Linux Audio

Check our new training course

Loading...
v4.17
 
  1/*
  2 * KVM paravirt_ops implementation
  3 *
  4 * This program is free software; you can redistribute it and/or modify
  5 * it under the terms of the GNU General Public License as published by
  6 * the Free Software Foundation; either version 2 of the License, or
  7 * (at your option) any later version.
  8 *
  9 * This program is distributed in the hope that it will be useful,
 10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 12 * GNU General Public License for more details.
 13 *
 14 * You should have received a copy of the GNU General Public License
 15 * along with this program; if not, write to the Free Software
 16 * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
 17 *
 18 * Copyright (C) 2007, Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
 19 * Copyright IBM Corporation, 2007
 20 *   Authors: Anthony Liguori <aliguori@us.ibm.com>
 21 */
 22
 
 
 23#include <linux/context_tracking.h>
 24#include <linux/init.h>
 
 25#include <linux/kernel.h>
 26#include <linux/kvm_para.h>
 27#include <linux/cpu.h>
 28#include <linux/mm.h>
 29#include <linux/highmem.h>
 30#include <linux/hardirq.h>
 31#include <linux/notifier.h>
 32#include <linux/reboot.h>
 33#include <linux/hash.h>
 34#include <linux/sched.h>
 35#include <linux/slab.h>
 36#include <linux/kprobes.h>
 37#include <linux/debugfs.h>
 38#include <linux/nmi.h>
 39#include <linux/swait.h>
 
 40#include <asm/timer.h>
 41#include <asm/cpu.h>
 42#include <asm/traps.h>
 43#include <asm/desc.h>
 44#include <asm/tlbflush.h>
 45#include <asm/apic.h>
 46#include <asm/apicdef.h>
 47#include <asm/hypervisor.h>
 48#include <asm/kvm_guest.h>
 
 
 
 
 
 
 49
 50static int kvmapf = 1;
 51
 52static int __init parse_no_kvmapf(char *arg)
 53{
 54        kvmapf = 0;
 55        return 0;
 56}
 57
 58early_param("no-kvmapf", parse_no_kvmapf);
 59
 60static int steal_acc = 1;
 61static int __init parse_no_stealacc(char *arg)
 62{
 63        steal_acc = 0;
 64        return 0;
 65}
 66
 67early_param("no-steal-acc", parse_no_stealacc);
 68
 69static int kvmclock_vsyscall = 1;
 70static int __init parse_no_kvmclock_vsyscall(char *arg)
 71{
 72        kvmclock_vsyscall = 0;
 73        return 0;
 74}
 75
 76early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
 77
 78static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
 79static DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64);
 80static int has_steal_clock = 0;
 81
 82/*
 83 * No need for any "IO delay" on KVM
 84 */
 85static void kvm_io_delay(void)
 86{
 87}
 88
 89#define KVM_TASK_SLEEP_HASHBITS 8
 90#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
 91
 92struct kvm_task_sleep_node {
 93	struct hlist_node link;
 94	struct swait_queue_head wq;
 95	u32 token;
 96	int cpu;
 97	bool halted;
 98};
 99
100static struct kvm_task_sleep_head {
101	raw_spinlock_t lock;
102	struct hlist_head list;
103} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
104
105static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
106						  u32 token)
107{
108	struct hlist_node *p;
109
110	hlist_for_each(p, &b->list) {
111		struct kvm_task_sleep_node *n =
112			hlist_entry(p, typeof(*n), link);
113		if (n->token == token)
114			return n;
115	}
116
117	return NULL;
118}
119
120/*
121 * @interrupt_kernel: Is this called from a routine which interrupts the kernel
122 * 		      (other than user space)?
123 */
124void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
125{
126	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
127	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
128	struct kvm_task_sleep_node n, *e;
129	DECLARE_SWAITQUEUE(wait);
130
131	rcu_irq_enter();
132
133	raw_spin_lock(&b->lock);
134	e = _find_apf_task(b, token);
135	if (e) {
136		/* dummy entry exist -> wake up was delivered ahead of PF */
137		hlist_del(&e->link);
138		kfree(e);
139		raw_spin_unlock(&b->lock);
140
141		rcu_irq_exit();
142		return;
143	}
144
145	n.token = token;
146	n.cpu = smp_processor_id();
147	n.halted = is_idle_task(current) ||
148		   (IS_ENABLED(CONFIG_PREEMPT_COUNT)
149		    ? preempt_count() > 1 || rcu_preempt_depth()
150		    : interrupt_kernel);
151	init_swait_queue_head(&n.wq);
152	hlist_add_head(&n.link, &b->list);
153	raw_spin_unlock(&b->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
154
155	for (;;) {
156		if (!n.halted)
157			prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
158		if (hlist_unhashed(&n.link))
159			break;
160
161		rcu_irq_exit();
162
163		if (!n.halted) {
164			local_irq_enable();
165			schedule();
166			local_irq_disable();
167		} else {
168			/*
169			 * We cannot reschedule. So halt.
170			 */
171			native_safe_halt();
172			local_irq_disable();
173		}
174
175		rcu_irq_enter();
176	}
177	if (!n.halted)
178		finish_swait(&n.wq, &wait);
179
180	rcu_irq_exit();
181	return;
182}
183EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
184
185static void apf_task_wake_one(struct kvm_task_sleep_node *n)
186{
187	hlist_del_init(&n->link);
188	if (n->halted)
189		smp_send_reschedule(n->cpu);
190	else if (swq_has_sleeper(&n->wq))
191		swake_up(&n->wq);
192}
193
194static void apf_task_wake_all(void)
195{
196	int i;
197
198	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
199		struct hlist_node *p, *next;
200		struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 
 
