1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * X86 specific Hyper-V initialization code.
4 *
5 * Copyright (C) 2016, Microsoft, Inc.
6 *
7 * Author : K. Y. Srinivasan <kys@microsoft.com>
8 */
9
10 #include <linux/efi.h>
11 #include <linux/types.h>
12 #include <linux/bitfield.h>
13 #include <linux/io.h>
14 #include <asm/apic.h>
15 #include <asm/desc.h>
16 #include <asm/hypervisor.h>
17 #include <asm/hyperv-tlfs.h>
18 #include <asm/mshyperv.h>
19 #include <asm/idtentry.h>
20 #include <linux/kexec.h>
21 #include <linux/version.h>
22 #include <linux/vmalloc.h>
23 #include <linux/mm.h>
24 #include <linux/hyperv.h>
25 #include <linux/slab.h>
26 #include <linux/kernel.h>
27 #include <linux/cpuhotplug.h>
28 #include <linux/syscore_ops.h>
29 #include <clocksource/hyperv_timer.h>
30 #include <linux/highmem.h>
31
32 int hyperv_init_cpuhp;
33 u64 hv_current_partition_id = ~0ull;
34 EXPORT_SYMBOL_GPL(hv_current_partition_id);
35
36 void *hv_hypercall_pg;
37 EXPORT_SYMBOL_GPL(hv_hypercall_pg);
38
39 union hv_ghcb __percpu **hv_ghcb_pg;
40
41 /* Storage to save the hypercall page temporarily for hibernation */
42 static void *hv_hypercall_pg_saved;
43
44 struct hv_vp_assist_page **hv_vp_assist_page;
45 EXPORT_SYMBOL_GPL(hv_vp_assist_page);
46
hyperv_init_ghcb(void)47 static int hyperv_init_ghcb(void)
48 {
49 u64 ghcb_gpa;
50 void *ghcb_va;
51 void **ghcb_base;
52
53 if (!hv_isolation_type_snp())
54 return 0;
55
56 if (!hv_ghcb_pg)
57 return -EINVAL;
58
59 /*
60 * GHCB page is allocated by paravisor. The address
61 * returned by MSR_AMD64_SEV_ES_GHCB is above shared
62 * memory boundary and map it here.
63 */
64 rdmsrl(MSR_AMD64_SEV_ES_GHCB, ghcb_gpa);
65 ghcb_va = memremap(ghcb_gpa, HV_HYP_PAGE_SIZE, MEMREMAP_WB);
66 if (!ghcb_va)
67 return -ENOMEM;
68
69 ghcb_base = (void **)this_cpu_ptr(hv_ghcb_pg);
70 *ghcb_base = ghcb_va;
71
72 return 0;
73 }
74
hv_cpu_init(unsigned int cpu)75 static int hv_cpu_init(unsigned int cpu)
76 {
77 union hv_vp_assist_msr_contents msr = { 0 };
78 struct hv_vp_assist_page **hvp = &hv_vp_assist_page[smp_processor_id()];
79 int ret;
80
81 ret = hv_common_cpu_init(cpu);
82 if (ret)
83 return ret;
84
85 if (!hv_vp_assist_page)
86 return 0;
87
88 if (!*hvp) {
89 if (hv_root_partition) {
90 /*
91 * For root partition we get the hypervisor provided VP assist
92 * page, instead of allocating a new page.
93 */
94 rdmsrl(HV_X64_MSR_VP_ASSIST_PAGE, msr.as_uint64);
95 *hvp = memremap(msr.pfn <<
96 HV_X64_MSR_VP_ASSIST_PAGE_ADDRESS_SHIFT,
97 PAGE_SIZE, MEMREMAP_WB);
98 } else {
99 /*
100 * The VP assist page is an "overlay" page (see Hyper-V TLFS's
101 * Section 5.2.1 "GPA Overlay Pages"). Here it must be zeroed
102 * out to make sure we always write the EOI MSR in
103 * hv_apic_eoi_write() *after* the EOI optimization is disabled
104 * in hv_cpu_die(), otherwise a CPU may not be stopped in the
105 * case of CPU offlining and the VM will hang.
