• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 <asm/timer.h>
40 #include <asm/cpu.h>
41 #include <asm/traps.h>
42 #include <asm/desc.h>
43 #include <asm/tlbflush.h>
44 #include <asm/idle.h>
45 #include <asm/apic.h>
46 #include <asm/apicdef.h>
47 #include <asm/hypervisor.h>
48 #include <asm/kvm_guest.h>
49 
50 static int kvmapf = 1;
51 
parse_no_kvmapf(char * arg)52 static int parse_no_kvmapf(char *arg)
53 {
54         kvmapf = 0;
55         return 0;
56 }
57 
58 early_param("no-kvmapf", parse_no_kvmapf);
59 
60 static int steal_acc = 1;
parse_no_stealacc(char * arg)61 static int parse_no_stealacc(char *arg)
62 {
63         steal_acc = 0;
64         return 0;
65 }
66 
67 early_param("no-steal-acc", parse_no_stealacc);
68 
69 static int kvmclock_vsyscall = 1;
parse_no_kvmclock_vsyscall(char * arg)70 static int parse_no_kvmclock_vsyscall(char *arg)
71 {
72         kvmclock_vsyscall = 0;
73         return 0;
74 }
75 
76 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
77 
78 static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
79 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
80 static int has_steal_clock = 0;
81 
82 /*
83  * No need for any "IO delay" on KVM
84  */
kvm_io_delay(void)85 static 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 
92 struct kvm_task_sleep_node {
93 	struct hlist_node link;
94 	wait_queue_head_t wq;
95 	u32 token;
96 	int cpu;
97 	bool halted;
98 };
99 
100 static struct kvm_task_sleep_head {
101 	spinlock_t lock;
102 	struct hlist_head list;
103 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
104 
_find_apf_task(struct kvm_task_sleep_head * b,u32 token)105 static 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 
kvm_async_pf_task_wait(u32 token)120 void kvm_async_pf_task_wait(u32 token)
121 {
122 	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
123 	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
124 	struct kvm_task_sleep_node n, *e;
125 	DEFINE_WAIT(wait);
126 
127 	rcu_irq_enter();
128 
129 	spin_lock(&b->lock);
130 	e = _find_apf_task(b, token);
131 	if (e) {
132 		/* dummy entry exist -> wake up was delivered ahead of PF */
133 		hlist_del(&e->link);
134 		kfree(e);
135 		spin_unlock(&b->lock);
136 
137 		rcu_irq_exit();
138 		return;
139 	}
140 
141 	n.token = token;
142 	n.cpu = smp_processor_id();
143 	n.halted = is_idle_task(current) || preempt_count() > 1;
144 	init_waitqueue_head(&n.wq);
145 	hlist_add_head(&n.link, &b->list);
146 	spin_unlock(&b->lock);
147 
148 	for (;;) {
149 		if (!n.halted)
150 			prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
151 		if (hlist_unhashed(&n.link))
152 			break;
153 
154 		rcu_irq_exit();
155 
156 		if (!n.halted) {
157 			local_irq_enable();
158 			schedule();
159 			local_irq_disable();
160 		} else {
161 			/*
162 			 * We cannot reschedule. So halt.