 
201		raw_spin_lock(&b->lock);
202		hlist_for_each_safe(p, next, &b->list) {
203			struct kvm_task_sleep_node *n =
204				hlist_entry(p, typeof(*n), link);
205			if (n->cpu == smp_processor_id())
206				apf_task_wake_one(n);
207		}
208		raw_spin_unlock(&b->lock);
209	}
210}
211
212void kvm_async_pf_task_wake(u32 token)
213{
214	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
215	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
216	struct kvm_task_sleep_node *n;
217
218	if (token == ~0) {
219		apf_task_wake_all();
220		return;
221	}
222
223again:
224	raw_spin_lock(&b->lock);
225	n = _find_apf_task(b, token);
226	if (!n) {
227		/*
228		 * async PF was not yet handled.
229		 * Add dummy entry for the token.
230		 */
231		n = kzalloc(sizeof(*n), GFP_ATOMIC);
232		if (!n) {
233			/*
234			 * Allocation failed! Busy wait while other cpu
235			 * handles async PF.
236			 */
237			raw_spin_unlock(&b->lock);
238			cpu_relax();
239			goto again;
240		}
241		n->token = token;
242		n->cpu = smp_processor_id();
243		init_swait_queue_head(&n->wq);
244		hlist_add_head(&n->link, &b->list);
245	} else
246		apf_task_wake_one(n);
 
247	raw_spin_unlock(&b->lock);
248	return;
249}
250EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
251
252u32 kvm_read_and_reset_pf_reason(void)
253{
254	u32 reason = 0;
255
256	if (__this_cpu_read(apf_reason.enabled)) {
257		reason = __this_cpu_read(apf_reason.reason);
258		__this_cpu_write(apf_reason.reason, 0);
259	}
260
261	return reason;
262}
263EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
264NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
265
266dotraplinkage void
267do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
268{
269	enum ctx_state prev_state;
 
270
271	switch (kvm_read_and_reset_pf_reason()) {
272	default:
273		do_page_fault(regs, error_code);
274		break;
275	case KVM_PV_REASON_PAGE_NOT_PRESENT:
276		/* page is swapped out by the host. */
277		prev_state = exception_enter();
278		kvm_async_pf_task_wait((u32)read_cr2(), !user_mode(regs));
279		exception_exit(prev_state);
280		break;
281	case KVM_PV_REASON_PAGE_READY:
282		rcu_irq_enter();
283		kvm_async_pf_task_wake((u32)read_cr2());
284		rcu_irq_exit();
285		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
286	}
 
 
287}
288NOKPROBE_SYMBOL(do_async_page_fault);
289
290static void __init paravirt_ops_setup(void)
291{
292	pv_info.name = "KVM";
293
294	if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
295		pv_cpu_ops.io_delay = kvm_io_delay;
296
297#ifdef CONFIG_X86_IO_APIC
298	no_timer_check = 1;
299#endif
300}
301
302static void kvm_register_steal_time(void)
303{
304	int cpu = smp_processor_id();
305	struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
306
307	if (!has_steal_clock)
308		return;
309
310	wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
311	pr_info("kvm-stealtime: cpu %d, msr %llx\n",
312		cpu, (unsigned long long) slow_virt_to_phys(st));
313}
314
315static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
316
317static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
318{
319	/**
320	 * This relies on __test_and_clear_bit to modify the memory
321	 * in a way that is atomic with respect to the local CPU.
322	 * The hypervisor only accesses this memory from the local CPU so
323	 * there's no need for lock or memory barriers.
324	 * An optimization barrier is implied in apic write.
325	 */
326	if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
327		return;
328	apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
329}
330
331static void kvm_guest_cpu_init(void)
332{
333	if (!kvm_para_available())
334		return;
335
336	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
337		u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
338
339#ifdef CONFIG_PREEMPT
340		pa |= KVM_ASYNC_PF_SEND_ALWAYS;
341#endif
342		pa |= KVM_ASYNC_PF_ENABLED;
343
344		if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
345			pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
346
 
 
347		wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
348		__this_cpu_write(apf_reason.enabled, 1);
349		printk(KERN_INFO"KVM setup async PF for cpu %d\n",
350		       smp_processor_id());
351	}
352
353	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
354		unsigned long pa;
 
355		/* Size alignment is implied but just to make it explicit. */
356		BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
357		__this_cpu_write(kvm_apic_eoi, 0);
358		pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
359			| KVM_MSR_ENABLED;
360		wrmsrl(MSR_KVM_PV_EOI_EN, pa);
361	}
362
363	if (has_steal_clock)
364		kvm_register_steal_time();
365}
366
367static void kvm_pv_disable_apf(void)
368{
369	if (!__this_cpu_read(apf_reason.enabled))
370		return;
371
372	wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
373	__this_cpu_write(apf_reason.enabled, 0);
374
375	printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
376	       smp_processor_id());
377}
378
379static void kvm_pv_guest_cpu_reboot(void *unused)
380{
381	/*
382	 * We disable PV EOI before we load a new kernel by kexec,
383	 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
384	 * New kernel can re-enable when it boots.
385	 */
386	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
387		wrmsrl(MSR_KVM_PV_EOI_EN, 0);
388	kvm_pv_disable_apf();
389	kvm_disable_steal_time();
390}
391
392static int kvm_pv_reboot_notify(struct notifier_block *nb,
393				unsigned long code, void *unused)
394{
395	if (code == SYS_RESTART)
396		on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
397	return NOTIFY_DONE;
398}
399
400static struct notifier_block kvm_pv_reboot_nb = {
401	.notifier_call = kvm_pv_reboot_notify,
402};
403
404static u64 kvm_steal_clock(int cpu)
405{
406	u64 steal;
407	struct kvm_steal_time *src;
408	int version;
409
410	src = &per_cpu(steal_time, cpu);
411	do {
412		version = src->version;
413		virt_rmb();
414		steal = src->steal;
415		virt_rmb();
416	} while ((version & 1) || (version != src->version));
417
418	return steal;
419}
420
421void kvm_disable_steal_time(void)
422{
423	if (!has_steal_clock)
424		return;
425
426	wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
427}
428
429static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
430{
431	early_set_memory_decrypted((unsigned long) ptr, size);
432}
433
434/*
435 * Iterate through all possible CPUs and map the memory region pointed
436 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
437 *
438 * Note: we iterate through all possible CPUs to ensure that CPUs
439 * hotplugged will have their per-cpu variable already mapped as
440 * decrypted.
441 */
442static void __init sev_map_percpu_data(void)
443{
444	int cpu;
445
446	if (!sev_active())
447		return;
448
449	for_each_possible_cpu(cpu) {
450		__set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
451		__set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
452		__set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
453	}
454}
455
456#ifdef CONFIG_SMP
457static void __init kvm_smp_prepare_cpus(unsigned int max_cpus)
458{
459	native_smp_prepare_cpus(max_cpus);
460	if (kvm_para_has_hint(KVM_HINTS_REALTIME))
461		static_branch_disable(&virt_spin_lock_key);
 
