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/xen.h> 33 #include <xen/interface/platform.h> 34 #include <asm/xen/hypercall.h> 35 36 #define DRV_NAME "xen-acpi-processor: " 37 38 static int no_hypercall; 39 MODULE_PARM_DESC(off, "Inhibit the hypercall."); 40 module_param_named(off, no_hypercall, int, 0400); 41 42 /* 43 * Note: Do not convert the acpi_id* below to cpumask_var_t or use cpumask_bit 44 * - as those shrink to nr_cpu_bits (which is dependent on possible_cpu), which 45 * can be less than what we want to put in. Instead use the 'nr_acpi_bits' 46 * which is dynamically computed based on the MADT or x2APIC table. 47 */ 48 static unsigned int nr_acpi_bits; 49 /* Mutex to protect the acpi_ids_done - for CPU hotplug use. */ 50 static DEFINE_MUTEX(acpi_ids_mutex); 51 /* Which ACPI ID we have processed from 'struct acpi_processor'. */ 52 static unsigned long *acpi_ids_done; 53 /* Which ACPI ID exist in the SSDT/DSDT processor definitions. */ 54 static unsigned long __initdata *acpi_id_present; 55 /* And if there is an _CST definition (or a PBLK) for the ACPI IDs */ 56 static unsigned long __initdata *acpi_id_cst_present; 57 58 static int push_cxx_to_hypervisor(struct acpi_processor *_pr) 59 { 60 struct xen_platform_op op = { 61 .cmd = XENPF_set_processor_pminfo, 62 .interface_version = XENPF_INTERFACE_VERSION, 63 .u.set_pminfo.id = _pr->acpi_id, 64 .u.set_pminfo.type = XEN_PM_CX, 65 }; 66 struct xen_processor_cx *dst_cx, *dst_cx_states = NULL; 67 struct acpi_processor_cx *cx; 68 unsigned int i, ok; 69 int ret = 0; 70 71 dst_cx_states = kcalloc(_pr->power.count, 72 sizeof(struct xen_processor_cx), GFP_KERNEL); 73 if (!dst_cx_states) 74 return -ENOMEM; 75 76 for (ok = 0, i = 1; i <= _pr->power.count; i++) { 77 cx = &_pr->power.states[i]; 78 if (!cx->valid) 79 continue; 80 81 dst_cx = &(dst_cx_states[ok++]); 82 83 dst_cx->reg.space_id = ACPI_ADR_SPACE_SYSTEM_IO; 84 if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) { 85 dst_cx->reg.bit_width = 8; 86 dst_cx->reg.bit_offset = 0; 87 dst_cx->reg.access_size = 1; 88 } else { 89 dst_cx->reg.space_id = ACPI_ADR_SPACE_FIXED_HARDWARE; 90 if (cx->entry_method == ACPI_CSTATE_FFH) { 91 /* NATIVE_CSTATE_BEYOND_HALT */ 92 dst_cx->reg.bit_offset = 2; 93 dst_cx->reg.bit_width = 1; /* VENDOR_INTEL */ 94 } 95 dst_cx->reg.access_size = 0; 96 } 97 dst_cx->reg.address = cx->address; 98 99 dst_cx->type = cx->type; 100 dst_cx->latency = cx->latency; 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 if (ret != -EINVAL) 132 /* EINVAL means the ACPI ID is incorrect - meaning the ACPI 133 * table is referencing a non-existing CPU - which can happen 134 * with broken ACPI tables. */ 135 pr_err(DRV_NAME "(CX): Hypervisor error (%d) for ACPI CPU%u\n", 136 ret, _pr->acpi_id); 137 138 kfree(dst_cx_states); 139 140 return ret; 141 } 142 static struct xen_processor_px * 143 xen_copy_pss_data(struct acpi_processor *_pr, 144 struct xen_processor_performance *dst_perf) 145 { 146 struct xen_processor_px *dst_states = NULL; 147 unsigned int i; 148 149 BUILD_BUG_ON(sizeof(struct xen_processor_px) != 150 sizeof(struct acpi_processor_px)); 151 152 dst_states = kcalloc(_pr->performance->state_count, 153 sizeof(struct xen_processor_px), GFP_KERNEL); 