Linux Audio

Check our new training course

Loading...
v6.13.7
  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(_pr->power.count,
 65				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(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);
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(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(_pr->performance->state_count,
146			     sizeof(struct xen_processor_px), GFP_KERNEL);
147	if (!dst_states)
148		return ERR_PTR(-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(&(_pr->performance->control_register),
221			  &dst_perf->control_register);
222	xen_copy_pct_data(&(_pr->performance->status_register),
223			  &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(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_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);
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(dst_states))
264		kfree(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(&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(&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);
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);
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(handle, &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(handle, NULL, NULL, &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, "_UID", NULL, &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(acpi_get_phys_id(handle,
358						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, &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(handle, "_CST", NULL, &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(nr_acpi_bits, GFP_KERNEL);
406	if (!acpi_id_present)
407		return -ENOMEM;
408
409	acpi_id_cst_present = bitmap_zalloc(nr_acpi_bits, GFP_KERNEL);
410	if (!acpi_id_cst_present) {
411		bitmap_free(acpi_id_present);
412		return -ENOMEM;
413	}
414
415	acpi_psd = kcalloc(nr_acpi_bits, sizeof(struct acpi_psd_package),
416			   GFP_KERNEL);
417	if (!acpi_psd) {
418		bitmap_free(acpi_id_present);
419		bitmap_free(acpi_id_cst_present);
420		return -ENOMEM;
421	}
422
423	acpi_walk_namespace(ACPI_TYPE_PROCESSOR, ACPI_ROOT_OBJECT,
424			    ACPI_UINT32_MAX,
425			    read_acpi_id, NULL, NULL, NULL);
426	acpi_get_devices(ACPI_PROCESSOR_DEVICE_HID, read_acpi_id, NULL, NULL);
427
428upload:
429	if (!bitmap_equal(acpi_id_present, acpi_ids_done, 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_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(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(_pr, sizeof(*_pr), GFP_KERNEL);
 
 
 
478		(void)upload_pm_data(_pr);
479	}
480
481	rc = check_acpi_ids(pr_backup);
482	kfree(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(acpi_ids_done, 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(&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(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(acpi_ids_done);
533		return -ENOMEM;
534	}
535	for_each_possible_cpu(i) {
536		if (!zalloc_cpumask_var_node(
537			&per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
538			GFP_KERNEL, cpu_to_node(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(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(&xap_syscore_ops);
567
568	return 0;
569err_unregister:
570	for_each_possible_cpu(i)
571		acpi_processor_unregister_performance(i);
572
573err_out:
574	/* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
575	free_acpi_perf_data();
576	bitmap_free(acpi_ids_done);
577	return rc;
578}
579static void __exit xen_acpi_processor_exit(void)
580{
581	int i;
582
583	unregister_syscore_ops(&xap_syscore_ops);
584	bitmap_free(acpi_ids_done);
585	bitmap_free(acpi_id_present);
586	bitmap_free(acpi_id_cst_present);
587	kfree(acpi_psd);
588	for_each_possible_cpu(i)
589		acpi_processor_unregister_performance(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);
v4.17
 