 
 
 
462}
463
464static void __init kvm_smp_prepare_boot_cpu(void)
465{
466	/*
467	 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
468	 * shares the guest physical address with the hypervisor.
469	 */
470	sev_map_percpu_data();
471
 
472	kvm_guest_cpu_init();
473	native_smp_prepare_boot_cpu();
474	kvm_spinlock_init();
475}
476
477static void kvm_guest_cpu_offline(void)
 
 
 
 
478{
479	kvm_disable_steal_time();
480	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
481		wrmsrl(MSR_KVM_PV_EOI_EN, 0);
482	kvm_pv_disable_apf();
483	apf_task_wake_all();
484}
485
486static int kvm_cpu_online(unsigned int cpu)
487{
488	local_irq_disable();
489	kvm_guest_cpu_init();
490	local_irq_enable();
491	return 0;
492}
493
494static int kvm_cpu_down_prepare(unsigned int cpu)
495{
496	local_irq_disable();
497	kvm_guest_cpu_offline();
498	local_irq_enable();
499	return 0;
500}
 
 
 
 
 
 
 
 
 
 
 
 
501#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
502
503static void __init kvm_apf_trap_init(void)
 
 
 
504{
505	update_intr_gate(X86_TRAP_PF, async_page_fault);
 
 
506}
507
508static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
 
 
 
 
 
 
 
 
 
 
 
 
 
509
510static void kvm_flush_tlb_others(const struct cpumask *cpumask,
511			const struct flush_tlb_info *info)
512{
513	u8 state;
514	int cpu;
515	struct kvm_steal_time *src;
516	struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_tlb_mask);
517
518	cpumask_copy(flushmask, cpumask);
519	/*
520	 * We have to call flush only on online vCPUs. And
521	 * queue flush_on_enter for pre-empted vCPUs
522	 */
523	for_each_cpu(cpu, flushmask) {
 
 
 
 
 
524		src = &per_cpu(steal_time, cpu);
525		state = READ_ONCE(src->preempted);
526		if ((state & KVM_VCPU_PREEMPTED)) {
527			if (try_cmpxchg(&src->preempted, &state,
528					state | KVM_VCPU_FLUSH_TLB))
529				__cpumask_clear_cpu(cpu, flushmask);
530		}
531	}
532
533	native_flush_tlb_others(flushmask, info);
534}
535
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
536static void __init kvm_guest_init(void)
537{
538	int i;
539
540	if (!kvm_para_available())
541		return;
542
543	paravirt_ops_setup();
544	register_reboot_notifier(&kvm_pv_reboot_nb);
545	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
546		raw_spin_lock_init(&async_pf_sleepers[i].lock);
547	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
548		x86_init.irqs.trap_init = kvm_apf_trap_init;
549
550	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
551		has_steal_clock = 1;
552		pv_time_ops.steal_clock = kvm_steal_clock;
553	}
554
555	if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
556	    !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
557	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME))
558		pv_mmu_ops.flush_tlb_others = kvm_flush_tlb_others;
559
560	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
561		apic_set_eoi_write(kvm_guest_apic_eoi_write);
562
563	if (kvmclock_vsyscall)
564		kvm_setup_vsyscall_timeinfo();
 
 
565
566#ifdef CONFIG_SMP
567	smp_ops.smp_prepare_cpus = kvm_smp_prepare_cpus;
 
 
 
 
 
568	smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 
 
 
 
569	if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
570				      kvm_cpu_online, kvm_cpu_down_prepare) < 0)
571		pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
572#else
573	sev_map_percpu_data();
574	kvm_guest_cpu_init();
575#endif
576
 
 
 
 
 
 
577	/*
578	 * Hard lockup detection is enabled by default. Disable it, as guests
579	 * can get false positives too easily, for example if the host is
580	 * overcommitted.
581	 */
582	hardlockup_detector_disable();
583}
584
585static noinline uint32_t __kvm_cpuid_base(void)
586{
587	if (boot_cpu_data.cpuid_level < 0)
588		return 0;	/* So we don't blow up on old processors */
589
590	if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
591		return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
592
593	return 0;
594}
595
596static inline uint32_t kvm_cpuid_base(void)
597{
598	static int kvm_cpuid_base = -1;
599
600	if (kvm_cpuid_base == -1)
601		kvm_cpuid_base = __kvm_cpuid_base();
602
603	return kvm_cpuid_base;
604}
605
606bool kvm_para_available(void)
607{
608	return kvm_cpuid_base() != 0;
609}
610EXPORT_SYMBOL_GPL(kvm_para_available);
611
612unsigned int kvm_arch_para_features(void)
613{
614	return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
615}
616
617unsigned int kvm_arch_para_hints(void)
618{
619	return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
620}
 
621
622static uint32_t __init kvm_detect(void)
623{
624	return kvm_cpuid_base();
625}
626
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
627const __initconst struct hypervisor_x86 x86_hyper_kvm = {
628	.name			= "KVM",
629	.detect			= kvm_detect,
630	.type			= X86_HYPER_KVM,
631	.init.guest_late_init	= kvm_guest_init,
632	.init.x2apic_available	= kvm_para_available,
 
 
 
 
 
 
633};
634
635static __init int activate_jump_labels(void)
636{
637	if (has_steal_clock) {
638		static_key_slow_inc(&paravirt_steal_enabled);
639		if (steal_acc)
640			static_key_slow_inc(&paravirt_steal_rq_enabled);
641	}
642
643	return 0;
644}
645arch_initcall(activate_jump_labels);
646
647static __init int kvm_setup_pv_tlb_flush(void)
648{
649	int cpu;
650
651	if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
652	    !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
653	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
654		for_each_possible_cpu(cpu) {
655			zalloc_cpumask_var_node(per_cpu_ptr(&__pv_tlb_mask, cpu),
656				GFP_KERNEL, cpu_to_node(cpu));
657		}
658		pr_info("KVM setup pv remote TLB flush\n");
659	}
660
661	return 0;
662}
663arch_initcall(kvm_setup_pv_tlb_flush);
664
665#ifdef CONFIG_PARAVIRT_SPINLOCKS
666
667/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
668static void kvm_kick_cpu(int cpu)
669{
670	int apicid;
671	unsigned long flags = 0;
672
673	apicid = per_cpu(x86_cpu_to_apicid, cpu);
674	kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
675}
676
677#include <asm/qspinlock.h>
678
679static void kvm_wait(u8 *ptr, u8 val)
680{
681	unsigned long flags;
682
683	if (in_nmi())
684		return;
685
686	local_irq_save(flags);
687
688	if (READ_ONCE(*ptr) != val)
689		goto out;
690
691	/*
692	 * halt until it's our turn and kicked. Note that we do safe halt
693	 * for irq enabled case to avoid hang when lock info is overwritten
694	 * in irq spinlock slowpath and no spurious interrupt occur to save us.
695	 */
696	if (arch_irqs_disabled_flags(flags))
697		halt();
698	else
699		safe_halt();
 
