2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2003-2008 Joseph Koshy
5 * Copyright (c) 2007 The FreeBSD Foundation
8 * Portions of this software were developed by A. Joseph Koshy under
9 * sponsorship from the FreeBSD Foundation and Google, Inc.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
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.
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
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
36 /* Support for the AMD K7 and later processors */
38 #include <sys/param.h>
40 #include <sys/malloc.h>
41 #include <sys/mutex.h>
43 #include <sys/pmckern.h>
45 #include <sys/systm.h>
47 #include <machine/cpu.h>
48 #include <machine/cpufunc.h>
49 #include <machine/md_var.h>
50 #include <machine/specialreg.h>
53 enum pmc_class amd_pmc_class;
56 /* AMD K7 & K8 PMCs */
58 struct pmc_descr pm_descr; /* "base class" */
59 uint32_t pm_evsel; /* address of EVSEL register */
60 uint32_t pm_perfctr; /* address of PERFCTR register */
63 static struct amd_descr amd_pmcdesc[AMD_NPMCS] =
70 .pd_caps = AMD_PMC_CAPS,
73 .pm_evsel = AMD_PMC_EVSEL_0,
74 .pm_perfctr = AMD_PMC_PERFCTR_0
81 .pd_caps = AMD_PMC_CAPS,
84 .pm_evsel = AMD_PMC_EVSEL_1,
85 .pm_perfctr = AMD_PMC_PERFCTR_1
92 .pd_caps = AMD_PMC_CAPS,
95 .pm_evsel = AMD_PMC_EVSEL_2,
96 .pm_perfctr = AMD_PMC_PERFCTR_2
103 .pd_caps = AMD_PMC_CAPS,
106 .pm_evsel = AMD_PMC_EVSEL_3,
107 .pm_perfctr = AMD_PMC_PERFCTR_3
114 .pd_caps = AMD_PMC_CAPS,
117 .pm_evsel = AMD_PMC_EVSEL_4,
118 .pm_perfctr = AMD_PMC_PERFCTR_4
125 .pd_caps = AMD_PMC_CAPS,
128 .pm_evsel = AMD_PMC_EVSEL_5,
129 .pm_perfctr = AMD_PMC_PERFCTR_5
136 .pd_caps = AMD_PMC_CAPS,
139 .pm_evsel = AMD_PMC_EVSEL_EP_L3_0,
140 .pm_perfctr = AMD_PMC_PERFCTR_EP_L3_0
147 .pd_caps = AMD_PMC_CAPS,
150 .pm_evsel = AMD_PMC_EVSEL_EP_L3_1,
151 .pm_perfctr = AMD_PMC_PERFCTR_EP_L3_1
158 .pd_caps = AMD_PMC_CAPS,
161 .pm_evsel = AMD_PMC_EVSEL_EP_L3_2,
162 .pm_perfctr = AMD_PMC_PERFCTR_EP_L3_2
169 .pd_caps = AMD_PMC_CAPS,
172 .pm_evsel = AMD_PMC_EVSEL_EP_L3_3,
173 .pm_perfctr = AMD_PMC_PERFCTR_EP_L3_3
180 .pd_caps = AMD_PMC_CAPS,
183 .pm_evsel = AMD_PMC_EVSEL_EP_L3_4,
184 .pm_perfctr = AMD_PMC_PERFCTR_EP_L3_4
191 .pd_caps = AMD_PMC_CAPS,
194 .pm_evsel = AMD_PMC_EVSEL_EP_L3_5,
195 .pm_perfctr = AMD_PMC_PERFCTR_EP_L3_5
202 .pd_caps = AMD_PMC_CAPS,
205 .pm_evsel = AMD_PMC_EVSEL_EP_DF_0,
206 .pm_perfctr = AMD_PMC_PERFCTR_EP_DF_0
213 .pd_caps = AMD_PMC_CAPS,
216 .pm_evsel = AMD_PMC_EVSEL_EP_DF_1,
217 .pm_perfctr = AMD_PMC_PERFCTR_EP_DF_1
224 .pd_caps = AMD_PMC_CAPS,
227 .pm_evsel = AMD_PMC_EVSEL_EP_DF_2,
228 .pm_perfctr = AMD_PMC_PERFCTR_EP_DF_2
235 .pd_caps = AMD_PMC_CAPS,
238 .pm_evsel = AMD_PMC_EVSEL_EP_DF_3,
239 .pm_perfctr = AMD_PMC_PERFCTR_EP_DF_3
243 struct amd_event_code_map {
244 enum pmc_event pe_ev; /* enum value */
245 uint16_t pe_code; /* encoded event mask */
246 uint8_t pe_mask; /* bits allowed in unit mask */
249 const struct amd_event_code_map amd_event_codes[] = {
250 #if defined(__i386__) /* 32 bit Athlon (K7) only */
