arch/x86/kernel/kvm.c

   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 <asm/timer.h>
  38 #include <asm/cpu.h>
  39 #include <asm/traps.h>
  40 #include <asm/desc.h>
  41 #include <asm/tlbflush.h>
  42 #include <asm/idle.h>
  43 #include <asm/apic.h>
  44 #include <asm/apicdef.h>
  45 #include <asm/hypervisor.h>
  46 #include <asm/kvm_guest.h>
  47
  48 static int kvmapf = 1;
  49
  50 static int parse_no_kvmapf(char *arg)
  51 {
  52         kvmapf = 0;
  53         return 0;
  54 }
  55
  56 early_param("no-kvmapf", parse_no_kvmapf);
  57
  58 static int steal_acc = 1;
  59 static int parse_no_stealacc(char *arg)
  60 {
  61         steal_acc = 0;
  62         return 0;
  63 }
  64
  65 early_param("no-steal-acc", parse_no_stealacc);
  66
  67 static int kvmclock_vsyscall = 1;
  68 static int parse_no_kvmclock_vsyscall(char *arg)
  69 {
  70         kvmclock_vsyscall = 0;
  71         return 0;
  72 }
  73
  74 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
  75
  76 static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  77 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
  78 static int has_steal_clock = 0;
  79
  80 /*
  81  * No need for any "IO delay" on KVM
  82  */
  83 static void kvm_io_delay(void)
  84 {
  85 }
  86
  87 #define KVM_TASK_SLEEP_HASHBITS 8
  88 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  89
  90 struct kvm_task_sleep_node {
  91         struct hlist_node link;
  92         wait_queue_head_t wq;
  93         u32 token;
  94         int cpu;
  95         bool halted;
  96 };
  97
  98 static struct kvm_task_sleep_head {
  99         spinlock_t lock;
 100         struct hlist_head list;
 101 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
 102
 103 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
 104                                                   u32 token)
 105 {
 106         struct hlist_node *p;
 107
 108         hlist_for_each(p, &b->list) {
 109                 struct kvm_task_sleep_node *n =
 110                         hlist_entry(p, typeof(*n), link);
 111                 if (n->token == token)
 112                         return n;
 113         }
 114
 115         return NULL;
 116 }
 117
 118 void kvm_async_pf_task_wait(u32 token)
 119 {
 120         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 121         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 122         struct kvm_task_sleep_node n, *e;
 123         DEFINE_WAIT(wait);
 124
 125         rcu_irq_enter();
 126
 127         spin_lock(&b->lock);
 128         e = _find_apf_task(b, token);
 129         if (e) {
 130                 /* dummy entry exist -> wake up was delivered ahead of PF */
 131                 hlist_del(&e->link);
 132                 kfree(e);
 133                 spin_unlock(&b->lock);
 134
 135                 rcu_irq_exit();
 136                 return;
 137         }
 138
 139         n.token = token;
 140         n.cpu = smp_processor_id();
 141         n.halted = is_idle_task(current) || preempt_count() > 1;
 142         init_waitqueue_head(&n.wq);
 143         hlist_add_head(&n.link, &b->list);
 144         spin_unlock(&b->lock);
 145
 146         for (;;) {
 147                 if (!n.halted)
 148                         prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 149                 if (hlist_unhashed(&n.link))
 150                         break;
 151
 152                 if (!n.halted) {
 153                         local_irq_enable();
 154                         schedule();
 155                         local_irq_disable();
 156                 } else {
 157                         /*
 158                          * We cannot reschedule. So halt.
 159                          */
 160                         rcu_irq_exit();
 161                         native_safe_halt();
 162                         rcu_irq_enter();
 163                         local_irq_disable();
 164                 }
 165         }
 166         if (!n.halted)
 167                 finish_wait(&n.wq, &wait);
 168
 169         rcu_irq_exit();
 170         return;
 171 }
 172 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 173
 174 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 175 {
 176         hlist_del_init(&n->link);
 177         if (n->halted)
 178                 smp_send_reschedule(n->cpu);
 179         else if (waitqueue_active(&n->wq))
 180                 wake_up(&n->wq);
 181 }
 182
 183 static void apf_task_wake_all(void)
 184 {
 185         int i;
 186
 187         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 188                 struct hlist_node *p, *next;
 189                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 190                 spin_lock(&b->lock);
 191                 hlist_for_each_safe(p, next, &b->list) {
 192                         struct kvm_task_sleep_node *n =
 193                                 hlist_entry(p, typeof(*n), link);
 194                         if (n->cpu == smp_processor_id())
 195                                 apf_task_wake_one(n);
 196                 }
 197                 spin_unlock(&b->lock);
 198         }
 199 }
 200
 201 void kvm_async_pf_task_wake(u32 token)
 202 {
 203         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 204         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 205         struct kvm_task_sleep_node *n;
 206
 207         if (token == ~0) {
 208                 apf_task_wake_all();
 209                 return;
 210         }
 211
 212 again:
 213         spin_lock(&b->lock);
 214         n = _find_apf_task(b, token);
 215         if (!n) {
 216                 /*
 217                  * async PF was not yet handled.
