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
v4.6
 
  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/module.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/idle.h>
 46#include <asm/apic.h>
 47#include <asm/apicdef.h>
 48#include <asm/hypervisor.h>
 49#include <asm/kvm_guest.h>
 50
 51static int kvmapf = 1;
 52
 53static int parse_no_kvmapf(char *arg)
 54{
 55        kvmapf = 0;
 56        return 0;
 57}
 58
 59early_param("no-kvmapf", parse_no_kvmapf);
 60
 61static int steal_acc = 1;
 62static int parse_no_stealacc(char *arg)
 63{
 64        steal_acc = 0;
 65        return 0;
 66}
 67
 68early_param("no-steal-acc", parse_no_stealacc);
 69
 70static int kvmclock_vsyscall = 1;
 71static int parse_no_kvmclock_vsyscall(char *arg)
 72{
 73        kvmclock_vsyscall = 0;
 74        return 0;
 75}
 76
 77early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
 78
 79static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
 80static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
 81static int has_steal_clock = 0;
 82
 83/*
 84 * No need for any "IO delay" on KVM
 85 */
 86static void kvm_io_delay(void)
 87{
 88}
 89
 90#define KVM_TASK_SLEEP_HASHBITS 8
 91#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
 92
 93struct kvm_task_sleep_node {
 94	struct hlist_node link;
 95	struct swait_queue_head wq;
 96	u32 token;
 97	int cpu;
 98	bool halted;
 99};
100
101static struct kvm_task_sleep_head {
102	raw_spinlock_t lock;
103	struct hlist_head list;
104} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
105
106static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
107						  u32 token)
108{
109	struct hlist_node *p;
110
111	hlist_for_each(p, &b->list) {
112		struct kvm_task_sleep_node *n =
113			hlist_entry(p, typeof(*n), link);
114		if (n->token == token)
115			return n;
116	}
117
118	return NULL;
119}
120
121void kvm_async_pf_task_wait(u32 token)
 
 
 
 
122{
123	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
124	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
125	struct kvm_task_sleep_node n, *e;
126	DECLARE_SWAITQUEUE(wait);
127
128	rcu_irq_enter();
129
130	raw_spin_lock(&b->lock);
131	e = _find_apf_task(b, token);
132	if (e) {
133		/* dummy entry exist -> wake up was delivered ahead of PF */
134		hlist_del(&e->link);
135		kfree(e);
136		raw_spin_unlock(&b->lock);
137
138		rcu_irq_exit();
139		return;
140	}
141
142	n.token = token;
143	n.cpu = smp_processor_id();
144	n.halted = is_idle_task(current) || preempt_count() > 1;
 
 
 
145	init_swait_queue_head(&n.wq);
146	hlist_add_head(&n.link, &b->list);
147	raw_spin_unlock(&b->lock);
148
149	for (;;) {
150		if (!n.halted)
151			prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
152		if (hlist_unhashed(&n.link))
153			break;
154
 
 
155		if (!n.halted) {
156			local_irq_enable();
157			schedule();
158			local_irq_disable();
159		} else {
160			/*
161			 * We cannot reschedule. So halt.
162			 */
163			rcu_irq_exit();
164			native_safe_halt();
165			rcu_irq_enter();
166			local_irq_disable();
167		}
 