154 if (!dst_states) 155 return ERR_PTR(-ENOMEM); 156 157 dst_perf->state_count = _pr->performance->state_count; 158 for (i = 0; i < _pr->performance->state_count; i++) { 159 /* Fortunatly for us, they are both the same size */ 160 memcpy(&(dst_states[i]), &(_pr->performance->states[i]), 161 sizeof(struct acpi_processor_px)); 162 } 163 return dst_states; 164 } 165 static int xen_copy_psd_data(struct acpi_processor *_pr, 166 struct xen_processor_performance *dst) 167 { 168 struct acpi_psd_package *pdomain; 169 170 BUILD_BUG_ON(sizeof(struct xen_psd_package) != 171 sizeof(struct acpi_psd_package)); 172 173 /* This information is enumerated only if acpi_processor_preregister_performance 174 * has been called. 175 */ 176 dst->shared_type = _pr->performance->shared_type; 177 178 pdomain = &(_pr->performance->domain_info); 179 180 /* 'acpi_processor_preregister_performance' does not parse if the 181 * num_processors <= 1, but Xen still requires it. Do it manually here. 182 */ 183 if (pdomain->num_processors <= 1) { 184 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL) 185 dst->shared_type = CPUFREQ_SHARED_TYPE_ALL; 186 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL) 187 dst->shared_type = CPUFREQ_SHARED_TYPE_HW; 188 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY) 189 dst->shared_type = CPUFREQ_SHARED_TYPE_ANY; 190 191 } 192 memcpy(&(dst->domain_info), pdomain, sizeof(struct acpi_psd_package)); 193 return 0; 194 } 195 static int xen_copy_pct_data(struct acpi_pct_register *pct, 196 struct xen_pct_register *dst_pct) 197 { 198 /* It would be nice if you could just do 'memcpy(pct, dst_pct') but 199 * sadly the Xen structure did not have the proper padding so the 200 * descriptor field takes two (dst_pct) bytes instead of one (pct). 201 */ 202 dst_pct->descriptor = pct->descriptor; 203 dst_pct->length = pct->length; 204 dst_pct->space_id = pct->space_id; 205 dst_pct->bit_width = pct->bit_width; 206 dst_pct->bit_offset = pct->bit_offset; 207 dst_pct->reserved = pct->reserved; 208 dst_pct->address = pct->address; 209 return 0; 210 } 211 static int push_pxx_to_hypervisor(struct acpi_processor *_pr) 212 { 213 int ret = 0; 214 struct xen_platform_op op = { 215 .cmd = XENPF_set_processor_pminfo, 216 .interface_version = XENPF_INTERFACE_VERSION, 217 .u.set_pminfo.id = _pr->acpi_id, 218 .u.set_pminfo.type = XEN_PM_PX, 219 }; 220 struct xen_processor_performance *dst_perf; 221 struct xen_processor_px *dst_states = NULL; 222 223 dst_perf = &op.u.set_pminfo.perf; 224 225 dst_perf->platform_limit = _pr->performance_platform_limit; 226 dst_perf->flags |= XEN_PX_PPC; 227 xen_copy_pct_data(&(_pr->performance->control_register), 228 &dst_perf->control_register); 229 xen_copy_pct_data(&(_pr->performance->status_register), 230 &dst_perf->status_register); 231 dst_perf->flags |= XEN_PX_PCT; 232 dst_states = xen_copy_pss_data(_pr, dst_perf); 233 if (!IS_ERR_OR_NULL(dst_states)) { 234 set_xen_guest_handle(dst_perf->states, dst_states); 235 dst_perf->flags |= XEN_PX_PSS; 236 } 237 if (!xen_copy_psd_data(_pr, dst_perf)) 238 dst_perf->flags |= XEN_PX_PSD; 239 240 if (dst_perf->flags != (XEN_PX_PSD | XEN_PX_PSS | XEN_PX_PCT | XEN_PX_PPC)) { 241 pr_warn(DRV_NAME "ACPI CPU%u missing some P-state data (%x), skipping.