 	{ PMC_EV_K8_DC_MICROARCHITECTURAL_EARLY_CANCEL, 0x49, 0x00 },
308 	{ PMC_EV_K8_DC_ONE_BIT_ECC_ERROR,		0x4A, 0x03 },
309 	{ PMC_EV_K8_DC_DISPATCHED_PREFETCH_INSTRUCTIONS, 0x4B, 0x07 },
310 	{ PMC_EV_K8_DC_DCACHE_ACCESSES_BY_LOCKS,	0x4C, 0x03 },
311 
312 	{ PMC_EV_K8_BU_CPU_CLK_UNHALTED,		0x76, 0x00 },
313 	{ PMC_EV_K8_BU_INTERNAL_L2_REQUEST,		0x7D, 0x1F },
314 	{ PMC_EV_K8_BU_FILL_REQUEST_L2_MISS,		0x7E, 0x07 },
315 	{ PMC_EV_K8_BU_FILL_INTO_L2,			0x7F, 0x03 },
316 
317 	{ PMC_EV_K8_IC_FETCH,				0x80, 0x00 },
318 	{ PMC_EV_K8_IC_MISS,				0x81, 0x00 },
319 	{ PMC_EV_K8_IC_REFILL_FROM_L2,			0x82, 0x00 },
320 	{ PMC_EV_K8_IC_REFILL_FROM_SYSTEM,		0x83, 0x00 },
321 	{ PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_HIT,	0x84, 0x00 },
322 	{ PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_MISS,	0x85, 0x00 },
323 	{ PMC_EV_K8_IC_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x86, 0x00 },
324 	{ PMC_EV_K8_IC_INSTRUCTION_FETCH_STALL,		0x87, 0x00 },
325 	{ PMC_EV_K8_IC_RETURN_STACK_HIT,		0x88, 0x00 },
326 	{ PMC_EV_K8_IC_RETURN_STACK_OVERFLOW,		0x89, 0x00 },
327 
328 	{ PMC_EV_K8_FR_RETIRED_X86_INSTRUCTIONS,	0xC0, 0x00 },
329 	{ PMC_EV_K8_FR_RETIRED_UOPS,			0xC1, 0x00 },
330 	{ PMC_EV_K8_FR_RETIRED_BRANCHES,		0xC2, 0x00 },
331 	{ PMC_EV_K8_FR_RETIRED_BRANCHES_MISPREDICTED,	0xC3, 0x00 },
332 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES,		0xC4, 0x00 },
333 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED, 0xC5, 0x00 },
334 	{ PMC_EV_K8_FR_RETIRED_FAR_CONTROL_TRANSFERS,	0xC6, 0x00 },
335 	{ PMC_EV_K8_FR_RETIRED_RESYNCS,			0xC7, 0x00 },
336 	{ PMC_EV_K8_FR_RETIRED_NEAR_RETURNS,		0xC8, 0x00 },
337 	{ PMC_EV_K8_FR_RETIRED_NEAR_RETURNS_MISPREDICTED, 0xC9, 0x00 },
338 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED_BY_ADDR_MISCOMPARE,
339 							0xCA, 0x00 },
340 	{ PMC_EV_K8_FR_RETIRED_FPU_INSTRUCTIONS,	0xCB, 0x0F },
341 	{ PMC_EV_K8_FR_RETIRED_FASTPATH_DOUBLE_OP_INSTRUCTIONS,
342 							0xCC, 0x07 },
343 	{ PMC_EV_K8_FR_INTERRUPTS_MASKED_CYCLES,	0xCD, 0x00 },
344 	{ PMC_EV_K8_FR_INTERRUPTS_MASKED_WHILE_PENDING_CYCLES, 0xCE, 0x00 },
345 	{ PMC_EV_K8_FR_TAKEN_HARDWARE_INTERRUPTS,	0xCF, 0x00 },
346 
347 	{ PMC_EV_K8_FR_DECODER_EMPTY,			0xD0, 0x00 },
348 	{ PMC_EV_K8_FR_DISPATCH_STALLS,			0xD1, 0x00 },
349 	{ PMC_EV_K8_FR_DISPATCH_STALL_FROM_BRANCH_ABORT_TO_RETIRE,
350 							0xD2, 0x00 },
351 	{ PMC_EV_K8_FR_DISPATCH_STALL_FOR_SERIALIZATION, 0xD3, 0x00 },
352 	{ PMC_EV_K8_FR_DISPATCH_STALL_FOR_SEGMENT_LOAD,	0xD4, 0x00 },
353 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_REORDER_BUFFER_IS_FULL,
354 							0xD5, 0x00 },
355 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_RESERVATION_STATIONS_ARE_FULL,
356 							0xD6, 0x00 },
357 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FPU_IS_FULL,	0xD7, 0x00 },
358 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_LS_IS_FULL,	0xD8, 0x00 },
359 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_WAITING_FOR_ALL_TO_BE_QUIET,
360 							0xD9, 0x00 },
361 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FAR_XFER_OR_RESYNC_BRANCH_PENDING,
362 							0xDA, 0x00 },
363 	{ PMC_EV_K8_FR_FPU_EXCEPTIONS,			0xDB, 0x0F },
364 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR0,	0xDC, 0x00 },
365 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR1,	0xDD, 0x00 },
366 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR2,	0xDE, 0x00 },
367 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR3,	0xDF, 0x00 },
368 
369 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_ACCESS_EVENT, 0xE0, 0x7 },
370 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_TABLE_OVERFLOW, 0xE1, 0x00 },