 
168	}
169	if (!n.halted)
170		finish_swait(&n.wq, &wait);
171
172	rcu_irq_exit();
173	return;
174}
175EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
176
177static void apf_task_wake_one(struct kvm_task_sleep_node *n)
178{
179	hlist_del_init(&n->link);
180	if (n->halted)
181		smp_send_reschedule(n->cpu);
182	else if (swait_active(&n->wq))
183		swake_up(&n->wq);
184}
185
186static void apf_task_wake_all(void)
187{
188	int i;
189
190	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
191		struct hlist_node *p, *next;
192		struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
193		raw_spin_lock(&b->lock);
194		hlist_for_each_safe(p, next, &b->list) {
195			struct kvm_task_sleep_node *n =
196				hlist_entry(p, typeof(*n), link);
197			if (n->cpu == smp_processor_id())
198				apf_task_wake_one(n);
199		}
200		raw_spin_unlock(&b->lock);
201	}
202}
203
204void kvm_async_pf_task_wake(u32 token)
205{
206	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
207	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
208	struct kvm_task_sleep_node *n;
209
210	if (token == ~0) {
211		apf_task_wake_all();
212		return;
213	}
214
215again:
216	raw_spin_lock(&b->lock);
217	n = _find_apf_task(b, token);
218	if (!n) {
219		/*
220		 * async PF was not yet handled.
221		 * Add dummy entry for the token.
222		 */
223		n = kzalloc(sizeof(*n), GFP_ATOMIC);
224		if (!n) {
225			/*
226			 * Allocation failed! Busy wait while other cpu
227			 * handles async PF.
228			 */
229			raw_spin_unlock(&b->lock);
230			cpu_relax();
231			goto again;
232		}
233		n->token = token;
234		n->cpu = smp_processor_id();
235		init_swait_queue_head(&n->wq);
236		hlist_add_head(&n->link, &b->list);
237	} else
238		apf_task_wake_one(n);
239	raw_spin_unlock(&b->lock);
240	return;
241}
242EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
243
244u32 kvm_read_and_reset_pf_reason(void)
245{
246	u32 reason = 0;
247
248	if (__this_cpu_read(apf_reason.enabled)) {
249		reason = __this_cpu_read(apf_reason.reason);
250		__this_cpu_write(apf_reason.reason, 0);
251	}
252
253	return reason;
254}
255EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
256NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
257
258dotraplinkage void
259do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
260{
261	enum ctx_state prev_state;
262
263	switch (kvm_read_and_reset_pf_reason()) {
264	default:
265		trace_do_page_fault(regs, error_code);
266		break;
267	case KVM_PV_REASON_PAGE_NOT_PRESENT:
268		/* page is swapped out by the host. */
269		prev_state = exception_enter();
270		exit_idle();
271		kvm_async_pf_task_wait((u32)read_cr2());
272		exception_exit(prev_state);
273		break;
274	case KVM_PV_REASON_PAGE_READY:
275		rcu_irq_enter();
276		exit_idle();
277		kvm_async_pf_task_wake((u32)read_cr2());
278		rcu_irq_exit();
279		break;
280	}
281}
282NOKPROBE_SYMBOL(do_async_page_fault);
283
284static void __init paravirt_ops_setup(void)
285{
286	pv_info.name = "KVM";
287
288	/*
289	 * KVM isn't paravirt in the sense of paravirt_enabled.  A KVM
290	 * guest kernel works like a bare metal kernel with additional
291	 * features, and paravirt_enabled is about features that are
292	 * missing.
293	 */
294	pv_info.paravirt_enabled = 0;
295
296	if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
297		pv_cpu_ops.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	memset(st, 0, sizeof(*st));
313
314	wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
315	pr_info("kvm-stealtime: cpu %d, msr %llx\n",
316		cpu, (unsigned long long) slow_virt_to_phys(st));
317}
318
319static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
320
321static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
322{
323	/**
324	 * This relies on __test_and_clear_bit to modify the memory
325	 * in a way that is atomic with respect to the local CPU.
326	 * The hypervisor only accesses this memory from the local CPU so
327	 * there's no need for lock or memory barriers.
328	 * An optimization barrier is implied in apic write.
329	 */
330	if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
331		return;
332	apic_write(APIC_EOI, APIC_EOI_ACK);
333}
334
335static void kvm_guest_cpu_init(void)
336{
337	if (!kvm_para_available())
338		return;
339
340	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
341		u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
342
343#ifdef CONFIG_PREEMPT
344		pa |= KVM_ASYNC_PF_SEND_ALWAYS;
345#endif
346		wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
 
