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/pmc.h> 43 #include <sys/pmckern.h> 44 #include <sys/smp.h> 45 #include <sys/systm.h> 46 47 #include <machine/cpu.h> 48 #include <machine/cpufunc.h> 49 #include <machine/md_var.h> 50 #include <machine/specialreg.h> 51 52 #ifdef HWPMC_DEBUG 53 enum pmc_class amd_pmc_class; 54 #endif 55 56 /* AMD K7 & K8 PMCs */ 57 struct amd_descr { 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 */ 61 }; 62 63 static struct amd_descr amd_pmcdesc[AMD_NPMCS] = 64 { 65 { 66 .pm_descr = 67 { 68 .pd_name = "", 69 .pd_class = -1, 70 .pd_caps = AMD_PMC_CAPS, 71 .pd_width = 48 72 }, 73 .pm_evsel = AMD_PMC_EVSEL_0, 74 .pm_perfctr = AMD_PMC_PERFCTR_0 75 }, 76 { 77 .pm_descr = 78 { 79 .pd_name = "", 80 .pd_class = -1, 81 .pd_caps = AMD_PMC_CAPS, 82 .pd_width = 48 83 }, 84 .pm_evsel = AMD_PMC_EVSEL_1, 85 .pm_perfctr = AMD_PMC_PERFCTR_1 86 }, 87 { 88 .pm_descr = 89 { 90 .pd_name = "", 91 .pd_class = -1, 92 .pd_caps = AMD_PMC_CAPS, 93 .pd_width = 48 94 }, 95 .pm_evsel = AMD_PMC_EVSEL_2, 96 .pm_perfctr = AMD_PMC_PERFCTR_2 97 }, 98 { 99 .pm_descr = 100 { 101 .pd_name = "", 102 .pd_class = -1, 103 .pd_caps = AMD_PMC_CAPS, 104 .pd_width = 48 105 }, 106 .pm_evsel = AMD_PMC_EVSEL_3, 107 .pm_perfctr = AMD_PMC_PERFCTR_3 108 } 109 }; 110 111 struct amd_event_code_map { 112 enum pmc_event pe_ev; /* enum value */ 113 uint16_t pe_code; /* encoded event mask */ 114 uint8_t pe_mask; /* bits allowed in unit mask */ 115 }; 116 117 const struct amd_event_code_map amd_event_codes[] = { 118 #if defined(__i386__) /* 32 bit Athlon (K7) only */ 119 { PMC_EV_K7_DC_ACCESSES, 0x40, 0 }, 120 { PMC_EV_K7_DC_MISSES, 0x41, 0 }, 121 { PMC_EV_K7_DC_REFILLS_FROM_L2, 0x42, AMD_PMC_UNITMASK_MOESI }, 122 { PMC_EV_K7_DC_REFILLS_FROM_SYSTEM, 0x43, AMD_PMC_UNITMASK_MOESI }, 123 { PMC_EV_K7_DC_WRITEBACKS, 0x44, AMD_PMC_UNITMASK_MOESI }, 124 { PMC_EV_K7_L1_DTLB_MISS_AND_L2_DTLB_HITS, 0x45, 0 }, 125 { PMC_EV_K7_L1_AND_L2_DTLB_MISSES, 0x46, 0 }, 126 { PMC_EV_K7_MISALIGNED_REFERENCES, 0x47, 0 }, 127 128 { PMC_EV_K7_IC_FETCHES, 0x80, 0 }, 129 { PMC_EV_K7_IC_MISSES, 0x81, 0 }, 130 131 { PMC_EV_K7_L1_ITLB_MISSES, 0x84, 0 }, 132 { PMC_EV_K7_L1_L2_ITLB_MISSES, 0x85, 0 }, 133 134 { PMC_EV_K7_RETIRED_INSTRUCTIONS, 0xC0, 0 }, 135 { PMC_EV_K7_RETIRED_OPS, 0xC1, 0 }, 136 { PMC_EV_K7_RETIRED_BRANCHES, 0xC2, 0 }, 137 { PMC_EV_K7_RETIRED_BRANCHES_MISPREDICTED, 0xC3, 0 }, 138 { PMC_EV_K7_RETIRED_TAKEN_BRANCHES, 0xC4, 0 }, 139 { PMC_EV_K7_RETIRED_TAKEN_BRANCHES_MISPREDICTED, 0xC5, 0 }, 140 { PMC_EV_K7_RETIRED_FAR_CONTROL_TRANSFERS, 0xC6, 0 }, 141 { PMC_EV_K7_RETIRED_RESYNC_BRANCHES, 0xC7, 0 }, 142 { PMC_EV_K7_INTERRUPTS_MASKED_CYCLES, 0xCD, 0 }, 143 { PMC_EV_K7_INTERRUPTS_MASKED_WHILE_PENDING_CYCLES, 0xCE, 0 }, 144 { PMC_EV_K7_HARDWARE_INTERRUPTS, 0xCF, 0 }, 145 #endif 146 147 { PMC_EV_K8_FP_DISPATCHED_FPU_OPS, 0x00, 0x3F }, 148 { PMC_EV_K8_FP_CYCLES_WITH_NO_FPU_OPS_RETIRED, 0x01, 0x00 }, 149 { PMC_EV_K8_FP_DISPATCHED_FPU_FAST_FLAG_OPS, 0x02, 0x00 }, 150 151 { PMC_EV_K8_LS_SEGMENT_REGISTER_LOAD, 0x20, 0x7F }, 152 { PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SELF_MODIFYING_CODE, 153 0x21, 0x00 }, 154 { PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x22, 0x00 }, 155 { PMC_EV_K8_LS_BUFFER2_FULL, 0x23, 0x00 }, 156 { PMC_EV_K8_LS_LOCKED_OPERATION, 0x24, 0x07 }, 157 { PMC_EV_K8_LS_MICROARCHITECTURAL_LATE_CANCEL, 0x25, 0x00 }, 158 { PMC_EV_K8_LS_RETIRED_CFLUSH_INSTRUCTIONS, 0x26, 0x00 }, 159 { PMC_EV_K8_LS_RETIRED_CPUID_INSTRUCTIONS, 0x27, 0x00 }, 160 161 { PMC_EV_K8_DC_ACCESS, 0x40, 0x00 }, 162 { PMC_EV_K8_DC_MISS, 0x41, 0x00 }, 163 { PMC_EV_K8_DC_REFILL_FROM_L2, 0x42, 0x1F }, 164 { PMC_EV_K8_DC_REFILL_FROM_SYSTEM, 0x43, 0x1F }, 165 { PMC_EV_K8_DC_COPYBACK, 0x44, 0x1F }, 166 { PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_HIT, 0x45, 0x00 }, 167 { PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_MISS, 0x46, 0x00 }, 168 { PMC_EV_K8_DC_MISALIGNED_DATA_REFERENCE, 0x47, 0x00 }, 169 { PMC_EV_K8_DC_MICROARCHITECTURAL_LATE_CANCEL, 0x48, 0x00 }, 170 { PMC_EV_K8_DC_MICROARCHITECTURAL_EARLY_CANCEL, 0x49, 0x00 }, 171 { PMC_EV_K8_DC_ONE_BIT_ECC_ERROR, 0x4A, 0x03 }, 172 { PMC_EV_K8_DC_DISPATCHED_PREFETCH_INSTRUCTIONS, 0x4B, 0x07 }, 173 { PMC_EV_K8_DC_DCACHE_ACCESSES_BY_LOCKS, 0x4C, 0x03 }, 174 175 { PMC_EV_K8_BU_CPU_CLK_UNHALTED, 0x76, 0x00 }, 176 { PMC_EV_K8_BU_INTERNAL_L2_REQUEST, 0x7D, 0x1F }, 177 { PMC_EV_K8_BU_FILL_REQUEST_L2_MISS, 0x7E, 0x07 }, 178 { PMC_EV_K8_BU_FILL_INTO_L2, 0x7F, 0x03 }, 179 180 { PMC_EV_K8_IC_FETCH, 0x80, 0x00 }, 181 { PMC_EV_K8_IC_MISS, 0x81, 0x00 }, 182 { PMC_EV_K8_IC_REFILL_FROM_L2, 0x82, 0x00 }, 183 { PMC_EV_K8_IC_REFILL_FROM_SYSTEM, 0x83, 0x00 }, 184 { PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_HIT, 0x84, 0x00 }, 185 { PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_MISS, 0x85, 0x00 }, 186 { PMC_EV_K8_IC_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x86, 0x00 }, 187 { PMC_EV_K8_IC_INSTRUCTION_FETCH_STALL, 0x87, 0x00 }, 188 { PMC_EV_K8_IC_RETURN_STACK_HIT, 0x88, 0x00 }, 189 { PMC_EV_K8_IC_RETURN_STACK_OVERFLOW, 0x89, 0x00 }, 190 191 { PMC_EV_K8_FR_RETIRED_X86_INSTRUCTIONS, 0xC0, 0x00 }, 192 { PMC_EV_K8_FR_RETIRED_UOPS, 0xC1, 0x00 }, 193 { PMC_EV_K8_FR_RETIRED_BRANCHES, 0xC2, 0x00 }, 194 { PMC_EV_K8_FR_RETIRED_BRANCHES_MISPREDICTED, 0xC3, 0x00 }, 195 { PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES, 0xC4, 0x00 }, 196 { PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED, 0xC5, 0x00 }, 197 { PMC_EV_K8_FR_RETIRED_FAR_CONTROL_TRANSFERS, 0xC6, 0x00 }, 198 { PMC_EV_K8_FR_RETIRED_RESYNCS, 0xC7, 0x00 }, 199 { PMC_EV_K8_FR_RETIRED_NEAR_RETURNS, 0xC8, 0x00 }, 200 { PMC_EV_K8_FR_RETIRED_NEAR_RETURNS_MISPREDICTED, 0xC9, 0x00 }, 201 { PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED_BY_ADDR_MISCOMPARE, 202 0xCA, 0x00 }, 203 { PMC_EV_K8_FR_RETIRED_FPU_INSTRUCTIONS, 0xCB, 0x0F }, 204 { PMC_EV_K8_FR_RETIRED_FASTPATH_DOUBLE_OP_INSTRUCTIONS, 205 0xCC, 0x07 }, 206 { PMC_EV_K8_FR_INTERRUPTS_MASKED_CYCLES, 0xCD, 0x00 }, 207 { PMC_EV_K8_FR_INTERRUPTS_MASKED_WHILE_PENDING_CYCLES, 0xCE, 0x00 }, 208 { PMC_EV_K8_FR_TAKEN_HARDWARE_INTERRUPTS, 0xCF, 0x00 }, 209 210 { PMC_EV_K8_FR_DECODER_EMPTY, 0xD0, 0x00 }, 211 { PMC_EV_K8_FR_DISPATCH_STALLS, 0xD1, 0x00 }, 212 { PMC_EV_K8_FR_DISPATCH_STALL_FROM_BRANCH_ABORT_TO_RETIRE, 213 0xD2, 0x00 }, 214 { PMC_EV_K8_FR_DISPATCH_STALL_FOR_SERIALIZATION, 0xD3, 0x00 }, 215 { PMC_EV_K8_FR_DISPATCH_STALL_FOR_SEGMENT_LOAD, 0xD4, 0x00 }, 216 { PMC_EV_K8_FR_DISPATCH_STALL_WHEN_REORDER_BUFFER_IS_FULL, 217 0xD5, 0x00 }, 218 { PMC_EV_K8_FR_DISPATCH_STALL_WHEN_RESERVATION_STATIONS_ARE_FULL, 219 0xD6, 0x00 }, 220 { PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FPU_IS_FULL, 0xD7, 0x00 }, 221 { PMC_EV_K8_FR_DISPATCH_STALL_WHEN_LS_IS_FULL, 0xD8, 0x00 }, 222 { PMC_EV_K8_FR_DISPATCH_STALL_WHEN_WAITING_FOR_ALL_TO_BE_QUIET, 223 0xD9, 0x00 }, 224 { PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FAR_XFER_OR_RESYNC_BRANCH_PENDING, 225 0xDA, 0x00 }, 226 { PMC_EV_K8_FR_FPU_EXCEPTIONS, 0xDB, 0x0F }, 227 { PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR0, 0xDC, 0x00 }, 228 { PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR1, 0xDD, 0x00 }, 229 { PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR2, 0xDE, 0x00 }, 230 { PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR3, 0xDF, 0x00 }, 231 232 { PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_ACCESS_EVENT, 0xE0, 0x7 }, 233 { PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_TABLE_OVERFLOW, 0xE1, 0x00 }, 234 { PMC_EV_K8_NB_MEMORY_CONTROLLER_DRAM_COMMAND_SLOTS_MISSED, 235 0xE2, 0x00 }, 236 { PMC_EV_K8_NB_MEMORY_CONTROLLER_TURNAROUND, 0xE3, 0x07 }, 237 { PMC_EV_K8_NB_MEMORY_CONTROLLER_BYPASS_SATURATION, 0xE4, 0x0F }, 238 { PMC_EV_K8_NB_SIZED_COMMANDS, 0xEB, 0x7F }, 239 { PMC_EV_K8_NB_PROBE_RESULT, 0xEC, 0x0F }, 240 { PMC_EV_K8_NB_HT_BUS0_BANDWIDTH, 0xF6, 0x0F }, 241 { PMC_EV_K8_NB_HT_BUS1_BANDWIDTH, 0xF7, 0x0F }, 242 { PMC_EV_K8_NB_HT_BUS2_BANDWIDTH, 0xF8, 0x0F } 243 244 }; 245 246 const int amd_event_codes_size = nitems(amd_event_codes); 247 248 /* 249 * Per-processor information 250 */ 251 252 struct amd_cpu { 253 struct pmc_hw pc_amdpmcs[AMD_NPMCS]; 254 }; 255 256 static struct amd_cpu **amd_pcpu; 257 258 /* 259 * read a pmc register 260 */ 261 262 static int 263 amd_read_pmc(int cpu, int ri, pmc_value_t *v) 264 { 265 enum pmc_mode mode; 266 const struct amd_descr *pd; 267 struct pmc *pm; 268 pmc_value_t tmp; 269 270 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 271 ("[amd,%d] illegal CPU value %d", __LINE__, cpu)); 272 KASSERT(ri >= 0 && ri < AMD_NPMCS, 273 ("[amd,%d] illegal row-index %d", __LINE__, ri)); 274 KASSERT(amd_pcpu[cpu], 275 ("[amd,%d] null per-cpu, cpu %d", __LINE__, cpu)); 276 277 pm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc; 278 pd = &amd_pmcdesc[ri]; 279 280 KASSERT(pm != NULL, 281 ("[amd,%d] No owner for HWPMC [cpu%d,pmc%d]", __LINE__, 282 cpu, ri)); 283 284 mode = PMC_TO_MODE(pm); 285 286 PMCDBG2(MDP,REA,1,"amd-read id=%d class=%d", ri, pd->pm_descr.pd_class); 287 288 #ifdef HWPMC_DEBUG 289 KASSERT(pd->pm_descr.pd_class == amd_pmc_class, 290 ("[amd,%d] unknown PMC class (%d)", __LINE__, 291 pd->pm_descr.pd_class)); 292 #endif 293 294 tmp = rdmsr(pd->pm_perfctr); /* RDMSR serializes */ 295 PMCDBG2(MDP,REA,2,"amd-read (pre-munge) id=%d -> %jd", ri, tmp); 296 if (PMC_IS_SAMPLING_MODE(mode)) { 297 /* Sign extend 48 bit value to 64 bits. */ 298 tmp = (pmc_value_t) (((int64_t) tmp << 16) >> 16); 299 tmp = AMD_PERFCTR_VALUE_TO_RELOAD_COUNT(tmp); 300 } 301 *v = tmp; 302 303 PMCDBG2(MDP,REA,2,"amd-read (post-munge) id=%d -> %jd", ri, *v); 304 305 return 0; 306 } 307 308 /* 309 * Write a PMC MSR. 310 */ 311 312 static int 313 amd_write_pmc(int cpu, int ri, pmc_value_t v) 314 { 315 const struct amd_descr *pd; 316 enum pmc_mode mode; 317 struct pmc *pm; 318 319 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 320 ("[amd,%d] illegal CPU value %d", __LINE__, cpu)); 321 KASSERT(ri >= 0 && ri < AMD_NPMCS, 322 ("[amd,%d] illegal row-index %d", __LINE__, ri)); 323 324 pm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc; 325 pd = &amd_pmcdesc[ri]; 326 327 KASSERT(pm != NULL, 328 ("[amd,%d] PMC not owned (cpu%d,pmc%d)", __LINE__, 329 cpu, ri)); 330 331 mode = PMC_TO_MODE(pm); 332 333 #ifdef HWPMC_DEBUG 334 KASSERT(pd->pm_descr.pd_class == amd_pmc_class, 335 ("[amd,%d] unknown PMC class (%d)", __LINE__, 336 pd->pm_descr.pd_class)); 337 #endif 338 339 /* use 2's complement of the count for sampling mode PMCs */ 340 if (PMC_IS_SAMPLING_MODE(mode)) 341 v = AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v); 342 343 PMCDBG3(MDP,WRI,1,"amd-write cpu=%d ri=%d v=%jx", cpu, ri, v); 344 345 /* write the PMC value */ 346 wrmsr(pd->pm_perfctr, v); 347 return 0; 348 } 349 350 /* 351 * configure hardware pmc according to the configuration recorded in 352 * pmc 'pm'. 