xref: /freebsd/sys/dev/hwpmc/hwpmc_amd.c (revision 2be1a816b9ff69588e55be0a84cbe2a31efc0f2f)
1 /*-
2  * Copyright (c) 2003-2007 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 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 /* Support for the AMD K7 and later processors */
36 
37 #include <sys/param.h>
38 #include <sys/lock.h>
39 #include <sys/malloc.h>
40 #include <sys/mutex.h>
41 #include <sys/pmc.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	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  = "TSC",
67 		.pd_class = PMC_CLASS_TSC,
68 		.pd_caps  = PMC_CAP_READ,
69 		.pd_width = 64
70 	},
71 	.pm_evsel   = MSR_TSC,
72 	.pm_perfctr = 0	/* unused */
73     },
74 
75     {
76 	.pm_descr =
77 	{
78 		.pd_name  = "",
79 		.pd_class = -1,
80 		.pd_caps  = AMD_PMC_CAPS,
81 		.pd_width = 48
82 	},
83 	.pm_evsel   = AMD_PMC_EVSEL_0,
84 	.pm_perfctr = AMD_PMC_PERFCTR_0
85     },
86     {
87 	.pm_descr =
88 	{
89 		.pd_name  = "",
90 		.pd_class = -1,
91 		.pd_caps  = AMD_PMC_CAPS,
92 		.pd_width = 48
93 	},
94 	.pm_evsel   = AMD_PMC_EVSEL_1,
95 	.pm_perfctr = AMD_PMC_PERFCTR_1
96     },
97     {
98 	.pm_descr =
99 	{
100 		.pd_name  = "",
101 		.pd_class = -1,
102 		.pd_caps  = AMD_PMC_CAPS,
103 		.pd_width = 48
104 	},
105 	.pm_evsel   = AMD_PMC_EVSEL_2,
106 	.pm_perfctr = AMD_PMC_PERFCTR_2
107     },
108     {
109 	.pm_descr =
110 	{
111 		.pd_name  = "",
112 		.pd_class = -1,
113 		.pd_caps  = AMD_PMC_CAPS,
114 		.pd_width = 48
115 	},
116 	.pm_evsel   = AMD_PMC_EVSEL_3,
117 	.pm_perfctr = AMD_PMC_PERFCTR_3
118     }
119 };
120 
121 struct amd_event_code_map {
122 	enum pmc_event	pe_ev;	 /* enum value */
123 	uint8_t		pe_code; /* encoded event mask */
124 	uint8_t		pe_mask; /* bits allowed in unit mask */
125 };
126 
127 const struct amd_event_code_map amd_event_codes[] = {
128 #if	defined(__i386__)	/* 32 bit Athlon (K7) only */
129 	{ PMC_EV_K7_DC_ACCESSES, 		0x40, 0 },
130 	{ PMC_EV_K7_DC_MISSES,			0x41, 0 },
131 	{ PMC_EV_K7_DC_REFILLS_FROM_L2,		0x42, AMD_PMC_UNITMASK_MOESI },
132 	{ PMC_EV_K7_DC_REFILLS_FROM_SYSTEM,	0x43, AMD_PMC_UNITMASK_MOESI },
133 	{ PMC_EV_K7_DC_WRITEBACKS,		0x44, AMD_PMC_UNITMASK_MOESI },
134 	{ PMC_EV_K7_L1_DTLB_MISS_AND_L2_DTLB_HITS, 0x45, 0 },
135 	{ PMC_EV_K7_L1_AND_L2_DTLB_MISSES,	0x46, 0 },
136 	{ PMC_EV_K7_MISALIGNED_REFERENCES,	0x47, 0 },
137 
138 	{ PMC_EV_K7_IC_FETCHES,			0x80, 0 },
139 	{ PMC_EV_K7_IC_MISSES,			0x81, 0 },
140 
141 	{ PMC_EV_K7_L1_ITLB_MISSES,		0x84, 0 },
142 	{ PMC_EV_K7_L1_L2_ITLB_MISSES,		0x85, 0 },
143 
144 	{ PMC_EV_K7_RETIRED_INSTRUCTIONS,	0xC0, 0 },
145 	{ PMC_EV_K7_RETIRED_OPS,		0xC1, 0 },
146 	{ PMC_EV_K7_RETIRED_BRANCHES,		0xC2, 0 },
147 	{ PMC_EV_K7_RETIRED_BRANCHES_MISPREDICTED, 0xC3, 0 },
148 	{ PMC_EV_K7_RETIRED_TAKEN_BRANCHES, 	0xC4, 0 },
149 	{ PMC_EV_K7_RETIRED_TAKEN_BRANCHES_MISPREDICTED, 0xC5, 0 },
150 	{ PMC_EV_K7_RETIRED_FAR_CONTROL_TRANSFERS, 0xC6, 0 },
151 	{ PMC_EV_K7_RETIRED_RESYNC_BRANCHES,	0xC7, 0 },
152 	{ PMC_EV_K7_INTERRUPTS_MASKED_CYCLES,	0xCD, 0 },
153 	{ PMC_EV_K7_INTERRUPTS_MASKED_WHILE_PENDING_CYCLES, 0xCE, 0 },
154 	{ PMC_EV_K7_HARDWARE_INTERRUPTS,	0xCF, 0 },
155 #endif
156 
157 	{ PMC_EV_K8_FP_DISPATCHED_FPU_OPS,		0x00, 0x3F },
158 	{ PMC_EV_K8_FP_CYCLES_WITH_NO_FPU_OPS_RETIRED,	0x01, 0x00 },
159 	{ PMC_EV_K8_FP_DISPATCHED_FPU_FAST_FLAG_OPS,	0x02, 0x00 },
160 
161 	{ PMC_EV_K8_LS_SEGMENT_REGISTER_LOAD, 		0x20, 0x7F },
162 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SELF_MODIFYING_CODE,
163 	  						0x21, 0x00 },
