1 /* 2 * Copyright 2012 by Oracle Inc 3 * Author: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> 4 * 5 * This code borrows ideas from https://lkml.org/lkml/2011/11/30/249 6 * so many thanks go to Kevin Tian <kevin.tian@intel.com> 7 * and Yu Ke <ke.yu@intel.com>. 8 * 9 * This program is free software; you can redistribute it and/or modify it 10 * under the terms and conditions of the GNU General Public License, 11 * version 2, as published by the Free Software Foundation. 12 * 13 * This program is distributed in the hope it will be useful, but WITHOUT 14 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 15 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 16 * more details. 17 * 18 */ 19 20 #include <linux/cpumask.h> 21 #include <linux/cpufreq.h> 22 #include <linux/freezer.h> 23 #include <linux/kernel.h> 24 #include <linux/kthread.h> 25 #include <linux/init.h> 26 #include <linux/module.h> 27 #include <linux/types.h> 28 #include <acpi/acpi_bus.h> 29 #include <acpi/acpi_drivers.h> 30 #include <acpi/processor.h> 31 32 #include <xen/interface/platform.h> 33 #include <asm/xen/hypercall.h> 34 35 #define DRV_NAME "xen-acpi-processor: " 36 37 static int no_hypercall; 38 MODULE_PARM_DESC(off, "Inhibit the hypercall."); 39 module_param_named(off, no_hypercall, int, 0400); 40 41 /* 42 * Note: Do not convert the acpi_id* below to cpumask_var_t or use cpumask_bit 43 * - as those shrink to nr_cpu_bits (which is dependent on possible_cpu), which 44 * can be less than what we want to put in. Instead use the 'nr_acpi_bits' 45 * which is dynamically computed based on the MADT or x2APIC table. 46 */ 47 static unsigned int nr_acpi_bits; 48 /* Mutex to protect the acpi_ids_done - for CPU hotplug use. */ 49 static DEFINE_MUTEX(acpi_ids_mutex); 50 /* Which ACPI ID we have processed from 'struct acpi_processor'. */ 51 static unsigned long *acpi_ids_done; 52 /* Which ACPI ID exist in the SSDT/DSDT processor definitions. */ 53 static unsigned long __initdata *acpi_id_present; 54 /* And if there is an _CST definition (or a PBLK) for the ACPI IDs */ 55 static unsigned long __initdata *acpi_id_cst_present; 56 57 static int push_cxx_to_hypervisor(struct acpi_processor *_pr) 58 { 59 struct xen_platform_op op = { 60 .cmd = XENPF_set_processor_pminfo, 61 .interface_version = XENPF_INTERFACE_VERSION, 62 .u.set_pminfo.id = _pr->acpi_id, 63 .u.set_pminfo.type = XEN_PM_CX, 64 }; 65 struct xen_processor_cx *dst_cx, *dst_cx_states = NULL; 66 struct acpi_processor_cx *cx; 67 unsigned int i, ok; 68 int ret = 0; 69 70 dst_cx_states = kcalloc(_pr->power.count, 71 sizeof(struct xen_processor_cx), GFP_KERNEL); 72 if (!dst_cx_states) 73 return -ENOMEM; 74 75 for (ok = 0, i = 1; i <= _pr->power.count; i++) { 76 cx = &_pr->power.states[i]; 77 if (!cx->valid) 78 continue; 79 80 dst_cx = &(dst_cx_states[ok++]); 81 82 dst_cx->reg.space_id = ACPI_ADR_SPACE_SYSTEM_IO; 83 if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) { 84 dst_cx->reg.bit_width = 8; 85 dst_cx->reg.bit_offset = 0; 86 dst_cx->reg.access_size = 1; 87 } else { 88 dst_cx->reg.space_id = ACPI_ADR_SPACE_FIXED_HARDWARE; 89 if (cx->entry_method == ACPI_CSTATE_FFH) { 90 /* NATIVE_CSTATE_BEYOND_HALT */ 91 dst_cx->reg.bit_offset = 2; 92 dst_cx->reg.bit_width = 1; /* VENDOR_INTEL */ 93 } 94 dst_cx->reg.access_size = 0; 95 } 96 dst_cx->reg.address = cx->address; 97 98 dst_cx->type = cx->type; 99 dst_cx->latency = cx->latency; 100 dst_cx->power = cx->power; 101 102 dst_cx->dpcnt = 0; 103 set_xen_guest_handle(dst_cx->dp, NULL); 104 } 105 if (!ok) { 106 pr_debug(DRV_NAME "No _Cx