353 */ 354 355 static int 356 amd_config_pmc(int cpu, int ri, struct pmc *pm) 357 { 358 struct pmc_hw *phw; 359 360 PMCDBG3(MDP,CFG,1, "cpu=%d ri=%d pm=%p", cpu, ri, pm); 361 362 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 363 ("[amd,%d] illegal CPU value %d", __LINE__, cpu)); 364 KASSERT(ri >= 0 && ri < AMD_NPMCS, 365 ("[amd,%d] illegal row-index %d", __LINE__, ri)); 366 367 phw = &amd_pcpu[cpu]->pc_amdpmcs[ri]; 368 369 KASSERT(pm == NULL || phw->phw_pmc == NULL, 370 ("[amd,%d] pm=%p phw->pm=%p hwpmc not unconfigured", 371 __LINE__, pm, phw->phw_pmc)); 372 373 phw->phw_pmc = pm; 374 return 0; 375 } 376 377 /* 378 * Retrieve a configured PMC pointer from hardware state. 379 */ 380 381 static int 382 amd_get_config(int cpu, int ri, struct pmc **ppm) 383 { 384 *ppm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc; 385 386 return 0; 387 } 388 389 /* 390 * Machine dependent actions taken during the context switch in of a 391 * thread. 392 */ 393 394 static int 395 amd_switch_in(struct pmc_cpu *pc, struct pmc_process *pp) 396 { 397 (void) pc; 398 399 PMCDBG3(MDP,SWI,1, "pc=%p pp=%p enable-msr=%d", pc, pp, 400 (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0); 401 402 /* enable the RDPMC instruction if needed */ 403 if (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) 404 load_cr4(rcr4() | CR4_PCE); 405 406 return 0; 407 } 408 409 /* 410 * Machine dependent actions taken during the context switch out of a 411 * thread. 412 */ 413 414 static int 415 amd_switch_out(struct pmc_cpu *pc, struct pmc_process *pp) 416 { 417 (void) pc; 418 (void) pp; /* can be NULL */ 419 420 PMCDBG3(MDP,SWO,1, "pc=%p pp=%p enable-msr=%d", pc, pp, pp ? 421 (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) == 1 : 0); 422 423 /* always turn off the RDPMC instruction */ 424 load_cr4(rcr4() & ~CR4_PCE); 425 426 return 0; 427 } 428 429 /* 430 * Check if a given allocation is feasible. 431 */ 432 433 static int 434 amd_allocate_pmc(int cpu, int ri, struct pmc *pm, 435 const struct pmc_op_pmcallocate *a) 436 { 437 int i; 438 uint32_t allowed_unitmask, caps, config, unitmask; 439 enum pmc_event pe; 440 const struct pmc_descr *pd; 441 442 (void) cpu; 443 444 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 445 ("[amd,%d] illegal CPU value %d", __LINE__, cpu)); 446 KASSERT(ri >= 0 && ri < AMD_NPMCS, 447 ("[amd,%d] illegal row index %d", __LINE__, ri)); 448 449 pd = &amd_pmcdesc[ri].pm_descr; 450 451 /* check class match */ 452 if (pd->pd_class != a->pm_class) 453 return EINVAL; 454 455 caps = pm->pm_caps; 456 457 PMCDBG2(MDP,ALL,1,"amd-allocate ri=%d caps=0x%x", ri, caps); 458 459 if ((pd->pd_caps & caps) != caps) 460 return EPERM; 461 462 pe = a->pm_ev; 463 464 /* map ev to the correct event mask code */ 465 config = allowed_unitmask = 0; 466 for (i = 0; i < amd_event_codes_size; i++) 467 if (amd_event_codes[i].pe_ev == pe) { 468 config = 469 AMD_PMC_TO_EVENTMASK(amd_event_codes[i].pe_code); 470 allowed_unitmask = 471 AMD_PMC_TO_UNITMASK(amd_event_codes[i].pe_mask); 472 break; 473 } 474 if (i == amd_event_codes_size) 475 return EINVAL; 476 477 unitmask = a->pm_md.pm_amd.pm_amd_config & AMD_PMC_UNITMASK; 478 if (unitmask & ~allowed_unitmask) /* disallow reserved bits */ 479 return EINVAL; 480 481 if (unitmask && (caps & PMC_CAP_QUALIFIER)) 482 config |= unitmask; 483 484 if (caps & PMC_CAP_THRESHOLD) 485 config |= a->pm_md.pm_amd.pm_amd_config & AMD_PMC_COUNTERMASK; 486 487 /* set at least one of