164 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x22, 0x00 },
165 	{ PMC_EV_K8_LS_BUFFER2_FULL,			0x23, 0x00 },
166 	{ PMC_EV_K8_LS_LOCKED_OPERATION,		0x24, 0x07 },
167 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_LATE_CANCEL,	0x25, 0x00 },
168 	{ PMC_EV_K8_LS_RETIRED_CFLUSH_INSTRUCTIONS,	0x26, 0x00 },
169 	{ PMC_EV_K8_LS_RETIRED_CPUID_INSTRUCTIONS,	0x27, 0x00 },
170 
171 	{ PMC_EV_K8_DC_ACCESS,				0x40, 0x00 },
172 	{ PMC_EV_K8_DC_MISS,				0x41, 0x00 },
173 	{ PMC_EV_K8_DC_REFILL_FROM_L2,			0x42, 0x1F },
174 	{ PMC_EV_K8_DC_REFILL_FROM_SYSTEM,		0x43, 0x1F },
175 	{ PMC_EV_K8_DC_COPYBACK,			0x44, 0x1F },
176 	{ PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_HIT,	0x45, 0x00 },
177 	{ PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_MISS,	0x46, 0x00 },
178 	{ PMC_EV_K8_DC_MISALIGNED_DATA_REFERENCE,	0x47, 0x00 },
179 	{ PMC_EV_K8_DC_MICROARCHITECTURAL_LATE_CANCEL,	0x48, 0x00 },
180 	{ PMC_EV_K8_DC_MICROARCHITECTURAL_EARLY_CANCEL, 0x49, 0x00 },
181 	{ PMC_EV_K8_DC_ONE_BIT_ECC_ERROR,		0x4A, 0x03 },
182 	{ PMC_EV_K8_DC_DISPATCHED_PREFETCH_INSTRUCTIONS, 0x4B, 0x07 },
183 	{ PMC_EV_K8_DC_DCACHE_ACCESSES_BY_LOCKS,	0x4C, 0x03 },
184 
185 	{ PMC_EV_K8_BU_CPU_CLK_UNHALTED,		0x76, 0x00 },
186 	{ PMC_EV_K8_BU_INTERNAL_L2_REQUEST,		0x7D, 0x1F },
187 	{ PMC_EV_K8_BU_FILL_REQUEST_L2_MISS,		0x7E, 0x07 },
188 	{ PMC_EV_K8_BU_FILL_INTO_L2,			0x7F, 0x03 },
189 
190 	{ PMC_EV_K8_IC_FETCH,				0x80, 0x00 },
191 	{ PMC_EV_K8_IC_MISS,				0x81, 0x00 },
192 	{ PMC_EV_K8_IC_REFILL_FROM_L2,			0x82, 0x00 },
193 	{ PMC_EV_K8_IC_REFILL_FROM_SYSTEM,		0x83, 0x00 },
194 	{ PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_HIT,	0x84, 0x00 },
195 	{ PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_MISS,	0x85, 0x00 },
196 	{ PMC_EV_K8_IC_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x86, 0x00 },
197 	{ PMC_EV_K8_IC_INSTRUCTION_FETCH_STALL,		0x87, 0x00 },
198 	{ PMC_EV_K8_IC_RETURN_STACK_HIT,		0x88, 0x00 },
199 	{ PMC_EV_K8_IC_RETURN_STACK_OVERFLOW,		0x89, 0x00 },
200 
201 	{ PMC_EV_K8_FR_RETIRED_X86_INSTRUCTIONS,	0xC0, 0x00 },
202 	{ PMC_EV_K8_FR_RETIRED_UOPS,			0xC1, 0x00 },
203 	{ PMC_EV_K8_FR_RETIRED_BRANCHES,		0xC2, 0x00 },
204 	{ PMC_EV_K8_FR_RETIRED_BRANCHES_MISPREDICTED,	0xC3, 0x00 },
205 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES,		0xC4, 0x00 },
206 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED, 0xC5, 0x00 },
207 	{ PMC_EV_K8_FR_RETIRED_FAR_CONTROL_TRANSFERS,	0xC6, 0x00 },
208 	{ PMC_EV_K8_FR_RETIRED_RESYNCS,			0xC7, 0x00 },
209 	{ PMC_EV_K8_FR_RETIRED_NEAR_RETURNS,		0xC8, 0x00 },
210 	{ PMC_EV_K8_FR_RETIRED_NEAR_RETURNS_MISPREDICTED, 0xC9, 0x00 },
211 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED_BY_ADDR_MISCOMPARE,
212 							0xCA, 0x00 },
213 	{ PMC_EV_K8_FR_RETIRED_FPU_INSTRUCTIONS,	0xCB, 0x0F },
214 	{ PMC_EV_K8_FR_RETIRED_FASTPATH_DOUBLE_OP_INSTRUCTIONS,
215 							0xCC, 0x07 },
216 	{ PMC_EV_K8_FR_INTERRUPTS_MASKED_CYCLES,	0xCD, 0x00 },
217 	{ PMC_EV_K8_FR_INTERRUPTS_MASKED_WHILE_PENDING_CYCLES, 0xCE, 0x00 },
218 	{ PMC_EV_K8_FR_TAKEN_HARDWARE_INTERRUPTS,	0xCF, 0x00 },
219 
220 	{ PMC_EV_K8_FR_DECODER_EMPTY,			0xD0, 0x00 },
221 	{ PMC_EV_K8_FR_DISPATCH_STALLS,			0xD1, 0x00 },
222 	{ PMC_EV_K8_FR_DISPATCH_STALL_FROM_BRANCH_ABORT_TO_RETIRE,
223 							0xD2, 0x00 },
224 	{ PMC_EV_K8_FR_DISPATCH_STALL_FOR_SERIALIZATION, 0xD3, 0x00 },
225 	{ PMC_EV_K8_FR_DISPATCH_STALL_FOR_SEGMENT_LOAD,	0xD4, 0x00 },
226 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_REORDER_BUFFER_IS_FULL,
227 							0xD5, 0x00 },
