1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright 2012 by Oracle Inc
4 * Author: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
5 *
6 * This code borrows ideas from
7 * https://lore.kernel.org/lkml/1322673664-14642-6-git-send-email-konrad.wilk@oracle.com
8 * so many thanks go to Kevin Tian <kevin.tian@intel.com>
9 * and Yu Ke <ke.yu@intel.com>.
10 */
11
12#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13
14#include <linux/cpumask.h>
15#include <linux/cpufreq.h>
16#include <linux/freezer.h>
17#include <linux/kernel.h>
18#include <linux/kthread.h>
19#include <linux/init.h>
20#include <linux/module.h>
21#include <linux/types.h>
22#include <linux/syscore_ops.h>
23#include <linux/acpi.h>
24#include <acpi/processor.h>
25#include <xen/xen.h>
26#include <xen/interface/platform.h>
27#include <asm/xen/hypercall.h>
28
29static int no_hypercall;
30MODULE_PARM_DESC(off, "Inhibit the hypercall.");
31module_param_named(off, no_hypercall, int, 0400);
32
33/*
34 * Note: Do not convert the acpi_id* below to cpumask_var_t or use cpumask_bit
35 * - as those shrink to nr_cpu_bits (which is dependent on possible_cpu), which
36 * can be less than what we want to put in. Instead use the 'nr_acpi_bits'
37 * which is dynamically computed based on the MADT or x2APIC table.
38 */
39static unsigned int nr_acpi_bits;
40/* Mutex to protect the acpi_ids_done - for CPU hotplug use. */
41static DEFINE_MUTEX(acpi_ids_mutex);
42/* Which ACPI ID we have processed from 'struct acpi_processor'. */
43static unsigned long *acpi_ids_done;
44/* Which ACPI ID exist in the SSDT/DSDT processor definitions. */
45static unsigned long *acpi_id_present;
46/* And if there is an _CST definition (or a PBLK) for the ACPI IDs */
47static unsigned long *acpi_id_cst_present;
48/* Which ACPI P-State dependencies for a enumerated processor */
49static struct acpi_psd_package *acpi_psd;
50
51static int push_cxx_to_hypervisor(struct acpi_processor *_pr)
52{
53 struct xen_platform_op op = {
54 .cmd = XENPF_set_processor_pminfo,
55 .interface_version = XENPF_INTERFACE_VERSION,
56 .u.set_pminfo.id = _pr->acpi_id,
57 .u.set_pminfo.type = XEN_PM_CX,
58 };
59 struct xen_processor_cx *dst_cx, *dst_cx_states = NULL;
60 struct acpi_processor_cx *cx;
61 unsigned int i, ok;
62 int ret = 0;
63
64 dst_cx_states = kcalloc(n: _pr->power.count,
65 size: sizeof(struct xen_processor_cx), GFP_KERNEL);
66 if (!dst_cx_states)
67 return -ENOMEM;
68
69 for (ok = 0, i = 1; i <= _pr->power.count; i++) {
70 cx = &_pr->power.states[i];
71 if (!cx->valid)
72 continue;
73
74 dst_cx = &(dst_cx_states[ok++]);
75
76 dst_cx->reg.space_id = ACPI_ADR_SPACE_SYSTEM_IO;
77 if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) {
78 dst_cx->reg.bit_width = 8;
79 dst_cx->reg.bit_offset = 0;
80 dst_cx->reg.access_size = 1;
81 } else {
82 dst_cx->reg.space_id = ACPI_ADR_SPACE_FIXED_HARDWARE;
83 if (cx->entry_method == ACPI_CSTATE_FFH) {
84 /* NATIVE_CSTATE_BEYOND_HALT */
85 dst_cx->reg.bit_offset = 2;
86 dst_cx->reg.bit_width = 1; /* VENDOR_INTEL */
87 }
88 dst_cx->reg.access_size = 0;
89 }
90 dst_cx->reg.address = cx->address;
91
92 dst_cx->type = cx->type;
93 dst_cx->latency = cx->latency;
94
95 dst_cx->dpcnt = 0;
96 set_xen_guest_handle(dst_cx->dp, NULL);
97 }
98 if (!ok) {
99 pr_debug("No _Cx for ACPI CPU %u\n", _pr->acpi_id);
100 kfree(objp: dst_cx_states);
