Linux Audio

Check our new training course

Loading...
v6.2
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Suspend support specific for i386/x86-64.
  4 *
  5 * Copyright (c) 2007 Rafael J. Wysocki <rjw@sisk.pl>
  6 * Copyright (c) 2002 Pavel Machek <pavel@ucw.cz>
  7 * Copyright (c) 2001 Patrick Mochel <mochel@osdl.org>
  8 */
  9
 10#include <linux/suspend.h>
 11#include <linux/export.h>
 12#include <linux/smp.h>
 13#include <linux/perf_event.h>
 14#include <linux/tboot.h>
 15#include <linux/dmi.h>
 16#include <linux/pgtable.h>
 17
 
 18#include <asm/proto.h>
 19#include <asm/mtrr.h>
 20#include <asm/page.h>
 21#include <asm/mce.h>
 22#include <asm/suspend.h>
 23#include <asm/fpu/api.h>
 24#include <asm/debugreg.h>
 25#include <asm/cpu.h>
 26#include <asm/cacheinfo.h>
 27#include <asm/mmu_context.h>
 28#include <asm/cpu_device_id.h>
 29#include <asm/microcode.h>
 30
 31#ifdef CONFIG_X86_32
 32__visible unsigned long saved_context_ebx;
 33__visible unsigned long saved_context_esp, saved_context_ebp;
 34__visible unsigned long saved_context_esi, saved_context_edi;
 35__visible unsigned long saved_context_eflags;
 36#endif
 37struct saved_context saved_context;
 38
 39static void msr_save_context(struct saved_context *ctxt)
 40{
 41	struct saved_msr *msr = ctxt->saved_msrs.array;
 42	struct saved_msr *end = msr + ctxt->saved_msrs.num;
 43
 44	while (msr < end) {
 45		if (msr->valid)
 46			rdmsrl(msr->info.msr_no, msr->info.reg.q);
 47		msr++;
 48	}
 49}
 50
 51static void msr_restore_context(struct saved_context *ctxt)
 52{
 53	struct saved_msr *msr = ctxt->saved_msrs.array;
 54	struct saved_msr *end = msr + ctxt->saved_msrs.num;
 55
 56	while (msr < end) {
 57		if (msr->valid)
 58			wrmsrl(msr->info.msr_no, msr->info.reg.q);
 59		msr++;
 60	}
 61}
 62
 63/**
 64 * __save_processor_state() - Save CPU registers before creating a
 65 *                             hibernation image and before restoring
 66 *                             the memory state from it
 67 * @ctxt: Structure to store the registers contents in.
 68 *
 69 * NOTE: If there is a CPU register the modification of which by the
 70 * boot kernel (ie. the kernel used for loading the hibernation image)
 71 * might affect the operations of the restored target kernel (ie. the one
 72 * saved in the hibernation image), then its contents must be saved by this
 73 * function.  In other words, if kernel A is hibernated and different
 74 * kernel B is used for loading the hibernation image into memory, the
 75 * kernel A's __save_processor_state() function must save all registers
 76 * needed by kernel A, so that it can operate correctly after the resume
 77 * regardless of what kernel B does in the meantime.
 78 */
 79static void __save_processor_state(struct saved_context *ctxt)
 80{
 81#ifdef CONFIG_X86_32
 82	mtrr_save_fixed_ranges(NULL);
 83#endif
 84	kernel_fpu_begin();
 85
 86	/*
 87	 * descriptor tables
 88	 */
 89	store_idt(&ctxt->idt);
 90
 91	/*
 92	 * We save it here, but restore it only in the hibernate case.
 93	 * For ACPI S3 resume, this is loaded via 'early_gdt_desc' in 64-bit
 94	 * mode in "secondary_startup_64". In 32-bit mode it is done via
 95	 * 'pmode_gdt' in wakeup_start.
 96	 */
 97	ctxt->gdt_desc.size = GDT_SIZE - 1;
 98	ctxt->gdt_desc.address = (unsigned long)get_cpu_gdt_rw(smp_processor_id());
 99
100	store_tr(ctxt->tr);
101
102	/* XMM0..XMM15 should be handled by kernel_fpu_begin(). */
103	/*
104	 * segment registers
105	 */
 
106	savesegment(gs, ctxt->gs);
 
