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
  42 #define MMU_QUEUE_SIZE 1024
  43
  44 static int kvmapf = 1;
  45
  46 static int parse_no_kvmapf(char *arg)
  47 {
  48         kvmapf = 0;
  49         return 0;
  50 }
  51
  52 early_param("no-kvmapf", parse_no_kvmapf);
  53
  54 static int steal_acc = 1;
  55 static int parse_no_stealacc(char *arg)
  56 {
  57         steal_acc = 0;
  58         return 0;
  59 }
  60
  61 early_param("no-steal-acc", parse_no_stealacc);
  62
  63 struct kvm_para_state {
  64         u8 mmu_queue[MMU_QUEUE_SIZE];
  65         int mmu_queue_len;
  66 };
  67
  68 static DEFINE_PER_CPU(struct kvm_para_state, para_state);
  69 static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  70 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
  71 static int has_steal_clock = 0;
  72
  73 static struct kvm_para_state *kvm_para_state(void)
  74 {
  75         return &per_cpu(para_state, raw_smp_processor_id());
  76 }
  77
  78 /*
  79  * No need for any "IO delay" on KVM
  80  */
  81 static void kvm_io_delay(void)
  82 {
  83 }
  84
  85 #define KVM_TASK_SLEEP_HASHBITS 8
  86 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  87
  88 struct kvm_task_sleep_node {
  89         struct hlist_node link;
  90         wait_queue_head_t wq;
  91         u32 token;
  92         int cpu;
  93         bool halted;
  94         struct mm_struct *mm;
  95 };
  96
  97 static struct kvm_task_sleep_head {
  98         spinlock_t lock;
  99         struct hlist_head list;
 100 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
 101
 102 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
 103                                                   u32 token)
 104 {
 105         struct hlist_node *p;
 106
 107         hlist_for_each(p, &b->list) {
 108                 struct kvm_task_sleep_node *n =
 109                         hlist_entry(p, typeof(*n), link);
 110                 if (n->token == token)
 111                         return n;
 112         }
 113
 114         return NULL;
 115 }
 116
 117 void kvm_async_pf_task_wait(u32 token)
 118 {
 119         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 120         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 121         struct kvm_task_sleep_node n, *e;
 122         DEFINE_WAIT(wait);
 123         int cpu, idle;
 124
 125         cpu = get_cpu();
 126         idle = idle_cpu(cpu);
 127         put_cpu();
 128
 129         spin_lock(&b->lock);
 130         e = _find_apf_task(b, token);
 131         if (e) {
 132                 /* dummy entry exist -> wake up was delivered ahead of PF */
 133                 hlist_del(&e->link);
 134                 kfree(e);
 135                 spin_unlock(&b->lock);
 136                 return;
 137         }
 138
 139         n.token = token;
 140         n.cpu = smp_processor_id();
 141         n.mm = current->active_mm;
 142         n.halted = idle || preempt_count() > 1;
 143         atomic_inc(&n.mm->mm_count);
 144         init_waitqueue_head(&n.wq);
 145         hlist_add_head(&n.link, &b->list);
 146         spin_unlock(&b->lock);
 147
 148         for (;;) {
 149                 if (!n.halted)
 150                         prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 151                 if (hlist_unhashed(&n.link))
 152                         break;
 153
 154                 if (!n.halted) {
 155                         local_irq_enable();
 156                         schedule();
 157                         local_irq_disable();
 158                 } else {
 159                         /*
 160                          * We cannot reschedule. So halt.
 161                          */
 162                         native_safe_halt();
 163                         local_irq_disable();
 164                 }
 165         }
 166         if (!n.halted)
 167                 finish_wait(&n.wq, &wait);
 168
 169         return;
 170 }
 171 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 172
 173 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 174 {
 175         hlist_del_init(&n->link);
 176         if (!n->mm)
 177                 return;
 178         mmdrop(n->mm);
 179         if (n->halted)
 180                 smp_send_reschedule(n->cpu);
 181         else if (waitqueue_active(&n->wq))
 182                 wake_up(&n->wq);
 183 }
 184
 185 static void apf_task_wake_all(void)
 186 {
 187         int i;
 188
 189         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 190                 struct hlist_node *p, *next;
 191                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 192                 spin_lock(&b->lock);
 193                 hlist_for_each_safe(p, next, &b->list) {
 194                         struct kvm_task_sleep_node *n =
 195                                 hlist_entry(p, typeof(*n), link);
 196                         if (n->cpu == smp_processor_id())
 197                                 apf_task_wake_one(n);
 198                 }
 199                 spin_unlock(&b->lock);
 200         }
 201 }
 202
 203 void kvm_async_pf_task_wake(u32 token)
 204 {
 205         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 206         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 207         struct kvm_task_sleep_node *n;
 208
 209         if (token == ~0) {
 210                 apf_task_wake_all();
 211                 return;
 212         }
 213
 214 again:
 215         spin_lock(&b->lock);
 216         n = _find_apf_task(b, token);
 217         if (!n) {
 218                 /*
 219                  * async PF was not yet handled.
