xref: /linux/arch/x86/power/cpu.c (revision c75c5ab575af7db707689cdbb5a5c458e9a034bb)
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 #include <linux/perf_event.h>
15 
16 #include <asm/pgtable.h>
17 #include <asm/proto.h>
18 #include <asm/mtrr.h>
19 #include <asm/page.h>
20 #include <asm/mce.h>
21 #include <asm/xcr.h>
22 #include <asm/suspend.h>
23 #include <asm/debugreg.h>
24 #include <asm/fpu-internal.h> /* pcntxt_mask */
25 #include <asm/cpu.h>
26 
27 #ifdef CONFIG_X86_32
28 static struct saved_context saved_context;
29 
30 unsigned long saved_context_ebx;
31 unsigned long saved_context_esp, saved_context_ebp;
32 unsigned long saved_context_esi, saved_context_edi;
33 unsigned long saved_context_eflags;
34 #else
35 /* CONFIG_X86_64 */
36 struct saved_context saved_context;
37 #endif
38 
39 /**
40  *	__save_processor_state - save CPU registers before creating a
41  *		hibernation image and before restoring the memory state from it
42  *	@ctxt - structure to store the registers contents in
43  *
44  *	NOTE: If there is a CPU register the modification of which by the
45  *	boot kernel (ie. the kernel used for loading the hibernation image)
46  *	might affect the operations of the restored target kernel (ie. the one
47  *	saved in the hibernation image), then its contents must be saved by this
48  *	function.  In other words, if kernel A is hibernated and different
49  *	kernel B is used for loading the hibernation image into memory, the
50  *	kernel A's __save_processor_state() function must save all registers
51  *	needed by kernel A, so that it can operate correctly after the resume
52  *	regardless of what kernel B does in the meantime.
53  */
54 static void __save_processor_state(struct saved_context *ctxt)
55 {
56 #ifdef CONFIG_X86_32
57 	mtrr_save_fixed_ranges(NULL);
58 #endif
59 	kernel_fpu_begin();
60 
61 	/*
62 	 * descriptor tables
63 	 */
64 #ifdef CONFIG_X86_32
65 	store_gdt(&ctxt->gdt);
66 	store_idt(&ctxt->idt);
67 #else
68 /* CONFIG_X86_64 */
69 	store_gdt((struct desc_ptr *)&ctxt->gdt_limit);
70 	store_idt((struct desc_ptr *)&ctxt->idt_limit);
71 #endif
72 	store_tr(ctxt->tr);
73 
74 	/* XMM0..XMM15 should be handled by kernel_fpu_begin(). */
75 	/*
76 	 * segment registers
77 	 */
78 #ifdef CONFIG_X86_32
79 	savesegment(es, ctxt->es);
80 	savesegment(fs, ctxt->fs);
81 	savesegment(gs, ctxt->gs);
82 	savesegment(ss, ctxt->ss);
83 #else
84 /* CONFIG_X86_64 */
85 	asm volatile ("movw %%ds, %0" : "=m" (ctxt->ds));
86 	asm volatile ("movw %%es, %0" : "=m" (ctxt->es));
87 	asm volatile ("movw %%fs, %0" : "=m" (ctxt->fs));
88 	asm volatile ("movw %%gs, %0" : "=m" (ctxt->gs));
89 	asm volatile ("movw %%ss, %0" : "=m" (ctxt->ss));
90 
91 	rdmsrl(MSR_FS_BASE, ctxt->fs_base);
92 	rdmsrl(MSR_GS_BASE, ctxt->gs_base);
93 	rdmsrl(MSR_KERNEL_GS_BASE, ctxt->gs_kernel_base);
94 	mtrr_save_fixed_ranges(NULL);
95 
96 	rdmsrl(MSR_EFER, ctxt->efer);
97 #endif
98 
99 	/*
100 	 * control registers
101 	 */
102 	ctxt->cr0 = read_cr0();
103 	ctxt->cr2 = read_cr2();
104 	ctxt->cr3 = read_cr3();
105 #ifdef CONFIG_X86_32
106 	ctxt->cr4 = read_cr4_safe();
107 #else
108 /* CONFIG_X86_64 */
109 	ctxt->cr4 = read_cr4();
110 	ctxt->cr8 = read_cr8();
111 #endif
112 	ctxt->misc_enable_saved = !rdmsrl_safe(MSR_IA32_MISC_ENABLE,
113 					       &ctxt->misc_enable);
114 }
115 
116 /* Needed by apm.c */
117 void save_processor_state(void)
118 {
119 	__save_processor_state(&saved_context);
120 	x86_platform.save_sched_clock_state();
121 }
122 #ifdef CONFIG_X86_32
123 EXPORT_SYMBOL(save_processor_state);
124 #endif
125 
126 static void do_fpu_end(void)
127 {
128 	/*
129 	 * Restore FPU regs if necessary.