251 { PMC_EV_K7_DC_ACCESSES, 0x40, 0 },
252 { PMC_EV_K7_DC_MISSES, 0x41, 0 },
253 { PMC_EV_K7_DC_REFILLS_FROM_L2, 0x42, AMD_PMC_UNITMASK_MOESI },
254 { PMC_EV_K7_DC_REFILLS_FROM_SYSTEM, 0x43, AMD_PMC_UNITMASK_MOESI },
255 { PMC_EV_K7_DC_WRITEBACKS, 0x44, AMD_PMC_UNITMASK_MOESI },
256 { PMC_EV_K7_L1_DTLB_MISS_AND_L2_DTLB_HITS, 0x45, 0 },
257 { PMC_EV_K7_L1_AND_L2_DTLB_MISSES, 0x46, 0 },
258 { PMC_EV_K7_MISALIGNED_REFERENCES, 0x47, 0 },
260 { PMC_EV_K7_IC_FETCHES, 0x80, 0 },
261 { PMC_EV_K7_IC_MISSES, 0x81, 0 },
263 { PMC_EV_K7_L1_ITLB_MISSES, 0x84, 0 },
264 { PMC_EV_K7_L1_L2_ITLB_MISSES, 0x85, 0 },
266 { PMC_EV_K7_RETIRED_INSTRUCTIONS, 0xC0, 0 },
267 { PMC_EV_K7_RETIRED_OPS, 0xC1, 0 },
268 { PMC_EV_K7_RETIRED_BRANCHES, 0xC2, 0 },
269 { PMC_EV_K7_RETIRED_BRANCHES_MISPREDICTED, 0xC3, 0 },
270 { PMC_EV_K7_RETIRED_TAKEN_BRANCHES, 0xC4, 0 },
271 { PMC_EV_K7_RETIRED_TAKEN_BRANCHES_MISPREDICTED, 0xC5, 0 },
272 { PMC_EV_K7_RETIRED_FAR_CONTROL_TRANSFERS, 0xC6, 0 },
273 { PMC_EV_K7_RETIRED_RESYNC_BRANCHES, 0xC7, 0 },
274 { PMC_EV_K7_INTERRUPTS_MASKED_CYCLES, 0xCD, 0 },
275 { PMC_EV_K7_INTERRUPTS_MASKED_WHILE_PENDING_CYCLES, 0xCE, 0 },
276 { PMC_EV_K7_HARDWARE_INTERRUPTS, 0xCF, 0 },
279 { PMC_EV_K8_FP_DISPATCHED_FPU_OPS, 0x00, 0x3F },
280 { PMC_EV_K8_FP_CYCLES_WITH_NO_FPU_OPS_RETIRED, 0x01, 0x00 },
281 { PMC_EV_K8_FP_DISPATCHED_FPU_FAST_FLAG_OPS, 0x02, 0x00 },
283 { PMC_EV_K8_LS_SEGMENT_REGISTER_LOAD, 0x20, 0x7F },
284 { PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SELF_MODIFYING_CODE,
286 { PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x22, 0x00 },
287 { PMC_EV_K8_LS_BUFFER2_FULL, 0x23, 0x00 },
288 { PMC_EV_K8_LS_LOCKED_OPERATION, 0x24, 0x07 },
289 { PMC_EV_K8_LS_MICROARCHITECTURAL_LATE_CANCEL, 0x25, 0x00 },
290 { PMC_EV_K8_LS_RETIRED_CFLUSH_INSTRUCTIONS, 0x26, 0x00 },
291 { PMC_EV_K8_LS_RETIRED_CPUID_INSTRUCTIONS, 0x27, 0x00 },
293 { PMC_EV_K8_DC_ACCESS, 0x40, 0x00 },
294 { PMC_EV_K8_DC_MISS, 0x41, 0x00 },
295 { PMC_EV_K8_DC_REFILL_FROM_L2, 0x42, 0x1F },
296 { PMC_EV_K8_DC_REFILL_FROM_SYSTEM, 0x43, 0x1F },
297 { PMC_EV_K8_DC_COPYBACK, 0x44, 0x1F },
298 { PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_HIT, 0x45, 0x00 },
299 { PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_MISS, 0x46, 0x00 },
300 { PMC_EV_K8_DC_MISALIGNED_DATA_REFERENCE, 0x47, 0x00 },
301 { PMC_EV_K8_DC_MICROARCHITECTURAL_LATE_CANCEL, 0x48, 0x00 },
302 { PMC_EV_K8_DC_MICROARCHITECTURAL_EARLY_CANCEL, 0x49, 0x00 },
303 { PMC_EV_K8_DC_ONE_BIT_ECC_ERROR, 0x4A, 0x03 },
304 { PMC_EV_K8_DC_DISPATCHED_PREFETCH_INSTRUCTIONS, 0x4B, 0x07 },
305 { PMC_EV_K8_DC_DCACHE_ACCESSES_BY_LOCKS, 0x4C, 0x03 },
307 { PMC_EV_K8_BU_CPU_CLK_UNHALTED, 0x76, 0x00 },
308 { PMC_EV_K8_BU_INTERNAL_L2_REQUEST, 0x7D, 0x1F },
309 { PMC_EV_K8_BU_FILL_REQUEST_L2_MISS, 0x7E, 0x07 },
310 { PMC_EV_K8_BU_FILL_INTO_L2, 0x7F, 0x03 },
312 { PMC_EV_K8_IC_FETCH, 0x80, 0x00 },
313 { PMC_EV_K8_IC_MISS, 0x81, 0x00 },
314 { PMC_EV_K8_IC_REFILL_FROM_L2, 0x82, 0x00 },
315 { PMC_EV_K8_IC_REFILL_FROM_SYSTEM, 0x83, 0x00 },
