Linux Audio

Check our new training course

Loading...
v5.4
  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#include <linux/context_tracking.h>
 11#include <linux/init.h>
 12#include <linux/kernel.h>
 13#include <linux/kvm_para.h>
 14#include <linux/cpu.h>
 15#include <linux/mm.h>
 16#include <linux/highmem.h>
 17#include <linux/hardirq.h>
 18#include <linux/notifier.h>
 19#include <linux/reboot.h>
 20#include <linux/hash.h>
 21#include <linux/sched.h>
 22#include <linux/slab.h>
 23#include <linux/kprobes.h>
 24#include <linux/debugfs.h>
 25#include <linux/nmi.h>
 26#include <linux/swait.h>
 27#include <asm/timer.h>
 28#include <asm/cpu.h>
 29#include <asm/traps.h>
 30#include <asm/desc.h>
 31#include <asm/tlbflush.h>
 32#include <asm/apic.h>
 33#include <asm/apicdef.h>
 34#include <asm/hypervisor.h>
 35#include <asm/tlb.h>
 36
 37static int kvmapf = 1;
 38
 39static int __init parse_no_kvmapf(char *arg)
 40{
 41        kvmapf = 0;
 42        return 0;
 43}
 44
 45early_param("no-kvmapf", parse_no_kvmapf);
 46
 47static int steal_acc = 1;
 48static int __init parse_no_stealacc(char *arg)
 49{
 50        steal_acc = 0;
 51        return 0;
 52}
 53
 54early_param("no-steal-acc", parse_no_stealacc);
 55
 
 
 
 
 
 
 
 
 