106 */
107 *hvp = __vmalloc(PAGE_SIZE, GFP_KERNEL | __GFP_ZERO);
108 if (*hvp)
109 msr.pfn = vmalloc_to_pfn(*hvp);
110 }
111 WARN_ON(!(*hvp));
112 if (*hvp) {
113 msr.enable = 1;
114 wrmsrl(HV_X64_MSR_VP_ASSIST_PAGE, msr.as_uint64);
115 }
116 }
117
118 return hyperv_init_ghcb();
119 }
120
121 static void (*hv_reenlightenment_cb)(void);
122
hv_reenlightenment_notify(struct work_struct * dummy)123 static void hv_reenlightenment_notify(struct work_struct *dummy)
124 {
125 struct hv_tsc_emulation_status emu_status;
126
127 rdmsrl(HV_X64_MSR_TSC_EMULATION_STATUS, *(u64 *)&emu_status);
128
129 /* Don't issue the callback if TSC accesses are not emulated */
130 if (hv_reenlightenment_cb && emu_status.inprogress)
131 hv_reenlightenment_cb();
132 }
133 static DECLARE_DELAYED_WORK(hv_reenlightenment_work, hv_reenlightenment_notify);
134
hyperv_stop_tsc_emulation(void)135 void hyperv_stop_tsc_emulation(void)
136 {
137 u64 freq;
138 struct hv_tsc_emulation_status emu_status;
139
140 rdmsrl(HV_X64_MSR_TSC_EMULATION_STATUS, *(u64 *)&emu_status);
141 emu_status.inprogress = 0;
142 wrmsrl(HV_X64_MSR_TSC_EMULATION_STATUS, *(u64 *)&emu_status);
143
144 rdmsrl(HV_X64_MSR_TSC_FREQUENCY, freq);
145 tsc_khz = div64_u64(freq, 1000);
146 }
147 EXPORT_SYMBOL_GPL(hyperv_stop_tsc_emulation);
148
hv_reenlightenment_available(void)149 static inline bool hv_reenlightenment_available(void)
150 {
151 /*
152 * Check for required features and privileges to make TSC frequency
153 * change notifications work.
154 */
155 return ms_hyperv.features & HV_ACCESS_FREQUENCY_MSRS &&
156 ms_hyperv.misc_features & HV_FEATURE_FREQUENCY_MSRS_AVAILABLE &&
157 ms_hyperv.features & HV_ACCESS_REENLIGHTENMENT;
158 }
159
DEFINE_IDTENTRY_SYSVEC(sysvec_hyperv_reenlightenment)160 DEFINE_IDTENTRY_SYSVEC(sysvec_hyperv_reenlightenment)
161 {
162 ack_APIC_irq();
163 inc_irq_stat(irq_hv_reenlightenment_count);
164 schedule_delayed_work(&hv_reenlightenment_work, HZ/10);
165 }
166
set_hv_tscchange_cb(void (* cb)(void))167 void set_hv_tscchange_cb(void (*cb)(void))
168 {
169 struct hv_reenlightenment_control re_ctrl = {
170 .vector = HYPERV_REENLIGHTENMENT_VECTOR,
171 .enabled = 1,
172 };
173 struct hv_tsc_emulation_control emu_ctrl = {.enabled = 1};
174
175 if (!hv_reenlightenment_available()) {
176 pr_warn("Hyper-V: reenlightenment support is unavailable\n");
177 return;
178 }
179
180 if (!hv_vp_index)
181 return;
182
183 hv_reenlightenment_cb = cb;
184
185 /* Make sure callback is registered before we write to MSRs */
186 wmb();
187
188 re_ctrl.target_vp = hv_vp_index[get_cpu()];
189
190 wrmsrl(HV_X64_MSR_REENLIGHTENMENT_CONTROL, *((u64 *)&re_ctrl));
191 wrmsrl(HV_X64_MSR_TSC_EMULATION_CONTROL, *((u64 *)&emu_ctrl));
192
193 put_cpu();
194 }
195 EXPORT_SYMBOL_GPL(set_hv_tscchange_cb);
196
clear_hv_tscchange_cb(void)197 void clear_hv_tscchange_cb(void)
198 {
