Commit | Line | Data |
---|---|---|
1da177e4 | 1 | /* |
6d48becd | 2 | * Suspend support specific for i386/x86-64. |
1da177e4 LT |
3 | * |
4 | * Distribute under GPLv2 | |
5 | * | |
cf7700fe | 6 | * Copyright (c) 2007 Rafael J. Wysocki <rjw@sisk.pl> |
a2531293 | 7 | * Copyright (c) 2002 Pavel Machek <pavel@ucw.cz> |
1da177e4 LT |
8 | * Copyright (c) 2001 Patrick Mochel <mochel@osdl.org> |
9 | */ | |
10 | ||
1da177e4 | 11 | #include <linux/suspend.h> |
69c60c88 | 12 | #include <linux/export.h> |
f6783d20 | 13 | #include <linux/smp.h> |
1d9d8639 | 14 | #include <linux/perf_event.h> |
406f992e | 15 | #include <linux/tboot.h> |
f6783d20 | 16 | |
3dd08325 | 17 | #include <asm/pgtable.h> |
f6783d20 | 18 | #include <asm/proto.h> |
3ebad590 | 19 | #include <asm/mtrr.h> |
f6783d20 SL |
20 | #include <asm/page.h> |
21 | #include <asm/mce.h> | |
a8af7898 | 22 | #include <asm/suspend.h> |
952f07ec | 23 | #include <asm/fpu/internal.h> |
1e350066 | 24 | #include <asm/debugreg.h> |
a71c8bc5 | 25 | #include <asm/cpu.h> |
37868fe1 | 26 | #include <asm/mmu_context.h> |
7a9c2dd0 | 27 | #include <linux/dmi.h> |
1da177e4 | 28 | |
833b2ca0 | 29 | #ifdef CONFIG_X86_32 |
d6efc2f7 AK |
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; | |
833b2ca0 | 34 | #endif |
cc456c4e | 35 | struct saved_context saved_context; |
1da177e4 | 36 | |
7a9c2dd0 CY |
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 | ||
5c9c9bec RW |
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 | */ | |
cae45957 | 75 | static void __save_processor_state(struct saved_context *ctxt) |
1da177e4 | 76 | { |
f9ebbe53 SL |
77 | #ifdef CONFIG_X86_32 |
78 | mtrr_save_fixed_ranges(NULL); | |
79 | #endif | |
1da177e4 LT |
80 | kernel_fpu_begin(); |
81 | ||
82 | /* | |
83 | * descriptor tables | |
84 | */ | |
f9ebbe53 | 85 | store_idt(&ctxt->idt); |
090edbe2 | 86 | |
cc456c4e KRW |
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; | |
69218e47 | 94 | ctxt->gdt_desc.address = (unsigned long)get_cpu_gdt_rw(smp_processor_id()); |
cc456c4e | 95 | |
9d1c6e7c | 96 | store_tr(ctxt->tr); |
1da177e4 LT |
97 | |
98 | /* XMM0..XMM15 should be handled by kernel_fpu_begin(). */ | |
1da177e4 LT |
99 | /* |
100 | * segment registers | |
101 | */ | |
7ee18d67 | 102 | #ifdef CONFIG_X86_32_LAZY_GS |
f9ebbe53 | 103 | savesegment(gs, ctxt->gs); |
7ee18d67 AL |
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); | |
1da177e4 LT |
110 | |
111 | rdmsrl(MSR_FS_BASE, ctxt->fs_base); | |
7ee18d67 AL |
112 | rdmsrl(MSR_GS_BASE, ctxt->kernelmode_gs_base); |
113 | rdmsrl(MSR_KERNEL_GS_BASE, ctxt->usermode_gs_base); | |
3ebad590 | 114 | mtrr_save_fixed_ranges(NULL); |
1da177e4 | 115 | |
f9ebbe53 SL |
116 | rdmsrl(MSR_EFER, ctxt->efer); |
117 | #endif | |
118 | ||
1da177e4 | 119 | /* |
cf7700fe | 120 | * control registers |
1da177e4 | 121 | */ |
f51c9452 GOC |
122 | ctxt->cr0 = read_cr0(); |
123 | ctxt->cr2 = read_cr2(); | |
6c690ee1 | 124 | ctxt->cr3 = __read_cr3(); |
1ef55be1 | 125 | ctxt->cr4 = __read_cr4(); |
1e02ce4c | 126 | #ifdef CONFIG_X86_64 |
f51c9452 | 127 | ctxt->cr8 = read_cr8(); |
