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/physdev.h>
  32 #include <xen/interface/vcpu.h>
  33 #include <xen/interface/sched.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/page.h>
  40 #include <asm/xen/hypercall.h>
  41 #include <asm/xen/hypervisor.h>
  42 #include <asm/fixmap.h>
  43 #include <asm/processor.h>
  44 #include <asm/setup.h>
  45 #include <asm/desc.h>
  46 #include <asm/pgtable.h>
  47 #include <asm/tlbflush.h>
  48 #include <asm/reboot.h>
  49 #include <asm/pgalloc.h>
  50
  51 #include "xen-ops.h"
  52 #include "mmu.h"
  53 #include "multicalls.h"
  54
  55 EXPORT_SYMBOL_GPL(hypercall_page);
  56
  57 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
  58 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
  59
  60 /*
  61  * Note about cr3 (pagetable base) values:
  62  *
  63  * xen_cr3 contains the current logical cr3 value; it contains the
  64  * last set cr3.  This may not be the current effective cr3, because
  65  * its update may be being lazily deferred.  However, a vcpu looking
  66  * at its own cr3 can use this value knowing that it everything will
  67  * be self-consistent.
  68  *
  69  * xen_current_cr3 contains the actual vcpu cr3; it is set once the
  70  * hypercall to set the vcpu cr3 is complete (so it may be a little
  71  * out of date, but it will never be set early).  If one vcpu is
  72  * looking at another vcpu's cr3 value, it should use this variable.
  73  */
  74 DEFINE_PER_CPU(unsigned long, xen_cr3);  /* cr3 stored as physaddr */
  75 DEFINE_PER_CPU(unsigned long, xen_current_cr3);  /* actual vcpu cr3 */
  76
  77 struct start_info *xen_start_info;
  78 EXPORT_SYMBOL_GPL(xen_start_info);
  79
  80 struct shared_info xen_dummy_shared_info;
  81
  82 /*
  83  * Point at some empty memory to start with. We map the real shared_info
  84  * page as soon as fixmap is up and running.
  85  */
  86 struct shared_info *HYPERVISOR_shared_info = (void *)&xen_dummy_shared_info;
  87
  88 /*
  89  * Flag to determine whether vcpu info placement is available on all
  90  * VCPUs.  We assume it is to start with, and then set it to zero on
  91  * the first failure.  This is because it can succeed on some VCPUs
  92  * and not others, since it can involve hypervisor memory allocation,
  93  * or because the guest failed to guarantee all the appropriate
  94  * constraints on all VCPUs (ie buffer can't cross a page boundary).
  95  *
  96  * Note that any particular CPU may be using a placed vcpu structure,
  97  * but we can only optimise if the all are.
  98  *
  99  * 0: not available, 1: available
 100  */
 101 static int have_vcpu_info_placement = 1;
 102
 103 static void xen_vcpu_setup(int cpu)
 104 {
 105         struct vcpu_register_vcpu_info info;
 106         int err;
 107         struct vcpu_info *vcpup;
 108
 109         BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
 110         per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
 111
 112         if (!have_vcpu_info_placement)
 113                 return;         /* already tested, not available */
 114
 115         vcpup = &per_cpu(xen_vcpu_info, cpu);
 116
 117         info.mfn = virt_to_mfn(vcpup);
 118         info.offset = offset_in_page(vcpup);
 119
 120         printk(KERN_DEBUG "trying to map vcpu_info %d at %p, mfn %llx, offset %d\n",
 121                cpu, vcpup, info.mfn, info.offset);
 122
 123         /* Check to see if the hypervisor will put the vcpu_info
 124            structure where we want it, which allows direct access via
 125            a percpu-variable. */
 126         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
 127
 128         if (err) {
 129                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
 130                 have_vcpu_info_placement = 0;
 131         } else {
 132                 /* This cpu is using the registered vcpu info, even if
 133                    later ones fail to. */
 134                 per_cpu(xen_vcpu, cpu) = vcpup;
 135
 136                 printk(KERN_DEBUG "cpu %d using vcpu_info at %p\n",
 137                        cpu, vcpup);
 138         }
 139 }
 140
 141 /*
 142  * On restore, set the vcpu placement up again.
 143  * If it fails, then we're in a bad state, since
 144  * we can't back out from using it...