130 	 */
131 	kernel_fpu_end();
132 }
133 
134 static void fix_processor_context(void)
135 {
136 	int cpu = smp_processor_id();
137 	struct tss_struct *t = &per_cpu(init_tss, cpu);
138 
139 	set_tss_desc(cpu, t);	/*
140 				 * This just modifies memory; should not be
141 				 * necessary. But... This is necessary, because
142 				 * 386 hardware has concept of busy TSS or some
143 				 * similar stupidity.
144 				 */
145 
146 #ifdef CONFIG_X86_64
147 	get_cpu_gdt_table(cpu)[GDT_ENTRY_TSS].type = 9;
148 
149 	syscall_init();				/* This sets MSR_*STAR and related */
150 #endif
151 	load_TR_desc();				/* This does ltr */
152 	load_LDT(&current->active_mm->context);	/* This does lldt */
153 }
154 
155 /**
156  *	__restore_processor_state - restore the contents of CPU registers saved
157  *		by __save_processor_state()
158  *	@ctxt - structure to load the registers contents from
159  */
160 static void __restore_processor_state(struct saved_context *ctxt)
161 {
162 	if (ctxt->misc_enable_saved)
163 		wrmsrl(MSR_IA32_MISC_ENABLE, ctxt->misc_enable);
164 	/*
165 	 * control registers
166 	 */
167 	/* cr4 was introduced in the Pentium CPU */
168 #ifdef CONFIG_X86_32
169 	if (ctxt->cr4)
170 		write_cr4(ctxt->cr4);
171 #else
172 /* CONFIG X86_64 */
173 	wrmsrl(MSR_EFER, ctxt->efer);
174 	write_cr8(ctxt->cr8);
175 	write_cr4(ctxt->cr4);
176 #endif
177 	write_cr3(ctxt->cr3);
178 	write_cr2(ctxt->cr2);
179 	write_cr0(ctxt->cr0);
180 
181 	/*
182 	 * now restore the descriptor tables to their proper values
183 	 * ltr is done i fix_processor_context().
184 	 */
185 #ifdef CONFIG_X86_32
186 	load_gdt(&ctxt->gdt);
187 	load_idt(&ctxt->idt);
188 #else
189 /* CONFIG_X86_64 */
190 	load_gdt((const struct desc_ptr *)&ctxt->gdt_limit);
191 	load_idt((const struct desc_ptr *)&ctxt->idt_limit);
192 #endif
193 
194 	/*
195 	 * segment registers
196 	 */
197 #ifdef CONFIG_X86_32
198 	loadsegment(es, ctxt->es);
199 	loadsegment(fs, ctxt->fs);
200 	loadsegment(gs, ctxt->gs);
201 	loadsegment(ss, ctxt->ss);
202 
203 	/*
204 	 * sysenter MSRs
205 	 */
206 	if (boot_cpu_has(X86_FEATURE_SEP))
207 		enable_sep_cpu();
208 #else
209 /* CONFIG_X86_64 */
210 	asm volatile ("movw %0, %%ds" :: "r" (ctxt->ds));
211 	asm volatile ("movw %0, %%es" :: "r" (ctxt->es));
212 	asm volatile ("movw %0, %%fs" :: "r" (ctxt->fs));
213 	load_gs_index(ctxt->gs);
214 	asm volatile ("movw %0, %%ss" :: "r" (ctxt->ss));
215 
216 	wrmsrl(MSR_FS_BASE, ctxt->fs_base);
217 	wrmsrl(MSR_GS_BASE, ctxt->gs_base);
218 	wrmsrl(MSR_KERNEL_GS_BASE, ctxt->gs_kernel_base);
219 #endif
220 
221 	/*
222 	 * restore XCR0 for xsave capable cpu's.