 56static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
 57DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
 58static int has_steal_clock = 0;
 59
 60/*
 61 * No need for any "IO delay" on KVM
 62 */
 63static void kvm_io_delay(void)
 64{
 65}
 66
 67#define KVM_TASK_SLEEP_HASHBITS 8
 68#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
 69
 70struct kvm_task_sleep_node {
 71	struct hlist_node link;
 72	struct swait_queue_head wq;
 73	u32 token;
 74	int cpu;
 75	bool halted;
 76};
 77
 78static struct kvm_task_sleep_head {
 79	raw_spinlock_t lock;
 80	struct hlist_head list;
 81} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
 82
 83static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
 84						  u32 token)
 85{
 86	struct hlist_node *p;
 87
 88	hlist_for_each(p, &b->list) {
 89		struct kvm_task_sleep_node *n =
 90			hlist_entry(p, typeof(*n), link);
 91		if (n->token == token)
 92			return n;
 93	}
 94
 95	return NULL;
 96}
 97
 98/*
 99 * @interrupt_kernel: Is this called from a routine which interrupts the kernel
100 * 		      (other than user space)?
101 */
102void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
103{
104	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
105	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
106	struct kvm_task_sleep_node n, *e;
107	DECLARE_SWAITQUEUE(wait);
108
109	rcu_irq_enter();
110
111	raw_spin_lock(&b->lock);
112	e = _find_apf_task(b, token);
113	if (e) {
114		/* dummy entry exist -> wake up was delivered ahead of PF */
115		hlist_del(&e->link);
116		kfree(e);
117		raw_spin_unlock(&b->lock);
118
119		rcu_irq_exit();
120		return;
121	}
122
123	n.token = token;
124	n.cpu = smp_processor_id();
125	n.halted = is_idle_task(current) ||
126		   (IS_ENABLED(CONFIG_PREEMPT_COUNT)
127		    ? preempt_count() > 1 || rcu_preempt_depth()
128		    : interrupt_kernel);
129	init_swait_queue_head(&n.wq);
130	hlist_add_head(&n.link, &b->list);
131	raw_spin_unlock(&b->lock);
132
133	for (;;) {
134		if (!n.halted)
135			prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
136		if (hlist_unhashed(&n.link))
137			break;
138
139		rcu_irq_exit();
140
141		if (!n.halted) {
142			local_irq_enable();
143			schedule();
144			local_irq_disable();
145		} else {
146			/*
147			 * We cannot reschedule. So halt.
148			 */
149			native_safe_halt();
150			local_irq_disable();
151		}
152
153		rcu_irq_enter();
154	}
155	if (!n.halted)
156		finish_swait(&n.wq, &wait);
157
158	rcu_irq_exit();
159	return;
160}
161EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
162
163static void apf_task_wake_one(struct kvm_task_sleep_node *n)
164{
165	hlist_del_init(&n->link);
166	if (n->halted)
167		smp_send_reschedule(n->cpu);
168	else if (swq_has_sleeper(&n->wq))
169		swake_up_one(&n->wq);
170}
171
172static void apf_task_wake_all(void)
173{
174	int i;
175
176	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
177		struct hlist_node *p, *next;
178		struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
179		raw_spin_lock(&b->lock);
180		hlist_for_each_safe(p, next, &b->list) {
181			struct kvm_task_sleep_node *n =
182				hlist_entry(p, typeof(*n), link);
183			if (n->cpu == smp_processor_id())
184				apf_task_wake_one(n);
185		}
186		raw_spin_unlock(&b->lock);
187	}
188}
189
190void kvm_async_pf_task_wake(u32 token)
191{
192	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
193	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
194	struct kvm_task_sleep_node *n;
195
196	if (token == ~0) {
197		apf_task_wake_all();
198		return;
199	}
200
201again:
202	raw_spin_lock(&b->lock);
203	n = _find_apf_task(b, token);
204	if (!n) {
205		/*
206		 * async PF was not yet handled.
207		 * Add dummy entry for the token.
208		 */
209		n = kzalloc(sizeof(*n), GFP_ATOMIC);
210		if (!n) {
211			/*
212			 * Allocation failed! Busy wait while other cpu
213			 * handles async PF.
214			 */
215			raw_spin_unlock(&b->lock);
216			cpu_relax();
217			goto again;
218		}
219		n->token = token;
220		n->cpu = smp_processor_id();
221		init_swait_queue_head(&n->wq);
222		hlist_add_head(&n->link, &b->list);
223	} else
224		apf_task_wake_one(n);
225	raw_spin_unlock(&b->lock);
226	return;
227}
228EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
229
230u32 kvm_read_and_reset_pf_reason(void)
231{
232	u32 reason = 0;
233
234	if (__this_cpu_read(apf_reason.enabled)) {
235		reason = __this_cpu_read(apf_reason.reason);
236		__this_cpu_write(apf_reason.reason, 0);
237	}
238
239	return reason;
240}
241EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
242NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
243
244dotraplinkage void
245do_async_page_fault(struct pt_regs *regs, unsigned long error_code, unsigned long address)
246{
247	enum ctx_state prev_state;
248
249	switch (kvm_read_and_reset_pf_reason()) {
250	default:
251		do_page_fault(regs, error_code, address);
252		break;
253	case KVM_PV_REASON_PAGE_NOT_PRESENT:
254		/* page is swapped out by the host. */
255		prev_state = exception_enter();
256		kvm_async_pf_task_wait((u32)address, !user_mode(regs));
257		exception_exit(prev_state);
258		break;
259	case KVM_PV_REASON_PAGE_READY:
260		rcu_irq_enter();
261		kvm_async_pf_task_wake((u32)address);
262		rcu_irq_exit();
263		break;
264	}
265}
266NOKPROBE_SYMBOL(do_async_page_fault);
267
268static void __init paravirt_ops_setup(void)
269{
270	pv_info.name = "KVM";
271
272	if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
273		pv_ops.cpu.io_delay = kvm_io_delay;
274
275#ifdef CONFIG_X86_IO_APIC
276	no_timer_check = 1;
277#endif
278}
279
280static void kvm_register_steal_time(void)
281{
282	int cpu = smp_processor_id();
283	struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
284
285	if (!has_steal_clock)
286		return;
287
288	wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
289	pr_info("kvm-stealtime: cpu %d, msr %llx\n",
290		cpu, (unsigned long long) slow_virt_to_phys(st));
291}
292
293static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
294
295static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
296{
297	/**
298	 * This relies on __test_and_clear_bit to modify the memory
299	 * in a way that is atomic with respect to the local CPU.
300	 * The hypervisor only accesses this memory from the local CPU so
301	 * there's no need for lock or memory barriers.
302	 * An optimization barrier is implied in apic write.
303	 */
304	if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
305		return;
306	apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
307}
308
309static void kvm_guest_cpu_init(void)
310{
 