 145  */
 146 void xen_vcpu_restore(void)
 147 {
 148         if (have_vcpu_info_placement) {
 149                 int cpu;
 150
 151                 for_each_online_cpu(cpu) {
 152                         bool other_cpu = (cpu != smp_processor_id());
 153
 154                         if (other_cpu &&
 155                             HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL))
 156                                 BUG();
 157
 158                         xen_vcpu_setup(cpu);
 159
 160                         if (other_cpu &&
 161                             HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL))
 162                                 BUG();
 163                 }
 164
 165                 BUG_ON(!have_vcpu_info_placement);
 166         }
 167 }
 168
 169 static void __init xen_banner(void)
 170 {
 171         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
 172                pv_info.name);
 173         printk(KERN_INFO "Hypervisor signature: %s%s\n",
 174                xen_start_info->magic,
 175                xen_feature(XENFEAT_mmu_pt_update_preserve_ad) ? " (preserve-AD)" : "");
 176 }
 177
 178 static void xen_cpuid(unsigned int *ax, unsigned int *bx,
 179                       unsigned int *cx, unsigned int *dx)
 180 {
 181         unsigned maskedx = ~0;
 182
 183         /*
 184          * Mask out inconvenient features, to try and disable as many
 185          * unsupported kernel subsystems as possible.
 186          */
 187         if (*ax == 1)
 188                 maskedx = ~((1 << X86_FEATURE_APIC) |  /* disable APIC */
 189                             (1 << X86_FEATURE_ACPI) |  /* disable ACPI */
 190                             (1 << X86_FEATURE_MCE)  |  /* disable MCE */
 191                             (1 << X86_FEATURE_MCA)  |  /* disable MCA */
 192                             (1 << X86_FEATURE_ACC));   /* thermal monitoring */
 193
 194         asm(XEN_EMULATE_PREFIX "cpuid"
 195                 : "=a" (*ax),
 196                   "=b" (*bx),
 197                   "=c" (*cx),
 198                   "=d" (*dx)
 199                 : "0" (*ax), "2" (*cx));
 200         *dx &= maskedx;
 201 }
 202
 203 static void xen_set_debugreg(int reg, unsigned long val)
 204 {
 205         HYPERVISOR_set_debugreg(reg, val);
 206 }
 207
 208 static unsigned long xen_get_debugreg(int reg)
 209 {
 210         return HYPERVISOR_get_debugreg(reg);
 211 }
 212
 213 static unsigned long xen_save_fl(void)
 214 {
 215         struct vcpu_info *vcpu;
 216         unsigned long flags;
 217
 218         vcpu = x86_read_percpu(xen_vcpu);
 219
 220         /* flag has opposite sense of mask */
 221         flags = !vcpu->evtchn_upcall_mask;
 222
 223         /* convert to IF type flag
 224            -0 -> 0x00000000
 225            -1 -> 0xffffffff
 226         */
 227         return (-flags) & X86_EFLAGS_IF;
 228 }
 229
 230 static void xen_restore_fl(unsigned long flags)
 231 {
 232         struct vcpu_info *vcpu;
 233
 234         /* convert from IF type flag */
 235         flags = !(flags & X86_EFLAGS_IF);
 236
 237         /* There's a one instruction preempt window here.  We need to
 238            make sure we're don't switch CPUs between getting the vcpu
 239            pointer and updating the mask. */
 240         preempt_disable();
 241         vcpu = x86_read_percpu(xen_vcpu);
 242         vcpu->evtchn_upcall_mask = flags;
 243         preempt_enable_no_resched();
 244
 245         /* Doesn't matter if we get preempted here, because any
 246            pending event will get dealt with anyway. */
 247
 248         if (flags == 0) {
 249                 preempt_check_resched();
 250                 barrier(); /* unmask then check (avoid races) */
 251                 if (unlikely(vcpu->evtchn_upcall_pending))
 252                         force_evtchn_callback();
 253         }
 254 }
 255
 256 static void xen_irq_disable(void)
 257 {
 258         /* There's a one instruction preempt window here.  We need to
 259            make sure we're don't switch CPUs between getting the vcpu
 260            pointer and updating the mask. */
 261         preempt_disable();
 262         x86_read_percpu(xen_vcpu)->evtchn_upcall_mask = 1;
