Commit | Line | Data |
---|---|---|
0cf1bfd2 MT |
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 | ||
56dd9470 | 23 | #include <linux/context_tracking.h> |
0cf1bfd2 MT |
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> | |
1da8a77b | 29 | #include <linux/highmem.h> |
096d14a3 | 30 | #include <linux/hardirq.h> |
fd10cde9 GN |
31 | #include <linux/notifier.h> |
32 | #include <linux/reboot.h> | |
631bc487 GN |
33 | #include <linux/hash.h> |
34 | #include <linux/sched.h> | |
35 | #include <linux/slab.h> | |
36 | #include <linux/kprobes.h> | |
92b75202 | 37 | #include <linux/debugfs.h> |
9919e39a | 38 | #include <linux/nmi.h> |
9db284f3 | 39 | #include <linux/swait.h> |
a90ede7b | 40 | #include <asm/timer.h> |
fd10cde9 | 41 | #include <asm/cpu.h> |
631bc487 GN |
42 | #include <asm/traps.h> |
43 | #include <asm/desc.h> | |
6c047cd9 | 44 | #include <asm/tlbflush.h> |
e0875921 | 45 | #include <asm/idle.h> |
ab9cf499 MT |
46 | #include <asm/apic.h> |
47 | #include <asm/apicdef.h> | |
fc73373b | 48 | #include <asm/hypervisor.h> |
3dc4f7cf | 49 | #include <asm/kvm_guest.h> |
096d14a3 | 50 | |
fd10cde9 GN |
51 | static int kvmapf = 1; |
52 | ||
53 | static int parse_no_kvmapf(char *arg) | |
54 | { | |
55 | kvmapf = 0; | |
56 | return 0; | |
57 | } | |
58 | ||
59 | early_param("no-kvmapf", parse_no_kvmapf); | |
60 | ||
d910f5c1 GC |
61 | static int steal_acc = 1; |
62 | static int parse_no_stealacc(char *arg) | |
63 | { | |
64 | steal_acc = 0; | |
65 | return 0; | |
66 | } | |
67 | ||
68 | early_param("no-steal-acc", parse_no_stealacc); | |
69 | ||
3dc4f7cf MT |
70 | static int kvmclock_vsyscall = 1; |
71 | static int parse_no_kvmclock_vsyscall(char *arg) | |
72 | { | |
73 | kvmclock_vsyscall = 0; | |
74 | return 0; | |
75 | } | |
76 | ||
77 | early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall); | |
78 | ||
fd10cde9 | 79 | static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64); |
d910f5c1 GC |
80 | static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64); |
81 | static int has_steal_clock = 0; | |
096d14a3 | 82 | |
0cf1bfd2 MT |
83 | /* |
84 | * No need for any "IO delay" on KVM | |
85 | */ | |
86 | static void kvm_io_delay(void) | |
87 | { | |
88 | } | |
89 | ||
631bc487 GN |
90 | #define KVM_TASK_SLEEP_HASHBITS 8 |
91 | #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS) | |
92 | ||
93 | struct kvm_task_sleep_node { | |
94 | struct hlist_node link; | |
9db284f3 | 95 | struct swait_queue_head wq; |
631bc487 GN |
96 | u32 token; |
97 | int cpu; | |
6c047cd9 | 98 | bool halted; |
631bc487 GN |
99 | }; |
100 | ||
101 | static struct kvm_task_sleep_head { | |
9db284f3 | 102 | raw_spinlock_t lock; |
631bc487 GN |
103 | struct hlist_head list; |
104 | } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE]; | |
105 | ||
106 | static 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 | ||
121 | void 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; | |
9db284f3 | 126 | DECLARE_SWAITQUEUE(wait); |
631bc487 | 127 | |
9b132fbe LZ |
128 | rcu_irq_enter(); |
129 | ||
9db284f3 | 130 | raw_spin_lock(&b->lock); |
631bc487 GN |
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); | |
9db284f3 | 136 | raw_spin_unlock(&b->lock); |
9b132fbe LZ |
137 | |
138 | rcu_irq_exit(); | |
631bc487 GN |
139 | return; |
140 | } | |
141 | ||
142 | n.token = token; | |