228 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_RESERVATION_STATIONS_ARE_FULL,
229 							0xD6, 0x00 },
230 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FPU_IS_FULL,	0xD7, 0x00 },
231 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_LS_IS_FULL,	0xD8, 0x00 },
232 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_WAITING_FOR_ALL_TO_BE_QUIET,
233 							0xD9, 0x00 },
234 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FAR_XFER_OR_RESYNC_BRANCH_PENDING,
235 							0xDA, 0x00 },
236 	{ PMC_EV_K8_FR_FPU_EXCEPTIONS,			0xDB, 0x0F },
237 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR0,	0xDC, 0x00 },
238 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR1,	0xDD, 0x00 },
239 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR2,	0xDE, 0x00 },
240 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR3,	0xDF, 0x00 },
241 
242 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_ACCESS_EVENT, 0xE0, 0x7 },
243 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_TABLE_OVERFLOW, 0xE1, 0x00 },
244 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_DRAM_COMMAND_SLOTS_MISSED,
245 							0xE2, 0x00 },
246 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_TURNAROUND,	0xE3, 0x07 },
247 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_BYPASS_SATURATION, 0xE4, 0x0F },
248 	{ PMC_EV_K8_NB_SIZED_COMMANDS,			0xEB, 0x7F },
249 	{ PMC_EV_K8_NB_PROBE_RESULT,			0xEC, 0x0F },
250 	{ PMC_EV_K8_NB_HT_BUS0_BANDWIDTH,		0xF6, 0x0F },
251 	{ PMC_EV_K8_NB_HT_BUS1_BANDWIDTH,		0xF7, 0x0F },
252 	{ PMC_EV_K8_NB_HT_BUS2_BANDWIDTH,		0xF8, 0x0F }
253 
254 };
255 
256 const int amd_event_codes_size =
257 	sizeof(amd_event_codes) / sizeof(amd_event_codes[0]);
258 
259 /*
260  * read a pmc register
261  */
262 
263 static int
264 amd_read_pmc(int cpu, int ri, pmc_value_t *v)
265 {
266 	enum pmc_mode mode;
267 	const struct amd_descr *pd;
268 	struct pmc *pm;
269 	const struct pmc_hw *phw;
270 	pmc_value_t tmp;
271 
272 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
273 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
274 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
275 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
276 
277 	phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
278 	pd  = &amd_pmcdesc[ri];
279 	pm  = phw->phw_pmc;
280 
281 	KASSERT(pm != NULL,
282 	    ("[amd,%d] No owner for HWPMC [cpu%d,pmc%d]", __LINE__,
283 		cpu, ri));
284 
285 	mode = PMC_TO_MODE(pm);
286 
287 	PMCDBG(MDP,REA,1,"amd-read id=%d class=%d", ri, pd->pm_descr.pd_class);
288 
289 	/* Reading the TSC is a special case */
290 	if (pd->pm_descr.pd_class == PMC_CLASS_TSC) {
291 		KASSERT(PMC_IS_COUNTING_MODE(mode),
292 		    ("[amd,%d] TSC counter in non-counting mode", __LINE__));
293 		*v = rdtsc();
294 		PMCDBG(MDP,REA,2,"amd-read id=%d -> %jd", ri, *v);
295 		return 0;
296 	}
297 
298 #ifdef	DEBUG
299 	KASSERT(pd->pm_descr.pd_class == amd_pmc_class,
300 	    ("[amd,%d] unknown PMC class (%d)", __LINE__,
301 		pd->pm_descr.pd_class));
302 #endif
303 
304 	tmp = rdmsr(pd->pm_perfctr); /* RDMSR serializes */
305 	PMCDBG(MDP,REA,2,"amd-read (pre-munge) id=%d -> %jd", ri, tmp);
306 	if (PMC_IS_SAMPLING_MODE(mode)) {
307 		/* Sign extend 48 bit value to 64 bits. */
308 		tmp = (pmc_value_t) (((int64_t) tmp << 16) >> 16);
309 		tmp = AMD_PERFCTR_VALUE_TO_RELOAD_COUNT(tmp);
310 	}
311 	*v = tmp;
312 
313 	PMCDBG(MDP,REA,2,"amd-read (post-munge) id=%d -> %jd", ri, *v);
314 
315 	return 0;
316 }
317 
318 /*
319  * Write a PMC MSR.