316 { PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_HIT, 0x84, 0x00 },
317 { PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_MISS, 0x85, 0x00 },
318 { PMC_EV_K8_IC_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x86, 0x00 },
319 { PMC_EV_K8_IC_INSTRUCTION_FETCH_STALL, 0x87, 0x00 },
320 { PMC_EV_K8_IC_RETURN_STACK_HIT, 0x88, 0x00 },
321 { PMC_EV_K8_IC_RETURN_STACK_OVERFLOW, 0x89, 0x00 },
323 { PMC_EV_K8_FR_RETIRED_X86_INSTRUCTIONS, 0xC0, 0x00 },
324 { PMC_EV_K8_FR_RETIRED_UOPS, 0xC1, 0x00 },
325 { PMC_EV_K8_FR_RETIRED_BRANCHES, 0xC2, 0x00 },
326 { PMC_EV_K8_FR_RETIRED_BRANCHES_MISPREDICTED, 0xC3, 0x00 },
327 { PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES, 0xC4, 0x00 },
328 { PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED, 0xC5, 0x00 },
329 { PMC_EV_K8_FR_RETIRED_FAR_CONTROL_TRANSFERS, 0xC6, 0x00 },
330 { PMC_EV_K8_FR_RETIRED_RESYNCS, 0xC7, 0x00 },
331 { PMC_EV_K8_FR_RETIRED_NEAR_RETURNS, 0xC8, 0x00 },
332 { PMC_EV_K8_FR_RETIRED_NEAR_RETURNS_MISPREDICTED, 0xC9, 0x00 },
333 { PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED_BY_ADDR_MISCOMPARE,
335 { PMC_EV_K8_FR_RETIRED_FPU_INSTRUCTIONS, 0xCB, 0x0F },
336 { PMC_EV_K8_FR_RETIRED_FASTPATH_DOUBLE_OP_INSTRUCTIONS,
338 { PMC_EV_K8_FR_INTERRUPTS_MASKED_CYCLES, 0xCD, 0x00 },
339 { PMC_EV_K8_FR_INTERRUPTS_MASKED_WHILE_PENDING_CYCLES, 0xCE, 0x00 },
340 { PMC_EV_K8_FR_TAKEN_HARDWARE_INTERRUPTS, 0xCF, 0x00 },
342 { PMC_EV_K8_FR_DECODER_EMPTY, 0xD0, 0x00 },
343 { PMC_EV_K8_FR_DISPATCH_STALLS, 0xD1, 0x00 },
344 { PMC_EV_K8_FR_DISPATCH_STALL_FROM_BRANCH_ABORT_TO_RETIRE,
346 { PMC_EV_K8_FR_DISPATCH_STALL_FOR_SERIALIZATION, 0xD3, 0x00 },
347 { PMC_EV_K8_FR_DISPATCH_STALL_FOR_SEGMENT_LOAD, 0xD4, 0x00 },
348 { PMC_EV_K8_FR_DISPATCH_STALL_WHEN_REORDER_BUFFER_IS_FULL,
350 { PMC_EV_K8_FR_DISPATCH_STALL_WHEN_RESERVATION_STATIONS_ARE_FULL,
352 { PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FPU_IS_FULL, 0xD7, 0x00 },
353 { PMC_EV_K8_FR_DISPATCH_STALL_WHEN_LS_IS_FULL, 0xD8, 0x00 },
354 { PMC_EV_K8_FR_DISPATCH_STALL_WHEN_WAITING_FOR_ALL_TO_BE_QUIET,
356 { PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FAR_XFER_OR_RESYNC_BRANCH_PENDING,
358 { PMC_EV_K8_FR_FPU_EXCEPTIONS, 0xDB, 0x0F },
359 { PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR0, 0xDC, 0x00 },
360 { PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR1, 0xDD, 0x00 },
361 { PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR2, 0xDE, 0x00 },
362 { PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR3, 0xDF, 0x00 },
364 { PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_ACCESS_EVENT, 0xE0, 0x7 },
365 { PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_TABLE_OVERFLOW, 0xE1, 0x00 },
366 { PMC_EV_K8_NB_MEMORY_CONTROLLER_DRAM_COMMAND_SLOTS_MISSED,
368 { PMC_EV_K8_NB_MEMORY_CONTROLLER_TURNAROUND, 0xE3, 0x07 },
369 { PMC_EV_K8_NB_MEMORY_CONTROLLER_BYPASS_SATURATION, 0xE4, 0x0F },
370 { PMC_EV_K8_NB_SIZED_COMMANDS, 0xEB, 0x7F },
371 { PMC_EV_K8_NB_PROBE_RESULT, 0xEC, 0x0F },
