1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * RISC-V SBI CPU idle driver. 4 * 5 * Copyright (c) 2021 Western Digital Corporation or its affiliates. 6 * Copyright (c) 2022 Ventana Micro Systems Inc. 7 */ 8 9 #define pr_fmt(fmt) "cpuidle-riscv-sbi: " fmt 10 11 #include <linux/cleanup.h> 12 #include <linux/cpuhotplug.h> 13 #include <linux/cpuidle.h> 14 #include <linux/cpumask.h> 15 #include <linux/cpu_pm.h> 16 #include <linux/cpu_cooling.h> 17 #include <linux/kernel.h> 18 #include <linux/module.h> 19 #include <linux/of.h> 20 #include <linux/slab.h> 21 #include <linux/platform_device.h> 22 #include <linux/pm_domain.h> 23 #include <linux/pm_runtime.h> 24 #include <asm/cpuidle.h> 25 #include <asm/sbi.h> 26 #include <asm/smp.h> 27 #include <asm/suspend.h> 28 29 #include "dt_idle_states.h" 30 #include "dt_idle_genpd.h" 31 32 struct sbi_cpuidle_data { 33 u32 *states; 34 struct device *dev; 35 }; 36 37 struct sbi_domain_state { 38 bool available; 39 u32 state; 40 }; 41 42 static DEFINE_PER_CPU_READ_MOSTLY(struct sbi_cpuidle_data, sbi_cpuidle_data); 43 static DEFINE_PER_CPU(struct sbi_domain_state, domain_state); 44 static bool sbi_cpuidle_use_osi; 45 static bool sbi_cpuidle_use_cpuhp; 46 static bool sbi_cpuidle_pd_allow_domain_state; 47 48 static inline void sbi_set_domain_state(u32 state) 49 { 50 struct sbi_domain_state *data = this_cpu_ptr(&domain_state); 51 52 data->available = true; 53 data->state = state; 54 } 55 56 static inline u32 sbi_get_domain_state(void) 57 { 58 struct sbi_domain_state *data = this_cpu_ptr(&domain_state); 59 60 return data->state; 61 } 62 63 static inline void sbi_clear_domain_state(void) 64 { 65 struct sbi_domain_state *data = this_cpu_ptr(&domain_state); 66 67 data->available = false; 68 } 69 70 static inline bool sbi_is_domain_state_available(void) 71 { 72 struct sbi_domain_state *data = this_cpu_ptr(&domain_state); 73 74 return data->available; 75 } 76 77 static int sbi_suspend_finisher(unsigned long suspend_type, 78 unsigned long resume_addr, 79 unsigned long opaque) 80 { 81 struct sbiret ret; 82 83 ret = sbi_ecall(SBI_EXT_HSM, SBI_EXT_HSM_HART_SUSPEND, 84 suspend_type, resume_addr, opaque, 0, 0, 0); 85 86 return (ret.error) ? sbi_err_map_linux_errno(ret.error) : 0; 87 } 88 89 static int sbi_suspend(u32 state) 90 { 91 if (state & SBI_HSM_SUSP_NON_RET_BIT) 92 return cpu_suspend(state, sbi_suspend_finisher); 93 else 94 return sbi_suspend_finisher(state, 0, 0); 95 } 96 97 static __cpuidle int sbi_cpuidle_enter_state(struct cpuidle_device *dev, 98 struct cpuidle_driver *drv, int idx) 99 { 100 u32 *states = __this_cpu_read(sbi_cpuidle_data.states); 101 u32 state = states[idx]; 102 103 if (state & SBI_HSM_SUSP_NON_RET_BIT) 104 return CPU_PM_CPU_IDLE_ENTER_PARAM(sbi_suspend, idx, state); 105 else 106 return CPU_PM_CPU_IDLE_ENTER_RETENTION_PARAM(sbi_suspend, 107 idx, state); 108 } 109 110 static __cpuidle int __sbi_enter_domain_idle_state(struct cpuidle_device *dev, 111 struct cpuidle_driver *drv, int idx, 112 bool s2idle) 113 { 114 struct sbi_cpuidle_data *data = this_cpu_ptr(&sbi_cpuidle_data); 115 u32 *states = data->states; 116 struct device *pd_dev = data->dev; 117 u32 state; 118 int ret; 119 120 ret = cpu_pm_enter(); 121 if (ret) 122 