Loading...
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 https://lkml.org/lkml/2011/11/30/249
7 * so many thanks go to Kevin Tian <kevin.tian@intel.com>
8 * and Yu Ke <ke.yu@intel.com>.
9 */
10
11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12
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>
24#include <xen/xen.h>
25#include <xen/interface/platform.h>
26#include <asm/xen/hypercall.h>
27
28static int no_hypercall;
29MODULE_PARM_DESC(off, "Inhibit the hypercall.");
30module_param_named(off, no_hypercall, int, 0400);
31
32/*
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.
37 */
38static unsigned int nr_acpi_bits;
39/* Mutex to protect the acpi_ids_done - for CPU hotplug use. */
40static DEFINE_MUTEX(acpi_ids_mutex);
41/* Which ACPI ID we have processed from 'struct acpi_processor'. */
42static unsigned long *acpi_ids_done;
43/* Which ACPI ID exist in the SSDT/DSDT processor definitions. */
44static unsigned long *acpi_id_present;
45/* And if there is an _CST definition (or a PBLK) for the ACPI IDs */
46static unsigned long *acpi_id_cst_present;
47/* Which ACPI P-State dependencies for a enumerated processor */
48static struct acpi_psd_package *acpi_psd;
49
50static int push_cxx_to_hypervisor(struct acpi_processor *_pr)
51{
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,
57 };
58 struct xen_processor_cx *dst_cx, *dst_cx_states = NULL;
59 struct acpi_processor_cx *cx;
60 unsigned int i, ok;
61 int ret = 0;
62
63 dst_cx_states = kcalloc(_pr->power.count,
64 sizeof(struct xen_processor_cx), GFP_KERNEL);
65 if (!dst_cx_states)
66 return -ENOMEM;
67
68 for (ok = 0, i = 1; i <= _pr->power.count; i++) {
69 cx = &_pr->power.states[i];
70 if (!cx->valid)
71 continue;
72
73 dst_cx = &(dst_cx_states[ok++]);
74
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;
80 } else {
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 */
86 }
87 dst_cx->reg.access_size = 0;
88 }
89 dst_cx->reg.address = cx->address;
90
91 dst_cx->type = cx->type;
92 dst_cx->latency = cx->latency;
93
94 dst_cx->dpcnt = 0;
95 set_xen_guest_handle(dst_cx->dp, NULL);
96 }
97 if (!ok) {
98 pr_debug("No _Cx for ACPI CPU %u\n", _pr->acpi_id);
99 kfree(dst_cx_states);
100 return -EINVAL;
101 }
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;
108
109 set_xen_guest_handle(op.u.set_pminfo.power.states, dst_cx_states);
110
111 if (!no_hypercall)
112 ret = HYPERVISOR_platform_op(&op);
113
114 if (!ret) {
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];
118 if (!cx->valid)
119 continue;
120 pr_debug(" C%d: %s %d uS\n",
121 cx->type, cx->desc, (u32)cx->latency);
122 }
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",
128 ret, _pr->acpi_id);
129
130 kfree(dst_cx_states);
131
132 return ret;
133}
134static struct xen_processor_px *
135xen_copy_pss_data(struct acpi_processor *_pr,
136 struct xen_processor_performance *dst_perf)
137{
138 struct xen_processor_px *dst_states = NULL;
139 unsigned int i;
140
141 BUILD_BUG_ON(sizeof(struct xen_processor_px) !=
142 sizeof(struct acpi_processor_px));
143
144 dst_states = kcalloc(_pr->performance->state_count,
145 sizeof(struct xen_processor_px), GFP_KERNEL);
146 if (!dst_states)
147 return ERR_PTR(-ENOMEM);
148
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));
154 }
155 return dst_states;
156}
157static int xen_copy_psd_data(struct acpi_processor *_pr,
158 struct xen_processor_performance *dst)
159{
160 struct acpi_psd_package *pdomain;
161
162 BUILD_BUG_ON(sizeof(struct xen_psd_package) !=
163 sizeof(struct acpi_psd_package));
164
165 /* This information is enumerated only if acpi_processor_preregister_performance
166 * has been called.
167 */
168 dst->shared_type = _pr->performance->shared_type;
169
170 pdomain = &(_pr->performance->domain_info);
171
172 /* 'acpi_processor_preregister_performance' does not parse if the
173 * num_processors <= 1, but Xen still requires it. Do it manually here.
174 */
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;
182
183 }
184 memcpy(&(dst->domain_info), pdomain, sizeof(struct acpi_psd_package));
185 return 0;
186}
187static int xen_copy_pct_data(struct acpi_pct_register *pct,
188 struct xen_pct_register *dst_pct)
189{
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).