 220                  * Add dummy entry for the token.
 221                  */
 222                 n = kmalloc(sizeof(*n), GFP_ATOMIC);
 223                 if (!n) {
 224                         /*
 225                          * Allocation failed! Busy wait while other cpu
 226                          * handles async PF.
 227                          */
 228                         spin_unlock(&b->lock);
 229                         cpu_relax();
 230                         goto again;
 231                 }
 232                 n->token = token;
 233                 n->cpu = smp_processor_id();
 234                 n->mm = NULL;
 235                 init_waitqueue_head(&n->wq);
 236                 hlist_add_head(&n->link, &b->list);
 237         } else
 238                 apf_task_wake_one(n);
 239         spin_unlock(&b->lock);
 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
 248         if (__get_cpu_var(apf_reason).enabled) {
 249                 reason = __get_cpu_var(apf_reason).reason;
 250                 __get_cpu_var(apf_reason).reason = 0;
 251         }
 252
 253         return reason;
 254 }
 255 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 256
 257 dotraplinkage void __kprobes
 258 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
 259 {
 260         switch (kvm_read_and_reset_pf_reason()) {
 261         default:
 262                 do_page_fault(regs, error_code);
 263                 break;
 264         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 265                 /* page is swapped out by the host. */
 266                 kvm_async_pf_task_wait((u32)read_cr2());
 267                 break;
 268         case KVM_PV_REASON_PAGE_READY:
 269                 kvm_async_pf_task_wake((u32)read_cr2());
 270                 break;
 271         }
 272 }
 273
 274 static void kvm_mmu_op(void *buffer, unsigned len)
 275 {
 276         int r;
 277         unsigned long a1, a2;
 278
 279         do {
 280                 a1 = __pa(buffer);
 281                 a2 = 0;   /* on i386 __pa() always returns <4G */
 282                 r = kvm_hypercall3(KVM_HC_MMU_OP, len, a1, a2);
 283                 buffer += r;
 284                 len -= r;
 285         } while (len);
 286 }
 287
 288 static void mmu_queue_flush(struct kvm_para_state *state)
 289 {
 290         if (state->mmu_queue_len) {
 291                 kvm_mmu_op(state->mmu_queue, state->mmu_queue_len);
 292                 state->mmu_queue_len = 0;
 293         }
 294 }
 295
 296 static void kvm_deferred_mmu_op(void *buffer, int len)
 297 {
 298         struct kvm_para_state *state = kvm_para_state();
 299
 300         if (paravirt_get_lazy_mode() != PARAVIRT_LAZY_MMU) {
 301                 kvm_mmu_op(buffer, len);
 302                 return;
 303         }
 304         if (state->mmu_queue_len + len > sizeof state->mmu_queue)
 305                 mmu_queue_flush(state);
 306         memcpy(state->mmu_queue + state->mmu_queue_len, buffer, len);
 307         state->mmu_queue_len += len;
 308 }
 309
 310 static void kvm_mmu_write(void *dest, u64 val)
 311 {
 312         __u64 pte_phys;
 313         struct kvm_mmu_op_write_pte wpte;
 314
 315 #ifdef CONFIG_HIGHPTE
 316         struct page *page;
 317         unsigned long dst = (unsigned long) dest;
 318
 319         page = kmap_atomic_to_page(dest);
 320         pte_phys = page_to_pfn(page);
 321         pte_phys <<= PAGE_SHIFT;
 322         pte_phys += (dst & ~(PAGE_MASK));
 323 #else
 324         pte_phys = (unsigned long)__pa(dest);
 325 #endif
 326         wpte.header.op = KVM_MMU_OP_WRITE_PTE;
 327         wpte.pte_val = val;
 328         wpte.pte_phys = pte_phys;
 329
 330         kvm_deferred_mmu_op(&wpte, sizeof wpte);
 331 }
 332
 333 /*
 334  * We only need to hook operations that are MMU writes.  We hook these so that
 335  * we can use lazy MMU mode to batch these operations.  We could probably
 336  * improve the performance of the host code if we used some of the information
 337  * here to simplify processing of batched writes.
