x86/hyperv: Restore VP assist page after cpu offlining/onlining
[linux-2.6-microblaze.git] / arch / x86 / hyperv / hv_init.c
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/sev.h>
17 #include <asm/hypervisor.h>
18 #include <asm/hyperv-tlfs.h>
19 #include <asm/mshyperv.h>
20 #include <asm/idtentry.h>
21 #include <linux/kexec.h>
22 #include <linux/version.h>
23 #include <linux/vmalloc.h>
24 #include <linux/mm.h>
25 #include <linux/hyperv.h>
26 #include <linux/slab.h>
27 #include <linux/kernel.h>
28 #include <linux/cpuhotplug.h>
29 #include <linux/syscore_ops.h>
30 #include <clocksource/hyperv_timer.h>
31 #include <linux/highmem.h>
32 #include <linux/swiotlb.h>
33
34 int hyperv_init_cpuhp;
35 u64 hv_current_partition_id = ~0ull;
36 EXPORT_SYMBOL_GPL(hv_current_partition_id);
37
38 void *hv_hypercall_pg;
39 EXPORT_SYMBOL_GPL(hv_hypercall_pg);
40
41 union hv_ghcb * __percpu *hv_ghcb_pg;
42
43 /* Storage to save the hypercall page temporarily for hibernation */
44 static void *hv_hypercall_pg_saved;
45
46 struct hv_vp_assist_page **hv_vp_assist_page;
47 EXPORT_SYMBOL_GPL(hv_vp_assist_page);
48
49 static int hyperv_init_ghcb(void)
50 {
51         u64 ghcb_gpa;
52         void *ghcb_va;
53         void **ghcb_base;
54
55         if (!hv_isolation_type_snp())
56                 return 0;
57
58         if (!hv_ghcb_pg)
59                 return -EINVAL;
60
61         /*
62          * GHCB page is allocated by paravisor. The address
63          * returned by MSR_AMD64_SEV_ES_GHCB is above shared
64          * memory boundary and map it here.
65          */
66         rdmsrl(MSR_AMD64_SEV_ES_GHCB, ghcb_gpa);
67         ghcb_va = memremap(ghcb_gpa, HV_HYP_PAGE_SIZE, MEMREMAP_WB);
68         if (!ghcb_va)
69                 return -ENOMEM;
70
71         ghcb_base = (void **)this_cpu_ptr(hv_ghcb_pg);
72         *ghcb_base = ghcb_va;
73
74         return 0;
75 }
76
77 static int hv_cpu_init(unsigned int cpu)
78 {
79         union hv_vp_assist_msr_contents msr = { 0 };
80         struct hv_vp_assist_page **hvp = &hv_vp_assist_page[cpu];
81         int ret;
82
83         ret = hv_common_cpu_init(cpu);
84         if (ret)
85                 return ret;
86
87         if (!hv_vp_assist_page)
88                 return 0;
89
90         if (hv_root_partition) {
91                 /*
92                  * For root partition we get the hypervisor provided VP assist
93                  * page, instead of allocating a new page.
94                  */
95                 rdmsrl(HV_X64_MSR_VP_ASSIST_PAGE, msr.as_uint64);
96                 *hvp = memremap(msr.pfn << 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                 if (!*hvp)
108                         *hvp = __vmalloc(PAGE_SIZE, GFP_KERNEL | __GFP_ZERO);
109                 if (*hvp)
110                         msr.pfn = vmalloc_to_pfn(*hvp);
111
112         }
113         if (!WARN_ON(!(*hvp))) {
114                 msr.enable = 1;
115                 wrmsrl(HV_X64_MSR_VP_ASSIST_PAGE, msr.as_uint64);
116         }
117
118         return hyperv_init_ghcb();
119 }
120
121 static void (*hv_reenlightenment_cb)(void);
122
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
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
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
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
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
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
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
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
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
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
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
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  */
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                 /* Negotiate GHCB Version. */
408                 if (!hv_ghcb_negotiate_protocol())
409                         hv_ghcb_terminate(SEV_TERM_SET_GEN,
410                                           GHCB_SEV_ES_PROT_UNSUPPORTED);
411
412                 hv_ghcb_pg = alloc_percpu(union hv_ghcb *);
413                 if (!hv_ghcb_pg)
414                         goto free_vp_assist_page;
415         }
416
417         cpuhp = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "x86/hyperv_init:online",
418                                   hv_cpu_init, hv_cpu_die);
419         if (cpuhp < 0)
420                 goto free_ghcb_page;
421
422         /*
423          * Setup the hypercall page and enable hypercalls.