199 struct hv_reenlightenment_control re_ctrl;
200
201 if (!hv_reenlightenment_available())
202 return;
203
204 rdmsrl(HV_X64_MSR_REENLIGHTENMENT_CONTROL, *(u64 *)&re_ctrl);
205 re_ctrl.enabled = 0;
206 wrmsrl(HV_X64_MSR_REENLIGHTENMENT_CONTROL, *(u64 *)&re_ctrl);
207
208 hv_reenlightenment_cb = NULL;
209 }
210 EXPORT_SYMBOL_GPL(clear_hv_tscchange_cb);
211
hv_cpu_die(unsigned int cpu)212 static int hv_cpu_die(unsigned int cpu)
213 {
214 struct hv_reenlightenment_control re_ctrl;
215 unsigned int new_cpu;
216 void **ghcb_va;
217
218 if (hv_ghcb_pg) {
219 ghcb_va = (void **)this_cpu_ptr(hv_ghcb_pg);
220 if (*ghcb_va)
221 memunmap(*ghcb_va);
222 *ghcb_va = NULL;
223 }
224
225 hv_common_cpu_die(cpu);
226
227 if (hv_vp_assist_page && hv_vp_assist_page[cpu]) {
228 union hv_vp_assist_msr_contents msr = { 0 };
229 if (hv_root_partition) {
230 /*
231 * For root partition the VP assist page is mapped to
232 * hypervisor provided page, and thus we unmap the
233 * page here and nullify it, so that in future we have
234 * correct page address mapped in hv_cpu_init.
235 */
236 memunmap(hv_vp_assist_page[cpu]);
237 hv_vp_assist_page[cpu] = NULL;
238 rdmsrl(HV_X64_MSR_VP_ASSIST_PAGE, msr.as_uint64);
239 msr.enable = 0;
240 }
241 wrmsrl(HV_X64_MSR_VP_ASSIST_PAGE, msr.as_uint64);
242 }
243
244 if (hv_reenlightenment_cb == NULL)
245 return 0;
246
247 rdmsrl(HV_X64_MSR_REENLIGHTENMENT_CONTROL, *((u64 *)&re_ctrl));
248 if (re_ctrl.target_vp == hv_vp_index[cpu]) {
249 /*
250 * Reassign reenlightenment notifications to some other online
251 * CPU or just disable the feature if there are no online CPUs
252 * left (happens on hibernation).
253 */
254 new_cpu = cpumask_any_but(cpu_online_mask, cpu);
255
256 if (new_cpu < nr_cpu_ids)
257 re_ctrl.target_vp = hv_vp_index[new_cpu];
258 else
259 re_ctrl.enabled = 0;
260
261 wrmsrl(HV_X64_MSR_REENLIGHTENMENT_CONTROL, *((u64 *)&re_ctrl));
262 }
263
264 return 0;
265 }
266
hv_pci_init(void)267 static int __init hv_pci_init(void)
268 {
269 int gen2vm = efi_enabled(EFI_BOOT);
270
271 /*
272 * For Generation-2 VM, we exit from pci_arch_init() by returning 0.
273 * The purpose is to suppress the harmless warning:
274 * "PCI: Fatal: No config space access function found"
275 */
276 if (gen2vm)
277 return 0;
278
279 /* For Generation-1 VM, we'll proceed in pci_arch_init(). */
280 return 1;
281 }
282
hv_suspend(void)283 static int hv_suspend(void)
284 {
285 union hv_x64_msr_hypercall_contents hypercall_msr;
286 int ret;
287
288 if (hv_root_partition)
289 return -EPERM;
290
291 /*
292 * Reset the hypercall page as it is going to be invalidated
293 * across hibernation. Setting hv_hypercall_pg to NULL ensures
294 * that any subsequent hypercall operation fails safely instead of
295 * crashing due to an access of an invalid page. The hypercall page
296 * pointer is restored on resume.