 263         preempt_enable_no_resched();
 264 }
 265
 266 static void xen_irq_enable(void)
 267 {
 268         struct vcpu_info *vcpu;
 269
 270         /* We don't need to worry about being preempted here, since
 271            either a) interrupts are disabled, so no preemption, or b)
 272            the caller is confused and is trying to re-enable interrupts
 273            on an indeterminate processor. */
 274
 275         vcpu = x86_read_percpu(xen_vcpu);
 276         vcpu->evtchn_upcall_mask = 0;
 277
 278         /* Doesn't matter if we get preempted here, because any
 279            pending event will get dealt with anyway. */
 280
 281         barrier(); /* unmask then check (avoid races) */
 282         if (unlikely(vcpu->evtchn_upcall_pending))
 283                 force_evtchn_callback();
 284 }
 285
 286 static void xen_safe_halt(void)
 287 {
 288         /* Blocking includes an implicit local_irq_enable(). */
 289         if (HYPERVISOR_sched_op(SCHEDOP_block, NULL) != 0)
 290                 BUG();
 291 }
 292
 293 static void xen_halt(void)
 294 {
 295         if (irqs_disabled())
 296                 HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
 297         else
 298                 xen_safe_halt();
 299 }
 300
 301 static void xen_leave_lazy(void)
 302 {
 303         paravirt_leave_lazy(paravirt_get_lazy_mode());
 304         xen_mc_flush();
 305 }
 306
 307 static unsigned long xen_store_tr(void)
 308 {
 309         return 0;
 310 }
 311
 312 static void xen_set_ldt(const void *addr, unsigned entries)
 313 {
 314         struct mmuext_op *op;
 315         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 316
 317         op = mcs.args;
 318         op->cmd = MMUEXT_SET_LDT;
 319         op->arg1.linear_addr = (unsigned long)addr;
 320         op->arg2.nr_ents = entries;
 321
 322         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 323
 324         xen_mc_issue(PARAVIRT_LAZY_CPU);
 325 }
 326
 327 static void xen_load_gdt(const struct desc_ptr *dtr)
 328 {
 329         unsigned long *frames;
 330         unsigned long va = dtr->address;
 331         unsigned int size = dtr->size + 1;
 332         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
 333         int f;
 334         struct multicall_space mcs;
 335
 336         /* A GDT can be up to 64k in size, which corresponds to 8192
 337            8-byte entries, or 16 4k pages.. */
 338
 339         BUG_ON(size > 65536);
 340         BUG_ON(va & ~PAGE_MASK);
 341
 342         mcs = xen_mc_entry(sizeof(*frames) * pages);
 343         frames = mcs.args;
 344
 345         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
 346                 frames[f] = virt_to_mfn(va);
 347                 make_lowmem_page_readonly((void *)va);
 348         }
 349
 350         MULTI_set_gdt(mcs.mc, frames, size / sizeof(struct desc_struct));
 351
 352         xen_mc_issue(PARAVIRT_LAZY_CPU);
 353 }
 354
 355 static void load_TLS_descriptor(struct thread_struct *t,
 356                                 unsigned int cpu, unsigned int i)
 357 {
 358         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
 359         xmaddr_t maddr = virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
 360         struct multicall_space mc = __xen_mc_entry(0);
 361
 362         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
 363 }
 364
 365 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
 366 {
 367         xen_mc_batch();
 368
 369         load_TLS_descriptor(t, cpu, 0);
 370         load_TLS_descriptor(t, cpu, 1);
 371         load_TLS_descriptor(t, cpu, 2);
 372
 373         xen_mc_issue(PARAVIRT_LAZY_CPU);
 374
 375         /*
 376          * XXX sleazy hack: If we're being called in a lazy-cpu zone,
 377          * it means we're in a context switch, and %gs has just been
 378          * saved.  This means we can zero it out to prevent faults on
 379          * exit from the hypervisor if the next process has no %gs.
 380          * Either way, it has been saved, and the new value will get
 381          * loaded properly.  This will go away as soon as Xen has been
 382          * modified to not save/restore %gs for normal hypercalls.