700
701out:
702	local_irq_restore(flags);
 
 
 
703}
704
705#ifdef CONFIG_X86_32
706__visible bool __kvm_vcpu_is_preempted(long cpu)
707{
708	struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
709
710	return !!(src->preempted & KVM_VCPU_PREEMPTED);
711}
712PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
713
714#else
715
716#include <asm/asm-offsets.h>
717
718extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
719
720/*
721 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
722 * restoring to/from the stack.
723 */
724asm(
725".pushsection .text;"
726".global __raw_callee_save___kvm_vcpu_is_preempted;"
727".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
728"__raw_callee_save___kvm_vcpu_is_preempted:"
729"movq	__per_cpu_offset(,%rdi,8), %rax;"
730"cmpb	$0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
731"setne	%al;"
732"ret;"
 
733".popsection");
734
735#endif
736
737/*
738 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
739 */
740void __init kvm_spinlock_init(void)
741{
742	if (!kvm_para_available())
743		return;
744	/* Does host kernel support KVM_FEATURE_PV_UNHALT? */
745	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 
 
 
746		return;
 
747
748	if (kvm_para_has_hint(KVM_HINTS_REALTIME))
749		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
750
751	__pv_init_lock_hash();
752	pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
753	pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
754	pv_lock_ops.wait = kvm_wait;
755	pv_lock_ops.kick = kvm_kick_cpu;
 
756
757	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
758		pv_lock_ops.vcpu_is_preempted =
759			PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
760	}
 
 
 
 
 
 
 
761}
762
763#endif	/* CONFIG_PARAVIRT_SPINLOCKS */
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * KVM paravirt_ops implementation
   4 *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
   5 * Copyright (C) 2007, Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
   6 * Copyright IBM Corporation, 2007
   7 *   Authors: Anthony Liguori <aliguori@us.ibm.com>
   8 */
   9
  10#define pr_fmt(fmt) "kvm-guest: " fmt
  11
  12#include <linux/context_tracking.h>
  13#include <linux/init.h>
  14#include <linux/irq.h>
  15#include <linux/kernel.h>
  16#include <linux/kvm_para.h>
  17#include <linux/cpu.h>
  18#include <linux/mm.h>
  19#include <linux/highmem.h>
  20#include <linux/hardirq.h>
  21#include <linux/notifier.h>
  22#include <linux/reboot.h>
  23#include <linux/hash.h>
  24#include <linux/sched.h>
  25#include <linux/slab.h>
  26#include <linux/kprobes.h>
 
  27#include <linux/nmi.h>
  28#include <linux/swait.h>
  29#include <linux/syscore_ops.h>
  30#include <asm/timer.h>
  31#include <asm/cpu.h>
  32#include <asm/traps.h>
  33#include <asm/desc.h>
  34#include <asm/tlbflush.h>
  35#include <asm/apic.h>
  36#include <asm/apicdef.h>
  37#include <asm/hypervisor.h>
  38#include <asm/tlb.h>
  39#include <asm/cpuidle_haltpoll.h>
  40#include <asm/ptrace.h>
  41#include <asm/reboot.h>
  42#include <asm/svm.h>
  43
  44DEFINE_STATIC_KEY_FALSE(kvm_async_pf_enabled);
  45
  46static int kvmapf = 1;
  47
  48static int __init parse_no_kvmapf(char *arg)
  49{
  50        kvmapf = 0;
  51        return 0;
  52}
  53
  54early_param("no-kvmapf", parse_no_kvmapf);
  55
  56static int steal_acc = 1;
  57static int __init parse_no_stealacc(char *arg)
  58{
  59        steal_acc = 0;
  60        return 0;
  61}
  62
  63early_param("no-steal-acc", parse_no_stealacc);
  64
 
 
 
 
 
 
 
 
 
  65static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  66DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
  67static int has_steal_clock = 0;
  68
  69/*
  70 * No need for any "IO delay" on KVM
  71 */
  72static void kvm_io_delay(void)
  73{
  74}
  75
  76#define KVM_TASK_SLEEP_HASHBITS 8
  77#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  78
  79struct kvm_task_sleep_node {
  80	struct hlist_node link;
  81	struct swait_queue_head wq;
  82	u32 token;
  83	int cpu;
 
  84};
  85
  86static struct kvm_task_sleep_head {
  87	raw_spinlock_t lock;
  88	struct hlist_head list;
  89} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
  90
  91static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
  92						  u32 token)
  93{
  94	struct hlist_node *p;
  95
  96	hlist_for_each(p, &b->list) {
  97		struct kvm_task_sleep_node *n =
  98			hlist_entry(p, typeof(*n), link);
  99		if (n->token == token)
 100			return n;
 101	}
 102
 103	return NULL;
 104}
 105
 106static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n)
 
 
 
 
 107{
 108	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 109	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 110	struct kvm_task_sleep_node *e;
 
 
 
 111
 112	raw_spin_lock(&b->lock);
 113	e = _find_apf_task(b, token);
 114	if (e) {
 115		/* dummy entry exist -> wake up was delivered ahead of PF */
 116		hlist_del(&e->link);
 
 117		raw_spin_unlock(&b->lock);
 118		kfree(e);
 119		return false;
 
 120	}
 121
 122	n->token = token;
 123	n->cpu = smp_processor_id();
 124	init_swait_queue_head(&n->wq);
 125	hlist_add_head(&n->link, &b->list);
 
 
 