 338  */
 339 static void kvm_set_pte(pte_t *ptep, pte_t pte)
 340 {
 341         kvm_mmu_write(ptep, pte_val(pte));
 342 }
 343
 344 static void kvm_set_pte_at(struct mm_struct *mm, unsigned long addr,
 345                            pte_t *ptep, pte_t pte)
 346 {
 347         kvm_mmu_write(ptep, pte_val(pte));
 348 }
 349
 350 static void kvm_set_pmd(pmd_t *pmdp, pmd_t pmd)
 351 {
 352         kvm_mmu_write(pmdp, pmd_val(pmd));
 353 }
 354
 355 #if PAGETABLE_LEVELS >= 3
 356 #ifdef CONFIG_X86_PAE
 357 static void kvm_set_pte_atomic(pte_t *ptep, pte_t pte)
 358 {
 359         kvm_mmu_write(ptep, pte_val(pte));
 360 }
 361
 362 static void kvm_pte_clear(struct mm_struct *mm,
 363                           unsigned long addr, pte_t *ptep)
 364 {
 365         kvm_mmu_write(ptep, 0);
 366 }
 367
 368 static void kvm_pmd_clear(pmd_t *pmdp)
 369 {
 370         kvm_mmu_write(pmdp, 0);
 371 }
 372 #endif
 373
 374 static void kvm_set_pud(pud_t *pudp, pud_t pud)
 375 {
 376         kvm_mmu_write(pudp, pud_val(pud));
 377 }
 378
 379 #if PAGETABLE_LEVELS == 4
 380 static void kvm_set_pgd(pgd_t *pgdp, pgd_t pgd)
 381 {
 382         kvm_mmu_write(pgdp, pgd_val(pgd));
 383 }
 384 #endif
 385 #endif /* PAGETABLE_LEVELS >= 3 */
 386
 387 static void kvm_flush_tlb(void)
 388 {
 389         struct kvm_mmu_op_flush_tlb ftlb = {
 390                 .header.op = KVM_MMU_OP_FLUSH_TLB,
 391         };
 392
 393         kvm_deferred_mmu_op(&ftlb, sizeof ftlb);
 394 }
 395
 396 static void kvm_release_pt(unsigned long pfn)
 397 {
 398         struct kvm_mmu_op_release_pt rpt = {
 399                 .header.op = KVM_MMU_OP_RELEASE_PT,
 400                 .pt_phys = (u64)pfn << PAGE_SHIFT,
 401         };
 402
 403         kvm_mmu_op(&rpt, sizeof rpt);
 404 }
 405
 406 static void kvm_enter_lazy_mmu(void)
 407 {
 408         paravirt_enter_lazy_mmu();
 409 }
 410
 411 static void kvm_leave_lazy_mmu(void)
 412 {
 413         struct kvm_para_state *state = kvm_para_state();
 414
 415         mmu_queue_flush(state);
 416         paravirt_leave_lazy_mmu();
 417 }
 418
 419 static void __init paravirt_ops_setup(void)
 420 {
 421         pv_info.name = "KVM";
 422         pv_info.paravirt_enabled = 1;
 423
 424         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
 425                 pv_cpu_ops.io_delay = kvm_io_delay;
 426
 427         if (kvm_para_has_feature(KVM_FEATURE_MMU_OP)) {
 428                 pv_mmu_ops.set_pte = kvm_set_pte;
 429                 pv_mmu_ops.set_pte_at = kvm_set_pte_at;
 430                 pv_mmu_ops.set_pmd = kvm_set_pmd;
 431 #if PAGETABLE_LEVELS >= 3
 432 #ifdef CONFIG_X86_PAE
 433                 pv_mmu_ops.set_pte_atomic = kvm_set_pte_atomic;
 434                 pv_mmu_ops.pte_clear = kvm_pte_clear;
 435                 pv_mmu_ops.pmd_clear = kvm_pmd_clear;
 436 #endif
 437                 pv_mmu_ops.set_pud = kvm_set_pud;
 438 #if PAGETABLE_LEVELS == 4
 439                 pv_mmu_ops.set_pgd = kvm_set_pgd;
 440 #endif
 441 #endif
 442                 pv_mmu_ops.flush_tlb_user = kvm_flush_tlb;
 443                 pv_mmu_ops.release_pte = kvm_release_pt;
 444                 pv_mmu_ops.release_pmd = kvm_release_pt;
 445                 pv_mmu_ops.release_pud = kvm_release_pt;
 446
 447                 pv_mmu_ops.lazy_mode.enter = kvm_enter_lazy_mmu;
 448                 pv_mmu_ops.lazy_mode.leave = kvm_leave_lazy_mmu;
 449         }
 450 #ifdef CONFIG_X86_IO_APIC
 451         no_timer_check = 1;
 452 #endif
 453 }
 454
 455 static void kvm_register_steal_time(void)
 456 {
 457         int cpu = smp_processor_id();
 458         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
 459
 460         if (!has_steal_clock)
 461                 return;
 462
 463         memset(st, 0, sizeof(*st));
 464
 465         wrmsrl(MSR_KVM_STEAL_TIME, (__pa(st) | KVM_MSR_ENABLED));
 466         printk(KERN_INFO "kvm-stealtime: cpu %d, msr %lx\n",
 467                 cpu, __pa(st));
 468 }
 469
 470 void __cpuinit kvm_guest_cpu_init(void)
 471 {
 472         if (!kvm_para_available())
 473                 return;
 474
 475         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
 476                 u64 pa = __pa(&__get_cpu_var(apf_reason));
 477
 478 #ifdef CONFIG_PREEMPT
 479                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
 480 #endif
 481                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
 482                 __get_cpu_var(apf_reason).enabled = 1;