223 	 */
224 	if (cpu_has_xsave)
225 		xsetbv(XCR_XFEATURE_ENABLED_MASK, pcntxt_mask);
226 
227 	fix_processor_context();
228 
229 	do_fpu_end();
230 	x86_platform.restore_sched_clock_state();
231 	mtrr_bp_restore();
232 	perf_restore_debug_store();
233 }
234 
235 /* Needed by apm.c */
236 void restore_processor_state(void)
237 {
238 	__restore_processor_state(&saved_context);
239 }
240 #ifdef CONFIG_X86_32
241 EXPORT_SYMBOL(restore_processor_state);
242 #endif
243 
244 /*
245  * When bsp_check() is called in hibernate and suspend, cpu hotplug
246  * is disabled already. So it's unnessary to handle race condition between
247  * cpumask query and cpu hotplug.
248  */
249 static int bsp_check(void)
250 {
251 	if (cpumask_first(cpu_online_mask) != 0) {
252 		pr_warn("CPU0 is offline.\n");
253 		return -ENODEV;
254 	}
255 
256 	return 0;
257 }
258 
259 static int bsp_pm_callback(struct notifier_block *nb, unsigned long action,
260 			   void *ptr)
261 {
262 	int ret = 0;
263 
264 	switch (action) {
265 	case PM_SUSPEND_PREPARE:
266 	case PM_HIBERNATION_PREPARE:
267 		ret = bsp_check();
268 		break;
269 #ifdef CONFIG_DEBUG_HOTPLUG_CPU0
270 	case PM_RESTORE_PREPARE:
271 		/*
272 		 * When system resumes from hibernation, online CPU0 because
273 		 * 1. it's required for resume and
274 		 * 2. the CPU was online before hibernation
275 		 */
276 		if (!cpu_online(0))
277 			_debug_hotplug_cpu(0, 1);
278 		break;
279 	case PM_POST_RESTORE:
280 		/*
281 		 * When a resume really happens, this code won't be called.
282 		 *
283 		 * This code is called only when user space hibernation software
284 		 * prepares for snapshot device during boot time. So we just
285 		 * call _debug_hotplug_cpu() to restore to CPU0's state prior to
286 		 * preparing the snapshot device.
287 		 *
288 		 * This works for normal boot case in our CPU0 hotplug debug
289 		 * mode, i.e. CPU0 is offline and user mode hibernation
290 		 * software initializes during boot time.
291 		 *
292 		 * If CPU0 is online and user application accesses snapshot
293 		 * device after boot time, this will offline CPU0 and user may
294 		 * see different CPU0 state before and after accessing
295 		 * the snapshot device. But hopefully this is not a case when
296 		 * user debugging CPU0 hotplug. Even if users hit this case,
297 		 * they can easily online CPU0 back.
298 		 *
299 		 * To simplify this debug code, we only consider normal boot
300 		 * case. Otherwise we need to remember CPU0's state and restore
301 		 * to that state and resolve racy conditions etc.
302 		 */
303 		_debug_hotplug_cpu(0, 0);
304 		break;
305 #endif
306 	default:
307 		break;
308 	}
309 	return notifier_from_errno(ret);
310 }
311 
312 static int __init bsp_pm_check_init(void)
313 {
314 	/*
315 	 * Set this bsp_pm_callback as lower priority than
316 	 * cpu_hotplug_pm_callback. So cpu_hotplug_pm_callback will be called
317 	 * earlier to disable cpu hotplug before bsp online check.
318 	 */
319 	pm_notifier(bsp_pm_callback, -INT_MAX);
320 	return 0;
321 }
322 
323 core_initcall(bsp_pm_check_init);
324