193 */
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;
201 return 0;
202}
203static int push_pxx_to_hypervisor(struct acpi_processor *_pr)
204{
205 int ret = 0;
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,
211 };
212 struct xen_processor_performance *dst_perf;
213 struct xen_processor_px *dst_states = NULL;
214
215 dst_perf = &op.u.set_pminfo.perf;
216
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;
228 }
229 if (!xen_copy_psd_data(_pr, dst_perf))
230 dst_perf->flags |= XEN_PX_PSD;
231
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);
235 ret = -ENODEV;
236 goto err_free;
237 }
238
239 if (!no_hypercall)
240 ret = HYPERVISOR_platform_op(&op);
241
242 if (!ret) {
243 struct acpi_processor_performance *perf;
244 unsigned int i;
245
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);
254 }
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",
260 ret, _pr->acpi_id);
261err_free:
262 if (!IS_ERR_OR_NULL(dst_states))
263 kfree(dst_states);
264
265 return ret;
266}
267static int upload_pm_data(struct acpi_processor *_pr)
268{
269 int err = 0;
270
271 mutex_lock(&acpi_ids_mutex);
272 if (__test_and_set_bit(_pr->acpi_id, acpi_ids_done)) {
273 mutex_unlock(&acpi_ids_mutex);
274 return -EBUSY;
275 }
276 if (_pr->flags.power)
277 err = push_cxx_to_hypervisor(_pr);
278
279 if (_pr->performance && _pr->performance->states)
280 err |= push_pxx_to_hypervisor(_pr);
281
282 mutex_unlock(&acpi_ids_mutex);
283 return err;
284}
285static unsigned int __init get_max_acpi_id(void)
286{
287 struct xenpf_pcpuinfo *info;
288 struct xen_platform_op op = {
289 .cmd = XENPF_get_cpuinfo,
290 .interface_version = XENPF_INTERFACE_VERSION,
291 };
292 int ret = 0;
293 unsigned int i, last_cpu, max_acpi_id = 0;
294
295 info = &op.u.pcpu_info;
296 info->xen_cpuid = 0;
297
298 ret = HYPERVISOR_platform_op(&op);
299 if (ret)
300 return NR_CPUS;
301
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++) {
305 info->xen_cpuid = i;
306 ret = HYPERVISOR_platform_op(&op);
307 if (ret)
308 continue;
309 max_acpi_id = max(info->acpi_id, max_acpi_id);
310 }
311 max_acpi_id *= 2; /* Slack for CPU hotplug support. */
312 pr_debug("Max ACPI ID: %u\n", max_acpi_id);
313 return max_acpi_id;
314}
315/*
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
322 * CPU amount.
323 */
324static acpi_status
325read_acpi_id(acpi_handle handle, u32 lvl, void *context, void **rv)
326{
327 u32 acpi_id;
328 acpi_status status;
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;
334
335 status = acpi_get_type(handle, &acpi_type);
336 if (ACPI_FAILURE(status))
337 return AE_OK;
338
339 switch (acpi_type) {
340 case ACPI_TYPE_PROCESSOR:
341 status = acpi_evaluate_object(handle, NULL, NULL, &buffer);
342 if (ACPI_FAILURE(status))
343 return AE_OK;
344 acpi_id = object.processor.proc_id;
345 pblk = object.processor.pblk_address;
346 break;
347 case ACPI_TYPE_DEVICE:
348 status = acpi_evaluate_integer(handle, "_UID", NULL, &tmp);
349 if (ACPI_FAILURE(status))
350 return AE_OK;
351 acpi_id = tmp;
352 break;
353 default:
354 return AE_OK;
355 }
356 if (invalid_phys_cpuid(acpi_get_phys_id(handle,
357 acpi_type == ACPI_TYPE_DEVICE,
358 acpi_id))) {
359 pr_debug("CPU with ACPI ID %u is unavailable\n", acpi_id);
360 return AE_OK;
361 }
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);
367 return AE_OK;
368 }
369 /* OK, There is a ACPI Processor object */
370 __set_bit(acpi_id, acpi_id_present);
371
372 pr_debug("ACPI CPU%u w/ PBLK:0x%lx\n", acpi_id, (unsigned long)pblk);
373
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);
379 }
380
381 status = acpi_evaluate_object(handle, "_CST", NULL, &buffer);
382 if (ACPI_FAILURE(status)) {
383 if (!pblk)
384 return AE_OK;
385 }
386 /* .. and it has a C-state */
387 __set_bit(acpi_id, acpi_id_cst_present);
388
389 return AE_OK;
390}
391static int check_acpi_ids(struct acpi_processor *pr_backup)
392{
393
394 if (!pr_backup)
395 return -ENODEV;
396
397 if (acpi_id_present && acpi_id_cst_present)
398 /* OK, done this once .. skip to uploading */
399 goto upload;
400
401 /* All online CPUs have been processed at this stage. Now verify
402 * whether in fact "online CPUs" == physical CPUs.