107#ifdef CONFIG_X86_64
 
108	savesegment(fs, ctxt->fs);
109	savesegment(ds, ctxt->ds);
110	savesegment(es, ctxt->es);
111
112	rdmsrl(MSR_FS_BASE, ctxt->fs_base);
113	rdmsrl(MSR_GS_BASE, ctxt->kernelmode_gs_base);
114	rdmsrl(MSR_KERNEL_GS_BASE, ctxt->usermode_gs_base);
115	mtrr_save_fixed_ranges(NULL);
116
117	rdmsrl(MSR_EFER, ctxt->efer);
118#endif
119
120	/*
121	 * control registers
122	 */
123	ctxt->cr0 = read_cr0();
124	ctxt->cr2 = read_cr2();
125	ctxt->cr3 = __read_cr3();
126	ctxt->cr4 = __read_cr4();
127	ctxt->misc_enable_saved = !rdmsrl_safe(MSR_IA32_MISC_ENABLE,
128					       &ctxt->misc_enable);
129	msr_save_context(ctxt);
130}
131
132/* Needed by apm.c */
133void save_processor_state(void)
134{
135	__save_processor_state(&saved_context);
136	x86_platform.save_sched_clock_state();
137}
138#ifdef CONFIG_X86_32
139EXPORT_SYMBOL(save_processor_state);
140#endif
141
142static void do_fpu_end(void)
143{
144	/*
145	 * Restore FPU regs if necessary.
146	 */
147	kernel_fpu_end();
148}
149
150static void fix_processor_context(void)
151{
152	int cpu = smp_processor_id();
153#ifdef CONFIG_X86_64
154	struct desc_struct *desc = get_cpu_gdt_rw(cpu);
155	tss_desc tss;
156#endif
157
158	/*
159	 * We need to reload TR, which requires that we change the
160	 * GDT entry to indicate "available" first.
161	 *
162	 * XXX: This could probably all be replaced by a call to
163	 * force_reload_TR().
164	 */
165	set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
166
167#ifdef CONFIG_X86_64
168	memcpy(&tss, &desc[GDT_ENTRY_TSS], sizeof(tss_desc));
169	tss.type = 0x9; /* The available 64-bit TSS (see AMD vol 2, pg 91 */
170	write_gdt_entry(desc, GDT_ENTRY_TSS, &tss, DESC_TSS);
171
172	syscall_init();				/* This sets MSR_*STAR and related */
173#else
174	if (boot_cpu_has(X86_FEATURE_SEP))
175		enable_sep_cpu();
176#endif
177	load_TR_desc();				/* This does ltr */
178	load_mm_ldt(current->active_mm);	/* This does lldt */
179	initialize_tlbstate_and_flush();
180
181	fpu__resume_cpu();
182
183	/* The processor is back on the direct GDT, load back the fixmap */
184	load_fixmap_gdt(cpu);
185}
186
187/**
188 * __restore_processor_state() - Restore the contents of CPU registers saved
189 *                               by __save_processor_state()
190 * @ctxt: Structure to load the registers contents from.
191 *
192 * The asm code that gets us here will have restored a usable GDT, although
193 * it will be pointing to the wrong alias.
194 */
195static void notrace __restore_processor_state(struct saved_context *ctxt)
196{
197	struct cpuinfo_x86 *c;
198
199	if (ctxt->misc_enable_saved)
200		wrmsrl(MSR_IA32_MISC_ENABLE, ctxt->misc_enable);
201	/*
202	 * control registers
203	 */
204	/* cr4 was introduced in the Pentium CPU */
205#ifdef CONFIG_X86_32
206	if (ctxt->cr4)
207		__write_cr4(ctxt->cr4);
208#else
209/* CONFIG X86_64 */
210	wrmsrl(MSR_EFER, ctxt->efer);
211	__write_cr4(ctxt->cr4);
212#endif
213	write_cr3(ctxt->cr3);
214	write_cr2(ctxt->cr2);
215	write_cr0(ctxt->cr0);
216
217	/* Restore the IDT. */
218	load_idt(&ctxt->idt);
219
220	/*
221	 * Just in case the asm code got us here with the SS, DS, or ES
222	 * out of sync with the GDT, update them.
223	 */
224	loadsegment(ss, __KERNEL_DS);
225	loadsegment(ds, __USER_DS);
226	loadsegment(es, __USER_DS);
227
228	/*
229	 * Restore percpu access.  Percpu access can happen in exception
230	 * handlers or in complicated helpers like load_gs_index().
231	 */
232#ifdef CONFIG_X86_64
233	wrmsrl(MSR_GS_BASE, ctxt->kernelmode_gs_base);
234#else
235	loadsegment(fs, __KERNEL_PERCPU);
 