143 | n.cpu = smp_processor_id(); | |
859f8450 | 144 | n.halted = is_idle_task(current) || preempt_count() > 1; |
9db284f3 | 145 | init_swait_queue_head(&n.wq); |
631bc487 | 146 | hlist_add_head(&n.link, &b->list); |
9db284f3 | 147 | raw_spin_unlock(&b->lock); |
631bc487 GN |
148 | |
149 | for (;;) { | |
6c047cd9 | 150 | if (!n.halted) |
9db284f3 | 151 | prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE); |
631bc487 GN |
152 | if (hlist_unhashed(&n.link)) |
153 | break; | |
6c047cd9 GN |
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 | */ | |
9b132fbe | 163 | rcu_irq_exit(); |
6c047cd9 | 164 | native_safe_halt(); |
9b132fbe | 165 | rcu_irq_enter(); |
6c047cd9 GN |
166 | local_irq_disable(); |
167 | } | |
631bc487 | 168 | } |
6c047cd9 | 169 | if (!n.halted) |
9db284f3 | 170 | finish_swait(&n.wq, &wait); |
631bc487 | 171 | |
9b132fbe | 172 | rcu_irq_exit(); |
631bc487 GN |
173 | return; |
174 | } | |
175 | EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait); | |
176 | ||
177 | static void apf_task_wake_one(struct kvm_task_sleep_node *n) | |
178 | { | |
179 | hlist_del_init(&n->link); | |
6c047cd9 GN |
180 | if (n->halted) |
181 | smp_send_reschedule(n->cpu); | |
9db284f3 RR |
182 | else if (swait_active(&n->wq)) |
183 | swake_up(&n->wq); | |
631bc487 GN |
184 | } |
185 | ||
186 | static 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]; | |
9db284f3 | 193 | raw_spin_lock(&b->lock); |
631bc487 GN |
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 | } | |
9db284f3 | 200 | raw_spin_unlock(&b->lock); |
631bc487 GN |
201 | } |
202 | } | |
203 | ||
204 | void 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 | ||
215 | again: | |
9db284f3 | 216 | raw_spin_lock(&b->lock); |
631bc487 GN |
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 | */ | |
62c49cc9 | 223 | n = kzalloc(sizeof(*n), GFP_ATOMIC); |
631bc487 GN |
224 | if (!n) { |
225 | /* | |
226 | * Allocation failed! Busy wait while other cpu | |
227 | * handles async PF. | |
228 | */ | |
9db284f3 | 229 | raw_spin_unlock(&b->lock); |
631bc487 GN |
230 | cpu_relax(); |
231 | goto again; | |
232 | } | |
233 | n->token = token; | |
234 | n->cpu = smp_processor_id(); | |
9db284f3 | 235 | init_swait_queue_head(&n->wq); |
631bc487 GN |
236 | hlist_add_head(&n->link, &b->list); |
237 | } else | |
238 | apf_task_wake_one(n); | |
9db284f3 | 239 | raw_spin_unlock(&b->lock); |
631bc487 GN |
240 | return; |
241 | } | |
242 | EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake); | |
243 | ||
244 | u32 kvm_read_and_reset_pf_reason(void) | |
245 | { | |
246 | u32 reason = 0; | |
247 | ||
89cbc767 CL |
248 | if (__this_cpu_read(apf_reason.enabled)) { |
249 | reason = __this_cpu_read(apf_reason.reason); | |
250 | __this_cpu_write(apf_reason.reason, 0); | |
631bc487 GN |
251 | } |
252 | ||
253 | return reason; | |
254 | } | |
255 | EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason); | |
9326638c | 256 | NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason); |
631bc487 | 257 | |
9326638c | 258 | dotraplinkage void |
631bc487 GN |
259 | do_async_page_fault(struct pt_regs *regs, unsigned long error_code) |
260 | { | |
6c1e0256 FW |
261 | enum ctx_state prev_state; |
262 | ||
631bc487 GN |
263 | switch (kvm_read_and_reset_pf_reason()) { |
264 | default: | |
65a7f03f | 265 | trace_do_page_fault(regs, error_code); |
631bc487 GN |
