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 #include <linux/tboot.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 <linux/dmi.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 35 struct saved_context saved_context; 36 37 static 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 48 static 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 */ 75 static 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 #ifdef CONFIG_X86_32 86 store_idt(&ctxt->idt); 87 #else 88 /* CONFIG_X86_64 */ 89 store_idt((struct desc_ptr *)&ctxt->idt_limit); 90 #endif 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 #ifdef CONFIG_X86_32 107 savesegment(es, ctxt->es); 108 savesegment(fs, ctxt->fs); 109 savesegment(gs, ctxt->gs); 110 savesegment(ss, ctxt->ss); 111 #else 112 /* CONFIG_X86_64 */ 113 asm volatile ("movw %%ds, %0" : "=m" (ctxt->ds)); 114 asm volatile ("movw %%es, %0" : "=m" (ctxt->es)); 115 asm volatile ("movw %%fs, %0" : "=m" (ctxt->fs)); 116 asm volatile ("movw %%gs, %0" : "=m" (ctxt->gs)); 117 asm volatile ("movw %%ss, %0" : "=m" (ctxt->ss)); 118 119 rdmsrl(MSR_FS_BASE, ctxt->fs_base); 120 rdmsrl(MSR_GS_BASE, ctxt->gs_base); 121 rdmsrl(MSR_KERNEL_GS_BASE, ctxt->gs_kernel_base); 122 mtrr_save_fixed_ranges(NULL); 123 124 rdmsrl(MSR_EFER, ctxt->efer); 125 #endif 126 127 /* 128 * control registers 129 */ 130 ctxt->cr0 = read_cr0(); 131 ctxt->cr2 = read_cr2(); 132 ctxt->cr3 = __read_cr3(); 133 ctxt->cr4 = __read_cr4(); 134 #ifdef CONFIG_X86_64 135 ctxt->cr8 = read_cr8(); 136 #endif 137 ctxt->misc_enable_saved = !rdmsrl_safe(MSR_IA32_MISC_ENABLE, 138 &ctxt->misc_enable); 139 msr_save_context(ctxt); 140 } 141 142 /* Needed by apm.c */ 143 void save_processor_state(void) 144 { 145 __save_processor_state(&saved_context); 146 x86_platform.save_sched_clock_state(); 147 } 148 #ifdef CONFIG_X86_32 149 EXPORT_SYMBOL(save_processor_state); 150 #endif 151 152 static void do_fpu_end(void) 153 { 154 /* 155 * Restore FPU regs if necessary. 156 */ 157 kernel_fpu_end(); 158 } 159 160 static void fix_processor_context(void) 161 { 162 int cpu = smp_processor_id(); 163 struct tss_struct *t = &per_cpu(cpu_tss, cpu); 164 #ifdef CONFIG_X86_64 165 struct desc_struct *desc = get_cpu_gdt_rw(cpu); 166 tss_desc tss; 167 #endif 168 set_tss_desc(cpu, t); /* 169 * This just modifies memory; should not be 170 * necessary. But... This is necessary, because 171 * 386 hardware has concept of busy TSS or some 172 * similar stupidity. 173 */ 174 175 #ifdef CONFIG_X86_64 176 memcpy(&tss, &desc[GDT_ENTRY_TSS], sizeof(tss_desc)); 177 tss.type = 0x9; /* The available 64-bit TSS (see AMD vol 2, pg 91 */ 178 write_gdt_entry(desc, GDT_ENTRY_TSS, &tss, DESC_TSS); 179 180 syscall_init(); /* This sets MSR_*STAR and related */ 181 #endif 182 load_TR_desc(); /* This does ltr */ 183 load_mm_ldt(current->active_mm); /* This does lldt */ 184 initialize_tlbstate_and_flush(); 185 186 fpu__resume_cpu(); 187 188 /* The processor is back on the direct GDT, load back the fixmap */ 189 load_fixmap_gdt(cpu); 190 } 191 192 /** 193 * __restore_processor_state - restore the contents of CPU registers saved 194 * by __save_processor_state() 195 * @ctxt - structure to load the registers contents from 196 */ 197 static void notrace __restore_processor_state(struct saved_context *ctxt) 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_cr8(ctxt->cr8); 212 __write_cr4(ctxt->cr4); 213 #endif 214 write_cr3(ctxt->cr3); 215 write_cr2(ctxt->cr2); 216 write_cr0(ctxt->cr0); 217 218 /* 219 * now restore the descriptor tables to their proper values 220 * ltr is done i fix_processor_context(). 