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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 21 22 #include <linux/cpumask.h> 23 #include <linux/cpufreq.h> 24 #include <linux/freezer.h> 25 #include <linux/kernel.h> 26 #include <linux/kthread.h> 27 #include <linux/init.h> 28 #include <linux/module.h> 29 #include <linux/types.h> 30 #include <linux/syscore_ops.h> 31 #include <linux/acpi.h> 32 #include <acpi/processor.h> 33 #include <xen/xen.h> 34 #include <xen/interface/platform.h> 35 #include <asm/xen/hypercall.h> 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 *acpi_id_present; 54 /* And if there is an _CST definition (or a PBLK) for the ACPI IDs */ 55 static unsigned long *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; 100 101 dst_cx->dpcnt = 0; 102 set_xen_guest_handle(dst_cx->dp, NULL); 103 } 104 if (!ok) { 105 pr_debug("No _Cx for ACPI CPU %u\n", _pr->acpi_id); 106 kfree(dst_cx_states); 107 return -EINVAL; 108 } 109 op.u.set_pminfo.power.count = ok; 110 op.u.set_pminfo.power.flags.bm_control = _pr->flags.bm_control; 111 op.u.set_pminfo.power.flags.bm_check = _pr->flags.bm_check; 112 op.u.set_pminfo.power.flags.has_cst = _pr->flags.has_cst; 113 op.u.set_pminfo.power.flags.power_setup_done = 114 _pr->flags.power_setup_done; 115 116 set_xen_guest_handle(op.u.set_pminfo.power.states, dst_cx_states); 117 118 if (!no_hypercall) 119 ret = HYPERVISOR_platform_op(&op); 120 121 if (!ret) { 122 pr_debug("ACPI CPU%u - C-states uploaded.\n", _pr->acpi_id); 123 for (i = 1; i <= _pr->power.count; i++) { 124 cx = &_pr->power.states[i]; 125 if (!cx->valid) 126 continue; 127 pr_debug(" C%d: %s %d uS\n", 128 cx->type, cx->desc, (u32)cx->latency); 129 } 130 } else if ((ret != -EINVAL) && (ret != -ENOSYS)) 131 /* EINVAL means the ACPI ID is incorrect - meaning the ACPI 132 * table is referencing a non-existing CPU - which can happen 133 * with broken ACPI tables. */ 134 pr_err("(CX): Hypervisor error (%d) for ACPI CPU%u\n", 135 ret, _pr->acpi_id); 136 137 kfree(dst_cx_states); 138 139 return ret; 140 } 141 static struct xen_processor_px * 142 xen_copy_pss_data(struct acpi_processor *_pr, 143 struct xen_processor_performance *dst_perf) 144 { 145 struct xen_processor_px *dst_states = NULL; 146 unsigned int i; 147 148 BUILD_BUG_ON(sizeof(struct xen_processor_px) != 149 sizeof(struct acpi_processor_px)); 150 151 dst_states = kcalloc(_pr->performance->state_count, 152 sizeof(struct xen_processor_px), GFP_KERNEL); 153 if (!dst_states) 154 return ERR_PTR(-ENOMEM); 155 156 dst_perf->state_count = _pr->performance->state_count; 157 for (i = 0; i < _pr->performance->state_count; i++) { 158 /* Fortunatly for us, they are both the same size */ 159 memcpy(&(dst_states[i]), &(_pr->performance->states[i]), 160 sizeof(struct acpi_processor_px)); 161 } 162 return dst_states; 163 } 164 static int xen_copy_psd_data(struct acpi_processor *_pr, 165 struct xen_processor_performance *dst) 166 { 167 struct acpi_psd_package *pdomain; 168 169 BUILD_BUG_ON(sizeof(struct xen_psd_package) != 170 sizeof(struct acpi_psd_package)); 171 172 /* This information is enumerated only if acpi_processor_preregister_performance 173 * has been called. 