 
 126	raw_spin_unlock(&b->lock);
 127	return true;
 128}
 129
 130/*
 131 * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
 132 * @token:	Token to identify the sleep node entry
 133 *
 134 * Invoked from the async pagefault handling code or from the VM exit page
 135 * fault handler. In both cases RCU is watching.
 136 */
 137void kvm_async_pf_task_wait_schedule(u32 token)
 138{
 139	struct kvm_task_sleep_node n;
 140	DECLARE_SWAITQUEUE(wait);
 141
 142	lockdep_assert_irqs_disabled();
 143
 144	if (!kvm_async_pf_queue_task(token, &n))
 145		return;
 146
 147	for (;;) {
 148		prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 
 149		if (hlist_unhashed(&n.link))
 150			break;
 151
 152		local_irq_enable();
 153		schedule();
 154		local_irq_disable();
 
 
 
 
 
 
 
 
 
 
 
 
 155	}
 156	finish_swait(&n.wq, &wait);
 
 
 
 
 157}
 158EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule);
 159
 160static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 161{
 162	hlist_del_init(&n->link);
 163	if (swq_has_sleeper(&n->wq))
 164		swake_up_one(&n->wq);
 
 
 165}
 166
 167static void apf_task_wake_all(void)
 168{
 169	int i;
 170
 171	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 
 172		struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 173		struct kvm_task_sleep_node *n;
 174		struct hlist_node *p, *next;
 175
 176		raw_spin_lock(&b->lock);
 177		hlist_for_each_safe(p, next, &b->list) {
 178			n = hlist_entry(p, typeof(*n), link);
 
 179			if (n->cpu == smp_processor_id())
 180				apf_task_wake_one(n);
 181		}
 182		raw_spin_unlock(&b->lock);
 183	}
 184}
 185
 186void kvm_async_pf_task_wake(u32 token)
 187{
 188	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 189	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 190	struct kvm_task_sleep_node *n;
 191
 192	if (token == ~0) {
 193		apf_task_wake_all();
 194		return;
 195	}
 196
 197again:
 198	raw_spin_lock(&b->lock);
 199	n = _find_apf_task(b, token);
 200	if (!n) {
 201		/*
 202		 * async PF was not yet handled.
 203		 * Add dummy entry for the token.
 204		 */
 205		n = kzalloc(sizeof(*n), GFP_ATOMIC);
 206		if (!n) {
 207			/*
 208			 * Allocation failed! Busy wait while other cpu
 209			 * handles async PF.
 210			 */
 211			raw_spin_unlock(&b->lock);
 212			cpu_relax();
 213			goto again;
 214		}
 215		n->token = token;
 216		n->cpu = smp_processor_id();
 217		init_swait_queue_head(&n->wq);
 218		hlist_add_head(&n->link, &b->list);
 219	} else {
 220		apf_task_wake_one(n);
 221	}
 222	raw_spin_unlock(&b->lock);
 223	return;
 224}
 225EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 226
 227noinstr u32 kvm_read_and_reset_apf_flags(void)
 228{
 229	u32 flags = 0;
 230
 231	if (__this_cpu_read(apf_reason.enabled)) {
 232		flags = __this_cpu_read(apf_reason.flags);
 233		__this_cpu_write(apf_reason.flags, 0);
 234	}
 235
 236	return flags;
 237}
 238EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
 
 239
 240noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
 
 241{
 242	u32 flags = kvm_read_and_reset_apf_flags();
 243	irqentry_state_t state;
 244
 245	if (!flags)
 246		return false;
 247
 248	state = irqentry_enter(regs);
 249	instrumentation_begin();
 250
 251	/*
 252	 * If the host managed to inject an async #PF into an interrupt
 253	 * disabled region, then die hard as this is not going to end well
 254	 * and the host side is seriously broken.
 255	 */
 256	if (unlikely(!(regs->flags & X86_EFLAGS_IF)))
 257		panic("Host injected async #PF in interrupt disabled region\n");
 258
 259	if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
 260		if (unlikely(!(user_mode(regs))))
 261			panic("Host injected async #PF in kernel mode\n");
 262		/* Page is swapped out by the host. */
 263		kvm_async_pf_task_wait_schedule(token);
 264	} else {
 265		WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
 266	}
 267
 268	instrumentation_end();
 269	irqentry_exit(regs, state);
 270	return true;
 271}
 272
 273DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
 274{
 275	struct pt_regs *old_regs = set_irq_regs(regs);
 276	u32 token;
 277
 278	ack_APIC_irq();
 279
 280	inc_irq_stat(irq_hv_callback_count);
 281
 282	if (__this_cpu_read(apf_reason.enabled)) {
 283		token = __this_cpu_read(apf_reason.token);
 284		kvm_async_pf_task_wake(token);
 285		__this_cpu_write(apf_reason.token, 0);
 286		wrmsrl(MSR_KVM_ASYNC_PF_ACK, 1);
 287	}
 288
 289	set_irq_regs(old_regs);
 290}
 
 291
 292static void __init paravirt_ops_setup(void)
 293{
 294	pv_info.name = "KVM";
 295
 296	if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
 297		pv_ops.cpu.io_delay = kvm_io_delay;
 298
 299#ifdef CONFIG_X86_IO_APIC
 300	no_timer_check = 1;
 301#endif
 302}
 303
 304static void kvm_register_steal_time(void)
 305{
 306	int cpu = smp_processor_id();
 307	struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
 308
 309	if (!has_steal_clock)
 310		return;
 311
 312	wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
 313	pr_info("stealtime: cpu %d, msr %llx\n", cpu,
 314		(unsigned long long) slow_virt_to_phys(st));
 315}
 316
 317static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
 318
 319static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
 320{
 321	/**
 322	 * This relies on __test_and_clear_bit to modify the memory
 323	 * in a way that is atomic with respect to the local CPU.
 324	 * The hypervisor only accesses this memory from the local CPU so
 325	 * there's no need for lock or memory barriers.
 326	 * An optimization barrier is implied in apic write.
 327	 */
 328	if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
 329		return;
 330	apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
 331}
 332
 333static void kvm_guest_cpu_init(void)
 334{
 335	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
 
 
 
 336		u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
 337
 338		WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled));
 339
 340		pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
 341		pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT;
 342
 343		if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
 344			pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
 345
 346		wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
 347
 348		wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
 349		__this_cpu_write(apf_reason.enabled, 1);
 350		pr_info("setup async PF for cpu %d\n", smp_processor_id());
 