236#endif
237
238	/* Restore the TSS, RO GDT, LDT, and usermode-relevant MSRs. */
239	fix_processor_context();
240
241	/*
242	 * Now that we have descriptor tables fully restored and working
243	 * exception handling, restore the usermode segments.
244	 */
245#ifdef CONFIG_X86_64
246	loadsegment(ds, ctxt->es);
247	loadsegment(es, ctxt->es);
248	loadsegment(fs, ctxt->fs);
249	load_gs_index(ctxt->gs);
250
251	/*
252	 * Restore FSBASE and GSBASE after restoring the selectors, since
253	 * restoring the selectors clobbers the bases.  Keep in mind
254	 * that MSR_KERNEL_GS_BASE is horribly misnamed.
255	 */
256	wrmsrl(MSR_FS_BASE, ctxt->fs_base);
257	wrmsrl(MSR_KERNEL_GS_BASE, ctxt->usermode_gs_base);
258#else
259	loadsegment(gs, ctxt->gs);
260#endif
261
262	do_fpu_end();
263	tsc_verify_tsc_adjust(true);
264	x86_platform.restore_sched_clock_state();
265	cache_bp_restore();
266	perf_restore_debug_store();
267
268	c = &cpu_data(smp_processor_id());
269	if (cpu_has(c, X86_FEATURE_MSR_IA32_FEAT_CTL))
270		init_ia32_feat_ctl(c);
271
272	microcode_bsp_resume();
273
274	/*
275	 * This needs to happen after the microcode has been updated upon resume
276	 * because some of the MSRs are "emulated" in microcode.
277	 */
278	msr_restore_context(ctxt);
279}
280
281/* Needed by apm.c */
282void notrace restore_processor_state(void)
283{
284	__restore_processor_state(&saved_context);
285}
286#ifdef CONFIG_X86_32
287EXPORT_SYMBOL(restore_processor_state);
288#endif
289
290#if defined(CONFIG_HIBERNATION) && defined(CONFIG_HOTPLUG_CPU)
291static void resume_play_dead(void)
292{
293	play_dead_common();
294	tboot_shutdown(TB_SHUTDOWN_WFS);
295	hlt_play_dead();
296}
297
298int hibernate_resume_nonboot_cpu_disable(void)
299{
300	void (*play_dead)(void) = smp_ops.play_dead;
301	int ret;
302
303	/*
304	 * Ensure that MONITOR/MWAIT will not be used in the "play dead" loop
305	 * during hibernate image restoration, because it is likely that the
306	 * monitored address will be actually written to at that time and then
307	 * the "dead" CPU will attempt to execute instructions again, but the
308	 * address in its instruction pointer may not be possible to resolve
309	 * any more at that point (the page tables used by it previously may
310	 * have been overwritten by hibernate image data).
311	 *
312	 * First, make sure that we wake up all the potentially disabled SMT
313	 * threads which have been initially brought up and then put into
314	 * mwait/cpuidle sleep.
315	 * Those will be put to proper (not interfering with hibernation
316	 * resume) sleep afterwards, and the resumed kernel will decide itself
317	 * what to do with them.
318	 */
319	ret = cpuhp_smt_enable();
320	if (ret)
321		return ret;
322	smp_ops.play_dead = resume_play_dead;
323	ret = freeze_secondary_cpus(0);
324	smp_ops.play_dead = play_dead;
325	return ret;
326}
327#endif
328
329/*
330 * When bsp_check() is called in hibernate and suspend, cpu hotplug
331 * is disabled already. So it's unnecessary to handle race condition between
332 * cpumask query and cpu hotplug.
333 */
334static int bsp_check(void)
335{
336	if (cpumask_first(cpu_online_mask) != 0) {
337		pr_warn("CPU0 is offline.\n");
338		return -ENODEV;
339	}
340
341	return 0;
342}
343
344static int bsp_pm_callback(struct notifier_block *nb, unsigned long action,
345			   void *ptr)
346{
347	int ret = 0;
348
349	switch (action) {
350	case PM_SUSPEND_PREPARE:
351	case PM_HIBERNATION_PREPARE:
352		ret = bsp_check();
353		break;
354#ifdef CONFIG_DEBUG_HOTPLUG_CPU0
355	case PM_RESTORE_PREPARE:
356		/*
357		 * When system resumes from hibernation, online CPU0 because
358		 * 1. it's required for resume and
359		 * 2. the CPU was online before hibernation
360		 */
361		if (!cpu_online(0))
362			_debug_hotplug_cpu(0, 1);
363		break;
364	case PM_POST_RESTORE:
365		/*
366		 * When a resume really happens, this code won't be called.
367		 *
368		 * This code is called only when user space hibernation software
369		 * prepares for snapshot device during boot time. So we just
370		 * call _debug_hotplug_cpu() to restore to CPU0's state prior to
371		 * preparing the snapshot device.
372		 *
373		 * This works for normal boot case in our CPU0 hotplug debug
374		 * mode, i.e. CPU0 is offline and user mode hibernation
375		 * software initializes during boot time.
376		 *
377		 * If CPU0 is online and user application accesses snapshot
378		 * device after boot time, this will offline CPU0 and user may
379		 * see different CPU0 state before and after accessing
380		 * the snapshot device. But hopefully this is not a case when
381		 * user debugging CPU0 hotplug. Even if users hit this case,
382		 * they can easily online CPU0 back.
383		 *
384		 * To simplify this debug code, we only consider normal boot
385		 * case. Otherwise we need to remember CPU0's state and restore
386		 * to that state and resolve racy conditions etc.
387		 */
388		_debug_hotplug_cpu(0, 0);
389		break;
390#endif
391	default:
392		break;
393	}
394	return notifier_from_errno(ret);
395}
396
397static int __init bsp_pm_check_init(void)
398{
399	/*
400	 * Set this bsp_pm_callback as lower priority than
401	 * cpu_hotplug_pm_callback. So cpu_hotplug_pm_callback will be called
402	 * earlier to disable cpu hotplug before bsp online check.
403	 */
404	pm_notifier(bsp_pm_callback, -INT_MAX);
405	return 0;
406}
407
408core_initcall(bsp_pm_check_init);
409
410static int msr_build_context(const u32 *msr_id, const int num)
411{
412	struct saved_msrs *saved_msrs = &saved_context.saved_msrs;
413	struct saved_msr *msr_array;
414	int total_num;
415	int i, j;
416
417	total_num = saved_msrs->num + num;
418
419	msr_array = kmalloc_array(total_num, sizeof(struct saved_msr), GFP_KERNEL);
420	if (!msr_array) {
421		pr_err("x86/pm: Can not allocate memory to save/restore MSRs during suspend.\n");
422		return -ENOMEM;
423	}
424
425	if (saved_msrs->array) {
426		/*
427		 * Multiple callbacks can invoke this function, so copy any
428		 * MSR save requests from previous invocations.
429		 */
430		memcpy(msr_array, saved_msrs->array,
431		       sizeof(struct saved_msr) * saved_msrs->num);
432
433		kfree(saved_msrs->array);
434	}
435
436	for (i = saved_msrs->num, j = 0; i < total_num; i++, j++) {
437		u64 dummy;
438
439		msr_array[i].info.msr_no	= msr_id[j];
440		msr_array[i].valid		= !rdmsrl_safe(msr_id[j], &dummy);
441		msr_array[i].info.reg.q		= 0;
442	}
443	saved_msrs->num   = total_num;
444	saved_msrs->array = msr_array;
445
446	return 0;
447}
448
449/*
450 * The following sections are a quirk framework for problematic BIOSen:
451 * Sometimes MSRs are modified by the BIOSen after suspended to
452 * RAM, this might cause unexpected behavior after wakeup.
453 * Thus we save/restore these specified MSRs across suspend/resume
454 * in order to work around it.
455 *
456 * For any further problematic BIOSen/platforms,
457 * please add your own function similar to msr_initialize_bdw.
458 */
459static int msr_initialize_bdw(const struct dmi_system_id *d)
460{
461	/* Add any extra MSR ids into this array. */
462	u32 bdw_msr_id[] = { MSR_IA32_THERM_CONTROL };
463
464	pr_info("x86/pm: %s detected, MSR saving is needed during suspending.\n", d->ident);
465	return msr_build_context(bdw_msr_id, ARRAY_SIZE(bdw_msr_id));
466}
467
468static const struct dmi_system_id msr_save_dmi_table[] = {
469	{
470	 .callback = msr_initialize_bdw,
471	 .ident = "BROADWELL BDX_EP",
472	 .matches = {
473		DMI_MATCH(DMI_PRODUCT_NAME, "GRANTLEY"),
474		DMI_MATCH(DMI_PRODUCT_VERSION, "E63448-400"),
475		},
476	},
477	{}
478};
479
480static int msr_save_cpuid_features(const struct x86_cpu_id *c)
481{
482	u32 cpuid_msr_id[] = {
483		MSR_AMD64_CPUID_FN_1,
484	};
485
486	pr_info("x86/pm: family %#hx cpu detected, MSR saving is needed during suspending.\n",
487		c->family);
488
489	return msr_build_context(cpuid_msr_id, ARRAY_SIZE(cpuid_msr_id));
490}
491
492static const struct x86_cpu_id msr_save_cpu_table[] = {
493	X86_MATCH_VENDOR_FAM(AMD, 0x15, &msr_save_cpuid_features),
494	X86_MATCH_VENDOR_FAM(AMD, 0x16, &msr_save_cpuid_features),
 