174 */ 175 dst->shared_type = _pr->performance->shared_type; 176 177 pdomain = &(_pr->performance->domain_info); 178 179 /* 'acpi_processor_preregister_performance' does not parse if the 180 * num_processors <= 1, but Xen still requires it. Do it manually here. 181 */ 182 if (pdomain->num_processors <= 1) { 183 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL) 184 dst->shared_type = CPUFREQ_SHARED_TYPE_ALL; 185 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL) 186 dst->shared_type = CPUFREQ_SHARED_TYPE_HW; 187 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY) 188 dst->shared_type = CPUFREQ_SHARED_TYPE_ANY; 189 190 } 191 memcpy(&(dst->domain_info), pdomain, sizeof(struct acpi_psd_package)); 192 return 0; 193 } 194 static int xen_copy_pct_data(struct acpi_pct_register *pct, 195 struct xen_pct_register *dst_pct) 196 { 197 /* It would be nice if you could just do 'memcpy(pct, dst_pct') but 198 * sadly the Xen structure did not have the proper padding so the 199 * descriptor field takes two (dst_pct) bytes instead of one (pct). 200 */ 201 dst_pct->descriptor = pct->descriptor; 202 dst_pct->length = pct->length; 203 dst_pct->space_id = pct->space_id; 204 dst_pct->bit_width = pct->bit_width; 205 dst_pct->bit_offset = pct->bit_offset; 206 dst_pct->reserved = pct->reserved; 207 dst_pct->address = pct->address; 208 return 0; 209 } 210 static int push_pxx_to_hypervisor(struct acpi_processor *_pr) 211 { 212 int ret = 0; 213 struct xen_platform_op op = { 214 .cmd = XENPF_set_processor_pminfo, 215 .interface_version = XENPF_INTERFACE_VERSION, 216 .u.set_pminfo.id = _pr->acpi_id, 217 .u.set_pminfo.type = XEN_PM_PX, 218 }; 219 struct xen_processor_performance *dst_perf; 220 struct xen_processor_px *dst_states = NULL; 221 222 dst_perf = &op.u.set_pminfo.perf; 223 224 dst_perf->platform_limit = _pr->performance_platform_limit; 225 dst_perf->flags |= XEN_PX_PPC; 226 xen_copy_pct_data(&(_pr->performance->control_register), 227 &dst_perf->control_register); 228 xen_copy_pct_data(&(_pr->performance->status_register), 229 &dst_perf->status_register); 230 dst_perf->flags |= XEN_PX_PCT; 231 dst_states = xen_copy_pss_data(_pr, dst_perf); 232 if (!IS_ERR_OR_NULL(dst_states)) { 233 set_xen_guest_handle(dst_perf->states, dst_states); 234 dst_perf->flags |= XEN_PX_PSS; 235 } 236 if (!xen_copy_psd_data(_pr, dst_perf)) 237 dst_perf->flags |= XEN_PX_PSD; 238 239 if (dst_perf->flags != (XEN_PX_PSD | XEN_PX_PSS | XEN_PX_PCT | XEN_PX_PPC)) { 240 pr_warn("ACPI CPU%u missing some P-state data (%x), skipping\n", 241 _pr->acpi_id, dst_perf->flags); 242 ret = -ENODEV; 243 goto err_free; 244 } 245 246 if (!no_hypercall) 247 ret = HYPERVISOR_platform_op(&op); 248 249 if (!ret) { 250 struct acpi_processor_performance *perf; 251 unsigned int i; 252 253 perf = _pr->performance; 254 pr_debug("ACPI CPU%u - P-states uploaded.