the 'usr' or 'os' caps */ 488 if (caps & PMC_CAP_USER) 489 config |= AMD_PMC_USR; 490 if (caps & PMC_CAP_SYSTEM) 491 config |= AMD_PMC_OS; 492 if ((caps & (PMC_CAP_USER|PMC_CAP_SYSTEM)) == 0) 493 config |= (AMD_PMC_USR|AMD_PMC_OS); 494 495 if (caps & PMC_CAP_EDGE) 496 config |= AMD_PMC_EDGE; 497 if (caps & PMC_CAP_INVERT) 498 config |= AMD_PMC_INVERT; 499 if (caps & PMC_CAP_INTERRUPT) 500 config |= AMD_PMC_INT; 501 502 pm->pm_md.pm_amd.pm_amd_evsel = config; /* save config value */ 503 504 PMCDBG2(MDP,ALL,2,"amd-allocate ri=%d -> config=0x%x", ri, config); 505 506 return 0; 507 } 508 509 /* 510 * Release machine dependent state associated with a PMC. This is a 511 * no-op on this architecture. 512 * 513 */ 514 515 /* ARGSUSED0 */ 516 static int 517 amd_release_pmc(int cpu, int ri, struct pmc *pmc) 518 { 519 #ifdef HWPMC_DEBUG 520 const struct amd_descr *pd; 521 #endif 522 struct pmc_hw *phw; 523 524 (void) pmc; 525 526 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 527 ("[amd,%d] illegal CPU value %d", __LINE__, cpu)); 528 KASSERT(ri >= 0 && ri < AMD_NPMCS, 529 ("[amd,%d] illegal row-index %d", __LINE__, ri)); 530 531 phw = &amd_pcpu[cpu]->pc_amdpmcs[ri]; 532 533 KASSERT(phw->phw_pmc == NULL, 534 ("[amd,%d] PHW pmc %p non-NULL", __LINE__, phw->phw_pmc)); 535 536 #ifdef HWPMC_DEBUG 537 pd = &amd_pmcdesc[ri]; 538 if (pd->pm_descr.pd_class == amd_pmc_class) 539 KASSERT(AMD_PMC_IS_STOPPED(pd->pm_evsel), 540 ("[amd,%d] PMC %d released while active", __LINE__, ri)); 541 #endif 542 543 return 0; 544 } 545 546 /* 547 * start a PMC. 548 */ 549 550 static int 551 amd_start_pmc(int cpu, int ri) 552 { 553 uint32_t config; 554 struct pmc *pm; 555 struct pmc_hw *phw; 556 const struct amd_descr *pd; 557 558 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 559 ("[amd,%d] illegal CPU value %d", __LINE__, cpu)); 560 KASSERT(ri >= 0 && ri < AMD_NPMCS, 561 ("[amd,%d] illegal row-index %d", __LINE__, ri)); 562 563 phw = &amd_pcpu[cpu]->pc_amdpmcs[ri]; 564 pm = phw->phw_pmc; 565 pd = &amd_pmcdesc[ri]; 566 567 KASSERT(pm != NULL, 568 ("[amd,%d] starting cpu%d,pmc%d with null pmc record", __LINE__, 569 cpu, ri)); 570 571 PMCDBG2(MDP,STA,1,"amd-start cpu=%d ri=%d", cpu, ri); 572 573 KASSERT(AMD_PMC_IS_STOPPED(pd->pm_evsel), 574 ("[amd,%d] pmc%d,cpu%d: Starting active PMC \"%s\"", __LINE__, 575 ri, cpu, pd->pm_descr.pd_name)); 576 577 /* turn on the PMC ENABLE bit */ 578 config = pm->pm_md.pm_amd.pm_amd_evsel | AMD_PMC_ENABLE; 579 580 PMCDBG1(MDP,STA,2,"amd-start config=0x%x", config); 581 582 wrmsr(pd->pm_evsel, config); 583 return 0; 584 } 585 586 /* 587 * Stop a PMC. 588 */ 589 590 static int 591 amd_stop_pmc(int cpu, int ri) 592 { 593 struct pmc *pm; 594 struct pmc_hw *phw; 595 const struct amd_descr *pd; 596 uint64_t config; 597 598 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 599 ("[amd,%d] illegal CPU value %d", __LINE__, cpu)); 600 KASSERT(ri >= 0 && ri < AMD_NPMCS, 601 ("[amd,%d] illegal row-index %d", __LINE__, ri)); 602 603 phw = &amd_pcpu[cpu]->pc_amdpmcs[ri]; 604 pm = phw->phw_pmc; 605 pd = &amd_pmcdesc[ri]; 606 607 KASSERT(pm != NULL, 608 ("[amd,%d] cpu%d,pmc%d no PMC to stop", __LINE__, 609 cpu, ri)); 610 KASSERT(!AMD_PMC_IS_STOPPED(pd->pm_evsel), 611 ("[amd,%d] PMC%d, CPU%d \"%s\" already stopped", 612 __LINE__, ri, cpu, pd->pm_descr.pd_name)); 613 614 PMCDBG1(MDP,STO,1,"amd-stop ri=%d", ri); 615 616 /* turn off the PMC ENABLE bit */ 617 config = pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE; 618 wrmsr(pd->pm_evsel, config); 619 return 0; 620 } 621 622 /* 623 * Interrupt handler. This