371 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_DRAM_COMMAND_SLOTS_MISSED,
372 							0xE2, 0x00 },
373 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_TURNAROUND,	0xE3, 0x07 },
374 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_BYPASS_SATURATION, 0xE4, 0x0F },
375 	{ PMC_EV_K8_NB_SIZED_COMMANDS,			0xEB, 0x7F },
376 	{ PMC_EV_K8_NB_PROBE_RESULT,			0xEC, 0x0F },
377 	{ PMC_EV_K8_NB_HT_BUS0_BANDWIDTH,		0xF6, 0x0F },
378 	{ PMC_EV_K8_NB_HT_BUS1_BANDWIDTH,		0xF7, 0x0F },
379 	{ PMC_EV_K8_NB_HT_BUS2_BANDWIDTH,		0xF8, 0x0F }
380 
381 };
382 
383 const int amd_event_codes_size = nitems(amd_event_codes);
384 
385 /*
386  * Per-processor information
387  */
388 
389 struct amd_cpu {
390 	struct pmc_hw	pc_amdpmcs[AMD_NPMCS];
391 };
392 
393 static struct amd_cpu **amd_pcpu;
394 
395 /*
396  * read a pmc register
397  */
398 
399 static int
400 amd_read_pmc(int cpu, int ri, pmc_value_t *v)
401 {
402 	enum pmc_mode mode;
403 	const struct amd_descr *pd;
404 	struct pmc *pm;
405 	pmc_value_t tmp;
406 
407 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
408 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
409 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
410 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
411 	KASSERT(amd_pcpu[cpu],
412 	    ("[amd,%d] null per-cpu, cpu %d", __LINE__, cpu));
413 
414 	pm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc;
415 	pd = &amd_pmcdesc[ri];
416 
417 	KASSERT(pm != NULL,
418 	    ("[amd,%d] No owner for HWPMC [cpu%d,pmc%d]", __LINE__,
419 		cpu, ri));
420 
421 	mode = PMC_TO_MODE(pm);
422 
423 	PMCDBG2(MDP,REA,1,"amd-read id=%d class=%d", ri, pd->pm_descr.pd_class);
424 
425 #ifdef	HWPMC_DEBUG
426 	KASSERT(pd->pm_descr.pd_class == amd_pmc_class,
427 	    ("[amd,%d] unknown PMC class (%d)", __LINE__,
428 		pd->pm_descr.pd_class));
429 #endif
430 
431 	tmp = rdmsr(pd->pm_perfctr); /* RDMSR serializes */
432 	PMCDBG2(MDP,REA,2,"amd-read (pre-munge) id=%d -> %jd", ri, tmp);
433 	if (PMC_IS_SAMPLING_MODE(mode)) {
434 		/* Sign extend 48 bit value to 64 bits. */
435 		tmp = (pmc_value_t) (((int64_t) tmp << 16) >> 16);
436 		tmp = AMD_PERFCTR_VALUE_TO_RELOAD_COUNT(tmp);
437 	}
438 	*v = tmp;
439 
440 	PMCDBG2(MDP,REA,2,"amd-read (post-munge) id=%d -> %jd", ri, *v);
441 
442 	return 0;
443 }
444 
445 /*
446  * Write a PMC MSR.
447  */
448 
449 static int
450 amd_write_pmc(int cpu, int ri, pmc_value_t v)
451 {
452 	const struct amd_descr *pd;
453 	enum pmc_mode mode;
454 	struct pmc *pm;
455 
456 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
457 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
458 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
459 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
460 
461 	pm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc;
462 	pd = &amd_pmcdesc[ri];
463 
464 	KASSERT(pm != NULL,
465 	    ("[amd,%d] PMC not owned (cpu%d,pmc%d)", __LINE__,
466 		cpu, ri));
467 
468 	mode = PMC_TO_MODE(pm);
469 
470 #ifdef	HWPMC_DEBUG
471 	KASSERT(pd->pm_descr.pd_class == amd_pmc_class,
472 	    ("[amd,%d] unknown PMC class (%d)", __LINE__,
473 		pd->pm_descr.pd_class));
474 #endif
475 
476 	/* use 2's complement of the count for sampling mode PMCs */
477 	if (PMC_IS_SAMPLING_MODE(mode))
478 		v = AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v);
479 
480 	PMCDBG3(MDP,WRI,1,"amd-write cpu=%d ri=%d v=%jx", cpu, ri, v);
481 
482 	/* write the PMC value */
483 	wrmsr(pd->pm_perfctr, v);
484 	return 0;
485 }
486 
487 /*
488  * configure hardware pmc according to the configuration recorded in
489  * pmc 'pm'.