 
 
 
 
 
 
 
 
 
 
 
495	{}
496};
497
498typedef int (*pm_cpu_match_t)(const struct x86_cpu_id *);
499static int pm_cpu_check(const struct x86_cpu_id *c)
500{
501	const struct x86_cpu_id *m;
502	int ret = 0;
503
504	m = x86_match_cpu(msr_save_cpu_table);
505	if (m) {
506		pm_cpu_match_t fn;
507
508		fn = (pm_cpu_match_t)m->driver_data;
509		ret = fn(m);
510	}
511
512	return ret;
513}
514
515static void pm_save_spec_msr(void)
516{
517	struct msr_enumeration {
518		u32 msr_no;
519		u32 feature;
520	} msr_enum[] = {
521		{ MSR_IA32_SPEC_CTRL,	 X86_FEATURE_MSR_SPEC_CTRL },
522		{ MSR_IA32_TSX_CTRL,	 X86_FEATURE_MSR_TSX_CTRL },
523		{ MSR_TSX_FORCE_ABORT,	 X86_FEATURE_TSX_FORCE_ABORT },
524		{ MSR_IA32_MCU_OPT_CTRL, X86_FEATURE_SRBDS_CTRL },
525		{ MSR_AMD64_LS_CFG,	 X86_FEATURE_LS_CFG_SSBD },
526		{ MSR_AMD64_DE_CFG,	 X86_FEATURE_LFENCE_RDTSC },
527	};
528	int i;
529
530	for (i = 0; i < ARRAY_SIZE(msr_enum); i++) {
531		if (boot_cpu_has(msr_enum[i].feature))
532			msr_build_context(&msr_enum[i].msr_no, 1);
533	}
534}
535
536static int pm_check_save_msr(void)
537{
538	dmi_check_system(msr_save_dmi_table);
539	pm_cpu_check(msr_save_cpu_table);
540	pm_save_spec_msr();
541
542	return 0;
543}
544
545device_initcall(pm_check_save_msr);
v5.4
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Suspend support specific for i386/x86-64.
  4 *
  5 * Copyright (c) 2007 Rafael J. Wysocki <rjw@sisk.pl>
  6 * Copyright (c) 2002 Pavel Machek <pavel@ucw.cz>
  7 * Copyright (c) 2001 Patrick Mochel <mochel@osdl.org>
  8 */
  9
 10#include <linux/suspend.h>
 11#include <linux/export.h>
 12#include <linux/smp.h>
 13#include <linux/perf_event.h>
 14#include <linux/tboot.h>
 15#include <linux/dmi.h>
 
