kvm.c revision a63512a4d711c9bd6a5d03847f45fcf88cdea0c6 
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/module.h>
24#include <linux/kernel.h>
25#include <linux/kvm_para.h>
26#include <linux/cpu.h>
27#include <linux/mm.h>
28#include <linux/highmem.h>
29#include <linux/hardirq.h>
30#include <linux/notifier.h>
31#include <linux/reboot.h>
32#include <linux/hash.h>
33#include <linux/sched.h>
34#include <linux/slab.h>
35#include <linux/kprobes.h>
36#include <asm/timer.h>
37#include <asm/cpu.h>
38#include <asm/traps.h>
39#include <asm/desc.h>
40#include <asm/tlbflush.h>
41
42#define MMU_QUEUE_SIZE 1024
43
44static int kvmapf = 1;
45
46static int parse_no_kvmapf(char *arg)
47{
48        kvmapf = 0;
49        return 0;
50}
51
52early_param("no-kvmapf", parse_no_kvmapf);
53
54struct kvm_para_state {
55	u8 mmu_queue[MMU_QUEUE_SIZE];
56	int mmu_queue_len;
57};
58
59static DEFINE_PER_CPU(struct kvm_para_state, para_state);
60static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
61
62static struct kvm_para_state *kvm_para_state(void)
63{
64	return &per_cpu(para_state, raw_smp_processor_id());
65}
66
67/*
68 * No need for any "IO delay" on KVM
69 */
70static void kvm_io_delay(void)
71{
72}
73
74#define KVM_TASK_SLEEP_HASHBITS 8
75#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
76
77struct kvm_task_sleep_node {
78	struct hlist_node link;
79	wait_queue_head_t wq;
80	u32 token;
81	int cpu;
82	bool halted;
83	struct mm_struct *mm;
84};
85
86static struct kvm_task_sleep_head {
87	spinlock_t lock;
88	struct hlist_head list;
89} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
90
91static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
92						  u32 token)
93{
94	struct hlist_node *p;
95
96	hlist_for_each(p, &b->list) {
97		struct kvm_task_sleep_node *n =
98			hlist_entry(p, typeof(*n), link);
99		if (n->token == token)
100			return n;
101	}
102
103	return NULL;
104}
105
106void kvm_async_pf_task_wait(u32 token)
107{
108	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
109	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
110	struct kvm_task_sleep_node n, *e;
111	DEFINE_WAIT(wait);
112	int cpu, idle;
113
114	cpu = get_cpu();
115	idle = idle_cpu(cpu);
116	put_cpu();
117
118	spin_lock(&b->lock);
119	e = _find_apf_task(b, token);
120	if (e) {
121		/* dummy entry exist -> wake up was delivered ahead of PF */
122		hlist_del(&e->link);
123		kfree(e);
124		spin_unlock(&b->lock);
125		return;
126	}
127
128	n.token = token;
129	n.cpu = smp_processor_id();
130	n.mm = current->active_mm;
131	n.halted = idle || preempt_count() > 1;
132	atomic_inc(&n.mm->mm_count);
133	init_waitqueue_head(&n.wq);
134	hlist_add_head(&n.link, &b->list);
135	spin_unlock(&b->lock);
136
137	for (;;) {
138		if (!n.halted)
139			prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
140		if (hlist_unhashed(&n.link))
141			break;
142
143		if (!n.halted) {
144			local_irq_enable();
145			schedule();
146			local_irq_disable();
147		} else {
148			/*
149			 * We cannot reschedule. So halt.
