1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (c) 2020 - 2022, NVIDIA CORPORATION. All rights reserved 4 */ 5 6 #include <linux/cpu.h> 7 #include <linux/cpufreq.h> 8 #include <linux/delay.h> 9 #include <linux/dma-mapping.h> 10 #include <linux/module.h> 11 #include <linux/of.h> 12 #include <linux/of_platform.h> 13 #include <linux/platform_device.h> 14 #include <linux/slab.h> 15 16 #include <asm/smp_plat.h> 17 18 #include <soc/tegra/bpmp.h> 19 #include <soc/tegra/bpmp-abi.h> 20 21 #define KHZ 1000 22 #define REF_CLK_MHZ 408 /* 408 MHz */ 23 #define US_DELAY 500 24 #define CPUFREQ_TBL_STEP_HZ (50 * KHZ * KHZ) 25 #define MAX_CNT ~0U 26 27 #define NDIV_MASK 0x1FF 28 29 #define CORE_OFFSET(cpu) (cpu * 8) 30 #define CMU_CLKS_BASE 0x2000 31 #define SCRATCH_FREQ_CORE_REG(data, cpu) (data->regs + CMU_CLKS_BASE + CORE_OFFSET(cpu)) 32 33 #define MMCRAB_CLUSTER_BASE(cl) (0x30000 + (cl * 0x10000)) 34 #define CLUSTER_ACTMON_BASE(data, cl) \ 35 (data->regs + (MMCRAB_CLUSTER_BASE(cl) + data->soc->actmon_cntr_base)) 36 #define CORE_ACTMON_CNTR_REG(data, cl, cpu) (CLUSTER_ACTMON_BASE(data, cl) + CORE_OFFSET(cpu)) 37 38 /* cpufreq transisition latency */ 39 #define TEGRA_CPUFREQ_TRANSITION_LATENCY (300 * 1000) /* unit in nanoseconds */ 40 41 enum cluster { 42 CLUSTER0, 43 CLUSTER1, 44 CLUSTER2, 45 CLUSTER3, 46 MAX_CLUSTERS, 47 }; 48 49 struct tegra_cpu_ctr { 50 u32 cpu; 51 u32 coreclk_cnt, last_coreclk_cnt; 52 u32 refclk_cnt, last_refclk_cnt; 53 }; 54 55 struct read_counters_work { 56 struct work_struct work; 57 struct tegra_cpu_ctr c; 58 }; 59 60 struct tegra_cpufreq_ops { 61 void (*read_counters)(struct tegra_cpu_ctr *c); 62 void (*set_cpu_ndiv)(struct cpufreq_policy *policy, u64 ndiv); 63 void (*get_cpu_cluster_id)(u32 cpu, u32 *cpuid, u32 *clusterid); 64 int (*get_cpu_ndiv)(u32 cpu, u32 cpuid, u32 clusterid, u64 *ndiv); 65 }; 66 67 struct tegra_cpufreq_soc { 68 struct tegra_cpufreq_ops *ops; 69 int maxcpus_per_cluster; 70 phys_addr_t actmon_cntr_base; 71 }; 72 73 struct tegra194_cpufreq_data { 74 void __iomem *regs; 75 size_t num_clusters; 76 struct cpufreq_frequency_table **tables; 77 const struct tegra_cpufreq_soc *soc; 78 }; 79 80 static struct workqueue_struct *read_counters_wq; 81 82 static void tegra_get_cpu_mpidr(void *mpidr) 83 { 84 *((u64 *)mpidr) = read_cpuid_mpidr() & MPIDR_HWID_BITMASK; 85 } 86 87 static void tegra234_get_cpu_cluster_id(u32 cpu, u32 *cpuid, u32 *clusterid) 88 { 89 u64 mpidr; 90 91 smp_call_function_single(cpu, tegra_get_cpu_mpidr, &mpidr, true); 92 93 if (cpuid) 94 *cpuid = MPIDR_AFFINITY_LEVEL(mpidr, 1); 95 if (clusterid) 96 *clusterid = MPIDR_AFFINITY_LEVEL(mpidr, 2); 97 } 98 99 static int tegra234_get_cpu_ndiv(u32 cpu, u32 cpuid, u32 clusterid, u64 *ndiv) 100 { 101 struct tegra194_cpufreq_data *data = cpufreq_get_driver_data(); 102 void __iomem *freq_core_reg; 103 u64 mpidr_id; 104 105 /* use physical id to get address of per core frequency register */ 106 mpidr_id = (clusterid * data->soc->maxcpus_per_cluster) + cpuid; 107 freq_core_reg = SCRATCH_FREQ_CORE_REG(data, mpidr_id); 108 109 *ndiv = readl(freq_core_reg) & NDIV_MASK; 110 