return -1; 123 124 /* Do runtime PM to manage a hierarchical CPU toplogy. */ 125 if (s2idle) 126 dev_pm_genpd_suspend(pd_dev); 127 else 128 pm_runtime_put_sync_suspend(pd_dev); 129 130 ct_cpuidle_enter(); 131 132 if (sbi_is_domain_state_available()) 133 state = sbi_get_domain_state(); 134 else 135 state = states[idx]; 136 137 ret = sbi_suspend(state) ? -1 : idx; 138 139 ct_cpuidle_exit(); 140 141 if (s2idle) 142 dev_pm_genpd_resume(pd_dev); 143 else 144 pm_runtime_get_sync(pd_dev); 145 146 cpu_pm_exit(); 147 148 /* Clear the domain state to start fresh when back from idle. */ 149 sbi_clear_domain_state(); 150 return ret; 151 } 152 153 static int sbi_enter_domain_idle_state(struct cpuidle_device *dev, 154 struct cpuidle_driver *drv, int idx) 155 { 156 return __sbi_enter_domain_idle_state(dev, drv, idx, false); 157 } 158 159 static int sbi_enter_s2idle_domain_idle_state(struct cpuidle_device *dev, 160 struct cpuidle_driver *drv, 161 int idx) 162 { 163 return __sbi_enter_domain_idle_state(dev, drv, idx, true); 164 } 165 166 static int sbi_cpuidle_cpuhp_up(unsigned int cpu) 167 { 168 struct device *pd_dev = __this_cpu_read(sbi_cpuidle_data.dev); 169 170 if (pd_dev) 171 pm_runtime_get_sync(pd_dev); 172 173 return 0; 174 } 175 176 static int sbi_cpuidle_cpuhp_down(unsigned int cpu) 177 { 178 struct device *pd_dev = __this_cpu_read(sbi_cpuidle_data.dev); 179 180 if (pd_dev) { 181 pm_runtime_put_sync(pd_dev); 182 /* Clear domain state to start fresh at next online. */ 183 sbi_clear_domain_state(); 184 } 185 186 return 0; 187 } 188 189 static void sbi_idle_init_cpuhp(void) 190 { 191 int err; 192 193 if (!sbi_cpuidle_use_cpuhp) 194 return; 195 196 err = cpuhp_setup_state_nocalls(CPUHP_AP_CPU_PM_STARTING, 197 "cpuidle/sbi:online", 198 sbi_cpuidle_cpuhp_up, 199 sbi_cpuidle_cpuhp_down); 200 if (err) 201 pr_warn("Failed %d while setup cpuhp state\n", err); 202 } 203 204 static const struct of_device_id sbi_cpuidle_state_match[] = { 205 { .compatible = "riscv,idle-state", 206 .data = sbi_cpuidle_enter_state }, 207 { }, 208 }; 209 210 static bool sbi_suspend_state_is_valid(u32 state) 211 { 212 if (state > SBI_HSM_SUSPEND_RET_DEFAULT && 213 state < SBI_HSM_SUSPEND_RET_PLATFORM) 214 return false; 215 if (state > SBI_HSM_SUSPEND_NON_RET_DEFAULT && 216 state < SBI_HSM_SUSPEND_NON_RET_PLATFORM) 217 return false; 218 return true; 219 } 220 221 static int sbi_dt_parse_state_node(struct device_node *np, u32 *state) 222 { 223 int err = of_property_read_u32(np, "riscv,sbi-suspend-param", state); 224 225 if (err) { 226 pr_warn("%pOF missing riscv,sbi-suspend-param property\n", np); 227 return err; 228 } 229 230 if (!sbi_suspend_state_is_valid(*state)) { 231 pr_warn("Invalid SBI suspend state %#x\n", *state); 232 return -EINVAL; 233 } 234 235 return 0; 236 } 237 238 static int sbi_dt_cpu_init_topology(struct cpuidle_driver *drv, 239 struct sbi_cpuidle_data *data, 240 unsigned int state_count, int cpu) 241 { 242 /* Currently limit the hierarchical topology to be used in OSI mode. */ 243 if (!sbi_cpuidle_use_osi) 244 return 0; 245 246 data->dev = dt_idle_attach_cpu(cpu, "sbi"); 247 if (IS_ERR_OR_NULL(data->dev)) 248 return PTR_ERR_OR_ZERO(data->dev); 249 250 /* 251 * Using the deepest state for the CPU to trigger a potential selection 252 * of a shared state for the domain, assumes the domain states are all 253 * deeper states. 