1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Hardware Feedback Interface Driver
4  *
5  * Copyright (c) 2021, Intel Corporation.
6  *
7  * Authors: Aubrey Li <aubrey.li@linux.intel.com>
8  *          Ricardo Neri <ricardo.neri-calderon@linux.intel.com>
9  *
10  *
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
16  * at boot.
17  *
18  * This file provides functionality to process HFI updates and relay these
19  * updates to userspace.
20  */
21 
22 #define pr_fmt(fmt)  "intel-hfi: " fmt
23 
24 #include <linux/bitops.h>
25 #include <linux/cpufeature.h>
26 #include <linux/cpumask.h>
27 #include <linux/gfp.h>
28 #include <linux/io.h>
29 #include <linux/kernel.h>
30 #include <linux/math.h>
31 #include <linux/mutex.h>
32 #include <linux/percpu-defs.h>
33 #include <linux/printk.h>
34 #include <linux/processor.h>
35 #include <linux/slab.h>
36 #include <linux/spinlock.h>
37 #include <linux/string.h>
38 #include <linux/topology.h>
39 #include <linux/workqueue.h>
40 
41 #include <asm/msr.h>
42 
43 #include "intel_hfi.h"
44 #include "thermal_interrupt.h"
45 
46 #include "../thermal_netlink.h"
47 
48 /* Hardware Feedback Interface MSR configuration bits */
49 #define HW_FEEDBACK_PTR_VALID_BIT		BIT(0)
50 #define HW_FEEDBACK_CONFIG_HFI_ENABLE_BIT	BIT(0)
51 
52 /* CPUID detection and enumeration definitions for HFI */
53 
54 #define CPUID_HFI_LEAF 6
55 
56 union hfi_capabilities {
57 	struct {
58 		u8	performance:1;
59 		u8	energy_efficiency:1;
60 		u8	__reserved:6;
61 	} split;
62 	u8 bits;
63 };
64 
65 union cpuid6_edx {
66 	struct {
67 		union hfi_capabilities	capabilities;
68 		u32			table_pages:4;
69 		u32			__reserved:4;
70 		s32			index:16;
71 	} split;
72 	u32 full;
73 };
74 
75 /**
76  * struct hfi_cpu_data - HFI capabilities per CPU
77  * @perf_cap:		Performance capability
78  * @ee_cap:		Energy efficiency capability
79  *
80  * Capabilities of a logical processor in the HFI table. These capabilities are
81  * unitless.
82  */
83 struct hfi_cpu_data {
84 	u8	perf_cap;
85 	u8	ee_cap;
86 } __packed;
87 
88 /**
89  * struct hfi_hdr - Header of the HFI table
90  * @perf_updated:	Hardware updated performance capabilities
91  * @ee_updated:		Hardware updated energy efficiency capabilities
92  *
93  * Properties of the data in an HFI table.
94  */
95 struct hfi_hdr {
96 	u8	perf_updated;
97 	u8	ee_updated;
98 } __packed;
99 
100 /**
101  * struct hfi_instance - Representation of an HFI instance (i.e., a table)
102  * @local_table:	Base of the local copy of the HFI table
103  * @timestamp:		Timestamp of the last update of the local table.
104  *			Located at the base of the local table.
105  * @hdr:		Base address of the header of the local table
106  * @data:		Base address of the data of the local table
107  * @cpus:		CPUs represented in this HFI table instance
108  * @hw_table:		Pointer to the HFI table of this instance
109  * @update_work:	Delayed work to process HFI updates
110  * @table_lock:		Lock to protect acceses to the table of this instance
111  * @event_lock:		Lock to process HFI interrupts
112  *
113  * A set of parameters to parse and navigate a specific HFI table.
114  */
115 struct hfi_instance {
116 	union {
117 		void			*local_table;
118 		u64			*timestamp;
119 	};
120 	void			*hdr;
121 	void			*data;
122 	cpumask_var_t		cpus;
123 	void			*hw_table;
124 	struct delayed_work	update_work;
125 	raw_spinlock_t		table_lock;
126 	raw_spinlock_t		event_lock;
127 };
128 
129 /**
130  * struct hfi_features - Supported HFI features
131  * @nr_table_pages:	Size of the HFI table in 4KB pages
132  * @cpu_stride:		Stride size to locate the capability data of a logical
133  *			processor within the table (i.e., row stride)
134  * @hdr_size:		Size of the table header
135  *
136  * Parameters and supported features that are common to all HFI instances
137  */
138 struct hfi_features {
139 	size_t		nr_table_pages;
140 	unsigned int	cpu_stride;
141 	unsigned int	hdr_size;
142 };
143 
144 /**
145  * struct hfi_cpu_info - Per-CPU attributes to consume HFI data
146  * @index:		Row of this CPU in its HFI table
147  * @hfi_instance:	Attributes of the HFI table to which this CPU belongs
148  *
149  * Parameters to link a logical processor to an HFI table and a row within it.