 
 
 
 
347		__this_cpu_write(apf_reason.enabled, 1);
348		printk(KERN_INFO"KVM setup async PF for cpu %d\n",
349		       smp_processor_id());
350	}
351
352	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
353		unsigned long pa;
354		/* Size alignment is implied but just to make it explicit. */
355		BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
356		__this_cpu_write(kvm_apic_eoi, 0);
357		pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
358			| KVM_MSR_ENABLED;
359		wrmsrl(MSR_KVM_PV_EOI_EN, pa);
360	}
361
362	if (has_steal_clock)
363		kvm_register_steal_time();
364}
365
366static void kvm_pv_disable_apf(void)
367{
368	if (!__this_cpu_read(apf_reason.enabled))
369		return;
370
371	wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
372	__this_cpu_write(apf_reason.enabled, 0);
373
374	printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
375	       smp_processor_id());
376}
377
378static void kvm_pv_guest_cpu_reboot(void *unused)
379{
380	/*
381	 * We disable PV EOI before we load a new kernel by kexec,
382	 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
383	 * New kernel can re-enable when it boots.
384	 */
385	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
386		wrmsrl(MSR_KVM_PV_EOI_EN, 0);
387	kvm_pv_disable_apf();
388	kvm_disable_steal_time();
389}
390
391static int kvm_pv_reboot_notify(struct notifier_block *nb,
392				unsigned long code, void *unused)
393{
394	if (code == SYS_RESTART)
395		on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
396	return NOTIFY_DONE;
397}
398
399static struct notifier_block kvm_pv_reboot_nb = {
400	.notifier_call = kvm_pv_reboot_notify,
401};
402
403static u64 kvm_steal_clock(int cpu)
404{
405	u64 steal;
406	struct kvm_steal_time *src;
407	int version;
408
409	src = &per_cpu(steal_time, cpu);
410	do {
411		version = src->version;
412		rmb();
413		steal = src->steal;
414		rmb();
415	} while ((version & 1) || (version != src->version));
416
417	return steal;
418}
419
420void kvm_disable_steal_time(void)
421{
422	if (!has_steal_clock)
423		return;
424
425	wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
426}
427
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
428#ifdef CONFIG_SMP
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
429static void __init kvm_smp_prepare_boot_cpu(void)
430{
 
 
 
 
 
 
431	kvm_guest_cpu_init();
432	native_smp_prepare_boot_cpu();
433	kvm_spinlock_init();
434}
435
436static void kvm_guest_cpu_online(void *dummy)
437{
438	kvm_guest_cpu_init();
439}
440
441static void kvm_guest_cpu_offline(void *dummy)
442{
443	kvm_disable_steal_time();
444	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
445		wrmsrl(MSR_KVM_PV_EOI_EN, 0);
446	kvm_pv_disable_apf();
447	apf_task_wake_all();
448}
449
450static int kvm_cpu_notify(struct notifier_block *self, unsigned long action,
451			  void *hcpu)
452{
453	int cpu = (unsigned long)hcpu;
454	switch (action) {
455	case CPU_ONLINE:
456	case CPU_DOWN_FAILED:
457	case CPU_ONLINE_FROZEN:
458		smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
459		break;
460	case CPU_DOWN_PREPARE:
461	case CPU_DOWN_PREPARE_FROZEN:
462		smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
463		break;
464	default:
465		break;
466	}
467	return NOTIFY_OK;
468}
469
470static struct notifier_block kvm_cpu_notifier = {
471        .notifier_call  = kvm_cpu_notify,
472};
 
 
 
 
473#endif
474
475static void __init kvm_apf_trap_init(void)
476{
477	set_intr_gate(14, async_page_fault);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
478}
479
480void __init kvm_guest_init(void)
481{
482	int i;
483
484	if (!kvm_para_available())
485		return;
486
487	paravirt_ops_setup();
488	register_reboot_notifier(&kvm_pv_reboot_nb);
489	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
490		raw_spin_lock_init(&async_pf_sleepers[i].lock);
491	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
492		x86_init.irqs.trap_init = kvm_apf_trap_init;
493
494	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
495		has_steal_clock = 1;
496		pv_time_ops.steal_clock = kvm_steal_clock;
 
