1 /* 2 * CPPC (Collaborative Processor Performance Control) methods used by CPUfreq drivers. 3 * 4 * (C) Copyright 2014, 2015 Linaro Ltd. 5 * Author: Ashwin Chaugule <ashwin.chaugule@linaro.org> 6 * 7 * This program is free software; you can redistribute it and/or 8 * modify it under the terms of the GNU General Public License 9 * as published by the Free Software Foundation; version 2 10 * of the License. 11 * 12 * CPPC describes a few methods for controlling CPU performance using 13 * information from a per CPU table called CPC. This table is described in 14 * the ACPI v5.0+ specification. The table consists of a list of 15 * registers which may be memory mapped or hardware registers and also may 16 * include some static integer values. 17 * 18 * CPU performance is on an abstract continuous scale as against a discretized 19 * P-state scale which is tied to CPU frequency only. In brief, the basic 20 * operation involves: 21 * 22 * - OS makes a CPU performance request. (Can provide min and max bounds) 23 * 24 * - Platform (such as BMC) is free to optimize request within requested bounds 25 * depending on power/thermal budgets etc. 26 * 27 * - Platform conveys its decision back to OS 28 * 29 * The communication between OS and platform occurs through another medium 30 * called (PCC) Platform Communication Channel. This is a generic mailbox like 31 * mechanism which includes doorbell semantics to indicate register updates. 32 * See drivers/mailbox/pcc.c for details on PCC. 33 * 34 * Finer details about the PCC and CPPC spec are available in the ACPI v5.1 and 35 * above specifications. 36 */ 37 38 #define pr_fmt(fmt) "ACPI CPPC: " fmt 39 40 #include <linux/cpufreq.h> 41 #include <linux/delay.h> 42 #include <linux/ktime.h> 43 #include <linux/rwsem.h> 44 #include <linux/wait.h> 45 46 #include <acpi/cppc_acpi.h> 47 48 struct cppc_pcc_data { 49 struct mbox_chan *pcc_channel; 50 void __iomem *pcc_comm_addr; 51 int pcc_subspace_idx; 52 bool pcc_channel_acquired; 53 ktime_t deadline; 54 unsigned int pcc_mpar, pcc_mrtt, pcc_nominal; 55 56 bool pending_pcc_write_cmd; /* Any pending/batched PCC write cmds? */ 57 bool platform_owns_pcc; /* Ownership of PCC subspace */ 58 unsigned int pcc_write_cnt; /* Running count of PCC write commands */ 59 60 /* 61 * Lock to provide controlled access to the PCC channel. 62 * 63 * For performance critical usecases(currently cppc_set_perf) 64 * We need to take read_lock and check if channel belongs to OSPM 65 * before reading or writing to PCC subspace 66 * We need to take write_lock before transferring the channel 67 * ownership to the platform via a Doorbell 68 * This allows us to batch a number of CPPC requests if they happen 69 * to originate in about the same time 70 * 71 * For non-performance critical usecases(init) 72 * Take write_lock for all purposes which gives exclusive access 73 */ 74 struct rw_semaphore pcc_lock; 75 76 /* Wait queue for CPUs whose requests were batched */ 77 wait_queue_head_t pcc_write_wait_q; 78 }; 79 80 /* Structure to represent the single PCC channel */ 81 static struct cppc_pcc_data pcc_data = { 82 .pcc_subspace_idx = -1, 83 .platform_owns_pcc = true, 84 }; 85 86 /* 87 * The cpc_desc structure contains the ACPI register details 88 * as described in the per CPU _CPC tables. The details 89 * include the type of register (e.g. PCC, System IO, FFH etc.) 90 * and destination addresses which lets us READ/WRITE CPU performance 91 * information using the appropriate I/O methods. 92 */ 93 static DEFINE_PER_CPU(struct cpc_desc *, cpc_desc_ptr); 94 95 /* pcc mapped address + header size + offset within PCC subspace */ 96 #define GET_PCC_VADDR(offs) (pcc_data.pcc_comm_addr + 0x8 + (offs)) 97 98 /* Check if a CPC regsiter is in PCC */ 99 #define CPC_IN_PCC(cpc) ((cpc)->type == ACPI_TYPE_BUFFER && \ 100 (cpc)->cpc_entry.reg.space_id == \ 101 ACPI_ADR_SPACE_PLATFORM_COMM) 102 103 /* Evalutes to True if reg is a NULL register descriptor */ 104 #define IS_NULL_REG(reg) ((reg)->space_id == ACPI_ADR_SPACE_SYSTEM_MEMORY && \ 105 (reg)->address == 0 && \ 106 (reg)->bit_width == 0 && \ 107 (reg)->bit_offset == 0 && \ 108 (reg)->access_width == 0) 109 110 /* Evalutes to True if an optional cpc field is supported */ 111 #define CPC_SUPPORTED(cpc) ((cpc)->type == ACPI_TYPE_INTEGER ? \ 112 !!