163 			 */
164 			native_safe_halt();
165 			local_irq_disable();
166 		}
167 
168 		rcu_irq_enter();
169 	}
170 	if (!n.halted)
171 		finish_wait(&n.wq, &wait);
172 
173 	rcu_irq_exit();
174 	return;
175 }
176 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
177 
apf_task_wake_one(struct kvm_task_sleep_node * n)178 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
179 {
180 	hlist_del_init(&n->link);
181 	if (n->halted)
182 		smp_send_reschedule(n->cpu);
183 	else if (waitqueue_active(&n->wq))
184 		wake_up(&n->wq);
185 }
186 
apf_task_wake_all(void)187 static void apf_task_wake_all(void)
188 {
189 	int i;
190 
191 	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
192 		struct hlist_node *p, *next;
193 		struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
194 		spin_lock(&b->lock);
195 		hlist_for_each_safe(p, next, &b->list) {
196 			struct kvm_task_sleep_node *n =
197 				hlist_entry(p, typeof(*n), link);
198 			if (n->cpu == smp_processor_id())
199 				apf_task_wake_one(n);
200 		}
201 		spin_unlock(&b->lock);
202 	}
203 }
204 
kvm_async_pf_task_wake(u32 token)205 void kvm_async_pf_task_wake(u32 token)
206 {
207 	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
208 	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
209 	struct kvm_task_sleep_node *n;
210 
211 	if (token == ~0) {
212 		apf_task_wake_all();
213 		return;
214 	}
215 
216 again:
217 	spin_lock(&b->lock);
218 	n = _find_apf_task(b, token);
219 	if (!n) {
220 		/*
221 		 * async PF was not yet handled.
222 		 * Add dummy entry for the token.
223 		 */
224 		n = kzalloc(sizeof(*n), GFP_ATOMIC);
225 		if (!n) {
226 			/*
227 			 * Allocation failed! Busy wait while other cpu
228 			 * handles async PF.
229 			 */
230 			spin_unlock(&b->lock);
231 			cpu_relax();
232 			goto again;
233 		}
234 		n->token = token;
235 		n->cpu = smp_processor_id();
236 		init_waitqueue_head(&n->wq);
237 		hlist_add_head(&n->link, &b->list);
238 	} else
239 		apf_task_wake_one(n);
240 	spin_unlock(&b->lock);
241 	return;
242 }
243 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
244 
kvm_read_and_reset_pf_reason(void)245 u32 kvm_read_and_reset_pf_reason(void)
246 {
247 	u32 reason = 0;
248 
249 	if (__this_cpu_read(apf_reason.enabled)) {
250 		reason = __this_cpu_read(apf_reason.reason);
251 		__this_cpu_write(apf_reason.reason, 0);
252 	}
253 
254 	return reason;
255 }
256 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
257 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
258 
259 dotraplinkage void
do_async_page_fault(struct pt_regs * regs,unsigned long error_code)260 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
261 {
262 	enum ctx_state prev_state;
263 
264 	switch (kvm_read_and_reset_pf_reason()) {
265 	default:
266 		trace_do_page_fault(regs, error_code);
267 		break;
268 	case KVM_PV_REASON_PAGE_NOT_PRESENT:
269 		/* page is swapped out by the host. */
270 		prev_state = exception_enter();
271 		exit_idle();
272 		kvm_async_pf_task_wait((u32)read_cr2());
273 		exception_exit(prev_state);
274 		break;
275 	case KVM_PV_REASON_PAGE_READY:
276 		rcu_irq_enter();
277 		exit_idle();
278 		kvm_async_pf_task_wake((u32)read_cr2());
279 		rcu_irq_exit();
280 		break;
281 	}
282 }
283 NOKPROBE_SYMBOL(do_async_page_fault);
284 
paravirt_ops_setup(void)285 static void __init paravirt_ops_setup(void)
286 {
287 	pv_info.name = "KVM";
288 
289 	/*
290 	 * KVM isn't paravirt in the sense of paravirt_enabled.  A KVM
291 	 * guest kernel works like a bare metal kernel with additional
292 	 * features, and paravirt_enabled is about features that are
293 	 * missing.