101 return -EINVAL;
102 }
103 op.u.set_pminfo.power.count = ok;
104 op.u.set_pminfo.power.flags.bm_control = _pr->flags.bm_control;
105 op.u.set_pminfo.power.flags.bm_check = _pr->flags.bm_check;
106 op.u.set_pminfo.power.flags.has_cst = _pr->flags.has_cst;
107 op.u.set_pminfo.power.flags.power_setup_done =
108 _pr->flags.power_setup_done;
109
110 set_xen_guest_handle(op.u.set_pminfo.power.states, dst_cx_states);
111
112 if (!no_hypercall)
113 ret = HYPERVISOR_platform_op(op: &op);
114
115 if (!ret) {
116 pr_debug("ACPI CPU%u - C-states uploaded.\n", _pr->acpi_id);
117 for (i = 1; i <= _pr->power.count; i++) {
118 cx = &_pr->power.states[i];
119 if (!cx->valid)
120 continue;
121 pr_debug(" C%d: %s %d uS\n",
122 cx->type, cx->desc, (u32)cx->latency);
123 }
124 } else if ((ret != -EINVAL) && (ret != -ENOSYS))
125 /* EINVAL means the ACPI ID is incorrect - meaning the ACPI
126 * table is referencing a non-existing CPU - which can happen
127 * with broken ACPI tables. */
128 pr_err("(CX): Hypervisor error (%d) for ACPI CPU%u\n",
129 ret, _pr->acpi_id);
130
131 kfree(objp: dst_cx_states);
132
133 return ret;
134}
135static struct xen_processor_px *
136xen_copy_pss_data(struct acpi_processor *_pr,
137 struct xen_processor_performance *dst_perf)
138{
139 struct xen_processor_px *dst_states = NULL;
140 unsigned int i;
141
142 BUILD_BUG_ON(sizeof(struct xen_processor_px) !=
143 sizeof(struct acpi_processor_px));
144
145 dst_states = kcalloc(n: _pr->performance->state_count,
146 size: sizeof(struct xen_processor_px), GFP_KERNEL);
147 if (!dst_states)
148 return ERR_PTR(error: -ENOMEM);
149
150 dst_perf->state_count = _pr->performance->state_count;
151 for (i = 0; i < _pr->performance->state_count; i++) {
152 /* Fortunatly for us, they are both the same size */
153 memcpy(&(dst_states[i]), &(_pr->performance->states[i]),
154 sizeof(struct acpi_processor_px));
155 }
156 return dst_states;
157}
158static int xen_copy_psd_data(struct acpi_processor *_pr,
159 struct xen_processor_performance *dst)
160{
161 struct acpi_psd_package *pdomain;
162
163 BUILD_BUG_ON(sizeof(struct xen_psd_package) !=
164 sizeof(struct acpi_psd_package));
165
166 /* This information is enumerated only if acpi_processor_preregister_performance
167 * has been called.
168 */
169 dst->shared_type = _pr->performance->shared_type;
170
171 pdomain = &(_pr->performance->domain_info);
172
173 /* 'acpi_processor_preregister_performance' does not parse if the
174 * num_processors <= 1, but Xen still requires it. Do it manually here.
175 */
176 if (pdomain->num_processors <= 1) {
177 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL)
178 dst->shared_type = CPUFREQ_SHARED_TYPE_ALL;
179 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL)
180 dst->shared_type = CPUFREQ_SHARED_TYPE_HW;
181 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY)
182 dst->shared_type = CPUFREQ_SHARED_TYPE_ANY;
183
184 }
185 memcpy(&(dst->domain_info), pdomain, sizeof(struct acpi_psd_package));
186 return 0;
187}
188static int xen_copy_pct_data(struct acpi_pct_register *pct,
189 struct xen_pct_register *dst_pct)
190{
191 /* It would be nice if you could just do 'memcpy(pct, dst_pct') but
192 * sadly the Xen structure did not have the proper padding so the
193 * descriptor field takes two (dst_pct) bytes instead of one (pct).