266 | break; |
267 | case KVM_PV_REASON_PAGE_NOT_PRESENT: | |
268 | /* page is swapped out by the host. */ | |
6c1e0256 | 269 | prev_state = exception_enter(); |
c5e015d4 | 270 | exit_idle(); |
631bc487 | 271 | kvm_async_pf_task_wait((u32)read_cr2()); |
6c1e0256 | 272 | exception_exit(prev_state); |
631bc487 GN |
273 | break; |
274 | case KVM_PV_REASON_PAGE_READY: | |
e0875921 GN |
275 | rcu_irq_enter(); |
276 | exit_idle(); | |
631bc487 | 277 | kvm_async_pf_task_wake((u32)read_cr2()); |
e0875921 | 278 | rcu_irq_exit(); |
631bc487 GN |
279 | break; |
280 | } | |
281 | } | |
9326638c | 282 | NOKPROBE_SYMBOL(do_async_page_fault); |
631bc487 | 283 | |
d3ac8815 | 284 | static void __init paravirt_ops_setup(void) |
0cf1bfd2 MT |
285 | { |
286 | pv_info.name = "KVM"; | |
29fa6825 AL |
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; | |
0cf1bfd2 MT |
295 | |
296 | if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY)) | |
297 | pv_cpu_ops.io_delay = kvm_io_delay; | |
298 | ||
a90ede7b MT |
299 | #ifdef CONFIG_X86_IO_APIC |
300 | no_timer_check = 1; | |
301 | #endif | |
0cf1bfd2 MT |
302 | } |
303 | ||
d910f5c1 GC |
304 | static 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 | ||
5dfd486c | 314 | wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED)); |
136867f5 SK |
315 | pr_info("kvm-stealtime: cpu %d, msr %llx\n", |
316 | cpu, (unsigned long long) slow_virt_to_phys(st)); | |
d910f5c1 GC |
317 | } |
318 | ||
ab9cf499 MT |
319 | static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED; |
320 | ||
321 | static 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 | */ | |
89cbc767 | 330 | if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi))) |
ab9cf499 | 331 | return; |
90536664 | 332 | apic_write(APIC_EOI, APIC_EOI_ACK); |
ab9cf499 MT |
333 | } |
334 | ||
ed3cf152 | 335 | static void kvm_guest_cpu_init(void) |
fd10cde9 GN |
336 | { |
337 | if (!kvm_para_available()) | |
338 | return; | |
339 | ||
340 | if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) { | |
89cbc767 | 341 | u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason)); |
fd10cde9 | 342 | |
6adba527 GN |
343 | #ifdef CONFIG_PREEMPT |
344 | pa |= KVM_ASYNC_PF_SEND_ALWAYS; | |
345 | #endif | |
fd10cde9 | 346 | wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED); |
89cbc767 | 347 | __this_cpu_write(apf_reason.enabled, 1); |
fd10cde9 GN |
348 | printk(KERN_INFO"KVM setup async PF for cpu %d\n", |
349 | smp_processor_id()); | |
350 | } | |
d910f5c1 | 351 | |
ab9cf499 MT |
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); | |
89cbc767 CL |
356 | __this_cpu_write(kvm_apic_eoi, 0); |
357 | pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi)) | |
5dfd486c | 358 | | KVM_MSR_ENABLED; |
ab9cf499 MT |
359 | wrmsrl(MSR_KVM_PV_EOI_EN, pa); |
360 | } | |
361 | ||
d910f5c1 GC |
362 | if (has_steal_clock) |
363 | kvm_register_steal_time(); | |
fd10cde9 GN |
364 | } |
365 | ||
ab9cf499 | 366 | static void kvm_pv_disable_apf(void) |
fd10cde9 | 367 | { |
89cbc767 | 368 | if (!__this_cpu_read(apf_reason.enabled)) |
fd10cde9 GN |
369 | return; |
370 | ||
371 | wrmsrl(MSR_KVM_ASYNC_PF_EN, 0); | |
89cbc767 | 372 | __this_cpu_write(apf_reason.enabled, 0); |
fd10cde9 GN |
373 | |
374 | printk(KERN_INFO"Unregister pv shared memory for cpu %d\n", | |
375 | smp_processor_id()); | |
376 | } | |
377 | ||