function needs to return '1' if the 624 * interrupt was this CPU's PMCs or '0' otherwise. It is not allowed 625 * to sleep or do anything a 'fast' interrupt handler is not allowed 626 * to do. 627 */ 628 629 static int 630 amd_intr(int cpu, struct trapframe *tf) 631 { 632 int i, error, retval; 633 uint32_t config, evsel, perfctr; 634 struct pmc *pm; 635 struct amd_cpu *pac; 636 pmc_value_t v; 637 638 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 639 ("[amd,%d] out of range CPU %d", __LINE__, cpu)); 640 641 PMCDBG3(MDP,INT,1, "cpu=%d tf=%p um=%d", cpu, (void *) tf, 642 TRAPF_USERMODE(tf)); 643 644 retval = 0; 645 646 pac = amd_pcpu[cpu]; 647 648 /* 649 * look for all PMCs that have interrupted: 650 * - look for a running, sampling PMC which has overflowed 651 * and which has a valid 'struct pmc' association 652 * 653 * If found, we call a helper to process the interrupt. 654 * 655 * If multiple PMCs interrupt at the same time, the AMD64 656 * processor appears to deliver as many NMIs as there are 657 * outstanding PMC interrupts. So we process only one NMI 658 * interrupt at a time. 659 */ 660 661 for (i = 0; retval == 0 && i < AMD_NPMCS; i++) { 662 663 if ((pm = pac->pc_amdpmcs[i].phw_pmc) == NULL || 664 !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 665 continue; 666 } 667 668 if (!AMD_PMC_HAS_OVERFLOWED(i)) 669 continue; 670 671 retval = 1; /* Found an interrupting PMC. */ 672 673 if (pm->pm_state != PMC_STATE_RUNNING) 674 continue; 675 676 /* Stop the PMC, reload count. */ 677 evsel = AMD_PMC_EVSEL_0 + i; 678 perfctr = AMD_PMC_PERFCTR_0 + i; 679 v = pm->pm_sc.pm_reloadcount; 680 config = rdmsr(evsel); 681 682 KASSERT((config & ~AMD_PMC_ENABLE) == 683 (pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE), 684 ("[amd,%d] config mismatch reg=0x%x pm=0x%x", __LINE__, 685 config, pm->pm_md.pm_amd.pm_amd_evsel)); 686 687 wrmsr(evsel, config & ~AMD_PMC_ENABLE); 688 wrmsr(perfctr, AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v)); 689 690 /* Restart the counter if logging succeeded. */ 691 error = pmc_process_interrupt(cpu, PMC_HR, pm, tf, 692 TRAPF_USERMODE(tf)); 693 if (error == 0) 694 wrmsr(evsel, config); 695 } 696 697 atomic_add_int(retval ? &pmc_stats.pm_intr_processed : 698 &pmc_stats.pm_intr_ignored, 1); 699 700 PMCDBG1(MDP,INT,2, "retval=%d", retval); 701 return (retval); 702 } 703 704 /* 705 * describe a PMC 706 */ 707 static int 708 amd_describe(int cpu, int ri, struct pmc_info *pi, struct pmc **ppmc) 709 { 710 int error; 711 size_t copied; 712 const struct amd_descr *pd; 713 struct pmc_hw *phw; 714 715 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 716 ("[amd,%d] illegal CPU %d", __LINE__, cpu)); 717 KASSERT(ri >= 0 && ri < AMD_NPMCS, 718 ("[amd,%d] row-index %d out of range", __LINE__, ri)); 719 720 phw = &amd_pcpu[cpu]->pc_amdpmcs[ri]; 721 pd = &amd_pmcdesc[ri]; 722 723 if ((error = copystr(pd->pm_descr.pd_name, pi->pm_name, 724 PMC_NAME_MAX, &copied)) != 0) 725 return error; 726 727 pi->pm_class = pd->pm_descr.pd_class; 728 729 if (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) { 730 pi->pm_enabled = TRUE; 731 *ppmc = phw->phw_pmc; 732 } else { 733 pi->pm_enabled = FALSE; 734 *ppmc = NULL; 735 } 736 737 return 0; 738 } 739 740 /* 741 * i386 specific entry points 742 */ 743 744 /* 745 * return the MSR address of the given PMC. 746 */ 747 748 static int 749 amd_get_msr(int ri, uint32_t *msr) 750 { 751 KASSERT(ri >= 0 && ri < AMD_NPMCS, 752 ("[amd,%d] ri %d out of range", __LINE__, ri)); 753 754 *msr = amd_pmcdesc[ri].pm_perfctr - AMD_PMC_PERFCTR_0; 755 756 return (0); 757 } 758 759 /* 760 * processor dependent initialization. 