for ACPI CPU %u\n", _pr->acpi_id); 107 kfree(dst_cx_states); 108 return -EINVAL; 109 } 110 op.u.set_pminfo.power.count = ok; 111 op.u.set_pminfo.power.flags.bm_control = _pr->flags.bm_control; 112 op.u.set_pminfo.power.flags.bm_check = _pr->flags.bm_check; 113 op.u.set_pminfo.power.flags.has_cst = _pr->flags.has_cst; 114 op.u.set_pminfo.power.flags.power_setup_done = 115 _pr->flags.power_setup_done; 116 117 set_xen_guest_handle(op.u.set_pminfo.power.states, dst_cx_states); 118 119 if (!no_hypercall) 120 ret = HYPERVISOR_dom0_op(&op); 121 122 if (!ret) { 123 pr_debug("ACPI CPU%u - C-states uploaded.\n", _pr->acpi_id); 124 for (i = 1; i <= _pr->power.count; i++) { 125 cx = &_pr->power.states[i]; 126 if (!cx->valid) 127 continue; 128 pr_debug(" C%d: %s %d uS\n", 129 cx->type, cx->desc, (u32)cx->latency); 130 } 131 } else 132 pr_err(DRV_NAME "(CX): Hypervisor error (%d) for ACPI CPU%u\n", 133 ret, _pr->acpi_id); 134 135 kfree(dst_cx_states); 136 137 return ret; 138 } 139 static struct xen_processor_px * 140 xen_copy_pss_data(struct acpi_processor *_pr, 141 struct xen_processor_performance *dst_perf) 142 { 143 struct xen_processor_px *dst_states = NULL; 144 unsigned int i; 145 146 BUILD_BUG_ON(sizeof(struct xen_processor_px) != 147 sizeof(struct acpi_processor_px)); 148 149 dst_states = kcalloc(_pr->performance->state_count, 150 sizeof(struct xen_processor_px), GFP_KERNEL); 151 if (!dst_states) 152 return ERR_PTR(-ENOMEM); 153 154 dst_perf->state_count = _pr->performance->state_count; 155 for (i = 0; i < _pr->performance->state_count; i++) { 156 /* Fortunatly for us, they are both the same size */ 157 memcpy(&(dst_states[i]), &(_pr->performance->states[i]), 158 sizeof(struct acpi_processor_px)); 159 } 160 return dst_states; 161 } 162 static int xen_copy_psd_data(struct acpi_processor *_pr, 163 struct xen_processor_performance *dst) 164 { 165 struct acpi_psd_package *pdomain; 166 167 BUILD_BUG_ON(sizeof(struct xen_psd_package) != 168 sizeof(struct acpi_psd_package)); 169 170 /* This information is enumerated only if acpi_processor_preregister_performance 171 * has been called. 172 */ 173 dst->shared_type = _pr->performance->shared_type; 174 175 pdomain = &(_pr->performance->domain_info); 176 177 /* 'acpi_processor_preregister_performance' does not parse if the 178 * num_processors <= 1, but Xen still requires it. Do it manually here. 179 */ 180 if (pdomain->num_processors <= 1) { 181 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL) 182 dst->shared_type = CPUFREQ_SHARED_TYPE_ALL; 183 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL) 184 dst->shared_type = CPUFREQ_SHARED_TYPE_HW; 185 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY) 186 dst->shared_type = CPUFREQ_SHARED_TYPE_ANY; 187 188 } 189 memcpy(&(dst->domain_info), pdomain, sizeof(struct acpi_psd_package)); 190 return 0; 191 } 192 static int xen_copy_pct_data(struct acpi_pct_register *pct, 193 struct xen_pct_register *dst_pct) 194 { 195 /* It would be nice if you could just do 'memcpy(pct, dst_pct') but 196 * sadly the Xen structure did not have the proper padding so the 197 * descriptor field takes two (dst_pct) bytes instead of one (pct). 