294 	 */
295 	pv_info.paravirt_enabled = 0;
296 
297 	if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
298 		pv_cpu_ops.io_delay = kvm_io_delay;
299 
300 #ifdef CONFIG_X86_IO_APIC
301 	no_timer_check = 1;
302 #endif
303 }
304 
kvm_register_steal_time(void)305 static void kvm_register_steal_time(void)
306 {
307 	int cpu = smp_processor_id();
308 	struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
309 
310 	if (!has_steal_clock)
311 		return;
312 
313 	memset(st, 0, sizeof(*st));
314 
315 	wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
316 	pr_info("kvm-stealtime: cpu %d, msr %llx\n",
317 		cpu, (unsigned long long) slow_virt_to_phys(st));
318 }
319 
320 static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
321 
kvm_guest_apic_eoi_write(u32 reg,u32 val)322 static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
323 {
324 	/**
325 	 * This relies on __test_and_clear_bit to modify the memory
326 	 * in a way that is atomic with respect to the local CPU.
327 	 * The hypervisor only accesses this memory from the local CPU so
328 	 * there's no need for lock or memory barriers.
329 	 * An optimization barrier is implied in apic write.
330 	 */
331 	if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
332 		return;
333 	apic_write(APIC_EOI, APIC_EOI_ACK);
334 }
335 
kvm_guest_cpu_init(void)336 void kvm_guest_cpu_init(void)
337 {
338 	if (!kvm_para_available())
339 		return;
340 
341 	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
342 		u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
343 
344 #ifdef CONFIG_PREEMPT
345 		pa |= KVM_ASYNC_PF_SEND_ALWAYS;
346 #endif
347 		wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
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 
kvm_pv_disable_apf(void)367 static 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 
kvm_pv_guest_cpu_reboot(void * unused)379 static 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 
kvm_pv_reboot_notify(struct notifier_block * nb,unsigned long code,void * unused)392 static 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 
400 static struct notifier_block kvm_pv_reboot_nb = {
401 	.notifier_call = kvm_pv_reboot_notify,
402 };
403 
kvm_steal_clock(int cpu)404 static 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 		rmb();
414 		steal = src->steal;
415 		rmb();
416 	} while ((version & 1) || (version != src->version));
417 
418 	return steal;
419 }
420 
kvm_disable_steal_time(void)421 void kvm_disable_steal_time(void)
422 {
423 	if (!has_steal_clock)
424 		return;
425 
426 	wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
427 }
428 
429 #ifdef CONFIG_SMP
kvm_smp_prepare_boot_cpu(void)430 static void __init kvm_smp_prepare_boot_cpu(void)
431 {
432 	kvm_guest_cpu_init();
433 	native_smp_prepare_boot_cpu();
434 	kvm_spinlock_init();
435 }
436 
kvm_guest_cpu_online(void * dummy)437 static void kvm_guest_cpu_online(void *dummy)
438 {
439 	kvm_guest_cpu_init();
440 }
441 
kvm_guest_cpu_offline(void * dummy)442 static void kvm_guest_cpu_offline(void *dummy)
443 {
444 	kvm_disable_steal_time();
445 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
446 		wrmsrl(MSR_KVM_PV_EOI_EN, 0);
447 	kvm_pv_disable_apf();
448 	apf_task_wake_all();
449 }
450 
kvm_cpu_notify(struct notifier_block * self,unsigned long action,void * hcpu)451 static int kvm_cpu_notify(struct notifier_block *self, unsigned long action,
452 			  void *hcpu)
453 {
454 	int cpu = (unsigned long)hcpu;
455 	switch (action) {
456 	case CPU_ONLINE:
457 	case CPU_DOWN_FAILED:
458 	case CPU_ONLINE_FROZEN:
459 		smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
460 		break;
461 	case CPU_DOWN_PREPARE:
462 	case CPU_DOWN_PREPARE_FROZEN:
463 		smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
464 		break;
465 	default:
466 		break;
467 	}
468 	return NOTIFY_OK;