194 */
195 dst_pct->descriptor = pct->descriptor;
196 dst_pct->length = pct->length;
197 dst_pct->space_id = pct->space_id;
198 dst_pct->bit_width = pct->bit_width;
199 dst_pct->bit_offset = pct->bit_offset;
200 dst_pct->reserved = pct->reserved;
201 dst_pct->address = pct->address;
202 return 0;
203}
204static int push_pxx_to_hypervisor(struct acpi_processor *_pr)
205{
206 int ret = 0;
207 struct xen_platform_op op = {
208 .cmd = XENPF_set_processor_pminfo,
209 .interface_version = XENPF_INTERFACE_VERSION,
210 .u.set_pminfo.id = _pr->acpi_id,
211 .u.set_pminfo.type = XEN_PM_PX,
212 };
213 struct xen_processor_performance *dst_perf;
214 struct xen_processor_px *dst_states = NULL;
215
216 dst_perf = &op.u.set_pminfo.perf;
217
218 dst_perf->platform_limit = _pr->performance_platform_limit;
219 dst_perf->flags |= XEN_PX_PPC;
220 xen_copy_pct_data(pct: &(_pr->performance->control_register),
221 dst_pct: &dst_perf->control_register);
222 xen_copy_pct_data(pct: &(_pr->performance->status_register),
223 dst_pct: &dst_perf->status_register);
224 dst_perf->flags |= XEN_PX_PCT;
225 dst_states = xen_copy_pss_data(_pr, dst_perf);
226 if (!IS_ERR_OR_NULL(ptr: dst_states)) {
227 set_xen_guest_handle(dst_perf->states, dst_states);
228 dst_perf->flags |= XEN_PX_PSS;
229 }
230 if (!xen_copy_psd_data(_pr, dst: dst_perf))
231 dst_perf->flags |= XEN_PX_PSD;
232
233 if (dst_perf->flags != (XEN_PX_PSD | XEN_PX_PSS | XEN_PX_PCT | XEN_PX_PPC)) {
234 pr_warn("ACPI CPU%u missing some P-state data (%x), skipping\n",
235 _pr->acpi_id, dst_perf->flags);
236 ret = -ENODEV;
237 goto err_free;
238 }
239
240 if (!no_hypercall)
241 ret = HYPERVISOR_platform_op(op: &op);
242
243 if (!ret) {
244 struct acpi_processor_performance *perf;
245 unsigned int i;
246
247 perf = _pr->performance;
248 pr_debug("ACPI CPU%u - P-states uploaded.\n", _pr->acpi_id);
249 for (i = 0; i < perf->state_count; i++) {
250 pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n",
251 (i == perf->state ? '*' : ' '), i,
252 (u32) perf->states[i].core_frequency,
253 (u32) perf->states[i].power,
254 (u32) perf->states[i].transition_latency);
255 }
256 } else if ((ret != -EINVAL) && (ret != -ENOSYS))
257 /* EINVAL means the ACPI ID is incorrect - meaning the ACPI
258 * table is referencing a non-existing CPU - which can happen
259 * with broken ACPI tables. */
260 pr_warn("(_PXX): Hypervisor error (%d) for ACPI CPU%u\n",
261 ret, _pr->acpi_id);
262err_free:
263 if (!IS_ERR_OR_NULL(ptr: dst_states))
264 kfree(objp: dst_states);
265
266 return ret;
267}
268static int upload_pm_data(struct acpi_processor *_pr)
269{
270 int err = 0;
271
272 mutex_lock(&acpi_ids_mutex);
273 if (__test_and_set_bit(_pr->acpi_id, acpi_ids_done)) {
274 mutex_unlock(lock: &acpi_ids_mutex);
275 return -EBUSY;
276 }
