xref: /linux/arch/x86/power/cpu.c (revision 861e10be08c69808065d755d3e3cab5d520a2d32)
1 /*
2  * Suspend support specific for i386/x86-64.
3  *
4  * Distribute under GPLv2
5  *
6  * Copyright (c) 2007 Rafael J. Wysocki <rjw@sisk.pl>
7  * Copyright (c) 2002 Pavel Machek <pavel@ucw.cz>
8  * Copyright (c) 2001 Patrick Mochel <mochel@osdl.org>
9  */
10 
11 #include <linux/suspend.h>
12 #include <linux/export.h>
13 #include <linux/smp.h>
14 
15 #include <asm/pgtable.h>
16 #include <asm/proto.h>
17 #include <asm/mtrr.h>
18 #include <asm/page.h>
19 #include <asm/mce.h>
20 #include <asm/xcr.h>
21 #include <asm/suspend.h>
22 #include <asm/debugreg.h>
23 #include <asm/fpu-internal.h> /* pcntxt_mask */
24 #include <asm/cpu.h>
25 
26 #ifdef CONFIG_X86_32
27 static struct saved_context saved_context;
28 
29 unsigned long saved_context_ebx;
30 unsigned long saved_context_esp, saved_context_ebp;
31 unsigned long saved_context_esi, saved_context_edi;
32 unsigned long saved_context_eflags;
33 #else
34 /* CONFIG_X86_64 */
35 struct saved_context saved_context;
36 #endif
37 
38 /**
39  *	__save_processor_state - save CPU registers before creating a
40  *		hibernation image and before restoring the memory state from it
41  *	@ctxt - structure to store the registers contents in
42  *
43  *	NOTE: If there is a CPU register the modification of which by the
44  *	boot kernel (ie. the kernel used for loading the hibernation image)
45  *	might affect the operations of the restored target kernel (ie. the one
46  *	saved in the hibernation image), then its contents must be saved by this
47  *	function.  In other words, if kernel A is hibernated and different
48  *	kernel B is used for loading the hibernation image into memory, the
49  *	kernel A's __save_processor_state() function must save all registers
50  *	needed by kernel A, so that it can operate correctly after the resume
51  *	regardless of what kernel B does in the meantime.
52  */
53 static void __save_processor_state(struct saved_context *ctxt)
54 {
55 #ifdef CONFIG_X86_32
56 	mtrr_save_fixed_ranges(NULL);
57 #endif
58 	kernel_fpu_begin();
59 
60 	/*
61 	 * descriptor tables
62 	 */
63 #ifdef CONFIG_X86_32
64 	store_gdt(&ctxt->gdt);
65 	store_idt(&ctxt->idt);
66 #else
67 /* CONFIG_X86_64 */
68 	store_gdt((struct desc_ptr *)&ctxt->gdt_limit);
69 	store_idt((struct desc_ptr *)&ctxt->idt_limit);
70 #endif
71 	store_tr(ctxt->tr);
72 
73 	/* XMM0..XMM15 should be handled by kernel_fpu_begin(). */
74 	/*
75 	 * segment registers
76 	 */
77 #ifdef CONFIG_X86_32
78 	savesegment(es, ctxt->es);
79 	savesegment(fs, ctxt->fs);
80 	savesegment(gs, ctxt->gs);
81 	savesegment(ss, ctxt->ss);
82 #else
83 /* CONFIG_X86_64 */
84 	asm volatile ("movw %%ds, %0" : "=m" (ctxt->ds));
85 	asm volatile ("movw %%es, %0" : "=m" (ctxt->es));
86 	asm volatile ("movw %%fs, %0" : "=m" (ctxt->fs));
87 	asm volatile ("movw %%gs, %0" : "=m" (ctxt->gs));
88 	asm volatile ("movw %%ss, %0" : "=m" (ctxt->ss));
89 
90 	rdmsrl(MSR_FS_BASE, ctxt->fs_base);
91 	rdmsrl(MSR_GS_BASE, ctxt->gs_base);
92 	rdmsrl(MSR_KERNEL_GS_BASE, ctxt->gs_kernel_base);
93 	mtrr_save_fixed_ranges(NULL);
94 
95 	rdmsrl(MSR_EFER, ctxt->efer);
96 #endif
97 
98 	/*
99 	 * control registers
100 	 */
101 	ctxt->cr0 = read_cr0();
102 	ctxt->cr2 = read_cr2();
103 	ctxt->cr3 = read_cr3();
104 #ifdef CONFIG_X86_32
105 	ctxt->cr4 = read_cr4_safe();
106 #else
107 /* CONFIG_X86_64 */
108 	ctxt->cr4 = read_cr4();
109 	ctxt->cr8 = read_cr8();
110 #endif
111 	ctxt->misc_enable_saved = !rdmsrl_safe(MSR_IA32_MISC_ENABLE,
112 					       &ctxt->misc_enable);
113 }
114 
115 /* Needed by apm.c */
116 void save_processor_state(void)
117 {
118 	__save_processor_state(&saved_context);
119 	x86_platform.save_sched_clock_state();
120 }
121 #ifdef CONFIG_X86_32
122 EXPORT_SYMBOL(save_processor_state);
123 #endif
124 
125 static void do_fpu_end(void)
126 {
127 	/*
128 	 * Restore FPU regs if necessary.