254 */ 255 drv->states[state_count - 1].flags |= CPUIDLE_FLAG_RCU_IDLE; 256 drv->states[state_count - 1].enter = sbi_enter_domain_idle_state; 257 drv->states[state_count - 1].enter_s2idle = 258 sbi_enter_s2idle_domain_idle_state; 259 sbi_cpuidle_use_cpuhp = true; 260 261 return 0; 262 } 263 264 static int sbi_cpuidle_dt_init_states(struct device *dev, 265 struct cpuidle_driver *drv, 266 unsigned int cpu, 267 unsigned int state_count) 268 { 269 struct sbi_cpuidle_data *data = per_cpu_ptr(&sbi_cpuidle_data, cpu); 270 struct device_node *state_node; 271 u32 *states; 272 int i, ret; 273 274 struct device_node *cpu_node __free(device_node) = of_cpu_device_node_get(cpu); 275 if (!cpu_node) 276 return -ENODEV; 277 278 states = devm_kcalloc(dev, state_count, sizeof(*states), GFP_KERNEL); 279 if (!states) 280 return -ENOMEM; 281 282 /* Parse SBI specific details from state DT nodes */ 283 for (i = 1; i < state_count; i++) { 284 state_node = of_get_cpu_state_node(cpu_node, i - 1); 285 if (!state_node) 286 break; 287 288 ret = sbi_dt_parse_state_node(state_node, &states[i]); 289 of_node_put(state_node); 290 291 if (ret) 292 return ret; 293 294 pr_debug("sbi-state %#x index %d\n", states[i], i); 295 } 296 if (i != state_count) 297 return -ENODEV; 298 299 /* Initialize optional data, used for the hierarchical topology. */ 300 ret = sbi_dt_cpu_init_topology(drv, data, state_count, cpu); 301 if (ret < 0) 302 return ret; 303 304 /* Store states in the per-cpu struct. */ 305 data->states = states; 306 307 return 0; 308 } 309 310 static void sbi_cpuidle_deinit_cpu(int cpu) 311 { 312 struct sbi_cpuidle_data *data = per_cpu_ptr(&sbi_cpuidle_data, cpu); 313 314 dt_idle_detach_cpu(data->dev); 315 sbi_cpuidle_use_cpuhp = false; 316 } 317 318 static int sbi_cpuidle_init_cpu(struct device *dev, int cpu) 319 { 320 struct cpuidle_driver *drv; 321 unsigned int state_count = 0; 322 int ret = 0; 323 324 drv = devm_kzalloc(dev, sizeof(*drv), GFP_KERNEL); 325 if (!drv) 326 return -ENOMEM; 327 328 drv->name = "sbi_cpuidle"; 329 drv->owner = THIS_MODULE; 330 drv->cpumask = (struct cpumask *)cpumask_of(cpu); 331 332 /* RISC-V architectural WFI to be represented as state index 0. */ 333 drv->states[0].enter = sbi_cpuidle_enter_state; 334 drv->states[0].exit_latency = 1; 335 drv->states[0].target_residency = 1; 336 drv->states[0].power_usage = UINT_MAX; 337 strcpy(drv->states[0].name, "WFI"); 338 strcpy(drv->states[0].desc, "RISC-V WFI"); 339 340 /* 341 * If no DT idle states are detected (ret == 0) let the driver 342 * initialization fail accordingly since there is no reason to 343 * initialize the idle driver if only wfi is supported, the 344 * default archictectural back-end already executes wfi 345 * on idle entry. 