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 #include <linux/pgtable.h>
17
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 struct cpuinfo_x86 *c;
197
198 if (ctxt->misc_enable_saved)
199 wrmsrl(MSR_IA32_MISC_ENABLE, ctxt->misc_enable);
200 /*
201 * control registers
202 */
203 /* cr4 was introduced in the Pentium CPU */
204 #ifdef CONFIG_X86_32
205 if (ctxt->cr4)
206 __write_cr4(ctxt->cr4);
207 #else
208 /* CONFIG X86_64 */
209 wrmsrl(MSR_EFER, ctxt->efer);
210 __write_cr4(ctxt->cr4);
211 #endif
212 write_cr3(ctxt->cr3);
213 write_cr2(ctxt->cr2);
214 write_cr0(ctxt->cr0);
215
216 /* Restore the IDT. */
217 load_idt(&ctxt->idt);
218
219 /*
220 * Just in case the asm code got us here with the SS, DS, or ES
221 * out of sync with the GDT, update them.
222 */
223 loadsegment(ss, __KERNEL_DS);
224 loadsegment(ds, __USER_DS);
225 loadsegment(es, __USER_DS);
226
227 /*
228 * Restore percpu access. Percpu access can happen in exception
229 * handlers or in complicated helpers like load_gs_index().
230 */
231 #ifdef CONFIG_X86_64
232 wrmsrl(MSR_GS_BASE, ctxt->kernelmode_gs_base);
233 #else
234 loadsegment(fs, __KERNEL_PERCPU);
235 loadsegment(gs, __KERNEL_STACK_CANARY);
236 #endif
237
238 /* Restore the TSS, RO GDT, LDT, and usermode-relevant MSRs. */
239 fix_processor_context();
240
241 /*
242 * Now that we have descriptor tables fully restored and working
243 * exception handling, restore the usermode segments.
244 */
245 #ifdef CONFIG_X86_64
246 loadsegment(ds, ctxt->es);
247 loadsegment(es, ctxt->es);
248 loadsegment(fs, ctxt->fs);
249 load_gs_index(ctxt->gs);
250
251 /*
252 * Restore FSBASE and GSBASE after restoring the selectors, since
253 * restoring the selectors clobbers the bases. Keep in mind
254 * that MSR_KERNEL_GS_BASE is horribly misnamed.
255 */
256 wrmsrl(MSR_FS_BASE, ctxt->fs_base);
257 wrmsrl(MSR_KERNEL_GS_BASE, ctxt->usermode_gs_base);
258 #elif defined(CONFIG_X86_32_LAZY_GS)
259 loadsegment(gs, ctxt->gs);
260 #endif
261
262 do_fpu_end();
263 tsc_verify_tsc_adjust(true);
264 x86_platform.restore_sched_clock_state();
265 mtrr_bp_restore();
266 perf_restore_debug_store();
267 msr_restore_context(ctxt);
268
269 c = &cpu_data(smp_processor_id());
270 if (cpu_has(c, X86_FEATURE_MSR_IA32_FEAT_CTL))
271 init_ia32_feat_ctl(c);
272 }
273
274 /* Needed by apm.c */
restore_processor_state(void)275 void notrace restore_processor_state(void)
276 {
277 __restore_processor_state(&saved_context);
278 }
279 #ifdef CONFIG_X86_32
280 EXPORT_SYMBOL(restore_processor_state);
281 #endif
282
283 #if defined(CONFIG_HIBERNATION) && defined(CONFIG_HOTPLUG_CPU)
resume_play_dead(void)284 static void resume_play_dead(void)
285 {
286 play_dead_common();
287 tboot_shutdown(TB_SHUTDOWN_WFS);
288 hlt_play_dead();
289 }
290
hibernate_resume_nonboot_cpu_disable(void)291 int hibernate_resume_nonboot_cpu_disable(void)
292 {
293 void (*play_dead)(void) = smp_ops.play_dead;
294 int ret;
295
296 /*
297 * Ensure that MONITOR/MWAIT will not be used in the "play dead" loop
298 * during hibernate image restoration, because it is likely that the
299 * monitored address will be actually written to at that time and then
300 * the "dead" CPU will attempt to execute instructions again, but the
301 * address in its instruction pointer may not be possible to resolve
302 * any more at that point (the page tables used by it previously may
303 * have been overwritten by hibernate image data).
