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/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 #include <asm/idle.h>
  42
  43 static int kvmapf = 1;
  44
  45 static int parse_no_kvmapf(char *arg)
  46 {
  47         kvmapf = 0;
  48         return 0;
  49 }
  50
  51 early_param("no-kvmapf", parse_no_kvmapf);
  52
  53 static int steal_acc = 1;
  54 static int parse_no_stealacc(char *arg)
  55 {
  56         steal_acc = 0;
  57         return 0;
  58 }
  59
  60 early_param("no-steal-acc", parse_no_stealacc);
  61
  62 static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  63 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
  64 static int has_steal_clock = 0;
  65
  66 /*
  67  * No need for any "IO delay" on KVM
  68  */
  69 static 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
  76 struct kvm_task_sleep_node {
  77         struct hlist_node link;
  78         wait_queue_head_t wq;
  79         u32 token;
  80         int cpu;
  81         bool halted;
  82         struct mm_struct *mm;
  83 };
  84
  85 static struct kvm_task_sleep_head {
  86         spinlock_t lock;
  87         struct hlist_head list;
  88 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
  89
  90 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
  91                                                   u32 token)
  92 {
  93         struct hlist_node *p;
  94
  95         hlist_for_each(p, &b->list) {
  96                 struct kvm_task_sleep_node *n =
  97                         hlist_entry(p, typeof(*n), link);
  98                 if (n->token == token)
  99                         return n;
 100         }
 101
 102         return NULL;
 103 }
 104
 105 void kvm_async_pf_task_wait(u32 token)
 106 {
 107         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 108         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 109         struct kvm_task_sleep_node n, *e;
 110         DEFINE_WAIT(wait);
 111         int cpu, idle;
 112
 113         cpu = get_cpu();
 114         idle = idle_cpu(cpu);
 115         put_cpu();
 116
 117         spin_lock(&b->lock);
 118         e = _find_apf_task(b, token);
 119         if (e) {
 120                 /* dummy entry exist -> wake up was delivered ahead of PF */
 121                 hlist_del(&e->link);
 122                 kfree(e);
 123                 spin_unlock(&b->lock);
 124                 return;
 125         }
 126
 127         n.token = token;
 128         n.cpu = smp_processor_id();
 129         n.mm = current->active_mm;
 130         n.halted = idle || preempt_count() > 1;
 131         atomic_inc(&n.mm->mm_count);
 132         init_waitqueue_head(&n.wq);
 133         hlist_add_head(&n.link, &b->list);
 134         spin_unlock(&b->lock);
 135
 136         for (;;) {
 137                 if (!n.halted)
 138                         prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 139                 if (hlist_unhashed(&n.link))
 140                         break;
 141
 142                 if (!n.halted) {
 143                         local_irq_enable();
 144                         schedule();
 145                         local_irq_disable();
 146                 } else {
 147                         /*
 148                          * We cannot reschedule. So halt.
 149                          */
 150                         native_safe_halt();
 151                         local_irq_disable();
 152                 }
 153         }
 154         if (!n.halted)
 155                 finish_wait(&n.wq, &wait);
 156
 157         return;
 158 }
 159 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 160
 161 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 162 {
 163         hlist_del_init(&n->link);
 164         if (!n->mm)
 165                 return;
 166         mmdrop(n->mm);
 167         if (n->halted)
 168                 smp_send_reschedule(n->cpu);
 169         else if (waitqueue_active(&n->wq))
 170                 wake_up(&n->wq);
 171 }
 172
 173 static void apf_task_wake_all(void)
 174 {
 175         int i;
 176
 177         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 178                 struct hlist_node *p, *next;
 179                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 180                 spin_lock(&b->lock);
 181                 hlist_for_each_safe(p, next, &b->list) {
 182                         struct kvm_task_sleep_node *n =
 183                                 hlist_entry(p, typeof(*n), link);
 184                         if (n->cpu == smp_processor_id())
 185                                 apf_task_wake_one(n);
 186                 }
 187                 spin_unlock(&b->lock);
 188         }
 189 }
 190
 191 void kvm_async_pf_task_wake(u32 token)
 192 {
 193         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 194         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 195         struct kvm_task_sleep_node *n;
 196
 197         if (token == ~0) {
 198                 apf_task_wake_all();
 199                 return;
 200         }
 201
 202 again:
 203         spin_lock(&b->lock);
 204         n = _find_apf_task(b, token);
 205         if (!n) {
 206                 /*
 207                  * async PF was not yet handled.