150			 */
151			native_safe_halt();
152			local_irq_disable();
153		}
154	}
155	if (!n.halted)
156		finish_wait(&n.wq, &wait);
157
158	return;
159}
160EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
161
162static void apf_task_wake_one(struct kvm_task_sleep_node *n)
163{
164	hlist_del_init(&n->link);
165	if (!n->mm)
166		return;
167	mmdrop(n->mm);
168	if (n->halted)
169		smp_send_reschedule(n->cpu);
170	else if (waitqueue_active(&n->wq))
171		wake_up(&n->wq);
172}
173
174static void apf_task_wake_all(void)
175{
176	int i;
177
178	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
179		struct hlist_node *p, *next;
180		struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
181		spin_lock(&b->lock);
182		hlist_for_each_safe(p, next, &b->list) {
183			struct kvm_task_sleep_node *n =
184				hlist_entry(p, typeof(*n), link);
185			if (n->cpu == smp_processor_id())
186				apf_task_wake_one(n);
187		}
188		spin_unlock(&b->lock);
189	}
190}
191
192void kvm_async_pf_task_wake(u32 token)
193{
194	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
195	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
196	struct kvm_task_sleep_node *n;
197
198	if (token == ~0) {
199		apf_task_wake_all();
200		return;
201	}
202
203again:
204	spin_lock(&b->lock);
205	n = _find_apf_task(b, token);
206	if (!n) {
207		/*
208		 * async PF was not yet handled.
209		 * Add dummy entry for the token.
210		 */
211		n = kmalloc(sizeof(*n), GFP_ATOMIC);
212		if (!n) {
213			/*
214			 * Allocation failed! Busy wait while other cpu
215			 * handles async PF.
216			 */
217			spin_unlock(&b->lock);
218			cpu_relax();
219			goto again;
220		}
221		n->token = token;
222		n->cpu = smp_processor_id();
223		n->mm = NULL;
224		init_waitqueue_head(&n->wq);
225		hlist_add_head(&n->link, &b->list);
226	} else
227		apf_task_wake_one(n);
228	spin_unlock(&b->lock);
229	return;
230}
231EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
232
233u32 kvm_read_and_reset_pf_reason(void)
234{
235	u32 reason = 0;
236
237	if (__get_cpu_var(apf_reason).enabled) {
238		reason = __get_cpu_var(apf_reason).reason;
239		__get_cpu_var(apf_reason).reason = 0;
240	}
241
242	return reason;
243}
244EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
245
246dotraplinkage void __kprobes
247do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
248{
249	switch (kvm_read_and_reset_pf_reason()) {
250	default:
251		do_page_fault(regs, error_code);
252		break;
253	case KVM_PV_REASON_PAGE_NOT_PRESENT:
254		/* page is swapped out by the host. */
255		kvm_async_pf_task_wait((u32)read_cr2());
256		break;
257	case KVM_PV_REASON_PAGE_READY:
258		kvm_async_pf_task_wake((u32)read_cr2());
259		break;
260	}
261}
262
263static void kvm_mmu_op(void *buffer, unsigned len)
264{
265	int r;
266	unsigned long a1, a2;
267
268	do {
269		a1 = __pa(buffer);
270		a2 = 0;   /* on i386 __pa() always returns <4G */
271		r = kvm_hypercall3(KVM_HC_MMU_OP, len, a1, a2);
272		buffer += r;
273		len -= r;
274	} while (len);
275}
276
277static void mmu_queue_flush(struct kvm_para_state *state)
278{
279	if (state->mmu_queue_len) {
280		kvm_mmu_op(state->mmu_queue, state->mmu_queue_len);
281		state->mmu_queue_len = 0;
282	}
283}
284
285static void kvm_deferred_mmu_op(void *buffer, int len)
286{
287	struct kvm_para_state *state = kvm_para_state();
288
289	if (paravirt_get_lazy_mode() != PARAVIRT_LAZY_MMU) {
290		kvm_mmu_op(buffer, len);
291		return;
292	}
293	if (state->mmu_queue_len + len > sizeof state->mmu_queue)
294		mmu_queue_flush(state);
295	memcpy(state->mmu_queue + state->mmu_queue_len, buffer, len);
296	state->mmu_queue_len += len;
297}
298
299static void kvm_mmu_write(void *dest, u64 val)
300{
301	__u64 pte_phys;
302	struct kvm_mmu_op_write_pte wpte;
303
304#ifdef CONFIG_HIGHPTE
305	struct page *page;
306	unsigned long dst = (unsigned long) dest;
307
308	page = kmap_atomic_to_page(dest);
309	pte_phys = page_to_pfn(page);
310	pte_phys <<= PAGE_SHIFT;
311	pte_phys += (dst & ~(PAGE_MASK));
312#else
313	pte_phys = (unsigned long)__pa(dest);
314#endif
315	wpte.header.op = KVM_MMU_OP_WRITE_PTE;
316	wpte.pte_val = val;
317	wpte.pte_phys = pte_phys;
318
319	kvm_deferred_mmu_op(&wpte, sizeof wpte);
320}
321
322/*
323 * We only need to hook operations that are MMU writes.  We hook these so that
324 * we can use lazy MMU mode to batch these operations.  We could probably
325 * improve the performance of the host code if we used some of the information
326 * here to simplify processing of batched writes.
