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/api.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 #include <asm/microcode.h>
29 
30 #ifdef CONFIG_X86_32
31 __visible unsigned long saved_context_ebx;
32 __visible unsigned long saved_context_esp, saved_context_ebp;
33 __visible unsigned long saved_context_esi, saved_context_edi;
34 __visible unsigned long saved_context_eflags;
35 #endif
36 struct saved_context saved_context;
37 
msr_save_context(struct saved_context * ctxt)38 static void msr_save_context(struct saved_context *ctxt)
39 {
40 	struct saved_msr *msr = ctxt->saved_msrs.array;
41 	struct saved_msr *end = msr + ctxt->saved_msrs.num;
42 
43 	while (msr < end) {
44 		if (msr->valid)
45 			rdmsrl(msr->info.msr_no, msr->info.reg.q);
46 		msr++;
47 	}
48 }
49 
msr_restore_context(struct saved_context * ctxt)50 static void msr_restore_context(struct saved_context *ctxt)
51 {
52 	struct saved_msr *msr = ctxt->saved_msrs.array;
53 	struct saved_msr *end = msr + ctxt->saved_msrs.num;
54 
55 	while (msr < end) {
56 		if (msr->valid)
57 			wrmsrl(msr->info.msr_no, msr->info.reg.q);
58 		msr++;
59 	}
60 }
61 
62 /**
63  * __save_processor_state() - Save CPU registers before creating a
64  *                             hibernation image and before restoring
65  *                             the memory state from it
66  * @ctxt: Structure to store the registers contents in.
67  *
68  * NOTE: If there is a CPU register the modification of which by the
69  * boot kernel (ie. the kernel used for loading the hibernation image)
70  * might affect the operations of the restored target kernel (ie. the one
71  * saved in the hibernation image), then its contents must be saved by this
72  * function.  In other words, if kernel A is hibernated and different
73  * kernel B is used for loading the hibernation image into memory, the
74  * kernel A's __save_processor_state() function must save all registers
75  * needed by kernel A, so that it can operate correctly after the resume
76  * regardless of what kernel B does in the meantime.
77  */
__save_processor_state(struct saved_context * ctxt)78 static void __save_processor_state(struct saved_context *ctxt)
79 {
80 #ifdef CONFIG_X86_32
81 	mtrr_save_fixed_ranges(NULL);
82 #endif
83 	kernel_fpu_begin();
84 
85 	/*
86 	 * descriptor tables
87 	 */
88 	store_idt(&ctxt->idt);
89 
90 	/*
91 	 * We save it here, but restore it only in the hibernate case.
92 	 * For ACPI S3 resume, this is loaded via 'early_gdt_desc' in 64-bit
93 	 * mode in "secondary_startup_64". In 32-bit mode it is done via
94 	 * 'pmode_gdt' in wakeup_start.
95 	 */
96 	ctxt->gdt_desc.size = GDT_SIZE - 1;
97 	ctxt->gdt_desc.address = (unsigned long)get_cpu_gdt_rw(smp_processor_id());
98 
99 	store_tr(ctxt->tr);
100 
101 	/* XMM0..XMM15 should be handled by kernel_fpu_begin(). */
102 	/*
103 	 * segment registers
104 	 */
105 	savesegment(gs, ctxt->gs);
106 #ifdef CONFIG_X86_64
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 #endif
236 
237 	/* Restore the TSS, RO GDT, LDT, and usermode-relevant MSRs. */
238 	fix_processor_context();
239 
240 	/*
241 	 * Now that we have descriptor tables fully restored and working
242 	 * exception handling, restore the usermode segments.
243 	 */
244 #ifdef CONFIG_X86_64
245 	loadsegment(ds, ctxt->es);
246 	loadsegment(es, ctxt->es);
247 	loadsegment(fs, ctxt->fs);
248 	load_gs_index(ctxt->gs);
249 
250 	/*
251 	 * Restore FSBASE and GSBASE after restoring the selectors, since
252 	 * restoring the selectors clobbers the bases.  Keep in mind
253 	 * that MSR_KERNEL_GS_BASE is horribly misnamed.