320  */
321 
322 static int
323 amd_write_pmc(int cpu, int ri, pmc_value_t v)
324 {
325 	const struct amd_descr *pd;
326 	struct pmc *pm;
327 	const struct pmc_hw *phw;
328 	enum pmc_mode mode;
329 
330 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
331 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
332 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
333 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
334 
335 	phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
336 	pd  = &amd_pmcdesc[ri];
337 	pm  = phw->phw_pmc;
338 
339 	KASSERT(pm != NULL,
340 	    ("[amd,%d] PMC not owned (cpu%d,pmc%d)", __LINE__,
341 		cpu, ri));
342 
343 	mode = PMC_TO_MODE(pm);
344 
345 	if (pd->pm_descr.pd_class == PMC_CLASS_TSC)
346 		return 0;
347 
348 #ifdef	DEBUG
349 	KASSERT(pd->pm_descr.pd_class == amd_pmc_class,
350 	    ("[amd,%d] unknown PMC class (%d)", __LINE__,
351 		pd->pm_descr.pd_class));
352 #endif
353 
354 	/* use 2's complement of the count for sampling mode PMCs */
355 	if (PMC_IS_SAMPLING_MODE(mode))
356 		v = AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v);
357 
358 	PMCDBG(MDP,WRI,1,"amd-write cpu=%d ri=%d v=%jx", cpu, ri, v);
359 
360 	/* write the PMC value */
361 	wrmsr(pd->pm_perfctr, v);
362 	return 0;
363 }
364 
365 /*
366  * configure hardware pmc according to the configuration recorded in
367  * pmc 'pm'.
368  */
369 
370 static int
371 amd_config_pmc(int cpu, int ri, struct pmc *pm)
372 {
373 	struct pmc_hw *phw;
374 
375 	PMCDBG(MDP,CFG,1, "cpu=%d ri=%d pm=%p", cpu, ri, pm);
376 
377 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
378 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
379 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
380 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
381 
382 	phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
383 
384 	KASSERT(pm == NULL || phw->phw_pmc == NULL,
385 	    ("[amd,%d] pm=%p phw->pm=%p hwpmc not unconfigured",
386 		__LINE__, pm, phw->phw_pmc));
387 
388 	phw->phw_pmc = pm;
389 	return 0;
390 }
391 
392 /*
393  * Retrieve a configured PMC pointer from hardware state.
394  */
395 
396 static int
397 amd_get_config(int cpu, int ri, struct pmc **ppm)
398 {
399 	*ppm = pmc_pcpu[cpu]->pc_hwpmcs[ri]->phw_pmc;
400 
401 	return 0;
402 }
403 
404 /*
405  * Machine dependent actions taken during the context switch in of a
406  * thread.
407  */
408 
409 static int
410 amd_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
411 {
412 	(void) pc;
413 
414 	PMCDBG(MDP,SWI,1, "pc=%p pp=%p enable-msr=%d", pc, pp,
415 	    (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0);
416 
417 	/* enable the RDPMC instruction if needed */
418 	if (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS)
419 		load_cr4(rcr4() | CR4_PCE);
420 
421 	return 0;
422 }
423 
424 /*
425  * Machine dependent actions taken during the context switch out of a
426  * thread.
427  */
428 
429 static int
430 amd_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
431 {
432 	(void) pc;
433 	(void) pp;		/* can be NULL */
434 
435 	PMCDBG(MDP,SWO,1, "pc=%p pp=%p enable-msr=%d", pc, pp, pp ?
436 	    (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) == 1 : 0);
437 
438 	/* always turn off the RDPMC instruction */
439 	load_cr4(rcr4() & ~CR4_PCE);
440 
441 	return 0;
442 }
443 
444 /*
445  * Check if a given allocation is feasible.
