arch/x86/xen/enlighten.c

   1 /*
   2  * Core of Xen paravirt_ops implementation.
   3  *
   4  * This file contains the xen_paravirt_ops structure itself, and the
   5  * implementations for:
   6  * - privileged instructions
   7  * - interrupt flags
   8  * - segment operations
   9  * - booting and setup
  10  *
  11  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  12  */
  13
  14 #include <linux/kernel.h>
  15 #include <linux/init.h>
  16 #include <linux/smp.h>
  17 #include <linux/preempt.h>
  18 #include <linux/hardirq.h>
  19 #include <linux/percpu.h>
  20 #include <linux/delay.h>
  21 #include <linux/start_kernel.h>
  22 #include <linux/sched.h>
  23 #include <linux/bootmem.h>
  24 #include <linux/module.h>
  25 #include <linux/mm.h>
  26 #include <linux/page-flags.h>
  27 #include <linux/highmem.h>
  28 #include <linux/console.h>
  29
  30 #include <xen/interface/xen.h>
  31 #include <xen/interface/version.h>
  32 #include <xen/interface/physdev.h>
  33 #include <xen/interface/vcpu.h>
  34 #include <xen/features.h>
  35 #include <xen/page.h>
  36 #include <xen/hvc-console.h>
  37
  38 #include <asm/paravirt.h>
  39 #include <asm/apic.h>
  40 #include <asm/page.h>
  41 #include <asm/xen/hypercall.h>
  42 #include <asm/xen/hypervisor.h>
  43 #include <asm/fixmap.h>
  44 #include <asm/processor.h>
  45 #include <asm/msr-index.h>
  46 #include <asm/setup.h>
  47 #include <asm/desc.h>
  48 #include <asm/pgtable.h>
  49 #include <asm/tlbflush.h>
  50 #include <asm/reboot.h>
  51
  52 #include "xen-ops.h"
  53 #include "mmu.h"
  54 #include "multicalls.h"
  55
  56 EXPORT_SYMBOL_GPL(hypercall_page);
  57
  58 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
  59 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
  60
  61 enum xen_domain_type xen_domain_type = XEN_NATIVE;
  62 EXPORT_SYMBOL_GPL(xen_domain_type);
  63
  64 struct start_info *xen_start_info;
  65 EXPORT_SYMBOL_GPL(xen_start_info);
  66
  67 struct shared_info xen_dummy_shared_info;
  68
  69 void *xen_initial_gdt;
  70
  71 /*
  72  * Point at some empty memory to start with. We map the real shared_info
  73  * page as soon as fixmap is up and running.
  74  */
  75 struct shared_info *HYPERVISOR_shared_info = (void *)&xen_dummy_shared_info;
  76
  77 /*
  78  * Flag to determine whether vcpu info placement is available on all
  79  * VCPUs.  We assume it is to start with, and then set it to zero on
  80  * the first failure.  This is because it can succeed on some VCPUs
  81  * and not others, since it can involve hypervisor memory allocation,
  82  * or because the guest failed to guarantee all the appropriate
  83  * constraints on all VCPUs (ie buffer can't cross a page boundary).
  84  *
  85  * Note that any particular CPU may be using a placed vcpu structure,
  86  * but we can only optimise if the all are.
  87  *
  88  * 0: not available, 1: available
  89  */
  90 static int have_vcpu_info_placement = 1;
  91
  92 static void xen_vcpu_setup(int cpu)
  93 {
  94         struct vcpu_register_vcpu_info info;
  95         int err;
  96         struct vcpu_info *vcpup;
  97
  98         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
  99         per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
 100
 101         if (!have_vcpu_info_placement)
 102                 return;         /* already tested, not available */
 103
 104         vcpup = &per_cpu(xen_vcpu_info, cpu);
 105
 106         info.mfn = arbitrary_virt_to_mfn(vcpup);
 107         info.offset = offset_in_page(vcpup);
 108
 109         printk(KERN_DEBUG "trying to map vcpu_info %d at %p, mfn %llx, offset %d\n",
 110                cpu, vcpup, info.mfn, info.offset);
 111
 112         /* Check to see if the hypervisor will put the vcpu_info
 113            structure where we want it, which allows direct access via
 114            a percpu-variable. */
 115         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
 116
 117         if (err) {
 118                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
 119                 have_vcpu_info_placement = 0;
 120         } else {
 121                 /* This cpu is using the registered vcpu info, even if
 122                    later ones fail to. */
 123                 per_cpu(xen_vcpu, cpu) = vcpup;
 124
 125                 printk(KERN_DEBUG "cpu %d using vcpu_info at %p\n",
 126                        cpu, vcpup);
 127         }
 128 }
 129
 130 /*
 131  * On restore, set the vcpu placement up again.