ab9cf499 MT |
378 | static 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(); | |
8fbe6a54 | 388 | kvm_disable_steal_time(); |
ab9cf499 MT |
389 | } |
390 | ||
fd10cde9 GN |
391 | static int kvm_pv_reboot_notify(struct notifier_block *nb, |
392 | unsigned long code, void *unused) | |
393 | { | |
394 | if (code == SYS_RESTART) | |
ab9cf499 | 395 | on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1); |
fd10cde9 GN |
396 | return NOTIFY_DONE; |
397 | } | |
398 | ||
399 | static struct notifier_block kvm_pv_reboot_nb = { | |
400 | .notifier_call = kvm_pv_reboot_notify, | |
401 | }; | |
402 | ||
d910f5c1 GC |
403 | static 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 | ||
420 | void kvm_disable_steal_time(void) | |
421 | { | |
422 | if (!has_steal_clock) | |
423 | return; | |
424 | ||
425 | wrmsr(MSR_KVM_STEAL_TIME, 0, 0); | |
426 | } | |
427 | ||
ca3f1017 GN |
428 | #ifdef CONFIG_SMP |
429 | static void __init kvm_smp_prepare_boot_cpu(void) | |
430 | { | |
fd10cde9 | 431 | kvm_guest_cpu_init(); |
ca3f1017 | 432 | native_smp_prepare_boot_cpu(); |
92b75202 | 433 | kvm_spinlock_init(); |
ca3f1017 | 434 | } |
fd10cde9 | 435 | |
148f9bb8 | 436 | static void kvm_guest_cpu_online(void *dummy) |
fd10cde9 GN |
437 | { |
438 | kvm_guest_cpu_init(); | |
439 | } | |
440 | ||
441 | static void kvm_guest_cpu_offline(void *dummy) | |
442 | { | |
d910f5c1 | 443 | kvm_disable_steal_time(); |
ab9cf499 MT |
444 | if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) |
445 | wrmsrl(MSR_KVM_PV_EOI_EN, 0); | |
446 | kvm_pv_disable_apf(); | |
631bc487 | 447 | apf_task_wake_all(); |
fd10cde9 GN |
448 | } |
449 | ||
148f9bb8 PG |
450 | static int kvm_cpu_notify(struct notifier_block *self, unsigned long action, |
451 | void *hcpu) | |
fd10cde9 GN |
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 | ||
148f9bb8 | 470 | static struct notifier_block kvm_cpu_notifier = { |
fd10cde9 GN |
471 | .notifier_call = kvm_cpu_notify, |
472 | }; | |
ca3f1017 GN |
473 | #endif |
474 | ||
631bc487 GN |
475 | static void __init kvm_apf_trap_init(void) |
476 | { | |
25c74b10 | 477 | set_intr_gate(14, async_page_fault); |
631bc487 GN |
478 | } |
479 | ||
0cf1bfd2 MT |
480 | void __init kvm_guest_init(void) |
481 | { | |
631bc487 GN |
482 | int i; |
483 | ||
0cf1bfd2 MT |
484 | if (!kvm_para_available()) |
485 | return; | |
486 | ||
487 | paravirt_ops_setup(); | |
fd10cde9 | 488 | register_reboot_notifier(&kvm_pv_reboot_nb); |
631bc487 | 489 | for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) |
9db284f3 | 490 | raw_spin_lock_init(&async_pf_sleepers[i].lock); |
631bc487 GN |
491 | if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF)) |
492 | x86_init.irqs.trap_init = kvm_apf_trap_init; | |
493 | ||
d910f5c1 GC |
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 | ||
90536664 MT |
499 | if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) |
500 | apic_set_eoi_write(kvm_guest_apic_eoi_write); | |
ab9cf499 | 501 | |
3dc4f7cf MT |
502 | if (kvmclock_vsyscall) |
503 | kvm_setup_vsyscall_timeinfo(); | |
504 | ||
ca3f1017 GN |
505 | #ifdef CONFIG_SMP |
506 | smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu; | |
fd10cde9 GN |
507 | register_cpu_notifier(&kvm_cpu_notifier); |
508 | #else | |
509 | kvm_guest_cpu_init(); | |
ca3f1017 | 510 | #endif |
9919e39a UO |
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 | */ | |
