xref: /freebsd/sys/dev/hwpmc/hwpmc_amd.c (revision d2ba5111c125104b09aa1acd1bfe8af2a24c79cc)
1 /*-
2  * Copyright (c) 2003-2008 Joseph Koshy
3  * Copyright (c) 2007 The FreeBSD Foundation
4  * All rights reserved.
5  *
6  * Portions of this software were developed by A. Joseph Koshy under
7  * sponsorship from the FreeBSD Foundation and Google, Inc.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 /* Support for the AMD K7 and later processors */
35 
36 #include <sys/param.h>
37 #include <sys/lock.h>
38 #include <sys/malloc.h>
39 #include <sys/mutex.h>
40 #include <sys/pmc.h>
41 #include <sys/pmckern.h>
42 #include <sys/smp.h>
43 #include <sys/systm.h>
44 
45 #include <machine/cpu.h>
46 #include <machine/cpufunc.h>
47 #include <machine/md_var.h>
48 #include <machine/specialreg.h>
49 
50 #ifdef	HWPMC_DEBUG
51 enum pmc_class	amd_pmc_class;
52 #endif
53 
54 /* AMD K7 & K8 PMCs */
55 struct amd_descr {
56 	struct pmc_descr pm_descr;  /* "base class" */
57 	uint32_t	pm_evsel;   /* address of EVSEL register */
58 	uint32_t	pm_perfctr; /* address of PERFCTR register */
59 };
60 
61 static  struct amd_descr amd_pmcdesc[AMD_NPMCS] =
62 {
63     {
64 	.pm_descr =
65 	{
66 		.pd_name  = "",
67 		.pd_class = -1,
68 		.pd_caps  = AMD_PMC_CAPS,
69 		.pd_width = 48
70 	},
71 	.pm_evsel   = AMD_PMC_EVSEL_0,
72 	.pm_perfctr = AMD_PMC_PERFCTR_0
73     },
74     {
75 	.pm_descr =
76 	{
77 		.pd_name  = "",
78 		.pd_class = -1,
79 		.pd_caps  = AMD_PMC_CAPS,
80 		.pd_width = 48
81 	},
82 	.pm_evsel   = AMD_PMC_EVSEL_1,
83 	.pm_perfctr = AMD_PMC_PERFCTR_1
84     },
85     {
86 	.pm_descr =
87 	{
88 		.pd_name  = "",
89 		.pd_class = -1,
90 		.pd_caps  = AMD_PMC_CAPS,
91 		.pd_width = 48
92 	},
93 	.pm_evsel   = AMD_PMC_EVSEL_2,
94 	.pm_perfctr = AMD_PMC_PERFCTR_2
95     },
96     {
97 	.pm_descr =
98 	{
99 		.pd_name  = "",
100 		.pd_class = -1,
101 		.pd_caps  = AMD_PMC_CAPS,
102 		.pd_width = 48
103 	},
104 	.pm_evsel   = AMD_PMC_EVSEL_3,
105 	.pm_perfctr = AMD_PMC_PERFCTR_3
106     }
107 };
108 
109 struct amd_event_code_map {
110 	enum pmc_event	pe_ev;	 /* enum value */
111 	uint8_t		pe_code; /* encoded event mask */
112 	uint8_t		pe_mask; /* bits allowed in unit mask */
113 };
114 
115 const struct amd_event_code_map amd_event_codes[] = {
116 #if	defined(__i386__)	/* 32 bit Athlon (K7) only */
117 	{ PMC_EV_K7_DC_ACCESSES, 		0x40, 0 },
118 	{ PMC_EV_K7_DC_MISSES,			0x41, 0 },
119 	{ PMC_EV_K7_DC_REFILLS_FROM_L2,		0x42, AMD_PMC_UNITMASK_MOESI },
120 	{ PMC_EV_K7_DC_REFILLS_FROM_SYSTEM,	0x43, AMD_PMC_UNITMASK_MOESI },
121 	{ PMC_EV_K7_DC_WRITEBACKS,		0x44, AMD_PMC_UNITMASK_MOESI },
122 	{ PMC_EV_K7_L1_DTLB_MISS_AND_L2_DTLB_HITS, 0x45, 0 },
123 	{ PMC_EV_K7_L1_AND_L2_DTLB_MISSES,	0x46, 0 },
124 	{ PMC_EV_K7_MISALIGNED_REFERENCES,	0x47, 0 },
125 
126 	{ PMC_EV_K7_IC_FETCHES,			0x80, 0 },
127 	{ PMC_EV_K7_IC_MISSES,			0x81, 0 },
128 
129 	{ PMC_EV_K7_L1_ITLB_MISSES,		0x84, 0 },
130 	{ PMC_EV_K7_L1_L2_ITLB_MISSES,		0x85, 0 },
131 
132 	{ PMC_EV_K7_RETIRED_INSTRUCTIONS,	0xC0, 0 },
133 	{ PMC_EV_K7_RETIRED_OPS,		0xC1, 0 },
134 	{ PMC_EV_K7_RETIRED_BRANCHES,		0xC2, 0 },
135 	{ PMC_EV_K7_RETIRED_BRANCHES_MISPREDICTED, 0xC3, 0 },
136 	{ PMC_EV_K7_RETIRED_TAKEN_BRANCHES, 	0xC4, 0 },
137 	{ PMC_EV_K7_RETIRED_TAKEN_BRANCHES_MISPREDICTED, 0xC5, 0 },
138 	{ PMC_EV_K7_RETIRED_FAR_CONTROL_TRANSFERS, 0xC6, 0 },
139 	{ PMC_EV_K7_RETIRED_RESYNC_BRANCHES,	0xC7, 0 },
140 	{ PMC_EV_K7_INTERRUPTS_MASKED_CYCLES,	0xCD, 0 },
141 	{ PMC_EV_K7_INTERRUPTS_MASKED_WHILE_PENDING_CYCLES, 0xCE, 0 },
142 	{ PMC_EV_K7_HARDWARE_INTERRUPTS,	0xCF, 0 },
143 #endif
144 
145 	{ PMC_EV_K8_FP_DISPATCHED_FPU_OPS,		0x00, 0x3F },