 383          */
 384         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU)
 385                 loadsegment(gs, 0);
 386 }
 387
 388 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
 389                                 const void *ptr)
 390 {
 391         unsigned long lp = (unsigned long)&dt[entrynum];
 392         xmaddr_t mach_lp = virt_to_machine(lp);
 393         u64 entry = *(u64 *)ptr;
 394
 395         preempt_disable();
 396
 397         xen_mc_flush();
 398         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
 399                 BUG();
 400
 401         preempt_enable();
 402 }
 403
 404 static int cvt_gate_to_trap(int vector, u32 low, u32 high,
 405                             struct trap_info *info)
 406 {
 407         u8 type, dpl;
 408
 409         type = (high >> 8) & 0x1f;
 410         dpl = (high >> 13) & 3;
 411
 412         if (type != 0xf && type != 0xe)
 413                 return 0;
 414
 415         info->vector = vector;
 416         info->address = (high & 0xffff0000) | (low & 0x0000ffff);
 417         info->cs = low >> 16;
 418         info->flags = dpl;
 419         /* interrupt gates clear IF */
 420         if (type == 0xe)
 421                 info->flags |= 4;
 422
 423         return 1;
 424 }
 425
 426 /* Locations of each CPU's IDT */
 427 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
 428
 429 /* Set an IDT entry.  If the entry is part of the current IDT, then
 430    also update Xen. */
 431 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
 432 {
 433         unsigned long p = (unsigned long)&dt[entrynum];
 434         unsigned long start, end;
 435
 436         preempt_disable();
 437
 438         start = __get_cpu_var(idt_desc).address;
 439         end = start + __get_cpu_var(idt_desc).size + 1;
 440
 441         xen_mc_flush();
 442
 443         native_write_idt_entry(dt, entrynum, g);
 444
 445         if (p >= start && (p + 8) <= end) {
 446                 struct trap_info info[2];
 447                 u32 *desc = (u32 *)g;
 448
 449                 info[1].address = 0;
 450
 451                 if (cvt_gate_to_trap(entrynum, desc[0], desc[1], &info[0]))
 452                         if (HYPERVISOR_set_trap_table(info))
 453                                 BUG();
 454         }
 455
 456         preempt_enable();
 457 }
 458
 459 static void xen_convert_trap_info(const struct desc_ptr *desc,
 460                                   struct trap_info *traps)
 461 {
 462         unsigned in, out, count;
 463
 464         count = (desc->size+1) / 8;
 465         BUG_ON(count > 256);
 466
 467         for (in = out = 0; in < count; in++) {
 468                 const u32 *entry = (u32 *)(desc->address + in * 8);
 469
 470                 if (cvt_gate_to_trap(in, entry[0], entry[1], &traps[out]))
 471                         out++;
 472         }
 473         traps[out].address = 0;
 474 }
 475
 476 void xen_copy_trap_info(struct trap_info *traps)
 477 {
 478         const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
 479
 480         xen_convert_trap_info(desc, traps);
 481 }
 482
 483 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
 484    hold a spinlock to protect the static traps[] array (static because
 485    it avoids allocation, and saves stack space). */
 486 static void xen_load_idt(const struct desc_ptr *desc)
 487 {
 488         static DEFINE_SPINLOCK(lock);
 489         static struct trap_info traps[257];
 490
 491         spin_lock(&lock);
 492
 493         __get_cpu_var(idt_desc) = *desc;
 494
 495         xen_convert_trap_info(desc, traps);
 496
 497         xen_mc_flush();
 498         if (HYPERVISOR_set_trap_table(traps))
 499                 BUG();
 500
 501         spin_unlock(&lock);
 502 }
 503
 504 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
 505    they're handled differently. */
 506 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
 507                                 const void *desc, int type)
 508 {
 509         preempt_disable();
 510
 511         switch (type) {
 512         case DESC_LDT:
 513         case DESC_TSS:
 514                 /* ignore */
 515                 break;
 516
 517         default: {
 518                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
 519
 520                 xen_mc_flush();
 521                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
 522                         BUG();
 523         }
 524
 525         }
 526