254 	 */
255 	wrmsrl(MSR_FS_BASE, ctxt->fs_base);
256 	wrmsrl(MSR_KERNEL_GS_BASE, ctxt->usermode_gs_base);
257 #else
258 	loadsegment(gs, ctxt->gs);
259 #endif
260 
261 	do_fpu_end();
262 	tsc_verify_tsc_adjust(true);
263 	x86_platform.restore_sched_clock_state();
264 	mtrr_bp_restore();
265 	perf_restore_debug_store();
266 
267 	c = &cpu_data(smp_processor_id());
268 	if (cpu_has(c, X86_FEATURE_MSR_IA32_FEAT_CTL))
269 		init_ia32_feat_ctl(c);
270 
271 	microcode_bsp_resume();
272 
273 	/*
274 	 * This needs to happen after the microcode has been updated upon resume
275 	 * because some of the MSRs are "emulated" in microcode.
276 	 */
277 	msr_restore_context(ctxt);
278 }
279 
280 /* Needed by apm.c */
restore_processor_state(void)281 void notrace restore_processor_state(void)
282 {
283 	__restore_processor_state(&saved_context);
284 }
285 #ifdef CONFIG_X86_32
286 EXPORT_SYMBOL(restore_processor_state);
287 #endif
288 
289 #if defined(CONFIG_HIBERNATION) && defined(CONFIG_HOTPLUG_CPU)
resume_play_dead(void)290 static void resume_play_dead(void)
291 {
292 	play_dead_common();
293 	tboot_shutdown(TB_SHUTDOWN_WFS);
294 	hlt_play_dead();
295 }
296 
hibernate_resume_nonboot_cpu_disable(void)297 int hibernate_resume_nonboot_cpu_disable(void)
298 {
299 	void (*play_dead)(void) = smp_ops.play_dead;
300 	int ret;
301 
302 	/*
303 	 * Ensure that MONITOR/MWAIT will not be used in the "play dead" loop
304 	 * during hibernate image restoration, because it is likely that the
305 	 * monitored address will be actually written to at that time and then
306 	 * the "dead" CPU will attempt to execute instructions again, but the
307 	 * address in its instruction pointer may not be possible to resolve
308 	 * any more at that point (the page tables used by it previously may
309 	 * have been overwritten by hibernate image data).
310 	 *
311 	 * First, make sure that we wake up all the potentially disabled SMT
312 	 * threads which have been initially brought up and then put into
313 	 * mwait/cpuidle sleep.
314 	 * Those will be put to proper (not interfering with hibernation
315 	 * resume) sleep afterwards, and the resumed kernel will decide itself
316 	 * what to do with them.
317 	 */
318 	ret = cpuhp_smt_enable();
319 	if (ret)
320 		return ret;
321 	smp_ops.play_dead = resume_play_dead;
322 	ret = freeze_secondary_cpus(0);
323 	smp_ops.play_dead = play_dead;
324 	return ret;
325 }
326 #endif
327 
328 /*
329  * When bsp_check() is called in hibernate and suspend, cpu hotplug
330  * is disabled already. So it's unnecessary to handle race condition between
331  * cpumask query and cpu hotplug.
332  */
bsp_check(void)333 static int bsp_check(void)
334 {
335 	if (cpumask_first(cpu_online_mask) != 0) {
336 		pr_warn("CPU0 is offline.\n");
337 		return -ENODEV;
338 	}
339 
340 	return 0;
341 }
342 
bsp_pm_callback(struct notifier_block * nb,unsigned long action,void * ptr)343 static int bsp_pm_callback(struct notifier_block *nb, unsigned long action,
344 			   void *ptr)
345 {
346 	int ret = 0;
347 
348 	switch (action) {
349 	case PM_SUSPEND_PREPARE:
350 	case PM_HIBERNATION_PREPARE:
351 		ret = bsp_check();
352 		break;
353 #ifdef CONFIG_DEBUG_HOTPLUG_CPU0
354 	case PM_RESTORE_PREPARE:
355 		/*
356 		 * When system resumes from hibernation, online CPU0 because
357 		 * 1. it's required for resume and
358 		 * 2. the CPU was online before hibernation
359 		 */
360 		if (!cpu_online(0))
361 			_debug_hotplug_cpu(0, 1);
362 		break;
363 	case PM_POST_RESTORE:
364 		/*
365 		 * When a resume really happens, this code won't be called.