 
 
 
 
 
 
497	}
498
499	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
500		apic_set_eoi_write(kvm_guest_apic_eoi_write);
501
502	if (kvmclock_vsyscall)
503		kvm_setup_vsyscall_timeinfo();
504
505#ifdef CONFIG_SMP
 
506	smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
507	register_cpu_notifier(&kvm_cpu_notifier);
 
 
 
 
 
 
 
 
508#else
 
509	kvm_guest_cpu_init();
510#endif
511
512	/*
513	 * Hard lockup detection is enabled by default. Disable it, as guests
514	 * can get false positives too easily, for example if the host is
515	 * overcommitted.
516	 */
517	hardlockup_detector_disable();
518}
519
520static noinline uint32_t __kvm_cpuid_base(void)
521{
522	if (boot_cpu_data.cpuid_level < 0)
523		return 0;	/* So we don't blow up on old processors */
524
525	if (cpu_has_hypervisor)
526		return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
527
528	return 0;
529}
530
531static inline uint32_t kvm_cpuid_base(void)
532{
533	static int kvm_cpuid_base = -1;
534
535	if (kvm_cpuid_base == -1)
536		kvm_cpuid_base = __kvm_cpuid_base();
537
538	return kvm_cpuid_base;
539}
540
541bool kvm_para_available(void)
542{
543	return kvm_cpuid_base() != 0;
544}
545EXPORT_SYMBOL_GPL(kvm_para_available);
546
547unsigned int kvm_arch_para_features(void)
548{
549	return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
550}
551
 
 
 
 
 
 
552static uint32_t __init kvm_detect(void)
553{
554	return kvm_cpuid_base();
555}
556
557const struct hypervisor_x86 x86_hyper_kvm __refconst = {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
558	.name			= "KVM",
559	.detect			= kvm_detect,
560	.x2apic_available	= kvm_para_available,
 
 
 
561};
562EXPORT_SYMBOL_GPL(x86_hyper_kvm);
563
564static __init int activate_jump_labels(void)
565{
566	if (has_steal_clock) {
567		static_key_slow_inc(&paravirt_steal_enabled);
568		if (steal_acc)
569			static_key_slow_inc(&paravirt_steal_rq_enabled);
570	}
571
572	return 0;
573}
574arch_initcall(activate_jump_labels);
575
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
576#ifdef CONFIG_PARAVIRT_SPINLOCKS
577
578/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
579static void kvm_kick_cpu(int cpu)
580{
581	int apicid;
582	unsigned long flags = 0;
583
584	apicid = per_cpu(x86_cpu_to_apicid, cpu);
585	kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
586}
587
588
589#ifdef CONFIG_QUEUED_SPINLOCKS
590
591#include <asm/qspinlock.h>
592
593static void kvm_wait(u8 *ptr, u8 val)
594{
595	unsigned long flags;
596
597	if (in_nmi())
598		return;
599
600	local_irq_save(flags);
601
602	if (READ_ONCE(*ptr) != val)
603		goto out;
604
605	/*
606	 * halt until it's our turn and kicked. Note that we do safe halt
607	 * for irq enabled case to avoid hang when lock info is overwritten
608	 * in irq spinlock slowpath and no spurious interrupt occur to save us.
609	 */
610	if (arch_irqs_disabled_flags(flags))
611		halt();
612	else
613		safe_halt();
614
615out:
616	local_irq_restore(flags);
617}
618
619#else /* !CONFIG_QUEUED_SPINLOCKS */
620
621enum kvm_contention_stat {
622	TAKEN_SLOW,
623	TAKEN_SLOW_PICKUP,
624	RELEASED_SLOW,
625	RELEASED_SLOW_KICKED,
626	NR_CONTENTION_STATS
627};
628
629#ifdef CONFIG_KVM_DEBUG_FS
630#define HISTO_BUCKETS	30
631
632static struct kvm_spinlock_stats
633{
634	u32 contention_stats[NR_CONTENTION_STATS];
635	u32 histo_spin_blocked[HISTO_BUCKETS+1];
636	u64 time_blocked;
637} spinlock_stats;
638
639static u8 zero_stats;
640
641static inline void check_zero(void)
642{
643	u8 ret;
644	u8 old;
645
646	old = READ_ONCE(zero_stats);
647	if (unlikely(old)) {
648		ret = cmpxchg(&zero_stats, old, 0);
649		/* This ensures only one fellow resets the stat */
650		if (ret == old)
651			memset(&spinlock_stats, 0, sizeof(spinlock_stats));
652	}
653}
 