(cpc)->cpc_entry.int_value : \ 113 !IS_NULL_REG(&(cpc)->cpc_entry.reg)) 114 /* 115 * Arbitrary Retries in case the remote processor is slow to respond 116 * to PCC commands. Keeping it high enough to cover emulators where 117 * the processors run painfully slow. 118 */ 119 #define NUM_RETRIES 500 120 121 struct cppc_attr { 122 struct attribute attr; 123 ssize_t (*show)(struct kobject *kobj, 124 struct attribute *attr, char *buf); 125 ssize_t (*store)(struct kobject *kobj, 126 struct attribute *attr, const char *c, ssize_t count); 127 }; 128 129 #define define_one_cppc_ro(_name) \ 130 static struct cppc_attr _name = \ 131 __ATTR(_name, 0444, show_##_name, NULL) 132 133 #define to_cpc_desc(a) container_of(a, struct cpc_desc, kobj) 134 135 static ssize_t show_feedback_ctrs(struct kobject *kobj, 136 struct attribute *attr, char *buf) 137 { 138 struct cpc_desc *cpc_ptr = to_cpc_desc(kobj); 139 struct cppc_perf_fb_ctrs fb_ctrs = {0}; 140 141 cppc_get_perf_ctrs(cpc_ptr->cpu_id, &fb_ctrs); 142 143 return scnprintf(buf, PAGE_SIZE, "ref:%llu del:%llu\n", 144 fb_ctrs.reference, fb_ctrs.delivered); 145 } 146 define_one_cppc_ro(feedback_ctrs); 147 148 static ssize_t show_reference_perf(struct kobject *kobj, 149 struct attribute *attr, char *buf) 150 { 151 struct cpc_desc *cpc_ptr = to_cpc_desc(kobj); 152 struct cppc_perf_fb_ctrs fb_ctrs = {0}; 153 154 cppc_get_perf_ctrs(cpc_ptr->cpu_id, &fb_ctrs); 155 156 return scnprintf(buf, PAGE_SIZE, "%llu\n", 157 fb_ctrs.reference_perf); 158 } 159 define_one_cppc_ro(reference_perf); 160 161 static ssize_t show_wraparound_time(struct kobject *kobj, 162 struct attribute *attr, char *buf) 163 { 164 struct cpc_desc *cpc_ptr = to_cpc_desc(kobj); 165 struct cppc_perf_fb_ctrs fb_ctrs = {0}; 166 167 cppc_get_perf_ctrs(cpc_ptr->cpu_id, &fb_ctrs); 168 169 return scnprintf(buf, PAGE_SIZE, "%llu\n", fb_ctrs.ctr_wrap_time); 170 171 } 172 define_one_cppc_ro(wraparound_time); 173 174 static struct attribute *cppc_attrs[] = { 175 &feedback_ctrs.attr, 176 &reference_perf.attr, 177 &wraparound_time.attr, 178 NULL 179 }; 180 181 static struct kobj_type cppc_ktype = { 182 .sysfs_ops = &kobj_sysfs_ops, 183 .default_attrs = cppc_attrs, 184 }; 185 186 static int check_pcc_chan(bool chk_err_bit) 187 { 188 int ret = -EIO, status = 0; 189 struct acpi_pcct_shared_memory __iomem *generic_comm_base = pcc_data.pcc_comm_addr; 190 ktime_t next_deadline = ktime_add(ktime_get(), pcc_data.deadline); 191 192 if (!pcc_data.platform_owns_pcc) 193 return 0; 194 195 /* Retry in case the remote processor was too slow to catch up. */ 196 while (!ktime_after(ktime_get(), next_deadline)) { 197 /* 198 * Per spec, prior to boot the PCC space wil be initialized by 199 * platform and should have set the command completion bit when 200 * PCC can be used by OSPM 201 */ 202 status = readw_relaxed(&generic_comm_base->status); 203 if (status & PCC_CMD_COMPLETE_MASK) { 204 ret = 0; 205 if (chk_err_bit && (status & PCC_ERROR_MASK)) 206 ret = -EIO; 207 break; 208 } 209 /* 210 * Reducing the bus traffic in case this loop takes longer than 211 * a few retries. 212 */ 213 udelay(3); 214 } 215 216 if (likely(!ret)) 217 pcc_data.platform_owns_pcc = false; 218 else 219 pr_err("PCC check channel failed. Status=%x\n", status); 220 221 return ret; 222 } 223 224 /* 225 * This function transfers the ownership of the PCC to the platform 226 * So it must be called while holding write_lock(pcc_lock) 227 */ 228 static int send_pcc_cmd(u16 cmd) 229 { 230 int ret = -EIO, i; 231 struct acpi_pcct_shared_memory *generic_comm_base = 232 (struct acpi_pcct_shared_memory *) pcc_data.pcc_comm_addr; 233 static ktime_t last_cmd_cmpl_time, last_mpar_reset; 234 static int mpar_count; 235 unsigned int time_delta; 236 237 /* 238 * For CMD_WRITE we know for a fact the caller should have checked 239 * the channel before writing to PCC space 240 */ 241 if (cmd == CMD_READ) { 242 /* 243 * If there are pending cpc_writes, then we stole the channel 244 * before write completion, so first send a WRITE command to 245 * platform 246 */ 247 if (pcc_data.pending_pcc_write_cmd) 248 send_pcc_cmd(CMD_WRITE); 249 250 ret = check_pcc_chan(false); 251 if (ret) 252 goto end; 253 } else /* CMD_WRITE */ 254 pcc_data.pending_pcc_write_cmd = FALSE; 255 256 /* 257 * Handle the Minimum Request Turnaround Time(MRTT) 258 * "The minimum amount of time that OSPM must wait after the completion 259 * of a command before issuing the next command, in microseconds" 260 */ 261 if (pcc_data.pcc_mrtt) { 262 time_delta = ktime_us_delta(ktime_get(), last_cmd_cmpl_time); 263 if (pcc_data.pcc_mrtt > time_delta) 264 udelay(pcc_data.pcc_mrtt - time_delta); 265 } 266 267 /* 268 * Handle the non-zero Maximum Periodic Access Rate(MPAR) 269 * "The maximum number of periodic requests that the subspace channel can 270 * support, reported in commands per minute. 0 indicates no limitation." 