150  */
151 struct hfi_cpu_info {
152 	s16			index;
153 	struct hfi_instance	*hfi_instance;
154 };
155 
156 static DEFINE_PER_CPU(struct hfi_cpu_info, hfi_cpu_info) = { .index = -1 };
157 
158 static int max_hfi_instances;
159 static struct hfi_instance *hfi_instances;
160 
161 static struct hfi_features hfi_features;
162 static DEFINE_MUTEX(hfi_instance_lock);
163 
164 static struct workqueue_struct *hfi_updates_wq;
165 #define HFI_UPDATE_INTERVAL		HZ
166 #define HFI_MAX_THERM_NOTIFY_COUNT	16
167 
168 static void get_hfi_caps(struct hfi_instance *hfi_instance,
169 			 struct thermal_genl_cpu_caps *cpu_caps)
170 {
171 	int cpu, i = 0;
172 
173 	raw_spin_lock_irq(&hfi_instance->table_lock);
174 	for_each_cpu(cpu, hfi_instance->cpus) {
175 		struct hfi_cpu_data *caps;
176 		s16 index;
177 
178 		index = per_cpu(hfi_cpu_info, cpu).index;
179 		caps = hfi_instance->data + index * hfi_features.cpu_stride;
180 		cpu_caps[i].cpu = cpu;
181 
182 		/*
183 		 * Scale performance and energy efficiency to
184 		 * the [0, 1023] interval that thermal netlink uses.
185 		 */
186 		cpu_caps[i].performance = caps->perf_cap << 2;
187 		cpu_caps[i].efficiency = caps->ee_cap << 2;
188 
189 		++i;
190 	}
191 	raw_spin_unlock_irq(&hfi_instance->table_lock);
192 }
193 
194 /*
195  * Call update_capabilities() when there are changes in the HFI table.
196  */
197 static void update_capabilities(struct hfi_instance *hfi_instance)
198 {
199 	struct thermal_genl_cpu_caps *cpu_caps;
200 	int i = 0, cpu_count;
201 
202 	/* CPUs may come online/offline while processing an HFI update. */
203 	mutex_lock(&hfi_instance_lock);
204 
205 	cpu_count = cpumask_weight(hfi_instance->cpus);
206 
207 	/* No CPUs to report in this hfi_instance. */
208 	if (!cpu_count)
209 		goto out;
210 
211 	cpu_caps = kcalloc(cpu_count, sizeof(*cpu_caps), GFP_KERNEL);
212 	if (!cpu_caps)
213 		goto out;
214 
215 	get_hfi_caps(hfi_instance, cpu_caps);
216 
217 	if (cpu_count < HFI_MAX_THERM_NOTIFY_COUNT)
218 		goto last_cmd;
219 
220 	/* Process complete chunks of HFI_MAX_THERM_NOTIFY_COUNT capabilities. */
221 	for (i = 0;
222 	     (i + HFI_MAX_THERM_NOTIFY_COUNT) <= cpu_count;
223 	     i += HFI_MAX_THERM_NOTIFY_COUNT)
224 		thermal_genl_cpu_capability_event(HFI_MAX_THERM_NOTIFY_COUNT,
225 						  &cpu_caps[i]);
226 
227 	cpu_count = cpu_count - i;
228 
229 last_cmd:
230 	/* Process the remaining capabilities if any. */
231 	if (cpu_count)
232 		thermal_genl_cpu_capability_event(cpu_count, &cpu_caps[i]);
233 
234 	kfree(cpu_caps);
235 out:
236 	mutex_unlock(&hfi_instance_lock);
237 }
238 
239 static void hfi_update_work_fn(struct work_struct *work)
240 {
241 	struct hfi_instance *hfi_instance;
242 
243 	hfi_instance = container_of(to_delayed_work(work), struct hfi_instance,
244 				    update_work);
245 
246 	update_capabilities(hfi_instance);
247 }
248 
249 void intel_hfi_process_event(__u64 pkg_therm_status_msr_val)
250 {
251 	struct hfi_instance *hfi_instance;
252 	int cpu = smp_processor_id();
253 	struct hfi_cpu_info *info;
254 	u64 new_timestamp, msr, hfi;
255 
256 	if (!pkg_therm_status_msr_val)
257 		return;
258 
259 	info = &per_cpu(hfi_cpu_info, cpu);
260 	if (!info)
261 		return;
262 
263 	/*
264 	 * A CPU is linked to its HFI instance before the thermal vector in the
265 	 * local APIC is unmasked. Hence, info->hfi_instance cannot be NULL
266 	 * when receiving an HFI event.