 132  * If it fails, then we're in a bad state, since
 133  * we can't back out from using it...
 134  */
 135 void xen_vcpu_restore(void)
 136 {
 137         if (have_vcpu_info_placement) {
 138                 int cpu;
 139
 140                 for_each_online_cpu(cpu) {
 141                         bool other_cpu = (cpu != smp_processor_id());
 142
 143                         if (other_cpu &&
 144                             HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
 145                                 BUG();
 146
 147                         xen_vcpu_setup(cpu);
 148
 149                         if (other_cpu &&
 150                             HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
 151                                 BUG();
 152                 }
 153
 154                 BUG_ON(!have_vcpu_info_placement);
 155         }
 156 }
 157
 158 static void __init xen_banner(void)
 159 {
 160         unsigned version = HYPERVISOR_xen_version(XENVER_version, NULL);
 161         struct xen_extraversion extra;
 162         HYPERVISOR_xen_version(XENVER_extraversion, &extra);
 163
 164         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
 165                pv_info.name);
 166         printk(KERN_INFO "Xen version: %d.%d%s%s\n",
 167                version >> 16, version & 0xffff, extra.extraversion,
 168                xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
 169 }
 170
 171 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
 172                       unsigned int *cx, unsigned int *dx)
 173 {
 174         unsigned maskedx = ~0;
 175
 176         /*
 177          * Mask out inconvenient features, to try and disable as many
 178          * unsupported kernel subsystems as possible.
 179          */
 180         if (*ax == 1)
 181                 maskedx = ~((1 << X86_FEATURE_APIC) |  /* disable APIC */
 182                             (1 << X86_FEATURE_ACPI) |  /* disable ACPI */
 183                             (1 << X86_FEATURE_MCE)  |  /* disable MCE */
 184                             (1 << X86_FEATURE_MCA)  |  /* disable MCA */
 185                             (1 << X86_FEATURE_ACC));   /* thermal monitoring */
 186
 187         asm(XEN_EMULATE_PREFIX "cpuid"
 188                 : "=a" (*ax),
 189                   "=b" (*bx),
 190                   "=c" (*cx),
 191                   "=d" (*dx)
 192                 : "0" (*ax), "2" (*cx));
 193         *dx &= maskedx;
 194 }
 195
 196 static void xen_set_debugreg(int reg, unsigned long val)
 197 {
 198         HYPERVISOR_set_debugreg(reg, val);
 199 }
 200
 201 static unsigned long xen_get_debugreg(int reg)
 202 {
 203         return HYPERVISOR_get_debugreg(reg);
 204 }
 205
 206 void xen_leave_lazy(void)
 207 {
 208         paravirt_leave_lazy(paravirt_get_lazy_mode());
 209         xen_mc_flush();
 210 }
 211
 212 static unsigned long xen_store_tr(void)
 213 {
 214         return 0;
 215 }
 216
 217 /*
 218  * Set the page permissions for a particular virtual address.  If the
 219  * address is a vmalloc mapping (or other non-linear mapping), then
 220  * find the linear mapping of the page and also set its protections to
 221  * match.