221 */ 222 #ifdef CONFIG_X86_32 223 load_idt(&ctxt->idt); 224 #else 225 /* CONFIG_X86_64 */ 226 load_idt((const struct desc_ptr *)&ctxt->idt_limit); 227 #endif 228 229 #ifdef CONFIG_X86_64 230 /* 231 * We need GSBASE restored before percpu access can work. 232 * percpu access can happen in exception handlers or in complicated 233 * helpers like load_gs_index(). 234 */ 235 wrmsrl(MSR_GS_BASE, ctxt->gs_base); 236 #endif 237 238 fix_processor_context(); 239 240 /* 241 * Restore segment registers. This happens after restoring the GDT 242 * and LDT, which happen in fix_processor_context(). 243 */ 244 #ifdef CONFIG_X86_32 245 loadsegment(es, ctxt->es); 246 loadsegment(fs, ctxt->fs); 247 loadsegment(gs, ctxt->gs); 248 loadsegment(ss, ctxt->ss); 249 250 /* 251 * sysenter MSRs 252 */ 253 if (boot_cpu_has(X86_FEATURE_SEP)) 254 enable_sep_cpu(); 255 #else 256 /* CONFIG_X86_64 */ 257 asm volatile ("movw %0, %%ds" :: "r" (ctxt->ds)); 258 asm volatile ("movw %0, %%es" :: "r" (ctxt->es)); 259 asm volatile ("movw %0, %%fs" :: "r" (ctxt->fs)); 260 load_gs_index(ctxt->gs); 261 asm volatile ("movw %0, %%ss" :: "r" (ctxt->ss)); 262 263 /* 264 * Restore FSBASE and user GSBASE after reloading the respective 265 * segment selectors. 266 */ 267 wrmsrl(MSR_FS_BASE, ctxt->fs_base); 268 wrmsrl(MSR_KERNEL_GS_BASE, ctxt->gs_kernel_base); 269 #endif 270 271 do_fpu_end(); 272 tsc_verify_tsc_adjust(true); 273 x86_platform.restore_sched_clock_state(); 274 mtrr_bp_restore(); 275 perf_restore_debug_store(); 276 msr_restore_context(ctxt); 277 } 278 279 /* Needed by apm.c */ 280 void notrace restore_processor_state(void) 281 { 282 __restore_processor_state(&saved_context); 283 } 284 #ifdef CONFIG_X86_32 285 EXPORT_SYMBOL(restore_processor_state); 286 #endif 287 288 #if defined(CONFIG_HIBERNATION) && defined(CONFIG_HOTPLUG_CPU) 289 static void resume_play_dead(void) 290 { 291 play_dead_common(); 292 tboot_shutdown(TB_SHUTDOWN_WFS); 293 hlt_play_dead(); 294 } 295 296 int hibernate_resume_nonboot_cpu_disable(void) 297 { 298 void (*play_dead)(void) = smp_ops.play_dead; 299 int ret; 300 301 /* 302 * Ensure that MONITOR/MWAIT will not be used in the "play dead" loop 303 * during hibernate image restoration, because it is likely that the 304 * monitored address will be actually written to at that time and then 305 * the "dead" CPU will attempt to execute instructions again, but the 306 * address in its instruction pointer may not be possible to resolve 307 * any more at that point (the page tables used by it previously may 308 * have been overwritten by hibernate image data). 309 */ 310 smp_ops.play_dead = resume_play_dead; 311 ret = disable_nonboot_cpus(); 312 smp_ops.play_dead = play_dead; 313 return ret; 314 } 315 #endif 316 317 /* 318 * When bsp_check() is called in hibernate and suspend, cpu hotplug 319 * is disabled already. So it's unnessary to handle race condition between 320 * cpumask query and cpu hotplug. 321 */ 322 static int bsp_check(void) 323 { 324 if (cpumask_first(cpu_online_mask) != 0) { 325 pr_warn("CPU0 is offline.\n"); 326 return -ENODEV; 327 } 328 329 return 0; 330 } 331 332 static int bsp_pm_callback(struct notifier_block *nb, unsigned long action, 333 void *ptr) 334 { 335 int ret = 0; 336 337 switch (action) { 338 case PM_SUSPEND_PREPARE: 339 case PM_HIBERNATION_PREPARE: 340 ret = bsp_check(); 341 break; 342 #ifdef CONFIG_DEBUG_HOTPLUG_CPU0 343 case PM_RESTORE_PREPARE: 344 /* 345 * When system resumes from hibernation, online CPU0 because 346 * 1. it's required for resume and 347 * 2. the CPU was online before hibernation 348 */ 349 if (!cpu_online(0)) 350 _debug_hotplug_cpu(0, 1); 351 break; 352 case PM_POST_RESTORE: 353 /* 354 * When a resume really happens, this code won't be called. 