277 if (_pr->flags.power)
278 err = push_cxx_to_hypervisor(_pr);
279
280 if (_pr->performance && _pr->performance->states)
281 err |= push_pxx_to_hypervisor(_pr);
282
283 mutex_unlock(lock: &acpi_ids_mutex);
284 return err;
285}
286static unsigned int __init get_max_acpi_id(void)
287{
288 struct xenpf_pcpuinfo *info;
289 struct xen_platform_op op = {
290 .cmd = XENPF_get_cpuinfo,
291 .interface_version = XENPF_INTERFACE_VERSION,
292 };
293 int ret = 0;
294 unsigned int i, last_cpu, max_acpi_id = 0;
295
296 info = &op.u.pcpu_info;
297 info->xen_cpuid = 0;
298
299 ret = HYPERVISOR_platform_op(op: &op);
300 if (ret)
301 return NR_CPUS;
302
303 /* The max_present is the same irregardless of the xen_cpuid */
304 last_cpu = op.u.pcpu_info.max_present;
305 for (i = 0; i <= last_cpu; i++) {
306 info->xen_cpuid = i;
307 ret = HYPERVISOR_platform_op(op: &op);
308 if (ret)
309 continue;
310 max_acpi_id = max(info->acpi_id, max_acpi_id);
311 }
312 max_acpi_id *= 2; /* Slack for CPU hotplug support. */
313 pr_debug("Max ACPI ID: %u\n", max_acpi_id);
314 return max_acpi_id;
315}
316/*
317 * The read_acpi_id and check_acpi_ids are there to support the Xen
318 * oddity of virtual CPUs != physical CPUs in the initial domain.
319 * The user can supply 'xen_max_vcpus=X' on the Xen hypervisor line
320 * which will band the amount of CPUs the initial domain can see.
321 * In general that is OK, except it plays havoc with any of the
322 * for_each_[present|online]_cpu macros which are banded to the virtual
323 * CPU amount.
324 */
325static acpi_status
326read_acpi_id(acpi_handle handle, u32 lvl, void *context, void **rv)
327{
328 u32 acpi_id;
329 acpi_status status;
330 acpi_object_type acpi_type;
331 unsigned long long tmp;
332 union acpi_object object = { 0 };
333 struct acpi_buffer buffer = { sizeof(union acpi_object), &object };
334 acpi_io_address pblk = 0;
335
336 status = acpi_get_type(object: handle, out_type: &acpi_type);
337 if (ACPI_FAILURE(status))
338 return AE_OK;
339
340 switch (acpi_type) {
341 case ACPI_TYPE_PROCESSOR:
342 status = acpi_evaluate_object(object: handle, NULL, NULL, return_object_buffer: &buffer);
343 if (ACPI_FAILURE(status))
344 return AE_OK;
345 acpi_id = object.processor.proc_id;
346 pblk = object.processor.pblk_address;
347 break;
348 case ACPI_TYPE_DEVICE:
349 status = acpi_evaluate_integer(handle, pathname: "_UID", NULL, data: &tmp);
350 if (ACPI_FAILURE(status))
351 return AE_OK;
352 acpi_id = tmp;
353 break;
354 default:
355 return AE_OK;
356 }
357 if (invalid_phys_cpuid(phys_id: acpi_get_phys_id(handle,
358 type: acpi_type == ACPI_TYPE_DEVICE,
359 acpi_id))) {
360 pr_debug("CPU with ACPI ID %u is unavailable\n", acpi_id);
361 return AE_OK;
362 }
363 /* There are more ACPI Processor objects than in x2APIC or MADT.