 208                  * Add dummy entry for the token.
 209                  */
 210                 n = kmalloc(sizeof(*n), GFP_ATOMIC);
 211                 if (!n) {
 212                         /*
 213                          * Allocation failed! Busy wait while other cpu
 214                          * handles async PF.
 215                          */
 216                         spin_unlock(&b->lock);
 217                         cpu_relax();
 218                         goto again;
 219                 }
 220                 n->token = token;
 221                 n->cpu = smp_processor_id();
 222                 n->mm = NULL;
 223                 init_waitqueue_head(&n->wq);
 224                 hlist_add_head(&n->link, &b->list);
 225         } else
 226                 apf_task_wake_one(n);
 227         spin_unlock(&b->lock);
 228         return;
 229 }
 230 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 231
 232 u32 kvm_read_and_reset_pf_reason(void)
 233 {
 234         u32 reason = 0;
 235
 236         if (__get_cpu_var(apf_reason).enabled) {
 237                 reason = __get_cpu_var(apf_reason).reason;
 238                 __get_cpu_var(apf_reason).reason = 0;
 239         }
 240
 241         return reason;
 242 }
 243 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 244
 245 dotraplinkage void __kprobes
 246 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
 247 {
 248         switch (kvm_read_and_reset_pf_reason()) {
 249         default:
 250                 do_page_fault(regs, error_code);
 251                 break;
 252         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 253                 /* page is swapped out by the host. */
 254                 kvm_async_pf_task_wait((u32)read_cr2());
 255                 break;
 256         case KVM_PV_REASON_PAGE_READY:
 257                 rcu_irq_enter();
 258                 exit_idle();
 259                 kvm_async_pf_task_wake((u32)read_cr2());
 260                 rcu_irq_exit();
 261                 break;
 262         }
 263 }
 264
 265 static void __init paravirt_ops_setup(void)
 266 {
 267         pv_info.name = "KVM";
 268         pv_info.paravirt_enabled = 1;
 269
 270         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
 271                 pv_cpu_ops.io_delay = kvm_io_delay;
 272
 273 #ifdef CONFIG_X86_IO_APIC
 274         no_timer_check = 1;
 275 #endif
 276 }
 277
 278 static void kvm_register_steal_time(void)
 279 {
 280         int cpu = smp_processor_id();
 281         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
 282
 283         if (!has_steal_clock)
 284                 return;
 285
 286         memset(st, 0, sizeof(*st));
 287
 288         wrmsrl(MSR_KVM_STEAL_TIME, (__pa(st) | KVM_MSR_ENABLED));
 289         printk(KERN_INFO "kvm-stealtime: cpu %d, msr %lx\n",
 290                 cpu, __pa(st));
 291 }
 292
 293 void __cpuinit kvm_guest_cpu_init(void)
 294 {
 295         if (!kvm_para_available())
 296                 return;
 297
 298         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
 299                 u64 pa = __pa(&__get_cpu_var(apf_reason));
 300
 301 #ifdef CONFIG_PREEMPT
 302                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
 303 #endif
 304                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
 305                 __get_cpu_var(apf_reason).enabled = 1;
 306                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 307                        smp_processor_id());
 308         }
 309
 310         if (has_steal_clock)
 311                 kvm_register_steal_time();
 312 }
 313
 314 static void kvm_pv_disable_apf(void *unused)
 315 {
 316         if (!__get_cpu_var(apf_reason).enabled)
 317                 return;
 318
 319         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 320         __get_cpu_var(apf_reason).enabled = 0;
 321
 322         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 323                smp_processor_id());