146 	{ PMC_EV_K8_FP_CYCLES_WITH_NO_FPU_OPS_RETIRED,	0x01, 0x00 },
147 	{ PMC_EV_K8_FP_DISPATCHED_FPU_FAST_FLAG_OPS,	0x02, 0x00 },
148 
149 	{ PMC_EV_K8_LS_SEGMENT_REGISTER_LOAD, 		0x20, 0x7F },
150 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SELF_MODIFYING_CODE,
151 	  						0x21, 0x00 },
152 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x22, 0x00 },
153 	{ PMC_EV_K8_LS_BUFFER2_FULL,			0x23, 0x00 },
154 	{ PMC_EV_K8_LS_LOCKED_OPERATION,		0x24, 0x07 },
155 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_LATE_CANCEL,	0x25, 0x00 },
156 	{ PMC_EV_K8_LS_RETIRED_CFLUSH_INSTRUCTIONS,	0x26, 0x00 },
157 	{ PMC_EV_K8_LS_RETIRED_CPUID_INSTRUCTIONS,	0x27, 0x00 },
158 
159 	{ PMC_EV_K8_DC_ACCESS,				0x40, 0x00 },
160 	{ PMC_EV_K8_DC_MISS,				0x41, 0x00 },
161 	{ PMC_EV_K8_DC_REFILL_FROM_L2,			0x42, 0x1F },
162 	{ PMC_EV_K8_DC_REFILL_FROM_SYSTEM,		0x43, 0x1F },
163 	{ PMC_EV_K8_DC_COPYBACK,			0x44, 0x1F },
164 	{ PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_HIT,	0x45, 0x00 },
165 	{ PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_MISS,	0x46, 0x00 },
166 	{ PMC_EV_K8_DC_MISALIGNED_DATA_REFERENCE,	0x47, 0x00 },
167 	{ PMC_EV_K8_DC_MICROARCHITECTURAL_LATE_CANCEL,	0x48, 0x00 },
168 	{ PMC_EV_K8_DC_MICROARCHITECTURAL_EARLY_CANCEL, 0x49, 0x00 },
169 	{ PMC_EV_K8_DC_ONE_BIT_ECC_ERROR,		0x4A, 0x03 },
170 	{ PMC_EV_K8_DC_DISPATCHED_PREFETCH_INSTRUCTIONS, 0x4B, 0x07 },
171 	{ PMC_EV_K8_DC_DCACHE_ACCESSES_BY_LOCKS,	0x4C, 0x03 },
172 
173 	{ PMC_EV_K8_BU_CPU_CLK_UNHALTED,		0x76, 0x00 },
174 	{ PMC_EV_K8_BU_INTERNAL_L2_REQUEST,		0x7D, 0x1F },
175 	{ PMC_EV_K8_BU_FILL_REQUEST_L2_MISS,		0x7E, 0x07 },
176 	{ PMC_EV_K8_BU_FILL_INTO_L2,			0x7F, 0x03 },
177 
178 	{ PMC_EV_K8_IC_FETCH,				0x80, 0x00 },
179 	{ PMC_EV_K8_IC_MISS,				0x81, 0x00 },
180 	{ PMC_EV_K8_IC_REFILL_FROM_L2,			0x82, 0x00 },
181 	{ PMC_EV_K8_IC_REFILL_FROM_SYSTEM,		0x83, 0x00 },
182 	{ PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_HIT,	0x84, 0x00 },
183 	{ PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_MISS,	0x85, 0x00 },
184 	{ PMC_EV_K8_IC_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x86, 0x00 },
185 	{ PMC_EV_K8_IC_INSTRUCTION_FETCH_STALL,		0x87, 0x00 },
186 	{ PMC_EV_K8_IC_RETURN_STACK_HIT,		0x88, 0x00 },
187 	{ PMC_EV_K8_IC_RETURN_STACK_OVERFLOW,		0x89, 0x00 },
188 
189 	{ PMC_EV_K8_FR_RETIRED_X86_INSTRUCTIONS,	0xC0, 0x00 },
190 	{ PMC_EV_K8_FR_RETIRED_UOPS,			0xC1, 0x00 },
191 	{ PMC_EV_K8_FR_RETIRED_BRANCHES,		0xC2, 0x00 },
192 	{ PMC_EV_K8_FR_RETIRED_BRANCHES_MISPREDICTED,	0xC3, 0x00 },
193 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES,		0xC4, 0x00 },
194 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED, 0xC5, 0x00 },
195 	{ PMC_EV_K8_FR_RETIRED_FAR_CONTROL_TRANSFERS,	0xC6, 0x00 },
196 	{ PMC_EV_K8_FR_RETIRED_RESYNCS,			0xC7, 0x00 },
197 	{ PMC_EV_K8_FR_RETIRED_NEAR_RETURNS,		0xC8, 0x00 },
198 	{ PMC_EV_K8_FR_RETIRED_NEAR_RETURNS_MISPREDICTED, 0xC9, 0x00 },
199 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED_BY_ADDR_MISCOMPARE,
200 							0xCA, 0x00 },
201 	{ PMC_EV_K8_FR_RETIRED_FPU_INSTRUCTIONS,	0xCB, 0x0F },
202 	{ PMC_EV_K8_FR_RETIRED_FASTPATH_DOUBLE_OP_INSTRUCTIONS,
203 							0xCC, 0x07 },
204 	{ PMC_EV_K8_FR_INTERRUPTS_MASKED_CYCLES,	0xCD, 0x00 },
205 	{ PMC_EV_K8_FR_INTERRUPTS_MASKED_WHILE_PENDING_CYCLES, 0xCE, 0x00 },
206 	{ PMC_EV_K8_FR_TAKEN_HARDWARE_INTERRUPTS,	0xCF, 0x00 },
207 
208 	{ PMC_EV_K8_FR_DECODER_EMPTY,			0xD0, 0x00 },
209 	{ PMC_EV_K8_FR_DISPATCH_STALLS,			0xD1, 0x00 },
210 	{ PMC_EV_K8_FR_DISPATCH_STALL_FROM_BRANCH_ABORT_TO_RETIRE,
211 							0xD2, 0x00 },