 
 
311	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
312		u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
313
314#ifdef CONFIG_PREEMPTION
315		pa |= KVM_ASYNC_PF_SEND_ALWAYS;
316#endif
317		pa |= KVM_ASYNC_PF_ENABLED;
318
319		if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
320			pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
321
322		wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
323		__this_cpu_write(apf_reason.enabled, 1);
324		printk(KERN_INFO"KVM setup async PF for cpu %d\n",
325		       smp_processor_id());
326	}
327
328	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
329		unsigned long pa;
330		/* Size alignment is implied but just to make it explicit. */
331		BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
332		__this_cpu_write(kvm_apic_eoi, 0);
333		pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
334			| KVM_MSR_ENABLED;
335		wrmsrl(MSR_KVM_PV_EOI_EN, pa);
336	}
337
338	if (has_steal_clock)
339		kvm_register_steal_time();
340}
341
342static void kvm_pv_disable_apf(void)
343{
344	if (!__this_cpu_read(apf_reason.enabled))
345		return;
346
347	wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
348	__this_cpu_write(apf_reason.enabled, 0);
349
350	printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
351	       smp_processor_id());
352}
353
354static void kvm_pv_guest_cpu_reboot(void *unused)
355{
356	/*
357	 * We disable PV EOI before we load a new kernel by kexec,
358	 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
359	 * New kernel can re-enable when it boots.
360	 */
361	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
362		wrmsrl(MSR_KVM_PV_EOI_EN, 0);
363	kvm_pv_disable_apf();
364	kvm_disable_steal_time();
365}
366
367static int kvm_pv_reboot_notify(struct notifier_block *nb,
368				unsigned long code, void *unused)
369{
370	if (code == SYS_RESTART)
371		on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
372	return NOTIFY_DONE;
373}
374
375static struct notifier_block kvm_pv_reboot_nb = {
376	.notifier_call = kvm_pv_reboot_notify,
377};
378
379static u64 kvm_steal_clock(int cpu)
380{
381	u64 steal;
382	struct kvm_steal_time *src;
383	int version;
384
385	src = &per_cpu(steal_time, cpu);
386	do {
387		version = src->version;
388		virt_rmb();
389		steal = src->steal;
390		virt_rmb();
391	} while ((version & 1) || (version != src->version));
392
393	return steal;
394}
395
396void kvm_disable_steal_time(void)
397{
398	if (!has_steal_clock)
399		return;
400
401	wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
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
431#ifdef CONFIG_SMP
432#define KVM_IPI_CLUSTER_SIZE	(2 * BITS_PER_LONG)
433
434static void __send_ipi_mask(const struct cpumask *mask, int vector)
435{
436	unsigned long flags;
437	int cpu, apic_id, icr;
438	int min = 0, max = 0;
439#ifdef CONFIG_X86_64
440	__uint128_t ipi_bitmap = 0;
441#else
442	u64 ipi_bitmap = 0;
443#endif
444	long ret;
445
446	if (cpumask_empty(mask))
447		return;
448
449	local_irq_save(flags);
450
451	switch (vector) {
452	default:
453		icr = APIC_DM_FIXED | vector;
454		break;
455	case NMI_VECTOR:
456		icr = APIC_DM_NMI;
457		break;
458	}
459
460	for_each_cpu(cpu, mask) {
461		apic_id = per_cpu(x86_cpu_to_apicid, cpu);
462		if (!ipi_bitmap) {
463			min = max = apic_id;
464		} else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
465			ipi_bitmap <<= min - apic_id;
466			min = apic_id;
467		} else if (apic_id < min + KVM_IPI_CLUSTER_SIZE) {
468			max = apic_id < max ? max : apic_id;
469		} else {
470			ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