\n", _pr->acpi_id); 255 for (i = 0; i < perf->state_count; i++) { 256 pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n", 257 (i == perf->state ? '*' : ' '), i, 258 (u32) perf->states[i].core_frequency, 259 (u32) perf->states[i].power, 260 (u32) perf->states[i].transition_latency); 261 } 262 } else if ((ret != -EINVAL) && (ret != -ENOSYS)) 263 /* EINVAL means the ACPI ID is incorrect - meaning the ACPI 264 * table is referencing a non-existing CPU - which can happen 265 * with broken ACPI tables. */ 266 pr_warn("(_PXX): Hypervisor error (%d) for ACPI CPU%u\n", 267 ret, _pr->acpi_id); 268 err_free: 269 if (!IS_ERR_OR_NULL(dst_states)) 270 kfree(dst_states); 271 272 return ret; 273 } 274 static int upload_pm_data(struct acpi_processor *_pr) 275 { 276 int err = 0; 277 278 mutex_lock(&acpi_ids_mutex); 279 if (__test_and_set_bit(_pr->acpi_id, acpi_ids_done)) { 280 mutex_unlock(&acpi_ids_mutex); 281 return -EBUSY; 282 } 283 if (_pr->flags.power) 284 err = push_cxx_to_hypervisor(_pr); 285 286 if (_pr->performance && _pr->performance->states) 287 err |= push_pxx_to_hypervisor(_pr); 288 289 mutex_unlock(&acpi_ids_mutex); 290 return err; 291 } 292 static unsigned int __init get_max_acpi_id(void) 293 { 294 struct xenpf_pcpuinfo *info; 295 struct xen_platform_op op = { 296 .cmd = XENPF_get_cpuinfo, 297 .interface_version = XENPF_INTERFACE_VERSION, 298 }; 299 int ret = 0; 300 unsigned int i, last_cpu, max_acpi_id = 0; 301 302 info = &op.u.pcpu_info; 303 info->xen_cpuid = 0; 304 305 ret = HYPERVISOR_platform_op(&op); 306 if (ret) 307 return NR_CPUS; 308 309 /* The max_present is the same irregardless of the xen_cpuid */ 310 last_cpu = op.u.pcpu_info.max_present; 311 for (i = 0; i <= last_cpu; i++) { 312 info->xen_cpuid = i; 313 ret = HYPERVISOR_platform_op(&op); 314 if (ret) 315 continue; 316 max_acpi_id = max(info->acpi_id, max_acpi_id); 317 } 318 max_acpi_id *= 2; /* Slack for CPU hotplug support. */ 319 pr_debug("Max ACPI ID: %u\n", max_acpi_id); 320 return max_acpi_id; 321 } 322 /* 323 * The read_acpi_id and check_acpi_ids are there to support the Xen 324 * oddity of virtual CPUs != physical CPUs in the initial domain. 325 * The user can supply 'xen_max_vcpus=X' on the Xen hypervisor line 326 * which will band the amount of CPUs the initial domain can see. 327 * In general that is OK, except it plays havoc with any of the 328 * for_each_[present|online]_cpu macros which are banded to the virtual 329 * CPU amount. 330 */ 331 static acpi_status 332 read_acpi_id(acpi_handle handle, u32 lvl, void *context, void **rv) 333 { 334 u32 acpi_id; 335 acpi_status status; 336 acpi_object_type acpi_type; 337 unsigned long long tmp; 338 union acpi_object object = { 0 }; 339 struct acpi_buffer buffer = { sizeof(union acpi_object), &object }; 340 acpi_io_address pblk = 0; 341 342 status = acpi_get_type(handle, &acpi_type); 343 if (ACPI_FAILURE(status)) 344 return AE_OK; 345 346 switch (acpi_type) { 347 case ACPI_TYPE_PROCESSOR: 348 status = acpi_evaluate_object(handle, NULL, NULL, &buffer); 349 if (ACPI_FAILURE(status)) 350 return AE_OK; 351 acpi_id = object.processor.proc_id; 352 pblk = object.processor.pblk_address; 353 break; 354 case ACPI_TYPE_DEVICE: 355 status = acpi_evaluate_integer(handle, "_UID", NULL, &tmp); 356 if (ACPI_FAILURE(status)) 357 return