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/smp.h> 26 #include <asm/suspend.h> 27 28 #include "dt_idle_states.h" 29 #include "dt_idle_genpd.h" 30 31 struct sbi_cpuidle_data { 32 u32 *states; 33 struct device *dev; 34 }; 35 36 struct sbi_domain_state { 37 bool available; 38 u32 state; 39 }; 40 41 static DEFINE_PER_CPU_READ_MOSTLY(struct sbi_cpuidle_data, sbi_cpuidle_data); 42 static DEFINE_PER_CPU(struct sbi_domain_state, domain_state); 43 static bool sbi_cpuidle_use_osi; 44 static bool sbi_cpuidle_use_cpuhp; 45 static bool sbi_cpuidle_pd_allow_domain_state; 46 47 static inline void sbi_set_domain_state(u32 state) 48 { 49 struct sbi_domain_state *data = this_cpu_ptr(&domain_state); 50 51 data->available = true; 52 data->state = state; 53 } 54 55 static inline u32 sbi_get_domain_state(void) 56 { 57 struct sbi_domain_state *data = this_cpu_ptr(&domain_state); 58 59 return data->state; 60 } 61 62 static inline void sbi_clear_domain_state(void) 63 { 64 struct sbi_domain_state *data = this_cpu_ptr(&domain_state); 65 66 data->available = false; 67 } 68 69 static inline bool sbi_is_domain_state_available(void) 70 { 71 struct sbi_domain_state *data = this_cpu_ptr(&domain_state); 72 73 return data->available; 74 } 75 76 static int sbi_suspend_finisher(unsigned long suspend_type, 77 unsigned long resume_addr, 78 unsigned long opaque) 79 { 80 struct sbiret ret; 81 82 ret = sbi_ecall(SBI_EXT_HSM, SBI_EXT_HSM_HART_SUSPEND, 83 suspend_type, resume_addr, opaque, 0, 0, 0); 84 85 return (ret.error) ? sbi_err_map_linux_errno(ret.error) : 0; 86 } 87 88 static int sbi_suspend(u32 state) 89 { 90 if (state & SBI_HSM_SUSP_NON_RET_BIT) 91 return cpu_suspend(state, sbi_suspend_finisher); 92 else 93 return sbi_suspend_finisher(state, 0, 0); 94 } 95 96 static int sbi_cpuidle_enter_state(struct cpuidle_device *dev, 97 struct cpuidle_driver *drv, int idx) 98 { 99 u32 *states = __this_cpu_read(sbi_cpuidle_data.states); 100 101 return CPU_PM_CPU_IDLE_ENTER_PARAM(sbi_suspend, idx, states[idx]); 102 } 103 104 static int __sbi_enter_domain_idle_state(struct cpuidle_device *dev, 105 struct cpuidle_driver *drv, int idx, 106 bool s2idle) 107 { 108 struct sbi_cpuidle_data *data = this_cpu_ptr(&sbi_cpuidle_data); 109 u32 *states = data->states; 110 struct device *pd_dev = data->dev; 111 u32 state; 112 int ret; 113 114 ret = cpu_pm_enter(); 115 if (ret) 116 return -1; 117 118 /* Do runtime PM to manage a hierarchical CPU toplogy. */ 119 ct_irq_enter_irqson(); 120 if (s2idle) 121 dev_pm_genpd_suspend(pd_dev); 122 else 123 pm_runtime_put_sync_suspend(pd_dev); 124 ct_irq_exit_irqson(); 125 126 if (sbi_is_domain_state_available()) 127 state = sbi_get_domain_state(); 128 else 129 state = states[idx]; 130 131 ret = sbi_suspend(state) ? -1 : idx; 132 133 ct_irq_enter_irqson(); 134 if (s2idle) 135 dev_pm_genpd_resume(pd_dev); 136 else 137 pm_runtime_get_sync(pd_dev); 138 ct_irq_exit_irqson(); 139 140 cpu_pm_exit(); 141 142 /* Clear the domain state to start fresh when back from idle. */ 143 sbi_clear_domain_state(); 144 return ret; 145 } 146 147 static