271 * 272 * This parameter should be ideally zero or large enough so that it can 273 * handle maximum number of requests that all the cores in the system can 274 * collectively generate. If it is not, we will follow the spec and just 275 * not send the request to the platform after hitting the MPAR limit in 276 * any 60s window 277 */ 278 if (pcc_data.pcc_mpar) { 279 if (mpar_count == 0) { 280 time_delta = ktime_ms_delta(ktime_get(), last_mpar_reset); 281 if (time_delta < 60 * MSEC_PER_SEC) { 282 pr_debug("PCC cmd not sent due to MPAR limit"); 283 ret = -EIO; 284 goto end; 285 } 286 last_mpar_reset = ktime_get(); 287 mpar_count = pcc_data.pcc_mpar; 288 } 289 mpar_count--; 290 } 291 292 /* Write to the shared comm region. */ 293 writew_relaxed(cmd, &generic_comm_base->command); 294 295 /* Flip CMD COMPLETE bit */ 296 writew_relaxed(0, &generic_comm_base->status); 297 298 pcc_data.platform_owns_pcc = true; 299 300 /* Ring doorbell */ 301 ret = mbox_send_message(pcc_data.pcc_channel, &cmd); 302 if (ret < 0) { 303 pr_err("Err sending PCC mbox message. cmd:%d, ret:%d\n", 304 cmd, ret); 305 goto end; 306 } 307 308 /* wait for completion and check for PCC errro bit */ 309 ret = check_pcc_chan(true); 310 311 if (pcc_data.pcc_mrtt) 312 last_cmd_cmpl_time = ktime_get(); 313 314 if (pcc_data.pcc_channel->mbox->txdone_irq) 315 mbox_chan_txdone(pcc_data.pcc_channel, ret); 316 else 317 mbox_client_txdone(pcc_data.pcc_channel, ret); 318 319 end: 320 if (cmd == CMD_WRITE) { 321 if (unlikely(ret)) { 322 for_each_possible_cpu(i) { 323 struct cpc_desc *desc = per_cpu(cpc_desc_ptr, i); 324 if (!desc) 325 continue; 326 327 if (desc->write_cmd_id == pcc_data.pcc_write_cnt) 328 desc->write_cmd_status = ret; 329 } 330 } 331 pcc_data.pcc_write_cnt++; 332 wake_up_all(&pcc_data.pcc_write_wait_q); 333 } 334 335 return ret; 336 } 337 338 static void cppc_chan_tx_done(struct mbox_client *cl, void *msg, int ret) 339 { 340 if (ret < 0) 341 pr_debug("TX did not complete: CMD sent:%x, ret:%d\n", 342 *(u16 *)msg, ret); 343 else 344 pr_debug("TX completed. CMD sent:%x, ret:%d\n", 345 *(u16 *)msg, ret); 346 } 347 348 struct mbox_client cppc_mbox_cl = { 349 .tx_done = cppc_chan_tx_done, 350 .knows_txdone = true, 351 }; 352 353 static int acpi_get_psd(struct cpc_desc *cpc_ptr, acpi_handle handle) 354 { 355 int result = -EFAULT; 356 acpi_status status = AE_OK; 357 struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL}; 358 struct acpi_buffer format = {sizeof("NNNNN"), "NNNNN"}; 359 struct acpi_buffer state = {0, NULL}; 360 union acpi_object *psd = NULL; 361 struct acpi_psd_package *pdomain; 362 363 status = acpi_evaluate_object_typed(handle, "_PSD", NULL, &buffer, 364 ACPI_TYPE_PACKAGE); 365 if (ACPI_FAILURE(status)) 366 return -ENODEV; 367 368 psd = buffer.pointer; 369 if (!psd || psd->package.count != 1) { 370 pr_debug("Invalid _PSD data\n"); 371 goto end; 372 } 373 374 pdomain = &(cpc_ptr->domain_info); 375 376 state.length = sizeof(struct acpi_psd_package); 377 state.pointer = pdomain; 378 379 status = acpi_extract_package(&(psd->package.elements[0]), 380 &format, &state); 381 if (ACPI_FAILURE(status)) { 382 pr_debug("Invalid _PSD data for CPU:%d\n", cpc_ptr->cpu_id); 383 goto end; 384 } 385 386 if (pdomain->num_entries != ACPI_PSD_REV0_ENTRIES) { 387 pr_debug("Unknown _PSD:num_entries for CPU:%d\n", cpc_ptr->cpu_id); 388 goto end; 389 } 390 391 if (pdomain->revision != ACPI_PSD_REV0_REVISION) { 392 pr_debug("Unknown _PSD:revision for CPU: %d\n", cpc_ptr->cpu_id); 393 goto end; 394 } 395 396 if (pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ALL && 397 pdomain->coord_type != DOMAIN_COORD_TYPE_SW_ANY && 398 pdomain->coord_type != DOMAIN_COORD_TYPE_HW_ALL) { 399 pr_debug("Invalid _PSD:coord_type for CPU:%d\n", cpc_ptr->cpu_id); 400 goto end; 401 } 402 403 result = 0; 404 end: 405 kfree(buffer.pointer); 406 return result; 407 } 408 409 /** 410 * acpi_get_psd_map - Map the CPUs in a common freq domain. 411 * @all_cpu_data: Ptrs to CPU specific CPPC data including PSD info. 412 * 413 * Return: 0 for success or negative value for err. 414 */ 415 int acpi_get_psd_map(struct cppc_cpudata **all_cpu_data) 416 { 417 int count_target; 418 int retval = 0; 419 unsigned int i, j; 420 cpumask_var_t covered_cpus; 421 struct cppc_cpudata *pr, *match_pr; 422 struct acpi_psd_package *pdomain; 423 struct acpi_psd_package *match_pdomain; 424 struct cpc_desc *cpc_ptr, *match_cpc_ptr; 425 426 if (!zalloc_cpumask_var(&covered_cpus, GFP_KERNEL)) 427 return -ENOMEM; 428 429 /* 430 * Now that we have _PSD data from all CPUs, lets setup P-state 431 * domain info. 432 */ 433 for_each_possible_cpu(i) { 434 pr = all_cpu_data[i]; 435 if (!pr) 436 continue; 437 438 if (cpumask_test_cpu(i, covered_cpus)) 439 continue; 440 441 cpc_ptr = per_cpu(cpc_desc_ptr, i); 442 if (!cpc_ptr) { 443 retval = -EFAULT; 444 goto err_ret; 445 } 446 447 pdomain = &(cpc_ptr->domain_info); 448 cpumask_set_cpu(i, pr->shared_cpu_map); 449 cpumask_set_cpu(i, covered_cpus); 450 if (pdomain->num_processors <= 1) 451 continue; 452 453 /* Validate the Domain info */ 454 count_target = pdomain->num_processors; 455 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL) 456 pr->shared_type = CPUFREQ_SHARED_TYPE_ALL; 457 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL) 458 pr->shared_type = CPUFREQ_SHARED_TYPE_HW; 459 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY) 