 222  */
 223 static void set_aliased_prot(void *v, pgprot_t prot)
 224 {
 225         int level;
 226         pte_t *ptep;
 227         pte_t pte;
 228         unsigned long pfn;
 229         struct page *page;
 230
 231         ptep = lookup_address((unsigned long)v, &level);
 232         BUG_ON(ptep == NULL);
 233
 234         pfn = pte_pfn(*ptep);
 235         page = pfn_to_page(pfn);
 236
 237         pte = pfn_pte(pfn, prot);
 238
 239         if (HYPERVISOR_update_va_mapping((unsigned long)v, pte, 0))
 240                 BUG();
 241
 242         if (!PageHighMem(page)) {
 243                 void *av = __va(PFN_PHYS(pfn));
 244
 245                 if (av != v)
 246                         if (HYPERVISOR_update_va_mapping((unsigned long)av, pte, 0))
 247                                 BUG();
 248         } else
 249                 kmap_flush_unused();
 250 }
 251
 252 static void xen_alloc_ldt(struct desc_struct *ldt, unsigned entries)
 253 {
 254         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
 255         int i;
 256
 257         for(i = 0; i < entries; i += entries_per_page)
 258                 set_aliased_prot(ldt + i, PAGE_KERNEL_RO);
 259 }
 260
 261 static void xen_free_ldt(struct desc_struct *ldt, unsigned entries)
 262 {
 263         const unsigned entries_per_page = PAGE_SIZE / LDT_ENTRY_SIZE;
 264         int i;
 265
 266         for(i = 0; i < entries; i += entries_per_page)
 267                 set_aliased_prot(ldt + i, PAGE_KERNEL);
 268 }
 269
 270 static void xen_set_ldt(const void *addr, unsigned entries)
 271 {
 272         struct mmuext_op *op;
 273         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 274
 275         op = mcs.args;
 276         op->cmd = MMUEXT_SET_LDT;
 277         op->arg1.linear_addr = (unsigned long)addr;
 278         op->arg2.nr_ents = entries;
 279
 280         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 281
 282         xen_mc_issue(PARAVIRT_LAZY_CPU);
 283 }
 284
 285 static void xen_load_gdt(const struct desc_ptr *dtr)
 286 {
 287         unsigned long *frames;
 288         unsigned long va = dtr->address;
 289         unsigned int size = dtr->size + 1;
 290         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
 291         int f;
 292         struct multicall_space mcs;
 293
 294         /* A GDT can be up to 64k in size, which corresponds to 8192
 295            8-byte entries, or 16 4k pages.. */
 296
 297         BUG_ON(size > 65536);
 298         BUG_ON(va & ~PAGE_MASK);
 299
 300         mcs = xen_mc_entry(sizeof(*frames) * pages);
 301         frames = mcs.args;
 302
 303         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
 304                 frames[f] = arbitrary_virt_to_mfn((void *)va);
 305
 306                 make_lowmem_page_readonly((void *)va);
 307                 make_lowmem_page_readonly(mfn_to_virt(frames[f]));
 308         }
 309
 310         MULTI_set_gdt(mcs.mc, frames, size / sizeof(struct desc_struct));
 311
 312         xen_mc_issue(PARAVIRT_LAZY_CPU);
 313 }
 314
 315 static void load_TLS_descriptor(struct thread_struct *t,
 316                                 unsigned int cpu, unsigned int i)
 317 {
 318         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
 319         xmaddr_t maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
 320         struct multicall_space mc = __xen_mc_entry(0);
 321
 322         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
 323 }
 324
 325 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
 326 {
 327         /*
 328          * XXX sleazy hack: If we're being called in a lazy-cpu zone
 329          * and lazy gs handling is enabled, it means we're in a
 330          * context switch, and %gs has just been saved.  This means we
 331          * can zero it out to prevent faults on exit from the
 332          * hypervisor if the next process has no %gs.  Either way, it
 333          * has been saved, and the new value will get loaded properly.
 334          * This will go away as soon as Xen has been modified to not
 335          * save/restore %gs for normal hypercalls.
 336          *
 337          * On x86_64, this hack is not used for %gs, because gs points
 338          * to KERNEL_GS_BASE (and uses it for PDA references), so we
 339          * must not zero %gs on x86_64
 340          *
 341          * For x86_64, we need to zero %fs, otherwise we may get an
 342          * exception between the new %fs descriptor being loaded and
 343          * %fs being effectively cleared at __switch_to().