403 */
404 acpi_id_present = bitmap_zalloc(nr_acpi_bits, GFP_KERNEL);
405 if (!acpi_id_present)
406 return -ENOMEM;
407
408 acpi_id_cst_present = bitmap_zalloc(nr_acpi_bits, GFP_KERNEL);
409 if (!acpi_id_cst_present) {
410 bitmap_free(acpi_id_present);
411 return -ENOMEM;
412 }
413
414 acpi_psd = kcalloc(nr_acpi_bits, sizeof(struct acpi_psd_package),
415 GFP_KERNEL);
416 if (!acpi_psd) {
417 bitmap_free(acpi_id_present);
418 bitmap_free(acpi_id_cst_present);
419 return -ENOMEM;
420 }
421
422 acpi_walk_namespace(ACPI_TYPE_PROCESSOR, ACPI_ROOT_OBJECT,
423 ACPI_UINT32_MAX,
424 read_acpi_id, NULL, NULL, NULL);
425 acpi_get_devices(ACPI_PROCESSOR_DEVICE_HID, read_acpi_id, NULL, NULL);
426
427upload:
428 if (!bitmap_equal(acpi_id_present, acpi_ids_done, nr_acpi_bits)) {
429 unsigned int i;
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,
437 &acpi_psd[i],
438 sizeof(struct acpi_psd_package));
439 }
440 (void)upload_pm_data(pr_backup);
441 }
442 }
443
444 return 0;
445}
446
447/* acpi_perf_data is a pointer to percpu data. */
448static struct acpi_processor_performance __percpu *acpi_perf_data;
449
450static void free_acpi_perf_data(void)
451{
452 unsigned int i;
453
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)
457 ->shared_cpu_map);
458 free_percpu(acpi_perf_data);
459}
460
461static int xen_upload_processor_pm_data(void)
462{
463 struct acpi_processor *pr_backup = NULL;
464 unsigned int i;
465 int rc = 0;
466
467 pr_info("Uploading Xen processor PM info\n");
468
469 for_each_possible_cpu(i) {
470 struct acpi_processor *_pr;
471 _pr = per_cpu(processors, i /* APIC ID */);
472 if (!_pr)
473 continue;
474
475 if (!pr_backup) {
476 pr_backup = kzalloc(sizeof(struct acpi_processor), GFP_KERNEL);
477 if (pr_backup)
478 memcpy(pr_backup, _pr, sizeof(struct acpi_processor));
479 }
480 (void)upload_pm_data(_pr);
481 }
482
483 rc = check_acpi_ids(pr_backup);
484 kfree(pr_backup);
485
486 return rc;
487}
488
489static void xen_acpi_processor_resume_worker(struct work_struct *dummy)
490{
491 int rc;
492
493 bitmap_zero(acpi_ids_done, nr_acpi_bits);
494
495 rc = xen_upload_processor_pm_data();
496 if (rc != 0)
497 pr_info("ACPI data upload failed, error = %d\n", rc);
498}
499
500static void xen_acpi_processor_resume(void)
501{
502 static DECLARE_WORK(wq, xen_acpi_processor_resume_worker);
503
504 /*
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.
508 *
509 * So defer the upload for some point safer.
510 */
511 schedule_work(&wq);
512}
513
514static struct syscore_ops xap_syscore_ops = {
515 .resume = xen_acpi_processor_resume,
516};
517
518static int __init xen_acpi_processor_init(void)
519{
520 unsigned int i;
521 int rc;
522
523 if (!xen_initial_domain())
524 return -ENODEV;
525
526 nr_acpi_bits = get_max_acpi_id() + 1;
527 acpi_ids_done = bitmap_zalloc(nr_acpi_bits, GFP_KERNEL);
528 if (!acpi_ids_done)
529 return -ENOMEM;
530
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);
535 return -ENOMEM;
536 }
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))) {
541 rc = -ENOMEM;
542 goto err_out;
543 }
544 }
545
546 /* Do initialization in ACPI core. It is OK to fail here. */
547 (void)acpi_processor_preregister_performance(acpi_perf_data);
548
549 for_each_possible_cpu(i) {
550 struct acpi_processor *pr;
551 struct acpi_processor_performance *perf;
552
553 pr = per_cpu(processors, i);
554 perf = per_cpu_ptr(acpi_perf_data, i);
555 if (!pr)
556 continue;
557
558 pr->performance = perf;
559 rc = acpi_processor_get_performance_info(pr);
560 if (rc)
561 goto err_out;
562 }
563
564 rc = xen_upload_processor_pm_data();
565 if (rc)
566 goto err_unregister;
567
568 register_syscore_ops(&xap_syscore_ops);
569
570 return 0;
571err_unregister:
572 for_each_possible_cpu(i)
573 acpi_processor_unregister_performance(i);
574
575err_out:
576 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
577 free_acpi_perf_data();
578 bitmap_free(acpi_ids_done);
579 return rc;
580}
581static void __exit xen_acpi_processor_exit(void)
582{
583 int i;
584
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);
589 kfree(acpi_psd);
590 for_each_possible_cpu(i)
591 acpi_processor_unregister_performance(i);
592
593 free_acpi_perf_data();
594}
595
596MODULE_AUTHOR("Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>");
597MODULE_DESCRIPTION("Xen ACPI Processor P-states (and Cx) driver which uploads PM data to Xen hypervisor");
598MODULE_LICENSE("GPL");
599
600/* We want to be loaded before the CPU freq scaling drivers are loaded.
