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/suspend.h> 22 #include <asm/fpu/internal.h> 23 #include <asm/debugreg.h> 24 #include <asm/cpu.h> 25 #include <asm/mmu_context.h> 26 27 #ifdef CONFIG_X86_32 28 __visible unsigned long saved_context_ebx; 29 __visible unsigned long saved_context_esp, saved_context_ebp; 30 __visible unsigned long saved_context_esi, saved_context_edi; 31 __visible unsigned long saved_context_eflags; 32 #endif 33 struct saved_context saved_context; 34 35 /** 36 * __save_processor_state - save CPU registers before creating a 37 * hibernation image and before restoring the memory state from it 38 * @ctxt - structure to store the registers contents in 39 * 40 * NOTE: If there is a CPU register the modification of which by the 41 * boot kernel (ie. the kernel used for loading the hibernation image) 42 * might affect the operations of the restored target kernel (ie. the one 43 * saved in the hibernation image), then its contents must be saved by this 44 * function. In other words, if kernel A is hibernated and different 45 * kernel B is used for loading the hibernation image into memory, the 46 * kernel A's __save_processor_state() function must save all registers 47 * needed by kernel A, so that it can operate correctly after the resume 48 * regardless of what kernel B does in the meantime. 49 */ 50 static void __save_processor_state(struct saved_context *ctxt) 51 { 52 #ifdef CONFIG_X86_32 53 mtrr_save_fixed_ranges(NULL); 54 #endif 55 kernel_fpu_begin(); 56 57 /* 58 * descriptor tables 59 */ 60 #ifdef CONFIG_X86_32 61 store_idt(&ctxt->idt); 62 #else 63 /* CONFIG_X86_64 */ 64 store_idt((struct desc_ptr *)&ctxt->idt_limit); 65 #endif 66 /* 67 * We save it here, but restore it only in the hibernate case. 68 * For ACPI S3 resume, this is loaded via 'early_gdt_desc' in 64-bit 69 * mode in "secondary_startup_64". In 32-bit mode it is done via 70 * 'pmode_gdt' in wakeup_start. 71 */ 72 ctxt->gdt_desc.size = GDT_SIZE - 1; 73 ctxt->gdt_desc.address = (unsigned long)get_cpu_gdt_table(smp_processor_id()); 74 75 store_tr(ctxt->tr); 76 77 /* XMM0..XMM15 should be handled by kernel_fpu_begin(). */ 78 /* 79 * segment registers 80 */ 81 #ifdef CONFIG_X86_32 82 savesegment(es, ctxt->es); 83 savesegment(fs, ctxt->fs); 84 savesegment(gs, ctxt->gs); 85 savesegment(ss, ctxt->ss); 86 #else 87 /* CONFIG_X86_64 */ 88 asm volatile ("movw %%ds, %0" : "=m" (ctxt->ds)); 89 asm volatile ("movw %%es, %0" : "=m" (ctxt->es)); 90 asm volatile ("movw %%fs, %0" : "=m" (ctxt->fs)); 91 asm volatile ("movw %%gs, %0" : "=m" (ctxt->gs)); 92 asm volatile ("movw %%ss, %0" : "=m" (ctxt->ss)); 93 94 rdmsrl(MSR_FS_BASE, ctxt->fs_base); 95 rdmsrl(MSR_GS_BASE, ctxt->gs_base); 96 rdmsrl(MSR_KERNEL_GS_BASE, ctxt->gs_kernel_base); 97 mtrr_save_fixed_ranges(NULL); 98 99 rdmsrl(MSR_EFER, ctxt->efer); 100 #endif 101 102 /* 103 * control registers 104 */ 105 ctxt->cr0 = read_cr0(); 106 ctxt->cr2 = read_cr2(); 107 ctxt->cr3 = read_cr3(); 108 ctxt->cr4 = __read_cr4_safe(); 109 #ifdef CONFIG_X86_64 110 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(cpu_tss, cpu); 138 #ifdef CONFIG_X86_64 139 struct desc_struct *desc = get_cpu_gdt_table(cpu); 140 tss_desc tss; 141 #endif 142 set_tss_desc(cpu, t); /* 143 * This just modifies memory; should not be 144 * necessary. But... This is necessary, because 145 * 386 hardware has concept of busy TSS or some 146 * similar stupidity. 147 */ 148 149 #ifdef CONFIG_X86_64 150 memcpy(&tss, &desc[GDT_ENTRY_TSS], sizeof(tss_desc)); 151 tss.type = 0x9; /* The available 64-bit TSS (see AMD vol 2, pg 91 */ 152 write_gdt_entry(desc, GDT_ENTRY_TSS, &tss, DESC_TSS); 153 154 syscall_init(); /* This sets MSR_*STAR and related */ 155 #endif 156 load_TR_desc(); /* This does ltr */ 157 load_mm_ldt(current->active_mm); /* This does lldt */ 158 159 fpu__resume_cpu(); 160 } 161 162 /** 163 * __restore_processor_state - restore the contents of CPU registers saved 164 * by __save_processor_state() 165 * @ctxt - structure to load the registers contents from 166 */ 167 static void notrace __restore_processor_state(struct saved_context *ctxt) 168 { 169 if (ctxt->misc_enable_saved) 170 wrmsrl(MSR_IA32_MISC_ENABLE, ctxt->misc_enable); 171 /* 172 * control registers 173 */ 174 /* cr4 was introduced in the Pentium CPU */ 175 #ifdef CONFIG_X86_32 176 if (ctxt->cr4) 177 __write_cr4(ctxt->cr4); 178 #else 179 /* CONFIG X86_64 */ 180 wrmsrl(MSR_EFER, ctxt->efer); 181 write_cr8(ctxt->cr8); 182 __write_cr4(ctxt->cr4); 183 #endif 184 write_cr3(ctxt->cr3); 185 write_cr2(ctxt->cr2); 186 write_cr0(ctxt->cr0); 187 188 /* 189 * now restore the descriptor tables to their proper values 190 * ltr is done i fix_processor_context(). 