460 pr->shared_type = CPUFREQ_SHARED_TYPE_ANY; 461 462 for_each_possible_cpu(j) { 463 if (i == j) 464 continue; 465 466 match_cpc_ptr = per_cpu(cpc_desc_ptr, j); 467 if (!match_cpc_ptr) { 468 retval = -EFAULT; 469 goto err_ret; 470 } 471 472 match_pdomain = &(match_cpc_ptr->domain_info); 473 if (match_pdomain->domain != pdomain->domain) 474 continue; 475 476 /* Here i and j are in the same domain */ 477 if (match_pdomain->num_processors != count_target) { 478 retval = -EFAULT; 479 goto err_ret; 480 } 481 482 if (pdomain->coord_type != match_pdomain->coord_type) { 483 retval = -EFAULT; 484 goto err_ret; 485 } 486 487 cpumask_set_cpu(j, covered_cpus); 488 cpumask_set_cpu(j, pr->shared_cpu_map); 489 } 490 491 for_each_possible_cpu(j) { 492 if (i == j) 493 continue; 494 495 match_pr = all_cpu_data[j]; 496 if (!match_pr) 497 continue; 498 499 match_cpc_ptr = per_cpu(cpc_desc_ptr, j); 500 if (!match_cpc_ptr) { 501 retval = -EFAULT; 502 goto err_ret; 503 } 504 505 match_pdomain = &(match_cpc_ptr->domain_info); 506 if (match_pdomain->domain != pdomain->domain) 507 continue; 508 509 match_pr->shared_type = pr->shared_type; 510 cpumask_copy(match_pr->shared_cpu_map, 511 pr->shared_cpu_map); 512 } 513 } 514 515 err_ret: 516 for_each_possible_cpu(i) { 517 pr = all_cpu_data[i]; 518 if (!pr) 519 continue; 520 521 /* Assume no coordination on any error parsing domain info */ 522 if (retval) { 523 cpumask_clear(pr->shared_cpu_map); 524 cpumask_set_cpu(i, pr->shared_cpu_map); 525 pr->shared_type = CPUFREQ_SHARED_TYPE_ALL; 526 } 527 } 528 529 free_cpumask_var(covered_cpus); 530 return retval; 531 } 532 EXPORT_SYMBOL_GPL(acpi_get_psd_map); 533 534 static int register_pcc_channel(int pcc_subspace_idx) 535 { 536 struct acpi_pcct_hw_reduced *cppc_ss; 537 u64 usecs_lat; 538 539 if (pcc_subspace_idx >= 0) { 540 pcc_data.pcc_channel = pcc_mbox_request_channel(&cppc_mbox_cl, 541 pcc_subspace_idx); 542 543 if (IS_ERR(pcc_data.pcc_channel)) { 544 pr_err("Failed to find PCC communication channel\n"); 545 return -ENODEV; 546 } 547 548 /* 549 * The PCC mailbox controller driver should 550 * have parsed the PCCT (global table of all 551 * PCC channels) and stored pointers to the 552 * subspace communication region in con_priv. 553 */ 554 cppc_ss = (pcc_data.pcc_channel)->con_priv; 555 556 if (!cppc_ss) { 557 pr_err("No PCC subspace found for CPPC\n"); 558 return -ENODEV; 559 } 560 561 /* 562 * cppc_ss->latency is just a Nominal value. In reality 563 * the remote processor could be much slower to reply. 564 * So add an arbitrary amount of wait on top of Nominal. 565 */ 566 usecs_lat = NUM_RETRIES * cppc_ss->latency; 567 pcc_data.deadline = ns_to_ktime(usecs_lat * NSEC_PER_USEC); 568 pcc_data.pcc_mrtt = cppc_ss->min_turnaround_time; 569 pcc_data.pcc_mpar = cppc_ss->max_access_rate; 570 pcc_data.pcc_nominal = cppc_ss->latency; 571 572 pcc_data.pcc_comm_addr = acpi_os_ioremap(cppc_ss->base_address, cppc_ss->length); 573 if (!pcc_data.pcc_comm_addr) { 574 pr_err("Failed to ioremap PCC comm region mem\n"); 575 return -ENOMEM; 576 } 577 578 /* Set flag so that we dont come here for each CPU. */ 579 pcc_data.pcc_channel_acquired = true; 580 } 581 582 return 0; 583 } 584 585 /** 586 * cpc_ffh_supported() - check if FFH reading supported 587 * 588 * Check if the architecture has support for functional fixed hardware 589 * read/write capability. 590 * 591 * Return: true for supported, false for not supported 592 */ 593 bool __weak cpc_ffh_supported(void) 594 { 595 return false; 596 } 597 598 /* 599 * An example CPC table looks like the following. 600 * 601 * Name(_CPC, Package() 602 * { 603 * 17, 604 * NumEntries 605 * 1, 606 * // Revision 607 * ResourceTemplate(){Register(PCC, 32, 0, 0x120, 2)}, 608 * // Highest Performance 609 * ResourceTemplate(){Register(PCC, 32, 0, 0x124, 2)}, 610 * // Nominal Performance 611 * ResourceTemplate(){Register(PCC, 32, 0, 0x128, 2)}, 612 * // Lowest Nonlinear Performance 613 * ResourceTemplate(){Register(PCC, 32, 0, 0x12C, 2)}, 614 * // Lowest Performance 615 * ResourceTemplate(){Register(PCC, 32, 0, 0x130, 2)}, 616 * // Guaranteed Performance Register 617 * ResourceTemplate(){Register(PCC, 32, 0, 0x110, 2)}, 618 * // Desired Performance Register 619 * ResourceTemplate(){Register(SystemMemory, 0, 0, 0, 0)}, 620 * .. 621 * .. 622 * .. 623 * 624 * } 625 * Each Register() encodes how to access that specific register. 626 * e.g. a sample PCC entry has the following encoding: 627 * 628 * Register ( 629 * PCC, 630 * AddressSpaceKeyword 631 * 8, 632 * //RegisterBitWidth 633 * 8, 634 * //RegisterBitOffset 635 * 0x30, 636 * //RegisterAddress 637 * 9 638 * //AccessSize (subspace ID) 639 * 0 640 * ) 641 * } 642 */ 643 644 /** 645 * acpi_cppc_processor_probe - Search for per CPU _CPC objects. 