761 */ 762 763 static int 764 amd_pcpu_init(struct pmc_mdep *md, int cpu) 765 { 766 int classindex, first_ri, n; 767 struct pmc_cpu *pc; 768 struct amd_cpu *pac; 769 struct pmc_hw *phw; 770 771 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 772 ("[amd,%d] insane cpu number %d", __LINE__, cpu)); 773 774 PMCDBG1(MDP,INI,1,"amd-init cpu=%d", cpu); 775 776 amd_pcpu[cpu] = pac = malloc(sizeof(struct amd_cpu), M_PMC, 777 M_WAITOK|M_ZERO); 778 779 /* 780 * Set the content of the hardware descriptors to a known 781 * state and initialize pointers in the MI per-cpu descriptor. 782 */ 783 pc = pmc_pcpu[cpu]; 784 #if defined(__amd64__) 785 classindex = PMC_MDEP_CLASS_INDEX_K8; 786 #elif defined(__i386__) 787 classindex = md->pmd_cputype == PMC_CPU_AMD_K8 ? 788 PMC_MDEP_CLASS_INDEX_K8 : PMC_MDEP_CLASS_INDEX_K7; 789 #endif 790 first_ri = md->pmd_classdep[classindex].pcd_ri; 791 792 KASSERT(pc != NULL, ("[amd,%d] NULL per-cpu pointer", __LINE__)); 793 794 for (n = 0, phw = pac->pc_amdpmcs; n < AMD_NPMCS; n++, phw++) { 795 phw->phw_state = PMC_PHW_FLAG_IS_ENABLED | 796 PMC_PHW_CPU_TO_STATE(cpu) | PMC_PHW_INDEX_TO_STATE(n); 797 phw->phw_pmc = NULL; 798 pc->pc_hwpmcs[n + first_ri] = phw; 799 } 800 801 return (0); 802 } 803 804 805 /* 806 * processor dependent cleanup prior to the KLD 807 * being unloaded 808 */ 809 810 static int 811 amd_pcpu_fini(struct pmc_mdep *md, int cpu) 812 { 813 int classindex, first_ri, i; 814 uint32_t evsel; 815 struct pmc_cpu *pc; 816 struct amd_cpu *pac; 817 818 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 819 ("[amd,%d] insane cpu number (%d)", __LINE__, cpu)); 820 821 PMCDBG1(MDP,INI,1,"amd-cleanup cpu=%d", cpu); 822 823 /* 824 * First, turn off all PMCs on this CPU. 825 */ 826 for (i = 0; i < 4; i++) { /* XXX this loop is now not needed */ 827 evsel = rdmsr(AMD_PMC_EVSEL_0 + i); 828 evsel &= ~AMD_PMC_ENABLE; 829 wrmsr(AMD_PMC_EVSEL_0 + i, evsel); 830 } 831 832 /* 833 * Next, free up allocated space. 834 */ 835 if ((pac = amd_pcpu[cpu]) == NULL) 836 return (0); 837 838 amd_pcpu[cpu] = NULL; 839 840 #ifdef HWPMC_DEBUG 841 for (i = 0; i < AMD_NPMCS; i++) { 842 KASSERT(pac->pc_amdpmcs[i].phw_pmc == NULL, 843 ("[amd,%d] CPU%d/PMC%d in use", __LINE__, cpu, i)); 844 KASSERT(AMD_PMC_IS_STOPPED(AMD_PMC_EVSEL_0 + i), 845 ("[amd,%d] CPU%d/PMC%d not stopped", __LINE__, cpu, i)); 846 } 847 #endif 848 849 pc = pmc_pcpu[cpu]; 850 KASSERT(pc != NULL, ("[amd,%d] NULL per-cpu state", __LINE__)); 851 852 #if defined(__amd64__) 853 classindex = PMC_MDEP_CLASS_INDEX_K8; 854 #elif defined(__i386__) 855 classindex = md->pmd_cputype == PMC_CPU_AMD_K8 ? PMC_MDEP_CLASS_INDEX_K8 : 856 PMC_MDEP_CLASS_INDEX_K7; 857 #endif 858 first_ri = md->pmd_classdep[classindex].pcd_ri; 859 860 /* 861 * Reset pointers in the MI 'per-cpu' state. 862 */ 863 for (i = 0; i < AMD_NPMCS; i++) { 864 pc->pc_hwpmcs[i + first_ri] = NULL; 865 } 866 867 868 free(pac, M_PMC); 869 870 return (0); 871 } 872 873 /* 874 * Initialize ourselves. 875 */ 876 877 struct pmc_mdep * 878 pmc_amd_initialize(void) 879 { 880 int classindex, error, i, ncpus; 881 struct pmc_classdep *pcd; 882 enum pmc_cputype cputype; 883 struct pmc_mdep *pmc_mdep; 884 enum pmc_class class; 885 char *name; 886 887 /* 888 * The presence of hardware performance counters on the AMD 889 * Athlon, Duron or later processors, is _not_ indicated by 890 * any of the processor feature flags set by the 'CPUID' 891 * instruction, so we only check the 'instruction family' 892 * field returned by CPUID for instruction family >= 6. 