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/cpuidle.h> 12 #include <linux/cpumask.h> 13 #include <linux/cpu_pm.h> 14 #include <linux/cpu_cooling.h> 15 #include <linux/kernel.h> 16 #include <linux/module.h> 17 #include <linux/of.h> 18 #include <linux/of_device.h> 19 #include <linux/slab.h> 20 #include <linux/platform_device.h> 21 #include <linux/pm_domain.h> 22 #include <linux/pm_runtime.h> 23 #include <asm/cpuidle.h> 24 #include <asm/sbi.h> 25 #include <asm/suspend.h> 26 27 #include "dt_idle_states.h" 28 #include "dt_idle_genpd.h" 29 30 struct sbi_cpuidle_data { 31 u32 *states; 32 struct device *dev; 33 }; 34 35 struct sbi_domain_state { 36 bool available; 37 u32 state; 38 }; 39 40 static DEFINE_PER_CPU_READ_MOSTLY(struct sbi_cpuidle_data, sbi_cpuidle_data); 41 static DEFINE_PER_CPU(struct sbi_domain_state, domain_state); 42 static bool sbi_cpuidle_use_osi; 43 static bool sbi_cpuidle_use_cpuhp; 44 static bool sbi_cpuidle_pd_allow_domain_state; 45 46 static inline void sbi_set_domain_state(u32 state) 47 { 48 struct sbi_domain_state *data = this_cpu_ptr(&domain_state); 49 50 data->available = true; 51 data->state = state; 52 } 53 54 static inline u32 sbi_get_domain_state(void) 55 { 56 struct sbi_domain_state *data = this_cpu_ptr(&domain_state); 57 58 return data->state; 59 } 60 61 static inline void sbi_clear_domain_state(void) 62 { 63 struct sbi_domain_state *data = this_cpu_ptr(&domain_state); 64 65 data->available = false; 66 } 67 68 static inline bool sbi_is_domain_state_available(void) 69 { 70 struct sbi_domain_state *data = this_cpu_ptr(&domain_state); 71 72 return data->available; 73 } 74 75 static int sbi_suspend_finisher(unsigned long suspend_type, 76 unsigned long resume_addr, 77 unsigned long opaque) 78 { 79 struct sbiret ret; 80 81 ret = sbi_ecall(SBI_EXT_HSM, SBI_EXT_HSM_HART_SUSPEND, 82 suspend_type, resume_addr, opaque, 0, 0, 0); 83 84 return (ret.error) ? sbi_err_map_linux_errno(ret.error) : 0; 85 } 86 87 static int sbi_suspend(u32 state) 88 { 89 if (state & SBI_HSM_SUSP_NON_RET_BIT) 90 return cpu_suspend(state, sbi_suspend_finisher); 91 else 92 return sbi_suspend_finisher(state, 0, 0); 93 } 94 95 static int sbi_cpuidle_enter_state(struct cpuidle_device *dev, 96 struct cpuidle_driver *drv, int idx) 97 { 98 u32 *states = __this_cpu_read(sbi_cpuidle_data.states); 99 100 return CPU_PM_CPU_IDLE_ENTER_PARAM(sbi_suspend, idx, states[idx]); 101 } 102 103 static int __sbi_enter_domain_idle_state(struct cpuidle_device *dev, 104 struct cpuidle_driver *drv, int idx, 105 bool s2idle) 106 { 107 struct sbi_cpuidle_data *data = this_cpu_ptr(&sbi_cpuidle_data); 108 u32 *states = data->states; 109 struct device *pd_dev = data->dev; 110 u32 state; 111 int ret; 112 113 ret = cpu_pm_enter(); 114 if (ret) 115 return -1; 116 117 /* Do runtime PM to manage a hierarchical CPU toplogy. */ 118 rcu_irq_enter_irqson(); 119 if (s2idle) 120 dev_pm_genpd_suspend(pd_dev); 121 else 122 pm_runtime_put_sync_suspend(pd_dev); 123 rcu_irq_exit_irqson(); 124 125 if (sbi_is_domain_state_available()) 126 state = sbi_get_domain_state(); 127 else 128 state = states[idx]; 129 130 ret = sbi_suspend(state) ? -1 : idx; 131 132 rcu_irq_enter_irqson(); 133 if (s2idle) 134 dev_pm_genpd_resume(pd_dev); 135 else 136 pm_runtime_get_sync(pd_dev); 137 rcu_irq_exit_irqson(); 138 139 cpu_pm_exit(); 140 141 /* Clear the domain state to start fresh when back from idle. */ 142 sbi_clear_domain_state(); 143 return ret; 144 } 145 146 static int sbi_enter_domain_idle_state(struct cpuidle_device *dev, 147 struct cpuidle_driver *drv, int idx) 148 { 149 return __sbi_enter_domain_idle_state(dev, drv, idx, false); 150 } 151 152 static int sbi_enter_s2idle_domain_idle_state(struct cpuidle_device *dev, 153 struct cpuidle_driver *drv, 154 int idx) 155 { 156 return __sbi_enter_domain_idle_state(dev, drv, idx, true); 157 } 158 159 static int sbi_cpuidle_cpuhp_up(unsigned int cpu) 160 { 161 struct device *pd_dev = __this_cpu_read(sbi_cpuidle_data.dev); 162 163 if (pd_dev) 164 pm_runtime_get_sync(pd_dev); 165 166 return 0; 167 } 168 169 static int sbi_cpuidle_cpuhp_down(unsigned int cpu) 170 { 171 struct device *pd_dev = __this_cpu_read(sbi_cpuidle_data.dev); 172 173 if (pd_dev) { 174 pm_runtime_put_sync(pd_dev); 175 /* Clear domain state to start fresh at next online. */ 176 sbi_clear_domain_state(); 177 } 178 179 return 0; 180 } 181 182 static void sbi_idle_init_cpuhp(void) 183 { 184 int err; 185 186 if (!sbi_cpuidle_use_cpuhp) 187 return; 188 189 err = cpuhp_setup_state_nocalls(CPUHP_AP_CPU_PM_STARTING, 190 "cpuidle/sbi:online", 191 sbi_cpuidle_cpuhp_up, 192 sbi_cpuidle_cpuhp_down); 193 if (err) 194 pr_warn("Failed %d while setup cpuhp state\n", err); 195 } 196 197 static const struct of_device_id sbi_cpuidle_state_match[] = { 198 { .compatible = "riscv,idle-state", 199 .data = sbi_cpuidle_enter_state }, 200 { }, 201 }; 202 203 static bool sbi_suspend_state_is_valid(u32 state) 204 { 205 if (state > SBI_HSM_SUSPEND_RET_DEFAULT && 206 state < SBI_HSM_SUSPEND_RET_PLATFORM) 207 return false; 208 if (state > SBI_HSM_SUSPEND_NON_RET_DEFAULT && 209 state < SBI_HSM_SUSPEND_NON_RET_PLATFORM) 210 return false; 211 return true; 212 } 213 214 static int sbi_dt_parse_state_node(struct device_node *np, u32 *state) 215 { 216 int err = of_property_read_u32(np, "riscv,sbi-suspend-param", state); 217 218 if (err) { 219 pr_warn("%pOF missing riscv,sbi-suspend-param property\n", np); 220 return err; 221 } 222 223 if (!sbi_suspend_state_is_valid(*state)) { 224 pr_warn("Invalid SBI suspend state %#x\n", *state); 225 return -EINVAL; 226 } 227 228 return 0; 229 } 230 231 static int sbi_dt_cpu_init_topology(struct cpuidle_driver *drv, 232 struct sbi_cpuidle_data *data, 233 unsigned int state_count, int cpu) 234 { 235 /* Currently limit the hierarchical topology to be used in OSI mode. */ 236 if (!sbi_cpuidle_use_osi) 237 return 0; 238 239 data->dev = dt_idle_attach_cpu(cpu, "sbi"); 240 if (IS_ERR_OR_NULL(data->dev)) 241 return PTR_ERR_OR_ZERO(data->dev); 242 243 /* 244 * Using the deepest state for the CPU to trigger a potential selection 245 * of a shared state for the domain, assumes the domain states are all 246 * deeper states. 247 */ 248 drv->states[state_count - 1].enter = sbi_enter_domain_idle_state; 249 drv->states[state_count - 1].enter_s2idle = 250 sbi_enter_s2idle_domain_idle_state; 251 sbi_cpuidle_use_cpuhp = true; 252 253 return 0; 254 } 255 256 static int sbi_cpuidle_dt_init_states(struct device *dev, 257 struct cpuidle_driver *drv, 258 unsigned int cpu, 259 unsigned int state_count) 260 { 261 struct sbi_cpuidle_data *data = per_cpu_ptr(&sbi_cpuidle_data, cpu); 262 struct device_node *state_node; 263 struct device_node *cpu_node; 264 u32 *states; 265 int i, ret; 266 267 cpu_node = of_cpu_device_node_get(cpu); 268 if (!cpu_node) 269 return -ENODEV; 270 271 states = devm_kcalloc(dev, state_count, sizeof(*states), GFP_KERNEL); 272 if (!states) { 273 ret = -ENOMEM; 274 goto fail; 275 } 276 277 /* Parse SBI specific details from state DT nodes */ 278 for (i = 1; i < state_count; i++) { 279 state_node = of_get_cpu_state_node(cpu_node, i - 1); 280 if (!state_node) 281 break; 282 283 ret = sbi_dt_parse_state_node(state_node, &states[i]); 284 of_node_put(state_node); 285 286 if (ret) 287 return ret; 288 289 pr_debug("sbi-state %#x index %d\n", states[i], i); 290 } 291 if (i != state_count) { 292 ret = -ENODEV; 293 goto fail; 294 } 295 296 /* Initialize optional data, used for the hierarchical topology. */ 297 ret = sbi_dt_cpu_init_topology(drv, data, state_count, cpu); 298 if (ret < 0) 299 return ret; 300 301 /* Store states in the per-cpu struct. */ 302 data->states = states; 303 304 fail: 305 of_node_put(cpu_node); 306 307 return ret; 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, state_count = 0; 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 = state_count > 0 ? &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_find_property(node, "#power-domain-cells", NULL)) 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_find_property(np, "power-domains", NULL) && 545 of_find_property(np, "power-domain-names", NULL)) { 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) <= 0) { 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