 218                  * Add dummy entry for the token.
 219                  */
 220                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 221                 if (!n) {
 222                         /*
 223                          * Allocation failed! Busy wait while other cpu
 224                          * handles async PF.
 225                          */
 226                         spin_unlock(&b->lock);
 227                         cpu_relax();
 228                         goto again;
 229                 }
 230                 n->token = token;
 231                 n->cpu = smp_processor_id();
 232                 init_waitqueue_head(&n->wq);
 233                 hlist_add_head(&n->link, &b->list);
 234         } else
 235                 apf_task_wake_one(n);
 236         spin_unlock(&b->lock);
 237         return;
 238 }
 239 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 240
 241 u32 kvm_read_and_reset_pf_reason(void)
 242 {
 243         u32 reason = 0;
 244
 245         if (__get_cpu_var(apf_reason).enabled) {
 246                 reason = __get_cpu_var(apf_reason).reason;
 247                 __get_cpu_var(apf_reason).reason = 0;
 248         }
 249
 250         return reason;
 251 }
 252 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 253
 254 dotraplinkage void __kprobes
 255 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
 256 {
 257         enum ctx_state prev_state;
 258
 259         switch (kvm_read_and_reset_pf_reason()) {
 260         default:
 261                 do_page_fault(regs, error_code);
 262                 break;
 263         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 264                 /* page is swapped out by the host. */
 265                 prev_state = exception_enter();
 266                 exit_idle();
 267                 kvm_async_pf_task_wait((u32)read_cr2());
 268                 exception_exit(prev_state);
 269                 break;
 270         case KVM_PV_REASON_PAGE_READY:
 271                 rcu_irq_enter();
 272                 exit_idle();
 273                 kvm_async_pf_task_wake((u32)read_cr2());
 274                 rcu_irq_exit();
 275                 break;
 276         }
 277 }
 278
 279 static void __init paravirt_ops_setup(void)
 280 {
 281         pv_info.name = "KVM";
 282
 283         /*
 284          * KVM isn't paravirt in the sense of paravirt_enabled.  A KVM
 285          * guest kernel works like a bare metal kernel with additional
 286          * features, and paravirt_enabled is about features that are
 287          * missing.
 288          */
 289         pv_info.paravirt_enabled = 0;
 290
 291         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
 292                 pv_cpu_ops.io_delay = kvm_io_delay;
 293
 294 #ifdef CONFIG_X86_IO_APIC
 295         no_timer_check = 1;
 296 #endif
 297 }
 298
 299 static void kvm_register_steal_time(void)
 300 {
 301         int cpu = smp_processor_id();
 302         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
 303
 304         if (!has_steal_clock)
 305                 return;
 306
 307         memset(st, 0, sizeof(*st));
 308
 309         wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
 310         pr_info("kvm-stealtime: cpu %d, msr %llx\n",
 311                 cpu, (unsigned long long) slow_virt_to_phys(st));
 312 }
 313
 314 static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
 315
 316 static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
 317 {
 318         /**
 319          * This relies on __test_and_clear_bit to modify the memory
 320          * in a way that is atomic with respect to the local CPU.
 321          * The hypervisor only accesses this memory from the local CPU so
 322          * there's no need for lock or memory barriers.
 323          * An optimization barrier is implied in apic write.
 324          */
 325         if (__test_and_clear_bit(KVM_PV_EOI_BIT, &__get_cpu_var(kvm_apic_eoi)))
 326                 return;
 327         apic_write(APIC_EOI, APIC_EOI_ACK);
 328 }
 329
 330 void __cpuinit kvm_guest_cpu_init(void)
 331 {
 332         if (!kvm_para_available())
 333                 return;
 334
 335         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
 336                 u64 pa = slow_virt_to_phys(&__get_cpu_var(apf_reason));
 337
 338 #ifdef CONFIG_PREEMPT
 339                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
 340 #endif
 341                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
 342                 __get_cpu_var(apf_reason).enabled = 1;
 343                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 344                        smp_processor_id());
 345         }
 346
 347         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 348                 unsigned long pa;
 349                 /* Size alignment is implied but just to make it explicit. */
 350                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 351                 __get_cpu_var(kvm_apic_eoi) = 0;
 352                 pa = slow_virt_to_phys(&__get_cpu_var(kvm_apic_eoi))
 353                         | KVM_MSR_ENABLED;
 354                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 355         }
 356
 357         if (has_steal_clock)
 358                 kvm_register_steal_time();
 359 }
 360
 361 static void kvm_pv_disable_apf(void)
 362 {
 363         if (!__get_cpu_var(apf_reason).enabled)
 364                 return;
 365
 366         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 367         __get_cpu_var(apf_reason).enabled = 0;
 368
 369         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 370                smp_processor_id());
 371 }
 372
 373 static void kvm_pv_guest_cpu_reboot(void *unused)
 374 {
 375         /*
 376          * We disable PV EOI before we load a new kernel by kexec,
 377          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
 378          * New kernel can re-enable when it boots.