 344          */
 345         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU) {
 346 #ifdef CONFIG_X86_32
 347                 lazy_load_gs(0);
 348 #else
 349                 loadsegment(fs, 0);
 350 #endif
 351         }
 352
 353         xen_mc_batch();
 354
 355         load_TLS_descriptor(t, cpu, 0);
 356         load_TLS_descriptor(t, cpu, 1);
 357         load_TLS_descriptor(t, cpu, 2);
 358
 359         xen_mc_issue(PARAVIRT_LAZY_CPU);
 360 }
 361
 362 #ifdef CONFIG_X86_64
 363 static void xen_load_gs_index(unsigned int idx)
 364 {
 365         if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
 366                 BUG();
 367 }
 368 #endif
 369
 370 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
 371                                 const void *ptr)
 372 {
 373         xmaddr_t mach_lp = arbitrary_virt_to_machine(&dt[entrynum]);
 374         u64 entry = *(u64 *)ptr;
 375
 376         preempt_disable();
 377
 378         xen_mc_flush();
 379         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
 380                 BUG();
 381
 382         preempt_enable();
 383 }
 384
 385 static int cvt_gate_to_trap(int vector, const gate_desc *val,
 386                             struct trap_info *info)
 387 {
 388         if (val->type != 0xf && val->type != 0xe)
 389                 return 0;
 390
 391         info->vector = vector;
 392         info->address = gate_offset(*val);
 393         info->cs = gate_segment(*val);
 394         info->flags = val->dpl;
 395         /* interrupt gates clear IF */
 396         if (val->type == 0xe)
 397                 info->flags |= 4;
 398
 399         return 1;
 400 }
 401
 402 /* Locations of each CPU's IDT */
 403 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
 404
 405 /* Set an IDT entry.  If the entry is part of the current IDT, then
 406    also update Xen. */
 407 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
 408 {
 409         unsigned long p = (unsigned long)&dt[entrynum];
 410         unsigned long start, end;
 411
 412         preempt_disable();
 413
 414         start = __get_cpu_var(idt_desc).address;
 415         end = start + __get_cpu_var(idt_desc).size + 1;
 416
 417         xen_mc_flush();
 418
 419         native_write_idt_entry(dt, entrynum, g);
 420
 421         if (p >= start && (p + 8) <= end) {
 422                 struct trap_info info[2];
 423
 424                 info[1].address = 0;
 425
 426                 if (cvt_gate_to_trap(entrynum, g, &info[0]))
 427                         if (HYPERVISOR_set_trap_table(info))
 428                                 BUG();
 429         }
 430
 431         preempt_enable();
 432 }
 433
 434 static void xen_convert_trap_info(const struct desc_ptr *desc,
 435                                   struct trap_info *traps)
 436 {
 437         unsigned in, out, count;
 438
 439         count = (desc->size+1) / sizeof(gate_desc);
 440         BUG_ON(count > 256);
 441
 442         for (in = out = 0; in < count; in++) {
 443                 gate_desc *entry = (gate_desc*)(desc->address) + in;
 444
 445                 if (cvt_gate_to_trap(in, entry, &traps[out]))
 446                         out++;
 447         }
 448         traps[out].address = 0;
 449 }
 450
 451 void xen_copy_trap_info(struct trap_info *traps)
 452 {
 453         const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
 454
 455         xen_convert_trap_info(desc, traps);
 456 }
 457
 458 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
 459    hold a spinlock to protect the static traps[] array (static because
 460    it avoids allocation, and saves stack space). */
 461 static void xen_load_idt(const struct desc_ptr *desc)
 462 {
 463         static DEFINE_SPINLOCK(lock);
 464         static struct trap_info traps[257];
 465
 466         spin_lock(&lock);
 467
 468         __get_cpu_var(idt_desc) = *desc;
 469
 470         xen_convert_trap_info(desc, traps);
 471
 472         xen_mc_flush();
 473         if (HYPERVISOR_set_trap_table(traps))
 474                 BUG();
 475
 476         spin_unlock(&lock);
 477 }
 478
 479 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
 480    they're handled differently. */
 481 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
 482                                 const void *desc, int type)
 483 {
 484         preempt_disable();
 485