446  */
447 
448 static int
449 amd_allocate_pmc(int cpu, int ri, struct pmc *pm,
450     const struct pmc_op_pmcallocate *a)
451 {
452 	int i;
453 	uint32_t allowed_unitmask, caps, config, unitmask;
454 	enum pmc_event pe;
455 	const struct pmc_descr *pd;
456 
457 	(void) cpu;
458 
459 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
460 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
461 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
462 	    ("[amd,%d] illegal row index %d", __LINE__, ri));
463 
464 	pd = &amd_pmcdesc[ri].pm_descr;
465 
466 	/* check class match */
467 	if (pd->pd_class != a->pm_class)
468 		return EINVAL;
469 
470 	caps = pm->pm_caps;
471 
472 	PMCDBG(MDP,ALL,1,"amd-allocate ri=%d caps=0x%x", ri, caps);
473 
474 	if ((pd->pd_caps & caps) != caps)
475 		return EPERM;
476 	if (pd->pd_class == PMC_CLASS_TSC) {
477 		/* TSC's are always allocated in system-wide counting mode */
478 		if (a->pm_ev != PMC_EV_TSC_TSC ||
479 		    a->pm_mode != PMC_MODE_SC)
480 			return EINVAL;
481 		return 0;
482 	}
483 
484 #ifdef	DEBUG
485 	KASSERT(pd->pd_class == amd_pmc_class,
486 	    ("[amd,%d] Unknown PMC class (%d)", __LINE__, pd->pd_class));
487 #endif
488 
489 	pe = a->pm_ev;
490 
491 	/* map ev to the correct event mask code */
492 	config = allowed_unitmask = 0;
493 	for (i = 0; i < amd_event_codes_size; i++)
494 		if (amd_event_codes[i].pe_ev == pe) {
495 			config =
496 			    AMD_PMC_TO_EVENTMASK(amd_event_codes[i].pe_code);
497 			allowed_unitmask =
498 			    AMD_PMC_TO_UNITMASK(amd_event_codes[i].pe_mask);
499 			break;
500 		}
501 	if (i == amd_event_codes_size)
502 		return EINVAL;
503 
504 	unitmask = a->pm_md.pm_amd.pm_amd_config & AMD_PMC_UNITMASK;
505 	if (unitmask & ~allowed_unitmask) /* disallow reserved bits */
506 		return EINVAL;
507 
508 	if (unitmask && (caps & PMC_CAP_QUALIFIER))
509 		config |= unitmask;
510 
511 	if (caps & PMC_CAP_THRESHOLD)
512 		config |= a->pm_md.pm_amd.pm_amd_config & AMD_PMC_COUNTERMASK;
513 
514 	/* set at least one of the 'usr' or 'os' caps */
515 	if (caps & PMC_CAP_USER)
516 		config |= AMD_PMC_USR;
517 	if (caps & PMC_CAP_SYSTEM)
518 		config |= AMD_PMC_OS;
519 	if ((caps & (PMC_CAP_USER|PMC_CAP_SYSTEM)) == 0)
520 		config |= (AMD_PMC_USR|AMD_PMC_OS);
521 
522 	if (caps & PMC_CAP_EDGE)
523 		config |= AMD_PMC_EDGE;
524 	if (caps & PMC_CAP_INVERT)
525 		config |= AMD_PMC_INVERT;
526 	if (caps & PMC_CAP_INTERRUPT)
527 		config |= AMD_PMC_INT;
528 
529 	pm->pm_md.pm_amd.pm_amd_evsel = config; /* save config value */
530 
531 	PMCDBG(MDP,ALL,2,"amd-allocate ri=%d -> config=0x%x", ri, config);
532 
533 	return 0;
534 }
535 
536 /*
537  * Release machine dependent state associated with a PMC.  This is a
538  * no-op on this architecture.
539  *
540  */
541 
542 /* ARGSUSED0 */
543 static int
544 amd_release_pmc(int cpu, int ri, struct pmc *pmc)
545 {
546 #ifdef	DEBUG
547 	const struct amd_descr *pd;
548 #endif
549 	struct pmc_hw *phw;
550 
551 	(void) pmc;
552 
553 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
554 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
555 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
556 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
557 
558 	phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
559 
560 	KASSERT(phw->phw_pmc == NULL,
561 	    ("[amd,%d] PHW pmc %p non-NULL", __LINE__, phw->phw_pmc));
562 
563 #ifdef	DEBUG
564 	pd = &amd_pmcdesc[ri];
565 	if (pd->pm_descr.pd_class == amd_pmc_class)
566 		KASSERT(AMD_PMC_IS_STOPPED(pd->pm_evsel),
567 		    ("[amd,%d] PMC %d released while active", __LINE__, ri));
568 #endif
569 
570 	return 0;
571 }
572 
573 /*
574  * start a PMC.
575  */
576 
577 static int
578 amd_start_pmc(int cpu, int ri)
579 {
580 	uint32_t config;
581 	struct pmc *pm;
582 	struct pmc_hw *phw;
583 	const struct amd_descr *pd;
584 
585 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
586 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
587 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
588 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
589 
590 	phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
591 	pm  = phw->phw_pmc;
592 	pd = &amd_pmcdesc[ri];
593 
594 	KASSERT(pm != NULL,
595 	    ("[amd,%d] starting cpu%d,pmc%d with null pmc record", __LINE__,
596 		cpu, ri));
597 
598 	PMCDBG(MDP,STA,1,"amd-start cpu=%d ri=%d", cpu, ri);
599 
600 	if (pd->pm_descr.pd_class == PMC_CLASS_TSC)
601 		return 0;	/* TSCs are always running */
602 
603 #ifdef	DEBUG
604 	KASSERT(pd->pm_descr.pd_class == amd_pmc_class,
605 	    ("[amd,%d] unknown PMC class (%d)", __LINE__,
606 		pd->pm_descr.pd_class));
607 #endif
608 
609 	KASSERT(AMD_PMC_IS_STOPPED(pd->pm_evsel),
610 	    ("[amd,%d] pmc%d,cpu%d: Starting active PMC \"%s\"", __LINE__,
611 	    ri, cpu, pd->pm_descr.pd_name));
612 
613 	/* turn on the PMC ENABLE bit */
614 	config = pm->pm_md.pm_amd.pm_amd_evsel | AMD_PMC_ENABLE;
615 
616 	PMCDBG(MDP,STA,2,"amd-start config=0x%x", config);
617 
618 	wrmsr(pd->pm_evsel, config);
619 	return 0;
620 }
621 
622 /*
623  * Stop a PMC.