AE_OK; 358 acpi_id = tmp; 359 break; 360 default: 361 return AE_OK; 362 } 363 /* There are more ACPI Processor objects than in x2APIC or MADT. 364 * This can happen with incorrect ACPI SSDT declerations. */ 365 if (acpi_id > nr_acpi_bits) { 366 pr_debug("We only have %u, trying to set %u\n", 367 nr_acpi_bits, acpi_id); 368 return AE_OK; 369 } 370 /* OK, There is a ACPI Processor object */ 371 __set_bit(acpi_id, acpi_id_present); 372 373 pr_debug("ACPI CPU%u w/ PBLK:0x%lx\n", acpi_id, (unsigned long)pblk); 374 375 status = acpi_evaluate_object(handle, "_CST", NULL, &buffer); 376 if (ACPI_FAILURE(status)) { 377 if (!pblk) 378 return AE_OK; 379 } 380 /* .. and it has a C-state */ 381 __set_bit(acpi_id, acpi_id_cst_present); 382 383 return AE_OK; 384 } 385 static int check_acpi_ids(struct acpi_processor *pr_backup) 386 { 387 388 if (!pr_backup) 389 return -ENODEV; 390 391 if (acpi_id_present && acpi_id_cst_present) 392 /* OK, done this once .. skip to uploading */ 393 goto upload; 394 395 /* All online CPUs have been processed at this stage. Now verify 396 * whether in fact "online CPUs" == physical CPUs. 397 */ 398 acpi_id_present = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL); 399 if (!acpi_id_present) 400 return -ENOMEM; 401 402 acpi_id_cst_present = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL); 403 if (!acpi_id_cst_present) { 404 kfree(acpi_id_present); 405 return -ENOMEM; 406 } 407 408 acpi_walk_namespace(ACPI_TYPE_PROCESSOR, ACPI_ROOT_OBJECT, 409 ACPI_UINT32_MAX, 410 read_acpi_id, NULL, NULL, NULL); 411 acpi_get_devices(ACPI_PROCESSOR_DEVICE_HID, read_acpi_id, NULL, NULL); 412 413 upload: 414 if (!bitmap_equal(acpi_id_present, acpi_ids_done, nr_acpi_bits)) { 415 unsigned int i; 416 for_each_set_bit(i, acpi_id_present, nr_acpi_bits) { 417 pr_backup->acpi_id = i; 418 /* Mask out C-states if there are no _CST or PBLK */ 419 pr_backup->flags.power = test_bit(i, acpi_id_cst_present); 420 (void)upload_pm_data(pr_backup); 421 } 422 } 423 424 return 0; 425 } 426 427 /* acpi_perf_data is a pointer to percpu data. */ 428 static struct acpi_processor_performance __percpu *acpi_perf_data; 429 430 static void free_acpi_perf_data(void) 431 { 432 unsigned int i; 433 434 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */ 435 for_each_possible_cpu(i) 436 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i) 437 ->shared_cpu_map); 438 free_percpu(acpi_perf_data); 439 } 440 441 static int xen_upload_processor_pm_data(void) 442 { 443 struct acpi_processor *pr_backup = NULL; 444 unsigned int i; 445 int rc = 0; 446 447 pr_info("Uploading Xen processor PM info\n"); 448 449 for_each_possible_cpu(i) { 450 struct acpi_processor *_pr; 451 _pr = per_cpu(processors, i /* APIC ID */); 452 if (!