490  */
491 
492 static int
493 amd_config_pmc(int cpu, int ri, struct pmc *pm)
494 {
495 	struct pmc_hw *phw;
496 
497 	PMCDBG3(MDP,CFG,1, "cpu=%d ri=%d pm=%p", cpu, ri, pm);
498 
499 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
500 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
501 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
502 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
503 
504 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
505 
506 	KASSERT(pm == NULL || phw->phw_pmc == NULL,
507 	    ("[amd,%d] pm=%p phw->pm=%p hwpmc not unconfigured",
508 		__LINE__, pm, phw->phw_pmc));
509 
510 	phw->phw_pmc = pm;
511 	return 0;
512 }
513 
514 /*
515  * Retrieve a configured PMC pointer from hardware state.
516  */
517 
518 static int
519 amd_get_config(int cpu, int ri, struct pmc **ppm)
520 {
521 	*ppm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc;
522 
523 	return 0;
524 }
525 
526 /*
527  * Machine dependent actions taken during the context switch in of a
528  * thread.
529  */
530 
531 static int
532 amd_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
533 {
534 	(void) pc;
535 
536 	PMCDBG3(MDP,SWI,1, "pc=%p pp=%p enable-msr=%d", pc, pp,
537 	    (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0);
538 
539 	/* enable the RDPMC instruction if needed */
540 	if (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS)
541 		load_cr4(rcr4() | CR4_PCE);
542 
543 	return 0;
544 }
545 
546 /*
547  * Machine dependent actions taken during the context switch out of a
548  * thread.
549  */
550 
551 static int
552 amd_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
553 {
554 	(void) pc;
555 	(void) pp;		/* can be NULL */
556 
557 	PMCDBG3(MDP,SWO,1, "pc=%p pp=%p enable-msr=%d", pc, pp, pp ?
558 	    (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) == 1 : 0);
559 
560 	/* always turn off the RDPMC instruction */
561 	load_cr4(rcr4() & ~CR4_PCE);
562 
563 	return 0;
564 }
565 
566 /*
567  * Check if a given allocation is feasible.
568  */
569 
570 static int
571 amd_allocate_pmc(int cpu, int ri, struct pmc *pm,
572     const struct pmc_op_pmcallocate *a)
573 {
574 	int i;
575 	uint64_t allowed_unitmask, caps, config, unitmask;
576 	enum pmc_event pe;
577 	const struct pmc_descr *pd;
578 
579 	(void) cpu;
580 
581 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
582 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
583 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
584 	    ("[amd,%d] illegal row index %d", __LINE__, ri));
585 
586 	pd = &amd_pmcdesc[ri].pm_descr;
587 
588 	/* check class match */
589 	if (pd->pd_class != a->pm_class)
590 		return EINVAL;
591 
592 	caps = pm->pm_caps;
593 
594 	PMCDBG2(MDP,ALL,1,"amd-allocate ri=%d caps=0x%x", ri, caps);
595 
596 	if((ri >= 0 && ri < 6) && !(a->pm_md.pm_amd.pm_amd_sub_class == PMC_AMD_SUB_CLASS_CORE))
597 		return EINVAL;
598 	if((ri >= 6 && ri < 12) && !(a->pm_md.pm_amd.pm_amd_sub_class == PMC_AMD_SUB_CLASS_L3_CACHE))
599 		return EINVAL;
600 	if((ri >= 12 && ri < 16) && !(a->pm_md.pm_amd.pm_amd_sub_class == PMC_AMD_SUB_CLASS_DATA_FABRIC))
601 		return EINVAL;
602 
603 	if (strlen(pmc_cpuid) != 0) {
604 		pm->pm_md.pm_amd.pm_amd_evsel =
605 			a->pm_md.pm_amd.pm_amd_config;
606 		PMCDBG2(MDP,ALL,2,"amd-allocate ri=%d -> config=0x%x", ri, a->pm_md.pm_amd.pm_amd_config);
607 		return (0);
608 	}
609 
610 	pe = a->pm_ev;
611 
612 	/* map ev to the correct event mask code */
613 	config = allowed_unitmask = 0;
