1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright 2012 by Oracle Inc
6 * This code borrows ideas from https://lkml.org/lkml/2011/11/30/249
11 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13 #include <linux/cpumask.h>
14 #include <linux/cpufreq.h>
15 #include <linux/freezer.h>
16 #include <linux/kernel.h>
17 #include <linux/kthread.h>
18 #include <linux/init.h>
19 #include <linux/module.h>
20 #include <linux/types.h>
21 #include <linux/syscore_ops.h>
22 #include <linux/acpi.h>
23 #include <acpi/processor.h>
25 #include <xen/interface/platform.h>
26 #include <asm/xen/hypercall.h>
28 static int no_hypercall;
29 MODULE_PARM_DESC(off, "Inhibit the hypercall.");
30 module_param_named(off, no_hypercall, int, 0400);
33 * Note: Do not convert the acpi_id* below to cpumask_var_t or use cpumask_bit
34 * - as those shrink to nr_cpu_bits (which is dependent on possible_cpu), which
35 * can be less than what we want to put in. Instead use the 'nr_acpi_bits'
36 * which is dynamically computed based on the MADT or x2APIC table.
38 static unsigned int nr_acpi_bits;
39 /* Mutex to protect the acpi_ids_done - for CPU hotplug use. */
40 static DEFINE_MUTEX(acpi_ids_mutex);
41 /* Which ACPI ID we have processed from 'struct acpi_processor'. */
42 static unsigned long *acpi_ids_done;
43 /* Which ACPI ID exist in the SSDT/DSDT processor definitions. */
44 static unsigned long *acpi_id_present;
45 /* And if there is an _CST definition (or a PBLK) for the ACPI IDs */
46 static unsigned long *acpi_id_cst_present;
47 /* Which ACPI P-State dependencies for a enumerated processor */
48 static struct acpi_psd_package *acpi_psd;
50 static int push_cxx_to_hypervisor(struct acpi_processor *_pr)
52 struct xen_platform_op op = {
53 .cmd = XENPF_set_processor_pminfo,
54 .interface_version = XENPF_INTERFACE_VERSION,
55 .u.set_pminfo.id = _pr->acpi_id,
56 .u.set_pminfo.type = XEN_PM_CX,
58 struct xen_processor_cx *dst_cx, *dst_cx_states = NULL;
59 struct acpi_processor_cx *cx;
63 dst_cx_states = kcalloc(_pr->power.count,
64 sizeof(struct xen_processor_cx), GFP_KERNEL);
68 for (ok = 0, i = 1; i <= _pr->power.count; i++) {
69 cx = &_pr->power.states[i];
73 dst_cx = &(dst_cx_states[ok++]);
75 dst_cx->reg.space_id = ACPI_ADR_SPACE_SYSTEM_IO;
76 if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) {
77 dst_cx->reg.bit_width = 8;
78 dst_cx->reg.bit_offset = 0;
79 dst_cx->reg.access_size = 1;
81 dst_cx->reg.space_id = ACPI_ADR_SPACE_FIXED_HARDWARE;
82 if (cx->entry_method == ACPI_CSTATE_FFH) {
83 /* NATIVE_CSTATE_BEYOND_HALT */
84 dst_cx->reg.bit_offset = 2;
85 dst_cx->reg.bit_width = 1; /* VENDOR_INTEL */
87 dst_cx->reg.access_size = 0;
89 dst_cx->reg.address = cx->address;
91 dst_cx->type = cx->type;
92 dst_cx->latency = cx->latency;
95 set_xen_guest_handle(dst_cx->dp, NULL);
98 pr_debug("No _Cx for ACPI CPU %u\n", _pr->acpi_id);
102 op.u.set_pminfo.power.count = ok;