646 * @pr: Ptr to acpi_processor containing this CPUs logical Id. 647 * 648 * Return: 0 for success or negative value for err. 649 */ 650 int acpi_cppc_processor_probe(struct acpi_processor *pr) 651 { 652 struct acpi_buffer output = {ACPI_ALLOCATE_BUFFER, NULL}; 653 union acpi_object *out_obj, *cpc_obj; 654 struct cpc_desc *cpc_ptr; 655 struct cpc_reg *gas_t; 656 struct device *cpu_dev; 657 acpi_handle handle = pr->handle; 658 unsigned int num_ent, i, cpc_rev; 659 acpi_status status; 660 int ret = -EFAULT; 661 662 /* Parse the ACPI _CPC table for this cpu. */ 663 status = acpi_evaluate_object_typed(handle, "_CPC", NULL, &output, 664 ACPI_TYPE_PACKAGE); 665 if (ACPI_FAILURE(status)) { 666 ret = -ENODEV; 667 goto out_buf_free; 668 } 669 670 out_obj = (union acpi_object *) output.pointer; 671 672 cpc_ptr = kzalloc(sizeof(struct cpc_desc), GFP_KERNEL); 673 if (!cpc_ptr) { 674 ret = -ENOMEM; 675 goto out_buf_free; 676 } 677 678 /* First entry is NumEntries. */ 679 cpc_obj = &out_obj->package.elements[0]; 680 if (cpc_obj->type == ACPI_TYPE_INTEGER) { 681 num_ent = cpc_obj->integer.value; 682 } else { 683 pr_debug("Unexpected entry type(%d) for NumEntries\n", 684 cpc_obj->type); 685 goto out_free; 686 } 687 688 /* Only support CPPCv2. Bail otherwise. */ 689 if (num_ent != CPPC_NUM_ENT) { 690 pr_debug("Firmware exports %d entries. Expected: %d\n", 691 num_ent, CPPC_NUM_ENT); 692 goto out_free; 693 } 694 695 cpc_ptr->num_entries = num_ent; 696 697 /* Second entry should be revision. */ 698 cpc_obj = &out_obj->package.elements[1]; 699 if (cpc_obj->type == ACPI_TYPE_INTEGER) { 700 cpc_rev = cpc_obj->integer.value; 701 } else { 702 pr_debug("Unexpected entry type(%d) for Revision\n", 703 cpc_obj->type); 704 goto out_free; 705 } 706 707 if (cpc_rev != CPPC_REV) { 708 pr_debug("Firmware exports revision:%d. Expected:%d\n", 709 cpc_rev, CPPC_REV); 710 goto out_free; 711 } 712 713 /* Iterate through remaining entries in _CPC */ 714 for (i = 2; i < num_ent; i++) { 715 cpc_obj = &out_obj->package.elements[i]; 716 717 if (cpc_obj->type == ACPI_TYPE_INTEGER) { 718 cpc_ptr->cpc_regs[i-2].type = ACPI_TYPE_INTEGER; 719 cpc_ptr->cpc_regs[i-2].cpc_entry.int_value = cpc_obj->integer.value; 720 } else if (cpc_obj->type == ACPI_TYPE_BUFFER) { 721 gas_t = (struct cpc_reg *) 722 cpc_obj->buffer.pointer; 723 724 /* 725 * The PCC Subspace index is encoded inside 726 * the CPC table entries. The same PCC index 727 * will be used for all the PCC entries, 728 * so extract it only once. 729 */ 730 if (gas_t->space_id == ACPI_ADR_SPACE_PLATFORM_COMM) { 731 if (pcc_data.pcc_subspace_idx < 0) 732 pcc_data.pcc_subspace_idx = gas_t->access_width; 733 else if (pcc_data.pcc_subspace_idx != gas_t->access_width) { 734 pr_debug("Mismatched PCC ids.\n"); 735 goto out_free; 736 } 737 } else if (gas_t->space_id == ACPI_ADR_SPACE_SYSTEM_MEMORY) { 738 if (gas_t->address) { 739 void __iomem *addr; 740 741 addr = ioremap(gas_t->address, gas_t->bit_width/8); 742 if (!addr) 743 goto out_free; 744 cpc_ptr->cpc_regs[i-2].sys_mem_vaddr = addr; 745 } 746 } else { 747 if (gas_t->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE || !cpc_ffh_supported()) { 748 /* Support only PCC ,SYS MEM and FFH type regs */ 749 pr_debug("Unsupported register type: %d\n", gas_t->space_id); 750 goto out_free; 751 } 752 } 753 754 cpc_ptr->cpc_regs[i-2].type = ACPI_TYPE_BUFFER; 755 memcpy(&cpc_ptr->cpc_regs[i-2].cpc_entry.reg, gas_t, sizeof(*gas_t)); 756 } else { 757 pr_debug("Err in entry:%d in CPC table of CPU:%d \n", i, pr->id); 758 goto out_free; 759 } 760 } 761 /* Store CPU Logical ID */ 762 cpc_ptr->cpu_id = pr->id; 763 764 /* Parse PSD data for this CPU */ 765 ret = acpi_get_psd(cpc_ptr, handle); 766 if (ret) 767 goto out_free; 768 769 /* Register PCC channel once for all CPUs. */ 770 if (!pcc_data.pcc_channel_acquired) { 771 ret = register_pcc_channel(pcc_data.pcc_subspace_idx); 772 if (ret) 773 goto out_free; 774 775 init_rwsem(&pcc_data.pcc_lock); 776 init_waitqueue_head(&pcc_data.pcc_write_wait_q); 777 } 778 779 /* Plug PSD data into this CPUs CPC descriptor. */ 780 per_cpu(cpc_desc_ptr, pr->id) = cpc_ptr; 781 782 /* Everything looks okay */ 783 pr_debug("Parsed CPC struct for CPU: %d\n", pr->id); 784 785 /* Add per logical CPU nodes for reading its feedback counters. */ 786 cpu_dev = get_cpu_device(pr->id); 787 if (!cpu_dev) 788 goto out_free; 789 790 ret = kobject_init_and_add(&cpc_ptr->kobj, &cppc_ktype, &cpu_dev->kobj, 791 "acpi_cppc"); 792 if (ret) 793 goto out_free; 794 795 kfree(output.pointer); 796 return 0; 797 798 out_free: 799 /* Free all the mapped sys mem areas for this CPU */ 800 for (i = 2; i < cpc_ptr->num_entries; i++) { 801 void __iomem *addr = cpc_ptr->cpc_regs[i-2].sys_mem_vaddr; 802 803 if (addr) 804 iounmap(addr); 805 } 806 kfree(cpc_ptr); 807 808 out_buf_free: 809 kfree(output.pointer); 810 return ret; 811 } 812 EXPORT_SYMBOL_GPL(acpi_cppc_processor_probe); 813 814 /** 815 * acpi_cppc_processor_exit - Cleanup CPC structs. 