624  */
625 
626 static int
627 amd_stop_pmc(int cpu, int ri)
628 {
629 	struct pmc *pm;
630 	struct pmc_hw *phw;
631 	const struct amd_descr *pd;
632 	uint64_t config;
633 
634 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
635 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
636 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
637 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
638 
639 	phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
640 	pm  = phw->phw_pmc;
641 	pd  = &amd_pmcdesc[ri];
642 
643 	KASSERT(pm != NULL,
644 	    ("[amd,%d] cpu%d,pmc%d no PMC to stop", __LINE__,
645 		cpu, ri));
646 
647 	/* can't stop a TSC */
648 	if (pd->pm_descr.pd_class == PMC_CLASS_TSC)
649 		return 0;
650 
651 #ifdef	DEBUG
652 	KASSERT(pd->pm_descr.pd_class == amd_pmc_class,
653 	    ("[amd,%d] unknown PMC class (%d)", __LINE__,
654 		pd->pm_descr.pd_class));
655 #endif
656 
657 	KASSERT(!AMD_PMC_IS_STOPPED(pd->pm_evsel),
658 	    ("[amd,%d] PMC%d, CPU%d \"%s\" already stopped",
659 		__LINE__, ri, cpu, pd->pm_descr.pd_name));
660 
661 	PMCDBG(MDP,STO,1,"amd-stop ri=%d", ri);
662 
663 	/* turn off the PMC ENABLE bit */
664 	config = pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE;
665 	wrmsr(pd->pm_evsel, config);
666 	return 0;
667 }
668 
669 /*
670  * Interrupt handler.  This function needs to return '1' if the
671  * interrupt was this CPU's PMCs or '0' otherwise.  It is not allowed
672  * to sleep or do anything a 'fast' interrupt handler is not allowed
673  * to do.
674  */
675 
676 static int
677 amd_intr(int cpu, struct trapframe *tf)
678 {
679 	int i, error, retval, ri;
680 	uint32_t config, evsel, perfctr;
681 	struct pmc *pm;
682 	struct pmc_cpu *pc;
683 	struct pmc_hw *phw;
684 	pmc_value_t v;
685 
686 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
687 	    ("[amd,%d] out of range CPU %d", __LINE__, cpu));
688 
689 	PMCDBG(MDP,INT,1, "cpu=%d tf=%p um=%d", cpu, (void *) tf,
690 	    TRAPF_USERMODE(tf));
691 
692 	retval = 0;
693 
694 	pc = pmc_pcpu[cpu];
695 
696 	/*
697 	 * look for all PMCs that have interrupted:
698 	 * - skip over the TSC [PMC#0]
699 	 * - look for a running, sampling PMC which has overflowed
700 	 *   and which has a valid 'struct pmc' association
701 	 *
702 	 * If found, we call a helper to process the interrupt.
703 	 *
704 	 * If multiple PMCs interrupt at the same time, the AMD64
705 	 * processor appears to deliver as many NMIs as there are
706 	 * outstanding PMC interrupts.  So we process only one NMI
707 	 * interrupt at a time.
708 	 */
709 
710 	for (i = 0; retval == 0 && i < AMD_NPMCS-1; i++) {
711 
712 		ri = i + 1;	/* row index; TSC is at ri == 0 */
713 
714 		if (!AMD_PMC_HAS_OVERFLOWED(i))
715 			continue;
716 
717 		phw = pc->pc_hwpmcs[ri];
718 
719 		KASSERT(phw != NULL, ("[amd,%d] null PHW pointer", __LINE__));
720 
721 		if ((pm = phw->phw_pmc) == NULL ||
722 		    pm->pm_state != PMC_STATE_RUNNING ||
723 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) {
724 			continue;
725 		}
726 
727 		retval = 1;	/* Found an interrupting PMC. */
728 
729 		/* Stop the PMC, reload count. */
730 		evsel   = AMD_PMC_EVSEL_0 + i;
731 		perfctr = AMD_PMC_PERFCTR_0 + i;
732 		v       = pm->pm_sc.pm_reloadcount;
733 		config  = rdmsr(evsel);
734 
735 		KASSERT((config & ~AMD_PMC_ENABLE) ==
736 		    (pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE),
737 		    ("[amd,%d] config mismatch reg=0x%x pm=0x%x", __LINE__,
738 			config, pm->pm_md.pm_amd.pm_amd_evsel));
739 
740 		wrmsr(evsel, config & ~AMD_PMC_ENABLE);
741 		wrmsr(perfctr, AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v));
742 
743 		/* Restart the counter if logging succeeded. */
744 		error = pmc_process_interrupt(cpu, pm, tf, TRAPF_USERMODE(tf));
745 		if (error == 0)
746 			wrmsr(evsel, config | AMD_PMC_ENABLE);
747 	}
748 
749 	atomic_add_int(retval ? &pmc_stats.pm_intr_processed :
750 	    &pmc_stats.pm_intr_ignored, 1);
751 
752 	return (retval);
753 }
754 
755 /*
756  * describe a PMC
757  */
758 static int
759 amd_describe(int cpu, int ri, struct pmc_info *pi, struct pmc **ppmc)
760 {
761 	int error;
762 	size_t copied;
763 	const struct amd_descr *pd;
764 	struct pmc_hw *phw;
765 
766 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
767 	    ("[amd,%d] illegal CPU %d", __LINE__, cpu));
768 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
769 	    ("[amd,%d] row-index %d out of range", __LINE__, ri));
770 
771 	phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
772 	pd  = &amd_pmcdesc[ri];
773 
774 	if ((error = copystr(pd->pm_descr.pd_name, pi->pm_name,
775 		 PMC_NAME_MAX, &copied)) != 0)
776 		return error;
777 
778 	pi->pm_class = pd->pm_descr.pd_class;
779 
780 	if (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) {
781 		pi->pm_enabled = TRUE;
782 		*ppmc          = phw->phw_pmc;
783 	} else {
784 		pi->pm_enabled = FALSE;
785 		*ppmc          = NULL;
786 	}
787 
788 	return 0;
789 }
790 
791 /*
792  * i386 specific entry points
793  */
794 
795 /*
796  * return the MSR address of the given PMC.