297 */
298 hv_hypercall_pg_saved = hv_hypercall_pg;
299 hv_hypercall_pg = NULL;
300
301 /* Disable the hypercall page in the hypervisor */
302 rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
303 hypercall_msr.enable = 0;
304 wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
305
306 ret = hv_cpu_die(0);
307 return ret;
308 }
309
hv_resume(void)310 static void hv_resume(void)
311 {
312 union hv_x64_msr_hypercall_contents hypercall_msr;
313 int ret;
314
315 ret = hv_cpu_init(0);
316 WARN_ON(ret);
317
318 /* Re-enable the hypercall page */
319 rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
320 hypercall_msr.enable = 1;
321 hypercall_msr.guest_physical_address =
322 vmalloc_to_pfn(hv_hypercall_pg_saved);
323 wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
324
325 hv_hypercall_pg = hv_hypercall_pg_saved;
326 hv_hypercall_pg_saved = NULL;
327
328 /*
329 * Reenlightenment notifications are disabled by hv_cpu_die(0),
330 * reenable them here if hv_reenlightenment_cb was previously set.
331 */
332 if (hv_reenlightenment_cb)
333 set_hv_tscchange_cb(hv_reenlightenment_cb);
334 }
335
336 /* Note: when the ops are called, only CPU0 is online and IRQs are disabled. */
337 static struct syscore_ops hv_syscore_ops = {
338 .suspend = hv_suspend,
339 .resume = hv_resume,
340 };
341
342 static void (* __initdata old_setup_percpu_clockev)(void);
343
hv_stimer_setup_percpu_clockev(void)344 static void __init hv_stimer_setup_percpu_clockev(void)
345 {
346 /*
347 * Ignore any errors in setting up stimer clockevents
348 * as we can run with the LAPIC timer as a fallback.
349 */
350 (void)hv_stimer_alloc(false);
351
352 /*
353 * Still register the LAPIC timer, because the direct-mode STIMER is
354 * not supported by old versions of Hyper-V. This also allows users
355 * to switch to LAPIC timer via /sys, if they want to.
356 */
357 if (old_setup_percpu_clockev)
358 old_setup_percpu_clockev();
359 }
360
hv_get_partition_id(void)361 static void __init hv_get_partition_id(void)
362 {
363 struct hv_get_partition_id *output_page;
364 u64 status;
365 unsigned long flags;
366
367 local_irq_save(flags);
368 output_page = *this_cpu_ptr(hyperv_pcpu_output_arg);
369 status = hv_do_hypercall(HVCALL_GET_PARTITION_ID, NULL, output_page);
370 if (!hv_result_success(status)) {
371 /* No point in proceeding if this failed */
372 pr_err("Failed to get partition ID: %lld\n", status);
373 BUG();
374 }
375 hv_current_partition_id = output_page->partition_id;
376 local_irq_restore(flags);
377 }
378
379 /*
380 * This function is to be invoked early in the boot sequence after the
381 * hypervisor has been detected.
382 *
383 * 1. Setup the hypercall page.
384 * 2. Register Hyper-V specific clocksource.
385 * 3. Setup Hyper-V specific APIC entry points.