267 	 */
268 	hfi_instance = info->hfi_instance;
269 	if (unlikely(!hfi_instance)) {
270 		pr_debug("Received event on CPU %d but instance was null", cpu);
271 		return;
272 	}
273 
274 	/*
275 	 * On most systems, all CPUs in the package receive a package-level
276 	 * thermal interrupt when there is an HFI update. It is sufficient to
277 	 * let a single CPU to acknowledge the update and queue work to
278 	 * process it. The remaining CPUs can resume their work.
279 	 */
280 	if (!raw_spin_trylock(&hfi_instance->event_lock))
281 		return;
282 
283 	rdmsrl(MSR_IA32_PACKAGE_THERM_STATUS, msr);
284 	hfi = msr & PACKAGE_THERM_STATUS_HFI_UPDATED;
285 	if (!hfi) {
286 		raw_spin_unlock(&hfi_instance->event_lock);
287 		return;
288 	}
289 
290 	/*
291 	 * Ack duplicate update. Since there is an active HFI
292 	 * status from HW, it must be a new event, not a case
293 	 * where a lagging CPU entered the locked region.
294 	 */
295 	new_timestamp = *(u64 *)hfi_instance->hw_table;
296 	if (*hfi_instance->timestamp == new_timestamp) {
297 		thermal_clear_package_intr_status(PACKAGE_LEVEL, PACKAGE_THERM_STATUS_HFI_UPDATED);
298 		raw_spin_unlock(&hfi_instance->event_lock);
299 		return;
300 	}
301 
302 	raw_spin_lock(&hfi_instance->table_lock);
303 
304 	/*
305 	 * Copy the updated table into our local copy. This includes the new
306 	 * timestamp.
307 	 */
308 	memcpy(hfi_instance->local_table, hfi_instance->hw_table,
309 	       hfi_features.nr_table_pages << PAGE_SHIFT);
310 
311 	/*
312 	 * Let hardware know that we are done reading the HFI table and it is
313 	 * free to update it again.
314 	 */
315 	thermal_clear_package_intr_status(PACKAGE_LEVEL, PACKAGE_THERM_STATUS_HFI_UPDATED);
316 
317 	raw_spin_unlock(&hfi_instance->table_lock);
318 	raw_spin_unlock(&hfi_instance->event_lock);
319 
320 	queue_delayed_work(hfi_updates_wq, &hfi_instance->update_work,
321 			   HFI_UPDATE_INTERVAL);
322 }
323 
324 static void init_hfi_cpu_index(struct hfi_cpu_info *info)
325 {
326 	union cpuid6_edx edx;
327 
328 	/* Do not re-read @cpu's index if it has already been initialized. */
329 	if (info->index > -1)
330 		return;
331 
332 	edx.full = cpuid_edx(CPUID_HFI_LEAF);
333 	info->index = edx.split.index;
334 }
335 
336 /*
337  * The format of the HFI table depends on the number of capabilities that the
338  * hardware supports. Keep a data structure to navigate the table.
339  */
340 static void init_hfi_instance(struct hfi_instance *hfi_instance)
341 {
342 	/* The HFI header is below the time-stamp. */
343 	hfi_instance->hdr = hfi_instance->local_table +
344 			    sizeof(*hfi_instance->timestamp);
345 
346 	/* The HFI data starts below the header. */
347 	hfi_instance->data = hfi_instance->hdr + hfi_features.hdr_size;
348 }
349 
350 /**
351  * intel_hfi_online() - Enable HFI on @cpu
352  * @cpu:	CPU in which the HFI will be enabled
353  *
354  * Enable the HFI to be used in @cpu. The HFI is enabled at the die/package
355  * level. The first CPU in the die/package to come online does the full HFI
356  * initialization. Subsequent CPUs will just link themselves to the HFI
357  * instance of their die/package.