 16
 17#include <asm/pgtable.h>
 18#include <asm/proto.h>
 19#include <asm/mtrr.h>
 20#include <asm/page.h>
 21#include <asm/mce.h>
 22#include <asm/suspend.h>
 23#include <asm/fpu/internal.h>
 24#include <asm/debugreg.h>
 25#include <asm/cpu.h>
 
 26#include <asm/mmu_context.h>
 27#include <asm/cpu_device_id.h>
 
 28
 29#ifdef CONFIG_X86_32
 30__visible unsigned long saved_context_ebx;
 31__visible unsigned long saved_context_esp, saved_context_ebp;
 32__visible unsigned long saved_context_esi, saved_context_edi;
 33__visible unsigned long saved_context_eflags;
 34#endif
 35struct saved_context saved_context;
 36
 37static void msr_save_context(struct saved_context *ctxt)
 38{
 39	struct saved_msr *msr = ctxt->saved_msrs.array;
 40	struct saved_msr *end = msr + ctxt->saved_msrs.num;
 41
 42	while (msr < end) {
 43		msr->valid = !rdmsrl_safe(msr->info.msr_no, &msr->info.reg.q);
 
 44		msr++;
 45	}
 46}
 47
 48static void msr_restore_context(struct saved_context *ctxt)
 49{
 50	struct saved_msr *msr = ctxt->saved_msrs.array;
 51	struct saved_msr *end = msr + ctxt->saved_msrs.num;
 52
 53	while (msr < end) {
 54		if (msr->valid)
 55			wrmsrl(msr->info.msr_no, msr->info.reg.q);
 56		msr++;
 57	}
 58}
 59
 60/**
 61 *	__save_processor_state - save CPU registers before creating a
 62 *		hibernation image and before restoring the memory state from it
 63 *	@ctxt - structure to store the registers contents in
 