198 */ 199 dst_pct->descriptor = pct->descriptor; 200 dst_pct->length = pct->length; 201 dst_pct->space_id = pct->space_id; 202 dst_pct->bit_width = pct->bit_width; 203 dst_pct->bit_offset = pct->bit_offset; 204 dst_pct->reserved = pct->reserved; 205 dst_pct->address = pct->address; 206 return 0; 207 } 208 static int push_pxx_to_hypervisor(struct acpi_processor *_pr) 209 { 210 int ret = 0; 211 struct xen_platform_op op = { 212 .cmd = XENPF_set_processor_pminfo, 213 .interface_version = XENPF_INTERFACE_VERSION, 214 .u.set_pminfo.id = _pr->acpi_id, 215 .u.set_pminfo.type = XEN_PM_PX, 216 }; 217 struct xen_processor_performance *dst_perf; 218 struct xen_processor_px *dst_states = NULL; 219 220 dst_perf = &op.u.set_pminfo.perf; 221 222 dst_perf->platform_limit = _pr->performance_platform_limit; 223 dst_perf->flags |= XEN_PX_PPC; 224 xen_copy_pct_data(&(_pr->performance->control_register), 225 &dst_perf->control_register); 226 xen_copy_pct_data(&(_pr->performance->status_register), 227 &dst_perf->status_register); 228 dst_perf->flags |= XEN_PX_PCT; 229 dst_states = xen_copy_pss_data(_pr, dst_perf); 230 if (!IS_ERR_OR_NULL(dst_states)) { 231 set_xen_guest_handle(dst_perf->states, dst_states); 232 dst_perf->flags |= XEN_PX_PSS; 233 } 234 if (!xen_copy_psd_data(_pr, dst_perf)) 235 dst_perf->flags |= XEN_PX_PSD; 236 237 if (dst_perf->flags != (XEN_PX_PSD | XEN_PX_PSS | XEN_PX_PCT | XEN_PX_PPC)) { 238 pr_warn(DRV_NAME "ACPI CPU%u missing some P-state data (%x), skipping.\n", 239 _pr->acpi_id, dst_perf->flags); 240 ret = -ENODEV; 241 goto err_free; 242 } 243 244 if (!no_hypercall) 245 ret = HYPERVISOR_dom0_op(&op); 246 247 if (!ret) { 248 struct acpi_processor_performance *perf; 249 unsigned int i; 250 251 perf = _pr->performance; 252 pr_debug("ACPI CPU%u - P-states uploaded.\n", _pr->acpi_id); 253 for (i = 0; i < perf->state_count; i++) { 254 pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n", 255 (i == perf->state ? '*' : ' '), i, 256 (u32) perf->states[i].core_frequency, 257 (u32) perf->states[i].power, 258 (u32) perf->states[i].transition_latency); 259 } 260 } else if (ret != -EINVAL) 261 /* EINVAL means the ACPI ID is incorrect - meaning the ACPI 262 * table is referencing a non-existing CPU - which can happen 263 * with broken ACPI tables. */ 264 pr_warn(DRV_NAME "(_PXX): Hypervisor error (%d) for ACPI CPU%u\n", 265 ret, _pr->acpi_id); 266 err_free: 267 if (!IS_ERR_OR_NULL(dst_states)) 268 kfree(dst_states); 269 270 return ret; 271 } 272 static int upload_pm_data(struct acpi_processor *_pr) 273 { 274 int err = 0; 275 276 mutex_lock(&acpi_ids_mutex); 277 if (__test_and_set_bit(_pr->acpi_id, acpi_ids_done)) { 278 mutex_unlock(&acpi_ids_mutex); 279 return -EBUSY; 280 } 281 if (_pr->flags.power) 282 err = push_cxx_to_hypervisor(_pr); 283 284 if (_pr->performance && _pr->performance->states) 285 err |= push_pxx_to_hypervisor(_pr); 286 287 mutex_unlock(&acpi_ids_mutex); 288 return err; 289 } 290 static unsigned int __init get_max_acpi_id(void) 291 { 292 struct xenpf_pcpuinfo *info; 293 struct xen_platform_op op = { 294 .cmd = XENPF_get_cpuinfo, 295 .interface_version = XENPF_INTERFACE_VERSION, 296 }; 297 int ret = 0; 298 unsigned int i, last_cpu, max_acpi_id = 0; 299 300 info = &op.u.pcpu_info; 301 info->xen_cpuid = 0; 302 303 ret = HYPERVISOR_dom0_op(&op); 304 if (ret) 305 return NR_CPUS; 306 307 /* The max_present is the same irregardless of the xen_cpuid */ 308 last_cpu = op.u.pcpu_info.max_present; 309 for (i = 0; i <= last_cpu; i++) { 310 info->xen_cpuid = i; 311 ret = HYPERVISOR_dom0_op(&op); 312 if (ret) 313 continue; 314 max_acpi_id = max(info->acpi_id, max_acpi_id); 315 } 316 max_acpi_id *= 2; /* Slack for CPU hotplug support. */ 317 pr_debug(DRV_NAME "Max ACPI ID: %u\n", max_acpi_id); 318 return max_acpi_id; 319 } 320 /* 321 * The read_acpi_id and check_acpi_ids are there to support the Xen 322 * oddity of virtual CPUs != physical CPUs in the initial domain. 