 486         switch (type) {
 487         case DESC_LDT:
 488         case DESC_TSS:
 489                 /* ignore */
 490                 break;
 491
 492         default: {
 493                 xmaddr_t maddr = arbitrary_virt_to_machine(&dt[entry]);
 494
 495                 xen_mc_flush();
 496                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
 497                         BUG();
 498         }
 499
 500         }
 501
 502         preempt_enable();
 503 }
 504
 505 static void xen_load_sp0(struct tss_struct *tss,
 506                          struct thread_struct *thread)
 507 {
 508         struct multicall_space mcs = xen_mc_entry(0);
 509         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
 510         xen_mc_issue(PARAVIRT_LAZY_CPU);
 511 }
 512
 513 static void xen_set_iopl_mask(unsigned mask)
 514 {
 515         struct physdev_set_iopl set_iopl;
 516
 517         /* Force the change at ring 0. */
 518         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
 519         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
 520 }
 521
 522 static void xen_io_delay(void)
 523 {
 524 }
 525
 526 #ifdef CONFIG_X86_LOCAL_APIC
 527 static u32 xen_apic_read(u32 reg)
 528 {
 529         return 0;
 530 }
 531
 532 static void xen_apic_write(u32 reg, u32 val)
 533 {
 534         /* Warn to see if there's any stray references */
 535         WARN_ON(1);
 536 }
 537
 538 static u64 xen_apic_icr_read(void)
 539 {
 540         return 0;
 541 }
 542
 543 static void xen_apic_icr_write(u32 low, u32 id)
 544 {
 545         /* Warn to see if there's any stray references */
 546         WARN_ON(1);
 547 }
 548
 549 static void xen_apic_wait_icr_idle(void)
 550 {
 551         return;
 552 }
 553
 554 static u32 xen_safe_apic_wait_icr_idle(void)
 555 {
 556         return 0;
 557 }
 558
 559 static void set_xen_basic_apic_ops(void)
 560 {
 561         apic->read = xen_apic_read;
 562         apic->write = xen_apic_write;
 563         apic->icr_read = xen_apic_icr_read;
 564         apic->icr_write = xen_apic_icr_write;
 565         apic->wait_icr_idle = xen_apic_wait_icr_idle;
 566         apic->safe_wait_icr_idle = xen_safe_apic_wait_icr_idle;
 567 }
 568
 569 #endif
 570
 571
 572 static void xen_clts(void)
 573 {
 574         struct multicall_space mcs;
 575
 576         mcs = xen_mc_entry(0);
 577
 578         MULTI_fpu_taskswitch(mcs.mc, 0);
 579
 580         xen_mc_issue(PARAVIRT_LAZY_CPU);
 581 }
 582
 583 static void xen_write_cr0(unsigned long cr0)
 584 {
 585         struct multicall_space mcs;
 586
 587         /* Only pay attention to cr0.TS; everything else is
 588            ignored. */
 589         mcs = xen_mc_entry(0);
 590
 591         MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
 592
 593         xen_mc_issue(PARAVIRT_LAZY_CPU);
 594 }
 595
 596 static void xen_write_cr4(unsigned long cr4)
 597 {
 598         cr4 &= ~X86_CR4_PGE;
 599         cr4 &= ~X86_CR4_PSE;
 600
 601         native_write_cr4(cr4);
 602 }
 603
 604 static int xen_write_msr_safe(unsigned int msr, unsigned low, unsigned high)
 605 {
 606         int ret;
 607
 608         ret = 0;
 609
 610         switch (msr) {
 611 #ifdef CONFIG_X86_64
 612                 unsigned which;
 613                 u64 base;
 614
 615         case MSR_FS_BASE:               which = SEGBASE_FS; goto set;
 616         case MSR_KERNEL_GS_BASE:        which = SEGBASE_GS_USER; goto set;
 617         case MSR_GS_BASE:               which = SEGBASE_GS_KERNEL; goto set;
 618
 619         set:
 620                 base = ((u64)high << 32) | low;
 621                 if (HYPERVISOR_set_segment_base(which, base) != 0)
 622                         ret = -EFAULT;
 623                 break;
 624 #endif
 625
 626         case MSR_STAR:
 627         case MSR_CSTAR:
 628         case MSR_LSTAR:
 629         case MSR_SYSCALL_MASK:
 630         case MSR_IA32_SYSENTER_CS:
 631         case MSR_IA32_SYSENTER_ESP:
 632         case MSR_IA32_SYSENTER_EIP:
 633                 /* Fast syscall setup is all done in hypercalls, so
 634                    these are all ignored.  Stub them out here to stop
 635                    Xen console noise. */
 636                 break;
 637
 638         default:
 639                 ret = native_write_msr_safe(msr, low, high);
 640         }
 641
 642         return ret;
 643 }
 644
