1 // SPDX-License-Identifier: GPL-2.0-only
3 * Hardware Feedback Interface Driver
5 * Copyright (c) 2021, Intel Corporation.
7 * Authors: Aubrey Li <aubrey.li@linux.intel.com>
8 * Ricardo Neri <ricardo.neri-calderon@linux.intel.com>
11 * The Hardware Feedback Interface provides a performance and energy efficiency
12 * capability information for each CPU in the system. Depending on the processor
13 * model, hardware may periodically update these capabilities as a result of
14 * changes in the operating conditions (e.g., power limits or thermal
15 * constraints). On other processor models, there is a single HFI update
18 * This file provides functionality to process HFI updates and relay these
19 * updates to userspace.
22 #define pr_fmt(fmt) "intel-hfi: " fmt
24 #include <linux/bitops.h>
25 #include <linux/cpufeature.h>
26 #include <linux/cpumask.h>
27 #include <linux/delay.h>
28 #include <linux/gfp.h>
30 #include <linux/kernel.h>
31 #include <linux/math.h>
32 #include <linux/mutex.h>
33 #include <linux/percpu-defs.h>
34 #include <linux/printk.h>
35 #include <linux/processor.h>
36 #include <linux/slab.h>
37 #include <linux/spinlock.h>
38 #include <linux/suspend.h>
39 #include <linux/string.h>
40 #include <linux/syscore_ops.h>
41 #include <linux/topology.h>
42 #include <linux/workqueue.h>
46 #include "intel_hfi.h"
47 #include "thermal_interrupt.h"
49 #include "../thermal_netlink.h"
51 /* Hardware Feedback Interface MSR configuration bits */
52 #define HW_FEEDBACK_PTR_VALID_BIT BIT(0)
53 #define HW_FEEDBACK_CONFIG_HFI_ENABLE_BIT BIT(0)
55 /* CPUID detection and enumeration definitions for HFI */
57 #define CPUID_HFI_LEAF 6
59 union hfi_capabilities {
62 u8 energy_efficiency:1;
70 union hfi_capabilities capabilities;
79 * struct hfi_cpu_data - HFI capabilities per CPU
80 * @perf_cap: Performance capability
81 * @ee_cap: Energy efficiency capability
83 * Capabilities of a logical processor in the HFI table. These capabilities are
92 * struct hfi_hdr - Header of the HFI table
93 * @perf_updated: Hardware updated performance capabilities
94 * @ee_updated: Hardware updated energy efficiency capabilities
96 * Properties of the data in an HFI table.
104 * struct hfi_instance - Representation of an HFI instance (i.e., a table)
105 * @local_table: Base of the local copy of the HFI table
106 * @timestamp: Timestamp of the last update of the local table.
107 * Located at the base of the local table.
108 * @hdr: Base address of the header of the local table
109 * @data: Base address of the data of the local table
110 * @cpus: CPUs represented in this HFI table instance
111 * @hw_table: Pointer to the HFI table of this instance
112 * @update_work: Delayed work to process HFI updates
113 * @table_lock: Lock to protect acceses to the table of this instance
114 * @event_lock: Lock to process HFI interrupts
116 * A set of parameters to parse and navigate a specific HFI table.
118 struct hfi_instance {
127 struct delayed_work update_work;
128 raw_spinlock_t table_lock;
129 raw_spinlock_t event_lock;
133 * struct hfi_features - Supported HFI features
134 * @nr_table_pages: Size of the HFI table in 4KB pages
135 * @cpu_stride: Stride size to locate the capability data of a logical
136 * processor within the table (i.e., row stride)
137 * @hdr_size: Size of the table header
139 * Parameters and supported features that are common to all HFI instances
141 struct hfi_features {
142 size_t nr_table_pages;
143 unsigned int cpu_stride;
144 unsigned int hdr_size;
148 * struct hfi_cpu_info - Per-CPU attributes to consume HFI data
149 * @index: Row of this CPU in its HFI table
150 * @hfi_instance: Attributes of the HFI table to which this CPU belongs
152 * Parameters to link a logical processor to an HFI table and a row within it.
154 struct hfi_cpu_info {
156 struct hfi_instance *hfi_instance;
159 static DEFINE_PER_CPU(struct hfi_cpu_info, hfi_cpu_info) = { .index = -1 };
161 static int max_hfi_instances;
162 static struct hfi_instance *hfi_instances;
164 static struct hfi_features hfi_features;
165 static DEFINE_MUTEX(hfi_instance_lock);
167 static struct workqueue_struct *hfi_updates_wq;
168 #define HFI_UPDATE_INTERVAL HZ
169 #define HFI_MAX_THERM_NOTIFY_COUNT 16
171 static void get_hfi_caps(struct hfi_instance *hfi_instance,
172 struct thermal_genl_cpu_caps *cpu_caps)
176 raw_spin_lock_irq(&hfi_instance->table_lock);
177 for_each_cpu(cpu, hfi_instance->cpus) {
178 struct hfi_cpu_data *caps;
181 index = per_cpu(hfi_cpu_info, cpu).index;
182 caps = hfi_instance->data + index * hfi_features.cpu_stride;
183 cpu_caps[i].cpu = cpu;
186 * Scale performance and energy efficiency to
187 * the [0, 1023] interval that thermal netlink uses.