601 * They are loaded in late_initcall. */
602device_initcall(xen_acpi_processor_init);
603module_exit(xen_acpi_processor_exit);
1/*
2 * Copyright 2012 by Oracle Inc
3 * Author: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
4 *
5 * This code borrows ideas from https://lkml.org/lkml/2011/11/30/249
6 * so many thanks go to Kevin Tian <kevin.tian@intel.com>
7 * and Yu Ke <ke.yu@intel.com>.
8 *
9 * This program is free software; you can redistribute it and/or modify it
10 * under the terms and conditions of the GNU General Public License,
11 * version 2, as published by the Free Software Foundation.
12 *
13 * This program is distributed in the hope it will be useful, but WITHOUT
14 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
16 * more details.
17 *
18 */
19
20#include <linux/cpumask.h>
21#include <linux/cpufreq.h>
22#include <linux/freezer.h>
23#include <linux/kernel.h>
24#include <linux/kthread.h>
25#include <linux/init.h>
26#include <linux/module.h>
27#include <linux/types.h>
28#include <acpi/acpi_bus.h>
29#include <acpi/acpi_drivers.h>
30#include <acpi/processor.h>
31
32#include <xen/xen.h>
33#include <xen/interface/platform.h>
34#include <asm/xen/hypercall.h>
35
36#define DRV_NAME "xen-acpi-processor: "
37
38static int no_hypercall;
39MODULE_PARM_DESC(off, "Inhibit the hypercall.");
40module_param_named(off, no_hypercall, int, 0400);
41
42/*
43 * Note: Do not convert the acpi_id* below to cpumask_var_t or use cpumask_bit
44 * - as those shrink to nr_cpu_bits (which is dependent on possible_cpu), which
45 * can be less than what we want to put in. Instead use the 'nr_acpi_bits'
46 * which is dynamically computed based on the MADT or x2APIC table.
47 */
48static unsigned int nr_acpi_bits;
49/* Mutex to protect the acpi_ids_done - for CPU hotplug use. */
50static DEFINE_MUTEX(acpi_ids_mutex);
51/* Which ACPI ID we have processed from 'struct acpi_processor'. */
52static unsigned long *acpi_ids_done;
53/* Which ACPI ID exist in the SSDT/DSDT processor definitions. */
54static unsigned long __initdata *acpi_id_present;
55/* And if there is an _CST definition (or a PBLK) for the ACPI IDs */
56static unsigned long __initdata *acpi_id_cst_present;
57
58static int push_cxx_to_hypervisor(struct acpi_processor *_pr)
59{
60 struct xen_platform_op op = {
61 .cmd = XENPF_set_processor_pminfo,
62 .interface_version = XENPF_INTERFACE_VERSION,
63 .u.set_pminfo.id = _pr->acpi_id,
64 .u.set_pminfo.type = XEN_PM_CX,
65 };
66 struct xen_processor_cx *dst_cx, *dst_cx_states = NULL;
67 struct acpi_processor_cx *cx;
68 unsigned int i, ok;
69 int ret = 0;
70
71 dst_cx_states = kcalloc(_pr->power.count,
72 sizeof(struct xen_processor_cx), GFP_KERNEL);
73 if (!dst_cx_states)
74 return -ENOMEM;
75
76 for (ok = 0, i = 1; i <= _pr->power.count; i++) {
77 cx = &_pr->power.states[i];
78 if (!cx->valid)
79 continue;
80
81 dst_cx = &(dst_cx_states[ok++]);
82
83 dst_cx->reg.space_id = ACPI_ADR_SPACE_SYSTEM_IO;
84 if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) {
85 dst_cx->reg.bit_width = 8;
86 dst_cx->reg.bit_offset = 0;
87 dst_cx->reg.access_size = 1;
88 } else {
89 dst_cx->reg.space_id = ACPI_ADR_SPACE_FIXED_HARDWARE;
90 if (cx->entry_method == ACPI_CSTATE_FFH) {
91 /* NATIVE_CSTATE_BEYOND_HALT */
92 dst_cx->reg.bit_offset = 2;
93 dst_cx->reg.bit_width = 1; /* VENDOR_INTEL */
94 }
95 dst_cx->reg.access_size = 0;
96 }
97 dst_cx->reg.address = cx->address;
98
99 dst_cx->type = cx->type;
100 dst_cx->latency = cx->latency;