  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#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
 21
 22#include <linux/cpumask.h>
 23#include <linux/cpufreq.h>
 24#include <linux/freezer.h>
 25#include <linux/kernel.h>
 26#include <linux/kthread.h>
 27#include <linux/init.h>
 28#include <linux/module.h>
 29#include <linux/types.h>
 30#include <linux/syscore_ops.h>
 31#include <linux/acpi.h>
 32#include <acpi/processor.h>
 33#include <xen/xen.h>
 34#include <xen/interface/platform.h>
 35#include <asm/xen/hypercall.h>
 36
 37static int no_hypercall;
 38MODULE_PARM_DESC(off, "Inhibit the hypercall.");
 39module_param_named(off, no_hypercall, int, 0400);
 40
 41/*
 42 * Note: Do not convert the acpi_id* below to cpumask_var_t or use cpumask_bit
 43 * - as those shrink to nr_cpu_bits (which is dependent on possible_cpu), which
 44 * can be less than what we want to put in. Instead use the 'nr_acpi_bits'
 45 * which is dynamically computed based on the MADT or x2APIC table.
 46 */
 47static unsigned int nr_acpi_bits;
 48/* Mutex to protect the acpi_ids_done - for CPU hotplug use. */
 49static DEFINE_MUTEX(acpi_ids_mutex);
 50/* Which ACPI ID we have processed from 'struct acpi_processor'. */
 51static unsigned long *acpi_ids_done;
 52/* Which ACPI ID exist in the SSDT/DSDT processor definitions. */
 53static unsigned long *acpi_id_present;
 54/* And if there is an _CST definition (or a PBLK) for the ACPI IDs */
 55static unsigned long *acpi_id_cst_present;
 56/* Which ACPI P-State dependencies for a enumerated processor */
 57static struct acpi_psd_package *acpi_psd;
 58
 59static int push_cxx_to_hypervisor(struct acpi_processor *_pr)
 60{
 61	struct xen_platform_op op = {
 62		.cmd			= XENPF_set_processor_pminfo,
 63		.interface_version	= XENPF_INTERFACE_VERSION,
 64		.u.set_pminfo.id	= _pr->acpi_id,
 65		.u.set_pminfo.type	= XEN_PM_CX,
 66	};
 67	struct xen_processor_cx *dst_cx, *dst_cx_states = NULL;
 68	struct acpi_processor_cx *cx;
 69	unsigned int i, ok;
 70	int ret = 0;
 71
 72	dst_cx_states = kcalloc(_pr->power.count,
 73				sizeof(struct xen_processor_cx), GFP_KERNEL);
 74	if (!dst_cx_states)
 75		return -ENOMEM;
 76
 77	for (ok = 0, i = 1; i <= _pr->power.count; i++) {
 78		cx = &_pr->power.states[i];
 79		if (!cx->valid)
 80			continue;
 81
 82		dst_cx = &(dst_cx_states[ok++]);
 83
 84		dst_cx->reg.space_id = ACPI_ADR_SPACE_SYSTEM_IO;
 85		if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) {
 86			dst_cx->reg.bit_width = 8;
 87			dst_cx->reg.bit_offset = 0;
 88			dst_cx->reg.access_size = 1;
 89		} else {
 90			dst_cx->reg.space_id = ACPI_ADR_SPACE_FIXED_HARDWARE;
 91			if (cx->entry_method == ACPI_CSTATE_FFH) {
 92				/* NATIVE_CSTATE_BEYOND_HALT */
 93				dst_cx->reg.bit_offset = 2;
 94				dst_cx->reg.bit_width = 1; /* VENDOR_INTEL */
 95			}
 96			dst_cx->reg.access_size = 0;
 97		}
 98		dst_cx->reg.address = cx->address;
 99
100		dst_cx->type = cx->type;
101		dst_cx->latency = cx->latency;
102
103		dst_cx->dpcnt = 0;
104		set_xen_guest_handle(dst_cx->dp, NULL);
105	}
106	if (!ok) {
107		pr_debug("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_platform_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) && (ret != -ENOSYS))
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("(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("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_platform_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) && (ret != -ENOSYS))
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("(_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_platform_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_platform_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("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
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("max acpi id %u, trying to set %u\n",
369			 nr_acpi_bits - 1, 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("ACPI CPU%u w/ PBLK:0x%lx\n", acpi_id, (unsigned long)pblk);
376
377	/* It has P-state dependencies */
378	if (!acpi_processor_get_psd(handle, &acpi_psd[acpi_id])) {
379		pr_debug("ACPI CPU%u w/ PST:coord_type = %llu domain = %llu\n",
380			 acpi_id, acpi_psd[acpi_id].coord_type,
381			 acpi_psd[acpi_id].domain);
382	}
383
384	status = acpi_evaluate_object(handle, "_CST", NULL, &buffer);
385	if (ACPI_FAILURE(status)) {
386		if (!pblk)
387			return AE_OK;
388	}
389	/* .. and it has a C-state */
390	__set_bit(acpi_id, acpi_id_cst_present);
391
392	return AE_OK;
393}
394static int check_acpi_ids(struct acpi_processor *pr_backup)
395{
396
397	if (!pr_backup)
398		return -ENODEV;
399
400	if (acpi_id_present && acpi_id_cst_present)
401		/* OK, done this once .. skip to uploading */
402		goto upload;
403
404	/* All online CPUs have been processed at this stage. Now verify
405	 * whether in fact "online CPUs" == physical CPUs.
406	 */
407	acpi_id_present = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL);
408	if (!acpi_id_present)
409		return -ENOMEM;
410
411	acpi_id_cst_present = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL);
412	if (!acpi_id_cst_present) {
413		kfree(acpi_id_present);
414		return -ENOMEM;
415	}
416
417	acpi_psd = kcalloc(nr_acpi_bits, sizeof(struct acpi_psd_package),