212 	{ PMC_EV_K8_FR_DISPATCH_STALL_FOR_SERIALIZATION, 0xD3, 0x00 },
213 	{ PMC_EV_K8_FR_DISPATCH_STALL_FOR_SEGMENT_LOAD,	0xD4, 0x00 },
214 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_REORDER_BUFFER_IS_FULL,
215 							0xD5, 0x00 },
216 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_RESERVATION_STATIONS_ARE_FULL,
217 							0xD6, 0x00 },
218 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FPU_IS_FULL,	0xD7, 0x00 },
219 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_LS_IS_FULL,	0xD8, 0x00 },
220 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_WAITING_FOR_ALL_TO_BE_QUIET,
221 							0xD9, 0x00 },
222 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FAR_XFER_OR_RESYNC_BRANCH_PENDING,
223 							0xDA, 0x00 },
224 	{ PMC_EV_K8_FR_FPU_EXCEPTIONS,			0xDB, 0x0F },
225 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR0,	0xDC, 0x00 },
226 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR1,	0xDD, 0x00 },
227 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR2,	0xDE, 0x00 },
228 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR3,	0xDF, 0x00 },
229 
230 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_ACCESS_EVENT, 0xE0, 0x7 },
231 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_TABLE_OVERFLOW, 0xE1, 0x00 },
232 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_DRAM_COMMAND_SLOTS_MISSED,
233 							0xE2, 0x00 },
234 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_TURNAROUND,	0xE3, 0x07 },
235 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_BYPASS_SATURATION, 0xE4, 0x0F },
236 	{ PMC_EV_K8_NB_SIZED_COMMANDS,			0xEB, 0x7F },
237 	{ PMC_EV_K8_NB_PROBE_RESULT,			0xEC, 0x0F },
238 	{ PMC_EV_K8_NB_HT_BUS0_BANDWIDTH,		0xF6, 0x0F },
239 	{ PMC_EV_K8_NB_HT_BUS1_BANDWIDTH,		0xF7, 0x0F },
240 	{ PMC_EV_K8_NB_HT_BUS2_BANDWIDTH,		0xF8, 0x0F }
241 
242 };
243 
244 const int amd_event_codes_size = nitems(amd_event_codes);
245 
246 /*
247  * Per-processor information
248  */
249 
250 struct amd_cpu {
251 	struct pmc_hw	pc_amdpmcs[AMD_NPMCS];
252 };
253 
254 static struct amd_cpu **amd_pcpu;
255 
256 /*
257  * read a pmc register
258  */
259 
260 static int
261 amd_read_pmc(int cpu, int ri, pmc_value_t *v)
262 {
263 	enum pmc_mode mode;
264 	const struct amd_descr *pd;
265 	struct pmc *pm;
266 	pmc_value_t tmp;
267 
268 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
269 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
270 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
271 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
272 	KASSERT(amd_pcpu[cpu],
273 	    ("[amd,%d] null per-cpu, cpu %d", __LINE__, cpu));
274 
275 	pm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc;
276 	pd = &amd_pmcdesc[ri];
277 
278 	KASSERT(pm != NULL,
279 	    ("[amd,%d] No owner for HWPMC [cpu%d,pmc%d]", __LINE__,
280 		cpu, ri));
281 
282 	mode = PMC_TO_MODE(pm);
283 
284 	PMCDBG2(MDP,REA,1,"amd-read id=%d class=%d", ri, pd->pm_descr.pd_class);
285 
286 #ifdef	HWPMC_DEBUG
287 	KASSERT(pd->pm_descr.pd_class == amd_pmc_class,
288 	    ("[amd,%d] unknown PMC class (%d)", __LINE__,
289 		pd->pm_descr.pd_class));
290 #endif
291 
292 	tmp = rdmsr(pd->pm_perfctr); /* RDMSR serializes */
293 	PMCDBG2(MDP,REA,2,"amd-read (pre-munge) id=%d -> %jd", ri, tmp);
294 	if (PMC_IS_SAMPLING_MODE(mode)) {
295 		/* Sign extend 48 bit value to 64 bits. */
296 		tmp = (pmc_value_t) (((int64_t) tmp << 16) >> 16);
297 		tmp = AMD_PERFCTR_VALUE_TO_RELOAD_COUNT(tmp);
298 	}
299 	*v = tmp;
300 
301 	PMCDBG2(MDP,REA,2,"amd-read (post-munge) id=%d -> %jd", ri, *v);
302 
303 	return 0;
304 }
305 
306 /*
307  * Write a PMC MSR.