304 *
305 * First, make sure that we wake up all the potentially disabled SMT
306 * threads which have been initially brought up and then put into
307 * mwait/cpuidle sleep.
308 * Those will be put to proper (not interfering with hibernation
309 * resume) sleep afterwards, and the resumed kernel will decide itself
310 * what to do with them.
311 */
312 ret = cpuhp_smt_enable();
313 if (ret)
314 return ret;
315 smp_ops.play_dead = resume_play_dead;
316 ret = freeze_secondary_cpus(0);
317 smp_ops.play_dead = play_dead;
318 return ret;
319 }
320 #endif
321
322 /*
323 * When bsp_check() is called in hibernate and suspend, cpu hotplug
324 * is disabled already. So it's unnessary to handle race condition between
325 * cpumask query and cpu hotplug.
326 */
bsp_check(void)327 static int bsp_check(void)
328 {
329 if (cpumask_first(cpu_online_mask) != 0) {
330 pr_warn("CPU0 is offline.\n");
331 return -ENODEV;
332 }
333
334 return 0;
335 }
336
bsp_pm_callback(struct notifier_block * nb,unsigned long action,void * ptr)337 static int bsp_pm_callback(struct notifier_block *nb, unsigned long action,
338 void *ptr)
339 {
340 int ret = 0;
341
342 switch (action) {
343 case PM_SUSPEND_PREPARE:
344 case PM_HIBERNATION_PREPARE:
345 ret = bsp_check();
346 break;
347 #ifdef CONFIG_DEBUG_HOTPLUG_CPU0
348 case PM_RESTORE_PREPARE:
349 /*
350 * When system resumes from hibernation, online CPU0 because
351 * 1. it's required for resume and
352 * 2. the CPU was online before hibernation
353 */
354 if (!cpu_online(0))
355 _debug_hotplug_cpu(0, 1);
356 break;
357 case PM_POST_RESTORE:
358 /*
359 * When a resume really happens, this code won't be called.
360 *
361 * This code is called only when user space hibernation software
362 * prepares for snapshot device during boot time. So we just
363 * call _debug_hotplug_cpu() to restore to CPU0's state prior to
364 * preparing the snapshot device.
365 *
366 * This works for normal boot case in our CPU0 hotplug debug
367 * mode, i.e. CPU0 is offline and user mode hibernation
368 * software initializes during boot time.
369 *
370 * If CPU0 is online and user application accesses snapshot
371 * device after boot time, this will offline CPU0 and user may
372 * see different CPU0 state before and after accessing
373 * the snapshot device. But hopefully this is not a case when
374 * user debugging CPU0 hotplug. Even if users hit this case,
375 * they can easily online CPU0 back.
376 *
377 * To simplify this debug code, we only consider normal boot
378 * case. Otherwise we need to remember CPU0's state and restore
379 * to that state and resolve racy conditions etc.
380 */
381 _debug_hotplug_cpu(0, 0);
382 break;
383 #endif
384 default:
385 break;
386 }
387 return notifier_from_errno(ret);
388 }
389
bsp_pm_check_init(void)390 static int __init bsp_pm_check_init(void)
391 {
392 /*
393 * Set this bsp_pm_callback as lower priority than
394 * cpu_hotplug_pm_callback. So cpu_hotplug_pm_callback will be called
395 * earlier to disable cpu hotplug before bsp online check.