364 * This can happen with incorrect ACPI SSDT declerations. */
365 if (acpi_id >= nr_acpi_bits) {
366 pr_debug("max acpi id %u, trying to set %u\n",
367 nr_acpi_bits - 1, acpi_id);
368 return AE_OK;
369 }
370 /* OK, There is a ACPI Processor object */
371 __set_bit(acpi_id, acpi_id_present);
372
373 pr_debug("ACPI CPU%u w/ PBLK:0x%lx\n", acpi_id, (unsigned long)pblk);
374
375 /* It has P-state dependencies */
376 if (!acpi_processor_get_psd(handle, pdomain: &acpi_psd[acpi_id])) {
377 pr_debug("ACPI CPU%u w/ PST:coord_type = %llu domain = %llu\n",
378 acpi_id, acpi_psd[acpi_id].coord_type,
379 acpi_psd[acpi_id].domain);
380 }
381
382 status = acpi_evaluate_object(object: handle, pathname: "_CST", NULL, return_object_buffer: &buffer);
383 if (ACPI_FAILURE(status)) {
384 if (!pblk)
385 return AE_OK;
386 }
387 /* .. and it has a C-state */
388 __set_bit(acpi_id, acpi_id_cst_present);
389
390 return AE_OK;
391}
392static int check_acpi_ids(struct acpi_processor *pr_backup)
393{
394
395 if (!pr_backup)
396 return -ENODEV;
397
398 if (acpi_id_present && acpi_id_cst_present)
399 /* OK, done this once .. skip to uploading */
400 goto upload;
401
402 /* All online CPUs have been processed at this stage. Now verify
403 * whether in fact "online CPUs" == physical CPUs.
404 */
405 acpi_id_present = bitmap_zalloc(nbits: nr_acpi_bits, GFP_KERNEL);
406 if (!acpi_id_present)
407 return -ENOMEM;
408
409 acpi_id_cst_present = bitmap_zalloc(nbits: nr_acpi_bits, GFP_KERNEL);
410 if (!acpi_id_cst_present) {
411 bitmap_free(bitmap: acpi_id_present);
412 return -ENOMEM;
413 }
414
415 acpi_psd = kcalloc(n: nr_acpi_bits, size: sizeof(struct acpi_psd_package),
416 GFP_KERNEL);
417 if (!acpi_psd) {
418 bitmap_free(bitmap: acpi_id_present);
419 bitmap_free(bitmap: acpi_id_cst_present);
420 return -ENOMEM;
421 }
422
423 acpi_walk_namespace(ACPI_TYPE_PROCESSOR, ACPI_ROOT_OBJECT,
424 ACPI_UINT32_MAX,
425 descending_callback: read_acpi_id, NULL, NULL, NULL);
426 acpi_get_devices(ACPI_PROCESSOR_DEVICE_HID, user_function: read_acpi_id, NULL, NULL);
427
428upload:
429 if (!bitmap_equal(src1: acpi_id_present, src2: acpi_ids_done, nbits: nr_acpi_bits)) {
430 unsigned int i;
431 for_each_set_bit(i, acpi_id_present, nr_acpi_bits) {
432 pr_backup->acpi_id = i;
433 /* Mask out C-states if there are no _CST or PBLK */
434 pr_backup->flags.power = test_bit(i, acpi_id_cst_present);
435 /* num_entries is non-zero if we evaluated _PSD */
436 if (acpi_psd[i].num_entries) {
437 memcpy(&pr_backup->performance->domain_info,
438 &acpi_psd[i],
439 sizeof(struct acpi_psd_package));
440 }
441 (void)upload_pm_data(pr: pr_backup);
442 }
443 }
444
445 return 0;
446}
447
448/* acpi_perf_data is a pointer to percpu data. */
449static struct acpi_processor_performance __percpu *acpi_perf_data;
450
451static void free_acpi_perf_data(void)
452{
453 int i;
454
455 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
456 for_each_possible_cpu(i)
457 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
458 ->shared_cpu_map);
459 free_percpu(pdata: acpi_perf_data);
460}
461
462static int xen_upload_processor_pm_data(void)
463{
464 struct acpi_processor *pr_backup = NULL;
465 int i;
466 int rc = 0;
467
468 pr_info("Uploading Xen processor PM info\n");
469
470 for_each_possible_cpu(i) {
471 struct acpi_processor *_pr;
472 _pr = per_cpu(processors, i /* APIC ID */);
473 if (!_pr)
474 continue;
475
476 if (!pr_backup)
477 pr_backup = kmemdup(p: _pr, size: sizeof(*_pr), GFP_KERNEL);
478 (void)upload_pm_data(_pr);
479 }
480
481 rc = check_acpi_ids(pr_backup);
482 kfree(objp: pr_backup);
483
484 return rc;
485}
486
487static void xen_acpi_processor_resume_worker(struct work_struct *dummy)
488{
489 int rc;
490
491 bitmap_zero(dst: acpi_ids_done, nbits: nr_acpi_bits);
492
493 rc = xen_upload_processor_pm_data();
494 if (rc != 0)
495 pr_info("ACPI data upload failed, error = %d\n", rc);
496}
497
498static void xen_acpi_processor_resume(void)
499{
500 static DECLARE_WORK(wq, xen_acpi_processor_resume_worker);
501
502 /*
503 * xen_upload_processor_pm_data() calls non-atomic code.