103 op.u.set_pminfo.power.flags.bm_control = _pr->flags.bm_control;
104 op.u.set_pminfo.power.flags.bm_check = _pr->flags.bm_check;
105 op.u.set_pminfo.power.flags.has_cst = _pr->flags.has_cst;
106 op.u.set_pminfo.power.flags.power_setup_done =
107 _pr->flags.power_setup_done;
109 set_xen_guest_handle(op.u.set_pminfo.power.states, dst_cx_states);
112 ret = HYPERVISOR_platform_op(&op);
115 pr_debug("ACPI CPU%u - C-states uploaded.\n", _pr->acpi_id);
116 for (i = 1; i <= _pr->power.count; i++) {
117 cx = &_pr->power.states[i];
120 pr_debug(" C%d: %s %d uS\n",
121 cx->type, cx->desc, (u32)cx->latency);
123 } else if ((ret != -EINVAL) && (ret != -ENOSYS))
124 /* EINVAL means the ACPI ID is incorrect - meaning the ACPI
125 * table is referencing a non-existing CPU - which can happen
126 * with broken ACPI tables. */
127 pr_err("(CX): Hypervisor error (%d) for ACPI CPU%u\n",
130 kfree(dst_cx_states);
134 static struct xen_processor_px *
135 xen_copy_pss_data(struct acpi_processor *_pr,
136 struct xen_processor_performance *dst_perf)
138 struct xen_processor_px *dst_states = NULL;
141 BUILD_BUG_ON(sizeof(struct xen_processor_px) !=
142 sizeof(struct acpi_processor_px));
144 dst_states = kcalloc(_pr->performance->state_count,
145 sizeof(struct xen_processor_px), GFP_KERNEL);
147 return ERR_PTR(-ENOMEM);
149 dst_perf->state_count = _pr->performance->state_count;
150 for (i = 0; i < _pr->performance->state_count; i++) {
151 /* Fortunatly for us, they are both the same size */
152 memcpy(&(dst_states[i]), &(_pr->performance->states[i]),
153 sizeof(struct acpi_processor_px));
157 static int xen_copy_psd_data(struct acpi_processor *_pr,
158 struct xen_processor_performance *dst)
160 struct acpi_psd_package *pdomain;
162 BUILD_BUG_ON(sizeof(struct xen_psd_package) !=
163 sizeof(struct acpi_psd_package));
165 /* This information is enumerated only if acpi_processor_preregister_performance
168 dst->shared_type = _pr->performance->shared_type;
170 pdomain = &(_pr->performance->domain_info);
172 /* 'acpi_processor_preregister_performance' does not parse if the
173 * num_processors <= 1, but Xen still requires it. Do it manually here.
175 if (pdomain->num_processors <= 1) {
176 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL)
177 dst->shared_type = CPUFREQ_SHARED_TYPE_ALL;
178 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL)
179 dst->shared_type = CPUFREQ_SHARED_TYPE_HW;
180 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY)
181 dst->shared_type = CPUFREQ_SHARED_TYPE_ANY;
184 memcpy(&(dst->domain_info), pdomain, sizeof(struct acpi_psd_package));
187 static int xen_copy_pct_data(struct acpi_pct_register *pct,
188 struct xen_pct_register *dst_pct)
190 /* It would be nice if you could just do 'memcpy(pct, dst_pct') but
191 * sadly the Xen structure did not have the proper padding so the
192 * descriptor field takes two (dst_pct) bytes instead of one (pct).