308  */
309 
310 static int
311 amd_write_pmc(int cpu, int ri, pmc_value_t v)
312 {
313 	const struct amd_descr *pd;
314 	enum pmc_mode mode;
315 	struct pmc *pm;
316 
317 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
318 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
319 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
320 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
321 
322 	pm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc;
323 	pd = &amd_pmcdesc[ri];
324 
325 	KASSERT(pm != NULL,
326 	    ("[amd,%d] PMC not owned (cpu%d,pmc%d)", __LINE__,
327 		cpu, ri));
328 
329 	mode = PMC_TO_MODE(pm);
330 
331 #ifdef	HWPMC_DEBUG
332 	KASSERT(pd->pm_descr.pd_class == amd_pmc_class,
333 	    ("[amd,%d] unknown PMC class (%d)", __LINE__,
334 		pd->pm_descr.pd_class));
335 #endif
336 
337 	/* use 2's complement of the count for sampling mode PMCs */
338 	if (PMC_IS_SAMPLING_MODE(mode))
339 		v = AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v);
340 
341 	PMCDBG3(MDP,WRI,1,"amd-write cpu=%d ri=%d v=%jx", cpu, ri, v);
342 
343 	/* write the PMC value */
344 	wrmsr(pd->pm_perfctr, v);
345 	return 0;
346 }
347 
348 /*
349  * configure hardware pmc according to the configuration recorded in
350  * pmc 'pm'.
351  */
352 
353 static int
354 amd_config_pmc(int cpu, int ri, struct pmc *pm)
355 {
356 	struct pmc_hw *phw;
357 
358 	PMCDBG3(MDP,CFG,1, "cpu=%d ri=%d pm=%p", cpu, ri, pm);
359 
360 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
361 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
362 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
363 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
364 
365 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
366 
367 	KASSERT(pm == NULL || phw->phw_pmc == NULL,
368 	    ("[amd,%d] pm=%p phw->pm=%p hwpmc not unconfigured",
369 		__LINE__, pm, phw->phw_pmc));
370 
371 	phw->phw_pmc = pm;
372 	return 0;
373 }
374 
375 /*
376  * Retrieve a configured PMC pointer from hardware state.
377  */
378 
379 static int
380 amd_get_config(int cpu, int ri, struct pmc **ppm)
381 {
382 	*ppm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc;
383 
384 	return 0;
385 }
386 
387 /*
388  * Machine dependent actions taken during the context switch in of a
389  * thread.
390  */
391 
392 static int
393 amd_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
394 {
395 	(void) pc;
396 
397 	PMCDBG3(MDP,SWI,1, "pc=%p pp=%p enable-msr=%d", pc, pp,
398 	    (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0);
399 
400 	/* enable the RDPMC instruction if needed */
401 	if (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS)
402 		load_cr4(rcr4() | CR4_PCE);
403 
404 	return 0;
405 }
406 
407 /*
408  * Machine dependent actions taken during the context switch out of a
409  * thread.
410  */
411 
412 static int
413 amd_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
414 {
415 	(void) pc;
416 	(void) pp;		/* can be NULL */
417 
418 	PMCDBG3(MDP,SWO,1, "pc=%p pp=%p enable-msr=%d", pc, pp, pp ?
419 	    (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) == 1 : 0);
420 
421 	/* always turn off the RDPMC instruction */
422 	load_cr4(rcr4() & ~CR4_PCE);
423 
424 	return 0;
425 }
426 
427 /*
428  * Check if a given allocation is feasible.