111 return 0; 112 } 113 114 static void tegra234_set_cpu_ndiv(struct cpufreq_policy *policy, u64 ndiv) 115 { 116 struct tegra194_cpufreq_data *data = cpufreq_get_driver_data(); 117 void __iomem *freq_core_reg; 118 u32 cpu, cpuid, clusterid; 119 u64 mpidr_id; 120 121 for_each_cpu_and(cpu, policy->cpus, cpu_online_mask) { 122 data->soc->ops->get_cpu_cluster_id(cpu, &cpuid, &clusterid); 123 124 /* use physical id to get address of per core frequency register */ 125 mpidr_id = (clusterid * data->soc->maxcpus_per_cluster) + cpuid; 126 freq_core_reg = SCRATCH_FREQ_CORE_REG(data, mpidr_id); 127 128 writel(ndiv, freq_core_reg); 129 } 130 } 131 132 /* 133 * This register provides access to two counter values with a single 134 * 64-bit read. The counter values are used to determine the average 135 * actual frequency a core has run at over a period of time. 136 * [63:32] PLLP counter: Counts at fixed frequency (408 MHz) 137 * [31:0] Core clock counter: Counts on every core clock cycle 138 */ 139 static void tegra234_read_counters(struct tegra_cpu_ctr *c) 140 { 141 struct tegra194_cpufreq_data *data = cpufreq_get_driver_data(); 142 void __iomem *actmon_reg; 143 u32 cpuid, clusterid; 144 u64 val; 145 146 data->soc->ops->get_cpu_cluster_id(c->cpu, &cpuid, &clusterid); 147 actmon_reg = CORE_ACTMON_CNTR_REG(data, clusterid, cpuid); 148 149 val = readq(actmon_reg); 150 c->last_refclk_cnt = upper_32_bits(val); 151 c->last_coreclk_cnt = lower_32_bits(val); 152 udelay(US_DELAY); 153 val = readq(actmon_reg); 154 c->refclk_cnt = upper_32_bits(val); 155 c->coreclk_cnt = lower_32_bits(val); 156 } 157 158 static struct tegra_cpufreq_ops tegra234_cpufreq_ops = { 159 .read_counters = tegra234_read_counters, 160 .get_cpu_cluster_id = tegra234_get_cpu_cluster_id, 161 .get_cpu_ndiv = tegra234_get_cpu_ndiv, 162 .set_cpu_ndiv = tegra234_set_cpu_ndiv, 163 }; 164 165 static const struct tegra_cpufreq_soc tegra234_cpufreq_soc = { 166 .ops = &tegra234_cpufreq_ops, 167 .actmon_cntr_base = 0x9000, 168 .maxcpus_per_cluster = 4, 169 }; 170 171 static void tegra194_get_cpu_cluster_id(u32 cpu, u32 *cpuid, u32 *clusterid) 172 { 173 u64 mpidr; 174 175 smp_call_function_single(cpu, tegra_get_cpu_mpidr, &mpidr, true); 176 177 if (cpuid) 178 *cpuid = MPIDR_AFFINITY_LEVEL(mpidr, 0); 179 if (clusterid) 180 *clusterid = MPIDR_AFFINITY_LEVEL(mpidr, 1); 181 } 182 183 /* 184 * Read per-core Read-only system register NVFREQ_FEEDBACK_EL1. 185 * The register provides frequency feedback information to 186 * determine the average actual frequency a core has run at over 187 * a period of time. 188 * [31:0] PLLP counter: Counts at fixed frequency (408 MHz) 189 * [63:32] Core clock counter: counts on every core clock cycle 190 * where the core is architecturally clocking 191 */ 192 static u64 read_freq_feedback(void) 193 { 194 u64 val = 0; 195 196 asm volatile("mrs %0, s3_0_c15_c0_5" : "=r" (val) : ); 197 198 return val; 199 } 200 201 static inline u32 map_ndiv_to_freq(struct mrq_cpu_ndiv_limits_response 202 *nltbl, u16 ndiv) 203 { 204 return nltbl->ref_clk_hz / KHZ * ndiv / (nltbl->pdiv * nltbl->mdiv); 205 } 206 