424          * 1. Register the guest ID
425          * 2. Enable the hypercall and register the hypercall page
426          */
427         guest_id = hv_generate_guest_id(LINUX_VERSION_CODE);
428         wrmsrl(HV_X64_MSR_GUEST_OS_ID, guest_id);
429
430         /* Hyper-V requires to write guest os id via ghcb in SNP IVM. */
431         hv_ghcb_msr_write(HV_X64_MSR_GUEST_OS_ID, guest_id);
432
433         hv_hypercall_pg = __vmalloc_node_range(PAGE_SIZE, 1, VMALLOC_START,
434                         VMALLOC_END, GFP_KERNEL, PAGE_KERNEL_ROX,
435                         VM_FLUSH_RESET_PERMS, NUMA_NO_NODE,
436                         __builtin_return_address(0));
437         if (hv_hypercall_pg == NULL)
438                 goto clean_guest_os_id;
439
440         rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
441         hypercall_msr.enable = 1;
442
443         if (hv_root_partition) {
444                 struct page *pg;
445                 void *src;
446
447                 /*
448                  * For the root partition, the hypervisor will set up its
449                  * hypercall page. The hypervisor guarantees it will not show
450                  * up in the root's address space. The root can't change the
451                  * location of the hypercall page.
452                  *
453                  * Order is important here. We must enable the hypercall page
454                  * so it is populated with code, then copy the code to an
455                  * executable page.
456                  */
457                 wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
458
459                 pg = vmalloc_to_page(hv_hypercall_pg);
460                 src = memremap(hypercall_msr.guest_physical_address << PAGE_SHIFT, PAGE_SIZE,
461                                 MEMREMAP_WB);
462                 BUG_ON(!src);
463                 memcpy_to_page(pg, 0, src, HV_HYP_PAGE_SIZE);
464                 memunmap(src);
465         } else {
466                 hypercall_msr.guest_physical_address = vmalloc_to_pfn(hv_hypercall_pg);
467                 wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
468         }
469
470         /*
471          * hyperv_init() is called before LAPIC is initialized: see
472          * apic_intr_mode_init() -> x86_platform.apic_post_init() and
473          * apic_bsp_setup() -> setup_local_APIC(). The direct-mode STIMER
474          * depends on LAPIC, so hv_stimer_alloc() should be called from
475          * x86_init.timers.setup_percpu_clockev.
476          */
477         old_setup_percpu_clockev = x86_init.timers.setup_percpu_clockev;
478         x86_init.timers.setup_percpu_clockev = hv_stimer_setup_percpu_clockev;
479
480         hv_apic_init();
481
482         x86_init.pci.arch_init = hv_pci_init;
483
484         register_syscore_ops(&hv_syscore_ops);
485
486         hyperv_init_cpuhp = cpuhp;
487
488         if (cpuid_ebx(HYPERV_CPUID_FEATURES) & HV_ACCESS_PARTITION_ID)
489                 hv_get_partition_id();
490
491         BUG_ON(hv_root_partition && hv_current_partition_id == ~0ull);
492
493 #ifdef CONFIG_PCI_MSI
494         /*
495          * If we're running as root, we want to create our own PCI MSI domain.
496          * We can't set this in hv_pci_init because that would be too late.
497          */
498         if (hv_root_partition)
499                 x86_init.irqs.create_pci_msi_domain = hv_create_pci_msi_domain;
500 #endif
501
502         /* Query the VMs extended capability once, so that it can be cached. */
503         hv_query_ext_cap(0);
504
505 #ifdef CONFIG_SWIOTLB
506         /*
507          * Swiotlb bounce buffer needs to be mapped in extra address
508          * space. Map function doesn't work in the early place and so
509          * call swiotlb_update_mem_attributes() here.