 645 void xen_setup_shared_info(void)
 646 {
 647         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 648                 set_fixmap(FIX_PARAVIRT_BOOTMAP,
 649                            xen_start_info->shared_info);
 650
 651                 HYPERVISOR_shared_info =
 652                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
 653         } else
 654                 HYPERVISOR_shared_info =
 655                         (struct shared_info *)__va(xen_start_info->shared_info);
 656
 657 #ifndef CONFIG_SMP
 658         /* In UP this is as good a place as any to set up shared info */
 659         xen_setup_vcpu_info_placement();
 660 #endif
 661
 662         xen_setup_mfn_list_list();
 663 }
 664
 665 /* This is called once we have the cpu_possible_map */
 666 void xen_setup_vcpu_info_placement(void)
 667 {
 668         int cpu;
 669
 670         for_each_possible_cpu(cpu)
 671                 xen_vcpu_setup(cpu);
 672
 673         /* xen_vcpu_setup managed to place the vcpu_info within the
 674            percpu area for all cpus, so make use of it */
 675         if (have_vcpu_info_placement) {
 676                 printk(KERN_INFO "Xen: using vcpu_info placement\n");
 677
 678                 pv_irq_ops.save_fl = __PV_IS_CALLEE_SAVE(xen_save_fl_direct);
 679                 pv_irq_ops.restore_fl = __PV_IS_CALLEE_SAVE(xen_restore_fl_direct);
 680                 pv_irq_ops.irq_disable = __PV_IS_CALLEE_SAVE(xen_irq_disable_direct);
 681                 pv_irq_ops.irq_enable = __PV_IS_CALLEE_SAVE(xen_irq_enable_direct);
 682                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
 683         }
 684 }
 685
 686 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
 687                           unsigned long addr, unsigned len)
 688 {
 689         char *start, *end, *reloc;
 690         unsigned ret;
 691
 692         start = end = reloc = NULL;
 693
 694 #define SITE(op, x)                                                     \
 695         case PARAVIRT_PATCH(op.x):                                      \
 696         if (have_vcpu_info_placement) {                                 \
 697                 start = (char *)xen_##x##_direct;                       \
 698                 end = xen_##x##_direct_end;                             \
 699                 reloc = xen_##x##_direct_reloc;                         \
 700         }                                                               \
 701         goto patch_site
 702
 703         switch (type) {
 704                 SITE(pv_irq_ops, irq_enable);
 705                 SITE(pv_irq_ops, irq_disable);
 706                 SITE(pv_irq_ops, save_fl);
 707                 SITE(pv_irq_ops, restore_fl);
 708 #undef SITE
 709
 710         patch_site:
 711                 if (start == NULL || (end-start) > len)
 712                         goto default_patch;
 713
 714                 ret = paravirt_patch_insns(insnbuf, len, start, end);
 715
 716                 /* Note: because reloc is assigned from something that
 717                    appears to be an array, gcc assumes it's non-null,
 718                    but doesn't know its relationship with start and
 719                    end. */
 720                 if (reloc > start && reloc < end) {
 721                         int reloc_off = reloc - start;
 722                         long *relocp = (long *)(insnbuf + reloc_off);
 723                         long delta = start - (char *)addr;
 724
 725                         *relocp += delta;
 726                 }
 727                 break;
 728
 729         default_patch:
 730         default:
 731                 ret = paravirt_patch_default(type, clobbers, insnbuf,
 732                                              addr, len);
 733                 break;
 734         }
 735
 736         return ret;
 737 }
 738
 739 static const struct pv_info xen_info __initdata = {
 740         .paravirt_enabled = 1,
 741         .shared_kernel_pmd = 0,
 742
 743         .name = "Xen",
 744 };
 745
 746 static const struct pv_init_ops xen_init_ops __initdata = {
 747         .patch = xen_patch,
 748
 749         .banner = xen_banner,
 750         .memory_setup = xen_memory_setup,
 751         .arch_setup = xen_arch_setup,
 752         .post_allocator_init = xen_post_allocator_init,
 753 };
 754
 755 static const struct pv_time_ops xen_time_ops __initdata = {