429  */
430 
431 static int
432 amd_allocate_pmc(int cpu, int ri, struct pmc *pm,
433     const struct pmc_op_pmcallocate *a)
434 {
435 	int i;
436 	uint32_t allowed_unitmask, caps, config, unitmask;
437 	enum pmc_event pe;
438 	const struct pmc_descr *pd;
439 
440 	(void) cpu;
441 
442 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
443 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
444 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
445 	    ("[amd,%d] illegal row index %d", __LINE__, ri));
446 
447 	pd = &amd_pmcdesc[ri].pm_descr;
448 
449 	/* check class match */
450 	if (pd->pd_class != a->pm_class)
451 		return EINVAL;
452 
453 	caps = pm->pm_caps;
454 
455 	PMCDBG2(MDP,ALL,1,"amd-allocate ri=%d caps=0x%x", ri, caps);
456 
457 	if ((pd->pd_caps & caps) != caps)
458 		return EPERM;
459 
460 	pe = a->pm_ev;
461 
462 	/* map ev to the correct event mask code */
463 	config = allowed_unitmask = 0;
464 	for (i = 0; i < amd_event_codes_size; i++)
465 		if (amd_event_codes[i].pe_ev == pe) {
466 			config =
467 			    AMD_PMC_TO_EVENTMASK(amd_event_codes[i].pe_code);
468 			allowed_unitmask =
469 			    AMD_PMC_TO_UNITMASK(amd_event_codes[i].pe_mask);
470 			break;
471 		}
472 	if (i == amd_event_codes_size)
473 		return EINVAL;
474 
475 	unitmask = a->pm_md.pm_amd.pm_amd_config & AMD_PMC_UNITMASK;
476 	if (unitmask & ~allowed_unitmask) /* disallow reserved bits */
477 		return EINVAL;
478 
479 	if (unitmask && (caps & PMC_CAP_QUALIFIER))
480 		config |= unitmask;
481 
482 	if (caps & PMC_CAP_THRESHOLD)
483 		config |= a->pm_md.pm_amd.pm_amd_config & AMD_PMC_COUNTERMASK;
484 
485 	/* set at least one of the 'usr' or 'os' caps */
486 	if (caps & PMC_CAP_USER)
487 		config |= AMD_PMC_USR;
488 	if (caps & PMC_CAP_SYSTEM)
489 		config |= AMD_PMC_OS;
490 	if ((caps & (PMC_CAP_USER|PMC_CAP_SYSTEM)) == 0)
491 		config |= (AMD_PMC_USR|AMD_PMC_OS);
492 
493 	if (caps & PMC_CAP_EDGE)
494 		config |= AMD_PMC_EDGE;
495 	if (caps & PMC_CAP_INVERT)
496 		config |= AMD_PMC_INVERT;
497 	if (caps & PMC_CAP_INTERRUPT)
498 		config |= AMD_PMC_INT;
499 
500 	pm->pm_md.pm_amd.pm_amd_evsel = config; /* save config value */
501 
502 	PMCDBG2(MDP,ALL,2,"amd-allocate ri=%d -> config=0x%x", ri, config);
503 
504 	return 0;
505 }
506 
507 /*
508  * Release machine dependent state associated with a PMC.  This is a
509  * no-op on this architecture.
510  *
511  */
512 
513 /* ARGSUSED0 */
514 static int
515 amd_release_pmc(int cpu, int ri, struct pmc *pmc)
516 {
517 #ifdef	HWPMC_DEBUG
518 	const struct amd_descr *pd;
519 #endif
520 	struct pmc_hw *phw;
521 
522 	(void) pmc;
523 
524 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
525 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
526 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
527 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
528 
529 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
530 
531 	KASSERT(phw->phw_pmc == NULL,
532 	    ("[amd,%d] PHW pmc %p non-NULL", __LINE__, phw->phw_pmc));
533 
534 #ifdef	HWPMC_DEBUG
535 	pd = &amd_pmcdesc[ri];
536 	if (pd->pm_descr.pd_class == amd_pmc_class)
537 		KASSERT(AMD_PMC_IS_STOPPED(pd->pm_evsel),
538 		    ("[amd,%d] PMC %d released while active", __LINE__, ri));
539 #endif
540 
541 	return 0;
542 }
543 
544 /*
545  * start a PMC.
546  */
547 
548 static int
549 amd_start_pmc(int cpu, int ri)
550 {
551 	uint32_t config;
552 	struct pmc *pm;
553 	struct pmc_hw *phw;
554 	const struct amd_descr *pd;
555 
556 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
557 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
558 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
559 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
560 
561 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
562 	pm  = phw->phw_pmc;
563 	pd = &amd_pmcdesc[ri];
564 
565 	KASSERT(pm != NULL,
566 	    ("[amd,%d] starting cpu%d,pmc%d with null pmc record", __LINE__,
567 		cpu, ri));
568 
569 	PMCDBG2(MDP,STA,1,"amd-start cpu=%d ri=%d", cpu, ri);
570 
571 	KASSERT(AMD_PMC_IS_STOPPED(pd->pm_evsel),
572 	    ("[amd,%d] pmc%d,cpu%d: Starting active PMC \"%s\"", __LINE__,
573 	    ri, cpu, pd->pm_descr.pd_name));
574 
575 	/* turn on the PMC ENABLE bit */
576 	config = pm->pm_md.pm_amd.pm_amd_evsel | AMD_PMC_ENABLE;
577 
578 	PMCDBG1(MDP,STA,2,"amd-start config=0x%x", config);
579 
580 	wrmsr(pd->pm_evsel, config);
581 	return 0;
582 }
583 
584 /*
585  * Stop a PMC.
586  */
587 
588 static int
589 amd_stop_pmc(int cpu, int ri)
590 {
591 	struct pmc *pm;
592 	struct pmc_hw *phw;
593 	const struct amd_descr *pd;
594 	uint64_t config;
595 
596 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
597 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
598 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
599 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
600 
601 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
602 	pm  = phw->phw_pmc;
603 	pd  = &amd_pmcdesc[ri];
604 
605 	KASSERT(pm != NULL,
606 	    ("[amd,%d] cpu%d,pmc%d no PMC to stop", __LINE__,
607 		cpu, ri));
608 	KASSERT(!AMD_PMC_IS_STOPPED(pd->pm_evsel),
609 	    ("[amd,%d] PMC%d, CPU%d \"%s\" already stopped",
610 		__LINE__, ri, cpu, pd->pm_descr.pd_name));
611 
612 	PMCDBG1(MDP,STO,1,"amd-stop ri=%d", ri);
613 
614 	/* turn off the PMC ENABLE bit */
615 	config = pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE;
616 	wrmsr(pd->pm_evsel, config);
617 	return 0;
618 }
619 
620 /*
621  * Interrupt handler.  This function needs to return '1' if the
622  * interrupt was this CPU's PMCs or '0' otherwise.  It is not allowed
623  * to sleep or do anything a 'fast' interrupt handler is not allowed
624  * to do.