323 * The user can supply 'xen_max_vcpus=X' on the Xen hypervisor line 324 * which will band the amount of CPUs the initial domain can see. 325 * In general that is OK, except it plays havoc with any of the 326 * for_each_[present|online]_cpu macros which are banded to the virtual 327 * CPU amount. 328 */ 329 static acpi_status __init 330 read_acpi_id(acpi_handle handle, u32 lvl, void *context, void **rv) 331 { 332 u32 acpi_id; 333 acpi_status status; 334 acpi_object_type acpi_type; 335 unsigned long long tmp; 336 union acpi_object object = { 0 }; 337 struct acpi_buffer buffer = { sizeof(union acpi_object), &object }; 338 acpi_io_address pblk = 0; 339 340 status = acpi_get_type(handle, &acpi_type); 341 if (ACPI_FAILURE(status)) 342 return AE_OK; 343 344 switch (acpi_type) { 345 case ACPI_TYPE_PROCESSOR: 346 status = acpi_evaluate_object(handle, NULL, NULL, &buffer); 347 if (ACPI_FAILURE(status)) 348 return AE_OK; 349 acpi_id = object.processor.proc_id; 350 pblk = object.processor.pblk_address; 351 break; 352 case ACPI_TYPE_DEVICE: 353 status = acpi_evaluate_integer(handle, "_UID", NULL, &tmp); 354 if (ACPI_FAILURE(status)) 355 return AE_OK; 356 acpi_id = tmp; 357 break; 358 default: 359 return AE_OK; 360 } 361 /* There are more ACPI Processor objects than in x2APIC or MADT. 362 * This can happen with incorrect ACPI SSDT declerations. */ 363 if (acpi_id > nr_acpi_bits) { 364 pr_debug(DRV_NAME "We only have %u, trying to set %u\n", 365 nr_acpi_bits, acpi_id); 366 return AE_OK; 367 } 368 /* OK, There is a ACPI Processor object */ 369 __set_bit(acpi_id, acpi_id_present); 370 371 pr_debug(DRV_NAME "ACPI CPU%u w/ PBLK:0x%lx\n", acpi_id, 372 (unsigned long)pblk); 373 374 status = acpi_evaluate_object(handle, "_CST", NULL, &buffer); 375 if (ACPI_FAILURE(status)) { 376 if (!pblk) 377 return AE_OK; 378 } 379 /* .. and it has a C-state */ 380 __set_bit(acpi_id, acpi_id_cst_present); 381 382 return AE_OK; 383 } 384 static int __init check_acpi_ids(struct acpi_processor *pr_backup) 385 { 386 387 if (!pr_backup) 388 return -ENODEV; 389 390 /* All online CPUs have been processed at this stage. Now verify 391 * whether in fact "online CPUs" == physical CPUs. 392 */ 393 acpi_id_present = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL); 394 if (!acpi_id_present) 395 return -ENOMEM; 396 397 acpi_id_cst_present = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL); 398 if (!acpi_id_cst_present) { 399 kfree(acpi_id_present); 400 return -ENOMEM; 401 } 402 403 acpi_walk_namespace(ACPI_TYPE_PROCESSOR, ACPI_ROOT_OBJECT, 404 ACPI_UINT32_MAX, 405 read_acpi_id, NULL, NULL, NULL); 406 acpi_get_devices("ACPI0007", read_acpi_id, NULL, NULL); 407 408 if (!bitmap_equal(acpi_id_present, acpi_ids_done, nr_acpi_bits)) { 409 unsigned int i; 410 for_each_set_bit(i, acpi_id_present, nr_acpi_bits) { 411 pr_backup->acpi_id = i; 412 /* Mask out C-states if there are no _CST or PBLK */ 413 pr_backup->flags.power = test_bit(i, acpi_id_cst_present); 414 (void)upload_pm_data(pr_backup); 415 } 416 } 417 kfree(acpi_id_present); 418 acpi_id_present = NULL; 419 kfree(acpi_id_cst_present); 420 acpi_id_cst_present = NULL; 421 return 0; 422 } 423 static int __init check_prereq(void) 424 { 425 struct cpuinfo_x86 *c = &cpu_data(0); 426 427 if (!xen_initial_domain()) 428 return -ENODEV; 429 430 if (!acpi_gbl_FADT.smi_command) 431 return -ENODEV; 432 433 if (c->x86_vendor == X86_VENDOR_INTEL) { 434 if (!cpu_has(c, X86_FEATURE_EST)) 435 return -ENODEV; 436 437 return 0; 438 } 439 if (c->x86_vendor == X86_VENDOR_AMD) { 440 /* Copied from powernow-k8.h, can't include ../cpufreq/powernow 441 * as we get compile warnings for the static functions. 