372 { PMC_EV_K8_NB_HT_BUS0_BANDWIDTH, 0xF6, 0x0F },
373 { PMC_EV_K8_NB_HT_BUS1_BANDWIDTH, 0xF7, 0x0F },
374 { PMC_EV_K8_NB_HT_BUS2_BANDWIDTH, 0xF8, 0x0F }
378 const int amd_event_codes_size = nitems(amd_event_codes);
381 * Per-processor information
385 struct pmc_hw pc_amdpmcs[AMD_NPMCS];
388 static struct amd_cpu **amd_pcpu;
391 * read a pmc register
395 amd_read_pmc(int cpu, int ri, pmc_value_t *v)
398 const struct amd_descr *pd;
402 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
403 ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
404 KASSERT(ri >= 0 && ri < AMD_NPMCS,
405 ("[amd,%d] illegal row-index %d", __LINE__, ri));
406 KASSERT(amd_pcpu[cpu],
407 ("[amd,%d] null per-cpu, cpu %d", __LINE__, cpu));
409 pm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc;
410 pd = &amd_pmcdesc[ri];
413 ("[amd,%d] No owner for HWPMC [cpu%d,pmc%d]", __LINE__,
416 mode = PMC_TO_MODE(pm);
418 PMCDBG2(MDP,REA,1,"amd-read id=%d class=%d", ri, pd->pm_descr.pd_class);
421 KASSERT(pd->pm_descr.pd_class == amd_pmc_class,
422 ("[amd,%d] unknown PMC class (%d)", __LINE__,
423 pd->pm_descr.pd_class));
426 tmp = rdmsr(pd->pm_perfctr); /* RDMSR serializes */
427 PMCDBG2(MDP,REA,2,"amd-read (pre-munge) id=%d -> %jd", ri, tmp);
428 if (PMC_IS_SAMPLING_MODE(mode)) {
429 /* Sign extend 48 bit value to 64 bits. */
430 tmp = (pmc_value_t) (((int64_t) tmp << 16) >> 16);
431 tmp = AMD_PERFCTR_VALUE_TO_RELOAD_COUNT(tmp);
435 PMCDBG2(MDP,REA,2,"amd-read (post-munge) id=%d -> %jd", ri, *v);
445 amd_write_pmc(int cpu, int ri, pmc_value_t v)
447 const struct amd_descr *pd;
451 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
452 ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
453 KASSERT(ri >= 0 && ri < AMD_NPMCS,
454 ("[amd,%d] illegal row-index %d", __LINE__, ri));
456 pm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc;
457 pd = &amd_pmcdesc[ri];
460 ("[amd,%d] PMC not owned (cpu%d,pmc%d)", __LINE__,
463 mode = PMC_TO_MODE(pm);
466 KASSERT(pd->pm_descr.pd_class == amd_pmc_class,
467 ("[amd,%d] unknown PMC class (%d)", __LINE__,
468 pd->pm_descr.pd_class));
471 /* use 2's complement of the count for sampling mode PMCs */
472 if (PMC_IS_SAMPLING_MODE(mode))
473 v = AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v);
475 PMCDBG3(MDP,WRI,1,"amd-write cpu=%d ri=%d v=%jx", cpu, ri, v);
477 /* write the PMC value */
478 wrmsr(pd->pm_perfctr, v);
483 * configure hardware pmc according to the configuration recorded in
488 amd_config_pmc(int cpu, int ri, struct pmc *pm)
492 PMCDBG3(MDP,CFG,1, "cpu=%d ri=%d pm=%p", cpu, ri, pm);
494 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
495 ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
496 KASSERT(ri >= 0 && ri < AMD_NPMCS,
497 ("[amd,%d] illegal row-index %d", __LINE__, ri));
499 phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
501 KASSERT(pm == NULL || phw->phw_pmc == NULL,
502 ("[amd,%d] pm=%p phw->pm=%p hwpmc not unconfigured",
503 __LINE__, pm, phw->phw_pmc));
510 * Retrieve a configured PMC pointer from hardware state.