189 cpu_caps[i].performance = caps->perf_cap << 2;
190 cpu_caps[i].efficiency = caps->ee_cap << 2;
194 raw_spin_unlock_irq(&hfi_instance->table_lock);
198 * Call update_capabilities() when there are changes in the HFI table.
200 static void update_capabilities(struct hfi_instance *hfi_instance)
202 struct thermal_genl_cpu_caps *cpu_caps;
203 int i = 0, cpu_count;
205 /* CPUs may come online/offline while processing an HFI update. */
206 mutex_lock(&hfi_instance_lock);
208 cpu_count = cpumask_weight(hfi_instance->cpus);
210 /* No CPUs to report in this hfi_instance. */
214 cpu_caps = kcalloc(cpu_count, sizeof(*cpu_caps), GFP_KERNEL);
218 get_hfi_caps(hfi_instance, cpu_caps);
220 if (cpu_count < HFI_MAX_THERM_NOTIFY_COUNT)
223 /* Process complete chunks of HFI_MAX_THERM_NOTIFY_COUNT capabilities. */
225 (i + HFI_MAX_THERM_NOTIFY_COUNT) <= cpu_count;
226 i += HFI_MAX_THERM_NOTIFY_COUNT)
227 thermal_genl_cpu_capability_event(HFI_MAX_THERM_NOTIFY_COUNT,
230 cpu_count = cpu_count - i;
233 /* Process the remaining capabilities if any. */
235 thermal_genl_cpu_capability_event(cpu_count, &cpu_caps[i]);
239 mutex_unlock(&hfi_instance_lock);
242 static void hfi_update_work_fn(struct work_struct *work)
244 struct hfi_instance *hfi_instance;
246 hfi_instance = container_of(to_delayed_work(work), struct hfi_instance,
249 update_capabilities(hfi_instance);
252 void intel_hfi_process_event(__u64 pkg_therm_status_msr_val)
254 struct hfi_instance *hfi_instance;
255 int cpu = smp_processor_id();
256 struct hfi_cpu_info *info;
257 u64 new_timestamp, msr, hfi;
259 if (!pkg_therm_status_msr_val)
262 info = &per_cpu(hfi_cpu_info, cpu);
267 * A CPU is linked to its HFI instance before the thermal vector in the
268 * local APIC is unmasked. Hence, info->hfi_instance cannot be NULL
269 * when receiving an HFI event.
271 hfi_instance = info->hfi_instance;
272 if (unlikely(!hfi_instance)) {
273 pr_debug("Received event on CPU %d but instance was null", cpu);
278 * On most systems, all CPUs in the package receive a package-level
279 * thermal interrupt when there is an HFI update. It is sufficient to
280 * let a single CPU to acknowledge the update and queue work to
281 * process it. The remaining CPUs can resume their work.
283 if (!raw_spin_trylock(&hfi_instance->event_lock))
286 rdmsrl(MSR_IA32_PACKAGE_THERM_STATUS, msr);
287 hfi = msr & PACKAGE_THERM_STATUS_HFI_UPDATED;
289 raw_spin_unlock(&hfi_instance->event_lock);
294 * Ack duplicate update. Since there is an active HFI
295 * status from HW, it must be a new event, not a case
296 * where a lagging CPU entered the locked region.
298 new_timestamp = *(u64 *)hfi_instance->hw_table;
299 if (*hfi_instance->timestamp == new_timestamp) {
300 thermal_clear_package_intr_status(PACKAGE_LEVEL, PACKAGE_THERM_STATUS_HFI_UPDATED);
301 raw_spin_unlock(&hfi_instance->event_lock);
305 raw_spin_lock(&hfi_instance->table_lock);
308 * Copy the updated table into our local copy. This includes the new
311 memcpy(hfi_instance->local_table, hfi_instance->hw_table,
312 hfi_features.nr_table_pages << PAGE_SHIFT);
315 * Let hardware know that we are done reading the HFI table and it is
316 * free to update it again.