614 	for (i = 0; i < amd_event_codes_size; i++)
615 		if (amd_event_codes[i].pe_ev == pe) {
616 			config =
617 			    AMD_PMC_TO_EVENTMASK(amd_event_codes[i].pe_code);
618 			allowed_unitmask =
619 			    AMD_PMC_TO_UNITMASK(amd_event_codes[i].pe_mask);
620 			break;
621 		}
622 	if (i == amd_event_codes_size)
623 		return EINVAL;
624 
625 	unitmask = a->pm_md.pm_amd.pm_amd_config & AMD_PMC_UNITMASK;
626 	if (unitmask & ~allowed_unitmask) /* disallow reserved bits */
627 		return EINVAL;
628 
629 	if (unitmask && (caps & PMC_CAP_QUALIFIER))
630 		config |= unitmask;
631 
632 	if (caps & PMC_CAP_THRESHOLD)
633 		config |= a->pm_md.pm_amd.pm_amd_config & AMD_PMC_COUNTERMASK;
634 
635 	/* set at least one of the 'usr' or 'os' caps */
636 	if (caps & PMC_CAP_USER)
637 		config |= AMD_PMC_USR;
638 	if (caps & PMC_CAP_SYSTEM)
639 		config |= AMD_PMC_OS;
640 	if ((caps & (PMC_CAP_USER|PMC_CAP_SYSTEM)) == 0)
641 		config |= (AMD_PMC_USR|AMD_PMC_OS);
642 
643 	if (caps & PMC_CAP_EDGE)
644 		config |= AMD_PMC_EDGE;
645 	if (caps & PMC_CAP_INVERT)
646 		config |= AMD_PMC_INVERT;
647 	if (caps & PMC_CAP_INTERRUPT)
648 		config |= AMD_PMC_INT;
649 
650 	pm->pm_md.pm_amd.pm_amd_evsel = config; /* save config value */
651 
652 	PMCDBG2(MDP,ALL,2,"amd-allocate ri=%d -> config=0x%x", ri, config);
653 
654 	return 0;
655 }
656 
657 /*
658  * Release machine dependent state associated with a PMC.  This is a
659  * no-op on this architecture.
660  *
661  */
662 
663 /* ARGSUSED0 */
664 static int
665 amd_release_pmc(int cpu, int ri, struct pmc *pmc)
666 {
667 #ifdef	HWPMC_DEBUG
668 	const struct amd_descr *pd;
669 #endif
670 	struct pmc_hw *phw __diagused;
671 
672 	(void) pmc;
673 
674 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
675 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
676 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
677 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
678 
679 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
680 
681 	KASSERT(phw->phw_pmc == NULL,
682 	    ("[amd,%d] PHW pmc %p non-NULL", __LINE__, phw->phw_pmc));
683 
684 #ifdef	HWPMC_DEBUG
685 	pd = &amd_pmcdesc[ri];
686 	if (pd->pm_descr.pd_class == amd_pmc_class)
687 		KASSERT(AMD_PMC_IS_STOPPED(pd->pm_evsel),
688 		    ("[amd,%d] PMC %d released while active", __LINE__, ri));
689 #endif
690 
691 	return 0;
692 }
693 
694 /*
695  * start a PMC.
696  */
697 
698 static int
699 amd_start_pmc(int cpu, int ri)
700 {
701 	uint64_t config;
702 	struct pmc *pm;
703 	struct pmc_hw *phw;
704 	const struct amd_descr *pd;
705 
706 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
707 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
708 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
709 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
710 
711 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
712 	pm  = phw->phw_pmc;
713 	pd = &amd_pmcdesc[ri];
714 
715 	KASSERT(pm != NULL,
716 	    ("[amd,%d] starting cpu%d,pmc%d with null pmc record", __LINE__,
717 		cpu, ri));
718 
719 	PMCDBG2(MDP,STA,1,"amd-start cpu=%d ri=%d", cpu, ri);
720 
721 	KASSERT(AMD_PMC_IS_STOPPED(pd->pm_evsel),
722 	    ("[amd,%d] pmc%d,cpu%d: Starting active PMC \"%s\"", __LINE__,
723 	    ri, cpu, pd->pm_descr.pd_name));
724 
725 	/* turn on the PMC ENABLE bit */
726 	config = pm->pm_md.pm_amd.pm_amd_evsel | AMD_PMC_ENABLE;
727 
728 	PMCDBG1(MDP,STA,2,"amd-start config=0x%x", config);
729 
730 	wrmsr(pd->pm_evsel, config);
731 	return 0;
732 }
733 
734 /*
735  * Stop a PMC.