101 dst_cx->power = cx->power;
102
103 dst_cx->dpcnt = 0;
104 set_xen_guest_handle(dst_cx->dp, NULL);
105 }
106 if (!ok) {
107 pr_debug(DRV_NAME "No _Cx for ACPI CPU %u\n", _pr->acpi_id);
108 kfree(dst_cx_states);
109 return -EINVAL;
110 }
111 op.u.set_pminfo.power.count = ok;
112 op.u.set_pminfo.power.flags.bm_control = _pr->flags.bm_control;
113 op.u.set_pminfo.power.flags.bm_check = _pr->flags.bm_check;
114 op.u.set_pminfo.power.flags.has_cst = _pr->flags.has_cst;
115 op.u.set_pminfo.power.flags.power_setup_done =
116 _pr->flags.power_setup_done;
117
118 set_xen_guest_handle(op.u.set_pminfo.power.states, dst_cx_states);
119
120 if (!no_hypercall)
121 ret = HYPERVISOR_dom0_op(&op);
122
123 if (!ret) {
124 pr_debug("ACPI CPU%u - C-states uploaded.\n", _pr->acpi_id);
125 for (i = 1; i <= _pr->power.count; i++) {
126 cx = &_pr->power.states[i];
127 if (!cx->valid)
128 continue;
129 pr_debug(" C%d: %s %d uS\n",
130 cx->type, cx->desc, (u32)cx->latency);
131 }
132 } else if (ret != -EINVAL)
133 /* EINVAL means the ACPI ID is incorrect - meaning the ACPI
134 * table is referencing a non-existing CPU - which can happen
135 * with broken ACPI tables. */
136 pr_err(DRV_NAME "(CX): Hypervisor error (%d) for ACPI CPU%u\n",
137 ret, _pr->acpi_id);
138
139 kfree(dst_cx_states);
140
141 return ret;
142}
143static struct xen_processor_px *
144xen_copy_pss_data(struct acpi_processor *_pr,
145 struct xen_processor_performance *dst_perf)
146{
147 struct xen_processor_px *dst_states = NULL;
148 unsigned int i;
149
150 BUILD_BUG_ON(sizeof(struct xen_processor_px) !=
151 sizeof(struct acpi_processor_px));
152
153 dst_states = kcalloc(_pr->performance->state_count,
154 sizeof(struct xen_processor_px), GFP_KERNEL);
155 if (!dst_states)
156 return ERR_PTR(-ENOMEM);
157
158 dst_perf->state_count = _pr->performance->state_count;
159 for (i = 0; i < _pr->performance->state_count; i++) {
160 /* Fortunatly for us, they are both the same size */
161 memcpy(&(dst_states[i]), &(_pr->performance->states[i]),
162 sizeof(struct acpi_processor_px));
163 }
164 return dst_states;
165}
166static int xen_copy_psd_data(struct acpi_processor *_pr,
167 struct xen_processor_performance *dst)
168{
169 struct acpi_psd_package *pdomain;
170
171 BUILD_BUG_ON(sizeof(struct xen_psd_package) !=
172 sizeof(struct acpi_psd_package));
173
174 /* This information is enumerated only if acpi_processor_preregister_performance
175 * has been called.
176 */
177 dst->shared_type = _pr->performance->shared_type;
178
179 pdomain = &(_pr->performance->domain_info);
180
181 /* 'acpi_processor_preregister_performance' does not parse if the
182 * num_processors <= 1, but Xen still requires it. Do it manually here.
183 */
184 if (pdomain->num_processors <= 1) {
185 if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ALL)
186 dst->shared_type = CPUFREQ_SHARED_TYPE_ALL;
187 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_HW_ALL)
188 dst->shared_type = CPUFREQ_SHARED_TYPE_HW;
189 else if (pdomain->coord_type == DOMAIN_COORD_TYPE_SW_ANY)
190 dst->shared_type = CPUFREQ_SHARED_TYPE_ANY;
191
192 }
193 memcpy(&(dst->domain_info), pdomain, sizeof(struct acpi_psd_package));
194 return 0;
195}
196static int xen_copy_pct_data(struct acpi_pct_register *pct,
197 struct xen_pct_register *dst_pct)
198{
199 /* It would be nice if you could just do 'memcpy(pct, dst_pct') but
200 * sadly the Xen structure did not have the proper padding so the
201 * descriptor field takes two (dst_pct) bytes instead of one (pct).