 756         .time_init = xen_time_init,
 757
 758         .set_wallclock = xen_set_wallclock,
 759         .get_wallclock = xen_get_wallclock,
 760         .get_tsc_khz = xen_tsc_khz,
 761         .sched_clock = xen_sched_clock,
 762 };
 763
 764 static const struct pv_cpu_ops xen_cpu_ops __initdata = {
 765         .cpuid = xen_cpuid,
 766
 767         .set_debugreg = xen_set_debugreg,
 768         .get_debugreg = xen_get_debugreg,
 769
 770         .clts = xen_clts,
 771
 772         .read_cr0 = native_read_cr0,
 773         .write_cr0 = xen_write_cr0,
 774
 775         .read_cr4 = native_read_cr4,
 776         .read_cr4_safe = native_read_cr4_safe,
 777         .write_cr4 = xen_write_cr4,
 778
 779         .wbinvd = native_wbinvd,
 780
 781         .read_msr = native_read_msr_safe,
 782         .write_msr = xen_write_msr_safe,
 783         .read_tsc = native_read_tsc,
 784         .read_pmc = native_read_pmc,
 785
 786         .iret = xen_iret,
 787         .irq_enable_sysexit = xen_sysexit,
 788 #ifdef CONFIG_X86_64
 789         .usergs_sysret32 = xen_sysret32,
 790         .usergs_sysret64 = xen_sysret64,
 791 #endif
 792
 793         .load_tr_desc = paravirt_nop,
 794         .set_ldt = xen_set_ldt,
 795         .load_gdt = xen_load_gdt,
 796         .load_idt = xen_load_idt,
 797         .load_tls = xen_load_tls,
 798 #ifdef CONFIG_X86_64
 799         .load_gs_index = xen_load_gs_index,
 800 #endif
 801
 802         .alloc_ldt = xen_alloc_ldt,
 803         .free_ldt = xen_free_ldt,
 804
 805         .store_gdt = native_store_gdt,
 806         .store_idt = native_store_idt,
 807         .store_tr = xen_store_tr,
 808
 809         .write_ldt_entry = xen_write_ldt_entry,
 810         .write_gdt_entry = xen_write_gdt_entry,
 811         .write_idt_entry = xen_write_idt_entry,
 812         .load_sp0 = xen_load_sp0,
 813
 814         .set_iopl_mask = xen_set_iopl_mask,
 815         .io_delay = xen_io_delay,
 816
 817         /* Xen takes care of %gs when switching to usermode for us */
 818         .swapgs = paravirt_nop,
 819
 820         .lazy_mode = {
 821                 .enter = paravirt_enter_lazy_cpu,
 822                 .leave = xen_leave_lazy,
 823         },
 824 };
 825
 826 static const struct pv_apic_ops xen_apic_ops __initdata = {
 827 #ifdef CONFIG_X86_LOCAL_APIC
 828         .setup_boot_clock = paravirt_nop,
 829         .setup_secondary_clock = paravirt_nop,
 830         .startup_ipi_hook = paravirt_nop,
 831 #endif
 832 };
 833
 834 static void xen_reboot(int reason)
 835 {
 836         struct sched_shutdown r = { .reason = reason };
 837
 838 #ifdef CONFIG_SMP
 839         smp_send_stop();
 840 #endif
 841
 842         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
 843                 BUG();
 844 }
 845
 846 static void xen_restart(char *msg)
 847 {
 848         xen_reboot(SHUTDOWN_reboot);
 849 }
 850
 851 static void xen_emergency_restart(void)
 852 {
 853         xen_reboot(SHUTDOWN_reboot);
 854 }
 855
 856 static void xen_machine_halt(void)
 857 {
 858         xen_reboot(SHUTDOWN_poweroff);
 859 }
 860
 861 static void xen_crash_shutdown(struct pt_regs *regs)
 862 {
 863         xen_reboot(SHUTDOWN_crash);
 864 }
 865
 866 static const struct machine_ops __initdata xen_machine_ops = {
 867         .restart = xen_restart,
 868         .halt = xen_machine_halt,
 869         .power_off = xen_machine_halt,
 870         .shutdown = xen_machine_halt,
 871         .crash_shutdown = xen_crash_shutdown,
 872         .emergency_restart = xen_emergency_restart,
 873 };
 874
 875
 876 /* First C function to be called on Xen boot */
 877 asmlinkage void __init xen_start_kernel(void)
 878 {
 879         pgd_t *pgd;
 880
 881         if (!xen_start_info)
 882                 return;
 883
 884         xen_domain_type = XEN_PV_DOMAIN;
 885
 886         BUG_ON(memcmp(xen_start_info->magic, "xen-3", 5) != 0);
 887
 888         xen_setup_features();
 889
 890         /* Install Xen paravirt ops */
 891         pv_info = xen_info;
 892         pv_init_ops = xen_init_ops;
 893         pv_time_ops = xen_time_ops;
 894         pv_cpu_ops = xen_cpu_ops;
 895         pv_apic_ops = xen_apic_ops;
 896         pv_mmu_ops = xen_mmu_ops;
 897
 898         xen_init_irq_ops();
 899
 900 #ifdef CONFIG_X86_LOCAL_APIC
 901         /*
 902          * set up the basic apic ops.