int sbi_enter_domain_idle_state(struct cpuidle_device *dev, 148 struct cpuidle_driver *drv, int idx) 149 { 150 return __sbi_enter_domain_idle_state(dev, drv, idx, false); 151 } 152 153 static int sbi_enter_s2idle_domain_idle_state(struct cpuidle_device *dev, 154 struct cpuidle_driver *drv, 155 int idx) 156 { 157 return __sbi_enter_domain_idle_state(dev, drv, idx, true); 158 } 159 160 static int sbi_cpuidle_cpuhp_up(unsigned int cpu) 161 { 162 struct device *pd_dev = __this_cpu_read(sbi_cpuidle_data.dev); 163 164 if (pd_dev) 165 pm_runtime_get_sync(pd_dev); 166 167 return 0; 168 } 169 170 static int sbi_cpuidle_cpuhp_down(unsigned int cpu) 171 { 172 struct device *pd_dev = __this_cpu_read(sbi_cpuidle_data.dev); 173 174 if (pd_dev) { 175 pm_runtime_put_sync(pd_dev); 176 /* Clear domain state to start fresh at next online. */ 177 sbi_clear_domain_state(); 178 } 179 180 return 0; 181 } 182 183 static void sbi_idle_init_cpuhp(void) 184 { 185 int err; 186 187 if (!sbi_cpuidle_use_cpuhp) 188 return; 189 190 err = cpuhp_setup_state_nocalls(CPUHP_AP_CPU_PM_STARTING, 191 "cpuidle/sbi:online", 192 sbi_cpuidle_cpuhp_up, 193 sbi_cpuidle_cpuhp_down); 194 if (err) 195 pr_warn("Failed %d while setup cpuhp state\n", err); 196 } 197 198 static const struct of_device_id sbi_cpuidle_state_match[] = { 199 { .compatible = "riscv,idle-state", 200 .data = sbi_cpuidle_enter_state }, 201 { }, 202 }; 203 204 static bool sbi_suspend_state_is_valid(u32 state) 205 { 206 if (state > SBI_HSM_SUSPEND_RET_DEFAULT && 207 state < SBI_HSM_SUSPEND_RET_PLATFORM) 208 return false; 209 if (state > SBI_HSM_SUSPEND_NON_RET_DEFAULT && 210 state < SBI_HSM_SUSPEND_NON_RET_PLATFORM) 211 return false; 212 return true; 213 } 214 215 static int sbi_dt_parse_state_node(struct device_node *np, u32 *state) 216 { 217 int err = of_property_read_u32(np, "riscv,sbi-suspend-param", state); 218 219 if (err) { 220 pr_warn("%pOF missing riscv,sbi-suspend-param property\n", np); 221 return err; 222 } 223 224 if (!sbi_suspend_state_is_valid(*state)) { 225 pr_warn("Invalid SBI suspend state %#x\n", *state); 226 return -EINVAL; 227 } 228 229 return 0; 230 } 231 232 static int sbi_dt_cpu_init_topology(struct cpuidle_driver *drv, 233 struct sbi_cpuidle_data *data, 234 unsigned int state_count, int cpu) 235 { 236 /* Currently limit the hierarchical topology to be used in OSI mode. */ 237 if (!sbi_cpuidle_use_osi) 238 return 0; 239 240 data->dev = dt_idle_attach_cpu(cpu, "sbi"); 241 if (IS_ERR_OR_NULL(data->dev)) 242 return PTR_ERR_OR_ZERO(data->dev); 243 244 /* 245 * Using the deepest state for the CPU to trigger a potential selection 246 * of a shared state for the domain, assumes the domain states are all 247 * deeper states. 