202 */
203 dst_pct->descriptor = pct->descriptor;
204 dst_pct->length = pct->length;
205 dst_pct->space_id = pct->space_id;
206 dst_pct->bit_width = pct->bit_width;
207 dst_pct->bit_offset = pct->bit_offset;
208 dst_pct->reserved = pct->reserved;
209 dst_pct->address = pct->address;
210 return 0;
211}
212static int push_pxx_to_hypervisor(struct acpi_processor *_pr)
213{
214 int ret = 0;
215 struct xen_platform_op op = {
216 .cmd = XENPF_set_processor_pminfo,
217 .interface_version = XENPF_INTERFACE_VERSION,
218 .u.set_pminfo.id = _pr->acpi_id,
219 .u.set_pminfo.type = XEN_PM_PX,
220 };
221 struct xen_processor_performance *dst_perf;
222 struct xen_processor_px *dst_states = NULL;
223
224 dst_perf = &op.u.set_pminfo.perf;
225
226 dst_perf->platform_limit = _pr->performance_platform_limit;
227 dst_perf->flags |= XEN_PX_PPC;
228 xen_copy_pct_data(&(_pr->performance->control_register),
229 &dst_perf->control_register);
230 xen_copy_pct_data(&(_pr->performance->status_register),
231 &dst_perf->status_register);
232 dst_perf->flags |= XEN_PX_PCT;
233 dst_states = xen_copy_pss_data(_pr, dst_perf);
234 if (!IS_ERR_OR_NULL(dst_states)) {
235 set_xen_guest_handle(dst_perf->states, dst_states);
236 dst_perf->flags |= XEN_PX_PSS;
237 }
238 if (!xen_copy_psd_data(_pr, dst_perf))
239 dst_perf->flags |= XEN_PX_PSD;
240
241 if (dst_perf->flags != (XEN_PX_PSD | XEN_PX_PSS | XEN_PX_PCT | XEN_PX_PPC)) {
242 pr_warn(DRV_NAME "ACPI CPU%u missing some P-state data (%x), skipping.\n",
243 _pr->acpi_id, dst_perf->flags);
244 ret = -ENODEV;
245 goto err_free;
246 }
247
248 if (!no_hypercall)
249 ret = HYPERVISOR_dom0_op(&op);
250
251 if (!ret) {
252 struct acpi_processor_performance *perf;
253 unsigned int i;
254
255 perf = _pr->performance;
256 pr_debug("ACPI CPU%u - P-states uploaded.\n", _pr->acpi_id);
257 for (i = 0; i < perf->state_count; i++) {
258 pr_debug(" %cP%d: %d MHz, %d mW, %d uS\n",
259 (i == perf->state ? '*' : ' '), i,
260 (u32) perf->states[i].core_frequency,
261 (u32) perf->states[i].power,
262 (u32) perf->states[i].transition_latency);
263 }
264 } else if (ret != -EINVAL)
265 /* EINVAL means the ACPI ID is incorrect - meaning the ACPI
266 * table is referencing a non-existing CPU - which can happen
267 * with broken ACPI tables. */
268 pr_warn(DRV_NAME "(_PXX): Hypervisor error (%d) for ACPI CPU%u\n",
269 ret, _pr->acpi_id);
270err_free:
271 if (!IS_ERR_OR_NULL(dst_states))
272 kfree(dst_states);
273
274 return ret;
275}
276static int upload_pm_data(struct acpi_processor *_pr)
277{
278 int err = 0;
279
280 mutex_lock(&acpi_ids_mutex);
281 if (__test_and_set_bit(_pr->acpi_id, acpi_ids_done)) {
282 mutex_unlock(&acpi_ids_mutex);
283 return -EBUSY;
284 }
285 if (_pr->flags.power)
286 err = push_cxx_to_hypervisor(_pr);
287
288 if (_pr->performance && _pr->performance->states)
289 err |= push_pxx_to_hypervisor(_pr);
290
291 mutex_unlock(&acpi_ids_mutex);
292 return err;
293}
294static unsigned int __init get_max_acpi_id(void)
295{
296 struct xenpf_pcpuinfo *info;
297 struct xen_platform_op op = {
298 .cmd = XENPF_get_cpuinfo,
299 .interface_version = XENPF_INTERFACE_VERSION,
300 };
301 int ret = 0;
302 unsigned int i, last_cpu, max_acpi_id = 0;
303
304 info = &op.u.pcpu_info;
305 info->xen_cpuid = 0;
306
307 ret = HYPERVISOR_dom0_op(&op);
308 if (ret)
309 return NR_CPUS;
310
311 /* The max_present is the same irregardless of the xen_cpuid */
312 last_cpu = op.u.pcpu_info.max_present;
313 for (i = 0; i <= last_cpu; i++) {
314 info->xen_cpuid = i;
315 ret = HYPERVISOR_dom0_op(&op);
316 if (ret)
317 continue;
318 max_acpi_id = max(info->acpi_id, max_acpi_id);
319 }
320 max_acpi_id *= 2; /* Slack for CPU hotplug support. */
321 pr_debug(DRV_NAME "Max ACPI ID: %u\n", max_acpi_id);
322 return max_acpi_id;
323}
324/*
325 * The read_acpi_id and check_acpi_ids are there to support the Xen
326 * oddity of virtual CPUs != physical CPUs in the initial domain.