 64 *
 65 *	NOTE: If there is a CPU register the modification of which by the
 66 *	boot kernel (ie. the kernel used for loading the hibernation image)
 67 *	might affect the operations of the restored target kernel (ie. the one
 68 *	saved in the hibernation image), then its contents must be saved by this
 69 *	function.  In other words, if kernel A is hibernated and different
 70 *	kernel B is used for loading the hibernation image into memory, the
 71 *	kernel A's __save_processor_state() function must save all registers
 72 *	needed by kernel A, so that it can operate correctly after the resume
 73 *	regardless of what kernel B does in the meantime.
 74 */
 75static void __save_processor_state(struct saved_context *ctxt)
 76{
 77#ifdef CONFIG_X86_32
 78	mtrr_save_fixed_ranges(NULL);
 79#endif
 80	kernel_fpu_begin();
 81
 82	/*
 83	 * descriptor tables
 84	 */
 85	store_idt(&ctxt->idt);
 86
 87	/*
 88	 * We save it here, but restore it only in the hibernate case.
 89	 * For ACPI S3 resume, this is loaded via 'early_gdt_desc' in 64-bit
 90	 * mode in "secondary_startup_64". In 32-bit mode it is done via
 91	 * 'pmode_gdt' in wakeup_start.
 92	 */
 93	ctxt->gdt_desc.size = GDT_SIZE - 1;
 94	ctxt->gdt_desc.address = (unsigned long)get_cpu_gdt_rw(smp_processor_id());
 95
 96	store_tr(ctxt->tr);
 97
 98	/* XMM0..XMM15 should be handled by kernel_fpu_begin(). */
 99	/*
100	 * segment registers
101	 */
102#ifdef CONFIG_X86_32_LAZY_GS
103	savesegment(gs, ctxt->gs);
104#endif
105#ifdef CONFIG_X86_64
106	savesegment(gs, ctxt->gs);
107	savesegment(fs, ctxt->fs);
108	savesegment(ds, ctxt->ds);
109	savesegment(es, ctxt->es);
110
111	rdmsrl(MSR_FS_BASE, ctxt->fs_base);
112	rdmsrl(MSR_GS_BASE, ctxt->kernelmode_gs_base);
113	rdmsrl(MSR_KERNEL_GS_BASE, ctxt->usermode_gs_base);
114	mtrr_save_fixed_ranges(NULL);
115
116	rdmsrl(MSR_EFER, ctxt->efer);
117#endif
118
119	/*
120	 * control registers
121	 */
122	ctxt->cr0 = read_cr0();
123	ctxt->cr2 = read_cr2();
124	ctxt->cr3 = __read_cr3();
125	ctxt->cr4 = __read_cr4();
126	ctxt->misc_enable_saved = !rdmsrl_safe(MSR_IA32_MISC_ENABLE,
127					       &ctxt->misc_enable);
128	msr_save_context(ctxt);
129}
130
131/* Needed by apm.c */
132void save_processor_state(void)
133{
134	__save_processor_state(&saved_context);
135	x86_platform.save_sched_clock_state();
136}
137#ifdef CONFIG_X86_32
138EXPORT_SYMBOL(save_processor_state);
139#endif
140
141static void do_fpu_end(void)
142{
143	/*
144	 * Restore FPU regs if necessary.
145	 */
146	kernel_fpu_end();
147}
148
149static void fix_processor_context(void)
150{
151	int cpu = smp_processor_id();
152#ifdef CONFIG_X86_64
153	struct desc_struct *desc = get_cpu_gdt_rw(cpu);
154	tss_desc tss;
155#endif
156
157	/*
158	 * We need to reload TR, which requires that we change the
159	 * GDT entry to indicate "available" first.
160	 *
161	 * XXX: This could probably all be replaced by a call to
162	 * force_reload_TR().
163	 */
164	set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
165
166#ifdef CONFIG_X86_64
167	memcpy(&tss, &desc[GDT_ENTRY_TSS], sizeof(tss_desc));
168	tss.type = 0x9; /* The available 64-bit TSS (see AMD vol 2, pg 91 */
169	write_gdt_entry(desc, GDT_ENTRY_TSS, &tss, DESC_TSS);
170
171	syscall_init();				/* This sets MSR_*STAR and related */
172#else
173	if (boot_cpu_has(X86_FEATURE_SEP))
174		enable_sep_cpu();
175#endif
176	load_TR_desc();				/* This does ltr */
177	load_mm_ldt(current->active_mm);	/* This does lldt */
178	initialize_tlbstate_and_flush();
179
180	fpu__resume_cpu();
181
182	/* The processor is back on the direct GDT, load back the fixmap */
183	load_fixmap_gdt(cpu);
184}
185
186/**
187 * __restore_processor_state - restore the contents of CPU registers saved
188 *                             by __save_processor_state()
189 * @ctxt - structure to load the registers contents from
190 *
191 * The asm code that gets us here will have restored a usable GDT, although
192 * it will be pointing to the wrong alias.
193 */
194static void notrace __restore_processor_state(struct saved_context *ctxt)
195{
 
 
196	if (ctxt->misc_enable_saved)
197		wrmsrl(MSR_IA32_MISC_ENABLE, ctxt->misc_enable);
198	/*
199	 * control registers
200	 */
201	/* cr4 was introduced in the Pentium CPU */
202#ifdef CONFIG_X86_32
203	if (ctxt->cr4)
204		__write_cr4(ctxt->cr4);
205#else
206/* CONFIG X86_64 */
207	wrmsrl(MSR_EFER, ctxt->efer);
208	__write_cr4(ctxt->cr4);
209#endif
210	write_cr3(ctxt->cr3);
211	write_cr2(ctxt->cr2);
212	write_cr0(ctxt->cr0);
213
214	/* Restore the IDT. */
215	load_idt(&ctxt->idt);
216
217	/*
218	 * Just in case the asm code got us here with the SS, DS, or ES
219	 * out of sync with the GDT, update them.
220	 */
221	loadsegment(ss, __KERNEL_DS);
222	loadsegment(ds, __USER_DS);
223	loadsegment(es, __USER_DS);
224
225	/*
226	 * Restore percpu access.  Percpu access can happen in exception
227	 * handlers or in complicated helpers like load_gs_index().
228	 */
229#ifdef CONFIG_X86_64
230	wrmsrl(MSR_GS_BASE, ctxt->kernelmode_gs_base);
231#else
232	loadsegment(fs, __KERNEL_PERCPU);
233	loadsegment(gs, __KERNEL_STACK_CANARY);
234#endif
235
236	/* Restore the TSS, RO GDT, LDT, and usermode-relevant MSRs. */
237	fix_processor_context();
238
239	/*
240	 * Now that we have descriptor tables fully restored and working
241	 * exception handling, restore the usermode segments.
242	 */
243#ifdef CONFIG_X86_64
244	loadsegment(ds, ctxt->es);
245	loadsegment(es, ctxt->es);
246	loadsegment(fs, ctxt->fs);
247	load_gs_index(ctxt->gs);
248
249	/*
250	 * Restore FSBASE and GSBASE after restoring the selectors, since
251	 * restoring the selectors clobbers the bases.  Keep in mind
252	 * that MSR_KERNEL_GS_BASE is horribly misnamed.
253	 */
254	wrmsrl(MSR_FS_BASE, ctxt->fs_base);
255	wrmsrl(MSR_KERNEL_GS_BASE, ctxt->usermode_gs_base);
256#elif defined(CONFIG_X86_32_LAZY_GS)
257	loadsegment(gs, ctxt->gs);
258#endif
259
260	do_fpu_end();
261	tsc_verify_tsc_adjust(true);
262	x86_platform.restore_sched_clock_state();
263	mtrr_bp_restore();
264	perf_restore_debug_store();
 