654
655static inline void add_stats(enum kvm_contention_stat var, u32 val)
656{
657	check_zero();
658	spinlock_stats.contention_stats[var] += val;
659}
660
 
661
662static inline u64 spin_time_start(void)
663{
664	return sched_clock();
665}
666
667static void __spin_time_accum(u64 delta, u32 *array)
668{
669	unsigned index;
 
 
 
 
 
 
 
 
 
 
 
 
670
671	index = ilog2(delta);
672	check_zero();
673
674	if (index < HISTO_BUCKETS)
675		array[index]++;
676	else
677		array[HISTO_BUCKETS]++;
678}
679
680static inline void spin_time_accum_blocked(u64 start)
681{
682	u32 delta;
 
 
683
684	delta = sched_clock() - start;
685	__spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
686	spinlock_stats.time_blocked += delta;
687}
688
689static struct dentry *d_spin_debug;
690static struct dentry *d_kvm_debug;
 
691
692static struct dentry *kvm_init_debugfs(void)
693{
694	d_kvm_debug = debugfs_create_dir("kvm-guest", NULL);
695	if (!d_kvm_debug)
696		printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n");
 
697
698	return d_kvm_debug;
 
 
 
699}
700
701static int __init kvm_spinlock_debugfs(void)
702{
703	struct dentry *d_kvm;
704
705	d_kvm = kvm_init_debugfs();
706	if (d_kvm == NULL)
707		return -ENOMEM;
708
709	d_spin_debug = debugfs_create_dir("spinlocks", d_kvm);
710
711	debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
712
713	debugfs_create_u32("taken_slow", 0444, d_spin_debug,
714		   &spinlock_stats.contention_stats[TAKEN_SLOW]);
715	debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
716		   &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
717
718	debugfs_create_u32("released_slow", 0444, d_spin_debug,
719		   &spinlock_stats.contention_stats[RELEASED_SLOW]);
720	debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
721		   &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
722
723	debugfs_create_u64("time_blocked", 0444, d_spin_debug,
724			   &spinlock_stats.time_blocked);
725
726	debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
727		     spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
728
729	return 0;
730}
731fs_initcall(kvm_spinlock_debugfs);
732#else  /* !CONFIG_KVM_DEBUG_FS */
733static inline void add_stats(enum kvm_contention_stat var, u32 val)
734{
735}
736
737static inline u64 spin_time_start(void)
738{
739	return 0;
740}
741
742static inline void spin_time_accum_blocked(u64 start)
743{
 