194 dst_pct->descriptor = pct->descriptor;
195 dst_pct->length = pct->length;
196 dst_pct->space_id = pct->space_id;
197 dst_pct->bit_width = pct->bit_width;
198 dst_pct->bit_offset = pct->bit_offset;
199 dst_pct->reserved = pct->reserved;
200 dst_pct->address = pct->address;
203 static int push_pxx_to_hypervisor(struct acpi_processor *_pr)
206 struct xen_platform_op op = {
207 .cmd = XENPF_set_processor_pminfo,
208 .interface_version = XENPF_INTERFACE_VERSION,
209 .u.set_pminfo.id = _pr->acpi_id,
210 .u.set_pminfo.type = XEN_PM_PX,
212 struct xen_processor_performance *dst_perf;
213 struct xen_processor_px *dst_states = NULL;
215 dst_perf = &op.u.set_pminfo.perf;
217 dst_perf->platform_limit = _pr->performance_platform_limit;
218 dst_perf->flags |= XEN_PX_PPC;
219 xen_copy_pct_data(&(_pr->performance->control_register),
220 &dst_perf->control_register);
221 xen_copy_pct_data(&(_pr->performance->status_register),
222 &dst_perf->status_register);
223 dst_perf->flags |= XEN_PX_PCT;
224 dst_states = xen_copy_pss_data(_pr, dst_perf);
225 if (!IS_ERR_OR_NULL(dst_states)) {
226 set_xen_guest_handle(dst_perf->states, dst_states);
227 dst_perf->flags |= XEN_PX_PSS;
229 if (!xen_copy_psd_data(_pr, dst_perf))
230 dst_perf->flags |= XEN_PX_PSD;
232 if (dst_perf->flags != (XEN_PX_PSD | XEN_PX_PSS | XEN_PX_PCT | XEN_PX_PPC)) {
233 pr_warn("ACPI CPU%u missing some P-state data (%x), skipping\n",
234 _pr->acpi_id, dst_perf->flags);
240 ret = HYPERVISOR_platform_op(&op);
243 struct acpi_processor_performance *perf;
246 perf = _pr->performance;
247 pr_debug("ACPI CPU%u - P-states uploaded.\n", _pr->acpi_id);
248 for (i = 0; i < perf->state_count; i++) {
249 pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n",
250 (i == perf->state ? '*' : ' '), i,
251 (u32) perf->states[i].core_frequency,
252 (u32) perf->states[i].power,
253 (u32) perf->states[i].transition_latency);
255 } else if ((ret != -EINVAL) && (ret != -ENOSYS))
256 /* EINVAL means the ACPI ID is incorrect - meaning the ACPI
257 * table is referencing a non-existing CPU - which can happen
258 * with broken ACPI tables. */
259 pr_warn("(_PXX): Hypervisor error (%d) for ACPI CPU%u\n",
262 if (!IS_ERR_OR_NULL(dst_states))
267 static int upload_pm_data(struct acpi_processor *_pr)
271 mutex_lock(&acpi_ids_mutex);
272 if (__test_and_set_bit(_pr->acpi_id, acpi_ids_done)) {
273 mutex_unlock(&acpi_ids_mutex);
276 if (_pr->flags.power)
277 err = push_cxx_to_hypervisor(_pr);
279 if (_pr->performance && _pr->performance->states)
280 err |= push_pxx_to_hypervisor(_pr);
282 mutex_unlock(&acpi_ids_mutex);
285 static unsigned int __init get_max_acpi_id(void)
287 struct xenpf_pcpuinfo *info;
288 struct xen_platform_op op = {
289 .cmd = XENPF_get_cpuinfo,
290 .interface_version = XENPF_INTERFACE_VERSION,
293 unsigned int i, last_cpu, max_acpi_id = 0;
295 info = &op.u.pcpu_info;
298 ret = HYPERVISOR_platform_op(&op);
302 /* The max_present is the same irregardless of the xen_cpuid */
303 last_cpu = op.u.pcpu_info.max_present;
304 for (i = 0; i <= last_cpu; i++) {
306 ret = HYPERVISOR_platform_op(&op);
309 max_acpi_id = max(info->acpi_id, max_acpi_id);
311 max_acpi_id *= 2; /* Slack for CPU hotplug support. */
312 pr_debug("Max ACPI ID: %u\n", max_acpi_id);
316 * The read_acpi_id and check_acpi_ids are there to support the Xen
317 * oddity of virtual CPUs != physical CPUs in the initial domain.
318 * The user can supply 'xen_max_vcpus=X' on the Xen hypervisor line
319 * which will band the amount of CPUs the initial domain can see.