327 */
328static void kvm_set_pte(pte_t *ptep, pte_t pte)
329{
330	kvm_mmu_write(ptep, pte_val(pte));
331}
332
333static void kvm_set_pte_at(struct mm_struct *mm, unsigned long addr,
334			   pte_t *ptep, pte_t pte)
335{
336	kvm_mmu_write(ptep, pte_val(pte));
337}
338
339static void kvm_set_pmd(pmd_t *pmdp, pmd_t pmd)
340{
341	kvm_mmu_write(pmdp, pmd_val(pmd));
342}
343
344#if PAGETABLE_LEVELS >= 3
345#ifdef CONFIG_X86_PAE
346static void kvm_set_pte_atomic(pte_t *ptep, pte_t pte)
347{
348	kvm_mmu_write(ptep, pte_val(pte));
349}
350
351static void kvm_pte_clear(struct mm_struct *mm,
352			  unsigned long addr, pte_t *ptep)
353{
354	kvm_mmu_write(ptep, 0);
355}
356
357static void kvm_pmd_clear(pmd_t *pmdp)
358{
359	kvm_mmu_write(pmdp, 0);
360}
361#endif
362
363static void kvm_set_pud(pud_t *pudp, pud_t pud)
364{
365	kvm_mmu_write(pudp, pud_val(pud));
366}
367
368#if PAGETABLE_LEVELS == 4
369static void kvm_set_pgd(pgd_t *pgdp, pgd_t pgd)
370{
371	kvm_mmu_write(pgdp, pgd_val(pgd));
372}
373#endif
374#endif /* PAGETABLE_LEVELS >= 3 */
375
376static void kvm_flush_tlb(void)
377{
378	struct kvm_mmu_op_flush_tlb ftlb = {
379		.header.op = KVM_MMU_OP_FLUSH_TLB,
380	};
381
382	kvm_deferred_mmu_op(&ftlb, sizeof ftlb);
383}
384
385static void kvm_release_pt(unsigned long pfn)
386{
387	struct kvm_mmu_op_release_pt rpt = {
388		.header.op = KVM_MMU_OP_RELEASE_PT,
389		.pt_phys = (u64)pfn << PAGE_SHIFT,
390	};
391
392	kvm_mmu_op(&rpt, sizeof rpt);
393}
394
395static void kvm_enter_lazy_mmu(void)
396{
397	paravirt_enter_lazy_mmu();
398}
399
400static void kvm_leave_lazy_mmu(void)
401{
402	struct kvm_para_state *state = kvm_para_state();
403
404	mmu_queue_flush(state);
405	paravirt_leave_lazy_mmu();
406}
407
408static void __init paravirt_ops_setup(void)
409{
410	pv_info.name = "KVM";
411	pv_info.paravirt_enabled = 1;
412
413	if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
414		pv_cpu_ops.io_delay = kvm_io_delay;
415
416	if (kvm_para_has_feature(KVM_FEATURE_MMU_OP)) {
417		pv_mmu_ops.set_pte = kvm_set_pte;
418		pv_mmu_ops.set_pte_at = kvm_set_pte_at;
419		pv_mmu_ops.set_pmd = kvm_set_pmd;
420#if PAGETABLE_LEVELS >= 3
421#ifdef CONFIG_X86_PAE
422		pv_mmu_ops.set_pte_atomic = kvm_set_pte_atomic;
423		pv_mmu_ops.pte_clear = kvm_pte_clear;
424		pv_mmu_ops.pmd_clear = kvm_pmd_clear;
425#endif
426		pv_mmu_ops.set_pud = kvm_set_pud;
427#if PAGETABLE_LEVELS == 4
428		pv_mmu_ops.set_pgd = kvm_set_pgd;
429#endif
430#endif
431		pv_mmu_ops.flush_tlb_user = kvm_flush_tlb;
432		pv_mmu_ops.release_pte = kvm_release_pt;
433		pv_mmu_ops.release_pmd = kvm_release_pt;
434		pv_mmu_ops.release_pud = kvm_release_pt;
435
436		pv_mmu_ops.lazy_mode.enter = kvm_enter_lazy_mmu;
437		pv_mmu_ops.lazy_mode.leave = kvm_leave_lazy_mmu;
438	}
439#ifdef CONFIG_X86_IO_APIC
440	no_timer_check = 1;
441#endif
442}
443
444void __cpuinit kvm_guest_cpu_init(void)
