kvm.c revision 631bc4878220932fe67fc46fc7cf7cccdb1ec597 
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
41#define MMU_QUEUE_SIZE 1024
42
43static int kvmapf = 1;
44
45static int parse_no_kvmapf(char *arg)
46{
47        kvmapf = 0;
48        return 0;
49}
50
51early_param("no-kvmapf", parse_no_kvmapf);
52
53struct kvm_para_state {
54	u8 mmu_queue[MMU_QUEUE_SIZE];
55	int mmu_queue_len;
56};
57
58static DEFINE_PER_CPU(struct kvm_para_state, para_state);
59static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
60
61static struct kvm_para_state *kvm_para_state(void)
62{
63	return &per_cpu(para_state, raw_smp_processor_id());
64}
65
66/*
67 * No need for any "IO delay" on KVM
68 */
69static void kvm_io_delay(void)
70{
71}
72
73#define KVM_TASK_SLEEP_HASHBITS 8
74#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
75
76struct kvm_task_sleep_node {
77	struct hlist_node link;
78	wait_queue_head_t wq;
79	u32 token;
80	int cpu;
81};
82
83static struct kvm_task_sleep_head {
84	spinlock_t lock;
85	struct hlist_head list;
86} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
87
88static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
89						  u32 token)
90{
91	struct hlist_node *p;
92
93	hlist_for_each(p, &b->list) {
94		struct kvm_task_sleep_node *n =
95			hlist_entry(p, typeof(*n), link);
96		if (n->token == token)
97			return n;
98	}
99
100	return NULL;
101}
102
103void kvm_async_pf_task_wait(u32 token)
104{
105	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
106	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
107	struct kvm_task_sleep_node n, *e;
108	DEFINE_WAIT(wait);
109
110	spin_lock(&b->lock);
111	e = _find_apf_task(b, token);
112	if (e) {
113		/* dummy entry exist -> wake up was delivered ahead of PF */
114		hlist_del(&e->link);
115		kfree(e);
116		spin_unlock(&b->lock);
117		return;
118	}
119
120	n.token = token;
121	n.cpu = smp_processor_id();
122	init_waitqueue_head(&n.wq);
123	hlist_add_head(&n.link, &b->list);
124	spin_unlock(&b->lock);
125
126	for (;;) {
127		prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
128		if (hlist_unhashed(&n.link))
129			break;
130		local_irq_enable();
131		schedule();
132		local_irq_disable();
133	}
134	finish_wait(&n.wq, &wait);
135
136	return;
137}
138EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
139
140static void apf_task_wake_one(struct kvm_task_sleep_node *n)
141{
142	hlist_del_init(&n->link);
143	if (waitqueue_active(&n->wq))
144		wake_up(&n->wq);
145}
146
147static void apf_task_wake_all(void)
148{
149	int i;
150
151	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
152		struct hlist_node *p, *next;
153		struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
154		spin_lock(&b->lock);
155		hlist_for_each_safe(p, next, &b->list) {
156			struct kvm_task_sleep_node *n =
157				hlist_entry(p, typeof(*n), link);
158			if (n->cpu == smp_processor_id())
159				apf_task_wake_one(n);
160		}
161		spin_unlock(&b->lock);
162	}
163}
164
165void kvm_async_pf_task_wake(u32 token)
166{
167	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
168	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
169	struct kvm_task_sleep_node *n;
170
171	if (token == ~0) {
172		apf_task_wake_all();
173		return;
174	}
175
176again:
177	spin_lock(&b->lock);
178	n = _find_apf_task(b, token);
179	if (!n) {
180		/*
181		 * async PF was not yet handled.
182		 * Add dummy entry for the token.
183		 */
184		n = kmalloc(sizeof(*n), GFP_ATOMIC);
185		if (!n) {
186			/*
187			 * Allocation failed! Busy wait while other cpu
188			 * handles async PF.