320 * In general that is OK, except it plays havoc with any of the
321 * for_each_[present|online]_cpu macros which are banded to the virtual
325 read_acpi_id(acpi_handle handle, u32 lvl, void *context, void **rv)
329 acpi_object_type acpi_type;
330 unsigned long long tmp;
331 union acpi_object object = { 0 };
332 struct acpi_buffer buffer = { sizeof(union acpi_object), &object };
333 acpi_io_address pblk = 0;
335 status = acpi_get_type(handle, &acpi_type);
336 if (ACPI_FAILURE(status))
340 case ACPI_TYPE_PROCESSOR:
341 status = acpi_evaluate_object(handle, NULL, NULL, &buffer);
342 if (ACPI_FAILURE(status))
344 acpi_id = object.processor.proc_id;
345 pblk = object.processor.pblk_address;
347 case ACPI_TYPE_DEVICE:
348 status = acpi_evaluate_integer(handle, "_UID", NULL, &tmp);
349 if (ACPI_FAILURE(status))
356 if (invalid_phys_cpuid(acpi_get_phys_id(handle,
357 acpi_type == ACPI_TYPE_DEVICE,
359 pr_debug("CPU with ACPI ID %u is unavailable\n", acpi_id);
362 /* There are more ACPI Processor objects than in x2APIC or MADT.
363 * This can happen with incorrect ACPI SSDT declerations. */
364 if (acpi_id >= nr_acpi_bits) {
365 pr_debug("max acpi id %u, trying to set %u\n",
366 nr_acpi_bits - 1, acpi_id);
369 /* OK, There is a ACPI Processor object */
370 __set_bit(acpi_id, acpi_id_present);
372 pr_debug("ACPI CPU%u w/ PBLK:0x%lx\n", acpi_id, (unsigned long)pblk);
374 /* It has P-state dependencies */
375 if (!acpi_processor_get_psd(handle, &acpi_psd[acpi_id])) {
376 pr_debug("ACPI CPU%u w/ PST:coord_type = %llu domain = %llu\n",
377 acpi_id, acpi_psd[acpi_id].coord_type,
378 acpi_psd[acpi_id].domain);
381 status = acpi_evaluate_object(handle, "_CST", NULL, &buffer);
382 if (ACPI_FAILURE(status)) {
386 /* .. and it has a C-state */
387 __set_bit(acpi_id, acpi_id_cst_present);
391 static int check_acpi_ids(struct acpi_processor *pr_backup)
397 if (acpi_id_present && acpi_id_cst_present)
398 /* OK, done this once .. skip to uploading */
401 /* All online CPUs have been processed at this stage. Now verify
402 * whether in fact "online CPUs" == physical CPUs.
404 acpi_id_present = bitmap_zalloc(nr_acpi_bits, GFP_KERNEL);
405 if (!acpi_id_present)
408 acpi_id_cst_present = bitmap_zalloc(nr_acpi_bits, GFP_KERNEL);
409 if (!acpi_id_cst_present) {
410 bitmap_free(acpi_id_present);
414 acpi_psd = kcalloc(nr_acpi_bits, sizeof(struct acpi_psd_package),
417 bitmap_free(acpi_id_present);
418 bitmap_free(acpi_id_cst_present);
422 acpi_walk_namespace(ACPI_TYPE_PROCESSOR, ACPI_ROOT_OBJECT,
424 read_acpi_id, NULL, NULL, NULL);
425 acpi_get_devices(ACPI_PROCESSOR_DEVICE_HID, read_acpi_id, NULL, NULL);
428 if (!bitmap_equal(acpi_id_present, acpi_ids_done, nr_acpi_bits)) {
430 for_each_set_bit(i, acpi_id_present, nr_acpi_bits) {
431 pr_backup->acpi_id = i;
432 /* Mask out C-states if there are no _CST or PBLK */
433 pr_backup->flags.power = test_bit(i, acpi_id_cst_present);
434 /* num_entries is non-zero if we evaluated _PSD */
435 if (acpi_psd[i].num_entries) {
436 memcpy(&pr_backup->performance->domain_info,
438 sizeof(struct acpi_psd_package));
440 (void)upload_pm_data(pr_backup);
447 /* acpi_perf_data is a pointer to percpu data. */
448 static struct acpi_processor_performance __percpu *acpi_perf_data;