346 */ 347 ret = dt_init_idle_driver(drv, sbi_cpuidle_state_match, 1); 348 if (ret <= 0) { 349 pr_debug("HART%ld: failed to parse DT idle states\n", 350 cpuid_to_hartid_map(cpu)); 351 return ret ? : -ENODEV; 352 } 353 state_count = ret + 1; /* Include WFI state as well */ 354 355 /* Initialize idle states from DT. */ 356 ret = sbi_cpuidle_dt_init_states(dev, drv, cpu, state_count); 357 if (ret) { 358 pr_err("HART%ld: failed to init idle states\n", 359 cpuid_to_hartid_map(cpu)); 360 return ret; 361 } 362 363 ret = cpuidle_register(drv, NULL); 364 if (ret) 365 goto deinit; 366 367 cpuidle_cooling_register(drv); 368 369 return 0; 370 deinit: 371 sbi_cpuidle_deinit_cpu(cpu); 372 return ret; 373 } 374 375 static void sbi_cpuidle_domain_sync_state(struct device *dev) 376 { 377 /* 378 * All devices have now been attached/probed to the PM domain 379 * topology, hence it's fine to allow domain states to be picked. 380 */ 381 sbi_cpuidle_pd_allow_domain_state = true; 382 } 383 384 #ifdef CONFIG_DT_IDLE_GENPD 385 386 static int sbi_cpuidle_pd_power_off(struct generic_pm_domain *pd) 387 { 388 struct genpd_power_state *state = &pd->states[pd->state_idx]; 389 u32 *pd_state; 390 391 if (!state->data) 392 return 0; 393 394 if (!sbi_cpuidle_pd_allow_domain_state) 395 return -EBUSY; 396 397 /* OSI mode is enabled, set the corresponding domain state. */ 398 pd_state = state->data; 399 sbi_set_domain_state(*pd_state); 400 401 return 0; 402 } 403 404 struct sbi_pd_provider { 405 struct list_head link; 406 struct device_node *node; 407 }; 408 409 static LIST_HEAD(sbi_pd_providers); 410 411 static int sbi_pd_init(struct device_node *np) 412 { 413 struct generic_pm_domain *pd; 414 struct sbi_pd_provider *pd_provider; 415 struct dev_power_governor *pd_gov; 416 int ret = -ENOMEM; 417 418 pd = dt_idle_pd_alloc(np, sbi_dt_parse_state_node); 419 if (!pd) 420 goto out; 421 422 pd_provider = kzalloc(sizeof(*pd_provider), GFP_KERNEL); 423 if (!pd_provider) 424 goto free_pd; 425 426 pd->flags |= GENPD_FLAG_IRQ_SAFE | GENPD_FLAG_CPU_DOMAIN; 427 428 /* Allow power off when OSI is available. */ 429 if (sbi_cpuidle_use_osi) 430 pd->power_off = sbi_cpuidle_pd_power_off; 431 else 432 pd->flags |= GENPD_FLAG_ALWAYS_ON; 433 434 /* Use governor for CPU PM domains if it has some states to manage. */ 435 pd_gov = pd->states ? &pm_domain_cpu_gov : NULL; 436 437 ret = pm_genpd_init(pd, pd_gov, false); 438 if (ret) 439 goto free_pd_prov; 440 441 ret = of_genpd_add_provider_simple(np, pd); 442 if (ret) 443 goto remove_pd; 444 445 pd_provider->node = of_node_get(np); 446 list_add(&pd_provider->link, &sbi_pd_providers); 447 448 pr_debug("init PM domain %s\n", pd->name); 449 return 0; 450 451 remove_pd: 452 pm_genpd_remove(pd); 453 free_pd_prov: 454 kfree(pd_provider); 455 free_pd: 456 dt_idle_pd_free(pd); 457 out: 458 pr_err("failed to init PM domain ret=%d %pOF\n", ret, np); 459 return ret; 460 } 461 462 static void sbi_pd_remove(void) 463 { 464 struct sbi_pd_provider *pd_provider, *it; 465 struct generic_pm_domain *genpd; 466 467 list_for_each_entry_safe(pd_provider, it, &sbi_pd_providers, link) { 468 of_genpd_del_provider(pd_provider->node); 469 470 genpd = of_genpd_remove_last(pd_provider->node); 471 if (!IS_ERR(genpd)) 472 kfree(genpd); 473 474 of_node_put(pd_provider->node); 475 list_del(&pd_provider->link); 476 kfree(pd_provider); 477 } 478 } 479 480 static int sbi_genpd_probe(struct device_node *np) 481 { 482 struct device_node *node; 483 int ret = 0, pd_count = 0; 484 485 if (!np) 486 return -ENODEV; 487 488 /* 489 * Parse child nodes for the "#power-domain-cells" property and 490 * initialize a genpd/genpd-of-provider pair when it's found. 