f9ebbe53 | 128 | #endif |
85a0e753 OZ |
129 | ctxt->misc_enable_saved = !rdmsrl_safe(MSR_IA32_MISC_ENABLE, |
130 | &ctxt->misc_enable); | |
7a9c2dd0 | 131 | msr_save_context(ctxt); |
1da177e4 LT |
132 | } |
133 | ||
f9ebbe53 | 134 | /* Needed by apm.c */ |
1da177e4 LT |
135 | void save_processor_state(void) |
136 | { | |
137 | __save_processor_state(&saved_context); | |
b74f05d6 | 138 | x86_platform.save_sched_clock_state(); |
1da177e4 | 139 | } |
f9ebbe53 SL |
140 | #ifdef CONFIG_X86_32 |
141 | EXPORT_SYMBOL(save_processor_state); | |
142 | #endif | |
1da177e4 | 143 | |
08967f94 | 144 | static void do_fpu_end(void) |
1da177e4 | 145 | { |
08967f94 | 146 | /* |
3134d04b | 147 | * Restore FPU regs if necessary. |
08967f94 SL |
148 | */ |
149 | kernel_fpu_end(); | |
1da177e4 LT |
150 | } |
151 | ||
3134d04b SL |
152 | static void fix_processor_context(void) |
153 | { | |
154 | int cpu = smp_processor_id(); | |
24933b82 | 155 | struct tss_struct *t = &per_cpu(cpu_tss, cpu); |
4d681be3 | 156 | #ifdef CONFIG_X86_64 |
69218e47 | 157 | struct desc_struct *desc = get_cpu_gdt_rw(cpu); |
4d681be3 | 158 | tss_desc tss; |
159 | #endif | |
3134d04b SL |
160 | set_tss_desc(cpu, t); /* |
161 | * This just modifies memory; should not be | |
162 | * necessary. But... This is necessary, because | |
163 | * 386 hardware has concept of busy TSS or some | |
164 | * similar stupidity. | |
165 | */ | |
166 | ||
167 | #ifdef CONFIG_X86_64 | |
4d681be3 | 168 | memcpy(&tss, &desc[GDT_ENTRY_TSS], sizeof(tss_desc)); |
169 | tss.type = 0x9; /* The available 64-bit TSS (see AMD vol 2, pg 91 */ | |
170 | write_gdt_entry(desc, GDT_ENTRY_TSS, &tss, DESC_TSS); | |
3134d04b SL |
171 | |
172 | syscall_init(); /* This sets MSR_*STAR and related */ | |
896c80be AL |
173 | #else |
174 | if (boot_cpu_has(X86_FEATURE_SEP)) | |
175 | enable_sep_cpu(); | |
3134d04b SL |
176 | #endif |
177 | load_TR_desc(); /* This does ltr */ | |
37868fe1 | 178 | load_mm_ldt(current->active_mm); /* This does lldt */ |
72c0098d | 179 | initialize_tlbstate_and_flush(); |
9254aaa0 IM |
180 | |
181 | fpu__resume_cpu(); | |
69218e47 TG |
182 | |
183 | /* The processor is back on the direct GDT, load back the fixmap */ | |
184 | load_fixmap_gdt(cpu); | |
3134d04b SL |
185 | } |
186 | ||
5c9c9bec | 187 | /** |
7ee18d67 AL |
188 | * __restore_processor_state - restore the contents of CPU registers saved |
189 | * by __save_processor_state() | |
190 | * @ctxt - structure to load the registers contents from | |
191 | * | |
192 | * The asm code that gets us here will have restored a usable GDT, although | |
193 | * it will be pointing to the wrong alias. | |
5c9c9bec | 194 | */ |
b8f99b3e | 195 | static void notrace __restore_processor_state(struct saved_context *ctxt) |
1da177e4 | 196 | { |
85a0e753 OZ |
197 | if (ctxt->misc_enable_saved) |
198 | wrmsrl(MSR_IA32_MISC_ENABLE, ctxt->misc_enable); | |
1da177e4 LT |
199 | /* |
200 | * control registers | |
201 | */ | |
3134d04b SL |
202 | /* cr4 was introduced in the Pentium CPU */ |
203 | #ifdef CONFIG_X86_32 | |
204 | if (ctxt->cr4) | |