 351	}
 352
 353	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 354		unsigned long pa;
 355
 356		/* Size alignment is implied but just to make it explicit. */
 357		BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 358		__this_cpu_write(kvm_apic_eoi, 0);
 359		pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
 360			| KVM_MSR_ENABLED;
 361		wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 362	}
 363
 364	if (has_steal_clock)
 365		kvm_register_steal_time();
 366}
 367
 368static void kvm_pv_disable_apf(void)
 369{
 370	if (!__this_cpu_read(apf_reason.enabled))
 371		return;
 372
 373	wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 374	__this_cpu_write(apf_reason.enabled, 0);
 375
 376	pr_info("disable async PF for cpu %d\n", smp_processor_id());
 
 377}
 378
 379static void kvm_disable_steal_time(void)
 380{
 381	if (!has_steal_clock)
 382		return;
 
 
 
 
 
 
 
 
 383
 384	wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 
 
 
 
 
 385}
 386
 
 
 
 
 387static u64 kvm_steal_clock(int cpu)
 388{
 389	u64 steal;
 390	struct kvm_steal_time *src;
 391	int version;
 392
 393	src = &per_cpu(steal_time, cpu);
 394	do {
 395		version = src->version;
 396		virt_rmb();
 397		steal = src->steal;
 398		virt_rmb();
 399	} while ((version & 1) || (version != src->version));
 400
 401	return steal;
 402}
 403
 
 
 
 
 
 
 
 
 404static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
 405{
 406	early_set_memory_decrypted((unsigned long) ptr, size);
 407}
 408
 409/*
 410 * Iterate through all possible CPUs and map the memory region pointed
 411 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
 412 *
 413 * Note: we iterate through all possible CPUs to ensure that CPUs
 414 * hotplugged will have their per-cpu variable already mapped as
 415 * decrypted.
 416 */
 417static void __init sev_map_percpu_data(void)
 418{
 419	int cpu;
 420
 421	if (!sev_active())
 422		return;
 423
 424	for_each_possible_cpu(cpu) {
 425		__set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
 426		__set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
 427		__set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
 428	}
 429}
 430
 431static void kvm_guest_cpu_offline(bool shutdown)
 
 432{
 433	kvm_disable_steal_time();
 434	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 435		wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 436	kvm_pv_disable_apf();
 437	if (!shutdown)
 438		apf_task_wake_all();
 439	kvmclock_disable();
 440}
 441
 442static int kvm_cpu_online(unsigned int cpu)
 443{
 444	unsigned long flags;
 
 
 
 
 445
 446	local_irq_save(flags);
 447	kvm_guest_cpu_init();
 448	local_irq_restore(flags);
 449	return 0;
 450}
 451
 452#ifdef CONFIG_SMP
 453
 454static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
 455
 456static bool pv_tlb_flush_supported(void)
 457{
 458	return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
 459		!kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 460		kvm_para_has_feature(KVM_FEATURE_STEAL_TIME));
 
 
 461}
 462
 463static bool pv_ipi_supported(void)
 464{
 465	return kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI);
 
 
 
 466}
 467
 468static bool pv_sched_yield_supported(void)
 469{
 470	return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
 471		!kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 472	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME));
 