491 */ 492 for_each_child_of_node(np, node) { 493 if (!of_property_present(node, "#power-domain-cells")) 494 continue; 495 496 ret = sbi_pd_init(node); 497 if (ret) 498 goto put_node; 499 500 pd_count++; 501 } 502 503 /* Bail out if not using the hierarchical CPU topology. */ 504 if (!pd_count) 505 goto no_pd; 506 507 /* Link genpd masters/subdomains to model the CPU topology. */ 508 ret = dt_idle_pd_init_topology(np); 509 if (ret) 510 goto remove_pd; 511 512 return 0; 513 514 put_node: 515 of_node_put(node); 516 remove_pd: 517 sbi_pd_remove(); 518 pr_err("failed to create CPU PM domains ret=%d\n", ret); 519 no_pd: 520 return ret; 521 } 522 523 #else 524 525 static inline int sbi_genpd_probe(struct device_node *np) 526 { 527 return 0; 528 } 529 530 #endif 531 532 static int sbi_cpuidle_probe(struct platform_device *pdev) 533 { 534 int cpu, ret; 535 struct cpuidle_driver *drv; 536 struct cpuidle_device *dev; 537 struct device_node *np, *pds_node; 538 539 /* Detect OSI support based on CPU DT nodes */ 540 sbi_cpuidle_use_osi = true; 541 for_each_possible_cpu(cpu) { 542 np = of_cpu_device_node_get(cpu); 543 if (np && 544 of_property_present(np, "power-domains") && 545 of_property_present(np, "power-domain-names")) { 546 continue; 547 } else { 548 sbi_cpuidle_use_osi = false; 549 break; 550 } 551 } 552 553 /* Populate generic power domains from DT nodes */ 554 pds_node = of_find_node_by_path("/cpus/power-domains"); 555 if (pds_node) { 556 ret = sbi_genpd_probe(pds_node); 557 of_node_put(pds_node); 558 if (ret) 559 return ret; 560 } 561 562 /* Initialize CPU idle driver for each CPU */ 563 for_each_possible_cpu(cpu) { 564 ret = sbi_cpuidle_init_cpu(&pdev->dev, cpu); 565 if (ret) { 566 pr_debug("HART%ld: idle driver init failed\n", 567 cpuid_to_hartid_map(cpu)); 568 goto out_fail; 569 } 570 } 571 572 /* Setup CPU hotplut notifiers */ 573 sbi_idle_init_cpuhp(); 574 575 pr_info("idle driver registered for all CPUs\n"); 576 577 return 0; 578 579 out_fail: 580 while (--cpu >= 0) { 581 dev = per_cpu(cpuidle_devices, cpu); 582 drv = cpuidle_get_cpu_driver(dev); 583 cpuidle_unregister(drv); 584 sbi_cpuidle_deinit_cpu(cpu); 585 } 586 587 return ret; 588 } 589 590 static struct platform_driver sbi_cpuidle_driver = { 591 .probe = sbi_cpuidle_probe, 592 .driver = { 593 .name = "sbi-cpuidle", 594 .sync_state = sbi_cpuidle_domain_sync_state, 595 }, 596 }; 597 598 static int __init sbi_cpuidle_init(void) 599 { 600 int ret; 601 struct platform_device *pdev; 602 603 /* 604 * The SBI HSM suspend function is only available when: 605 * 1) SBI version is 0.3 or higher 606 * 2) SBI HSM extension is available 607 */ 608 if ((sbi_spec_version < sbi_mk_version(0, 3)) || 609 !sbi_probe_extension(SBI_EXT_HSM)) { 610 pr_info("HSM suspend not available\n"); 611 return 0; 612 } 613 614 ret = platform_driver_register(&sbi_cpuidle_driver); 615 if (ret) 616 return ret; 617 618 pdev = platform_device_register_simple("sbi-cpuidle", 619 -1, NULL, 0); 620 if (IS_ERR(pdev)) { 621 platform_driver_unregister(&sbi_cpuidle_driver); 622 return PTR_ERR(pdev); 623 } 624 625 return 0; 626 } 627 device_initcall(sbi_cpuidle_init); 628