207 static void tegra194_read_counters(struct tegra_cpu_ctr *c) 208 { 209 u64 val; 210 211 val = read_freq_feedback(); 212 c->last_refclk_cnt = lower_32_bits(val); 213 c->last_coreclk_cnt = upper_32_bits(val); 214 udelay(US_DELAY); 215 val = read_freq_feedback(); 216 c->refclk_cnt = lower_32_bits(val); 217 c->coreclk_cnt = upper_32_bits(val); 218 } 219 220 static void tegra_read_counters(struct work_struct *work) 221 { 222 struct tegra194_cpufreq_data *data = cpufreq_get_driver_data(); 223 struct read_counters_work *read_counters_work; 224 struct tegra_cpu_ctr *c; 225 226 /* 227 * ref_clk_counter(32 bit counter) runs on constant clk, 228 * pll_p(408MHz). 229 * It will take = 2 ^ 32 / 408 MHz to overflow ref clk counter 230 * = 10526880 usec = 10.527 sec to overflow 231 * 232 * Like wise core_clk_counter(32 bit counter) runs on core clock. 233 * It's synchronized to crab_clk (cpu_crab_clk) which runs at 234 * freq of cluster. Assuming max cluster clock ~2000MHz, 235 * It will take = 2 ^ 32 / 2000 MHz to overflow core clk counter 236 * = ~2.147 sec to overflow 237 */ 238 read_counters_work = container_of(work, struct read_counters_work, 239 work); 240 c = &read_counters_work->c; 241 242 data->soc->ops->read_counters(c); 243 } 244 245 /* 246 * Return instantaneous cpu speed 247 * Instantaneous freq is calculated as - 248 * -Takes sample on every query of getting the freq. 249 * - Read core and ref clock counters; 250 * - Delay for X us 251 * - Read above cycle counters again 252 * - Calculates freq by subtracting current and previous counters 253 * divided by the delay time or eqv. of ref_clk_counter in delta time 254 * - Return Kcycles/second, freq in KHz 255 * 256 * delta time period = x sec 257 * = delta ref_clk_counter / (408 * 10^6) sec 258 * freq in Hz = cycles/sec 259 * = (delta cycles / x sec 260 * = (delta cycles * 408 * 10^6) / delta ref_clk_counter 261 * in KHz = (delta cycles * 408 * 10^3) / delta ref_clk_counter 262 * 263 * @cpu - logical cpu whose freq to be updated 264 * Returns freq in KHz on success, 0 if cpu is offline 265 */ 266 static unsigned int tegra194_calculate_speed(u32 cpu) 267 { 268 struct read_counters_work read_counters_work; 269 struct tegra_cpu_ctr c; 270 u32 delta_refcnt; 271 u32 delta_ccnt; 272 u32 rate_mhz; 273 274 /* 275 * udelay() is required to reconstruct cpu frequency over an 276 * observation window. Using workqueue to call udelay() with 277 * interrupts enabled. 278 */ 279 read_counters_work.c.cpu = cpu; 280 INIT_WORK_ONSTACK(&read_counters_work.work, tegra_read_counters); 281 queue_work_on(cpu, read_counters_wq, &read_counters_work.work); 282 flush_work(&read_counters_work.work); 283 c = read_counters_work.c; 284 285 if (c.coreclk_cnt < c.last_coreclk_cnt) 286 delta_ccnt = c.coreclk_cnt + (MAX_CNT - c.last_coreclk_cnt); 287 else 288 delta_ccnt = c.coreclk_cnt - c.last_coreclk_cnt; 289 if (!delta_ccnt) 290 return 0; 291 292 /* ref clock is 32 bits */ 293 if (c.refclk_cnt < c.last_refclk_cnt) 294 delta_refcnt = c.refclk_cnt + (MAX_CNT - c.last_refclk_cnt); 295 else 296 delta_refcnt = c.refclk_cnt - c.last_refclk_cnt; 