816 * @pr: Ptr to acpi_processor containing this CPUs logical Id. 817 * 818 * Return: Void 819 */ 820 void acpi_cppc_processor_exit(struct acpi_processor *pr) 821 { 822 struct cpc_desc *cpc_ptr; 823 unsigned int i; 824 void __iomem *addr; 825 826 cpc_ptr = per_cpu(cpc_desc_ptr, pr->id); 827 828 /* Free all the mapped sys mem areas for this CPU */ 829 for (i = 2; i < cpc_ptr->num_entries; i++) { 830 addr = cpc_ptr->cpc_regs[i-2].sys_mem_vaddr; 831 if (addr) 832 iounmap(addr); 833 } 834 835 kobject_put(&cpc_ptr->kobj); 836 kfree(cpc_ptr); 837 } 838 EXPORT_SYMBOL_GPL(acpi_cppc_processor_exit); 839 840 /** 841 * cpc_read_ffh() - Read FFH register 842 * @cpunum: cpu number to read 843 * @reg: cppc register information 844 * @val: place holder for return value 845 * 846 * Read bit_width bits from a specified address and bit_offset 847 * 848 * Return: 0 for success and error code 849 */ 850 int __weak cpc_read_ffh(int cpunum, struct cpc_reg *reg, u64 *val) 851 { 852 return -ENOTSUPP; 853 } 854 855 /** 856 * cpc_write_ffh() - Write FFH register 857 * @cpunum: cpu number to write 858 * @reg: cppc register information 859 * @val: value to write 860 * 861 * Write value of bit_width bits to a specified address and bit_offset 862 * 863 * Return: 0 for success and error code 864 */ 865 int __weak cpc_write_ffh(int cpunum, struct cpc_reg *reg, u64 val) 866 { 867 return -ENOTSUPP; 868 } 869 870 /* 871 * Since cpc_read and cpc_write are called while holding pcc_lock, it should be 872 * as fast as possible. We have already mapped the PCC subspace during init, so 873 * we can directly write to it. 874 */ 875 876 static int cpc_read(int cpu, struct cpc_register_resource *reg_res, u64 *val) 877 { 878 int ret_val = 0; 879 void __iomem *vaddr = 0; 880 struct cpc_reg *reg = ®_res->cpc_entry.reg; 881 882 if (reg_res->type == ACPI_TYPE_INTEGER) { 883 *val = reg_res->cpc_entry.int_value; 884 return ret_val; 885 } 886 887 *val = 0; 888 if (reg->space_id == ACPI_ADR_SPACE_PLATFORM_COMM) 889 vaddr = GET_PCC_VADDR(reg->address); 890 else if (reg->space_id == ACPI_ADR_SPACE_SYSTEM_MEMORY) 891 vaddr = reg_res->sys_mem_vaddr; 892 else if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) 893 return cpc_read_ffh(cpu, reg, val); 894 else 895 return acpi_os_read_memory((acpi_physical_address)reg->address, 896 val, reg->bit_width); 897 898 switch (reg->bit_width) { 899 case 8: 900 *val = readb_relaxed(vaddr); 901 break; 902 case 16: 903 *val = readw_relaxed(vaddr); 904 break; 905 case 32: 906 *val = readl_relaxed(vaddr); 907 break; 908 case 64: 909 *val = readq_relaxed(vaddr); 910 break; 911 default: 912 pr_debug("Error: Cannot read %u bit width from PCC\n", 913 reg->bit_width); 914 ret_val = -EFAULT; 915 } 916 917 return ret_val; 918 } 919 920 static int cpc_write(int cpu, struct cpc_register_resource *reg_res, u64 val) 921 { 922 int ret_val = 0; 923 void __iomem *vaddr = 0; 924 struct cpc_reg *reg = ®_res->cpc_entry.reg; 925 926 if (reg->space_id == ACPI_ADR_SPACE_PLATFORM_COMM) 927 vaddr = GET_PCC_VADDR(reg->address); 928 else if (reg->space_id == ACPI_ADR_SPACE_SYSTEM_MEMORY) 929 vaddr = reg_res->sys_mem_vaddr; 930 else if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) 931 return cpc_write_ffh(cpu, reg, val); 932 else 933 return acpi_os_write_memory((acpi_physical_address)reg->address, 934 val, reg->bit_width); 935 936 switch (reg->bit_width) { 937 case 8: 938 writeb_relaxed(val, vaddr); 939 break; 940 case 16: 941 writew_relaxed(val, vaddr); 942 break; 943 case 32: 944 writel_relaxed(val, vaddr); 945 break; 946 case 64: 947 writeq_relaxed(val, vaddr); 948 break; 949 default: 950 pr_debug("Error: Cannot write %u bit width to PCC\n", 951 reg->bit_width); 952 ret_val = -EFAULT; 953 break; 954 } 955 956 return ret_val; 957 } 958 959 /** 960 * cppc_get_perf_caps - Get a CPUs performance capabilities. 961 * @cpunum: CPU from which to get capabilities info. 962 * @perf_caps: ptr to cppc_perf_caps. See cppc_acpi.h 963 * 964 * Return: 0 for success with perf_caps populated else -ERRNO. 965 */ 966 int cppc_get_perf_caps(int cpunum, struct cppc_perf_caps *perf_caps) 967 { 968 struct cpc_desc *cpc_desc = per_cpu(cpc_desc_ptr, cpunum); 969 struct cpc_register_resource *highest_reg, *lowest_reg, *ref_perf, 970 *nom_perf; 971 u64 high, low, nom; 972 int ret = 0, regs_in_pcc = 0; 973 974 if (!cpc_desc) { 975 pr_debug("No CPC descriptor for CPU:%d\n", cpunum); 976 return -ENODEV; 977 } 978 979 highest_reg = &cpc_desc->cpc_regs[HIGHEST_PERF]; 980 lowest_reg = &cpc_desc->cpc_regs[LOWEST_PERF]; 981 ref_perf = &cpc_desc->cpc_regs[REFERENCE_PERF]; 982 nom_perf = &cpc_desc->cpc_regs[NOMINAL_PERF]; 983 984 /* Are any of the regs PCC ?