471				(unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
472			WARN_ONCE(ret < 0, "KVM: failed to send PV IPI: %ld", ret);
473			min = max = apic_id;
474			ipi_bitmap = 0;
475		}
476		__set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
477	}
478
479	if (ipi_bitmap) {
480		ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
481			(unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
482		WARN_ONCE(ret < 0, "KVM: failed to send PV IPI: %ld", ret);
483	}
484
485	local_irq_restore(flags);
486}
487
488static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
489{
490	__send_ipi_mask(mask, vector);
491}
492
493static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
494{
495	unsigned int this_cpu = smp_processor_id();
496	struct cpumask new_mask;
497	const struct cpumask *local_mask;
498
499	cpumask_copy(&new_mask, mask);
500	cpumask_clear_cpu(this_cpu, &new_mask);
501	local_mask = &new_mask;
502	__send_ipi_mask(local_mask, vector);
503}
504
505/*
506 * Set the IPI entry points
507 */
508static void kvm_setup_pv_ipi(void)
509{
510	apic->send_IPI_mask = kvm_send_ipi_mask;
511	apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
512	pr_info("KVM setup pv IPIs\n");
513}
514
515static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
516{
517	int cpu;
518
519	native_send_call_func_ipi(mask);
520
521	/* Make sure other vCPUs get a chance to run if they need to. */
522	for_each_cpu(cpu, mask) {
523		if (vcpu_is_preempted(cpu)) {
524			kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
525			break;
526		}
527	}
528}
529
530static void __init kvm_smp_prepare_cpus(unsigned int max_cpus)
531{
532	native_smp_prepare_cpus(max_cpus);
533	if (kvm_para_has_hint(KVM_HINTS_REALTIME))
534		static_branch_disable(&virt_spin_lock_key);
535}
536
537static void __init kvm_smp_prepare_boot_cpu(void)
538{
539	/*
540	 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
541	 * shares the guest physical address with the hypervisor.
542	 */
543	sev_map_percpu_data();
544
545	kvm_guest_cpu_init();
546	native_smp_prepare_boot_cpu();
547	kvm_spinlock_init();
548}
549
550static void kvm_guest_cpu_offline(void)
551{
552	kvm_disable_steal_time();
553	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
554		wrmsrl(MSR_KVM_PV_EOI_EN, 0);
555	kvm_pv_disable_apf();
556	apf_task_wake_all();
557}
558
559static int kvm_cpu_online(unsigned int cpu)
560{
561	local_irq_disable();
562	kvm_guest_cpu_init();
563	local_irq_enable();
564	return 0;
565}
566
567static int kvm_cpu_down_prepare(unsigned int cpu)
568{
569	local_irq_disable();
570	kvm_guest_cpu_offline();
571	local_irq_enable();
572	return 0;
573}
574#endif
575
576static void __init kvm_apf_trap_init(void)
577{
578	update_intr_gate(X86_TRAP_PF, async_page_fault);
579}
580
581static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
582
583static void kvm_flush_tlb_others(const struct cpumask *cpumask,
584			const struct flush_tlb_info *info)
585{
586	u8 state;
587	int cpu;
588	struct kvm_steal_time *src;
589	struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_tlb_mask);
590
591	cpumask_copy(flushmask, cpumask);
592	/*
593	 * We have to call flush only on online vCPUs. And
594	 * queue flush_on_enter for pre-empted vCPUs
595	 */
596	for_each_cpu(cpu, flushmask) {
597		src = &per_cpu(steal_time, cpu);
598		state = READ_ONCE(src->preempted);
599		if ((state & KVM_VCPU_PREEMPTED)) {
600			if (try_cmpxchg(&src->preempted, &state,
601					state | KVM_VCPU_FLUSH_TLB))
602				__cpumask_clear_cpu(cpu, flushmask);
603		}
604	}
605
606	native_flush_tlb_others(flushmask, info);
607}
608
609static void __init kvm_guest_init(void)
610{
611	int i;
612
 