736  */
737 
738 static int
739 amd_stop_pmc(int cpu, int ri)
740 {
741 	struct pmc *pm;
742 	struct pmc_hw *phw;
743 	const struct amd_descr *pd;
744 	uint64_t config;
745 	int i;
746 
747 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
748 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
749 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
750 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
751 
752 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
753 	pm  = phw->phw_pmc;
754 	pd  = &amd_pmcdesc[ri];
755 
756 	KASSERT(pm != NULL,
757 	    ("[amd,%d] cpu%d,pmc%d no PMC to stop", __LINE__,
758 		cpu, ri));
759 	KASSERT(!AMD_PMC_IS_STOPPED(pd->pm_evsel),
760 	    ("[amd,%d] PMC%d, CPU%d \"%s\" already stopped",
761 		__LINE__, ri, cpu, pd->pm_descr.pd_name));
762 
763 	PMCDBG1(MDP,STO,1,"amd-stop ri=%d", ri);
764 
765 	/* turn off the PMC ENABLE bit */
766 	config = pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE;
767 	wrmsr(pd->pm_evsel, config);
768 
769 	/*
770 	 * Due to NMI latency on newer AMD processors
771 	 * NMI interrupts are ignored, which leads to
772 	 * panic or messages based on kernel configuration
773 	 */
774 
775 	/* Wait for the count to be reset */
776 	for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
777 		if (rdmsr(pd->pm_perfctr) & (1 << (pd->pm_descr.pd_width - 1)))
778 			break;
779 
780 		DELAY(1);
781 	}
782 
783 	return 0;
784 }
785 
786 /*
787  * Interrupt handler.  This function needs to return '1' if the
788  * interrupt was this CPU's PMCs or '0' otherwise.  It is not allowed
789  * to sleep or do anything a 'fast' interrupt handler is not allowed
790  * to do.
791  */
792 
793 static int
794 amd_intr(struct trapframe *tf)
795 {
796 	int i, error, retval, cpu;
797 	uint64_t config, evsel, perfctr;
798 	struct pmc *pm;
799 	struct amd_cpu *pac;
800 	pmc_value_t v;
801 	uint32_t active = 0, count = 0;
802 
803 	cpu = curcpu;
804 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
805 	    ("[amd,%d] out of range CPU %d", __LINE__, cpu));
806 
807 	PMCDBG3(MDP,INT,1, "cpu=%d tf=%p um=%d", cpu, (void *) tf,
808 	    TRAPF_USERMODE(tf));
809 
810 	retval = 0;
811 
812 	pac = amd_pcpu[cpu];
813 
814 	/*
815 	 * look for all PMCs that have interrupted:
816 	 * - look for a running, sampling PMC which has overflowed
817 	 *   and which has a valid 'struct pmc' association
818 	 *
819 	 * If found, we call a helper to process the interrupt.
820 	 *
821 	 * PMCs interrupting at the same time are collapsed into
822 	 * a single interrupt. Check all the valid pmcs for
823 	 * overflow.
824 	 */
825 
826 	for (i = 0; i < AMD_CORE_NPMCS; i++) {
827 
828 		if ((pm = pac->pc_amdpmcs[i].phw_pmc) == NULL ||
829 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) {
830 			continue;
831 		}
832 
833 		/* Consider pmc with valid handle as active */
834 		active++;
835 
836 		if (!AMD_PMC_HAS_OVERFLOWED(i))
837 			continue;
838 
839 		retval = 1;	/* Found an interrupting PMC. */
840 
841 		if (pm->pm_state != PMC_STATE_RUNNING)
842 			continue;
843 
844 		/* Stop the PMC, reload count. */
845 		evsel	= amd_pmcdesc[i].pm_evsel;
846 		perfctr	= amd_pmcdesc[i].pm_perfctr;
847 		v       = pm->pm_sc.pm_reloadcount;
848 		config  = rdmsr(evsel);
849 
850 		KASSERT((config & ~AMD_PMC_ENABLE) ==
851 		    (pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE),
852 		    ("[amd,%d] config mismatch reg=0x%jx pm=0x%jx", __LINE__,
853 			 (uintmax_t)config, (uintmax_t)pm->pm_md.pm_amd.pm_amd_evsel));
854 
855 		wrmsr(evsel, config & ~AMD_PMC_ENABLE);
856 		wrmsr(perfctr, AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v));
857 
858 		/* Restart the counter if logging succeeded. */
859 		error = pmc_process_interrupt(PMC_HR, pm, tf);
860 		if (error == 0)
861 			wrmsr(evsel, config);
862 	}
863 
864 	/*
865 	 * Due to NMI latency, there can be a scenario in which
866 	 * multiple pmcs gets serviced in an earlier NMI and we
867 	 * do not find an overflow in the subsequent NMI.
868 	 *
869 	 * For such cases we keep a per-cpu count of active NMIs
870 	 * and compare it with min(active pmcs, 2) to determine
871 	 * if this NMI was for a pmc overflow which was serviced
872 	 * in an earlier request or should be ignored.