 324 }
 325
 326 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 327                                 unsigned long code, void *unused)
 328 {
 329         if (code == SYS_RESTART)
 330                 on_each_cpu(kvm_pv_disable_apf, NULL, 1);
 331         return NOTIFY_DONE;
 332 }
 333
 334 static struct notifier_block kvm_pv_reboot_nb = {
 335         .notifier_call = kvm_pv_reboot_notify,
 336 };
 337
 338 static u64 kvm_steal_clock(int cpu)
 339 {
 340         u64 steal;
 341         struct kvm_steal_time *src;
 342         int version;
 343
 344         src = &per_cpu(steal_time, cpu);
 345         do {
 346                 version = src->version;
 347                 rmb();
 348                 steal = src->steal;
 349                 rmb();
 350         } while ((version & 1) || (version != src->version));
 351
 352         return steal;
 353 }
 354
 355 void kvm_disable_steal_time(void)
 356 {
 357         if (!has_steal_clock)
 358                 return;
 359
 360         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 361 }
 362
 363 #ifdef CONFIG_SMP
 364 static void __init kvm_smp_prepare_boot_cpu(void)
 365 {
 366 #ifdef CONFIG_KVM_CLOCK
 367         WARN_ON(kvm_register_clock("primary cpu clock"));
 368 #endif
 369         kvm_guest_cpu_init();
 370         native_smp_prepare_boot_cpu();
 371 }
 372
 373 static void __cpuinit kvm_guest_cpu_online(void *dummy)
 374 {
 375         kvm_guest_cpu_init();
 376 }
 377
 378 static void kvm_guest_cpu_offline(void *dummy)
 379 {
 380         kvm_disable_steal_time();
 381         kvm_pv_disable_apf(NULL);
 382         apf_task_wake_all();
 383 }
 384
 385 static int __cpuinit kvm_cpu_notify(struct notifier_block *self,
 386                                     unsigned long action, void *hcpu)
 387 {
 388         int cpu = (unsigned long)hcpu;
 389         switch (action) {
 390         case CPU_ONLINE:
 391         case CPU_DOWN_FAILED:
 392         case CPU_ONLINE_FROZEN:
 393                 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
 394                 break;
 395         case CPU_DOWN_PREPARE:
 396         case CPU_DOWN_PREPARE_FROZEN:
 397                 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
 398                 break;
 399         default:
 400                 break;
 401         }
 402         return NOTIFY_OK;
 403 }
 404
 405 static struct notifier_block __cpuinitdata kvm_cpu_notifier = {
 406         .notifier_call  = kvm_cpu_notify,
 407 };
 408 #endif
 409
 410 static void __init kvm_apf_trap_init(void)
 411 {
 412         set_intr_gate(14, &async_page_fault);
 413 }
 414
 415 void __init kvm_guest_init(void)
 416 {
 417         int i;
 418
 419         if (!kvm_para_available())
 420                 return;
 421
 422         paravirt_ops_setup();
 423         register_reboot_notifier(&kvm_pv_reboot_nb);
 424         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 425                 spin_lock_init(&async_pf_sleepers[i].lock);
 426         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 427                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 428
 429         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 430                 has_steal_clock = 1;
 431                 pv_time_ops.steal_clock = kvm_steal_clock;
 432         }
 433
 434 #ifdef CONFIG_SMP
 435         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 436         register_cpu_notifier(&kvm_cpu_notifier);
 437 #else
 438         kvm_guest_cpu_init();
 439 #endif
 440 }
 441
 442 static __init int activate_jump_labels(void)
 443 {
 444         if (has_steal_clock) {
 445                 static_key_slow_inc(&paravirt_steal_enabled);
 446                 if (steal_acc)
 447                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 448         }
 449
 450         return 0;
 451 }
 452 arch_initcall(activate_jump_labels);