 483                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 484                        smp_processor_id());
 485         }
 486
 487         if (has_steal_clock)
 488                 kvm_register_steal_time();
 489 }
 490
 491 static void kvm_pv_disable_apf(void *unused)
 492 {
 493         if (!__get_cpu_var(apf_reason).enabled)
 494                 return;
 495
 496         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 497         __get_cpu_var(apf_reason).enabled = 0;
 498
 499         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 500                smp_processor_id());
 501 }
 502
 503 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 504                                 unsigned long code, void *unused)
 505 {
 506         if (code == SYS_RESTART)
 507                 on_each_cpu(kvm_pv_disable_apf, NULL, 1);
 508         return NOTIFY_DONE;
 509 }
 510
 511 static struct notifier_block kvm_pv_reboot_nb = {
 512         .notifier_call = kvm_pv_reboot_notify,
 513 };
 514
 515 static u64 kvm_steal_clock(int cpu)
 516 {
 517         u64 steal;
 518         struct kvm_steal_time *src;
 519         int version;
 520
 521         src = &per_cpu(steal_time, cpu);
 522         do {
 523                 version = src->version;
 524                 rmb();
 525                 steal = src->steal;
 526                 rmb();
 527         } while ((version & 1) || (version != src->version));
 528
 529         return steal;
 530 }
 531
 532 void kvm_disable_steal_time(void)
 533 {
 534         if (!has_steal_clock)
 535                 return;
 536
 537         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 538 }
 539
 540 #ifdef CONFIG_SMP
 541 static void __init kvm_smp_prepare_boot_cpu(void)
 542 {
 543 #ifdef CONFIG_KVM_CLOCK
 544         WARN_ON(kvm_register_clock("primary cpu clock"));
 545 #endif
 546         kvm_guest_cpu_init();
 547         native_smp_prepare_boot_cpu();
 548 }
 549
 550 static void __cpuinit kvm_guest_cpu_online(void *dummy)
 551 {
 552         kvm_guest_cpu_init();
 553 }
 554
 555 static void kvm_guest_cpu_offline(void *dummy)
 556 {
 557         kvm_disable_steal_time();
 558         kvm_pv_disable_apf(NULL);
 559         apf_task_wake_all();
 560 }
 561
 562 static int __cpuinit kvm_cpu_notify(struct notifier_block *self,
 563                                     unsigned long action, void *hcpu)
 564 {
 565         int cpu = (unsigned long)hcpu;
 566         switch (action) {
 567         case CPU_ONLINE:
 568         case CPU_DOWN_FAILED:
 569         case CPU_ONLINE_FROZEN:
 570                 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
 571                 break;
 572         case CPU_DOWN_PREPARE:
 573         case CPU_DOWN_PREPARE_FROZEN:
 574                 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
 575                 break;
 576         default:
 577                 break;
 578         }
 579         return NOTIFY_OK;
 580 }
 581
 582 static struct notifier_block __cpuinitdata kvm_cpu_notifier = {
 583         .notifier_call  = kvm_cpu_notify,
 584 };
 585 #endif
 586
 587 static void __init kvm_apf_trap_init(void)
 588 {
 589         set_intr_gate(14, &async_page_fault);
 590 }
 591
 592 void __init kvm_guest_init(void)
 593 {
 594         int i;
 595
 596         if (!kvm_para_available())
 597                 return;
 598
 599         paravirt_ops_setup();
 600         register_reboot_notifier(&kvm_pv_reboot_nb);
 601         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 602                 spin_lock_init(&async_pf_sleepers[i].lock);
 603         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 604                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 605
 606         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 607                 has_steal_clock = 1;
 608                 pv_time_ops.steal_clock = kvm_steal_clock;
 609         }
 610
 611 #ifdef CONFIG_SMP
 612         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 613         register_cpu_notifier(&kvm_cpu_notifier);
 614 #else
 615         kvm_guest_cpu_init();
 616 #endif
 617 }
 618
 619 static __init int activate_jump_labels(void)
 620 {
 621         if (has_steal_clock) {
 622                 jump_label_inc(&paravirt_steal_enabled);
 623                 if (steal_acc)
 624                         jump_label_inc(&paravirt_steal_rq_enabled);
 625         }
 626
 627         return 0;
 628 }
 629 arch_initcall(activate_jump_labels);