469 }
470 
471 static struct notifier_block kvm_cpu_notifier = {
472         .notifier_call  = kvm_cpu_notify,
473 };
474 #endif
475 
kvm_apf_trap_init(void)476 static void __init kvm_apf_trap_init(void)
477 {
478 	set_intr_gate(14, async_page_fault);
479 }
480 
kvm_guest_init(void)481 void __init kvm_guest_init(void)
482 {
483 	int i;
484 
485 	if (!kvm_para_available())
486 		return;
487 
488 	paravirt_ops_setup();
489 	register_reboot_notifier(&kvm_pv_reboot_nb);
490 	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
491 		spin_lock_init(&async_pf_sleepers[i].lock);
492 	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
493 		x86_init.irqs.trap_init = kvm_apf_trap_init;
494 
495 	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
496 		has_steal_clock = 1;
497 		pv_time_ops.steal_clock = kvm_steal_clock;
498 	}
499 
500 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
501 		apic_set_eoi_write(kvm_guest_apic_eoi_write);
502 
503 	if (kvmclock_vsyscall)
504 		kvm_setup_vsyscall_timeinfo();
505 
506 #ifdef CONFIG_SMP
507 	smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
508 	register_cpu_notifier(&kvm_cpu_notifier);
509 #else
510 	kvm_guest_cpu_init();
511 #endif
512 
513 	/*
514 	 * Hard lockup detection is enabled by default. Disable it, as guests
515 	 * can get false positives too easily, for example if the host is
516 	 * overcommitted.
517 	 */
518 	watchdog_enable_hardlockup_detector(false);
519 }
520 
__kvm_cpuid_base(void)521 static noinline uint32_t __kvm_cpuid_base(void)
522 {
523 	if (boot_cpu_data.cpuid_level < 0)
524 		return 0;	/* So we don't blow up on old processors */
525 
526 	if (cpu_has_hypervisor)
527 		return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
528 
529 	return 0;
530 }
531 
kvm_cpuid_base(void)532 static inline uint32_t kvm_cpuid_base(void)
533 {
534 	static int kvm_cpuid_base = -1;
535 
536 	if (kvm_cpuid_base == -1)
537 		kvm_cpuid_base = __kvm_cpuid_base();
538 
539 	return kvm_cpuid_base;
540 }
541 
kvm_para_available(void)542 bool kvm_para_available(void)
543 {
544 	return kvm_cpuid_base() != 0;
545 }
546 EXPORT_SYMBOL_GPL(kvm_para_available);
547 
kvm_arch_para_features(void)548 unsigned int kvm_arch_para_features(void)
549 {
550 	return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
551 }
552 
kvm_detect(void)553 static uint32_t __init kvm_detect(void)
554 {
555 	return kvm_cpuid_base();
556 }
557 
558 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
559 	.name			= "KVM",
560 	.detect			= kvm_detect,
561 	.x2apic_available	= kvm_para_available,
562 };
563 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
564 
activate_jump_labels(void)565 static __init int activate_jump_labels(void)
566 {
567 	if (has_steal_clock) {
568 		static_key_slow_inc(&paravirt_steal_enabled);
569 		if (steal_acc)
570 			static_key_slow_inc(&paravirt_steal_rq_enabled);
571 	}
572 
573 	return 0;
574 }
575 arch_initcall(activate_jump_labels);
576 
577 #ifdef CONFIG_PARAVIRT_SPINLOCKS
578 
579 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
kvm_kick_cpu(int cpu)580 static void kvm_kick_cpu(int cpu)
581 {
582 	int apicid;
583 	unsigned long flags = 0;
584 
585 	apicid = per_cpu(x86_cpu_to_apicid, cpu);
586 	kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
587 }
588 
589 enum kvm_contention_stat {
590 	TAKEN_SLOW,
591 	TAKEN_SLOW_PICKUP,
592 	RELEASED_SLOW,
593 	RELEASED_SLOW_KICKED,
594 	NR_CONTENTION_STATS
595 };
596 
597 #ifdef CONFIG_KVM_DEBUG_FS
598 #define HISTO_BUCKETS	30
599 
600 static struct kvm_spinlock_stats
601 {
602 	u32 contention_stats[NR_CONTENTION_STATS];
603 	u32 histo_spin_blocked[HISTO_BUCKETS+1];
604 	u64 time_blocked;
605 } spinlock_stats;
606 
607 static u8 zero_stats;
608 
check_zero(void)609 static inline void check_zero(void)
610 {
611 	u8 ret;
612 	u8 old;