514 amd_get_config(int cpu, int ri, struct pmc **ppm)
516 *ppm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc;
522 * Machine dependent actions taken during the context switch in of a
527 amd_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
531 PMCDBG3(MDP,SWI,1, "pc=%p pp=%p enable-msr=%d", pc, pp,
532 (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0);
534 /* enable the RDPMC instruction if needed */
535 if (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS)
536 load_cr4(rcr4() | CR4_PCE);
542 * Machine dependent actions taken during the context switch out of a
547 amd_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
550 (void) pp; /* can be NULL */
552 PMCDBG3(MDP,SWO,1, "pc=%p pp=%p enable-msr=%d", pc, pp, pp ?
553 (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) == 1 : 0);
555 /* always turn off the RDPMC instruction */
556 load_cr4(rcr4() & ~CR4_PCE);
562 * Check if a given allocation is feasible.
566 amd_allocate_pmc(int cpu, int ri, struct pmc *pm,
567 const struct pmc_op_pmcallocate *a)
570 uint64_t allowed_unitmask, caps, config, unitmask;
572 const struct pmc_descr *pd;
576 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
577 ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
578 KASSERT(ri >= 0 && ri < AMD_NPMCS,
579 ("[amd,%d] illegal row index %d", __LINE__, ri));
581 pd = &amd_pmcdesc[ri].pm_descr;
583 /* check class match */
584 if (pd->pd_class != a->pm_class)
589 PMCDBG2(MDP,ALL,1,"amd-allocate ri=%d caps=0x%x", ri, caps);
591 if((ri >= 0 && ri < 6) && !(a->pm_md.pm_amd.pm_amd_sub_class == PMC_AMD_SUB_CLASS_CORE))
593 if((ri >= 6 && ri < 12) && !(a->pm_md.pm_amd.pm_amd_sub_class == PMC_AMD_SUB_CLASS_L3_CACHE))
595 if((ri >= 12 && ri < 16) && !(a->pm_md.pm_amd.pm_amd_sub_class == PMC_AMD_SUB_CLASS_DATA_FABRIC))
598 if ((pd->pd_caps & caps) != caps)
600 if (strlen(pmc_cpuid) != 0) {
601 pm->pm_md.pm_amd.pm_amd_evsel =
602 a->pm_md.pm_amd.pm_amd_config;
603 PMCDBG2(MDP,ALL,2,"amd-allocate ri=%d -> config=0x%x", ri, a->pm_md.pm_amd.pm_amd_config);
609 /* map ev to the correct event mask code */
610 config = allowed_unitmask = 0;
611 for (i = 0; i < amd_event_codes_size; i++)
612 if (amd_event_codes[i].pe_ev == pe) {
614 AMD_PMC_TO_EVENTMASK(amd_event_codes[i].pe_code);
616 AMD_PMC_TO_UNITMASK(amd_event_codes[i].pe_mask);
619 if (i == amd_event_codes_size)
622 unitmask = a->pm_md.pm_amd.pm_amd_config & AMD_PMC_UNITMASK;
623 if (unitmask & ~allowed_unitmask) /* disallow reserved bits */
626 if (unitmask && (caps & PMC_CAP_QUALIFIER))
629 if (caps & PMC_CAP_THRESHOLD)
630 config |= a->pm_md.pm_amd.pm_amd_config & AMD_PMC_COUNTERMASK;
632 /* set at least one of the 'usr' or 'os' caps */
633 if (caps & PMC_CAP_USER)
634 config |= AMD_PMC_USR;
635 if (caps & PMC_CAP_SYSTEM)
636 config |= AMD_PMC_OS;
637 if ((caps & (PMC_CAP_USER|PMC_CAP_SYSTEM)) == 0)
638 config |= (AMD_PMC_USR|AMD_PMC_OS);
640 if (caps & PMC_CAP_EDGE)
641 config |= AMD_PMC_EDGE;
642 if (caps & PMC_CAP_INVERT)
643 config |= AMD_PMC_INVERT;
644 if (caps & PMC_CAP_INTERRUPT)
645 config |= AMD_PMC_INT;
647 pm->pm_md.pm_amd.pm_amd_evsel = config; /* save config value */
649 PMCDBG2(MDP,ALL,2,"amd-allocate ri=%d -> config=0x%x", ri, config);
655 * Release machine dependent state associated with a PMC. This is a
656 * no-op on this architecture.