*/ 985 if (CPC_IN_PCC(highest_reg) || CPC_IN_PCC(lowest_reg) || 986 CPC_IN_PCC(ref_perf) || CPC_IN_PCC(nom_perf)) { 987 regs_in_pcc = 1; 988 down_write(&pcc_data.pcc_lock); 989 /* Ring doorbell once to update PCC subspace */ 990 if (send_pcc_cmd(CMD_READ) < 0) { 991 ret = -EIO; 992 goto out_err; 993 } 994 } 995 996 cpc_read(cpunum, highest_reg, &high); 997 perf_caps->highest_perf = high; 998 999 cpc_read(cpunum, lowest_reg, &low); 1000 perf_caps->lowest_perf = low; 1001 1002 cpc_read(cpunum, nom_perf, &nom); 1003 perf_caps->nominal_perf = nom; 1004 1005 if (!high || !low || !nom) 1006 ret = -EFAULT; 1007 1008 out_err: 1009 if (regs_in_pcc) 1010 up_write(&pcc_data.pcc_lock); 1011 return ret; 1012 } 1013 EXPORT_SYMBOL_GPL(cppc_get_perf_caps); 1014 1015 /** 1016 * cppc_get_perf_ctrs - Read a CPUs performance feedback counters. 1017 * @cpunum: CPU from which to read counters. 1018 * @perf_fb_ctrs: ptr to cppc_perf_fb_ctrs. See cppc_acpi.h 1019 * 1020 * Return: 0 for success with perf_fb_ctrs populated else -ERRNO. 1021 */ 1022 int cppc_get_perf_ctrs(int cpunum, struct cppc_perf_fb_ctrs *perf_fb_ctrs) 1023 { 1024 struct cpc_desc *cpc_desc = per_cpu(cpc_desc_ptr, cpunum); 1025 struct cpc_register_resource *delivered_reg, *reference_reg, 1026 *ref_perf_reg, *ctr_wrap_reg; 1027 u64 delivered, reference, ref_perf, ctr_wrap_time; 1028 int ret = 0, regs_in_pcc = 0; 1029 1030 if (!cpc_desc) { 1031 pr_debug("No CPC descriptor for CPU:%d\n", cpunum); 1032 return -ENODEV; 1033 } 1034 1035 delivered_reg = &cpc_desc->cpc_regs[DELIVERED_CTR]; 1036 reference_reg = &cpc_desc->cpc_regs[REFERENCE_CTR]; 1037 ref_perf_reg = &cpc_desc->cpc_regs[REFERENCE_PERF]; 1038 ctr_wrap_reg = &cpc_desc->cpc_regs[CTR_WRAP_TIME]; 1039 1040 /* 1041 * If refernce perf register is not supported then we should 1042 * use the nominal perf value 1043 */ 1044 if (!CPC_SUPPORTED(ref_perf_reg)) 1045 ref_perf_reg = &cpc_desc->cpc_regs[NOMINAL_PERF]; 1046 1047 /* Are any of the regs PCC ?*/ 1048 if (CPC_IN_PCC(delivered_reg) || CPC_IN_PCC(reference_reg) || 1049 CPC_IN_PCC(ctr_wrap_reg) || CPC_IN_PCC(ref_perf_reg)) { 1050 down_write(&pcc_data.pcc_lock); 1051 regs_in_pcc = 1; 1052 /* Ring doorbell once to update PCC subspace */ 1053 if (send_pcc_cmd(CMD_READ) < 0) { 1054 ret = -EIO; 1055 goto out_err; 1056 } 1057 } 1058 1059 cpc_read(cpunum, delivered_reg, &delivered); 1060 cpc_read(cpunum, reference_reg, &reference); 1061 cpc_read(cpunum, ref_perf_reg, &ref_perf); 1062 1063 /* 1064 * Per spec, if ctr_wrap_time optional register is unsupported, then the 1065 * performance counters are assumed to never wrap during the lifetime of 1066 * platform 1067 */ 1068 ctr_wrap_time = (u64)(~((u64)0)); 1069 if (CPC_SUPPORTED(ctr_wrap_reg)) 1070 cpc_read(cpunum, ctr_wrap_reg, &ctr_wrap_time); 1071 1072 if (!delivered || !reference || !ref_perf) { 1073 ret = -EFAULT; 1074 goto out_err; 1075 } 1076 1077 perf_fb_ctrs->delivered = delivered; 1078 perf_fb_ctrs->reference = reference; 1079 perf_fb_ctrs->reference_perf = ref_perf; 1080 perf_fb_ctrs->ctr_wrap_time = ctr_wrap_time; 1081 out_err: 1082 if (regs_in_pcc) 1083 up_write(&pcc_data.pcc_lock); 1084 return ret; 1085 } 1086 EXPORT_SYMBOL_GPL(cppc_get_perf_ctrs); 1087 1088 /** 1089 * cppc_set_perf - Set a CPUs performance controls. 1090 * @cpu: CPU for which to set performance controls. 1091 * @perf_ctrls: ptr to cppc_perf_ctrls. See cppc_acpi.h 1092 * 1093 * Return: 0 for success, -ERRNO otherwise. 1094 */ 1095 int cppc_set_perf(int cpu, struct cppc_perf_ctrls *perf_ctrls) 1096 { 1097 struct cpc_desc *cpc_desc = per_cpu(cpc_desc_ptr, cpu); 1098 struct cpc_register_resource *desired_reg; 1099 int ret = 0; 1100 1101 if (!cpc_desc) { 1102 pr_debug("No CPC descriptor for CPU:%d\n", cpu); 1103 return -ENODEV; 1104 } 1105 1106 desired_reg = &cpc_desc->cpc_regs[DESIRED_PERF]; 1107 1108 /* 1109 * This is Phase-I where we want to write to CPC registers 1110 * -> We want all CPUs to be able to execute this phase in parallel 1111 * 1112 * Since read_lock can be acquired by multiple CPUs simultaneously we 1113 * achieve that goal here 1114 */ 1115 if (CPC_IN_PCC(desired_reg)) { 1116 down_read(&pcc_data.pcc_lock); /* BEGIN Phase-I */ 1117 if (pcc_data.platform_owns_pcc) { 1118 ret = check_pcc_chan(false); 1119 if (ret) { 1120 up_read(&pcc_data.pcc_lock); 1121 return ret; 1122 } 1123 } 1124 /* 1125 * Update the pending_write to make sure a PCC CMD_READ will not 1126 * arrive and steal the channel during the switch to write lock 1127 */ 1128 pcc_data.pending_pcc_write_cmd = true; 1129 cpc_desc->write_cmd_id = pcc_data.pcc_write_cnt; 1130 cpc_desc->write_cmd_status = 0; 1131 } 1132 1133 /* 1134 * Skip writing MIN/MAX until Linux knows how to come up with 1135 * useful values. 1136 */ 1137 cpc_write(cpu, desired_reg, perf_ctrls->desired_perf); 1138 1139 if (CPC_IN_PCC(desired_reg)) 1140 up_read(&pcc_data.pcc_lock); /* END Phase-I */ 1141 /* 1142 * This is Phase-II where we transfer the ownership of PCC to Platform 1143 * 1144 * Short Summary: Basically if we think of a group of cppc_set_perf 1145 * requests that happened in short overlapping interval. The last CPU to 1146 * come out of Phase-I will enter Phase-II and ring the doorbell. 