297 if (!delta_refcnt) { 298 pr_debug("cpufreq: %d is idle, delta_refcnt: 0\n", cpu); 299 return 0; 300 } 301 rate_mhz = ((unsigned long)(delta_ccnt * REF_CLK_MHZ)) / delta_refcnt; 302 303 return (rate_mhz * KHZ); /* in KHz */ 304 } 305 306 static void tegra194_get_cpu_ndiv_sysreg(void *ndiv) 307 { 308 u64 ndiv_val; 309 310 asm volatile("mrs %0, s3_0_c15_c0_4" : "=r" (ndiv_val) : ); 311 312 *(u64 *)ndiv = ndiv_val; 313 } 314 315 static int tegra194_get_cpu_ndiv(u32 cpu, u32 cpuid, u32 clusterid, u64 *ndiv) 316 { 317 int ret; 318 319 ret = smp_call_function_single(cpu, tegra194_get_cpu_ndiv_sysreg, &ndiv, true); 320 321 return ret; 322 } 323 324 static void tegra194_set_cpu_ndiv_sysreg(void *data) 325 { 326 u64 ndiv_val = *(u64 *)data; 327 328 asm volatile("msr s3_0_c15_c0_4, %0" : : "r" (ndiv_val)); 329 } 330 331 static void tegra194_set_cpu_ndiv(struct cpufreq_policy *policy, u64 ndiv) 332 { 333 on_each_cpu_mask(policy->cpus, tegra194_set_cpu_ndiv_sysreg, &ndiv, true); 334 } 335 336 static unsigned int tegra194_get_speed(u32 cpu) 337 { 338 struct tegra194_cpufreq_data *data = cpufreq_get_driver_data(); 339 struct cpufreq_frequency_table *pos; 340 u32 cpuid, clusterid; 341 unsigned int rate; 342 u64 ndiv; 343 int ret; 344 345 data->soc->ops->get_cpu_cluster_id(cpu, &cpuid, &clusterid); 346 347 /* reconstruct actual cpu freq using counters */ 348 rate = tegra194_calculate_speed(cpu); 349 350 /* get last written ndiv value */ 351 ret = data->soc->ops->get_cpu_ndiv(cpu, cpuid, clusterid, &ndiv); 352 if (WARN_ON_ONCE(ret)) 353 return rate; 354 355 /* 356 * If the reconstructed frequency has acceptable delta from 357 * the last written value, then return freq corresponding 358 * to the last written ndiv value from freq_table. This is 359 * done to return consistent value. 360 */ 361 cpufreq_for_each_valid_entry(pos, data->tables[clusterid]) { 362 if (pos->driver_data != ndiv) 363 continue; 364 365 if (abs(pos->frequency - rate) > 115200) { 366 pr_warn("cpufreq: cpu%d,cur:%u,set:%u,set ndiv:%llu\n", 367 cpu, rate, pos->frequency, ndiv); 368 } else { 369 rate = pos->frequency; 370 } 371 break; 372 } 373 return rate; 374 } 375 376 static int tegra194_cpufreq_init(struct cpufreq_policy *policy) 377 { 378 struct tegra194_cpufreq_data *data = cpufreq_get_driver_data(); 379 int maxcpus_per_cluster = data->soc->maxcpus_per_cluster; 380 u32 start_cpu, cpu; 381 u32 clusterid; 382 383 data->soc->ops->get_cpu_cluster_id(policy->cpu, NULL, &clusterid); 384 385 if (clusterid >= data->num_clusters || !data->tables[clusterid]) 386 return -EINVAL; 387 388 start_cpu = rounddown(policy->cpu, maxcpus_per_cluster); 389 /* set same policy for all cpus in a cluster */ 390 for (cpu = start_cpu; cpu < (start_cpu + maxcpus_per_cluster); cpu++) { 391 if (cpu_possible(cpu)) 392 cpumask_set_cpu(cpu, policy->cpus); 393 } 394 policy->freq_table = data->tables[clusterid]; 395 policy->cpuinfo.transition_latency = TEGRA_CPUFREQ_TRANSITION_LATENCY; 396 397 return 0; 398 } 399 400 static int tegra194_cpufreq_set_target(struct cpufreq_policy *policy, 401 unsigned int index) 402 { 403 struct cpufreq_frequency_table *tbl = policy->freq_table + index; 404 struct tegra194_cpufreq_data *data = cpufreq_get_driver_data(); 405 406 /* 407 * Each core writes frequency in per core register. Then both cores 408 * in a cluster run at same frequency which is the maximum frequency 409 * request out of the values requested by both cores in that cluster. 