 
 
 
 
 
 
 
 
 
 
265	msr_restore_context(ctxt);
266}
267
268/* Needed by apm.c */
269void notrace restore_processor_state(void)
270{
271	__restore_processor_state(&saved_context);
272}
273#ifdef CONFIG_X86_32
274EXPORT_SYMBOL(restore_processor_state);
275#endif
276
277#if defined(CONFIG_HIBERNATION) && defined(CONFIG_HOTPLUG_CPU)
278static void resume_play_dead(void)
279{
280	play_dead_common();
281	tboot_shutdown(TB_SHUTDOWN_WFS);
282	hlt_play_dead();
283}
284
285int hibernate_resume_nonboot_cpu_disable(void)
286{
287	void (*play_dead)(void) = smp_ops.play_dead;
288	int ret;
289
290	/*
291	 * Ensure that MONITOR/MWAIT will not be used in the "play dead" loop
292	 * during hibernate image restoration, because it is likely that the
293	 * monitored address will be actually written to at that time and then
294	 * the "dead" CPU will attempt to execute instructions again, but the
295	 * address in its instruction pointer may not be possible to resolve
296	 * any more at that point (the page tables used by it previously may
297	 * have been overwritten by hibernate image data).
298	 *
299	 * First, make sure that we wake up all the potentially disabled SMT
300	 * threads which have been initially brought up and then put into
301	 * mwait/cpuidle sleep.
302	 * Those will be put to proper (not interfering with hibernation
303	 * resume) sleep afterwards, and the resumed kernel will decide itself
304	 * what to do with them.
305	 */
306	ret = cpuhp_smt_enable();
307	if (ret)
308		return ret;
309	smp_ops.play_dead = resume_play_dead;
310	ret = disable_nonboot_cpus();
311	smp_ops.play_dead = play_dead;
312	return ret;
313}
314#endif
315
316/*
317 * When bsp_check() is called in hibernate and suspend, cpu hotplug
318 * is disabled already. So it's unnessary to handle race condition between
319 * cpumask query and cpu hotplug.
320 */
321static int bsp_check(void)
322{
323	if (cpumask_first(cpu_online_mask) != 0) {
324		pr_warn("CPU0 is offline.\n");
325		return -ENODEV;
326	}
327
328	return 0;
329}
330
331static int bsp_pm_callback(struct notifier_block *nb, unsigned long action,
332			   void *ptr)
333{
334	int ret = 0;
335
336	switch (action) {
337	case PM_SUSPEND_PREPARE:
338	case PM_HIBERNATION_PREPARE:
339		ret = bsp_check();
340		break;
341#ifdef CONFIG_DEBUG_HOTPLUG_CPU0
342	case PM_RESTORE_PREPARE:
343		/*
344		 * When system resumes from hibernation, online CPU0 because
345		 * 1. it's required for resume and
346		 * 2. the CPU was online before hibernation
347		 */
348		if (!cpu_online(0))
349			_debug_hotplug_cpu(0, 1);
350		break;
351	case PM_POST_RESTORE:
352		/*
353		 * When a resume really happens, this code won't be called.
354		 *
355		 * This code is called only when user space hibernation software
356		 * prepares for snapshot device during boot time. So we just
357		 * call _debug_hotplug_cpu() to restore to CPU0's state prior to
358		 * preparing the snapshot device.
359		 *
360		 * This works for normal boot case in our CPU0 hotplug debug
361		 * mode, i.e. CPU0 is offline and user mode hibernation
362		 * software initializes during boot time.
363		 *
364		 * If CPU0 is online and user application accesses snapshot
365		 * device after boot time, this will offline CPU0 and user may
366		 * see different CPU0 state before and after accessing
367		 * the snapshot device. But hopefully this is not a case when
368		 * user debugging CPU0 hotplug. Even if users hit this case,
369		 * they can easily online CPU0 back.
370		 *
371		 * To simplify this debug code, we only consider normal boot
372		 * case. Otherwise we need to remember CPU0's state and restore
373		 * to that state and resolve racy conditions etc.
374		 */
375		_debug_hotplug_cpu(0, 0);
376		break;
377#endif
378	default:
379		break;
380	}
381	return notifier_from_errno(ret);
382}
383
384static int __init bsp_pm_check_init(void)
385{
386	/*
387	 * Set this bsp_pm_callback as lower priority than
388	 * cpu_hotplug_pm_callback. So cpu_hotplug_pm_callback will be called
389	 * earlier to disable cpu hotplug before bsp online check.
390	 */
391	pm_notifier(bsp_pm_callback, -INT_MAX);
392	return 0;
393}
394
395core_initcall(bsp_pm_check_init);
396
397static int msr_build_context(const u32 *msr_id, const int num)
398{
399	struct saved_msrs *saved_msrs = &saved_context.saved_msrs;
400	struct saved_msr *msr_array;
401	int total_num;
402	int i, j;
403
404	total_num = saved_msrs->num + num;
405
406	msr_array = kmalloc_array(total_num, sizeof(struct saved_msr), GFP_KERNEL);
407	if (!msr_array) {
408		pr_err("x86/pm: Can not allocate memory to save/restore MSRs during suspend.\n");
409		return -ENOMEM;
410	}
411
412	if (saved_msrs->array) {
413		/*
414		 * Multiple callbacks can invoke this function, so copy any
415		 * MSR save requests from previous invocations.
416		 */
417		memcpy(msr_array, saved_msrs->array,
418		       sizeof(struct saved_msr) * saved_msrs->num);
419
420		kfree(saved_msrs->array);
421	}
422
423	for (i = saved_msrs->num, j = 0; i < total_num; i++, j++) {
 