613 
614 	old = ACCESS_ONCE(zero_stats);
615 	if (unlikely(old)) {
616 		ret = cmpxchg(&zero_stats, old, 0);
617 		/* This ensures only one fellow resets the stat */
618 		if (ret == old)
619 			memset(&spinlock_stats, 0, sizeof(spinlock_stats));
620 	}
621 }
622 
add_stats(enum kvm_contention_stat var,u32 val)623 static inline void add_stats(enum kvm_contention_stat var, u32 val)
624 {
625 	check_zero();
626 	spinlock_stats.contention_stats[var] += val;
627 }
628 
629 
spin_time_start(void)630 static inline u64 spin_time_start(void)
631 {
632 	return sched_clock();
633 }
634 
__spin_time_accum(u64 delta,u32 * array)635 static void __spin_time_accum(u64 delta, u32 *array)
636 {
637 	unsigned index;
638 
639 	index = ilog2(delta);
640 	check_zero();
641 
642 	if (index < HISTO_BUCKETS)
643 		array[index]++;
644 	else
645 		array[HISTO_BUCKETS]++;
646 }
647 
spin_time_accum_blocked(u64 start)648 static inline void spin_time_accum_blocked(u64 start)
649 {
650 	u32 delta;
651 
652 	delta = sched_clock() - start;
653 	__spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
654 	spinlock_stats.time_blocked += delta;
655 }
656 
657 static struct dentry *d_spin_debug;
658 static struct dentry *d_kvm_debug;
659 
kvm_init_debugfs(void)660 struct dentry *kvm_init_debugfs(void)
661 {
662 	d_kvm_debug = debugfs_create_dir("kvm-guest", NULL);
663 	if (!d_kvm_debug)
664 		printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n");
665 
666 	return d_kvm_debug;
667 }
668 
kvm_spinlock_debugfs(void)669 static int __init kvm_spinlock_debugfs(void)
670 {
671 	struct dentry *d_kvm;
672 
673 	d_kvm = kvm_init_debugfs();
674 	if (d_kvm == NULL)
675 		return -ENOMEM;
676 
677 	d_spin_debug = debugfs_create_dir("spinlocks", d_kvm);
678 
679 	debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
680 
681 	debugfs_create_u32("taken_slow", 0444, d_spin_debug,
682 		   &spinlock_stats.contention_stats[TAKEN_SLOW]);
683 	debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
684 		   &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
685 
686 	debugfs_create_u32("released_slow", 0444, d_spin_debug,
687 		   &spinlock_stats.contention_stats[RELEASED_SLOW]);
688 	debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
689 		   &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
690 
691 	debugfs_create_u64("time_blocked", 0444, d_spin_debug,
692 			   &spinlock_stats.time_blocked);
693 
694 	debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
695 		     spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
696 
697 	return 0;
698 }
699 fs_initcall(kvm_spinlock_debugfs);
700 #else  /* !CONFIG_KVM_DEBUG_FS */
add_stats(enum kvm_contention_stat var,u32 val)701 static inline void add_stats(enum kvm_contention_stat var, u32 val)
702 {
703 }
704 
spin_time_start(void)705 static inline u64 spin_time_start(void)
706 {
707 	return 0;
708 }
709 
spin_time_accum_blocked(u64 start)710 static inline void spin_time_accum_blocked(u64 start)
711 {
712 }
713 #endif  /* CONFIG_KVM_DEBUG_FS */
714 
715 struct kvm_lock_waiting {
716 	struct arch_spinlock *lock;
717 	__ticket_t want;
718 };
719 
720 /* cpus 'waiting' on a spinlock to become available */
721 static cpumask_t waiting_cpus;
722 
723 /* Track spinlock on which a cpu is waiting */
724 static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting);
725 
kvm_lock_spinning(struct arch_spinlock * lock,__ticket_t want)726 __visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
727 {
728 	struct kvm_lock_waiting *w;
729 	int cpu;
730 	u64 start;
731 	unsigned long flags;
732 
733 	if (in_nmi())
734 		return;
735 
736 	w = this_cpu_ptr(&klock_waiting);
737 	cpu = smp_processor_id();
738 	start = spin_time_start();
739 
740 	/*
741 	 * Make sure an interrupt handler can't upset things in a
742 	 * partially setup state.