410 */ 411 data->soc->ops->set_cpu_ndiv(policy, (u64)tbl->driver_data); 412 413 return 0; 414 } 415 416 static struct cpufreq_driver tegra194_cpufreq_driver = { 417 .name = "tegra194", 418 .flags = CPUFREQ_CONST_LOOPS | CPUFREQ_NEED_INITIAL_FREQ_CHECK, 419 .verify = cpufreq_generic_frequency_table_verify, 420 .target_index = tegra194_cpufreq_set_target, 421 .get = tegra194_get_speed, 422 .init = tegra194_cpufreq_init, 423 .attr = cpufreq_generic_attr, 424 }; 425 426 static struct tegra_cpufreq_ops tegra194_cpufreq_ops = { 427 .read_counters = tegra194_read_counters, 428 .get_cpu_cluster_id = tegra194_get_cpu_cluster_id, 429 .get_cpu_ndiv = tegra194_get_cpu_ndiv, 430 .set_cpu_ndiv = tegra194_set_cpu_ndiv, 431 }; 432 433 static const struct tegra_cpufreq_soc tegra194_cpufreq_soc = { 434 .ops = &tegra194_cpufreq_ops, 435 .maxcpus_per_cluster = 2, 436 }; 437 438 static void tegra194_cpufreq_free_resources(void) 439 { 440 destroy_workqueue(read_counters_wq); 441 } 442 443 static struct cpufreq_frequency_table * 444 init_freq_table(struct platform_device *pdev, struct tegra_bpmp *bpmp, 445 unsigned int cluster_id) 446 { 447 struct cpufreq_frequency_table *freq_table; 448 struct mrq_cpu_ndiv_limits_response resp; 449 unsigned int num_freqs, ndiv, delta_ndiv; 450 struct mrq_cpu_ndiv_limits_request req; 451 struct tegra_bpmp_message msg; 452 u16 freq_table_step_size; 453 int err, index; 454 455 memset(&req, 0, sizeof(req)); 456 req.cluster_id = cluster_id; 457 458 memset(&msg, 0, sizeof(msg)); 459 msg.mrq = MRQ_CPU_NDIV_LIMITS; 460 msg.tx.data = &req; 461 msg.tx.size = sizeof(req); 462 msg.rx.data = &resp; 463 msg.rx.size = sizeof(resp); 464 465 err = tegra_bpmp_transfer(bpmp, &msg); 466 if (err) 467 return ERR_PTR(err); 468 if (msg.rx.ret == -BPMP_EINVAL) { 469 /* Cluster not available */ 470 return NULL; 471 } 472 if (msg.rx.ret) 473 return ERR_PTR(-EINVAL); 474 475 /* 476 * Make sure frequency table step is a multiple of mdiv to match 477 * vhint table granularity. 478 */ 479 freq_table_step_size = resp.mdiv * 480 DIV_ROUND_UP(CPUFREQ_TBL_STEP_HZ, resp.ref_clk_hz); 481 482 dev_dbg(&pdev->dev, "cluster %d: frequency table step size: %d\n", 483 cluster_id, freq_table_step_size); 484 485 delta_ndiv = resp.ndiv_max - resp.ndiv_min; 486 487 if (unlikely(delta_ndiv == 0)) { 488 num_freqs = 1; 489 } else { 490 /* We store both ndiv_min and ndiv_max hence the +1 */ 491 num_freqs = delta_ndiv / freq_table_step_size + 1; 492 } 493 494 num_freqs += (delta_ndiv % freq_table_step_size) ? 