\n", 242 _pr->acpi_id, dst_perf->flags); 243 ret = -ENODEV; 244 goto err_free; 245 } 246 247 if (!no_hypercall) 248 ret = HYPERVISOR_dom0_op(&op); 249 250 if (!ret) { 251 struct acpi_processor_performance *perf; 252 unsigned int i; 253 254 perf = _pr->performance; 255 pr_debug("ACPI CPU%u - P-states uploaded.\n", _pr->acpi_id); 256 for (i = 0; i < perf->state_count; i++) { 257 pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n", 258 (i == perf->state ? '*' : ' '), i, 259 (u32) perf->states[i].core_frequency, 260 (u32) perf->states[i].power, 261 (u32) perf->states[i].transition_latency); 262 } 263 } else if (ret != -EINVAL) 264 /* EINVAL means the ACPI ID is incorrect - meaning the ACPI 265 * table is referencing a non-existing CPU - which can happen 266 * with broken ACPI tables. */ 267 pr_warn(DRV_NAME "(_PXX): Hypervisor error (%d) for ACPI CPU%u\n", 268 ret, _pr->acpi_id); 269 err_free: 270 if (!IS_ERR_OR_NULL(dst_states)) 271 kfree(dst_states); 272 273 return ret; 274 } 275 static int upload_pm_data(struct acpi_processor *_pr) 276 { 277 int err = 0; 278 279 mutex_lock(&acpi_ids_mutex); 280 if (__test_and_set_bit(_pr->acpi_id, acpi_ids_done)) { 281 mutex_unlock(&acpi_ids_mutex); 282 return -EBUSY; 283 } 284 if (_pr->flags.power) 285 err = push_cxx_to_hypervisor(_pr); 286 287 if (_pr->performance && _pr->performance->states) 288 err |= push_pxx_to_hypervisor(_pr); 289 290 mutex_unlock(&acpi_ids_mutex); 291 return err; 292 } 293 static unsigned int __init get_max_acpi_id(void) 294 { 295 struct xenpf_pcpuinfo *info; 296 struct xen_platform_op op = { 297 .cmd = XENPF_get_cpuinfo, 298 .interface_version = XENPF_INTERFACE_VERSION, 299 }; 300 int ret = 0; 301 unsigned int i, last_cpu, max_acpi_id = 0; 302 303 info = &op.u.pcpu_info; 304 info->xen_cpuid = 0; 305 306 ret = HYPERVISOR_dom0_op(&op); 307 if (ret) 308 return NR_CPUS; 309 310 /* The max_present is the same irregardless of the xen_cpuid */ 311 last_cpu = op.u.pcpu_info.max_present; 312 for (i = 0; i <= last_cpu; i++) { 313 info->xen_cpuid = i; 314 ret = HYPERVISOR_dom0_op(&op); 315 if (ret) 316 continue; 317 max_acpi_id = max(info->acpi_id, max_acpi_id); 318 } 319 max_acpi_id *= 2; /* Slack for CPU hotplug support. */ 320 pr_debug(DRV_NAME "Max ACPI ID: %u\n", max_acpi_id); 321 return max_acpi_id; 322 } 323 /* 324 * The read_acpi_id and check_acpi_ids are there to support the Xen 325 * oddity of virtual CPUs != physical CPUs in the initial domain. 326 * The user can supply 'xen_max_vcpus=X' on the Xen hypervisor line 327 * which will band the amount of CPUs the initial domain can see. 328 * In general that is OK, except it plays havoc with any of the 329 * for_each_[present|online]_cpu macros which are banded to the virtual 330 * CPU amount. 