625  */
626 
627 static int
628 amd_intr(int cpu, struct trapframe *tf)
629 {
630 	int i, error, retval;
631 	uint32_t config, evsel, perfctr;
632 	struct pmc *pm;
633 	struct amd_cpu *pac;
634 	pmc_value_t v;
635 
636 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
637 	    ("[amd,%d] out of range CPU %d", __LINE__, cpu));
638 
639 	PMCDBG3(MDP,INT,1, "cpu=%d tf=%p um=%d", cpu, (void *) tf,
640 	    TRAPF_USERMODE(tf));
641 
642 	retval = 0;
643 
644 	pac = amd_pcpu[cpu];
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 	 * If multiple PMCs interrupt at the same time, the AMD64
654 	 * processor appears to deliver as many NMIs as there are
655 	 * outstanding PMC interrupts.  So we process only one NMI
656 	 * interrupt at a time.
657 	 */
658 
659 	for (i = 0; retval == 0 && i < AMD_NPMCS; i++) {
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 		if (!AMD_PMC_HAS_OVERFLOWED(i))
667 			continue;
668 
669 		retval = 1;	/* Found an interrupting PMC. */
670 
671 		if (pm->pm_state != PMC_STATE_RUNNING)
672 			continue;
673 
674 		/* Stop the PMC, reload count. */
675 		evsel   = AMD_PMC_EVSEL_0 + i;
676 		perfctr = AMD_PMC_PERFCTR_0 + i;
677 		v       = pm->pm_sc.pm_reloadcount;
678 		config  = rdmsr(evsel);
679 
680 		KASSERT((config & ~AMD_PMC_ENABLE) ==
681 		    (pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE),
682 		    ("[amd,%d] config mismatch reg=0x%x pm=0x%x", __LINE__,
683 			config, pm->pm_md.pm_amd.pm_amd_evsel));
684 
685 		wrmsr(evsel, config & ~AMD_PMC_ENABLE);
686 		wrmsr(perfctr, AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v));
687 
688 		/* Restart the counter if logging succeeded. */
689 		error = pmc_process_interrupt(cpu, PMC_HR, pm, tf,
690 		    TRAPF_USERMODE(tf));
691 		if (error == 0)
692 			wrmsr(evsel, config);
693 	}
694 
695 	atomic_add_int(retval ? &pmc_stats.pm_intr_processed :
696 	    &pmc_stats.pm_intr_ignored, 1);
697 
698 	PMCDBG1(MDP,INT,2, "retval=%d", retval);
699 	return (retval);
700 }
701 
702 /*
703  * describe a PMC
704  */
705 static int
706 amd_describe(int cpu, int ri, struct pmc_info *pi, struct pmc **ppmc)
707 {
708 	int error;
709 	size_t copied;
710 	const struct amd_descr *pd;
711 	struct pmc_hw *phw;
712 
713 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
714 	    ("[amd,%d] illegal CPU %d", __LINE__, cpu));
715 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
716 	    ("[amd,%d] row-index %d out of range", __LINE__, ri));
717 
718 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
719 	pd  = &amd_pmcdesc[ri];
720 
721 	if ((error = copystr(pd->pm_descr.pd_name, pi->pm_name,
722 		 PMC_NAME_MAX, &copied)) != 0)
723 		return error;
724 
725 	pi->pm_class = pd->pm_descr.pd_class;
726 
727 	if (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) {
728 		pi->pm_enabled = TRUE;
729 		*ppmc          = phw->phw_pmc;
730 	} else {
731 		pi->pm_enabled = FALSE;
732 		*ppmc          = NULL;
733 	}
734 
735 	return 0;
736 }
737 
738 /*
739  * i386 specific entry points
740  */
741 
742 /*
743  * return the MSR address of the given PMC.
744  */
745 
746 static int
747 amd_get_msr(int ri, uint32_t *msr)
748 {
749 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
750 	    ("[amd,%d] ri %d out of range", __LINE__, ri));
751 
752 	*msr = amd_pmcdesc[ri].pm_perfctr - AMD_PMC_PERFCTR_0;
753 
754 	return (0);
755 }
756 
757 /*
758  * processor dependent initialization.
759  */
760 
761 static int
762 amd_pcpu_init(struct pmc_mdep *md, int cpu)
763 {
764 	int classindex, first_ri, n;
765 	struct pmc_cpu *pc;
766 	struct amd_cpu *pac;
767 	struct pmc_hw  *phw;
768 
769 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
770 	    ("[amd,%d] insane cpu number %d", __LINE__, cpu));
771 
772 	PMCDBG1(MDP,INI,1,"amd-init cpu=%d", cpu);
773 
774 	amd_pcpu[cpu] = pac = malloc(sizeof(struct amd_cpu), M_PMC,
775 	    M_WAITOK|M_ZERO);
776 
777 	/*
778 	 * Set the content of the hardware descriptors to a known
779 	 * state and initialize pointers in the MI per-cpu descriptor.