 
 
613	paravirt_ops_setup();
614	register_reboot_notifier(&kvm_pv_reboot_nb);
615	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
616		raw_spin_lock_init(&async_pf_sleepers[i].lock);
617	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
618		x86_init.irqs.trap_init = kvm_apf_trap_init;
619
620	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
621		has_steal_clock = 1;
622		pv_ops.time.steal_clock = kvm_steal_clock;
623	}
624
625	if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
626	    !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
627	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
628		pv_ops.mmu.flush_tlb_others = kvm_flush_tlb_others;
629		pv_ops.mmu.tlb_remove_table = tlb_remove_table;
630	}
631
632	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
633		apic_set_eoi_write(kvm_guest_apic_eoi_write);
634
 
 
 
635#ifdef CONFIG_SMP
636	smp_ops.smp_prepare_cpus = kvm_smp_prepare_cpus;
637	smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
638	if (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
639	    !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
640	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
641		smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
642		pr_info("KVM setup pv sched yield\n");
643	}
644	if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
645				      kvm_cpu_online, kvm_cpu_down_prepare) < 0)
646		pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
647#else
648	sev_map_percpu_data();
649	kvm_guest_cpu_init();
650#endif
651
652	/*
653	 * Hard lockup detection is enabled by default. Disable it, as guests
654	 * can get false positives too easily, for example if the host is
655	 * overcommitted.
656	 */
657	hardlockup_detector_disable();
658}
659
660static noinline uint32_t __kvm_cpuid_base(void)
661{
662	if (boot_cpu_data.cpuid_level < 0)
663		return 0;	/* So we don't blow up on old processors */
664
665	if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
666		return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
667
668	return 0;
669}
670
671static inline uint32_t kvm_cpuid_base(void)
672{
673	static int kvm_cpuid_base = -1;
674
675	if (kvm_cpuid_base == -1)
676		kvm_cpuid_base = __kvm_cpuid_base();
677
678	return kvm_cpuid_base;
679}
680
681bool kvm_para_available(void)
682{
683	return kvm_cpuid_base() != 0;
684}
685EXPORT_SYMBOL_GPL(kvm_para_available);
686
687unsigned int kvm_arch_para_features(void)
688{
689	return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
690}
691
692unsigned int kvm_arch_para_hints(void)
693{
694	return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
695}
696EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
697
698static uint32_t __init kvm_detect(void)
699{
700	return kvm_cpuid_base();
701}
702
703static void __init kvm_apic_init(void)
704{
705#if defined(CONFIG_SMP)
706	if (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI))
707		kvm_setup_pv_ipi();
708#endif
709}
710
711static void __init kvm_init_platform(void)
712{
713	kvmclock_init();
714	x86_platform.apic_post_init = kvm_apic_init;
715}
716
717const __initconst struct hypervisor_x86 x86_hyper_kvm = {
718	.name			= "KVM",
719	.detect			= kvm_detect,
720	.type			= X86_HYPER_KVM,
721	.init.guest_late_init	= kvm_guest_init,
722	.init.x2apic_available	= kvm_para_available,
723	.init.init_platform	= kvm_init_platform,
724};
725
726static __init int activate_jump_labels(void)
727{
728	if (has_steal_clock) {
729		static_key_slow_inc(&paravirt_steal_enabled);
730		if (steal_acc)
731			static_key_slow_inc(&paravirt_steal_rq_enabled);
732	}
733
734	return 0;
735}
736arch_initcall(activate_jump_labels);
737
738static __init int kvm_setup_pv_tlb_flush(void)
739{
740	int cpu;
741
742	if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
743	    !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
744	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
745		for_each_possible_cpu(cpu) {
746			zalloc_cpumask_var_node(per_cpu_ptr(&__pv_tlb_mask, cpu),
747				GFP_KERNEL, cpu_to_node(cpu));
748		}
749		pr_info("KVM setup pv remote TLB flush\n");
750	}
751
752	return 0;
753}
754arch_initcall(kvm_setup_pv_tlb_flush);
755
756#ifdef CONFIG_PARAVIRT_SPINLOCKS
757
758/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
759static void kvm_kick_cpu(int cpu)
760{
761	int apicid;
762	unsigned long flags = 0;
763
764	apicid = per_cpu(x86_cpu_to_apicid, cpu);
765	kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
766}
767
768#include <asm/qspinlock.h>
769
770static void kvm_wait(u8 *ptr, u8 val)
771{
772	unsigned long flags;
773
774	if (in_nmi())
775		return;
776
777	local_irq_save(flags);
778
779	if (READ_ONCE(*ptr) != val)
780		goto out;
781
782	/*
783	 * halt until it's our turn and kicked. Note that we do safe halt
784	 * for irq enabled case to avoid hang when lock info is overwritten
785	 * in irq spinlock slowpath and no spurious interrupt occur to save us.
786	 */
787	if (arch_irqs_disabled_flags(flags))
788		halt();
789	else
790		safe_halt();
791
792out:
793	local_irq_restore(flags);
794}
795
796#ifdef CONFIG_X86_32
797__visible bool __kvm_vcpu_is_preempted(long cpu)
798{
799	struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
800
801	return !!(src->preempted & KVM_VCPU_PREEMPTED);
802}
803PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
804
805#else
806
807#include <asm/asm-offsets.h>
808
809extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
810
811/*
812 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
813 * restoring to/from the stack.
814 */
815asm(
816".pushsection .text;"
817".global __raw_callee_save___kvm_vcpu_is_preempted;"
818".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
819"__raw_callee_save___kvm_vcpu_is_preempted:"
820"movq	__per_cpu_offset(,%rdi,8), %rax;"
821"cmpb	$0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
822"setne	%al;"
823"ret;"
824".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
825".popsection");
826
827#endif
828
829/*
830 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
831 */
832void __init kvm_spinlock_init(void)
833{
 