 379          */
 380         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 381                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 382         kvm_pv_disable_apf();
 383         kvm_disable_steal_time();
 384 }
 385
 386 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 387                                 unsigned long code, void *unused)
 388 {
 389         if (code == SYS_RESTART)
 390                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 391         return NOTIFY_DONE;
 392 }
 393
 394 static struct notifier_block kvm_pv_reboot_nb = {
 395         .notifier_call = kvm_pv_reboot_notify,
 396 };
 397
 398 static u64 kvm_steal_clock(int cpu)
 399 {
 400         u64 steal;
 401         struct kvm_steal_time *src;
 402         int version;
 403
 404         src = &per_cpu(steal_time, cpu);
 405         do {
 406                 version = src->version;
 407                 rmb();
 408                 steal = src->steal;
 409                 rmb();
 410         } while ((version & 1) || (version != src->version));
 411
 412         return steal;
 413 }
 414
 415 void kvm_disable_steal_time(void)
 416 {
 417         if (!has_steal_clock)
 418                 return;
 419
 420         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 421 }
 422
 423 #ifdef CONFIG_SMP
 424 static void __init kvm_smp_prepare_boot_cpu(void)
 425 {
 426         WARN_ON(kvm_register_clock("primary cpu clock"));
 427         kvm_guest_cpu_init();
 428         native_smp_prepare_boot_cpu();
 429 }
 430
 431 static void __cpuinit kvm_guest_cpu_online(void *dummy)
 432 {
 433         kvm_guest_cpu_init();
 434 }
 435
 436 static void kvm_guest_cpu_offline(void *dummy)
 437 {
 438         kvm_disable_steal_time();
 439         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 440                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 441         kvm_pv_disable_apf();
 442         apf_task_wake_all();
 443 }
 444
 445 static int __cpuinit kvm_cpu_notify(struct notifier_block *self,
 446                                     unsigned long action, void *hcpu)
 447 {
 448         int cpu = (unsigned long)hcpu;
 449         switch (action) {
 450         case CPU_ONLINE:
 451         case CPU_DOWN_FAILED:
 452         case CPU_ONLINE_FROZEN:
 453                 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
 454                 break;
 455         case CPU_DOWN_PREPARE:
 456         case CPU_DOWN_PREPARE_FROZEN:
 457                 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
 458                 break;
 459         default:
 460                 break;
 461         }
 462         return NOTIFY_OK;
 463 }
 464
 465 static struct notifier_block __cpuinitdata kvm_cpu_notifier = {
 466         .notifier_call  = kvm_cpu_notify,
 467 };
 468 #endif
 469
 470 static void __init kvm_apf_trap_init(void)
 471 {
 472         set_intr_gate(14, &async_page_fault);
 473 }
 474
 475 void __init kvm_guest_init(void)
 476 {
 477         int i;
 478
 479         if (!kvm_para_available())
 480                 return;
 481
 482         paravirt_ops_setup();
 483         register_reboot_notifier(&kvm_pv_reboot_nb);
 484         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 485                 spin_lock_init(&async_pf_sleepers[i].lock);
 486         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 487                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 488
 489         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 490                 has_steal_clock = 1;
 491                 pv_time_ops.steal_clock = kvm_steal_clock;
 492         }
 493
 494         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 495                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 496
 497         if (kvmclock_vsyscall)
 498                 kvm_setup_vsyscall_timeinfo();
 499
 500 #ifdef CONFIG_SMP
 501         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 502         register_cpu_notifier(&kvm_cpu_notifier);
 503 #else
 504         kvm_guest_cpu_init();
 505 #endif
 506 }
 507
 508 static bool __init kvm_detect(void)
 509 {
 510         if (!kvm_para_available())
 511                 return false;
 512         return true;
 513 }
 514
 515 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
 516         .name                   = "KVM",
 517         .detect                 = kvm_detect,
 518         .x2apic_available       = kvm_para_available,
 519 };
 520 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
 521
 522 static __init int activate_jump_labels(void)
 523 {
 524         if (has_steal_clock) {
 525                 static_key_slow_inc(&paravirt_steal_enabled);
 526                 if (steal_acc)
 527                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 528         }
 529
 530         return 0;
 531 }
 532 arch_initcall(activate_jump_labels);