327 * The user can supply 'xen_max_vcpus=X' on the Xen hypervisor line
328 * which will band the amount of CPUs the initial domain can see.
329 * In general that is OK, except it plays havoc with any of the
330 * for_each_[present|online]_cpu macros which are banded to the virtual
331 * CPU amount.
332 */
333static acpi_status __init
334read_acpi_id(acpi_handle handle, u32 lvl, void *context, void **rv)
335{
336 u32 acpi_id;
337 acpi_status status;
338 acpi_object_type acpi_type;
339 unsigned long long tmp;
340 union acpi_object object = { 0 };
341 struct acpi_buffer buffer = { sizeof(union acpi_object), &object };
342 acpi_io_address pblk = 0;
343
344 status = acpi_get_type(handle, &acpi_type);
345 if (ACPI_FAILURE(status))
346 return AE_OK;
347
348 switch (acpi_type) {
349 case ACPI_TYPE_PROCESSOR:
350 status = acpi_evaluate_object(handle, NULL, NULL, &buffer);
351 if (ACPI_FAILURE(status))
352 return AE_OK;
353 acpi_id = object.processor.proc_id;
354 pblk = object.processor.pblk_address;
355 break;
356 case ACPI_TYPE_DEVICE:
357 status = acpi_evaluate_integer(handle, "_UID", NULL, &tmp);
358 if (ACPI_FAILURE(status))
359 return AE_OK;
360 acpi_id = tmp;
361 break;
362 default:
363 return AE_OK;
364 }
365 /* There are more ACPI Processor objects than in x2APIC or MADT.
366 * This can happen with incorrect ACPI SSDT declerations. */
367 if (acpi_id > nr_acpi_bits) {
368 pr_debug(DRV_NAME "We only have %u, trying to set %u\n",
369 nr_acpi_bits, acpi_id);
370 return AE_OK;
371 }
372 /* OK, There is a ACPI Processor object */
373 __set_bit(acpi_id, acpi_id_present);
374
375 pr_debug(DRV_NAME "ACPI CPU%u w/ PBLK:0x%lx\n", acpi_id,
376 (unsigned long)pblk);
377
378 status = acpi_evaluate_object(handle, "_CST", NULL, &buffer);
379 if (ACPI_FAILURE(status)) {
380 if (!pblk)
381 return AE_OK;
382 }
383 /* .. and it has a C-state */
384 __set_bit(acpi_id, acpi_id_cst_present);
385
386 return AE_OK;
387}
388static int __init check_acpi_ids(struct acpi_processor *pr_backup)
389{
390
391 if (!pr_backup)
392 return -ENODEV;
393
394 /* All online CPUs have been processed at this stage. Now verify
395 * whether in fact "online CPUs" == physical CPUs.
396 */
397 acpi_id_present = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL);
398 if (!acpi_id_present)
399 return -ENOMEM;
400
401 acpi_id_cst_present = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL);
402 if (!acpi_id_cst_present) {
403 kfree(acpi_id_present);
404 return -ENOMEM;
405 }
406
407 acpi_walk_namespace(ACPI_TYPE_PROCESSOR, ACPI_ROOT_OBJECT,
408 ACPI_UINT32_MAX,
409 read_acpi_id, NULL, NULL, NULL);
410 acpi_get_devices("ACPI0007", read_acpi_id, NULL, NULL);
411
412 if (!bitmap_equal(acpi_id_present, acpi_ids_done, nr_acpi_bits)) {
413 unsigned int i;
414 for_each_set_bit(i, acpi_id_present, nr_acpi_bits) {
415 pr_backup->acpi_id = i;
416 /* Mask out C-states if there are no _CST or PBLK */
417 pr_backup->flags.power = test_bit(i, acpi_id_cst_present);
418 (void)upload_pm_data(pr_backup);
419 }
420 }
421 kfree(acpi_id_present);
422 acpi_id_present = NULL;
423 kfree(acpi_id_cst_present);
424 acpi_id_cst_present = NULL;
425 return 0;
426}
427static int __init check_prereq(void)
428{
429 struct cpuinfo_x86 *c = &cpu_data(0);
430
431 if (!xen_initial_domain())
432 return -ENODEV;
433
434 if (!acpi_gbl_FADT.smi_command)
435 return -ENODEV;
436
437 if (c->x86_vendor == X86_VENDOR_INTEL) {
438 if (!cpu_has(c, X86_FEATURE_EST))
439 return -ENODEV;
440
441 return 0;
442 }
443 if (c->x86_vendor == X86_VENDOR_AMD) {
444 /* Copied from powernow-k8.h, can't include ../cpufreq/powernow
445 * as we get compile warnings for the static functions.