662 amd_release_pmc(int cpu, int ri, struct pmc *pmc)
665 const struct amd_descr *pd;
671 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
672 ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
673 KASSERT(ri >= 0 && ri < AMD_NPMCS,
674 ("[amd,%d] illegal row-index %d", __LINE__, ri));
676 phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
678 KASSERT(phw->phw_pmc == NULL,
679 ("[amd,%d] PHW pmc %p non-NULL", __LINE__, phw->phw_pmc));
682 pd = &amd_pmcdesc[ri];
683 if (pd->pm_descr.pd_class == amd_pmc_class)
684 KASSERT(AMD_PMC_IS_STOPPED(pd->pm_evsel),
685 ("[amd,%d] PMC %d released while active", __LINE__, ri));
696 amd_start_pmc(int cpu, int ri)
701 const struct amd_descr *pd;
703 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
704 ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
705 KASSERT(ri >= 0 && ri < AMD_NPMCS,
706 ("[amd,%d] illegal row-index %d", __LINE__, ri));
708 phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
710 pd = &amd_pmcdesc[ri];
713 ("[amd,%d] starting cpu%d,pmc%d with null pmc record", __LINE__,
716 PMCDBG2(MDP,STA,1,"amd-start cpu=%d ri=%d", cpu, ri);
718 KASSERT(AMD_PMC_IS_STOPPED(pd->pm_evsel),
719 ("[amd,%d] pmc%d,cpu%d: Starting active PMC \"%s\"", __LINE__,
720 ri, cpu, pd->pm_descr.pd_name));
722 /* turn on the PMC ENABLE bit */
723 config = pm->pm_md.pm_amd.pm_amd_evsel | AMD_PMC_ENABLE;
725 PMCDBG1(MDP,STA,2,"amd-start config=0x%x", config);
727 wrmsr(pd->pm_evsel, config);
736 amd_stop_pmc(int cpu, int ri)
740 const struct amd_descr *pd;
743 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
744 ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
745 KASSERT(ri >= 0 && ri < AMD_NPMCS,
746 ("[amd,%d] illegal row-index %d", __LINE__, ri));
748 phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
750 pd = &amd_pmcdesc[ri];
753 ("[amd,%d] cpu%d,pmc%d no PMC to stop", __LINE__,
755 KASSERT(!AMD_PMC_IS_STOPPED(pd->pm_evsel),
756 ("[amd,%d] PMC%d, CPU%d \"%s\" already stopped",
757 __LINE__, ri, cpu, pd->pm_descr.pd_name));
759 PMCDBG1(MDP,STO,1,"amd-stop ri=%d", ri);
761 /* turn off the PMC ENABLE bit */
762 config = pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE;
763 wrmsr(pd->pm_evsel, config);
768 * Interrupt handler. This function needs to return '1' if the
769 * interrupt was this CPU's PMCs or '0' otherwise. It is not allowed
770 * to sleep or do anything a 'fast' interrupt handler is not allowed
775 amd_intr(struct trapframe *tf)
777 int i, error, retval, cpu;
778 uint64_t config, evsel, perfctr;
784 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
785 ("[amd,%d] out of range CPU %d", __LINE__, cpu));
787 PMCDBG3(MDP,INT,1, "cpu=%d tf=%p um=%d", cpu, (void *) tf,
795 * look for all PMCs that have interrupted:
796 * - look for a running, sampling PMC which has overflowed
797 * and which has a valid 'struct pmc' association
799 * If found, we call a helper to process the interrupt.
801 * If multiple PMCs interrupt at the same time, the AMD64
802 * processor appears to deliver as many NMIs as there are
803 * outstanding PMC interrupts. So we process only one NMI
804 * interrupt at a time.
807 for (i = 0; retval == 0 && i < AMD_NPMCS; i++) {
809 if ((pm = pac->pc_amdpmcs[i].phw_pmc) == NULL ||
810 !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) {
814 if (!AMD_PMC_HAS_OVERFLOWED(i))
817 retval = 1; /* Found an interrupting PMC. */
819 if (pm->pm_state != PMC_STATE_RUNNING)
822 /* Stop the PMC, reload count. */
823 evsel = AMD_PMC_EVSEL_0 + i;
824 perfctr = AMD_PMC_PERFCTR_0 + i;
825 v = pm->pm_sc.pm_reloadcount;
826 config = rdmsr(evsel);
828 KASSERT((config & ~AMD_PMC_ENABLE) ==
829 (pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE),
830 ("[amd,%d] config mismatch reg=0x%jx pm=0x%jx", __LINE__,
831 (uintmax_t)config, (uintmax_t)pm->pm_md.pm_amd.pm_amd_evsel));
833 wrmsr(evsel, config & ~AMD_PMC_ENABLE);
834 wrmsr(perfctr, AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v));
836 /* Restart the counter if logging succeeded. */
837 error = pmc_process_interrupt(PMC_HR, pm, tf);
839 wrmsr(evsel, config);
843 counter_u64_add(pmc_stats.pm_intr_processed, 1);
845 counter_u64_add(pmc_stats.pm_intr_ignored, 1);
847 PMCDBG1(MDP,INT,2, "retval=%d", retval);
855 amd_describe(int cpu, int ri, struct pmc_info *pi, struct pmc **ppmc)
859 const struct amd_descr *pd;
862 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
863 ("[amd,%d] illegal CPU %d", __LINE__, cpu));
864 KASSERT(ri >= 0 && ri < AMD_NPMCS,
865 ("[amd,%d] row-index %d out of range", __LINE__, ri));
867 phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
868 pd = &amd_pmcdesc[ri];
870 if ((error = copystr(pd->pm_descr.pd_name, pi->pm_name,
871 PMC_NAME_MAX, &copied)) != 0)
874 pi->pm_class = pd->pm_descr.pd_class;
876 if (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) {
877 pi->pm_enabled = TRUE;
878 *ppmc = phw->phw_pmc;
880 pi->pm_enabled = FALSE;
888 * i386 specific entry points
892 * return the MSR address of the given PMC.