780 	 */
781 	pc = pmc_pcpu[cpu];
782 #if	defined(__amd64__)
783 	classindex = PMC_MDEP_CLASS_INDEX_K8;
784 #elif	defined(__i386__)
785 	classindex = md->pmd_cputype == PMC_CPU_AMD_K8 ?
786 	    PMC_MDEP_CLASS_INDEX_K8 : PMC_MDEP_CLASS_INDEX_K7;
787 #endif
788 	first_ri = md->pmd_classdep[classindex].pcd_ri;
789 
790 	KASSERT(pc != NULL, ("[amd,%d] NULL per-cpu pointer", __LINE__));
791 
792 	for (n = 0, phw = pac->pc_amdpmcs; n < AMD_NPMCS; n++, phw++) {
793 		phw->phw_state 	  = PMC_PHW_FLAG_IS_ENABLED |
794 		    PMC_PHW_CPU_TO_STATE(cpu) | PMC_PHW_INDEX_TO_STATE(n);
795 		phw->phw_pmc	  = NULL;
796 		pc->pc_hwpmcs[n + first_ri]  = phw;
797 	}
798 
799 	return (0);
800 }
801 
802 
803 /*
804  * processor dependent cleanup prior to the KLD
805  * being unloaded
806  */
807 
808 static int
809 amd_pcpu_fini(struct pmc_mdep *md, int cpu)
810 {
811 	int classindex, first_ri, i;
812 	uint32_t evsel;
813 	struct pmc_cpu *pc;
814 	struct amd_cpu *pac;
815 
816 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
817 	    ("[amd,%d] insane cpu number (%d)", __LINE__, cpu));
818 
819 	PMCDBG1(MDP,INI,1,"amd-cleanup cpu=%d", cpu);
820 
821 	/*
822 	 * First, turn off all PMCs on this CPU.
823 	 */
824 	for (i = 0; i < 4; i++) { /* XXX this loop is now not needed */
825 		evsel = rdmsr(AMD_PMC_EVSEL_0 + i);
826 		evsel &= ~AMD_PMC_ENABLE;
827 		wrmsr(AMD_PMC_EVSEL_0 + i, evsel);
828 	}
829 
830 	/*
831 	 * Next, free up allocated space.
832 	 */
833 	if ((pac = amd_pcpu[cpu]) == NULL)
834 		return (0);
835 
836 	amd_pcpu[cpu] = NULL;
837 
838 #ifdef	HWPMC_DEBUG
839 	for (i = 0; i < AMD_NPMCS; i++) {
840 		KASSERT(pac->pc_amdpmcs[i].phw_pmc == NULL,
841 		    ("[amd,%d] CPU%d/PMC%d in use", __LINE__, cpu, i));
842 		KASSERT(AMD_PMC_IS_STOPPED(AMD_PMC_EVSEL_0 + i),
843 		    ("[amd,%d] CPU%d/PMC%d not stopped", __LINE__, cpu, i));
844 	}
845 #endif
846 
847 	pc = pmc_pcpu[cpu];
848 	KASSERT(pc != NULL, ("[amd,%d] NULL per-cpu state", __LINE__));
849 
850 #if	defined(__amd64__)
851 	classindex = PMC_MDEP_CLASS_INDEX_K8;
852 #elif	defined(__i386__)
853 	classindex = md->pmd_cputype == PMC_CPU_AMD_K8 ? PMC_MDEP_CLASS_INDEX_K8 :
854 	    PMC_MDEP_CLASS_INDEX_K7;
855 #endif
856 	first_ri = md->pmd_classdep[classindex].pcd_ri;
857 
858 	/*
859 	 * Reset pointers in the MI 'per-cpu' state.
860 	 */
861 	for (i = 0; i < AMD_NPMCS; i++) {
862 		pc->pc_hwpmcs[i + first_ri] = NULL;
863 	}
864 
865 
866 	free(pac, M_PMC);
867 
868 	return (0);
869 }
870 
871 /*
872  * Initialize ourselves.
873  */
874 
875 struct pmc_mdep *
876 pmc_amd_initialize(void)
877 {
878 	int classindex, error, i, ncpus;
879 	struct pmc_classdep *pcd;
880 	enum pmc_cputype cputype;
881 	struct pmc_mdep *pmc_mdep;
882 	enum pmc_class class;
883 	char *name;
884 
885 	/*
886 	 * The presence of hardware performance counters on the AMD
887 	 * Athlon, Duron or later processors, is _not_ indicated by
888 	 * any of the processor feature flags set by the 'CPUID'
889 	 * instruction, so we only check the 'instruction family'
890 	 * field returned by CPUID for instruction family >= 6.