331 */ 332 static acpi_status __init 333 read_acpi_id(acpi_handle handle, u32 lvl, void *context, void **rv) 334 { 335 u32 acpi_id; 336 acpi_status status; 337 acpi_object_type acpi_type; 338 unsigned long long tmp; 339 union acpi_object object = { 0 }; 340 struct acpi_buffer buffer = { sizeof(union acpi_object), &object }; 341 acpi_io_address pblk = 0; 342 343 status = acpi_get_type(handle, &acpi_type); 344 if (ACPI_FAILURE(status)) 345 return AE_OK; 346 347 switch (acpi_type) { 348 case ACPI_TYPE_PROCESSOR: 349 status = acpi_evaluate_object(handle, NULL, NULL, &buffer); 350 if (ACPI_FAILURE(status)) 351 return AE_OK; 352 acpi_id = object.processor.proc_id; 353 pblk = object.processor.pblk_address; 354 break; 355 case ACPI_TYPE_DEVICE: 356 status = acpi_evaluate_integer(handle, "_UID", NULL, &tmp); 357 if (ACPI_FAILURE(status)) 358 return AE_OK; 359 acpi_id = tmp; 360 break; 361 default: 362 return AE_OK; 363 } 364 /* There are more ACPI Processor objects than in x2APIC or MADT. 365 * This can happen with incorrect ACPI SSDT declerations. */ 366 if (acpi_id > nr_acpi_bits) { 367 pr_debug(DRV_NAME "We only have %u, trying to set %u\n", 368 nr_acpi_bits, acpi_id); 369 return AE_OK; 370 } 371 /* OK, There is a ACPI Processor object */ 372 __set_bit(acpi_id, acpi_id_present); 373 374 pr_debug(DRV_NAME "ACPI CPU%u w/ PBLK:0x%lx\n", acpi_id, 375 (unsigned long)pblk); 376 377 status = acpi_evaluate_object(handle, "_CST", NULL, &buffer); 378 if (ACPI_FAILURE(status)) { 379 if (!pblk) 380 return AE_OK; 381 } 382 /* .. and it has a C-state */ 383 __set_bit(acpi_id, acpi_id_cst_present); 384 385 return AE_OK; 386 } 387 static int __init check_acpi_ids(struct acpi_processor *pr_backup) 388 { 389 390 if (!pr_backup) 391 return -ENODEV; 392 393 /* All online CPUs have been processed at this stage. Now verify 394 * whether in fact "online CPUs" == physical CPUs. 395 */ 396 acpi_id_present = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL); 397 if (!acpi_id_present) 398 return -ENOMEM; 399 400 acpi_id_cst_present = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL); 401 if (!acpi_id_cst_present) { 402 kfree(acpi_id_present); 403 return -ENOMEM; 404 } 405 406 acpi_walk_namespace(ACPI_TYPE_PROCESSOR, ACPI_ROOT_OBJECT, 407 ACPI_UINT32_MAX, 408 read_acpi_id, NULL, NULL, NULL); 409 acpi_get_devices("ACPI0007", read_acpi_id, NULL, NULL); 410 411 if (!bitmap_equal(acpi_id_present, acpi_ids_done, nr_acpi_bits)) { 412 unsigned int i; 413 for_each_set_bit(i, acpi_id_present, nr_acpi_bits) { 414 pr_backup->acpi_id = i; 415 /* Mask out C-states if there are no _CST or PBLK */ 416 pr_backup->flags.power = test_bit(i, acpi_id_cst_present); 417 (void)upload_pm_data(pr_backup); 418 } 419 } 420 kfree(acpi_id_present); 421 acpi_id_present = NULL; 422 kfree(acpi_id_cst_present); 423 acpi_id_cst_present = NULL; 424 return 0; 425 } 426 static int __init check_prereq(void) 427 { 428 struct cpuinfo_x86 *c = &cpu_data(0); 429 430 if (!xen_initial_domain()) 431 return -ENODEV; 432 433 if (!acpi_gbl_FADT.smi_command) 434 return -ENODEV; 435 436 if (c->x86_vendor == X86_VENDOR_INTEL) { 437 if (!cpu_has(c, X86_FEATURE_EST)) 438 return -ENODEV; 439 440 return 0; 441 } 442 if (c->x86_vendor == X86_VENDOR_AMD) { 443 /* Copied from powernow-k8.h, can't include ../cpufreq/powernow 444 * as we get compile warnings for the static functions. 