189			 */
190			spin_unlock(&b->lock);
191			cpu_relax();
192			goto again;
193		}
194		n->token = token;
195		n->cpu = smp_processor_id();
196		init_waitqueue_head(&n->wq);
197		hlist_add_head(&n->link, &b->list);
198	} else
199		apf_task_wake_one(n);
200	spin_unlock(&b->lock);
201	return;
202}
203EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
204
205u32 kvm_read_and_reset_pf_reason(void)
206{
207	u32 reason = 0;
208
209	if (__get_cpu_var(apf_reason).enabled) {
210		reason = __get_cpu_var(apf_reason).reason;
211		__get_cpu_var(apf_reason).reason = 0;
212	}
213
214	return reason;
215}
216EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
217
218dotraplinkage void __kprobes
219do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
220{
221	switch (kvm_read_and_reset_pf_reason()) {
222	default:
223		do_page_fault(regs, error_code);
224		break;
225	case KVM_PV_REASON_PAGE_NOT_PRESENT:
226		/* page is swapped out by the host. */
227		kvm_async_pf_task_wait((u32)read_cr2());
228		break;
229	case KVM_PV_REASON_PAGE_READY:
230		kvm_async_pf_task_wake((u32)read_cr2());
231		break;
232	}
233}
234
235static void kvm_mmu_op(void *buffer, unsigned len)
236{
237	int r;
238	unsigned long a1, a2;
239
240	do {
241		a1 = __pa(buffer);
242		a2 = 0;   /* on i386 __pa() always returns <4G */
243		r = kvm_hypercall3(KVM_HC_MMU_OP, len, a1, a2);
244		buffer += r;
245		len -= r;
246	} while (len);
247}
248
249static void mmu_queue_flush(struct kvm_para_state *state)
250{
251	if (state->mmu_queue_len) {
252		kvm_mmu_op(state->mmu_queue, state->mmu_queue_len);
253		state->mmu_queue_len = 0;
254	}
255}
256
257static void kvm_deferred_mmu_op(void *buffer, int len)
258{
259	struct kvm_para_state *state = kvm_para_state();
260
261	if (paravirt_get_lazy_mode() != PARAVIRT_LAZY_MMU) {
262		kvm_mmu_op(buffer, len);
263		return;
264	}
265	if (state->mmu_queue_len + len > sizeof state->mmu_queue)
266		mmu_queue_flush(state);
267	memcpy(state->mmu_queue + state->mmu_queue_len, buffer, len);
268	state->mmu_queue_len += len;
269}
270
271static void kvm_mmu_write(void *dest, u64 val)
272{
273	__u64 pte_phys;
274	struct kvm_mmu_op_write_pte wpte;
275
276#ifdef CONFIG_HIGHPTE
277	struct page *page;
278	unsigned long dst = (unsigned long) dest;
279
280	page = kmap_atomic_to_page(dest);
281	pte_phys = page_to_pfn(page);
282	pte_phys <<= PAGE_SHIFT;
283	pte_phys += (dst & ~(PAGE_MASK));
284#else
285	pte_phys = (unsigned long)__pa(dest);
286#endif
287	wpte.header.op = KVM_MMU_OP_WRITE_PTE;
288	wpte.pte_val = val;
289	wpte.pte_phys = pte_phys;
290
291	kvm_deferred_mmu_op(&wpte, sizeof wpte);
292}
293
294/*
295 * We only need to hook operations that are MMU writes.  We hook these so that
296 * we can use lazy MMU mode to batch these operations.  We could probably
297 * improve the performance of the host code if we used some of the information
298 * here to simplify processing of batched writes.
299 */
300static void kvm_set_pte(pte_t *ptep, pte_t pte)
301{
302	kvm_mmu_write(ptep, pte_val(pte));
303}
304
305static void kvm_set_pte_at(struct mm_struct *mm, unsigned long addr,
306			   pte_t *ptep, pte_t pte)
307{
308	kvm_mmu_write(ptep, pte_val(pte));
309}
310
311static void kvm_set_pmd(pmd_t *pmdp, pmd_t pmd)
312{
313	kvm_mmu_write(pmdp, pmd_val(pmd));
314}
315
316#if PAGETABLE_LEVELS >= 3
317#ifdef CONFIG_X86_PAE
318static void kvm_set_pte_atomic(pte_t *ptep, pte_t pte)
319{
320	kvm_mmu_write(ptep, pte_val(pte));
321}
322
323static void kvm_pte_clear(struct mm_struct *mm,
324			  unsigned long addr, pte_t *ptep)
325{
326	kvm_mmu_write(ptep, 0);
327}
328
329static void kvm_pmd_clear(pmd_t *pmdp)
330{
331	kvm_mmu_write(pmdp, 0);
332}
333#endif
334
335static void kvm_set_pud(pud_t *pudp, pud_t pud)
336{
337	kvm_mmu_write(pudp, pud_val(pud));
338}
339
340#if PAGETABLE_LEVELS == 4
341static void kvm_set_pgd(pgd_t *pgdp, pgd_t pgd)
342{
343	kvm_mmu_write(pgdp, pgd_val(pgd));
344}
345#endif
346#endif /* PAGETABLE_LEVELS >= 3 */
347
348static void kvm_flush_tlb(void)
349{
350	struct kvm_mmu_op_flush_tlb ftlb = {
351		.header.op = KVM_MMU_OP_FLUSH_TLB,
352	};
353
354	kvm_deferred_mmu_op(&ftlb, sizeof ftlb);
355}
356
357static void kvm_release_pt(unsigned long pfn)
358{
359	struct kvm_mmu_op_release_pt rpt = {
360		.header.op = KVM_MMU_OP_RELEASE_PT,