 473}
 474
 475#define KVM_IPI_CLUSTER_SIZE	(2 * BITS_PER_LONG)
 476
 477static void __send_ipi_mask(const struct cpumask *mask, int vector)
 478{
 479	unsigned long flags;
 480	int cpu, apic_id, icr;
 481	int min = 0, max = 0;
 482#ifdef CONFIG_X86_64
 483	__uint128_t ipi_bitmap = 0;
 484#else
 485	u64 ipi_bitmap = 0;
 486#endif
 487	long ret;
 488
 489	if (cpumask_empty(mask))
 490		return;
 491
 492	local_irq_save(flags);
 493
 494	switch (vector) {
 495	default:
 496		icr = APIC_DM_FIXED | vector;
 497		break;
 498	case NMI_VECTOR:
 499		icr = APIC_DM_NMI;
 500		break;
 501	}
 502
 503	for_each_cpu(cpu, mask) {
 504		apic_id = per_cpu(x86_cpu_to_apicid, cpu);
 505		if (!ipi_bitmap) {
 506			min = max = apic_id;
 507		} else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
 508			ipi_bitmap <<= min - apic_id;
 509			min = apic_id;
 510		} else if (apic_id < min + KVM_IPI_CLUSTER_SIZE) {
 511			max = apic_id < max ? max : apic_id;
 512		} else {
 513			ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
 514				(unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
 515			WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
 516				  ret);
 517			min = max = apic_id;
 518			ipi_bitmap = 0;
 519		}
 520		__set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
 521	}
 522
 523	if (ipi_bitmap) {
 524		ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
 525			(unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
 526		WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
 527			  ret);
 528	}
 529
 530	local_irq_restore(flags);
 531}
 532
 533static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
 534{
 535	__send_ipi_mask(mask, vector);
 536}
 537
 538static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
 539{
 540	unsigned int this_cpu = smp_processor_id();
 541	struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
 542	const struct cpumask *local_mask;
 543
 544	cpumask_copy(new_mask, mask);
 545	cpumask_clear_cpu(this_cpu, new_mask);
 546	local_mask = new_mask;
 547	__send_ipi_mask(local_mask, vector);
 548}
 549
 550/*
 551 * Set the IPI entry points
 552 */
 553static void kvm_setup_pv_ipi(void)
 554{
 555	apic->send_IPI_mask = kvm_send_ipi_mask;
 556	apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
 557	pr_info("setup PV IPIs\n");
 558}
 559
 560static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
 561{
 562	int cpu;
 563
 564	native_send_call_func_ipi(mask);
 565
 566	/* Make sure other vCPUs get a chance to run if they need to. */
 567	for_each_cpu(cpu, mask) {
 568		if (vcpu_is_preempted(cpu)) {
 569			kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
 570			break;
 571		}
 572	}
 573}
 574
 575static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
 576			const struct flush_tlb_info *info)
 577{
 578	u8 state;
 579	int cpu;
 580	struct kvm_steal_time *src;
 581	struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
 582
 583	cpumask_copy(flushmask, cpumask);
 584	/*
 585	 * We have to call flush only on online vCPUs. And
 586	 * queue flush_on_enter for pre-empted vCPUs
 587	 */
 588	for_each_cpu(cpu, flushmask) {
 589		/*
 590		 * The local vCPU is never preempted, so we do not explicitly
 591		 * skip check for local vCPU - it will never be cleared from
 592		 * flushmask.
 593		 */
 594		src = &per_cpu(steal_time, cpu);
 595		state = READ_ONCE(src->preempted);
 596		if ((state & KVM_VCPU_PREEMPTED)) {
 597			if (try_cmpxchg(&src->preempted, &state,
 598					state | KVM_VCPU_FLUSH_TLB))
 599				__cpumask_clear_cpu(cpu, flushmask);
 600		}
 601	}
 602
 603	native_flush_tlb_multi(flushmask, info);
 604}
 605
 606static __init int kvm_alloc_cpumask(void)
 607{
 608	int cpu;
 609
 610	if (!kvm_para_available() || nopv)
 611		return 0;
 612
 613	if (pv_tlb_flush_supported() || pv_ipi_supported())
 614		for_each_possible_cpu(cpu) {
 615			zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
 616				GFP_KERNEL, cpu_to_node(cpu));
 617		}
 618
 619	return 0;
 620}
 621arch_initcall(kvm_alloc_cpumask);
 622
 623static void __init kvm_smp_prepare_boot_cpu(void)
 624{
 625	/*
 626	 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
 627	 * shares the guest physical address with the hypervisor.
 628	 */
 629	sev_map_percpu_data();
 630
 631	kvm_guest_cpu_init();
 632	native_smp_prepare_boot_cpu();
 633	kvm_spinlock_init();
 634}
 635
 636static int kvm_cpu_down_prepare(unsigned int cpu)
 637{
 638	unsigned long flags;
 639
 640	local_irq_save(flags);
 641	kvm_guest_cpu_offline(false);
 642	local_irq_restore(flags);
 643	return 0;
 644}
 645
 646#endif
 647
 648static int kvm_suspend(void)
 649{
 650	kvm_guest_cpu_offline(false);
 651
 652	return 0;
 653}
 654
 655static void kvm_resume(void)
 656{
 657	kvm_cpu_online(raw_smp_processor_id());
 658}
 659
 660static struct syscore_ops kvm_syscore_ops = {
 661	.suspend	= kvm_suspend,
 662	.resume		= kvm_resume,
 663};
 664
 665static void kvm_pv_guest_cpu_reboot(void *unused)
 666{
 667	kvm_guest_cpu_offline(true);
 668}
 669
 670static int kvm_pv_reboot_notify(struct notifier_block *nb,
 671				unsigned long code, void *unused)
 672{
 673	if (code == SYS_RESTART)
 674		on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 675	return NOTIFY_DONE;
 676}
 677
 678static struct notifier_block kvm_pv_reboot_nb = {
 679	.notifier_call = kvm_pv_reboot_notify,
 680};
 681
 682/*
 683 * After a PV feature is registered, the host will keep writing to the
 684 * registered memory location. If the guest happens to shutdown, this memory
 685 * won't be valid. In cases like kexec, in which you install a new kernel, this
 686 * means a random memory location will be kept being written.
 687 */
 688#ifdef CONFIG_KEXEC_CORE
 689static void kvm_crash_shutdown(struct pt_regs *regs)
 690{
 691	kvm_guest_cpu_offline(true);
 692	native_machine_crash_shutdown(regs);
 693}
 694#endif
 695
 696static void __init kvm_guest_init(void)
 697{
 698	int i;
 699
 
 
 
 700	paravirt_ops_setup();
 701	register_reboot_notifier(&kvm_pv_reboot_nb);
 702	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 703		raw_spin_lock_init(&async_pf_sleepers[i].lock);
 
 
 704
 705	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 706		has_steal_clock = 1;
 707		static_call_update(pv_steal_clock, kvm_steal_clock);
 708	}
 709
 
 
 
 
 
 710	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 711		apic_set_eoi_write(kvm_guest_apic_eoi_write);
 712
 713	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
 714		static_branch_enable(&kvm_async_pf_enabled);
 715		alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_kvm_asyncpf_interrupt);
 716	}
 717
 718#ifdef CONFIG_SMP
 719	if (pv_tlb_flush_supported()) {
 720		pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
 721		pv_ops.mmu.tlb_remove_table = tlb_remove_table;
 722		pr_info("KVM setup pv remote TLB flush\n");
 723	}
 724
 725	smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 726	if (pv_sched_yield_supported()) {
 727		smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
 728		pr_info("setup PV sched yield\n");
 729	}
 730	if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 731				      kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 732		pr_err("failed to install cpu hotplug callbacks\n");
 733#else
 734	sev_map_percpu_data();
 735	kvm_guest_cpu_init();
 736#endif
 737
 738#ifdef CONFIG_KEXEC_CORE
 739	machine_ops.crash_shutdown = kvm_crash_shutdown;
 740#endif
 741
 742	register_syscore_ops(&kvm_syscore_ops);
 743
 744	/*
 745	 * Hard lockup detection is enabled by default. Disable it, as guests
 746	 * can get false positives too easily, for example if the host is
 747	 * overcommitted.
 748	 */
 749	hardlockup_detector_disable();
 750}
 751
 752static noinline uint32_t __kvm_cpuid_base(void)
 753{
 754	if (boot_cpu_data.cpuid_level < 0)
 755		return 0;	/* So we don't blow up on old processors */
 756
 757	if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
 758		return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
 759
 760	return 0;
 761}
 762
 763static inline uint32_t kvm_cpuid_base(void)
 764{
 765	static int kvm_cpuid_base = -1;
 766
 767	if (kvm_cpuid_base == -1)
 768		kvm_cpuid_base = __kvm_cpuid_base();
 769
 770	return kvm_cpuid_base;
 771}
 772
 773bool kvm_para_available(void)
 774{
 775	return kvm_cpuid_base() != 0;
 776}
 777EXPORT_SYMBOL_GPL(kvm_para_available);
 778
 779unsigned int kvm_arch_para_features(void)
 780{
 781	return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 782}
 783
 784unsigned int kvm_arch_para_hints(void)
 785{
 786	return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 787}
 788EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
 789
 790static uint32_t __init kvm_detect(void)
 791{
 792	return kvm_cpuid_base();
 793}
 794
 795static void __init kvm_apic_init(void)
 796{
 797#ifdef CONFIG_SMP
 798	if (pv_ipi_supported())
 799		kvm_setup_pv_ipi();
 800#endif
 801}
 802
 803static bool __init kvm_msi_ext_dest_id(void)
 804{
 805	return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
 806}
 807
 808static void __init kvm_init_platform(void)
 809{
 810	kvmclock_init();
 811	x86_platform.apic_post_init = kvm_apic_init;
 812}
 813
 814#if defined(CONFIG_AMD_MEM_ENCRYPT)
 815static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
 816{
 817	/* RAX and CPL are already in the GHCB */
 818	ghcb_set_rbx(ghcb, regs->bx);
 819	ghcb_set_rcx(ghcb, regs->cx);
 820	ghcb_set_rdx(ghcb, regs->dx);
 821	ghcb_set_rsi(ghcb, regs->si);
 822}
 823
 824static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
 825{
 826	/* No checking of the return state needed */
 827	return true;
 828}
 829#endif
 830
 831const __initconst struct hypervisor_x86 x86_hyper_kvm = {
 832	.name				= "KVM",
 833	.detect				= kvm_detect,
 834	.type				= X86_HYPER_KVM,
 835	.init.guest_late_init		= kvm_guest_init,
 836	.init.x2apic_available		= kvm_para_available,
 837	.init.msi_ext_dest_id		= kvm_msi_ext_dest_id,
 838	.init.init_platform		= kvm_init_platform,
 839#if defined(CONFIG_AMD_MEM_ENCRYPT)
 840	.runtime.sev_es_hcall_prepare	= kvm_sev_es_hcall_prepare,
 841	.runtime.sev_es_hcall_finish	= kvm_sev_es_hcall_finish,
 842#endif
 843};
 844
 845static __init int activate_jump_labels(void)
 846{
 847	if (has_steal_clock) {
 848		static_key_slow_inc(&paravirt_steal_enabled);
 849		if (steal_acc)
 850			static_key_slow_inc(&paravirt_steal_rq_enabled);
 851	}
 852
 853	return 0;
 854}
 855arch_initcall(activate_jump_labels);
 856
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 857#ifdef CONFIG_PARAVIRT_SPINLOCKS
 858
 859/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 860static void kvm_kick_cpu(int cpu)
 861{
 862	int apicid;
 863	unsigned long flags = 0;
 864
 865	apicid = per_cpu(x86_cpu_to_apicid, cpu);
 866	kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 867}
 868
 869#include <asm/qspinlock.h>
 870
 871static void kvm_wait(u8 *ptr, u8 val)
 872{
 