744}
745#endif  /* CONFIG_KVM_DEBUG_FS */
746
747struct kvm_lock_waiting {
748	struct arch_spinlock *lock;
749	__ticket_t want;
750};
751
752/* cpus 'waiting' on a spinlock to become available */
753static cpumask_t waiting_cpus;
754
755/* Track spinlock on which a cpu is waiting */
756static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting);
757
758__visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
759{
760	struct kvm_lock_waiting *w;
761	int cpu;
762	u64 start;
763	unsigned long flags;
764	__ticket_t head;
765
766	if (in_nmi())
767		return;
768
769	w = this_cpu_ptr(&klock_waiting);
770	cpu = smp_processor_id();
771	start = spin_time_start();
772
773	/*
774	 * Make sure an interrupt handler can't upset things in a
775	 * partially setup state.
776	 */
777	local_irq_save(flags);
778
779	/*
780	 * The ordering protocol on this is that the "lock" pointer
781	 * may only be set non-NULL if the "want" ticket is correct.
782	 * If we're updating "want", we must first clear "lock".
783	 */
784	w->lock = NULL;
785	smp_wmb();
786	w->want = want;
787	smp_wmb();
788	w->lock = lock;
789
790	add_stats(TAKEN_SLOW, 1);
791
792	/*
793	 * This uses set_bit, which is atomic but we should not rely on its
794	 * reordering gurantees. So barrier is needed after this call.
795	 */
796	cpumask_set_cpu(cpu, &waiting_cpus);
797
798	barrier();
799
800	/*
801	 * Mark entry to slowpath before doing the pickup test to make
802	 * sure we don't deadlock with an unlocker.
803	 */
804	__ticket_enter_slowpath(lock);
805
806	/* make sure enter_slowpath, which is atomic does not cross the read */
807	smp_mb__after_atomic();
808
809	/*
810	 * check again make sure it didn't become free while
811	 * we weren't looking.
812	 */
813	head = READ_ONCE(lock->tickets.head);
814	if (__tickets_equal(head, want)) {
815		add_stats(TAKEN_SLOW_PICKUP, 1);
816		goto out;
817	}
818
819	/*
820	 * halt until it's our turn and kicked. Note that we do safe halt
821	 * for irq enabled case to avoid hang when lock info is overwritten
822	 * in irq spinlock slowpath and no spurious interrupt occur to save us.
823	 */
824	if (arch_irqs_disabled_flags(flags))
825		halt();
826	else
827		safe_halt();
828
829out:
830	cpumask_clear_cpu(cpu, &waiting_cpus);
831	w->lock = NULL;
832	local_irq_restore(flags);
833	spin_time_accum_blocked(start);
834}
835PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning);
836
837/* Kick vcpu waiting on @lock->head to reach value @ticket */
838static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket)
839{
840	int cpu;
 
841
842	add_stats(RELEASED_SLOW, 1);
843	for_each_cpu(cpu, &waiting_cpus) {
844		const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu);
845		if (READ_ONCE(w->lock) == lock &&
846		    READ_ONCE(w->want) == ticket) {
847			add_stats(RELEASED_SLOW_KICKED, 1);
848			kvm_kick_cpu(cpu);
849			break;
850		}
851	}
852}
853
854#endif /* !CONFIG_QUEUED_SPINLOCKS */
855
856/*
857 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
858 */
859void __init kvm_spinlock_init(void)
860{
861	if (!kvm_para_available())
862		return;
863	/* Does host kernel support KVM_FEATURE_PV_UNHALT? */
864	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
865		return;
866
867#ifdef CONFIG_QUEUED_SPINLOCKS
868	__pv_init_lock_hash();
869	pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
870	pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
871	pv_lock_ops.wait = kvm_wait;
872	pv_lock_ops.kick = kvm_kick_cpu;
873#else /* !CONFIG_QUEUED_SPINLOCKS */
874	pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
875	pv_lock_ops.unlock_kick = kvm_unlock_kick;
876#endif
877}
878
879static __init int kvm_spinlock_init_jump(void)
880{
881	if (!kvm_para_available())
882		return 0;
883	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
884		return 0;
885
886	static_key_slow_inc(&paravirt_ticketlocks_enabled);
887	printk(KERN_INFO "KVM setup paravirtual spinlock\n");
888
889	return 0;
890}
891early_initcall(kvm_spinlock_init_jump);
892
893#endif	/* CONFIG_PARAVIRT_SPINLOCKS */