797  */
798 
799 static int
800 amd_get_msr(int ri, uint32_t *msr)
801 {
802 	KASSERT(ri >= 0 && ri < AMD_NPMCS,
803 	    ("[amd,%d] ri %d out of range", __LINE__, ri));
804 
805 	*msr = amd_pmcdesc[ri].pm_perfctr - AMD_PMC_PERFCTR_0;
806 	return 0;
807 }
808 
809 /*
810  * processor dependent initialization.
811  */
812 
813 /*
814  * Per-processor data structure
815  *
816  * [common stuff]
817  * [5 struct pmc_hw pointers]
818  * [5 struct pmc_hw structures]
819  */
820 
821 struct amd_cpu {
822 	struct pmc_cpu	pc_common;
823 	struct pmc_hw	*pc_hwpmcs[AMD_NPMCS];
824 	struct pmc_hw	pc_amdpmcs[AMD_NPMCS];
825 };
826 
827 
828 static int
829 amd_init(int cpu)
830 {
831 	int n;
832 	struct amd_cpu *pcs;
833 	struct pmc_hw  *phw;
834 
835 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
836 	    ("[amd,%d] insane cpu number %d", __LINE__, cpu));
837 
838 	PMCDBG(MDP,INI,1,"amd-init cpu=%d", cpu);
839 
840 	MALLOC(pcs, struct amd_cpu *, sizeof(struct amd_cpu), M_PMC,
841 	    M_WAITOK|M_ZERO);
842 
843 	phw = &pcs->pc_amdpmcs[0];
844 
845 	/*
846 	 * Initialize the per-cpu mutex and set the content of the
847 	 * hardware descriptors to a known state.
848 	 */
849 
850 	for (n = 0; n < AMD_NPMCS; n++, phw++) {
851 		phw->phw_state 	  = PMC_PHW_FLAG_IS_ENABLED |
852 		    PMC_PHW_CPU_TO_STATE(cpu) | PMC_PHW_INDEX_TO_STATE(n);
853 		phw->phw_pmc	  = NULL;
854 		pcs->pc_hwpmcs[n] = phw;
855 	}
856 
857 	/* Mark the TSC as shareable */
858 	pcs->pc_hwpmcs[0]->phw_state |= PMC_PHW_FLAG_IS_SHAREABLE;
859 
860 	pmc_pcpu[cpu] = (struct pmc_cpu *) pcs;
861 
862 	return 0;
863 }
864 
865 
866 /*
867  * processor dependent cleanup prior to the KLD
868  * being unloaded
869  */
870 
871 static int
872 amd_cleanup(int cpu)
873 {
874 	int i;
875 	uint32_t evsel;
876 	struct pmc_cpu *pcs;
877 
878 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
879 	    ("[amd,%d] insane cpu number (%d)", __LINE__, cpu));
880 
881 	PMCDBG(MDP,INI,1,"amd-cleanup cpu=%d", cpu);
882 
883 	/*
884 	 * First, turn off all PMCs on this CPU.
885 	 */
886 
887 	for (i = 0; i < 4; i++) { /* XXX this loop is now not needed */
888 		evsel = rdmsr(AMD_PMC_EVSEL_0 + i);
889 		evsel &= ~AMD_PMC_ENABLE;
890 		wrmsr(AMD_PMC_EVSEL_0 + i, evsel);
891 	}
892 
893 	/*
894 	 * Next, free up allocated space.