1147 * 1148 * We have the following requirements for Phase-II: 1149 * 1. We want to execute Phase-II only when there are no CPUs 1150 * currently executing in Phase-I 1151 * 2. Once we start Phase-II we want to avoid all other CPUs from 1152 * entering Phase-I. 1153 * 3. We want only one CPU among all those who went through Phase-I 1154 * to run phase-II 1155 * 1156 * If write_trylock fails to get the lock and doesn't transfer the 1157 * PCC ownership to the platform, then one of the following will be TRUE 1158 * 1. There is at-least one CPU in Phase-I which will later execute 1159 * write_trylock, so the CPUs in Phase-I will be responsible for 1160 * executing the Phase-II. 1161 * 2. Some other CPU has beaten this CPU to successfully execute the 1162 * write_trylock and has already acquired the write_lock. We know for a 1163 * fact it(other CPU acquiring the write_lock) couldn't have happened 1164 * before this CPU's Phase-I as we held the read_lock. 1165 * 3. Some other CPU executing pcc CMD_READ has stolen the 1166 * down_write, in which case, send_pcc_cmd will check for pending 1167 * CMD_WRITE commands by checking the pending_pcc_write_cmd. 1168 * So this CPU can be certain that its request will be delivered 1169 * So in all cases, this CPU knows that its request will be delivered 1170 * by another CPU and can return 1171 * 1172 * After getting the down_write we still need to check for 1173 * pending_pcc_write_cmd to take care of the following scenario 1174 * The thread running this code could be scheduled out between 1175 * Phase-I and Phase-II. Before it is scheduled back on, another CPU 1176 * could have delivered the request to Platform by triggering the 1177 * doorbell and transferred the ownership of PCC to platform. So this 1178 * avoids triggering an unnecessary doorbell and more importantly before 1179 * triggering the doorbell it makes sure that the PCC channel ownership 1180 * is still with OSPM. 1181 * pending_pcc_write_cmd can also be cleared by a different CPU, if 1182 * there was a pcc CMD_READ waiting on down_write and it steals the lock 1183 * before the pcc CMD_WRITE is completed. pcc_send_cmd checks for this 1184 * case during a CMD_READ and if there are pending writes it delivers 1185 * the write command before servicing the read command 1186 */ 1187 if (CPC_IN_PCC(desired_reg)) { 1188 if (down_write_trylock(&pcc_data.pcc_lock)) { /* BEGIN Phase-II */ 1189 /* Update only if there are pending write commands */ 1190 if (pcc_data.pending_pcc_write_cmd) 1191 send_pcc_cmd(CMD_WRITE); 1192 up_write(&pcc_data.pcc_lock); /* END Phase-II */ 1193 } else 1194 /* Wait until pcc_write_cnt is updated by send_pcc_cmd */ 1195 wait_event(pcc_data.pcc_write_wait_q, 1196 cpc_desc->write_cmd_id != pcc_data.pcc_write_cnt); 1197 1198 /* send_pcc_cmd updates the status in case of failure */ 1199 ret = cpc_desc->write_cmd_status; 1200 } 1201 return ret; 1202 } 1203 EXPORT_SYMBOL_GPL(cppc_set_perf); 1204 1205 /** 1206 * cppc_get_transition_latency - returns frequency transition latency in ns 1207 * 1208 * ACPI CPPC does not explicitly specifiy how a platform can specify the 1209 * transition latency for perfromance change requests. The closest we have 1210 * is the timing information from the PCCT tables which provides the info 1211 * on the number and frequency of PCC commands the platform can handle. 1212 */ 1213 unsigned int cppc_get_transition_latency(int cpu_num) 1214 { 1215 /* 1216 * Expected transition latency is based on the PCCT timing values 1217 * Below are definition from ACPI spec: 1218 * pcc_nominal- Expected latency to process a command, in microseconds 1219 * pcc_mpar - The maximum number of periodic requests that the subspace 1220 * channel can support, reported in commands per minute. 0 1221 * indicates no limitation. 1222 * pcc_mrtt - The minimum amount of time that OSPM must wait after the 1223 * completion of a command before issuing the next command, 1224 * in microseconds. 1225 */ 1226 unsigned int latency_ns = 0; 1227 struct cpc_desc *cpc_desc; 1228 struct cpc_register_resource *desired_reg; 1229 1230 cpc_desc = per_cpu(cpc_desc_ptr, cpu_num); 1231 if (!cpc_desc) 1232 return CPUFREQ_ETERNAL; 1233 1234 desired_reg = &cpc_desc->cpc_regs[DESIRED_PERF]; 1235 if (!CPC_IN_PCC(desired_reg)) 1236 return CPUFREQ_ETERNAL; 1237 1238 if (pcc_data.pcc_mpar) 1239 latency_ns = 60 * (1000 * 1000 * 1000 / pcc_data.pcc_mpar); 1240 1241 latency_ns = max(latency_ns, pcc_data.pcc_nominal * 1000); 1242 latency_ns = max(latency_ns, pcc_data.pcc_mrtt * 1000); 1243 1244 return latency_ns; 1245 } 1246 EXPORT_SYMBOL_GPL(cppc_get_transition_latency); 1247