504 * However, the context for xen_acpi_processor_resume is syscore
505 * with only the boot CPU online and in an atomic context.
506 *
507 * So defer the upload for some point safer.
508 */
509 schedule_work(work: &wq);
510}
511
512static struct syscore_ops xap_syscore_ops = {
513 .resume = xen_acpi_processor_resume,
514};
515
516static int __init xen_acpi_processor_init(void)
517{
518 int i;
519 int rc;
520
521 if (!xen_initial_domain())
522 return -ENODEV;
523
524 nr_acpi_bits = get_max_acpi_id() + 1;
525 acpi_ids_done = bitmap_zalloc(nbits: nr_acpi_bits, GFP_KERNEL);
526 if (!acpi_ids_done)
527 return -ENOMEM;
528
529 acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
530 if (!acpi_perf_data) {
531 pr_debug("Memory allocation error for acpi_perf_data\n");
532 bitmap_free(bitmap: acpi_ids_done);
533 return -ENOMEM;
534 }
535 for_each_possible_cpu(i) {
536 if (!zalloc_cpumask_var_node(
537 mask: &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
538 GFP_KERNEL, cpu_to_node(cpu: i))) {
539 rc = -ENOMEM;
540 goto err_out;
541 }
542 }
543
544 /* Do initialization in ACPI core. It is OK to fail here. */
545 (void)acpi_processor_preregister_performance(performance: acpi_perf_data);
546
547 for_each_possible_cpu(i) {
548 struct acpi_processor *pr;
549 struct acpi_processor_performance *perf;
550
551 pr = per_cpu(processors, i);
552 perf = per_cpu_ptr(acpi_perf_data, i);
553 if (!pr)
554 continue;
555
556 pr->performance = perf;
557 rc = acpi_processor_get_performance_info(pr);
558 if (rc)
559 goto err_out;
560 }
561
562 rc = xen_upload_processor_pm_data();
563 if (rc)
564 goto err_unregister;
565
566 register_syscore_ops(ops: &xap_syscore_ops);
567
568 return 0;
569err_unregister:
570 for_each_possible_cpu(i)
571 acpi_processor_unregister_performance(cpu: i);
572
573err_out:
574 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
575 free_acpi_perf_data();
576 bitmap_free(bitmap: acpi_ids_done);
577 return rc;
578}
579static void __exit xen_acpi_processor_exit(void)
580{
581 int i;
582
583 unregister_syscore_ops(ops: &xap_syscore_ops);
584 bitmap_free(bitmap: acpi_ids_done);
585 bitmap_free(bitmap: acpi_id_present);
586 bitmap_free(bitmap: acpi_id_cst_present);
587 kfree(objp: acpi_psd);
588 for_each_possible_cpu(i)
589 acpi_processor_unregister_performance(cpu: i);
590
591 free_acpi_perf_data();
592}
593
594MODULE_AUTHOR("Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>");
595MODULE_DESCRIPTION("Xen ACPI Processor P-states (and Cx) driver which uploads PM data to Xen hypervisor");
596MODULE_LICENSE("GPL");
597
598/* We want to be loaded before the CPU freq scaling drivers are loaded.
599 * They are loaded in late_initcall. */
600device_initcall(xen_acpi_processor_init);
601module_exit(xen_acpi_processor_exit);
602

source code of linux/drivers/xen/xen-acpi-processor.c