 
424		msr_array[i].info.msr_no	= msr_id[j];
425		msr_array[i].valid		= false;
426		msr_array[i].info.reg.q		= 0;
427	}
428	saved_msrs->num   = total_num;
429	saved_msrs->array = msr_array;
430
431	return 0;
432}
433
434/*
435 * The following sections are a quirk framework for problematic BIOSen:
436 * Sometimes MSRs are modified by the BIOSen after suspended to
437 * RAM, this might cause unexpected behavior after wakeup.
438 * Thus we save/restore these specified MSRs across suspend/resume
439 * in order to work around it.
440 *
441 * For any further problematic BIOSen/platforms,
442 * please add your own function similar to msr_initialize_bdw.
443 */
444static int msr_initialize_bdw(const struct dmi_system_id *d)
445{
446	/* Add any extra MSR ids into this array. */
447	u32 bdw_msr_id[] = { MSR_IA32_THERM_CONTROL };
448
449	pr_info("x86/pm: %s detected, MSR saving is needed during suspending.\n", d->ident);
450	return msr_build_context(bdw_msr_id, ARRAY_SIZE(bdw_msr_id));
451}
452
453static const struct dmi_system_id msr_save_dmi_table[] = {
454	{
455	 .callback = msr_initialize_bdw,
456	 .ident = "BROADWELL BDX_EP",
457	 .matches = {
458		DMI_MATCH(DMI_PRODUCT_NAME, "GRANTLEY"),
459		DMI_MATCH(DMI_PRODUCT_VERSION, "E63448-400"),
460		},
461	},
462	{}
463};
464
465static int msr_save_cpuid_features(const struct x86_cpu_id *c)
466{
467	u32 cpuid_msr_id[] = {
468		MSR_AMD64_CPUID_FN_1,
469	};
470
471	pr_info("x86/pm: family %#hx cpu detected, MSR saving is needed during suspending.\n",
472		c->family);
473
474	return msr_build_context(cpuid_msr_id, ARRAY_SIZE(cpuid_msr_id));
475}
476
477static const struct x86_cpu_id msr_save_cpu_table[] = {
478	{
479		.vendor = X86_VENDOR_AMD,
480		.family = 0x15,
481		.model = X86_MODEL_ANY,
482		.feature = X86_FEATURE_ANY,
483		.driver_data = (kernel_ulong_t)msr_save_cpuid_features,
484	},
485	{
486		.vendor = X86_VENDOR_AMD,
487		.family = 0x16,
488		.model = X86_MODEL_ANY,
489		.feature = X86_FEATURE_ANY,
490		.driver_data = (kernel_ulong_t)msr_save_cpuid_features,
491	},
492	{}
493};
494
495typedef int (*pm_cpu_match_t)(const struct x86_cpu_id *);
496static int pm_cpu_check(const struct x86_cpu_id *c)
497{
498	const struct x86_cpu_id *m;
499	int ret = 0;
500
501	m = x86_match_cpu(msr_save_cpu_table);
502	if (m) {
503		pm_cpu_match_t fn;
504
505		fn = (pm_cpu_match_t)m->driver_data;
506		ret = fn(m);
507	}
508
509	return ret;
510}
511
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
512static int pm_check_save_msr(void)
513{
514	dmi_check_system(msr_save_dmi_table);
515	pm_cpu_check(msr_save_cpu_table);
 
516
517	return 0;
518}
519
520device_initcall(pm_check_save_msr);