692297d8 | 517 | hardlockup_detector_disable(); |
0cf1bfd2 | 518 | } |
d910f5c1 | 519 | |
1c300a40 PB |
520 | static 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 | ||
531 | static 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 | ||
541 | bool kvm_para_available(void) | |
542 | { | |
543 | return kvm_cpuid_base() != 0; | |
544 | } | |
545 | EXPORT_SYMBOL_GPL(kvm_para_available); | |
546 | ||
77f01bdf PB |
547 | unsigned int kvm_arch_para_features(void) |
548 | { | |
549 | return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES); | |
550 | } | |
551 | ||
9df56f19 | 552 | static uint32_t __init kvm_detect(void) |
fc73373b | 553 | { |
9df56f19 | 554 | return kvm_cpuid_base(); |
fc73373b PB |
555 | } |
556 | ||
557 | const struct hypervisor_x86 x86_hyper_kvm __refconst = { | |
558 | .name = "KVM", | |
559 | .detect = kvm_detect, | |
4cca6ea0 | 560 | .x2apic_available = kvm_para_available, |
fc73373b PB |
561 | }; |
562 | EXPORT_SYMBOL_GPL(x86_hyper_kvm); | |
563 | ||
d910f5c1 GC |
564 | static __init int activate_jump_labels(void) |
565 | { | |
566 | if (has_steal_clock) { | |
c5905afb | 567 | static_key_slow_inc(¶virt_steal_enabled); |
d910f5c1 | 568 | if (steal_acc) |
c5905afb | 569 | static_key_slow_inc(¶virt_steal_rq_enabled); |
d910f5c1 GC |
570 | } |
571 | ||
572 | return 0; | |
573 | } | |
574 | arch_initcall(activate_jump_labels); | |
92b75202 SV |
575 | |
576 | #ifdef CONFIG_PARAVIRT_SPINLOCKS | |
577 | ||
578 | /* Kick a cpu by its apicid. Used to wake up a halted vcpu */ | |
36bd6213 | 579 | static void kvm_kick_cpu(int cpu) |
92b75202 SV |
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 | ||
bf0c7c34 | 588 | |
62c7a1e9 | 589 | #ifdef CONFIG_QUEUED_SPINLOCKS |
bf0c7c34 WL |
590 | |
591 | #include <asm/qspinlock.h> | |
592 | ||
593 | static 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 | ||
615 | out: | |
616 | local_irq_restore(flags); | |
617 | } | |
618 | ||
62c7a1e9 | 619 | #else /* !CONFIG_QUEUED_SPINLOCKS */ |
bf0c7c34 | 620 | |
92b75202 SV |
621 | enum 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 | ||
632 | static 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 | ||
639 | static u8 zero_stats; | |
640 | ||
641 | static inline void check_zero(void) | |
642 | { | |
643 | u8 ret; | |
644 | u8 old; | |
645 | ||
d6abfdb2 | 646 | old = READ_ONCE(zero_stats); |
92b75202 SV |
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 | ||
655 | static 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 | ||
662 | static inline u64 spin_time_start(void) | |
663 | { | |
664 | return sched_clock(); | |
665 | } | |
666 | ||
667 | static 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 | ||
680 | static 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 | ||
689 | static struct dentry *d_spin_debug; | |
690 | static struct dentry *d_kvm_debug; | |
691 | ||
ed3cf152 | 692 | static struct dentry *kvm_init_debugfs(void) |
92b75202 | 693 | { |
d780a312 | 694 | d_kvm_debug = debugfs_create_dir("kvm-guest", NULL); |
92b75202 SV |
695 | if (!d_kvm_debug) |
696 | printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n"); | |
697 | ||
698 | return d_kvm_debug; | |
699 | } | |
700 | ||
701 | static 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 | } | |
731 | fs_initcall(kvm_spinlock_debugfs); | |
732 | #else /* !CONFIG_KVM_DEBUG_FS */ | |