318 thermal_clear_package_intr_status(PACKAGE_LEVEL, PACKAGE_THERM_STATUS_HFI_UPDATED);
320 raw_spin_unlock(&hfi_instance->table_lock);
321 raw_spin_unlock(&hfi_instance->event_lock);
323 queue_delayed_work(hfi_updates_wq, &hfi_instance->update_work,
324 HFI_UPDATE_INTERVAL);
327 static void init_hfi_cpu_index(struct hfi_cpu_info *info)
329 union cpuid6_edx edx;
331 /* Do not re-read @cpu's index if it has already been initialized. */
332 if (info->index > -1)
335 edx.full = cpuid_edx(CPUID_HFI_LEAF);
336 info->index = edx.split.index;
340 * The format of the HFI table depends on the number of capabilities that the
341 * hardware supports. Keep a data structure to navigate the table.
343 static void init_hfi_instance(struct hfi_instance *hfi_instance)
345 /* The HFI header is below the time-stamp. */
346 hfi_instance->hdr = hfi_instance->local_table +
347 sizeof(*hfi_instance->timestamp);
349 /* The HFI data starts below the header. */
350 hfi_instance->data = hfi_instance->hdr + hfi_features.hdr_size;
353 /* Caller must hold hfi_instance_lock. */
354 static void hfi_enable(void)
358 rdmsrl(MSR_IA32_HW_FEEDBACK_CONFIG, msr_val);
359 msr_val |= HW_FEEDBACK_CONFIG_HFI_ENABLE_BIT;
360 wrmsrl(MSR_IA32_HW_FEEDBACK_CONFIG, msr_val);
363 static void hfi_set_hw_table(struct hfi_instance *hfi_instance)
365 phys_addr_t hw_table_pa;
368 hw_table_pa = virt_to_phys(hfi_instance->hw_table);
369 msr_val = hw_table_pa | HW_FEEDBACK_PTR_VALID_BIT;
370 wrmsrl(MSR_IA32_HW_FEEDBACK_PTR, msr_val);
373 /* Caller must hold hfi_instance_lock. */
374 static void hfi_disable(void)
379 rdmsrl(MSR_IA32_HW_FEEDBACK_CONFIG, msr_val);
380 msr_val &= ~HW_FEEDBACK_CONFIG_HFI_ENABLE_BIT;
381 wrmsrl(MSR_IA32_HW_FEEDBACK_CONFIG, msr_val);
384 * Wait for hardware to acknowledge the disabling of HFI. Some
385 * processors may not do it. Wait for ~2ms. This is a reasonable
386 * time for hardware to complete any pending actions on the HFI
389 for (i = 0; i < 2000; i++) {
390 rdmsrl(MSR_IA32_PACKAGE_THERM_STATUS, msr_val);
391 if (msr_val & PACKAGE_THERM_STATUS_HFI_UPDATED)
400 * intel_hfi_online() - Enable HFI on @cpu
401 * @cpu: CPU in which the HFI will be enabled
403 * Enable the HFI to be used in @cpu. The HFI is enabled at the die/package
404 * level. The first CPU in the die/package to come online does the full HFI
405 * initialization. Subsequent CPUs will just link themselves to the HFI
406 * instance of their die/package.
408 * This function is called before enabling the thermal vector in the local APIC
409 * in order to ensure that @cpu has an associated HFI instance when it receives
412 void intel_hfi_online(unsigned int cpu)
414 struct hfi_instance *hfi_instance;
415 struct hfi_cpu_info *info;
418 /* Nothing to do if hfi_instances are missing. */
423 * Link @cpu to the HFI instance of its package/die. It does not
424 * matter whether the instance has been initialized.
426 info = &per_cpu(hfi_cpu_info, cpu);
427 die_id = topology_logical_die_id(cpu);
428 hfi_instance = info->hfi_instance;
430 if (die_id >= max_hfi_instances)
433 hfi_instance = &hfi_instances[die_id];
434 info->hfi_instance = hfi_instance;
437 init_hfi_cpu_index(info);
440 * Now check if the HFI instance of the package/die of @cpu has been
441 * initialized (by checking its header). In such case, all we have to
442 * do is to add @cpu to this instance's cpumask.
444 mutex_lock(&hfi_instance_lock);
445 if (hfi_instance->hdr) {
446 cpumask_set_cpu(cpu, hfi_instance->cpus);
451 * Hardware is programmed with the physical address of the first page
452 * frame of the table. Hence, the allocated memory must be page-aligned.
454 * Some processors do not forget the initial address of the HFI table
455 * even after having been reprogrammed. Keep using the same pages. Do
458 hfi_instance->hw_table = alloc_pages_exact(hfi_features.nr_table_pages,
459 GFP_KERNEL | __GFP_ZERO);
460 if (!hfi_instance->hw_table)
464 * Allocate memory to keep a local copy of the table that
465 * hardware generates.