873 	 */
874 
875 	if (retval) {
876 		DPCPU_SET(nmi_counter, min(2, active));
877 	} else {
878 		if ((count = DPCPU_GET(nmi_counter))) {
879 			retval = 1;
880 			DPCPU_SET(nmi_counter, --count);
881 		}
882 	}
883 
884 	if (retval)
885 		counter_u64_add(pmc_stats.pm_intr_processed, 1);
886 	else
887 		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
888 
889 	PMCDBG1(MDP,INT,2, "retval=%d", retval);
890 	return (retval);
891 }
892 
893 /*
894  * describe a PMC
895  */
896 static int
897 amd_describe(int cpu, int ri, struct pmc_info *pi, struct pmc **ppmc)
898 {
899 	int error;
900 	size_t copied;
901 	const struct amd_descr *pd;
902 	struct pmc_hw *phw;
903 
904 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
905 	    ("[amd,%d] illegal CPU %d", __LINE__, cpu));
906 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
907 	    ("[amd,%d] row-index %d out of range", __LINE__, ri));
908 
909 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
910 	pd  = &amd_pmcdesc[ri];
911 
912 	if ((error = copystr(pd->pm_descr.pd_name, pi->pm_name,
913 		 PMC_NAME_MAX, &copied)) != 0)
914 		return error;
915 
916 	pi->pm_class = pd->pm_descr.pd_class;
917 
918 	if (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) {
919 		pi->pm_enabled = TRUE;
920 		*ppmc          = phw->phw_pmc;
921 	} else {
922 		pi->pm_enabled = FALSE;
923 		*ppmc          = NULL;
924 	}
925 
926 	return 0;
927 }
928 
929 /*
930  * i386 specific entry points
931  */
932 
933 /*
934  * return the MSR address of the given PMC.
935  */
936 
937 static int
938 amd_get_msr(int ri, uint32_t *msr)
939 {
940 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
941 	    ("[amd,%d] ri %d out of range", __LINE__, ri));
942 
943 	*msr = amd_pmcdesc[ri].pm_perfctr - AMD_PMC_PERFCTR_0;
944 
945 	return (0);
946 }
947 
948 /*
949  * processor dependent initialization.
950  */
951 
952 static int
953 amd_pcpu_init(struct pmc_mdep *md, int cpu)
954 {
955 	int classindex, first_ri, n;
956 	struct pmc_cpu *pc;
957 	struct amd_cpu *pac;
958 	struct pmc_hw  *phw;
959 
960 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
961 	    ("[amd,%d] insane cpu number %d", __LINE__, cpu));
962 
963 	PMCDBG1(MDP,INI,1,"amd-init cpu=%d", cpu);
964 
965 	amd_pcpu[cpu] = pac = malloc(sizeof(struct amd_cpu), M_PMC,
966 	    M_WAITOK|M_ZERO);
967 
968 	/*
969 	 * Set the content of the hardware descriptors to a known
970 	 * state and initialize pointers in the MI per-cpu descriptor.
971 	 */
972 	pc = pmc_pcpu[cpu];
973 #if	defined(__amd64__)
974 	classindex = PMC_MDEP_CLASS_INDEX_K8;
975 #elif	defined(__i386__)
976 	classindex = md->pmd_cputype == PMC_CPU_AMD_K8 ?
977 	    PMC_MDEP_CLASS_INDEX_K8 : PMC_MDEP_CLASS_INDEX_K7;
978 #endif
979 	first_ri = md->pmd_classdep[classindex].pcd_ri;
980 
981 	KASSERT(pc != NULL, ("[amd,%d] NULL per-cpu pointer", __LINE__));
982 
983 	for (n = 0, phw = pac->pc_amdpmcs; n < AMD_NPMCS; n++, phw++) {
984 		phw->phw_state 	  = PMC_PHW_FLAG_IS_ENABLED |
985 		    PMC_PHW_CPU_TO_STATE(cpu) | PMC_PHW_INDEX_TO_STATE(n);
986 		phw->phw_pmc	  = NULL;
987 		pc->pc_hwpmcs[n + first_ri]  = phw;
988 	}
989 
990 	return (0);
991 }
992 
993 
994 /*
995  * processor dependent cleanup prior to the KLD
996  * being unloaded
997  */
998 
999 static int
1000 amd_pcpu_fini(struct pmc_mdep *md, int cpu)
1001 {
1002 	int classindex, first_ri, i;
1003 	uint32_t evsel;
1004 	struct pmc_cpu *pc;
1005 	struct amd_cpu *pac;
1006 
1007 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1008 	    ("[amd,%d] insane cpu number (%d)", __LINE__, cpu));
1009 
1010 	PMCDBG1(MDP,INI,1,"amd-cleanup cpu=%d", cpu);
1011 
1012 	/*
1013 	 * First, turn off all PMCs on this CPU.