891 	 */
892 
893 	name = NULL;
894 	switch (cpu_id & 0xF00) {
895 #if	defined(__i386__)
896 	case 0x600:		/* Athlon(tm) processor */
897 		classindex = PMC_MDEP_CLASS_INDEX_K7;
898 		cputype = PMC_CPU_AMD_K7;
899 		class = PMC_CLASS_K7;
900 		name = "K7";
901 		break;
902 #endif
903 	case 0xF00:		/* Athlon64/Opteron processor */
904 		classindex = PMC_MDEP_CLASS_INDEX_K8;
905 		cputype = PMC_CPU_AMD_K8;
906 		class = PMC_CLASS_K8;
907 		name = "K8";
908 		break;
909 
910 	default:
911 		(void) printf("pmc: Unknown AMD CPU.\n");
912 		return NULL;
913 	}
914 
915 #ifdef	HWPMC_DEBUG
916 	amd_pmc_class = class;
917 #endif
918 
919 	/*
920 	 * Allocate space for pointers to PMC HW descriptors and for
921 	 * the MDEP structure used by MI code.
922 	 */
923 	amd_pcpu = malloc(sizeof(struct amd_cpu *) * pmc_cpu_max(), M_PMC,
924 	    M_WAITOK|M_ZERO);
925 
926 	/*
927 	 * These processors have two classes of PMCs: the TSC and
928 	 * programmable PMCs.
929 	 */
930 	pmc_mdep = pmc_mdep_alloc(2);
931 
932 	pmc_mdep->pmd_cputype = cputype;
933 
934 	ncpus = pmc_cpu_max();
935 
936 	/* Initialize the TSC. */
937 	error = pmc_tsc_initialize(pmc_mdep, ncpus);
938 	if (error)
939 		goto error;
940 
941 	/* Initialize AMD K7 and K8 PMC handling. */
942 	pcd = &pmc_mdep->pmd_classdep[classindex];
943 
944 	pcd->pcd_caps		= AMD_PMC_CAPS;
945 	pcd->pcd_class		= class;
946 	pcd->pcd_num		= AMD_NPMCS;
947 	pcd->pcd_ri		= pmc_mdep->pmd_npmc;
948 	pcd->pcd_width		= 48;
949 
950 	/* fill in the correct pmc name and class */
951 	for (i = 0; i < AMD_NPMCS; i++) {
952 		(void) snprintf(amd_pmcdesc[i].pm_descr.pd_name,
953 		    sizeof(amd_pmcdesc[i].pm_descr.pd_name), "%s-%d",
954 		    name, i);
955 		amd_pmcdesc[i].pm_descr.pd_class = class;
956 	}
957 
958 	pcd->pcd_allocate_pmc	= amd_allocate_pmc;
959 	pcd->pcd_config_pmc	= amd_config_pmc;
960 	pcd->pcd_describe	= amd_describe;
961 	pcd->pcd_get_config	= amd_get_config;
962 	pcd->pcd_get_msr	= amd_get_msr;
963 	pcd->pcd_pcpu_fini	= amd_pcpu_fini;
964 	pcd->pcd_pcpu_init	= amd_pcpu_init;
965 	pcd->pcd_read_pmc	= amd_read_pmc;
966 	pcd->pcd_release_pmc	= amd_release_pmc;
967 	pcd->pcd_start_pmc	= amd_start_pmc;
968 	pcd->pcd_stop_pmc	= amd_stop_pmc;
969 	pcd->pcd_write_pmc	= amd_write_pmc;
970 
971 	pmc_mdep->pmd_pcpu_init = NULL;
972 	pmc_mdep->pmd_pcpu_fini = NULL;
973 	pmc_mdep->pmd_intr	= amd_intr;
974 	pmc_mdep->pmd_switch_in = amd_switch_in;
975 	pmc_mdep->pmd_switch_out = amd_switch_out;
976 
977 	pmc_mdep->pmd_npmc     += AMD_NPMCS;
978 
979 	PMCDBG0(MDP,INI,0,"amd-initialize");
980 
981 	return (pmc_mdep);
982 
983   error:
984 	if (error) {
985 		free(pmc_mdep, M_PMC);
986 		pmc_mdep = NULL;
987 	}
988 
989 	return (NULL);
990 }
991 
992 /*
993  * Finalization code for AMD CPUs.
994  */
995 
996 void
997 pmc_amd_finalize(struct pmc_mdep *md)
998 {
999 #if	defined(INVARIANTS)
1000 	int classindex, i, ncpus, pmcclass;
1001 #endif
1002 
1003 	pmc_tsc_finalize(md);
1004 
1005 	KASSERT(amd_pcpu != NULL, ("[amd,%d] NULL per-cpu array pointer",
1006 	    __LINE__));
1007 
1008 #if	defined(INVARIANTS)
1009 	switch (md->pmd_cputype) {
1010 #if	defined(__i386__)
1011 	case PMC_CPU_AMD_K7:
1012 		classindex = PMC_MDEP_CLASS_INDEX_K7;
1013 		pmcclass = PMC_CLASS_K7;
1014 		break;
1015 #endif
1016 	default:
1017 		classindex = PMC_MDEP_CLASS_INDEX_K8;
1018 		pmcclass = PMC_CLASS_K8;
1019 	}
1020 
1021 	KASSERT(md->pmd_classdep[classindex].pcd_class == pmcclass,
1022 	    ("[amd,%d] pmc class mismatch", __LINE__));
1023 
1024 	ncpus = pmc_cpu_max();
1025 
1026 	for (i = 0; i < ncpus; i++)
1027 		KASSERT(amd_pcpu[i] == NULL, ("[amd,%d] non-null pcpu",
1028 		    __LINE__));
1029 #endif
1030 
1031 	free(amd_pcpu, M_PMC);
1032 	amd_pcpu = NULL;
1033 }
1034