396 */
397 pm_notifier(bsp_pm_callback, -INT_MAX);
398 return 0;
399 }
400
401 core_initcall(bsp_pm_check_init);
402
msr_build_context(const u32 * msr_id,const int num)403 static int msr_build_context(const u32 *msr_id, const int num)
404 {
405 struct saved_msrs *saved_msrs = &saved_context.saved_msrs;
406 struct saved_msr *msr_array;
407 int total_num;
408 int i, j;
409
410 total_num = saved_msrs->num + num;
411
412 msr_array = kmalloc_array(total_num, sizeof(struct saved_msr), GFP_KERNEL);
413 if (!msr_array) {
414 pr_err("x86/pm: Can not allocate memory to save/restore MSRs during suspend.\n");
415 return -ENOMEM;
416 }
417
418 if (saved_msrs->array) {
419 /*
420 * Multiple callbacks can invoke this function, so copy any
421 * MSR save requests from previous invocations.
422 */
423 memcpy(msr_array, saved_msrs->array,
424 sizeof(struct saved_msr) * saved_msrs->num);
425
426 kfree(saved_msrs->array);
427 }
428
429 for (i = saved_msrs->num, j = 0; i < total_num; i++, j++) {
430 msr_array[i].info.msr_no = msr_id[j];
431 msr_array[i].valid = false;
432 msr_array[i].info.reg.q = 0;
433 }
434 saved_msrs->num = total_num;
435 saved_msrs->array = msr_array;
436
437 return 0;
438 }
439
440 /*
441 * The following sections are a quirk framework for problematic BIOSen:
442 * Sometimes MSRs are modified by the BIOSen after suspended to
443 * RAM, this might cause unexpected behavior after wakeup.
444 * Thus we save/restore these specified MSRs across suspend/resume
445 * in order to work around it.
446 *
447 * For any further problematic BIOSen/platforms,
448 * please add your own function similar to msr_initialize_bdw.
449 */
msr_initialize_bdw(const struct dmi_system_id * d)450 static int msr_initialize_bdw(const struct dmi_system_id *d)
451 {
452 /* Add any extra MSR ids into this array. */
453 u32 bdw_msr_id[] = { MSR_IA32_THERM_CONTROL };
454
455 pr_info("x86/pm: %s detected, MSR saving is needed during suspending.\n", d->ident);
456 return msr_build_context(bdw_msr_id, ARRAY_SIZE(bdw_msr_id));
457 }
458
459 static const struct dmi_system_id msr_save_dmi_table[] = {
460 {
461 .callback = msr_initialize_bdw,
462 .ident = "BROADWELL BDX_EP",
463 .matches = {
464 DMI_MATCH(DMI_PRODUCT_NAME, "GRANTLEY"),
465 DMI_MATCH(DMI_PRODUCT_VERSION, "E63448-400"),
466 },
467 },
468 {}
469 };
470
msr_save_cpuid_features(const struct x86_cpu_id * c)471 static int msr_save_cpuid_features(const struct x86_cpu_id *c)
472 {
473 u32 cpuid_msr_id[] = {
474 MSR_AMD64_CPUID_FN_1,
475 };
476
477 pr_info("x86/pm: family %#hx cpu detected, MSR saving is needed during suspending.\n",
478 c->family);
479
480 return msr_build_context(cpuid_msr_id, ARRAY_SIZE(cpuid_msr_id));
481 }
482
483 static const struct x86_cpu_id msr_save_cpu_table[] = {
484 X86_MATCH_VENDOR_FAM(AMD, 0x15, &msr_save_cpuid_features),
485 X86_MATCH_VENDOR_FAM(AMD, 0x16, &msr_save_cpuid_features),
486 {}
487 };
488
489 typedef int (*pm_cpu_match_t)(const struct x86_cpu_id *);
pm_cpu_check(const struct x86_cpu_id * c)490 static int pm_cpu_check(const struct x86_cpu_id *c)
491 {
492 const struct x86_cpu_id *m;
493 int ret = 0;
494
495 m = x86_match_cpu(msr_save_cpu_table);
496 if (m) {
497 pm_cpu_match_t fn;
498
499 fn = (pm_cpu_match_t)m->driver_data;
500 ret = fn(m);
501 }
502
503 return ret;
504 }
505
pm_check_save_msr(void)506 static int pm_check_save_msr(void)
507 {
508 dmi_check_system(msr_save_dmi_table);
509 pm_cpu_check(msr_save_cpu_table);
510
511 return 0;
512 }
513
514 device_initcall(pm_check_save_msr);
515