1e02ce4c | 205 | __write_cr4(ctxt->cr4); |
3134d04b SL |
206 | #else |
207 | /* CONFIG X86_64 */ | |
3c321bce | 208 | wrmsrl(MSR_EFER, ctxt->efer); |
f51c9452 | 209 | write_cr8(ctxt->cr8); |
1e02ce4c | 210 | __write_cr4(ctxt->cr4); |
3134d04b | 211 | #endif |
f51c9452 GOC |
212 | write_cr3(ctxt->cr3); |
213 | write_cr2(ctxt->cr2); | |
214 | write_cr0(ctxt->cr0); | |
1da177e4 | 215 | |
7ee18d67 AL |
216 | /* Restore the IDT. */ |
217 | load_idt(&ctxt->idt); | |
218 | ||
8d783b3e | 219 | /* |
7ee18d67 AL |
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. | |
8d783b3e | 222 | */ |
7ee18d67 AL |
223 | loadsegment(ss, __KERNEL_DS); |
224 | loadsegment(ds, __USER_DS); | |
225 | loadsegment(es, __USER_DS); | |
8d783b3e | 226 | |
1da177e4 | 227 | /* |
7ee18d67 AL |
228 | * Restore percpu access. Percpu access can happen in exception |
229 | * handlers or in complicated helpers like load_gs_index(). | |
5b06bbcf | 230 | */ |
7ee18d67 AL |
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); | |
5b06bbcf AL |
236 | #endif |
237 | ||
7ee18d67 | 238 | /* Restore the TSS, RO GDT, LDT, and usermode-relevant MSRs. */ |
5b06bbcf AL |
239 | fix_processor_context(); |
240 | ||
241 | /* | |
7ee18d67 AL |
242 | * Now that we have descriptor tables fully restored and working |
243 | * exception handling, restore the usermode segments. | |
1da177e4 | 244 | */ |
7ee18d67 AL |
245 | #ifdef CONFIG_X86_64 |
246 | loadsegment(ds, ctxt->es); | |
3134d04b SL |
247 | loadsegment(es, ctxt->es); |
248 | loadsegment(fs, ctxt->fs); | |
1da177e4 | 249 | load_gs_index(ctxt->gs); |
1da177e4 | 250 | |
5b06bbcf | 251 | /* |
7ee18d67 AL |
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. | |
5b06bbcf | 255 | */ |
1da177e4 | 256 | wrmsrl(MSR_FS_BASE, ctxt->fs_base); |
7ee18d67 AL |
257 | wrmsrl(MSR_KERNEL_GS_BASE, ctxt->usermode_gs_base); |
258 | #elif defined(CONFIG_X86_32_LAZY_GS) | |
259 | loadsegment(gs, ctxt->gs); | |
3134d04b | 260 | #endif |
1da177e4 | 261 | |
1da177e4 | 262 | do_fpu_end(); |
6a369583 | 263 | tsc_verify_tsc_adjust(true); |
dba69d10 | 264 | x86_platform.restore_sched_clock_state(); |
d0af9eed | 265 | mtrr_bp_restore(); |
1d9d8639 | 266 | perf_restore_debug_store(); |
7a9c2dd0 | 267 | msr_restore_context(ctxt); |
1da177e4 LT |
268 | } |
269 | ||
3134d04b | 270 | /* Needed by apm.c */ |
b8f99b3e | 271 | void notrace restore_processor_state(void) |
1da177e4 LT |
272 | { |
273 | __restore_processor_state(&saved_context); | |
274 | } | |
3134d04b SL |
275 | #ifdef CONFIG_X86_32 |
276 | EXPORT_SYMBOL(restore_processor_state); | |
277 | #endif | |
209efae1 | 278 | |
406f992e RW |
279 | #if defined(CONFIG_HIBERNATION) && defined(CONFIG_HOTPLUG_CPU) |
280 | static void resume_play_dead(void) | |
281 | { | |
282 | play_dead_common(); | |
283 | tboot_shutdown(TB_SHUTDOWN_WFS); | |
284 | hlt_play_dead(); | |
285 | } | |
286 | ||
287 | int hibernate_resume_nonboot_cpu_disable(void) | |
288 | { | |
289 | void (*play_dead)(void) = smp_ops.play_dead; | |
290 | int ret; | |
291 | ||
292 | /* | |