896 amd_get_msr(int ri, uint32_t *msr)
898 KASSERT(ri >= 0 && ri < AMD_NPMCS,
899 ("[amd,%d] ri %d out of range", __LINE__, ri));
901 *msr = amd_pmcdesc[ri].pm_perfctr - AMD_PMC_PERFCTR_0;
907 * processor dependent initialization.
911 amd_pcpu_init(struct pmc_mdep *md, int cpu)
913 int classindex, first_ri, n;
918 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
919 ("[amd,%d] insane cpu number %d", __LINE__, cpu));
921 PMCDBG1(MDP,INI,1,"amd-init cpu=%d", cpu);
923 amd_pcpu[cpu] = pac = malloc(sizeof(struct amd_cpu), M_PMC,
927 * Set the content of the hardware descriptors to a known
928 * state and initialize pointers in the MI per-cpu descriptor.
931 #if defined(__amd64__)
932 classindex = PMC_MDEP_CLASS_INDEX_K8;
933 #elif defined(__i386__)
934 classindex = md->pmd_cputype == PMC_CPU_AMD_K8 ?
935 PMC_MDEP_CLASS_INDEX_K8 : PMC_MDEP_CLASS_INDEX_K7;
937 first_ri = md->pmd_classdep[classindex].pcd_ri;
939 KASSERT(pc != NULL, ("[amd,%d] NULL per-cpu pointer", __LINE__));
941 for (n = 0, phw = pac->pc_amdpmcs; n < AMD_NPMCS; n++, phw++) {
942 phw->phw_state = PMC_PHW_FLAG_IS_ENABLED |
943 PMC_PHW_CPU_TO_STATE(cpu) | PMC_PHW_INDEX_TO_STATE(n);
945 pc->pc_hwpmcs[n + first_ri] = phw;
953 * processor dependent cleanup prior to the KLD
958 amd_pcpu_fini(struct pmc_mdep *md, int cpu)
960 int classindex, first_ri, i;
965 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
966 ("[amd,%d] insane cpu number (%d)", __LINE__, cpu));
968 PMCDBG1(MDP,INI,1,"amd-cleanup cpu=%d", cpu);
971 * First, turn off all PMCs on this CPU.
973 for (i = 0; i < 4; i++) { /* XXX this loop is now not needed */
974 evsel = rdmsr(AMD_PMC_EVSEL_0 + i);
975 evsel &= ~AMD_PMC_ENABLE;
976 wrmsr(AMD_PMC_EVSEL_0 + i, evsel);
980 * Next, free up allocated space.
982 if ((pac = amd_pcpu[cpu]) == NULL)
985 amd_pcpu[cpu] = NULL;
988 for (i = 0; i < AMD_NPMCS; i++) {
989 KASSERT(pac->pc_amdpmcs[i].phw_pmc == NULL,
990 ("[amd,%d] CPU%d/PMC%d in use", __LINE__, cpu, i));
991 KASSERT(AMD_PMC_IS_STOPPED(AMD_PMC_EVSEL_0 + i),
992 ("[amd,%d] CPU%d/PMC%d not stopped", __LINE__, cpu, i));
997 KASSERT(pc != NULL, ("[amd,%d] NULL per-cpu state", __LINE__));
999 #if defined(__amd64__)
1000 classindex = PMC_MDEP_CLASS_INDEX_K8;
1001 #elif defined(__i386__)
1002 classindex = md->pmd_cputype == PMC_CPU_AMD_K8 ? PMC_MDEP_CLASS_INDEX_K8 :
1003 PMC_MDEP_CLASS_INDEX_K7;
1005 first_ri = md->pmd_classdep[classindex].pcd_ri;
1008 * Reset pointers in the MI 'per-cpu' state.
1010 for (i = 0; i < AMD_NPMCS; i++) {
1011 pc->pc_hwpmcs[i + first_ri] = NULL;
1021 * Initialize ourselves.
1025 pmc_amd_initialize(void)
1027 int classindex, error, i, ncpus;
1028 struct pmc_classdep *pcd;
1029 enum pmc_cputype cputype;
1030 struct pmc_mdep *pmc_mdep;
1031 enum pmc_class class;
1036 * The presence of hardware performance counters on the AMD
1037 * Athlon, Duron or later processors, is _not_ indicated by
1038 * any of the processor feature flags set by the 'CPUID'
1039 * instruction, so we only check the 'instruction family'
1040 * field returned by CPUID for instruction family >= 6.