386 */
hyperv_init(void)387 void __init hyperv_init(void)
388 {
389 u64 guest_id;
390 union hv_x64_msr_hypercall_contents hypercall_msr;
391 int cpuhp;
392
393 if (x86_hyper_type != X86_HYPER_MS_HYPERV)
394 return;
395
396 if (hv_common_init())
397 return;
398
399 hv_vp_assist_page = kcalloc(num_possible_cpus(),
400 sizeof(*hv_vp_assist_page), GFP_KERNEL);
401 if (!hv_vp_assist_page) {
402 ms_hyperv.hints &= ~HV_X64_ENLIGHTENED_VMCS_RECOMMENDED;
403 goto common_free;
404 }
405
406 if (hv_isolation_type_snp()) {
407 hv_ghcb_pg = alloc_percpu(union hv_ghcb *);
408 if (!hv_ghcb_pg)
409 goto free_vp_assist_page;
410 }
411
412 cpuhp = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "x86/hyperv_init:online",
413 hv_cpu_init, hv_cpu_die);
414 if (cpuhp < 0)
415 goto free_ghcb_page;
416
417 /*
418 * Setup the hypercall page and enable hypercalls.
419 * 1. Register the guest ID
420 * 2. Enable the hypercall and register the hypercall page
421 */
422 guest_id = generate_guest_id(0, LINUX_VERSION_CODE, 0);
423 wrmsrl(HV_X64_MSR_GUEST_OS_ID, guest_id);
424
425 /* Hyper-V requires to write guest os id via ghcb in SNP IVM. */
426 hv_ghcb_msr_write(HV_X64_MSR_GUEST_OS_ID, guest_id);
427
428 hv_hypercall_pg = __vmalloc_node_range(PAGE_SIZE, 1, VMALLOC_START,
429 VMALLOC_END, GFP_KERNEL, PAGE_KERNEL_ROX,
430 VM_FLUSH_RESET_PERMS, NUMA_NO_NODE,
431 __builtin_return_address(0));
432 if (hv_hypercall_pg == NULL)
433 goto clean_guest_os_id;
434
435 rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
436 hypercall_msr.enable = 1;
437
438 if (hv_root_partition) {
439 struct page *pg;
440 void *src, *dst;
441
442 /*
443 * For the root partition, the hypervisor will set up its
444 * hypercall page. The hypervisor guarantees it will not show
445 * up in the root's address space. The root can't change the
446 * location of the hypercall page.
447 *
448 * Order is important here. We must enable the hypercall page
449 * so it is populated with code, then copy the code to an
450 * executable page.
451 */
452 wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
453
454 pg = vmalloc_to_page(hv_hypercall_pg);
455 dst = kmap(pg);
456 src = memremap(hypercall_msr.guest_physical_address << PAGE_SHIFT, PAGE_SIZE,
457 MEMREMAP_WB);
458 BUG_ON(!(src && dst));
459 memcpy(dst, src, HV_HYP_PAGE_SIZE);
460 memunmap(src);
461 kunmap(pg);
462 } else {
463 hypercall_msr.guest_physical_address = vmalloc_to_pfn(hv_hypercall_pg);
464 wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
465 }
466
467 /*
468 * hyperv_init() is called before LAPIC is initialized: see
469 * apic_intr_mode_init() -> x86_platform.apic_post_init() and
470 * apic_bsp_setup() -> setup_local_APIC(). The direct-mode STIMER
471 * depends on LAPIC, so hv_stimer_alloc() should be called from
472 * x86_init.timers.setup_percpu_clockev.
473 */
474 old_setup_percpu_clockev = x86_init.timers.setup_percpu_clockev;
475 x86_init.timers.setup_percpu_clockev = hv_stimer_setup_percpu_clockev;
476
477 hv_apic_init();
478
479 x86_init.pci.arch_init = hv_pci_init;
480
481 register_syscore_ops(&hv_syscore_ops);
482
483 hyperv_init_cpuhp = cpuhp;
484
485 if (cpuid_ebx(HYPERV_CPUID_FEATURES) & HV_ACCESS_PARTITION_ID)
486 hv_get_partition_id();
487
488 BUG_ON(hv_root_partition && hv_current_partition_id == ~0ull);
489
490 #ifdef CONFIG_PCI_MSI
491 /*
492 * If we're running as root, we want to create our own PCI MSI domain.
493 * We can't set this in hv_pci_init because that would be too late.