895 	 */
896 
897 	if ((pcs = pmc_pcpu[cpu]) == NULL)
898 		return 0;
899 
900 #ifdef	DEBUG
901 	/* check the TSC */
902 	KASSERT(pcs->pc_hwpmcs[0]->phw_pmc == NULL,
903 	    ("[amd,%d] CPU%d,PMC0 still in use", __LINE__, cpu));
904 	for (i = 1; i < AMD_NPMCS; i++) {
905 		KASSERT(pcs->pc_hwpmcs[i]->phw_pmc == NULL,
906 		    ("[amd,%d] CPU%d/PMC%d in use", __LINE__, cpu, i));
907 		KASSERT(AMD_PMC_IS_STOPPED(AMD_PMC_EVSEL_0 + (i-1)),
908 		    ("[amd,%d] CPU%d/PMC%d not stopped", __LINE__, cpu, i));
909 	}
910 #endif
911 
912 	pmc_pcpu[cpu] = NULL;
913 	FREE(pcs, M_PMC);
914 	return 0;
915 }
916 
917 /*
918  * Initialize ourselves.
919  */
920 
921 struct pmc_mdep *
922 pmc_amd_initialize(void)
923 {
924 	enum pmc_cputype cputype;
925 	enum pmc_class class;
926 	struct pmc_mdep *pmc_mdep;
927 	char *name;
928 	int i;
929 
930 	/*
931 	 * The presence of hardware performance counters on the AMD
932 	 * Athlon, Duron or later processors, is _not_ indicated by
933 	 * any of the processor feature flags set by the 'CPUID'
934 	 * instruction, so we only check the 'instruction family'
935 	 * field returned by CPUID for instruction family >= 6.
936 	 */
937 
938 	class = cputype = -1;
939 	name = NULL;
940 	switch (cpu_id & 0xF00) {
941 	case 0x600:		/* Athlon(tm) processor */
942 		cputype = PMC_CPU_AMD_K7;
943 		class = PMC_CLASS_K7;
944 		name = "K7";
945 		break;
946 	case 0xF00:		/* Athlon64/Opteron processor */
947 		cputype = PMC_CPU_AMD_K8;
948 		class = PMC_CLASS_K8;
949 		name = "K8";
950 		break;
951 	}
952 
953 	if ((int) cputype == -1) {
954 		(void) printf("pmc: Unknown AMD CPU.\n");
955 		return NULL;
956 	}
957 
958 #ifdef	DEBUG
959 	amd_pmc_class = class;
960 #endif
961 
962 	MALLOC(pmc_mdep, struct pmc_mdep *, sizeof(struct pmc_mdep),
963 	    M_PMC, M_WAITOK|M_ZERO);
964 
965 	pmc_mdep->pmd_cputype	   = cputype;
966 	pmc_mdep->pmd_npmc 	   = AMD_NPMCS;
967 
968 	/* this processor has two classes of usable PMCs */
969 	pmc_mdep->pmd_nclass       = 2;
970 
971 	/* TSC */
972 	pmc_mdep->pmd_classes[0].pm_class   = PMC_CLASS_TSC;
973 	pmc_mdep->pmd_classes[0].pm_caps    = PMC_CAP_READ;
974 	pmc_mdep->pmd_classes[0].pm_width   = 64;
975 
976 	/* AMD K7/K8 PMCs */
977 	pmc_mdep->pmd_classes[1].pm_class   = class;
978 	pmc_mdep->pmd_classes[1].pm_caps    = AMD_PMC_CAPS;
979 	pmc_mdep->pmd_classes[1].pm_width   = 48;
980 
981 	pmc_mdep->pmd_nclasspmcs[0] = 1;
982 	pmc_mdep->pmd_nclasspmcs[1] = (AMD_NPMCS-1);
983 
984 	/* fill in the correct pmc name and class */
985 	for (i = 1; i < AMD_NPMCS; i++) {
986 		(void) snprintf(amd_pmcdesc[i].pm_descr.pd_name,
987 		    sizeof(amd_pmcdesc[i].pm_descr.pd_name), "%s-%d",
988 		    name, i-1);
989 		amd_pmcdesc[i].pm_descr.pd_class = class;
990 	}
991 
992 	pmc_mdep->pmd_init    	   = amd_init;
993 	pmc_mdep->pmd_cleanup 	   = amd_cleanup;
994 	pmc_mdep->pmd_switch_in    = amd_switch_in;
995 	pmc_mdep->pmd_switch_out   = amd_switch_out;
996 	pmc_mdep->pmd_read_pmc 	   = amd_read_pmc;
997 	pmc_mdep->pmd_write_pmc    = amd_write_pmc;
998 	pmc_mdep->pmd_config_pmc   = amd_config_pmc;
999 	pmc_mdep->pmd_get_config   = amd_get_config;
1000 	pmc_mdep->pmd_allocate_pmc = amd_allocate_pmc;
1001 	pmc_mdep->pmd_release_pmc  = amd_release_pmc;
1002 	pmc_mdep->pmd_start_pmc    = amd_start_pmc;
1003 	pmc_mdep->pmd_stop_pmc     = amd_stop_pmc;
1004 	pmc_mdep->pmd_intr	   = amd_intr;
1005 	pmc_mdep->pmd_describe     = amd_describe;
1006 	pmc_mdep->pmd_get_msr  	   = amd_get_msr; /* i386 */
1007 
1008 	PMCDBG(MDP,INI,0,"%s","amd-initialize");
1009 
1010 	return pmc_mdep;
1011 }
1012