366 		 *
367 		 * This code is called only when user space hibernation software
368 		 * prepares for snapshot device during boot time. So we just
369 		 * call _debug_hotplug_cpu() to restore to CPU0's state prior to
370 		 * preparing the snapshot device.
371 		 *
372 		 * This works for normal boot case in our CPU0 hotplug debug
373 		 * mode, i.e. CPU0 is offline and user mode hibernation
374 		 * software initializes during boot time.
375 		 *
376 		 * If CPU0 is online and user application accesses snapshot
377 		 * device after boot time, this will offline CPU0 and user may
378 		 * see different CPU0 state before and after accessing
379 		 * the snapshot device. But hopefully this is not a case when
380 		 * user debugging CPU0 hotplug. Even if users hit this case,
381 		 * they can easily online CPU0 back.
382 		 *
383 		 * To simplify this debug code, we only consider normal boot
384 		 * case. Otherwise we need to remember CPU0's state and restore
385 		 * to that state and resolve racy conditions etc.
386 		 */
387 		_debug_hotplug_cpu(0, 0);
388 		break;
389 #endif
390 	default:
391 		break;
392 	}
393 	return notifier_from_errno(ret);
394 }
395 
bsp_pm_check_init(void)396 static int __init bsp_pm_check_init(void)
397 {
398 	/*
399 	 * Set this bsp_pm_callback as lower priority than
400 	 * cpu_hotplug_pm_callback. So cpu_hotplug_pm_callback will be called
401 	 * earlier to disable cpu hotplug before bsp online check.
402 	 */
403 	pm_notifier(bsp_pm_callback, -INT_MAX);
404 	return 0;
405 }
406 
407 core_initcall(bsp_pm_check_init);
408 
msr_build_context(const u32 * msr_id,const int num)409 static int msr_build_context(const u32 *msr_id, const int num)
410 {
411 	struct saved_msrs *saved_msrs = &saved_context.saved_msrs;
412 	struct saved_msr *msr_array;
413 	int total_num;
414 	int i, j;
415 
416 	total_num = saved_msrs->num + num;
417 
418 	msr_array = kmalloc_array(total_num, sizeof(struct saved_msr), GFP_KERNEL);
419 	if (!msr_array) {
420 		pr_err("x86/pm: Can not allocate memory to save/restore MSRs during suspend.\n");
421 		return -ENOMEM;
422 	}
423 
424 	if (saved_msrs->array) {
425 		/*
426 		 * Multiple callbacks can invoke this function, so copy any
427 		 * MSR save requests from previous invocations.
428 		 */
429 		memcpy(msr_array, saved_msrs->array,
430 		       sizeof(struct saved_msr) * saved_msrs->num);
431 
432 		kfree(saved_msrs->array);
433 	}
434 
435 	for (i = saved_msrs->num, j = 0; i < total_num; i++, j++) {
436 		u64 dummy;
437 
438 		msr_array[i].info.msr_no	= msr_id[j];
439 		msr_array[i].valid		= !rdmsrl_safe(msr_id[j], &dummy);
440 		msr_array[i].info.reg.q		= 0;
441 	}
442 	saved_msrs->num   = total_num;
443 	saved_msrs->array = msr_array;
444 
445 	return 0;
446 }
447 
448 /*
449  * The following sections are a quirk framework for problematic BIOSen:
450  * Sometimes MSRs are modified by the BIOSen after suspended to
451  * RAM, this might cause unexpected behavior after wakeup.
452  * Thus we save/restore these specified MSRs across suspend/resume
453  * in order to work around it.
454  *
455  * For any further problematic BIOSen/platforms,
456  * please add your own function similar to msr_initialize_bdw.