248 */ 249 drv->states[state_count - 1].enter = sbi_enter_domain_idle_state; 250 drv->states[state_count - 1].enter_s2idle = 251 sbi_enter_s2idle_domain_idle_state; 252 sbi_cpuidle_use_cpuhp = true; 253 254 return 0; 255 } 256 257 static int sbi_cpuidle_dt_init_states(struct device *dev, 258 struct cpuidle_driver *drv, 259 unsigned int cpu, 260 unsigned int state_count) 261 { 262 struct sbi_cpuidle_data *data = per_cpu_ptr(&sbi_cpuidle_data, cpu); 263 struct device_node *state_node; 264 struct device_node *cpu_node; 265 u32 *states; 266 int i, ret; 267 268 cpu_node = of_cpu_device_node_get(cpu); 269 if (!cpu_node) 270 return -ENODEV; 271 272 states = devm_kcalloc(dev, state_count, sizeof(*states), GFP_KERNEL); 273 if (!states) { 274 ret = -ENOMEM; 275 goto fail; 276 } 277 278 /* Parse SBI specific details from state DT nodes */ 279 for (i = 1; i < state_count; i++) { 280 state_node = of_get_cpu_state_node(cpu_node, i - 1); 281 if (!state_node) 282 break; 283 284 ret = sbi_dt_parse_state_node(state_node, &states[i]); 285 of_node_put(state_node); 286 287 if (ret) 288 return ret; 289 290 pr_debug("sbi-state %#x index %d\n", states[i], i); 291 } 292 if (i != state_count) { 293 ret = -ENODEV; 294 goto fail; 295 } 296 297 /* Initialize optional data, used for the hierarchical topology. */ 298 ret = sbi_dt_cpu_init_topology(drv, data, state_count, cpu); 299 if (ret < 0) 300 return ret; 301 302 /* Store states in the per-cpu struct. */ 303 data->states = states; 304 305 fail: 306 of_node_put(cpu_node); 307 308 return ret; 309 } 310 311 static void sbi_cpuidle_deinit_cpu(int cpu) 312 { 313 struct sbi_cpuidle_data *data = per_cpu_ptr(&sbi_cpuidle_data, cpu); 314 315 dt_idle_detach_cpu(data->dev); 316 sbi_cpuidle_use_cpuhp = false; 317 } 318 319 static int sbi_cpuidle_init_cpu(struct device *dev, int cpu) 320 { 321 struct cpuidle_driver *drv; 322 unsigned int state_count = 0; 323 int ret = 0; 324 325 drv = devm_kzalloc(dev, sizeof(*drv), GFP_KERNEL); 326 if (!drv) 327 return -ENOMEM; 328 329 drv->name = "sbi_cpuidle"; 330 drv->owner = THIS_MODULE; 331 drv->cpumask = (struct cpumask *)cpumask_of(cpu); 332 333 /* RISC-V architectural WFI to be represented as state index 0. */ 334 drv->states[0].enter = sbi_cpuidle_enter_state; 335 drv->states[0].exit_latency = 1; 336 drv->states[0].target_residency = 1; 337 drv->states[0].power_usage = UINT_MAX; 338 strcpy(drv->states[0].name, "WFI"); 339 strcpy(drv->states[0].desc, "RISC-V WFI"); 340 341 /* 342 * If no DT idle states are detected (ret == 0) let the driver 343 * initialization fail accordingly since there is no reason to 344 * initialize the idle driver if only wfi is supported, the 345 * default archictectural back-end already executes wfi 346 * on idle entry. 347 */ 348 ret = dt_init_idle_driver(drv, sbi_cpuidle_state_match, 1); 349 if (ret <= 0) { 350 pr_debug("HART%ld: failed to parse DT idle states\n", 351 cpuid_to_hartid_map(cpu)); 352 return ret ? : -ENODEV; 353 } 354 state_count = ret + 1; /* Include WFI state as well */ 355 356 /* Initialize idle states from DT. */ 357 ret = sbi_cpuidle_dt_init_states(dev, drv, cpu, state_count); 358 if (ret) { 359 pr_err("HART%ld: failed to init idle states\n", 360 cpuid_to_hartid_map(cpu)); 361 return ret; 362 } 363 364 ret = cpuidle_register(drv, NULL); 365 if (ret) 366 goto deinit; 367 368 cpuidle_cooling_register(drv); 369 370 return 0; 371 deinit: 372 sbi_cpuidle_deinit_cpu(cpu); 373 return ret; 374 } 375 376 static void sbi_cpuidle_domain_sync_state(struct device *dev) 377 { 378 /* 379 * All devices have now been attached/probed to the PM domain 380 * topology, hence it's fine to allow domain states to be picked. 