1 : 0; 495 496 freq_table = devm_kcalloc(&pdev->dev, num_freqs + 1, 497 sizeof(*freq_table), GFP_KERNEL); 498 if (!freq_table) 499 return ERR_PTR(-ENOMEM); 500 501 for (index = 0, ndiv = resp.ndiv_min; 502 ndiv < resp.ndiv_max; 503 index++, ndiv += freq_table_step_size) { 504 freq_table[index].driver_data = ndiv; 505 freq_table[index].frequency = map_ndiv_to_freq(&resp, ndiv); 506 } 507 508 freq_table[index].driver_data = resp.ndiv_max; 509 freq_table[index++].frequency = map_ndiv_to_freq(&resp, resp.ndiv_max); 510 freq_table[index].frequency = CPUFREQ_TABLE_END; 511 512 return freq_table; 513 } 514 515 static int tegra194_cpufreq_probe(struct platform_device *pdev) 516 { 517 const struct tegra_cpufreq_soc *soc; 518 struct tegra194_cpufreq_data *data; 519 struct tegra_bpmp *bpmp; 520 int err, i; 521 522 data = devm_kzalloc(&pdev->dev, sizeof(*data), GFP_KERNEL); 523 if (!data) 524 return -ENOMEM; 525 526 soc = of_device_get_match_data(&pdev->dev); 527 528 if (soc->ops && soc->maxcpus_per_cluster) { 529 data->soc = soc; 530 } else { 531 dev_err(&pdev->dev, "soc data missing\n"); 532 return -EINVAL; 533 } 534 535 data->num_clusters = MAX_CLUSTERS; 536 data->tables = devm_kcalloc(&pdev->dev, data->num_clusters, 537 sizeof(*data->tables), GFP_KERNEL); 538 if (!data->tables) 539 return -ENOMEM; 540 541 if (soc->actmon_cntr_base) { 542 /* mmio registers are used for frequency request and re-construction */ 543 data->regs = devm_platform_ioremap_resource(pdev, 0); 544 if (IS_ERR(data->regs)) 545 return PTR_ERR(data->regs); 546 } 547 548 platform_set_drvdata(pdev, data); 549 550 bpmp = tegra_bpmp_get(&pdev->dev); 551 if (IS_ERR(bpmp)) 552 return PTR_ERR(bpmp); 553 554 read_counters_wq = alloc_workqueue("read_counters_wq", __WQ_LEGACY, 1); 555 if (!read_counters_wq) { 556 dev_err(&pdev->dev, "fail to create_workqueue\n"); 557 err = -EINVAL; 558 goto put_bpmp; 559 } 560 561 for (i = 0; i < data->num_clusters; i++) { 562 data->tables[i] = init_freq_table(pdev, bpmp, i); 563 if (IS_ERR(data->tables[i])) { 564 err = PTR_ERR(data->tables[i]); 565 goto err_free_res; 566 } 567 } 568 569 tegra194_cpufreq_driver.driver_data = data; 570 571 err = cpufreq_register_driver(&tegra194_cpufreq_driver); 572 if (!err) 573 goto put_bpmp; 574 575 err_free_res: 576 tegra194_cpufreq_free_resources(); 577 put_bpmp: 578 tegra_bpmp_put(bpmp); 579 return err; 580 } 581 582 static int tegra194_cpufreq_remove(struct platform_device *pdev) 583 { 584 cpufreq_unregister_driver(&tegra194_cpufreq_driver); 585 tegra194_cpufreq_free_resources(); 586 587 return 0; 588 } 589 590 static const struct of_device_id tegra194_cpufreq_of_match[] = { 591 { .compatible = "nvidia,tegra194-ccplex", .data = &tegra194_cpufreq_soc }, 592 { .compatible = "nvidia,tegra234-ccplex-cluster", .data = &tegra234_cpufreq_soc }, 593 { /* sentinel */ } 594 }; 595 596 static struct platform_driver tegra194_ccplex_driver = { 597 .driver = { 598 .name = "tegra194-cpufreq", 599 .of_match_table = tegra194_cpufreq_of_match, 600 }, 601 .probe = tegra194_cpufreq_probe, 602 .remove = tegra194_cpufreq_remove, 603 }; 604 module_platform_driver(tegra194_ccplex_driver); 605 606 MODULE_AUTHOR("Mikko Perttunen <mperttunen@nvidia.com>"); 607 MODULE_AUTHOR("Sumit Gupta <sumitg@nvidia.com>"); 608 MODULE_DESCRIPTION("NVIDIA Tegra194 cpufreq driver"); 609 MODULE_LICENSE("GPL v2"); 610