293 | * Ensure that MONITOR/MWAIT will not be used in the "play dead" loop | |
294 | * during hibernate image restoration, because it is likely that the | |
295 | * monitored address will be actually written to at that time and then | |
296 | * the "dead" CPU will attempt to execute instructions again, but the | |
297 | * address in its instruction pointer may not be possible to resolve | |
298 | * any more at that point (the page tables used by it previously may | |
299 | * have been overwritten by hibernate image data). | |
300 | */ | |
301 | smp_ops.play_dead = resume_play_dead; | |
302 | ret = disable_nonboot_cpus(); | |
303 | smp_ops.play_dead = play_dead; | |
304 | return ret; | |
305 | } | |
306 | #endif | |
307 | ||
209efae1 FY |
308 | /* |
309 | * When bsp_check() is called in hibernate and suspend, cpu hotplug | |
310 | * is disabled already. So it's unnessary to handle race condition between | |
311 | * cpumask query and cpu hotplug. | |
312 | */ | |
313 | static int bsp_check(void) | |
314 | { | |
315 | if (cpumask_first(cpu_online_mask) != 0) { | |
316 | pr_warn("CPU0 is offline.\n"); | |
317 | return -ENODEV; | |
318 | } | |
319 | ||
320 | return 0; | |
321 | } | |
322 | ||
323 | static int bsp_pm_callback(struct notifier_block *nb, unsigned long action, | |
324 | void *ptr) | |
325 | { | |
326 | int ret = 0; | |
327 | ||
328 | switch (action) { | |
329 | case PM_SUSPEND_PREPARE: | |
330 | case PM_HIBERNATION_PREPARE: | |
331 | ret = bsp_check(); | |
332 | break; | |
a71c8bc5 FY |
333 | #ifdef CONFIG_DEBUG_HOTPLUG_CPU0 |
334 | case PM_RESTORE_PREPARE: | |
335 | /* | |
336 | * When system resumes from hibernation, online CPU0 because | |
337 | * 1. it's required for resume and | |
338 | * 2. the CPU was online before hibernation | |
339 | */ | |
340 | if (!cpu_online(0)) | |
341 | _debug_hotplug_cpu(0, 1); | |
342 | break; | |
343 | case PM_POST_RESTORE: | |
344 | /* | |
345 | * When a resume really happens, this code won't be called. | |
346 | * | |
347 | * This code is called only when user space hibernation software | |
348 | * prepares for snapshot device during boot time. So we just | |
349 | * call _debug_hotplug_cpu() to restore to CPU0's state prior to | |
350 | * preparing the snapshot device. | |
351 | * | |
352 | * This works for normal boot case in our CPU0 hotplug debug | |
353 | * mode, i.e. CPU0 is offline and user mode hibernation | |
354 | * software initializes during boot time. | |
355 | * | |
356 | * If CPU0 is online and user application accesses snapshot | |
357 | * device after boot time, this will offline CPU0 and user may | |
358 | * see different CPU0 state before and after accessing | |
359 | * the snapshot device. But hopefully this is not a case when | |
360 | * user debugging CPU0 hotplug. Even if users hit this case, | |
361 | * they can easily online CPU0 back. | |
362 | * | |
363 | * To simplify this debug code, we only consider normal boot | |
364 | * case. Otherwise we need to remember CPU0's state and restore | |
365 | * to that state and resolve racy conditions etc. | |
366 | */ | |