191 */ 192 #ifdef CONFIG_X86_32 193 load_idt(&ctxt->idt); 194 #else 195 /* CONFIG_X86_64 */ 196 load_idt((const struct desc_ptr *)&ctxt->idt_limit); 197 #endif 198 199 /* 200 * segment registers 201 */ 202 #ifdef CONFIG_X86_32 203 loadsegment(es, ctxt->es); 204 loadsegment(fs, ctxt->fs); 205 loadsegment(gs, ctxt->gs); 206 loadsegment(ss, ctxt->ss); 207 208 /* 209 * sysenter MSRs 210 */ 211 if (boot_cpu_has(X86_FEATURE_SEP)) 212 enable_sep_cpu(); 213 #else 214 /* CONFIG_X86_64 */ 215 asm volatile ("movw %0, %%ds" :: "r" (ctxt->ds)); 216 asm volatile ("movw %0, %%es" :: "r" (ctxt->es)); 217 asm volatile ("movw %0, %%fs" :: "r" (ctxt->fs)); 218 load_gs_index(ctxt->gs); 219 asm volatile ("movw %0, %%ss" :: "r" (ctxt->ss)); 220 221 wrmsrl(MSR_FS_BASE, ctxt->fs_base); 222 wrmsrl(MSR_GS_BASE, ctxt->gs_base); 223 wrmsrl(MSR_KERNEL_GS_BASE, ctxt->gs_kernel_base); 224 #endif 225 226 fix_processor_context(); 227 228 do_fpu_end(); 229 x86_platform.restore_sched_clock_state(); 230 mtrr_bp_restore(); 231 perf_restore_debug_store(); 232 } 233 234 /* Needed by apm.c */ 235 void notrace restore_processor_state(void) 236 { 237 __restore_processor_state(&saved_context); 238 } 239 #ifdef CONFIG_X86_32 240 EXPORT_SYMBOL(restore_processor_state); 241 #endif 242 243 /* 244 * When bsp_check() is called in hibernate and suspend, cpu hotplug 245 * is disabled already. So it's unnessary to handle race condition between 246 * cpumask query and cpu hotplug. 247 */ 248 static int bsp_check(void) 249 { 250 if (cpumask_first(cpu_online_mask) != 0) { 251 pr_warn("CPU0 is offline.\n"); 252 return -ENODEV; 253 } 254 255 return 0; 256 } 257 258 static int bsp_pm_callback(struct notifier_block *nb, unsigned long action, 259 void *ptr) 260 { 261 int ret = 0; 262 263 switch (action) { 264 case PM_SUSPEND_PREPARE: 265 case PM_HIBERNATION_PREPARE: 266 ret = bsp_check(); 267 break; 268 #ifdef CONFIG_DEBUG_HOTPLUG_CPU0 269 case PM_RESTORE_PREPARE: 270 /* 271 * When system resumes from hibernation, online CPU0 because 272 * 1. it's required for resume and 273 * 2. the CPU was online before hibernation 274 */ 275 if (!cpu_online(0)) 276 _debug_hotplug_cpu(0, 1); 277 break; 278 case PM_POST_RESTORE: 279 /* 280 * When a resume really happens, this code won't be called. 281 * 282 * This code is called only when user space hibernation software 283 * prepares for snapshot device during boot time. So we just 284 * call _debug_hotplug_cpu() to restore to CPU0's state prior to 285 * preparing the snapshot device. 286 * 287 * This works for normal boot case in our CPU0 hotplug debug 288 * mode, i.e. CPU0 is offline and user mode hibernation 289 * software initializes during boot time. 290 * 291 * If CPU0 is online and user application accesses snapshot 292 * device after boot time, this will offline CPU0 and user may 293 * see different CPU0 state before and after accessing 294 * the snapshot device. But hopefully this is not a case when 295 * user debugging CPU0 hotplug. Even if users hit this case, 296 * they can easily online CPU0 back. 297 * 298 * To simplify this debug code, we only consider normal boot 299 * case. Otherwise we need to remember CPU0's state and restore 300 * to that state and resolve racy conditions etc. 301 */ 302 _debug_hotplug_cpu(0, 0); 303 break; 304 #endif 305 default: 306 break; 307 } 308 return notifier_from_errno(ret); 309 } 310 311 static int __init bsp_pm_check_init(void) 312 { 313 /* 314 * Set this bsp_pm_callback as lower priority than 315 * cpu_hotplug_pm_callback. So cpu_hotplug_pm_callback will be called 316 * earlier to disable cpu hotplug before bsp online check. 317 */ 318 pm_notifier(bsp_pm_callback, -INT_MAX); 319 return 0; 320 } 321 322 core_initcall(bsp_pm_check_init); 323