 903          */
 904         set_xen_basic_apic_ops();
 905 #endif
 906
 907         if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
 908                 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
 909                 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
 910         }
 911
 912         machine_ops = xen_machine_ops;
 913
 914 #ifdef CONFIG_X86_64
 915         /*
 916          * Setup percpu state.  We only need to do this for 64-bit
 917          * because 32-bit already has %fs set properly.
 918          */
 919         load_percpu_segment(0);
 920 #endif
 921         /*
 922          * The only reliable way to retain the initial address of the
 923          * percpu gdt_page is to remember it here, so we can go and
 924          * mark it RW later, when the initial percpu area is freed.
 925          */
 926         xen_initial_gdt = &per_cpu(gdt_page, 0);
 927
 928         xen_smp_init();
 929
 930         /* Get mfn list */
 931         if (!xen_feature(XENFEAT_auto_translated_physmap))
 932                 xen_build_dynamic_phys_to_machine();
 933
 934         pgd = (pgd_t *)xen_start_info->pt_base;
 935
 936         /* Prevent unwanted bits from being set in PTEs. */
 937         __supported_pte_mask &= ~_PAGE_GLOBAL;
 938         if (!xen_initial_domain())
 939                 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
 940
 941         /* Don't do the full vcpu_info placement stuff until we have a
 942            possible map and a non-dummy shared_info. */
 943         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
 944
 945         local_irq_disable();
 946         early_boot_irqs_off();
 947
 948         xen_raw_console_write("mapping kernel into physical memory\n");
 949         pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages);
 950
 951         init_mm.pgd = pgd;
 952
 953         /* keep using Xen gdt for now; no urgent need to change it */
 954
 955         pv_info.kernel_rpl = 1;
 956         if (xen_feature(XENFEAT_supervisor_mode_kernel))
 957                 pv_info.kernel_rpl = 0;
 958
 959         /* set the limit of our address space */
 960         xen_reserve_top();
 961
 962 #ifdef CONFIG_X86_32
 963         /* set up basic CPUID stuff */
 964         cpu_detect(&new_cpu_data);
 965         new_cpu_data.hard_math = 1;
 966         new_cpu_data.x86_capability[0] = cpuid_edx(1);
 967 #endif
 968
 969         /* Poke various useful things into boot_params */
 970         boot_params.hdr.type_of_loader = (9 << 4) | 0;
 971         boot_params.hdr.ramdisk_image = xen_start_info->mod_start
 972                 ? __pa(xen_start_info->mod_start) : 0;
 973         boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
 974         boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
 975
 976         if (!xen_initial_domain()) {
 977                 add_preferred_console("xenboot", 0, NULL);
 978                 add_preferred_console("tty", 0, NULL);
 979                 add_preferred_console("hvc", 0, NULL);
 980         }
 981
 982         xen_raw_console_write("about to get started...\n");
 983
 984         /* Start the world */
 985 #ifdef CONFIG_X86_32
 986         i386_start_kernel();
 987 #else
 988         x86_64_start_reservations((char *)__pa_symbol(&boot_params));
 989 #endif
 990 }