367 | _debug_hotplug_cpu(0, 0); | |
368 | break; | |
369 | #endif | |
209efae1 FY |
370 | default: |
371 | break; | |
372 | } | |
373 | return notifier_from_errno(ret); | |
374 | } | |
375 | ||
376 | static int __init bsp_pm_check_init(void) | |
377 | { | |
378 | /* | |
379 | * Set this bsp_pm_callback as lower priority than | |
380 | * cpu_hotplug_pm_callback. So cpu_hotplug_pm_callback will be called | |
381 | * earlier to disable cpu hotplug before bsp online check. | |
382 | */ | |
383 | pm_notifier(bsp_pm_callback, -INT_MAX); | |
384 | return 0; | |
385 | } | |
386 | ||
387 | core_initcall(bsp_pm_check_init); | |
7a9c2dd0 CY |
388 | |
389 | static int msr_init_context(const u32 *msr_id, const int total_num) | |
390 | { | |
391 | int i = 0; | |
392 | struct saved_msr *msr_array; | |
393 | ||
394 | if (saved_context.saved_msrs.array || saved_context.saved_msrs.num > 0) { | |
395 | pr_err("x86/pm: MSR quirk already applied, please check your DMI match table.\n"); | |
396 | return -EINVAL; | |
397 | } | |
398 | ||
399 | msr_array = kmalloc_array(total_num, sizeof(struct saved_msr), GFP_KERNEL); | |
400 | if (!msr_array) { | |
401 | pr_err("x86/pm: Can not allocate memory to save/restore MSRs during suspend.\n"); | |
402 | return -ENOMEM; | |
403 | } | |
404 | ||
405 | for (i = 0; i < total_num; i++) { | |
406 | msr_array[i].info.msr_no = msr_id[i]; | |
407 | msr_array[i].valid = false; | |
408 | msr_array[i].info.reg.q = 0; | |
409 | } | |
410 | saved_context.saved_msrs.num = total_num; | |
411 | saved_context.saved_msrs.array = msr_array; | |
412 | ||
413 | return 0; | |
414 | } | |
415 | ||
416 | /* | |
417 | * The following section is a quirk framework for problematic BIOSen: | |
418 | * Sometimes MSRs are modified by the BIOSen after suspended to | |
419 | * RAM, this might cause unexpected behavior after wakeup. | |
420 | * Thus we save/restore these specified MSRs across suspend/resume | |
421 | * in order to work around it. | |
422 | * | |
423 | * For any further problematic BIOSen/platforms, | |
424 | * please add your own function similar to msr_initialize_bdw. | |
425 | */ | |
426 | static int msr_initialize_bdw(const struct dmi_system_id *d) | |
427 | { | |
428 | /* Add any extra MSR ids into this array. */ | |
429 | u32 bdw_msr_id[] = { MSR_IA32_THERM_CONTROL }; | |
430 | ||
431 | pr_info("x86/pm: %s detected, MSR saving is needed during suspending.\n", d->ident); | |
432 | return msr_init_context(bdw_msr_id, ARRAY_SIZE(bdw_msr_id)); | |
433 | } | |
434 | ||
6faadbbb | 435 | static const struct dmi_system_id msr_save_dmi_table[] = { |
7a9c2dd0 CY |
436 | { |
437 | .callback = msr_initialize_bdw, | |
438 | .ident = "BROADWELL BDX_EP", | |
439 | .matches = { | |
440 | DMI_MATCH(DMI_PRODUCT_NAME, "GRANTLEY"), | |
441 | DMI_MATCH(DMI_PRODUCT_VERSION, "E63448-400"), | |
442 | }, | |
443 | }, | |
444 | {} | |
445 | }; | |
446 | ||
447 | static int pm_check_save_msr(void) | |
448 | { | |
449 | dmi_check_system(msr_save_dmi_table); | |
450 | return 0; | |
451 | } | |
452 | ||
453 | device_initcall(pm_check_save_msr); |