1014 	 */
1015 	for (i = 0; i < 4; i++) { /* XXX this loop is now not needed */
1016 		evsel = rdmsr(AMD_PMC_EVSEL_0 + i);
1017 		evsel &= ~AMD_PMC_ENABLE;
1018 		wrmsr(AMD_PMC_EVSEL_0 + i, evsel);
1019 	}
1020 
1021 	/*
1022 	 * Next, free up allocated space.
1023 	 */
1024 	if ((pac = amd_pcpu[cpu]) == NULL)
1025 		return (0);
1026 
1027 	amd_pcpu[cpu] = NULL;
1028 
1029 #ifdef	HWPMC_DEBUG
1030 	for (i = 0; i < AMD_NPMCS; i++) {
1031 		KASSERT(pac->pc_amdpmcs[i].phw_pmc == NULL,
1032 		    ("[amd,%d] CPU%d/PMC%d in use", __LINE__, cpu, i));
1033 		KASSERT(AMD_PMC_IS_STOPPED(AMD_PMC_EVSEL_0 + i),
1034 		    ("[amd,%d] CPU%d/PMC%d not stopped", __LINE__, cpu, i));
1035 	}
1036 #endif
1037 
1038 	pc = pmc_pcpu[cpu];
1039 	KASSERT(pc != NULL, ("[amd,%d] NULL per-cpu state", __LINE__));
1040 
1041 #if	defined(__amd64__)
1042 	classindex = PMC_MDEP_CLASS_INDEX_K8;
1043 #elif	defined(__i386__)
1044 	classindex = md->pmd_cputype == PMC_CPU_AMD_K8 ? PMC_MDEP_CLASS_INDEX_K8 :
1045 	    PMC_MDEP_CLASS_INDEX_K7;
1046 #endif
1047 	first_ri = md->pmd_classdep[classindex].pcd_ri;
1048 
1049 	/*
1050 	 * Reset pointers in the MI 'per-cpu' state.
1051 	 */
1052 	for (i = 0; i < AMD_NPMCS; i++) {
1053 		pc->pc_hwpmcs[i + first_ri] = NULL;
1054 	}
1055 
1056 
1057 	free(pac, M_PMC);
1058 
1059 	return (0);
1060 }
1061 
1062 /*
1063  * Initialize ourselves.
1064  */
1065 
1066 struct pmc_mdep *
1067 pmc_amd_initialize(void)
1068 {
1069 	int classindex, error, i, ncpus;
1070 	struct pmc_classdep *pcd;
1071 	enum pmc_cputype cputype;
1072 	struct pmc_mdep *pmc_mdep;
1073 	enum pmc_class class;
1074 	int family, model, stepping;
1075 	char *name;
1076 
1077 	/*
1078 	 * The presence of hardware performance counters on the AMD
1079 	 * Athlon, Duron or later processors, is _not_ indicated by
1080 	 * any of the processor feature flags set by the 'CPUID'
1081 	 * instruction, so we only check the 'instruction family'
1082 	 * field returned by CPUID for instruction family >= 6.
1083 	 */
1084 
1085 	name = NULL;
1086 	family = CPUID_TO_FAMILY(cpu_id);
1087 	model = CPUID_TO_MODEL(cpu_id);
1088 	stepping = CPUID_TO_STEPPING(cpu_id);
1089 
1090 	if (family == 0x18)
1091 		snprintf(pmc_cpuid, sizeof(pmc_cpuid), "HygonGenuine-%d-%02X-%X",
1092 		    family, model, stepping);
1093 	else
1094 		snprintf(pmc_cpuid, sizeof(pmc_cpuid), "AuthenticAMD-%d-%02X-%X",
1095 		    family, model, stepping);
1096 
1097 	switch (cpu_id & 0xF00) {
1098 #if	defined(__i386__)
1099 	case 0x600:		/* Athlon(tm) processor */
1100 		classindex = PMC_MDEP_CLASS_INDEX_K7;
1101 		cputype = PMC_CPU_AMD_K7;
1102 		class = PMC_CLASS_K7;
1103 		name = "K7";
1104 		break;
1105 #endif
1106 	case 0xF00:		/* Athlon64/Opteron processor */
1107 		classindex = PMC_MDEP_CLASS_INDEX_K8;
1108 		cputype = PMC_CPU_AMD_K8;
1109 		class = PMC_CLASS_K8;
1110 		name = "K8";
1111 		break;
1112 
1113 	default:
1114 		(void) printf("pmc: Unknown AMD CPU %x %d-%d.\n", cpu_id, (cpu_id & 0xF00) >> 8, model);
1115 		return NULL;
1116 	}
1117 
1118 #ifdef	HWPMC_DEBUG
1119 	amd_pmc_class = class;
1120 #endif
1121 
1122 	/*
1123 	 * Allocate space for pointers to PMC HW descriptors and for
1124 	 * the MDEP structure used by MI code.