445 */ 446 #define CPUID_FREQ_VOLT_CAPABILITIES 0x80000007 447 #define USE_HW_PSTATE 0x00000080 448 u32 eax, ebx, ecx, edx; 449 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx); 450 if ((edx & USE_HW_PSTATE) != USE_HW_PSTATE) 451 return -ENODEV; 452 return 0; 453 } 454 return -ENODEV; 455 } 456 /* acpi_perf_data is a pointer to percpu data. */ 457 static struct acpi_processor_performance __percpu *acpi_perf_data; 458 459 static void free_acpi_perf_data(void) 460 { 461 unsigned int i; 462 463 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */ 464 for_each_possible_cpu(i) 465 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i) 466 ->shared_cpu_map); 467 free_percpu(acpi_perf_data); 468 } 469 470 static int __init xen_acpi_processor_init(void) 471 { 472 struct acpi_processor *pr_backup = NULL; 473 unsigned int i; 474 int rc = check_prereq(); 475 476 if (rc) 477 return rc; 478 479 nr_acpi_bits = get_max_acpi_id() + 1; 480 acpi_ids_done = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL); 481 if (!acpi_ids_done) 482 return -ENOMEM; 483 484 acpi_perf_data = alloc_percpu(struct acpi_processor_performance); 485 if (!acpi_perf_data) { 486 pr_debug(DRV_NAME "Memory allocation error for acpi_perf_data.\n"); 487 kfree(acpi_ids_done); 488 return -ENOMEM; 489 } 490 for_each_possible_cpu(i) { 491 if (!zalloc_cpumask_var_node( 492 &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map, 493 GFP_KERNEL, cpu_to_node(i))) { 494 rc = -ENOMEM; 495 goto err_out; 496 } 497 } 498 499 /* Do initialization in ACPI core. It is OK to fail here. */ 500 (void)acpi_processor_preregister_performance(acpi_perf_data); 501 502 for_each_possible_cpu(i) { 503 struct acpi_processor_performance *perf; 504 505 perf = per_cpu_ptr(acpi_perf_data, i); 506 rc = acpi_processor_register_performance(perf, i); 507 if (rc) 508 goto err_out; 509 } 510 rc = acpi_processor_notify_smm(THIS_MODULE); 511 if (rc) 512 goto err_unregister; 513 514 for_each_possible_cpu(i) { 515 struct acpi_processor *_pr; 516 _pr = per_cpu(processors, i /* APIC ID */); 517 if (!_pr) 518 continue; 519 520 if (!pr_backup) { 521 pr_backup = kzalloc(sizeof(struct acpi_processor), GFP_KERNEL); 522 if (pr_backup) 523 memcpy(pr_backup, _pr, sizeof(struct acpi_processor)); 524 } 525 (void)upload_pm_data(_pr); 526 } 527 rc = check_acpi_ids(pr_backup); 528 529 kfree(pr_backup); 530 pr_backup = NULL; 531 532 if (rc) 533 goto err_unregister; 534 535 return 0; 536 err_unregister: 537 for_each_possible_cpu(i) { 538 struct acpi_processor_performance *perf; 539 perf = per_cpu_ptr(acpi_perf_data, i); 540 acpi_processor_unregister_performance(perf, i); 541 } 542 err_out: 543 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */ 544 free_acpi_perf_data(); 545 kfree(acpi_ids_done); 546 return rc; 547 } 548 static void __exit xen_acpi_processor_exit(void) 549 { 550 int i; 551 552 kfree(acpi_ids_done); 553 for_each_possible_cpu(i) { 554 struct acpi_processor_performance *perf; 555 perf = per_cpu_ptr(acpi_perf_data, i); 556 acpi_processor_unregister_performance(perf, i); 557 } 558 free_acpi_perf_data(); 559 } 560 561 MODULE_AUTHOR("Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>"); 562 MODULE_DESCRIPTION("Xen ACPI Processor P-states (and Cx) driver which uploads PM data to Xen hypervisor"); 563 MODULE_LICENSE("GPL"); 564 565 /* We want to be loaded before the CPU freq scaling drivers are loaded. 566 * They are loaded in late_initcall. */ 567 device_initcall(xen_acpi_processor_init); 568 module_exit(xen_acpi_processor_exit); 569