361		.pt_phys = (u64)pfn << PAGE_SHIFT,
362	};
363
364	kvm_mmu_op(&rpt, sizeof rpt);
365}
366
367static void kvm_enter_lazy_mmu(void)
368{
369	paravirt_enter_lazy_mmu();
370}
371
372static void kvm_leave_lazy_mmu(void)
373{
374	struct kvm_para_state *state = kvm_para_state();
375
376	mmu_queue_flush(state);
377	paravirt_leave_lazy_mmu();
378}
379
380static void __init paravirt_ops_setup(void)
381{
382	pv_info.name = "KVM";
383	pv_info.paravirt_enabled = 1;
384
385	if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
386		pv_cpu_ops.io_delay = kvm_io_delay;
387
388	if (kvm_para_has_feature(KVM_FEATURE_MMU_OP)) {
389		pv_mmu_ops.set_pte = kvm_set_pte;
390		pv_mmu_ops.set_pte_at = kvm_set_pte_at;
391		pv_mmu_ops.set_pmd = kvm_set_pmd;
392#if PAGETABLE_LEVELS >= 3
393#ifdef CONFIG_X86_PAE
394		pv_mmu_ops.set_pte_atomic = kvm_set_pte_atomic;
395		pv_mmu_ops.pte_clear = kvm_pte_clear;
396		pv_mmu_ops.pmd_clear = kvm_pmd_clear;
397#endif
398		pv_mmu_ops.set_pud = kvm_set_pud;
399#if PAGETABLE_LEVELS == 4
400		pv_mmu_ops.set_pgd = kvm_set_pgd;
401#endif
402#endif
403		pv_mmu_ops.flush_tlb_user = kvm_flush_tlb;
404		pv_mmu_ops.release_pte = kvm_release_pt;
405		pv_mmu_ops.release_pmd = kvm_release_pt;
406		pv_mmu_ops.release_pud = kvm_release_pt;
407
408		pv_mmu_ops.lazy_mode.enter = kvm_enter_lazy_mmu;
409		pv_mmu_ops.lazy_mode.leave = kvm_leave_lazy_mmu;
410	}
411#ifdef CONFIG_X86_IO_APIC
412	no_timer_check = 1;
413#endif
414}
415
416void __cpuinit kvm_guest_cpu_init(void)
417{
418	if (!kvm_para_available())
419		return;
420
421	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
422		u64 pa = __pa(&__get_cpu_var(apf_reason));
423
424		wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
425		__get_cpu_var(apf_reason).enabled = 1;
426		printk(KERN_INFO"KVM setup async PF for cpu %d\n",
427		       smp_processor_id());
428	}
429}
430
431static void kvm_pv_disable_apf(void *unused)
432{
433	if (!__get_cpu_var(apf_reason).enabled)
434		return;
435
436	wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
437	__get_cpu_var(apf_reason).enabled = 0;
438
439	printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
440	       smp_processor_id());
441}
442
443static int kvm_pv_reboot_notify(struct notifier_block *nb,
444				unsigned long code, void *unused)
445{
446	if (code == SYS_RESTART)
447		on_each_cpu(kvm_pv_disable_apf, NULL, 1);
448	return NOTIFY_DONE;
449}
450
451static struct notifier_block kvm_pv_reboot_nb = {
452	.notifier_call = kvm_pv_reboot_notify,
453};
454
455#ifdef CONFIG_SMP
456static void __init kvm_smp_prepare_boot_cpu(void)
457{
458	WARN_ON(kvm_register_clock("primary cpu clock"));
459	kvm_guest_cpu_init();
460	native_smp_prepare_boot_cpu();
461}
462
463static void kvm_guest_cpu_online(void *dummy)
464{
465	kvm_guest_cpu_init();
466}
467
468static void kvm_guest_cpu_offline(void *dummy)
469{
470	kvm_pv_disable_apf(NULL);
471	apf_task_wake_all();
472}
473
474static int __cpuinit kvm_cpu_notify(struct notifier_block *self,
475				    unsigned long action, void *hcpu)
476{
477	int cpu = (unsigned long)hcpu;
478	switch (action) {
479	case CPU_ONLINE:
480	case CPU_DOWN_FAILED:
481	case CPU_ONLINE_FROZEN:
482		smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
483		break;
484	case CPU_DOWN_PREPARE:
485	case CPU_DOWN_PREPARE_FROZEN:
486		smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
487		break;
488	default:
489		break;
490	}
491	return NOTIFY_OK;
492}
493
494static struct notifier_block __cpuinitdata kvm_cpu_notifier = {
495        .notifier_call  = kvm_cpu_notify,
496};
497#endif
498
499static void __init kvm_apf_trap_init(void)
500{
501	set_intr_gate(14, &async_page_fault);
502}
503
504void __init kvm_guest_init(void)
505{
506	int i;
507
508	if (!kvm_para_available())
509		return;
510
511	paravirt_ops_setup();
512	register_reboot_notifier(&kvm_pv_reboot_nb);
513	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
514		spin_lock_init(&async_pf_sleepers[i].lock);
515	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
516		x86_init.irqs.trap_init = kvm_apf_trap_init;
517
518#ifdef CONFIG_SMP
519	smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
520	register_cpu_notifier(&kvm_cpu_notifier);
521#else
522	kvm_guest_cpu_init();
523#endif
524}
525