1125 	 */
1126 	amd_pcpu = malloc(sizeof(struct amd_cpu *) * pmc_cpu_max(), M_PMC,
1127 	    M_WAITOK|M_ZERO);
1128 
1129 	/*
1130 	 * These processors have two classes of PMCs: the TSC and
1131 	 * programmable PMCs.
1132 	 */
1133 	pmc_mdep = pmc_mdep_alloc(2);
1134 
1135 	pmc_mdep->pmd_cputype = cputype;
1136 
1137 	ncpus = pmc_cpu_max();
1138 
1139 	/* Initialize the TSC. */
1140 	error = pmc_tsc_initialize(pmc_mdep, ncpus);
1141 	if (error)
1142 		goto error;
1143 
1144 	/* Initialize AMD K7 and K8 PMC handling. */
1145 	pcd = &pmc_mdep->pmd_classdep[classindex];
1146 
1147 	pcd->pcd_caps		= AMD_PMC_CAPS;
1148 	pcd->pcd_class		= class;
1149 	pcd->pcd_num		= AMD_NPMCS;
1150 	pcd->pcd_ri		= pmc_mdep->pmd_npmc;
1151 	pcd->pcd_width		= 48;
1152 
1153 	/* fill in the correct pmc name and class */
1154 	for (i = 0; i < AMD_NPMCS; i++) {
1155 		(void) snprintf(amd_pmcdesc[i].pm_descr.pd_name,
1156 		    sizeof(amd_pmcdesc[i].pm_descr.pd_name), "%s-%d",
1157 		    name, i);
1158 		amd_pmcdesc[i].pm_descr.pd_class = class;
1159 	}
1160 
1161 	pcd->pcd_allocate_pmc	= amd_allocate_pmc;
1162 	pcd->pcd_config_pmc	= amd_config_pmc;
1163 	pcd->pcd_describe	= amd_describe;
1164 	pcd->pcd_get_config	= amd_get_config;
1165 	pcd->pcd_get_msr	= amd_get_msr;
1166 	pcd->pcd_pcpu_fini	= amd_pcpu_fini;
1167 	pcd->pcd_pcpu_init	= amd_pcpu_init;
1168 	pcd->pcd_read_pmc	= amd_read_pmc;
1169 	pcd->pcd_release_pmc	= amd_release_pmc;
1170 	pcd->pcd_start_pmc	= amd_start_pmc;
1171 	pcd->pcd_stop_pmc	= amd_stop_pmc;
1172 	pcd->pcd_write_pmc	= amd_write_pmc;
1173 
1174 	pmc_mdep->pmd_pcpu_init = NULL;
1175 	pmc_mdep->pmd_pcpu_fini = NULL;
1176 	pmc_mdep->pmd_intr	= amd_intr;
1177 	pmc_mdep->pmd_switch_in = amd_switch_in;
1178 	pmc_mdep->pmd_switch_out = amd_switch_out;
1179 
1180 	pmc_mdep->pmd_npmc     += AMD_NPMCS;
1181 
1182 	PMCDBG0(MDP,INI,0,"amd-initialize");
1183 
1184 	return (pmc_mdep);
1185 
1186   error:
1187 	if (error) {
1188 		free(pmc_mdep, M_PMC);
1189 		pmc_mdep = NULL;
1190 	}
1191 
1192 	return (NULL);
1193 }
1194 
1195 /*
1196  * Finalization code for AMD CPUs.
1197  */
1198 
1199 void
1200 pmc_amd_finalize(struct pmc_mdep *md)
1201 {
1202 #if	defined(INVARIANTS)
1203 	int classindex, i, ncpus, pmcclass;
1204 #endif
1205 
1206 	pmc_tsc_finalize(md);
1207 
1208 	KASSERT(amd_pcpu != NULL, ("[amd,%d] NULL per-cpu array pointer",
1209 	    __LINE__));
1210 
1211 #if	defined(INVARIANTS)
1212 	switch (md->pmd_cputype) {
1213 #if	defined(__i386__)
1214 	case PMC_CPU_AMD_K7:
1215 		classindex = PMC_MDEP_CLASS_INDEX_K7;
1216 		pmcclass = PMC_CLASS_K7;
1217 		break;
1218 #endif
1219 	default:
1220 		classindex = PMC_MDEP_CLASS_INDEX_K8;
1221 		pmcclass = PMC_CLASS_K8;
1222 	}
1223 
1224 	KASSERT(md->pmd_classdep[classindex].pcd_class == pmcclass,
1225 	    ("[amd,%d] pmc class mismatch", __LINE__));
1226 
1227 	ncpus = pmc_cpu_max();
1228 
1229 	for (i = 0; i < ncpus; i++)
1230 		KASSERT(amd_pcpu[i] == NULL, ("[amd,%d] non-null pcpu",
1231 		    __LINE__));
1232 #endif
1233 
1234 	free(amd_pcpu, M_PMC);
1235 	amd_pcpu = NULL;
1236 }
1237