457  */
msr_initialize_bdw(const struct dmi_system_id * d)458 static int msr_initialize_bdw(const struct dmi_system_id *d)
459 {
460 	/* Add any extra MSR ids into this array. */
461 	u32 bdw_msr_id[] = { MSR_IA32_THERM_CONTROL };
462 
463 	pr_info("x86/pm: %s detected, MSR saving is needed during suspending.\n", d->ident);
464 	return msr_build_context(bdw_msr_id, ARRAY_SIZE(bdw_msr_id));
465 }
466 
467 static const struct dmi_system_id msr_save_dmi_table[] = {
468 	{
469 	 .callback = msr_initialize_bdw,
470 	 .ident = "BROADWELL BDX_EP",
471 	 .matches = {
472 		DMI_MATCH(DMI_PRODUCT_NAME, "GRANTLEY"),
473 		DMI_MATCH(DMI_PRODUCT_VERSION, "E63448-400"),
474 		},
475 	},
476 	{}
477 };
478 
msr_save_cpuid_features(const struct x86_cpu_id * c)479 static int msr_save_cpuid_features(const struct x86_cpu_id *c)
480 {
481 	u32 cpuid_msr_id[] = {
482 		MSR_AMD64_CPUID_FN_1,
483 	};
484 
485 	pr_info("x86/pm: family %#hx cpu detected, MSR saving is needed during suspending.\n",
486 		c->family);
487 
488 	return msr_build_context(cpuid_msr_id, ARRAY_SIZE(cpuid_msr_id));
489 }
490 
491 static const struct x86_cpu_id msr_save_cpu_table[] = {
492 	X86_MATCH_VENDOR_FAM(AMD, 0x15, &msr_save_cpuid_features),
493 	X86_MATCH_VENDOR_FAM(AMD, 0x16, &msr_save_cpuid_features),
494 	{}
495 };
496 
497 typedef int (*pm_cpu_match_t)(const struct x86_cpu_id *);
pm_cpu_check(const struct x86_cpu_id * c)498 static int pm_cpu_check(const struct x86_cpu_id *c)
499 {
500 	const struct x86_cpu_id *m;
501 	int ret = 0;
502 
503 	m = x86_match_cpu(msr_save_cpu_table);
504 	if (m) {
505 		pm_cpu_match_t fn;
506 
507 		fn = (pm_cpu_match_t)m->driver_data;
508 		ret = fn(m);
509 	}
510 
511 	return ret;
512 }
513 
pm_save_spec_msr(void)514 static void pm_save_spec_msr(void)
515 {
516 	struct msr_enumeration {
517 		u32 msr_no;
518 		u32 feature;
519 	} msr_enum[] = {
520 		{ MSR_IA32_SPEC_CTRL,	 X86_FEATURE_MSR_SPEC_CTRL },
521 		{ MSR_IA32_TSX_CTRL,	 X86_FEATURE_MSR_TSX_CTRL },
522 		{ MSR_TSX_FORCE_ABORT,	 X86_FEATURE_TSX_FORCE_ABORT },
523 		{ MSR_IA32_MCU_OPT_CTRL, X86_FEATURE_SRBDS_CTRL },
524 		{ MSR_AMD64_LS_CFG,	 X86_FEATURE_LS_CFG_SSBD },
525 		{ MSR_AMD64_DE_CFG,	 X86_FEATURE_LFENCE_RDTSC },
526 	};
527 	int i;
528 
529 	for (i = 0; i < ARRAY_SIZE(msr_enum); i++) {
530 		if (boot_cpu_has(msr_enum[i].feature))
531 			msr_build_context(&msr_enum[i].msr_no, 1);
532 	}
533 }
534 
pm_check_save_msr(void)535 static int pm_check_save_msr(void)
536 {
537 	dmi_check_system(msr_save_dmi_table);
538 	pm_cpu_check(msr_save_cpu_table);
539 	pm_save_spec_msr();
540 
541 	return 0;
542 }
543 
544 device_initcall(pm_check_save_msr);
545