743 	 */
744 	local_irq_save(flags);
745 
746 	/*
747 	 * The ordering protocol on this is that the "lock" pointer
748 	 * may only be set non-NULL if the "want" ticket is correct.
749 	 * If we're updating "want", we must first clear "lock".
750 	 */
751 	w->lock = NULL;
752 	smp_wmb();
753 	w->want = want;
754 	smp_wmb();
755 	w->lock = lock;
756 
757 	add_stats(TAKEN_SLOW, 1);
758 
759 	/*
760 	 * This uses set_bit, which is atomic but we should not rely on its
761 	 * reordering gurantees. So barrier is needed after this call.
762 	 */
763 	cpumask_set_cpu(cpu, &waiting_cpus);
764 
765 	barrier();
766 
767 	/*
768 	 * Mark entry to slowpath before doing the pickup test to make
769 	 * sure we don't deadlock with an unlocker.
770 	 */
771 	__ticket_enter_slowpath(lock);
772 
773 	/*
774 	 * check again make sure it didn't become free while
775 	 * we weren't looking.
776 	 */
777 	if (ACCESS_ONCE(lock->tickets.head) == want) {
778 		add_stats(TAKEN_SLOW_PICKUP, 1);
779 		goto out;
780 	}
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 
792 out:
793 	cpumask_clear_cpu(cpu, &waiting_cpus);
794 	w->lock = NULL;
795 	local_irq_restore(flags);
796 	spin_time_accum_blocked(start);
797 }
798 PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning);
799 
800 /* Kick vcpu waiting on @lock->head to reach value @ticket */
kvm_unlock_kick(struct arch_spinlock * lock,__ticket_t ticket)801 static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket)
802 {
803 	int cpu;
804 
805 	add_stats(RELEASED_SLOW, 1);
806 	for_each_cpu(cpu, &waiting_cpus) {
807 		const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu);
808 		if (ACCESS_ONCE(w->lock) == lock &&
809 		    ACCESS_ONCE(w->want) == ticket) {
810 			add_stats(RELEASED_SLOW_KICKED, 1);
811 			kvm_kick_cpu(cpu);
812 			break;
813 		}
814 	}
815 }
816 
817 /*
818  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
819  */
kvm_spinlock_init(void)820 void __init kvm_spinlock_init(void)
821 {
822 	if (!kvm_para_available())
823 		return;
824 	/* Does host kernel support KVM_FEATURE_PV_UNHALT? */
825 	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
826 		return;
827 
828 	pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
829 	pv_lock_ops.unlock_kick = kvm_unlock_kick;
830 }
831 
kvm_spinlock_init_jump(void)832 static __init int kvm_spinlock_init_jump(void)
833 {
834 	if (!kvm_para_available())
835 		return 0;
836 	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
837 		return 0;
838 
839 	static_key_slow_inc(&paravirt_ticketlocks_enabled);
840 	printk(KERN_INFO "KVM setup paravirtual spinlock\n");
841 
842 	return 0;
843 }
844 early_initcall(kvm_spinlock_init_jump);
845 
846 #endif	/* CONFIG_PARAVIRT_SPINLOCKS */
847