450 static void free_acpi_perf_data(void)
454 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
455 for_each_possible_cpu(i)
456 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
458 free_percpu(acpi_perf_data);
461 static int xen_upload_processor_pm_data(void)
463 struct acpi_processor *pr_backup = NULL;
467 pr_info("Uploading Xen processor PM info\n");
469 for_each_possible_cpu(i) {
470 struct acpi_processor *_pr;
471 _pr = per_cpu(processors, i /* APIC ID */);
476 pr_backup = kzalloc(sizeof(struct acpi_processor), GFP_KERNEL);
478 memcpy(pr_backup, _pr, sizeof(struct acpi_processor));
480 (void)upload_pm_data(_pr);
483 rc = check_acpi_ids(pr_backup);
489 static void xen_acpi_processor_resume_worker(struct work_struct *dummy)
493 bitmap_zero(acpi_ids_done, nr_acpi_bits);
495 rc = xen_upload_processor_pm_data();
497 pr_info("ACPI data upload failed, error = %d\n", rc);
500 static void xen_acpi_processor_resume(void)
502 static DECLARE_WORK(wq, xen_acpi_processor_resume_worker);
505 * xen_upload_processor_pm_data() calls non-atomic code.
506 * However, the context for xen_acpi_processor_resume is syscore
507 * with only the boot CPU online and in an atomic context.
509 * So defer the upload for some point safer.
514 static struct syscore_ops xap_syscore_ops = {
515 .resume = xen_acpi_processor_resume,
518 static int __init xen_acpi_processor_init(void)
523 if (!xen_initial_domain())
526 nr_acpi_bits = get_max_acpi_id() + 1;
527 acpi_ids_done = bitmap_zalloc(nr_acpi_bits, GFP_KERNEL);
531 acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
532 if (!acpi_perf_data) {
533 pr_debug("Memory allocation error for acpi_perf_data\n");
534 bitmap_free(acpi_ids_done);
537 for_each_possible_cpu(i) {
538 if (!zalloc_cpumask_var_node(
539 &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
540 GFP_KERNEL, cpu_to_node(i))) {
546 /* Do initialization in ACPI core. It is OK to fail here. */
547 (void)acpi_processor_preregister_performance(acpi_perf_data);
549 for_each_possible_cpu(i) {
550 struct acpi_processor *pr;
551 struct acpi_processor_performance *perf;
553 pr = per_cpu(processors, i);
554 perf = per_cpu_ptr(acpi_perf_data, i);
558 pr->performance = perf;
559 rc = acpi_processor_get_performance_info(pr);
564 rc = xen_upload_processor_pm_data();
568 register_syscore_ops(&xap_syscore_ops);
572 for_each_possible_cpu(i)
573 acpi_processor_unregister_performance(i);
576 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
577 free_acpi_perf_data();
578 bitmap_free(acpi_ids_done);
581 static void __exit xen_acpi_processor_exit(void)
585 unregister_syscore_ops(&xap_syscore_ops);
586 bitmap_free(acpi_ids_done);
587 bitmap_free(acpi_id_present);
588 bitmap_free(acpi_id_cst_present);
590 for_each_possible_cpu(i)
591 acpi_processor_unregister_performance(i);
593 free_acpi_perf_data();
597 MODULE_DESCRIPTION("Xen ACPI Processor P-states (and Cx) driver which uploads PM data to Xen hypervisor");
598 MODULE_LICENSE("GPL");
600 /* We want to be loaded before the CPU freq scaling drivers are loaded.
601 * They are loaded in late_initcall. */
602 device_initcall(xen_acpi_processor_init);
603 module_exit(xen_acpi_processor_exit);