129 	 */
130 	kernel_fpu_end();
131 }
132 
133 static void fix_processor_context(void)
134 {
135 	int cpu = smp_processor_id();
136 	struct tss_struct *t = &per_cpu(init_tss, cpu);
137 
138 	set_tss_desc(cpu, t);	/*
139 				 * This just modifies memory; should not be
140 				 * necessary. But... This is necessary, because
141 				 * 386 hardware has concept of busy TSS or some
142 				 * similar stupidity.
143 				 */
144 
145 #ifdef CONFIG_X86_64
146 	get_cpu_gdt_table(cpu)[GDT_ENTRY_TSS].type = 9;
147 
148 	syscall_init();				/* This sets MSR_*STAR and related */
149 #endif
150 	load_TR_desc();				/* This does ltr */
151 	load_LDT(&current->active_mm->context);	/* This does lldt */
152 }
153 
154 /**
155  *	__restore_processor_state - restore the contents of CPU registers saved
156  *		by __save_processor_state()
157  *	@ctxt - structure to load the registers contents from
158  */
159 static void __restore_processor_state(struct saved_context *ctxt)
160 {
161 	if (ctxt->misc_enable_saved)
162 		wrmsrl(MSR_IA32_MISC_ENABLE, ctxt->misc_enable);
163 	/*
164 	 * control registers
165 	 */
166 	/* cr4 was introduced in the Pentium CPU */
167 #ifdef CONFIG_X86_32
168 	if (ctxt->cr4)
169 		write_cr4(ctxt->cr4);
170 #else
171 /* CONFIG X86_64 */
172 	wrmsrl(MSR_EFER, ctxt->efer);
173 	write_cr8(ctxt->cr8);
174 	write_cr4(ctxt->cr4);
175 #endif
176 	write_cr3(ctxt->cr3);
177 	write_cr2(ctxt->cr2);
178 	write_cr0(ctxt->cr0);
179 
180 	/*
181 	 * now restore the descriptor tables to their proper values
182 	 * ltr is done i fix_processor_context().
183 	 */
184 #ifdef CONFIG_X86_32
185 	load_gdt(&ctxt->gdt);
186 	load_idt(&ctxt->idt);
187 #else
188 /* CONFIG_X86_64 */
189 	load_gdt((const struct desc_ptr *)&ctxt->gdt_limit);
190 	load_idt((const struct desc_ptr *)&ctxt->idt_limit);
191 #endif
192 
193 	/*
194 	 * segment registers
195 	 */
196 #ifdef CONFIG_X86_32
197 	loadsegment(es, ctxt->es);
198 	loadsegment(fs, ctxt->fs);
199 	loadsegment(gs, ctxt->gs);
200 	loadsegment(ss, ctxt->ss);
201 
202 	/*
203 	 * sysenter MSRs
204 	 */
205 	if (boot_cpu_has(X86_FEATURE_SEP))
206 		enable_sep_cpu();
207 #else
208 /* CONFIG_X86_64 */
209 	asm volatile ("movw %0, %%ds" :: "r" (ctxt->ds));
210 	asm volatile ("movw %0, %%es" :: "r" (ctxt->es));
211 	asm volatile ("movw %0, %%fs" :: "r" (ctxt->fs));
212 	load_gs_index(ctxt->gs);
213 	asm volatile ("movw %0, %%ss" :: "r" (ctxt->ss));
214 
215 	wrmsrl(MSR_FS_BASE, ctxt->fs_base);
216 	wrmsrl(MSR_GS_BASE, ctxt->gs_base);
217 	wrmsrl(MSR_KERNEL_GS_BASE, ctxt->gs_kernel_base);
218 #endif
219 
220 	/*
221 	 * restore XCR0 for xsave capable cpu's.