381 */ 382 sbi_cpuidle_pd_allow_domain_state = true; 383 } 384 385 #ifdef CONFIG_DT_IDLE_GENPD 386 387 static int sbi_cpuidle_pd_power_off(struct generic_pm_domain *pd) 388 { 389 struct genpd_power_state *state = &pd->states[pd->state_idx]; 390 u32 *pd_state; 391 392 if (!state->data) 393 return 0; 394 395 if (!sbi_cpuidle_pd_allow_domain_state) 396 return -EBUSY; 397 398 /* OSI mode is enabled, set the corresponding domain state. */ 399 pd_state = state->data; 400 sbi_set_domain_state(*pd_state); 401 402 return 0; 403 } 404 405 struct sbi_pd_provider { 406 struct list_head link; 407 struct device_node *node; 408 }; 409 410 static LIST_HEAD(sbi_pd_providers); 411 412 static int sbi_pd_init(struct device_node *np) 413 { 414 struct generic_pm_domain *pd; 415 struct sbi_pd_provider *pd_provider; 416 struct dev_power_governor *pd_gov; 417 int ret = -ENOMEM; 418 419 pd = dt_idle_pd_alloc(np, sbi_dt_parse_state_node); 420 if (!pd) 421 goto out; 422 423 pd_provider = kzalloc(sizeof(*pd_provider), GFP_KERNEL); 424 if (!pd_provider) 425 goto free_pd; 426 427 pd->flags |= GENPD_FLAG_IRQ_SAFE | GENPD_FLAG_CPU_DOMAIN; 428 429 /* Allow power off when OSI is available. */ 430 if (sbi_cpuidle_use_osi) 431 pd->power_off = sbi_cpuidle_pd_power_off; 432 else 433 pd->flags |= GENPD_FLAG_ALWAYS_ON; 434 435 /* Use governor for CPU PM domains if it has some states to manage. */ 436 pd_gov = pd->states ? &pm_domain_cpu_gov : NULL; 437 438 ret = pm_genpd_init(pd, pd_gov, false); 439 if (ret) 440 goto free_pd_prov; 441 442 ret = of_genpd_add_provider_simple(np, pd); 443 if (ret) 444 goto remove_pd; 445 446 pd_provider->node = of_node_get(np); 447 list_add(&pd_provider->link, &sbi_pd_providers); 448 449 pr_debug("init PM domain %s\n", pd->name); 450 return 0; 451 452 remove_pd: 453 pm_genpd_remove(pd); 454 free_pd_prov: 455 kfree(pd_provider); 456 free_pd: 457 dt_idle_pd_free(pd); 458 out: 459 pr_err("failed to init PM domain ret=%d %pOF\n", ret, np); 460 return ret; 461 } 462 463 static void sbi_pd_remove(void) 464 { 465 struct sbi_pd_provider *pd_provider, *it; 466 struct generic_pm_domain *genpd; 467 468 list_for_each_entry_safe(pd_provider, it, &sbi_pd_providers, link) { 469 of_genpd_del_provider(pd_provider->node); 470 471 genpd = of_genpd_remove_last(pd_provider->node); 472 if (!IS_ERR(genpd)) 473 kfree(genpd); 474 475 of_node_put(pd_provider->node); 476 list_del(&pd_provider->link); 477 kfree(pd_provider); 478 } 479 } 480 481 static int sbi_genpd_probe(struct device_node *np) 482 { 483 struct device_node *node; 484 int ret = 0, pd_count = 0; 485 486 if (!np) 487 return -ENODEV; 488 489 /* 490 * Parse child nodes for the "#power-domain-cells" property and 491 * initialize a genpd/genpd-of-provider pair when it's found. 