494 */
495 if (hv_root_partition)
496 x86_init.irqs.create_pci_msi_domain = hv_create_pci_msi_domain;
497 #endif
498
499 /* Query the VMs extended capability once, so that it can be cached. */
500 hv_query_ext_cap(0);
501 return;
502
503 clean_guest_os_id:
504 wrmsrl(HV_X64_MSR_GUEST_OS_ID, 0);
505 hv_ghcb_msr_write(HV_X64_MSR_GUEST_OS_ID, 0);
506 cpuhp_remove_state(cpuhp);
507 free_ghcb_page:
508 free_percpu(hv_ghcb_pg);
509 free_vp_assist_page:
510 kfree(hv_vp_assist_page);
511 hv_vp_assist_page = NULL;
512 common_free:
513 hv_common_free();
514 }
515
516 /*
517 * This routine is called before kexec/kdump, it does the required cleanup.
518 */
hyperv_cleanup(void)519 void hyperv_cleanup(void)
520 {
521 union hv_x64_msr_hypercall_contents hypercall_msr;
522
523 unregister_syscore_ops(&hv_syscore_ops);
524
525 /* Reset our OS id */
526 wrmsrl(HV_X64_MSR_GUEST_OS_ID, 0);
527 hv_ghcb_msr_write(HV_X64_MSR_GUEST_OS_ID, 0);
528
529 /*
530 * Reset hypercall page reference before reset the page,
531 * let hypercall operations fail safely rather than
532 * panic the kernel for using invalid hypercall page
533 */
534 hv_hypercall_pg = NULL;
535
536 /* Reset the hypercall page */
537 hypercall_msr.as_uint64 = 0;
538 wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
539
540 /* Reset the TSC page */
541 hypercall_msr.as_uint64 = 0;
542 wrmsrl(HV_X64_MSR_REFERENCE_TSC, hypercall_msr.as_uint64);
543 }
544
hyperv_report_panic(struct pt_regs * regs,long err,bool in_die)545 void hyperv_report_panic(struct pt_regs *regs, long err, bool in_die)
546 {
547 static bool panic_reported;
548 u64 guest_id;
549
550 if (in_die && !panic_on_oops)
551 return;
552
553 /*
554 * We prefer to report panic on 'die' chain as we have proper
555 * registers to report, but if we miss it (e.g. on BUG()) we need
556 * to report it on 'panic'.
557 */
558 if (panic_reported)
559 return;
560 panic_reported = true;
561
562 rdmsrl(HV_X64_MSR_GUEST_OS_ID, guest_id);
563
564 wrmsrl(HV_X64_MSR_CRASH_P0, err);
565 wrmsrl(HV_X64_MSR_CRASH_P1, guest_id);
566 wrmsrl(HV_X64_MSR_CRASH_P2, regs->ip);
567 wrmsrl(HV_X64_MSR_CRASH_P3, regs->ax);
568 wrmsrl(HV_X64_MSR_CRASH_P4, regs->sp);
569
570 /*
571 * Let Hyper-V know there is crash data available
572 */
573 wrmsrl(HV_X64_MSR_CRASH_CTL, HV_CRASH_CTL_CRASH_NOTIFY);
574 }
575 EXPORT_SYMBOL_GPL(hyperv_report_panic);
576
hv_is_hyperv_initialized(void)577 bool hv_is_hyperv_initialized(void)
578 {
579 union hv_x64_msr_hypercall_contents hypercall_msr;
580
581 /*
582 * Ensure that we're really on Hyper-V, and not a KVM or Xen
583 * emulation of Hyper-V
584 */
585 if (x86_hyper_type != X86_HYPER_MS_HYPERV)
586 return false;
587
588 /*
589 * Verify that earlier initialization succeeded by checking
590 * that the hypercall page is setup
591 */
592 hypercall_msr.as_uint64 = 0;
593 rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
594
595 return hypercall_msr.enable;
596 }
597 EXPORT_SYMBOL_GPL(hv_is_hyperv_initialized);
598