358  *
359  * This function is called before enabling the thermal vector in the local APIC
360  * in order to ensure that @cpu has an associated HFI instance when it receives
361  * an HFI event.
362  */
363 void intel_hfi_online(unsigned int cpu)
364 {
365 	struct hfi_instance *hfi_instance;
366 	struct hfi_cpu_info *info;
367 	phys_addr_t hw_table_pa;
368 	u64 msr_val;
369 	u16 die_id;
370 
371 	/* Nothing to do if hfi_instances are missing. */
372 	if (!hfi_instances)
373 		return;
374 
375 	/*
376 	 * Link @cpu to the HFI instance of its package/die. It does not
377 	 * matter whether the instance has been initialized.
378 	 */
379 	info = &per_cpu(hfi_cpu_info, cpu);
380 	die_id = topology_logical_die_id(cpu);
381 	hfi_instance = info->hfi_instance;
382 	if (!hfi_instance) {
383 		if (die_id >= max_hfi_instances)
384 			return;
385 
386 		hfi_instance = &hfi_instances[die_id];
387 		info->hfi_instance = hfi_instance;
388 	}
389 
390 	init_hfi_cpu_index(info);
391 
392 	/*
393 	 * Now check if the HFI instance of the package/die of @cpu has been
394 	 * initialized (by checking its header). In such case, all we have to
395 	 * do is to add @cpu to this instance's cpumask.
396 	 */
397 	mutex_lock(&hfi_instance_lock);
398 	if (hfi_instance->hdr) {
399 		cpumask_set_cpu(cpu, hfi_instance->cpus);
400 		goto unlock;
401 	}
402 
403 	/*
404 	 * Hardware is programmed with the physical address of the first page
405 	 * frame of the table. Hence, the allocated memory must be page-aligned.
406 	 */
407 	hfi_instance->hw_table = alloc_pages_exact(hfi_features.nr_table_pages,
408 						   GFP_KERNEL | __GFP_ZERO);
409 	if (!hfi_instance->hw_table)
410 		goto unlock;
411 
412 	hw_table_pa = virt_to_phys(hfi_instance->hw_table);
413 
414 	/*
415 	 * Allocate memory to keep a local copy of the table that
416 	 * hardware generates.
417 	 */
418 	hfi_instance->local_table = kzalloc(hfi_features.nr_table_pages << PAGE_SHIFT,
419 					    GFP_KERNEL);
420 	if (!hfi_instance->local_table)
421 		goto free_hw_table;
422 
423 	/*
424 	 * Program the address of the feedback table of this die/package. On
425 	 * some processors, hardware remembers the old address of the HFI table
426 	 * even after having been reprogrammed and re-enabled. Thus, do not free
427 	 * the pages allocated for the table or reprogram the hardware with a
428 	 * new base address. Namely, program the hardware only once.
429 	 */
430 	msr_val = hw_table_pa | HW_FEEDBACK_PTR_VALID_BIT;
431 	wrmsrl(MSR_IA32_HW_FEEDBACK_PTR, msr_val);
432 
433 	init_hfi_instance(hfi_instance);
434 
435 	INIT_DELAYED_WORK(&hfi_instance->update_work, hfi_update_work_fn);
436 	raw_spin_lock_init(&hfi_instance->table_lock);
437 	raw_spin_lock_init(&hfi_instance->event_lock);
438 
439 	cpumask_set_cpu(cpu, hfi_instance->cpus);
440 
441 	/*
442 	 * Enable the hardware feedback interface and never disable it. See
443 	 * comment on programming the address of the table.
444 	 */
445 	rdmsrl(MSR_IA32_HW_FEEDBACK_CONFIG, msr_val);
446 	msr_val |= HW_FEEDBACK_CONFIG_HFI_ENABLE_BIT;
447 	wrmsrl(MSR_IA32_HW_FEEDBACK_CONFIG, msr_val);
448 
449 unlock:
450 	mutex_unlock(&hfi_instance_lock);
451 	return;
452 
453 free_hw_table:
454 	free_pages_exact(hfi_instance->hw_table, hfi_features.nr_table_pages);
455 	goto unlock;
456 }
457 
458 /**
459  * intel_hfi_offline() - Disable HFI on @cpu
460  * @cpu:	CPU in which the HFI will be disabled
461  *
462  * Remove @cpu from those covered by its HFI instance.