510          */
511         if (hv_is_isolation_supported())
512                 swiotlb_update_mem_attributes();
513 #endif
514
515         return;
516
517 clean_guest_os_id:
518         wrmsrl(HV_X64_MSR_GUEST_OS_ID, 0);
519         hv_ghcb_msr_write(HV_X64_MSR_GUEST_OS_ID, 0);
520         cpuhp_remove_state(cpuhp);
521 free_ghcb_page:
522         free_percpu(hv_ghcb_pg);
523 free_vp_assist_page:
524         kfree(hv_vp_assist_page);
525         hv_vp_assist_page = NULL;
526 common_free:
527         hv_common_free();
528 }
529
530 /*
531  * This routine is called before kexec/kdump, it does the required cleanup.
532  */
533 void hyperv_cleanup(void)
534 {
535         union hv_x64_msr_hypercall_contents hypercall_msr;
536         union hv_reference_tsc_msr tsc_msr;
537
538         unregister_syscore_ops(&hv_syscore_ops);
539
540         /* Reset our OS id */
541         wrmsrl(HV_X64_MSR_GUEST_OS_ID, 0);
542         hv_ghcb_msr_write(HV_X64_MSR_GUEST_OS_ID, 0);
543
544         /*
545          * Reset hypercall page reference before reset the page,
546          * let hypercall operations fail safely rather than
547          * panic the kernel for using invalid hypercall page
548          */
549         hv_hypercall_pg = NULL;
550
551         /* Reset the hypercall page */
552         hypercall_msr.as_uint64 = hv_get_register(HV_X64_MSR_HYPERCALL);
553         hypercall_msr.enable = 0;
554         hv_set_register(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
555
556         /* Reset the TSC page */
557         tsc_msr.as_uint64 = hv_get_register(HV_X64_MSR_REFERENCE_TSC);
558         tsc_msr.enable = 0;
559         hv_set_register(HV_X64_MSR_REFERENCE_TSC, tsc_msr.as_uint64);
560 }
561
562 void hyperv_report_panic(struct pt_regs *regs, long err, bool in_die)
563 {
564         static bool panic_reported;
565         u64 guest_id;
566
567         if (in_die && !panic_on_oops)
568                 return;
569
570         /*
571          * We prefer to report panic on 'die' chain as we have proper
572          * registers to report, but if we miss it (e.g. on BUG()) we need
573          * to report it on 'panic'.
574          */
575         if (panic_reported)
576                 return;
577         panic_reported = true;
578
579         rdmsrl(HV_X64_MSR_GUEST_OS_ID, guest_id);
580
581         wrmsrl(HV_X64_MSR_CRASH_P0, err);
582         wrmsrl(HV_X64_MSR_CRASH_P1, guest_id);
583         wrmsrl(HV_X64_MSR_CRASH_P2, regs->ip);
584         wrmsrl(HV_X64_MSR_CRASH_P3, regs->ax);
585         wrmsrl(HV_X64_MSR_CRASH_P4, regs->sp);
586
587         /*
588          * Let Hyper-V know there is crash data available
589          */
590         wrmsrl(HV_X64_MSR_CRASH_CTL, HV_CRASH_CTL_CRASH_NOTIFY);
591 }
592 EXPORT_SYMBOL_GPL(hyperv_report_panic);
593
594 bool hv_is_hyperv_initialized(void)
595 {
596         union hv_x64_msr_hypercall_contents hypercall_msr;
597
598         /*
599          * Ensure that we're really on Hyper-V, and not a KVM or Xen
600          * emulation of Hyper-V
601          */
602         if (x86_hyper_type != X86_HYPER_MS_HYPERV)
603                 return false;
604
605         /*
606          * Verify that earlier initialization succeeded by checking
607          * that the hypercall page is setup
608          */
609         hypercall_msr.as_uint64 = 0;
610         rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
611
612         return hypercall_msr.enable;
613 }
614 EXPORT_SYMBOL_GPL(hv_is_hyperv_initialized);