445{
446	if (!kvm_para_available())
447		return;
448
449	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
450		u64 pa = __pa(&__get_cpu_var(apf_reason));
451
452#ifdef CONFIG_PREEMPT
453		pa |= KVM_ASYNC_PF_SEND_ALWAYS;
454#endif
455		wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
456		__get_cpu_var(apf_reason).enabled = 1;
457		printk(KERN_INFO"KVM setup async PF for cpu %d\n",
458		       smp_processor_id());
459	}
460}
461
462static void kvm_pv_disable_apf(void *unused)
463{
464	if (!__get_cpu_var(apf_reason).enabled)
465		return;
466
467	wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
468	__get_cpu_var(apf_reason).enabled = 0;
469
470	printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
471	       smp_processor_id());
472}
473
474static int kvm_pv_reboot_notify(struct notifier_block *nb,
475				unsigned long code, void *unused)
476{
477	if (code == SYS_RESTART)
478		on_each_cpu(kvm_pv_disable_apf, NULL, 1);
479	return NOTIFY_DONE;
480}
481
482static struct notifier_block kvm_pv_reboot_nb = {
483	.notifier_call = kvm_pv_reboot_notify,
484};
485
486#ifdef CONFIG_SMP
487static void __init kvm_smp_prepare_boot_cpu(void)
488{
489#ifdef CONFIG_KVM_CLOCK
490	WARN_ON(kvm_register_clock("primary cpu clock"));
491#endif
492	kvm_guest_cpu_init();
493	native_smp_prepare_boot_cpu();
494}
495
496static void kvm_guest_cpu_online(void *dummy)
497{
498	kvm_guest_cpu_init();
499}
500
501static void kvm_guest_cpu_offline(void *dummy)
502{
503	kvm_pv_disable_apf(NULL);
504	apf_task_wake_all();
505}
506
507static int __cpuinit kvm_cpu_notify(struct notifier_block *self,
508				    unsigned long action, void *hcpu)
509{
510	int cpu = (unsigned long)hcpu;
511	switch (action) {
512	case CPU_ONLINE:
513	case CPU_DOWN_FAILED:
514	case CPU_ONLINE_FROZEN:
515		smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
516		break;
517	case CPU_DOWN_PREPARE:
518	case CPU_DOWN_PREPARE_FROZEN:
519		smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
520		break;
521	default:
522		break;
523	}
524	return NOTIFY_OK;
525}
526
527static struct notifier_block __cpuinitdata kvm_cpu_notifier = {
528        .notifier_call  = kvm_cpu_notify,
529};
530#endif
531
532static void __init kvm_apf_trap_init(void)
533{
534	set_intr_gate(14, &async_page_fault);
535}
536
537void __init kvm_guest_init(void)
538{
539	int i;
540
541	if (!kvm_para_available())
542		return;
543
544	paravirt_ops_setup();
545	register_reboot_notifier(&kvm_pv_reboot_nb);
546	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
547		spin_lock_init(&async_pf_sleepers[i].lock);
548	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
549		x86_init.irqs.trap_init = kvm_apf_trap_init;
550
551#ifdef CONFIG_SMP
552	smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
553	register_cpu_notifier(&kvm_cpu_notifier);
554#else
555	kvm_guest_cpu_init();
556#endif
557}
558