418			   GFP_KERNEL);
419	if (!acpi_psd) {
420		kfree(acpi_id_present);
421		kfree(acpi_id_cst_present);
422		return -ENOMEM;
423	}
424
425	acpi_walk_namespace(ACPI_TYPE_PROCESSOR, ACPI_ROOT_OBJECT,
426			    ACPI_UINT32_MAX,
427			    read_acpi_id, NULL, NULL, NULL);
428	acpi_get_devices(ACPI_PROCESSOR_DEVICE_HID, read_acpi_id, NULL, NULL);
429
430upload:
431	if (!bitmap_equal(acpi_id_present, acpi_ids_done, nr_acpi_bits)) {
432		unsigned int i;
433		for_each_set_bit(i, acpi_id_present, nr_acpi_bits) {
434			pr_backup->acpi_id = i;
435			/* Mask out C-states if there are no _CST or PBLK */
436			pr_backup->flags.power = test_bit(i, acpi_id_cst_present);
437			/* num_entries is non-zero if we evaluated _PSD */
438			if (acpi_psd[i].num_entries) {
439				memcpy(&pr_backup->performance->domain_info,
440				       &acpi_psd[i],
441				       sizeof(struct acpi_psd_package));
442			}
443			(void)upload_pm_data(pr_backup);
444		}
445	}
446
447	return 0;
448}
449
450/* acpi_perf_data is a pointer to percpu data. */
451static struct acpi_processor_performance __percpu *acpi_perf_data;
452
453static void free_acpi_perf_data(void)
454{
455	unsigned int i;
456
457	/* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
458	for_each_possible_cpu(i)
459		free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
460				 ->shared_cpu_map);
461	free_percpu(acpi_perf_data);
462}
463
464static int xen_upload_processor_pm_data(void)
465{
466	struct acpi_processor *pr_backup = NULL;
467	unsigned int i;
468	int rc = 0;
469
470	pr_info("Uploading Xen processor PM info\n");
471
472	for_each_possible_cpu(i) {
473		struct acpi_processor *_pr;
474		_pr = per_cpu(processors, i /* APIC ID */);
475		if (!_pr)
476			continue;
477
478		if (!pr_backup) {
479			pr_backup = kzalloc(sizeof(struct acpi_processor), GFP_KERNEL);
480			if (pr_backup)
481				memcpy(pr_backup, _pr, sizeof(struct acpi_processor));
482		}
483		(void)upload_pm_data(_pr);
484	}
485
486	rc = check_acpi_ids(pr_backup);
487	kfree(pr_backup);
488
489	return rc;
490}
491
492static void xen_acpi_processor_resume_worker(struct work_struct *dummy)
493{
494	int rc;
495
496	bitmap_zero(acpi_ids_done, nr_acpi_bits);
497
498	rc = xen_upload_processor_pm_data();
499	if (rc != 0)
500		pr_info("ACPI data upload failed, error = %d\n", rc);
501}
502
503static void xen_acpi_processor_resume(void)
504{
505	static DECLARE_WORK(wq, xen_acpi_processor_resume_worker);
506
507	/*
508	 * xen_upload_processor_pm_data() calls non-atomic code.
509	 * However, the context for xen_acpi_processor_resume is syscore
510	 * with only the boot CPU online and in an atomic context.
511	 *
512	 * So defer the upload for some point safer.
513	 */
514	schedule_work(&wq);
515}
516
517static struct syscore_ops xap_syscore_ops = {
518	.resume	= xen_acpi_processor_resume,
519};
520
521static int __init xen_acpi_processor_init(void)
522{
523	unsigned int i;
524	int rc;
525
526	if (!xen_initial_domain())
527		return -ENODEV;
528
529	nr_acpi_bits = get_max_acpi_id() + 1;
530	acpi_ids_done = kcalloc(BITS_TO_LONGS(nr_acpi_bits), sizeof(unsigned long), GFP_KERNEL);
531	if (!acpi_ids_done)
532		return -ENOMEM;
533
534	acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
535	if (!acpi_perf_data) {
536		pr_debug("Memory allocation error for acpi_perf_data\n");
537		kfree(acpi_ids_done);
538		return -ENOMEM;
539	}
540	for_each_possible_cpu(i) {
541		if (!zalloc_cpumask_var_node(
542			&per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
543			GFP_KERNEL, cpu_to_node(i))) {
544			rc = -ENOMEM;
545			goto err_out;
546		}
547	}
548
549	/* Do initialization in ACPI core. It is OK to fail here. */
550	(void)acpi_processor_preregister_performance(acpi_perf_data);
551
552	for_each_possible_cpu(i) {
553		struct acpi_processor *pr;
554		struct acpi_processor_performance *perf;
555
556		pr = per_cpu(processors, i);
557		perf = per_cpu_ptr(acpi_perf_data, i);
558		if (!pr)
559			continue;
560
561		pr->performance = perf;
562		rc = acpi_processor_get_performance_info(pr);
563		if (rc)
564			goto err_out;
565	}
566
567	rc = xen_upload_processor_pm_data();
568	if (rc)
569		goto err_unregister;
570
571	register_syscore_ops(&xap_syscore_ops);
572
573	return 0;
574err_unregister:
575	for_each_possible_cpu(i)
576		acpi_processor_unregister_performance(i);
577
578err_out:
579	/* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
580	free_acpi_perf_data();
581	kfree(acpi_ids_done);
582	return rc;
583}
584static void __exit xen_acpi_processor_exit(void)
585{
586	int i;
587
588	unregister_syscore_ops(&xap_syscore_ops);
589	kfree(acpi_ids_done);
590	kfree(acpi_id_present);
591	kfree(acpi_id_cst_present);
592	kfree(acpi_psd);
593	for_each_possible_cpu(i)
594		acpi_processor_unregister_performance(i);
595
596	free_acpi_perf_data();
597}
598
599MODULE_AUTHOR("Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>");
600MODULE_DESCRIPTION("Xen ACPI Processor P-states (and Cx) driver which uploads PM data to Xen hypervisor");
601MODULE_LICENSE("GPL");
602
603/* We want to be loaded before the CPU freq scaling drivers are loaded.
604 * They are loaded in late_initcall. */
605device_initcall(xen_acpi_processor_init);
606module_exit(xen_acpi_processor_exit);