355 * 356 * This code is called only when user space hibernation software 357 * prepares for snapshot device during boot time. So we just 358 * call _debug_hotplug_cpu() to restore to CPU0's state prior to 359 * preparing the snapshot device. 360 * 361 * This works for normal boot case in our CPU0 hotplug debug 362 * mode, i.e. CPU0 is offline and user mode hibernation 363 * software initializes during boot time. 364 * 365 * If CPU0 is online and user application accesses snapshot 366 * device after boot time, this will offline CPU0 and user may 367 * see different CPU0 state before and after accessing 368 * the snapshot device. But hopefully this is not a case when 369 * user debugging CPU0 hotplug. Even if users hit this case, 370 * they can easily online CPU0 back. 371 * 372 * To simplify this debug code, we only consider normal boot 373 * case. Otherwise we need to remember CPU0's state and restore 374 * to that state and resolve racy conditions etc. 375 */ 376 _debug_hotplug_cpu(0, 0); 377 break; 378 #endif 379 default: 380 break; 381 } 382 return notifier_from_errno(ret); 383 } 384 385 static int __init bsp_pm_check_init(void) 386 { 387 /* 388 * Set this bsp_pm_callback as lower priority than 389 * cpu_hotplug_pm_callback. So cpu_hotplug_pm_callback will be called 390 * earlier to disable cpu hotplug before bsp online check. 391 */ 392 pm_notifier(bsp_pm_callback, -INT_MAX); 393 return 0; 394 } 395 396 core_initcall(bsp_pm_check_init); 397 398 static int msr_init_context(const u32 *msr_id, const int total_num) 399 { 400 int i = 0; 401 struct saved_msr *msr_array; 402 403 if (saved_context.saved_msrs.array || saved_context.saved_msrs.num > 0) { 404 pr_err("x86/pm: MSR quirk already applied, please check your DMI match table.\n"); 405 return -EINVAL; 406 } 407 408 msr_array = kmalloc_array(total_num, sizeof(struct saved_msr), GFP_KERNEL); 409 if (!msr_array) { 410 pr_err("x86/pm: Can not allocate memory to save/restore MSRs during suspend.\n"); 411 return -ENOMEM; 412 } 413 414 for (i = 0; i < total_num; i++) { 415 msr_array[i].info.msr_no = msr_id[i]; 416 msr_array[i].valid = false; 417 msr_array[i].info.reg.q = 0; 418 } 419 saved_context.saved_msrs.num = total_num; 420 saved_context.saved_msrs.array = msr_array; 421 422 return 0; 423 } 424 425 /* 426 * The following section is a quirk framework for problematic BIOSen: 427 * Sometimes MSRs are modified by the BIOSen after suspended to 428 * RAM, this might cause unexpected behavior after wakeup. 429 * Thus we save/restore these specified MSRs across suspend/resume 430 * in order to work around it. 431 * 432 * For any further problematic BIOSen/platforms, 433 * please add your own function similar to msr_initialize_bdw. 434 */ 435 static int msr_initialize_bdw(const struct dmi_system_id *d) 436 { 437 /* Add any extra MSR ids into this array. */ 438 u32 bdw_msr_id[] = { MSR_IA32_THERM_CONTROL }; 439 440 pr_info("x86/pm: %s detected, MSR saving is needed during suspending.\n", d->ident); 441 return msr_init_context(bdw_msr_id, ARRAY_SIZE(bdw_msr_id)); 442 } 443 444 static const struct dmi_system_id msr_save_dmi_table[] = { 445 { 446 .callback = msr_initialize_bdw, 447 .ident = "BROADWELL BDX_EP", 448 .matches = { 449 DMI_MATCH(DMI_PRODUCT_NAME, "GRANTLEY"), 450 DMI_MATCH(DMI_PRODUCT_VERSION, "E63448-400"), 451 }, 452 }, 453 {} 454 }; 455 456 static int pm_check_save_msr(void) 457 { 458 dmi_check_system(msr_save_dmi_table); 459 return 0; 460 } 461 462 device_initcall(pm_check_save_msr); 463