467 hfi_instance->local_table = kzalloc(hfi_features.nr_table_pages << PAGE_SHIFT,
469 if (!hfi_instance->local_table)
472 init_hfi_instance(hfi_instance);
474 INIT_DELAYED_WORK(&hfi_instance->update_work, hfi_update_work_fn);
475 raw_spin_lock_init(&hfi_instance->table_lock);
476 raw_spin_lock_init(&hfi_instance->event_lock);
478 cpumask_set_cpu(cpu, hfi_instance->cpus);
480 hfi_set_hw_table(hfi_instance);
484 mutex_unlock(&hfi_instance_lock);
488 free_pages_exact(hfi_instance->hw_table, hfi_features.nr_table_pages);
493 * intel_hfi_offline() - Disable HFI on @cpu
494 * @cpu: CPU in which the HFI will be disabled
496 * Remove @cpu from those covered by its HFI instance.
498 * On some processors, hardware remembers previous programming settings even
499 * after being reprogrammed. Thus, keep HFI enabled even if all CPUs in the
500 * die/package of @cpu are offline. See note in intel_hfi_online().
502 void intel_hfi_offline(unsigned int cpu)
504 struct hfi_cpu_info *info = &per_cpu(hfi_cpu_info, cpu);
505 struct hfi_instance *hfi_instance;
508 * Check if @cpu as an associated, initialized (i.e., with a non-NULL
509 * header). Also, HFI instances are only initialized if X86_FEATURE_HFI
512 hfi_instance = info->hfi_instance;
516 if (!hfi_instance->hdr)
519 mutex_lock(&hfi_instance_lock);
520 cpumask_clear_cpu(cpu, hfi_instance->cpus);
522 if (!cpumask_weight(hfi_instance->cpus))
525 mutex_unlock(&hfi_instance_lock);
528 static __init int hfi_parse_features(void)
530 unsigned int nr_capabilities;
531 union cpuid6_edx edx;
533 if (!boot_cpu_has(X86_FEATURE_HFI))
537 * If we are here we know that CPUID_HFI_LEAF exists. Parse the
538 * supported capabilities and the size of the HFI table.
540 edx.full = cpuid_edx(CPUID_HFI_LEAF);
542 if (!edx.split.capabilities.split.performance) {
543 pr_debug("Performance reporting not supported! Not using HFI\n");
548 * The number of supported capabilities determines the number of
549 * columns in the HFI table. Exclude the reserved bits.
551 edx.split.capabilities.split.__reserved = 0;
552 nr_capabilities = hweight8(edx.split.capabilities.bits);
554 /* The number of 4KB pages required by the table */
555 hfi_features.nr_table_pages = edx.split.table_pages + 1;
558 * The header contains change indications for each supported feature.
559 * The size of the table header is rounded up to be a multiple of 8
562 hfi_features.hdr_size = DIV_ROUND_UP(nr_capabilities, 8) * 8;
565 * Data of each logical processor is also rounded up to be a multiple
568 hfi_features.cpu_stride = DIV_ROUND_UP(nr_capabilities, 8) * 8;
573 static void hfi_do_enable(void)
575 /* This code runs only on the boot CPU. */
576 struct hfi_cpu_info *info = &per_cpu(hfi_cpu_info, 0);
577 struct hfi_instance *hfi_instance = info->hfi_instance;
579 /* No locking needed. There is no concurrency with CPU online. */
580 hfi_set_hw_table(hfi_instance);
584 static int hfi_do_disable(void)
586 /* No locking needed. There is no concurrency with CPU offline. */
592 static struct syscore_ops hfi_pm_ops = {
593 .resume = hfi_do_enable,
594 .suspend = hfi_do_disable,
597 void __init intel_hfi_init(void)
599 struct hfi_instance *hfi_instance;
602 if (hfi_parse_features())
605 /* There is one HFI instance per die/package. */
606 max_hfi_instances = topology_max_packages() *
607 topology_max_die_per_package();
610 * This allocation may fail. CPU hotplug callbacks must check
611 * for a null pointer.
613 hfi_instances = kcalloc(max_hfi_instances, sizeof(*hfi_instances),
618 for (i = 0; i < max_hfi_instances; i++) {
619 hfi_instance = &hfi_instances[i];
620 if (!zalloc_cpumask_var(&hfi_instance->cpus, GFP_KERNEL))
624 hfi_updates_wq = create_singlethread_workqueue("hfi-updates");
628 register_syscore_ops(&hfi_pm_ops);
633 for (j = 0; j < i; ++j) {
634 hfi_instance = &hfi_instances[j];
635 free_cpumask_var(hfi_instance->cpus);
638 kfree(hfi_instances);
639 hfi_instances = NULL;