1044 model = ((cpu_id & 0xF0000) >> 12) | ((cpu_id & 0xF0) >> 4);
1045 if (CPUID_TO_FAMILY(cpu_id) == 0x17)
1046 snprintf(pmc_cpuid, sizeof(pmc_cpuid), "AuthenticAMD-%d-%02X",
1047 CPUID_TO_FAMILY(cpu_id), model);
1049 switch (cpu_id & 0xF00) {
1050 #if defined(__i386__)
1051 case 0x600: /* Athlon(tm) processor */
1052 classindex = PMC_MDEP_CLASS_INDEX_K7;
1053 cputype = PMC_CPU_AMD_K7;
1054 class = PMC_CLASS_K7;
1058 case 0xF00: /* Athlon64/Opteron processor */
1059 classindex = PMC_MDEP_CLASS_INDEX_K8;
1060 cputype = PMC_CPU_AMD_K8;
1061 class = PMC_CLASS_K8;
1066 (void) printf("pmc: Unknown AMD CPU %x %d-%d.\n", cpu_id, (cpu_id & 0xF00) >> 8, model);
1071 amd_pmc_class = class;
1075 * Allocate space for pointers to PMC HW descriptors and for
1076 * the MDEP structure used by MI code.
1078 amd_pcpu = malloc(sizeof(struct amd_cpu *) * pmc_cpu_max(), M_PMC,
1082 * These processors have two classes of PMCs: the TSC and
1083 * programmable PMCs.
1085 pmc_mdep = pmc_mdep_alloc(2);
1087 pmc_mdep->pmd_cputype = cputype;
1089 ncpus = pmc_cpu_max();
1091 /* Initialize the TSC. */
1092 error = pmc_tsc_initialize(pmc_mdep, ncpus);
1096 /* Initialize AMD K7 and K8 PMC handling. */
1097 pcd = &pmc_mdep->pmd_classdep[classindex];
1099 pcd->pcd_caps = AMD_PMC_CAPS;
1100 pcd->pcd_class = class;
1101 pcd->pcd_num = AMD_NPMCS;
1102 pcd->pcd_ri = pmc_mdep->pmd_npmc;
1103 pcd->pcd_width = 48;
1105 /* fill in the correct pmc name and class */
1106 for (i = 0; i < AMD_NPMCS; i++) {
1107 (void) snprintf(amd_pmcdesc[i].pm_descr.pd_name,
1108 sizeof(amd_pmcdesc[i].pm_descr.pd_name), "%s-%d",
1110 amd_pmcdesc[i].pm_descr.pd_class = class;
1113 pcd->pcd_allocate_pmc = amd_allocate_pmc;
1114 pcd->pcd_config_pmc = amd_config_pmc;
1115 pcd->pcd_describe = amd_describe;
1116 pcd->pcd_get_config = amd_get_config;
1117 pcd->pcd_get_msr = amd_get_msr;
1118 pcd->pcd_pcpu_fini = amd_pcpu_fini;
1119 pcd->pcd_pcpu_init = amd_pcpu_init;
1120 pcd->pcd_read_pmc = amd_read_pmc;
1121 pcd->pcd_release_pmc = amd_release_pmc;
1122 pcd->pcd_start_pmc = amd_start_pmc;
1123 pcd->pcd_stop_pmc = amd_stop_pmc;
1124 pcd->pcd_write_pmc = amd_write_pmc;
1126 pmc_mdep->pmd_pcpu_init = NULL;
1127 pmc_mdep->pmd_pcpu_fini = NULL;
1128 pmc_mdep->pmd_intr = amd_intr;
1129 pmc_mdep->pmd_switch_in = amd_switch_in;
1130 pmc_mdep->pmd_switch_out = amd_switch_out;
1132 pmc_mdep->pmd_npmc += AMD_NPMCS;
1134 PMCDBG0(MDP,INI,0,"amd-initialize");
1140 free(pmc_mdep, M_PMC);
1148 * Finalization code for AMD CPUs.
1152 pmc_amd_finalize(struct pmc_mdep *md)
1154 #if defined(INVARIANTS)
1155 int classindex, i, ncpus, pmcclass;
1158 pmc_tsc_finalize(md);
1160 KASSERT(amd_pcpu != NULL, ("[amd,%d] NULL per-cpu array pointer",
1163 #if defined(INVARIANTS)
1164 switch (md->pmd_cputype) {
1165 #if defined(__i386__)
1166 case PMC_CPU_AMD_K7:
1167 classindex = PMC_MDEP_CLASS_INDEX_K7;
1168 pmcclass = PMC_CLASS_K7;
1172 classindex = PMC_MDEP_CLASS_INDEX_K8;
1173 pmcclass = PMC_CLASS_K8;
1176 KASSERT(md->pmd_classdep[classindex].pcd_class == pmcclass,
1177 ("[amd,%d] pmc class mismatch", __LINE__));
1179 ncpus = pmc_cpu_max();
1181 for (i = 0; i < ncpus; i++)
1182 KASSERT(amd_pcpu[i] == NULL, ("[amd,%d] non-null pcpu",
1186 free(amd_pcpu, M_PMC);