 
834	/* Does host kernel support KVM_FEATURE_PV_UNHALT? */
835	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
836		return;
837
838	if (kvm_para_has_hint(KVM_HINTS_REALTIME))
839		return;
840
841	/* Don't use the pvqspinlock code if there is only 1 vCPU. */
842	if (num_possible_cpus() == 1)
843		return;
844
845	__pv_init_lock_hash();
846	pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
847	pv_ops.lock.queued_spin_unlock =
848		PV_CALLEE_SAVE(__pv_queued_spin_unlock);
849	pv_ops.lock.wait = kvm_wait;
850	pv_ops.lock.kick = kvm_kick_cpu;
851
852	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
853		pv_ops.lock.vcpu_is_preempted =
854			PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
855	}
856}
857
858#endif	/* CONFIG_PARAVIRT_SPINLOCKS */
859
860#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
861
862static void kvm_disable_host_haltpoll(void *i)
863{
864	wrmsrl(MSR_KVM_POLL_CONTROL, 0);
865}
866
867static void kvm_enable_host_haltpoll(void *i)
868{
869	wrmsrl(MSR_KVM_POLL_CONTROL, 1);
870}
871
872void arch_haltpoll_enable(unsigned int cpu)
873{
874	if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
875		pr_err_once("kvm: host does not support poll control\n");
876		pr_err_once("kvm: host upgrade recommended\n");
877		return;
878	}
879
880	/* Enable guest halt poll disables host halt poll */
881	smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
882}
883EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
884
885void arch_haltpoll_disable(unsigned int cpu)
886{
887	if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
888		return;
889
890	/* Enable guest halt poll disables host halt poll */
891	smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
892}
893EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
894#endif
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 */