733 | static inline void add_stats(enum kvm_contention_stat var, u32 val) | |
734 | { | |
735 | } | |
736 | ||
737 | static inline u64 spin_time_start(void) | |
738 | { | |
739 | return 0; | |
740 | } | |
741 | ||
742 | static inline void spin_time_accum_blocked(u64 start) | |
743 | { | |
744 | } | |
745 | #endif /* CONFIG_KVM_DEBUG_FS */ | |
746 | ||
747 | struct kvm_lock_waiting { | |
748 | struct arch_spinlock *lock; | |
749 | __ticket_t want; | |
750 | }; | |
751 | ||
752 | /* cpus 'waiting' on a spinlock to become available */ | |
753 | static cpumask_t waiting_cpus; | |
754 | ||
755 | /* Track spinlock on which a cpu is waiting */ | |
756 | static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting); | |
757 | ||
dd41f818 | 758 | __visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want) |
92b75202 SV |
759 | { |
760 | struct kvm_lock_waiting *w; | |
761 | int cpu; | |
762 | u64 start; | |
763 | unsigned long flags; | |
d6abfdb2 | 764 | __ticket_t head; |
92b75202 SV |
765 | |
766 | if (in_nmi()) | |
767 | return; | |
768 | ||
89cbc767 | 769 | w = this_cpu_ptr(&klock_waiting); |
92b75202 SV |
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 | ||
d6abfdb2 R |
806 | /* make sure enter_slowpath, which is atomic does not cross the read */ |
807 | smp_mb__after_atomic(); | |
808 | ||
92b75202 SV |
809 | /* |
810 | * check again make sure it didn't become free while | |
811 | * we weren't looking. | |
812 | */ | |
d6abfdb2 R |
813 | head = READ_ONCE(lock->tickets.head); |
814 | if (__tickets_equal(head, want)) { | |
92b75202 SV |
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 | ||
829 | out: | |
830 | cpumask_clear_cpu(cpu, &waiting_cpus); | |
831 | w->lock = NULL; | |
832 | local_irq_restore(flags); | |
833 | spin_time_accum_blocked(start); | |
834 | } | |
835 | PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning); | |
836 | ||
837 | /* Kick vcpu waiting on @lock->head to reach value @ticket */ | |
838 | static 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); | |
d6abfdb2 R |
845 | if (READ_ONCE(w->lock) == lock && |
846 | READ_ONCE(w->want) == ticket) { | |
92b75202 SV |
847 | add_stats(RELEASED_SLOW_KICKED, 1); |
848 | kvm_kick_cpu(cpu); | |
849 | break; | |
850 | } | |
851 | } | |
852 | } | |
853 | ||
62c7a1e9 | 854 | #endif /* !CONFIG_QUEUED_SPINLOCKS */ |
bf0c7c34 | 855 | |
92b75202 SV |
856 | /* |
857 | * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present. | |
858 | */ | |
859 | void __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 | ||
62c7a1e9 | 867 | #ifdef CONFIG_QUEUED_SPINLOCKS |
bf0c7c34 WL |
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; | |
62c7a1e9 | 873 | #else /* !CONFIG_QUEUED_SPINLOCKS */ |
3dbef3e3 R |
874 | pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning); |
875 | pv_lock_ops.unlock_kick = kvm_unlock_kick; | |
bf0c7c34 | 876 | #endif |
3dbef3e3 R |
877 | } |
878 | ||
879 | static __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; | |
92b75202 SV |
885 | |
886 | static_key_slow_inc(¶virt_ticketlocks_enabled); | |
3dbef3e3 | 887 | printk(KERN_INFO "KVM setup paravirtual spinlock\n"); |
92b75202 | 888 | |
3dbef3e3 | 889 | return 0; |
92b75202 | 890 | } |
3dbef3e3 R |
891 | early_initcall(kvm_spinlock_init_jump); |
892 | ||
92b75202 | 893 | #endif /* CONFIG_PARAVIRT_SPINLOCKS */ |