442 */ 443 #define CPUID_FREQ_VOLT_CAPABILITIES 0x80000007 444 #define USE_HW_PSTATE 0x00000080 445 u32 eax, ebx, ecx, edx; 446 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx); 447 if ((edx & USE_HW_PSTATE) != USE_HW_PSTATE) 448 return -ENODEV; 449 return 0; 450 } 451 return -ENODEV; 452 } 453 /* acpi_perf_data is a pointer to percpu data. */ 454 static struct acpi_processor_performance __percpu *acpi_perf_data; 455 456 static void free_acpi_perf_data(void) 457 { 458 unsigned int i; 459 460 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */ 461 for_each_possible_cpu(i) 462 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i) 463 ->shared_cpu_map); 464 free_percpu(acpi_perf_data); 465 } 466 467 static int __init xen_acpi_processor_init(void) 468 { 469 struct acpi_processor *pr_backup = NULL; 470 unsigned int i; 471 int rc = check_prereq(); 472 473 if (rc) 474 return rc; 475 476 nr_acpi_bits = get_max_acpi_id() + 1; 477 acpi_ids_done = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL); 478 if (!acpi_ids_done) 479 return -ENOMEM; 480 481 acpi_perf_data = alloc_percpu(struct acpi_processor_performance); 482 if (!acpi_perf_data) { 483 pr_debug(DRV_NAME "Memory allocation error for acpi_perf_data.\n"); 484 kfree(acpi_ids_done); 485 return -ENOMEM; 486 } 487 for_each_possible_cpu(i) { 488 if (!zalloc_cpumask_var_node( 489 &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map, 490 GFP_KERNEL, cpu_to_node(i))) { 491 rc = -ENOMEM; 492 goto err_out; 493 } 494 } 495 496 /* Do initialization in ACPI core. It is OK to fail here. */ 497 (void)acpi_processor_preregister_performance(acpi_perf_data); 498 499 for_each_possible_cpu(i) { 500 struct acpi_processor_performance *perf; 501 502 perf = per_cpu_ptr(acpi_perf_data, i); 503 rc = acpi_processor_register_performance(perf, i); 504 if (WARN_ON(rc)) 505 goto err_out; 506 } 507 rc = acpi_processor_notify_smm(THIS_MODULE); 508 if (WARN_ON(rc)) 509 goto err_unregister; 510 511 for_each_possible_cpu(i) { 512 struct acpi_processor *_pr; 513 _pr = per_cpu(processors, i /* APIC ID */); 514 if (!_pr) 515 continue; 516 517 if (!pr_backup) { 518 pr_backup = kzalloc(sizeof(struct acpi_processor), GFP_KERNEL); 519 memcpy(pr_backup, _pr, sizeof(struct acpi_processor)); 520 } 521 (void)upload_pm_data(_pr); 522 } 523 rc = check_acpi_ids(pr_backup); 524 if (rc) 525 goto err_unregister; 526 527 kfree(pr_backup); 528 529 return 0; 530 err_unregister: 531 for_each_possible_cpu(i) { 532 struct acpi_processor_performance *perf; 533 perf = per_cpu_ptr(acpi_perf_data, i); 534 acpi_processor_unregister_performance(perf, i); 535 } 536 err_out: 537 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */ 538 free_acpi_perf_data(); 539 kfree(acpi_ids_done); 540 return rc; 541 } 542 static void __exit xen_acpi_processor_exit(void) 543 { 544 int i; 545 546 kfree(acpi_ids_done); 547 for_each_possible_cpu(i) { 548 struct acpi_processor_performance *perf; 549 perf = per_cpu_ptr(acpi_perf_data, i); 550 acpi_processor_unregister_performance(perf, i); 551 } 552 free_acpi_perf_data(); 553 } 554 555 MODULE_AUTHOR("Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>"); 556 MODULE_DESCRIPTION("Xen ACPI Processor P-states (and Cx) driver which uploads PM data to Xen hypervisor"); 557 MODULE_LICENSE("GPL"); 558 559 /* We want to be loaded before the CPU freq scaling drivers are loaded. 560 * They are loaded in late_initcall. */ 561 device_initcall(xen_acpi_processor_init); 562 module_exit(xen_acpi_processor_exit); 563