463  *
464  * On some processors, hardware remembers previous programming settings even
465  * after being reprogrammed. Thus, keep HFI enabled even if all CPUs in the
466  * die/package of @cpu are offline. See note in intel_hfi_online().
467  */
468 void intel_hfi_offline(unsigned int cpu)
469 {
470 	struct hfi_cpu_info *info = &per_cpu(hfi_cpu_info, cpu);
471 	struct hfi_instance *hfi_instance;
472 
473 	/*
474 	 * Check if @cpu as an associated, initialized (i.e., with a non-NULL
475 	 * header). Also, HFI instances are only initialized if X86_FEATURE_HFI
476 	 * is present.
477 	 */
478 	hfi_instance = info->hfi_instance;
479 	if (!hfi_instance)
480 		return;
481 
482 	if (!hfi_instance->hdr)
483 		return;
484 
485 	mutex_lock(&hfi_instance_lock);
486 	cpumask_clear_cpu(cpu, hfi_instance->cpus);
487 	mutex_unlock(&hfi_instance_lock);
488 }
489 
490 static __init int hfi_parse_features(void)
491 {
492 	unsigned int nr_capabilities;
493 	union cpuid6_edx edx;
494 
495 	if (!boot_cpu_has(X86_FEATURE_HFI))
496 		return -ENODEV;
497 
498 	/*
499 	 * If we are here we know that CPUID_HFI_LEAF exists. Parse the
500 	 * supported capabilities and the size of the HFI table.
501 	 */
502 	edx.full = cpuid_edx(CPUID_HFI_LEAF);
503 
504 	if (!edx.split.capabilities.split.performance) {
505 		pr_debug("Performance reporting not supported! Not using HFI\n");
506 		return -ENODEV;
507 	}
508 
509 	/*
510 	 * The number of supported capabilities determines the number of
511 	 * columns in the HFI table. Exclude the reserved bits.
512 	 */
513 	edx.split.capabilities.split.__reserved = 0;
514 	nr_capabilities = hweight8(edx.split.capabilities.bits);
515 
516 	/* The number of 4KB pages required by the table */
517 	hfi_features.nr_table_pages = edx.split.table_pages + 1;
518 
519 	/*
520 	 * The header contains change indications for each supported feature.
521 	 * The size of the table header is rounded up to be a multiple of 8
522 	 * bytes.
523 	 */
524 	hfi_features.hdr_size = DIV_ROUND_UP(nr_capabilities, 8) * 8;
525 
526 	/*
527 	 * Data of each logical processor is also rounded up to be a multiple
528 	 * of 8 bytes.
529 	 */
530 	hfi_features.cpu_stride = DIV_ROUND_UP(nr_capabilities, 8) * 8;
531 
532 	return 0;
533 }
534 
535 void __init intel_hfi_init(void)
536 {
537 	struct hfi_instance *hfi_instance;
538 	int i, j;
539 
540 	if (hfi_parse_features())
541 		return;
542 
543 	/* There is one HFI instance per die/package. */
544 	max_hfi_instances = topology_max_packages() *
545 			    topology_max_die_per_package();
546 
547 	/*
548 	 * This allocation may fail. CPU hotplug callbacks must check
549 	 * for a null pointer.
550 	 */
551 	hfi_instances = kcalloc(max_hfi_instances, sizeof(*hfi_instances),
552 				GFP_KERNEL);
553 	if (!hfi_instances)
554 		return;
555 
556 	for (i = 0; i < max_hfi_instances; i++) {
557 		hfi_instance = &hfi_instances[i];
558 		if (!zalloc_cpumask_var(&hfi_instance->cpus, GFP_KERNEL))
559 			goto err_nomem;
560 	}
561 
562 	hfi_updates_wq = create_singlethread_workqueue("hfi-updates");
563 	if (!hfi_updates_wq)
564 		goto err_nomem;
565 
566 	return;
567 
568 err_nomem:
569 	for (j = 0; j < i; ++j) {
570 		hfi_instance = &hfi_instances[j];
571 		free_cpumask_var(hfi_instance->cpus);
572 	}
573 
574 	kfree(hfi_instances);
575 	hfi_instances = NULL;
576 }
577