446 */
447#define CPUID_FREQ_VOLT_CAPABILITIES 0x80000007
448#define USE_HW_PSTATE 0x00000080
449 u32 eax, ebx, ecx, edx;
450 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
451 if ((edx & USE_HW_PSTATE) != USE_HW_PSTATE)
452 return -ENODEV;
453 return 0;
454 }
455 return -ENODEV;
456}
457/* acpi_perf_data is a pointer to percpu data. */
458static struct acpi_processor_performance __percpu *acpi_perf_data;
459
460static void free_acpi_perf_data(void)
461{
462 unsigned int i;
463
464 /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
465 for_each_possible_cpu(i)
466 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
467 ->shared_cpu_map);
468 free_percpu(acpi_perf_data);
469}
470
471static int __init xen_acpi_processor_init(void)
472{
473 struct acpi_processor *pr_backup = NULL;
474 unsigned int i;
475 int rc = check_prereq();
476
477 if (rc)
478 return rc;
479
480 nr_acpi_bits = get_max_acpi_id() + 1;
481 acpi_ids_done = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL);
482 if (!acpi_ids_done)
483 return -ENOMEM;
484
485 acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
486 if (!acpi_perf_data) {
487 pr_debug(DRV_NAME "Memory allocation error for acpi_perf_data.\n");
488 kfree(acpi_ids_done);
489 return -ENOMEM;
490 }
491 for_each_possible_cpu(i) {
492 if (!zalloc_cpumask_var_node(
493 &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
494 GFP_KERNEL, cpu_to_node(i))) {
495 rc = -ENOMEM;
496 goto err_out;
497 }
498 }
499
500 /* Do initialization in ACPI core. It is OK to fail here. */
501 (void)acpi_processor_preregister_performance(acpi_perf_data);
502
503 for_each_possible_cpu(i) {
504 struct acpi_processor_performance *perf;
505
506 perf = per_cpu_ptr(acpi_perf_data, i);
507 rc = acpi_processor_register_performance(perf, i);
508 if (rc)
509 goto err_out;
510 }
511 rc = acpi_processor_notify_smm(THIS_MODULE);
512 if (rc)
513 goto err_unregister;
514
515 for_each_possible_cpu(i) {
516 struct acpi_processor *_pr;
517 _pr = per_cpu(processors, i /* APIC ID */);
518 if (!_pr)
519 continue;
520
521 if (!pr_backup) {
522 pr_backup = kzalloc(sizeof(struct acpi_processor), GFP_KERNEL);
523 memcpy(pr_backup, _pr, sizeof(struct acpi_processor));
524 }
525 (void)upload_pm_data(_pr);
526 }
527 rc = check_acpi_ids(pr_backup);
528 if (rc)
529 goto err_unregister;
530
531 kfree(pr_backup);
532
533 return 0;
534err_unregister:
535 for_each_possible_cpu(i) {
536 struct acpi_processor_performance *perf;
537 perf = per_cpu_ptr(acpi_perf_data, i);
538 acpi_processor_unregister_performance(perf, i);
539 }
540err_out:
541 /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
542 free_acpi_perf_data();
543 kfree(acpi_ids_done);
544 return rc;
545}
546static void __exit xen_acpi_processor_exit(void)
547{
548 int i;
549
550 kfree(acpi_ids_done);
551 for_each_possible_cpu(i) {
552 struct acpi_processor_performance *perf;
553 perf = per_cpu_ptr(acpi_perf_data, i);
554 acpi_processor_unregister_performance(perf, i);
555 }
556 free_acpi_perf_data();
557}
558
559MODULE_AUTHOR("Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>");
560MODULE_DESCRIPTION("Xen ACPI Processor P-states (and Cx) driver which uploads PM data to Xen hypervisor");
561MODULE_LICENSE("GPL");
562
563/* We want to be loaded before the CPU freq scaling drivers are loaded.
564 * They are loaded in late_initcall. */
565device_initcall(xen_acpi_processor_init);
566module_exit(xen_acpi_processor_exit);