222 	 */
223 	if (cpu_has_xsave)
224 		xsetbv(XCR_XFEATURE_ENABLED_MASK, pcntxt_mask);
225 
226 	fix_processor_context();
227 
228 	do_fpu_end();
229 	x86_platform.restore_sched_clock_state();
230 	mtrr_bp_restore();
231 }
232 
233 /* Needed by apm.c */
234 void restore_processor_state(void)
235 {
236 	__restore_processor_state(&saved_context);
237 }
238 #ifdef CONFIG_X86_32
239 EXPORT_SYMBOL(restore_processor_state);
240 #endif
241 
242 /*
243  * When bsp_check() is called in hibernate and suspend, cpu hotplug
244  * is disabled already. So it's unnessary to handle race condition between
245  * cpumask query and cpu hotplug.
246  */
247 static int bsp_check(void)
248 {
249 	if (cpumask_first(cpu_online_mask) != 0) {
250 		pr_warn("CPU0 is offline.\n");
251 		return -ENODEV;
252 	}
253 
254 	return 0;
255 }
256 
257 static int bsp_pm_callback(struct notifier_block *nb, unsigned long action,
258 			   void *ptr)
259 {
260 	int ret = 0;
261 
262 	switch (action) {
263 	case PM_SUSPEND_PREPARE:
264 	case PM_HIBERNATION_PREPARE:
265 		ret = bsp_check();
266 		break;
267 #ifdef CONFIG_DEBUG_HOTPLUG_CPU0
268 	case PM_RESTORE_PREPARE:
269 		/*
270 		 * When system resumes from hibernation, online CPU0 because
271 		 * 1. it's required for resume and
272 		 * 2. the CPU was online before hibernation
273 		 */
274 		if (!cpu_online(0))
275 			_debug_hotplug_cpu(0, 1);
276 		break;
277 	case PM_POST_RESTORE:
278 		/*
279 		 * When a resume really happens, this code won't be called.
280 		 *
281 		 * This code is called only when user space hibernation software
282 		 * prepares for snapshot device during boot time. So we just
283 		 * call _debug_hotplug_cpu() to restore to CPU0's state prior to
284 		 * preparing the snapshot device.
285 		 *
286 		 * This works for normal boot case in our CPU0 hotplug debug
287 		 * mode, i.e. CPU0 is offline and user mode hibernation
288 		 * software initializes during boot time.
289 		 *
290 		 * If CPU0 is online and user application accesses snapshot
291 		 * device after boot time, this will offline CPU0 and user may
292 		 * see different CPU0 state before and after accessing
293 		 * the snapshot device. But hopefully this is not a case when
294 		 * user debugging CPU0 hotplug. Even if users hit this case,
295 		 * they can easily online CPU0 back.
296 		 *
297 		 * To simplify this debug code, we only consider normal boot
298 		 * case. Otherwise we need to remember CPU0's state and restore
299 		 * to that state and resolve racy conditions etc.
300 		 */
301 		_debug_hotplug_cpu(0, 0);
302 		break;
303 #endif
304 	default:
305 		break;
306 	}
307 	return notifier_from_errno(ret);
308 }
309 
310 static int __init bsp_pm_check_init(void)
311 {
312 	/*
313 	 * Set this bsp_pm_callback as lower priority than
314 	 * cpu_hotplug_pm_callback. So cpu_hotplug_pm_callback will be called
315 	 * earlier to disable cpu hotplug before bsp online check.
316 	 */
317 	pm_notifier(bsp_pm_callback, -INT_MAX);
318 	return 0;
319 }
320 
321 core_initcall(bsp_pm_check_init);
322