492 */ 493 for_each_child_of_node(np, node) { 494 if (!of_find_property(node, "#power-domain-cells", NULL)) 495 continue; 496 497 ret = sbi_pd_init(node); 498 if (ret) 499 goto put_node; 500 501 pd_count++; 502 } 503 504 /* Bail out if not using the hierarchical CPU topology. */ 505 if (!pd_count) 506 goto no_pd; 507 508 /* Link genpd masters/subdomains to model the CPU topology. */ 509 ret = dt_idle_pd_init_topology(np); 510 if (ret) 511 goto remove_pd; 512 513 return 0; 514 515 put_node: 516 of_node_put(node); 517 remove_pd: 518 sbi_pd_remove(); 519 pr_err("failed to create CPU PM domains ret=%d\n", ret); 520 no_pd: 521 return ret; 522 } 523 524 #else 525 526 static inline int sbi_genpd_probe(struct device_node *np) 527 { 528 return 0; 529 } 530 531 #endif 532 533 static int sbi_cpuidle_probe(struct platform_device *pdev) 534 { 535 int cpu, ret; 536 struct cpuidle_driver *drv; 537 struct cpuidle_device *dev; 538 struct device_node *np, *pds_node; 539 540 /* Detect OSI support based on CPU DT nodes */ 541 sbi_cpuidle_use_osi = true; 542 for_each_possible_cpu(cpu) { 543 np = of_cpu_device_node_get(cpu); 544 if (np && 545 of_find_property(np, "power-domains", NULL) && 546 of_find_property(np, "power-domain-names", NULL)) { 547 continue; 548 } else { 549 sbi_cpuidle_use_osi = false; 550 break; 551 } 552 } 553 554 /* Populate generic power domains from DT nodes */ 555 pds_node = of_find_node_by_path("/cpus/power-domains"); 556 if (pds_node) { 557 ret = sbi_genpd_probe(pds_node); 558 of_node_put(pds_node); 559 if (ret) 560 return ret; 561 } 562 563 /* Initialize CPU idle driver for each CPU */ 564 for_each_possible_cpu(cpu) { 565 ret = sbi_cpuidle_init_cpu(&pdev->dev, cpu); 566 if (ret) { 567 pr_debug("HART%ld: idle driver init failed\n", 568 cpuid_to_hartid_map(cpu)); 569 goto out_fail; 570 } 571 } 572 573 /* Setup CPU hotplut notifiers */ 574 sbi_idle_init_cpuhp(); 575 576 pr_info("idle driver registered for all CPUs\n"); 577 578 return 0; 579 580 out_fail: 581 while (--cpu >= 0) { 582 dev = per_cpu(cpuidle_devices, cpu); 583 drv = cpuidle_get_cpu_driver(dev); 584 cpuidle_unregister(drv); 585 sbi_cpuidle_deinit_cpu(cpu); 586 } 587 588 return ret; 589 } 590 591 static struct platform_driver sbi_cpuidle_driver = { 592 .probe = sbi_cpuidle_probe, 593 .driver = { 594 .name = "sbi-cpuidle", 595 .sync_state = sbi_cpuidle_domain_sync_state, 596 }, 597 }; 598 599 static int __init sbi_cpuidle_init(void) 600 { 601 int ret; 602 struct platform_device *pdev; 603 604 /* 605 * The SBI HSM suspend function is only available when: 606 * 1) SBI version is 0.3 or higher 607 * 2) SBI HSM extension is available 608 */ 609 if ((sbi_spec_version < sbi_mk_version(0, 3)) || 610 sbi_probe_extension(SBI_EXT_HSM) <= 0) { 611 pr_info("HSM suspend not available\n"); 612 return 0; 613 } 614 615 ret = platform_driver_register(&sbi_cpuidle_driver); 616 if (ret) 617 return ret; 618 619 pdev = platform_device_register_simple("sbi-cpuidle", 620 -1, NULL, 0); 621 if (IS_ERR(pdev)) { 622 platform_driver_unregister(&sbi_cpuidle_driver); 623 return PTR_ERR(pdev); 624 } 625 626 return 0; 627 } 628 device_initcall(sbi_cpuidle_init); 629