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
35 struct saved_context saved_context;
36
msr_save_context(struct saved_context * ctxt)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
msr_restore_context(struct saved_context * ctxt)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 */
__save_processor_state(struct saved_context * ctxt)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 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 */
save_processor_state(void)132 void save_processor_state(void)
133 {
134 __save_processor_state(&saved_context);
135 x86_platform.save_sched_clock_state();
136 }
137 #ifdef CONFIG_X86_32
138 EXPORT_SYMBOL(save_processor_state);
139 #endif
140
do_fpu_end(void)141 static void do_fpu_end(void)
142 {
143 /*
144 * Restore FPU regs if necessary.
145 */
146 kernel_fpu_end();
147 }
148
fix_processor_context(void)149 static 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 */
__restore_processor_state(struct saved_context * ctxt)194 static 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 */
restore_processor_state(void)269 void notrace restore_processor_state(void)
270 {
271 __restore_processor_state(&saved_context);
272 }
273 #ifdef CONFIG_X86_32
274 EXPORT_SYMBOL(restore_processor_state);
275 #endif
276
277 #if defined(CONFIG_HIBERNATION) && defined(CONFIG_HOTPLUG_CPU)
resume_play_dead(void)278 static void resume_play_dead(void)
279 {
280 play_dead_common();
281 tboot_shutdown(TB_SHUTDOWN_WFS);
282 hlt_play_dead();
283 }
284
hibernate_resume_nonboot_cpu_disable(void)285 int 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 */
bsp_check(void)321 static 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
bsp_pm_callback(struct notifier_block * nb,unsigned long action,void * ptr)331 static 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
bsp_pm_check_init(void)384 static 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
395 core_initcall(bsp_pm_check_init);
396
msr_build_context(const u32 * msr_id,const int num)397 static 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 */
msr_initialize_bdw(const struct dmi_system_id * d)444 static 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
453 static 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
msr_save_cpuid_features(const struct x86_cpu_id * c)465 static 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
477 static 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
495 typedef int (*pm_cpu_match_t)(const struct x86_cpu_id *);
pm_cpu_check(const struct x86_cpu_id * c)496 static 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
pm_check_save_msr(void)512 static 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
520 device_initcall(pm_check_save_msr);
521