_pr) 453 continue; 454 455 if (!pr_backup) { 456 pr_backup = kzalloc(sizeof(struct acpi_processor), GFP_KERNEL); 457 if (pr_backup) 458 memcpy(pr_backup, _pr, sizeof(struct acpi_processor)); 459 } 460 (void)upload_pm_data(_pr); 461 } 462 463 rc = check_acpi_ids(pr_backup); 464 kfree(pr_backup); 465 466 return rc; 467 } 468 469 static void xen_acpi_processor_resume_worker(struct work_struct *dummy) 470 { 471 int rc; 472 473 bitmap_zero(acpi_ids_done, nr_acpi_bits); 474 475 rc = xen_upload_processor_pm_data(); 476 if (rc != 0) 477 pr_info("ACPI data upload failed, error = %d\n", rc); 478 } 479 480 static void xen_acpi_processor_resume(void) 481 { 482 static DECLARE_WORK(wq, xen_acpi_processor_resume_worker); 483 484 /* 485 * xen_upload_processor_pm_data() calls non-atomic code. 486 * However, the context for xen_acpi_processor_resume is syscore 487 * with only the boot CPU online and in an atomic context. 488 * 489 * So defer the upload for some point safer. 490 */ 491 schedule_work(&wq); 492 } 493 494 static struct syscore_ops xap_syscore_ops = { 495 .resume = xen_acpi_processor_resume, 496 }; 497 498 static int __init xen_acpi_processor_init(void) 499 { 500 unsigned int i; 501 int rc; 502 503 if (!xen_initial_domain()) 504 return -ENODEV; 505 506 nr_acpi_bits = get_max_acpi_id() + 1; 507 acpi_ids_done = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL); 508 if (!acpi_ids_done) 509 return -ENOMEM; 510 511 acpi_perf_data = alloc_percpu(struct acpi_processor_performance); 512 if (!acpi_perf_data) { 513 pr_debug("Memory allocation error for acpi_perf_data\n"); 514 kfree(acpi_ids_done); 515 return -ENOMEM; 516 } 517 for_each_possible_cpu(i) { 518 if (!zalloc_cpumask_var_node( 519 &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map, 520 GFP_KERNEL, cpu_to_node(i))) { 521 rc = -ENOMEM; 522 goto err_out; 523 } 524 } 525 526 /* Do initialization in ACPI core. It is OK to fail here. */ 527 (void)acpi_processor_preregister_performance(acpi_perf_data); 528 529 for_each_possible_cpu(i) { 530 struct acpi_processor *pr; 531 struct acpi_processor_performance *perf; 532 533 pr = per_cpu(processors, i); 534 perf = per_cpu_ptr(acpi_perf_data, i); 535 if (!pr) 536 continue; 537 538 pr->performance = perf; 539 rc = acpi_processor_get_performance_info(pr); 540 if (rc) 541 goto err_out; 542 } 543 544 rc = xen_upload_processor_pm_data(); 545 if (rc) 546 goto err_unregister; 547 548 register_syscore_ops(&xap_syscore_ops); 549 550 return 0; 551 err_unregister: 552 for_each_possible_cpu(i) 553 acpi_processor_unregister_performance(i); 554 555 err_out: 556 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */ 557 free_acpi_perf_data(); 558 kfree(acpi_ids_done); 559 return rc; 560 } 561 static void __exit xen_acpi_processor_exit(void) 562 { 563 int i; 564 565 unregister_syscore_ops(&xap_syscore_ops); 566 kfree(acpi_ids_done); 567 kfree(acpi_id_present); 568 kfree(acpi_id_cst_present); 569 for_each_possible_cpu(i) 570 acpi_processor_unregister_performance(i); 571 572 free_acpi_perf_data(); 573 } 574 575 MODULE_AUTHOR("Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>"); 576 MODULE_DESCRIPTION("Xen ACPI Processor P-states (and Cx) driver which uploads PM data to Xen hypervisor"); 577 MODULE_LICENSE("GPL"); 578 579 /* We want to be loaded before the CPU freq scaling drivers are loaded. 580 * They are loaded in late_initcall. */ 581 device_initcall(xen_acpi_processor_init); 582 module_exit(xen_acpi_processor_exit); 583