893 */ 894 895 name = NULL; 896 switch (cpu_id & 0xF00) { 897 #if defined(__i386__) 898 case 0x600: /* Athlon(tm) processor */ 899 classindex = PMC_MDEP_CLASS_INDEX_K7; 900 cputype = PMC_CPU_AMD_K7; 901 class = PMC_CLASS_K7; 902 name = "K7"; 903 break; 904 #endif 905 case 0xF00: /* Athlon64/Opteron processor */ 906 classindex = PMC_MDEP_CLASS_INDEX_K8; 907 cputype = PMC_CPU_AMD_K8; 908 class = PMC_CLASS_K8; 909 name = "K8"; 910 break; 911 912 default: 913 (void) printf("pmc: Unknown AMD CPU.\n"); 914 return NULL; 915 } 916 917 #ifdef HWPMC_DEBUG 918 amd_pmc_class = class; 919 #endif 920 921 /* 922 * Allocate space for pointers to PMC HW descriptors and for 923 * the MDEP structure used by MI code. 924 */ 925 amd_pcpu = malloc(sizeof(struct amd_cpu *) * pmc_cpu_max(), M_PMC, 926 M_WAITOK|M_ZERO); 927 928 /* 929 * These processors have two classes of PMCs: the TSC and 930 * programmable PMCs. 931 */ 932 pmc_mdep = pmc_mdep_alloc(2); 933 934 pmc_mdep->pmd_cputype = cputype; 935 936 ncpus = pmc_cpu_max(); 937 938 /* Initialize the TSC. */ 939 error = pmc_tsc_initialize(pmc_mdep, ncpus); 940 if (error) 941 goto error; 942 943 /* Initialize AMD K7 and K8 PMC handling. */ 944 pcd = &pmc_mdep->pmd_classdep[classindex]; 945 946 pcd->pcd_caps = AMD_PMC_CAPS; 947 pcd->pcd_class = class; 948 pcd->pcd_num = AMD_NPMCS; 949 pcd->pcd_ri = pmc_mdep->pmd_npmc; 950 pcd->pcd_width = 48; 951 952 /* fill in the correct pmc name and class */ 953 for (i = 0; i < AMD_NPMCS; i++) { 954 (void) snprintf(amd_pmcdesc[i].pm_descr.pd_name, 955 sizeof(amd_pmcdesc[i].pm_descr.pd_name), "%s-%d", 956 name, i); 957 amd_pmcdesc[i].pm_descr.pd_class = class; 958 } 959 960 pcd->pcd_allocate_pmc = amd_allocate_pmc; 961 pcd->pcd_config_pmc = amd_config_pmc; 962 pcd->pcd_describe = amd_describe; 963 pcd->pcd_get_config = amd_get_config; 964 pcd->pcd_get_msr = amd_get_msr; 965 pcd->pcd_pcpu_fini = amd_pcpu_fini; 966 pcd->pcd_pcpu_init = amd_pcpu_init; 967 pcd->pcd_read_pmc = amd_read_pmc; 968 pcd->pcd_release_pmc = amd_release_pmc; 969 pcd->pcd_start_pmc = amd_start_pmc; 970 pcd->pcd_stop_pmc = amd_stop_pmc; 971 pcd->pcd_write_pmc = amd_write_pmc; 972 973 pmc_mdep->pmd_pcpu_init = NULL; 974 pmc_mdep->pmd_pcpu_fini = NULL; 975 pmc_mdep->pmd_intr = amd_intr; 976 pmc_mdep->pmd_switch_in = amd_switch_in; 977 pmc_mdep->pmd_switch_out = amd_switch_out; 978 979 pmc_mdep->pmd_npmc += AMD_NPMCS; 980 981 PMCDBG0(MDP,INI,0,"amd-initialize"); 982 983 return (pmc_mdep); 984 985 error: 986 if (error) { 987 free(pmc_mdep, M_PMC); 988 pmc_mdep = NULL; 989 } 990 991 return (NULL); 992 } 993 994 /* 995 * Finalization code for AMD CPUs. 996 */ 997 998 void 999 pmc_amd_finalize(struct pmc_mdep *md) 1000 { 1001 #if defined(INVARIANTS) 1002 int classindex, i, ncpus, pmcclass; 1003 #endif 1004 1005 pmc_tsc_finalize(md); 1006 1007 KASSERT(amd_pcpu != NULL, ("[amd,%d] NULL per-cpu array pointer", 1008 __LINE__)); 1009 1010 #if defined(INVARIANTS) 1011 switch (md->pmd_cputype) { 1012 #if defined(__i386__) 1013 case PMC_CPU_AMD_K7: 1014 classindex = PMC_MDEP_CLASS_INDEX_K7; 1015 pmcclass = PMC_CLASS_K7; 1016 break; 1017 #endif 1018 default: 1019 classindex = PMC_MDEP_CLASS_INDEX_K8; 1020 pmcclass = PMC_CLASS_K8; 1021 } 1022 1023 KASSERT(md->pmd_classdep[classindex].pcd_class == pmcclass, 1024 ("[amd,%d] pmc class mismatch", __LINE__)); 1025 1026 ncpus = pmc_cpu_max(); 1027 1028 for (i = 0; i < ncpus; i++) 1029 KASSERT(amd_pcpu[i] == NULL, ("[amd,%d] non-null pcpu", 1030 __LINE__)); 1031 #endif 1032 1033 free(amd_pcpu, M_PMC); 1034 amd_pcpu = NULL; 1035 } 1036