 
 873	if (in_nmi())
 874		return;
 875
 
 
 
 
 
 876	/*
 877	 * halt until it's our turn and kicked. Note that we do safe halt
 878	 * for irq enabled case to avoid hang when lock info is overwritten
 879	 * in irq spinlock slowpath and no spurious interrupt occur to save us.
 880	 */
 881	if (irqs_disabled()) {
 882		if (READ_ONCE(*ptr) == val)
 883			halt();
 884	} else {
 885		local_irq_disable();
 886
 887		if (READ_ONCE(*ptr) == val)
 888			safe_halt();
 889
 890		local_irq_enable();
 891	}
 892}
 893
 894#ifdef CONFIG_X86_32
 895__visible bool __kvm_vcpu_is_preempted(long cpu)
 896{
 897	struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
 898
 899	return !!(src->preempted & KVM_VCPU_PREEMPTED);
 900}
 901PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
 902
 903#else
 904
 905#include <asm/asm-offsets.h>
 906
 907extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
 908
 909/*
 910 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
 911 * restoring to/from the stack.
 912 */
 913asm(
 914".pushsection .text;"
 915".global __raw_callee_save___kvm_vcpu_is_preempted;"
 916".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
 917"__raw_callee_save___kvm_vcpu_is_preempted:"
 918"movq	__per_cpu_offset(,%rdi,8), %rax;"
 919"cmpb	$0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
 920"setne	%al;"
 921"ret;"
 922".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
 923".popsection");
 924
 925#endif
 926
 927/*
 928 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 929 */
 930void __init kvm_spinlock_init(void)
 931{
 932	/*
 933	 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
 934	 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
 935	 * preferred over native qspinlock when vCPU is preempted.
 936	 */
 937	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
 938		pr_info("PV spinlocks disabled, no host support\n");
 939		return;
 940	}
 941
 942	/*
 943	 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
 944	 * are available.
 945	 */
 946	if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
 947		pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
 948		goto out;
 949	}
 950
 951	if (num_possible_cpus() == 1) {
 952		pr_info("PV spinlocks disabled, single CPU\n");
 953		goto out;
 954	}
 955
 956	if (nopvspin) {
 957		pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
 958		goto out;
 959	}
 960
 961	pr_info("PV spinlocks enabled\n");
 962
 963	__pv_init_lock_hash();
 964	pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 965	pv_ops.lock.queued_spin_unlock =
 966		PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 967	pv_ops.lock.wait = kvm_wait;
 968	pv_ops.lock.kick = kvm_kick_cpu;
 969
 970	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 971		pv_ops.lock.vcpu_is_preempted =
 972			PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
 973	}
 974	/*
 975	 * When PV spinlock is enabled which is preferred over
 976	 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
 977	 * Just disable it anyway.
 978	 */
 979out:
 980	static_branch_disable(&virt_spin_lock_key);
 981}
 982
 983#endif	/* CONFIG_PARAVIRT_SPINLOCKS */
 984
 985#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
 986
 987static void kvm_disable_host_haltpoll(void *i)
 988{
 989	wrmsrl(MSR_KVM_POLL_CONTROL, 0);
 990}
 991
 992static void kvm_enable_host_haltpoll(void *i)
 993{
 994	wrmsrl(MSR_KVM_POLL_CONTROL, 1);
 995}
 996
 997void arch_haltpoll_enable(unsigned int cpu)
 998{
 999	if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1000		pr_err_once("host does not support poll control\n");
1001		pr_err_once("host upgrade recommended\n");
1002		return;
1003	}
1004
1005	/* Enable guest halt poll disables host halt poll */
1006	smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
1007}
1008EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1009
1010void arch_haltpoll_disable(unsigned int cpu)
1011{
1012	if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1013		return;
1014
1015	/* Disable guest halt poll enables host halt poll */
1016	smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
1017}
1018EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
1019#endif