 527         preempt_enable();
 528 }
 529
 530 static void xen_load_sp0(struct tss_struct *tss,
 531                           struct thread_struct *thread)
 532 {
 533         struct multicall_space mcs = xen_mc_entry(0);
 534         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
 535         xen_mc_issue(PARAVIRT_LAZY_CPU);
 536 }
 537
 538 static void xen_set_iopl_mask(unsigned mask)
 539 {
 540         struct physdev_set_iopl set_iopl;
 541
 542         /* Force the change at ring 0. */
 543         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
 544         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
 545 }
 546
 547 static void xen_io_delay(void)
 548 {
 549 }
 550
 551 #ifdef CONFIG_X86_LOCAL_APIC
 552 static u32 xen_apic_read(unsigned long reg)
 553 {
 554         return 0;
 555 }
 556
 557 static void xen_apic_write(unsigned long reg, u32 val)
 558 {
 559         /* Warn to see if there's any stray references */
 560         WARN_ON(1);
 561 }
 562 #endif
 563
 564 static void xen_flush_tlb(void)
 565 {
 566         struct mmuext_op *op;
 567         struct multicall_space mcs;
 568
 569         preempt_disable();
 570
 571         mcs = xen_mc_entry(sizeof(*op));
 572
 573         op = mcs.args;
 574         op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
 575         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 576
 577         xen_mc_issue(PARAVIRT_LAZY_MMU);
 578
 579         preempt_enable();
 580 }
 581
 582 static void xen_flush_tlb_single(unsigned long addr)
 583 {
 584         struct mmuext_op *op;
 585         struct multicall_space mcs;
 586
 587         preempt_disable();
 588
 589         mcs = xen_mc_entry(sizeof(*op));
 590         op = mcs.args;
 591         op->cmd = MMUEXT_INVLPG_LOCAL;
 592         op->arg1.linear_addr = addr & PAGE_MASK;
 593         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 594
 595         xen_mc_issue(PARAVIRT_LAZY_MMU);
 596
 597         preempt_enable();
 598 }
 599
 600 static void xen_flush_tlb_others(const cpumask_t *cpus, struct mm_struct *mm,
 601                                  unsigned long va)
 602 {
 603         struct {
 604                 struct mmuext_op op;
 605                 cpumask_t mask;
 606         } *args;
 607         cpumask_t cpumask = *cpus;
 608         struct multicall_space mcs;
 609
 610         /*
 611          * A couple of (to be removed) sanity checks:
 612          *
 613          * - current CPU must not be in mask
 614          * - mask must exist :)
 615          */
 616         BUG_ON(cpus_empty(cpumask));
 617         BUG_ON(cpu_isset(smp_processor_id(), cpumask));
 618         BUG_ON(!mm);
 619
 620         /* If a CPU which we ran on has gone down, OK. */
 621         cpus_and(cpumask, cpumask, cpu_online_map);
 622         if (cpus_empty(cpumask))
 623                 return;
 624
 625         mcs = xen_mc_entry(sizeof(*args));
 626         args = mcs.args;
 627         args->mask = cpumask;
 628         args->op.arg2.vcpumask = &args->mask;
 629
 630         if (va == TLB_FLUSH_ALL) {
 631                 args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
 632         } else {
 633                 args->op.cmd = MMUEXT_INVLPG_MULTI;
 634                 args->op.arg1.linear_addr = va;
 635         }
 636
 637         MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
 638
 639         xen_mc_issue(PARAVIRT_LAZY_MMU);
 640 }
 641
 642 static void xen_clts(void)
 643 {
 644         struct multicall_space mcs;
 645
 646         mcs = xen_mc_entry(0);
 647
 648         MULTI_fpu_taskswitch(mcs.mc, 0);
 649
 650         xen_mc_issue(PARAVIRT_LAZY_CPU);
 651 }
 652
 653 static void xen_write_cr0(unsigned long cr0)
 654 {
 655         struct multicall_space mcs;
 656
 657         /* Only pay attention to cr0.TS; everything else is
 658            ignored. */
 659         mcs = xen_mc_entry(0);
 660
 661         MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
 662
 663         xen_mc_issue(PARAVIRT_LAZY_CPU);
 664 }
 665
 666 static void xen_write_cr2(unsigned long cr2)
 667 {
 668         x86_read_percpu(xen_vcpu)->arch.cr2 = cr2;
 669 }
 670
 671 static unsigned long xen_read_cr2(void)
 672 {
 673         return x86_read_percpu(xen_vcpu)->arch.cr2;
 674 }
 675
 676 static unsigned long xen_read_cr2_direct(void)
 677 {
 678         return x86_read_percpu(xen_vcpu_info.arch.cr2);
 679 }
 680
 681 static void xen_write_cr4(unsigned long cr4)
 682 {
 683         cr4 &= ~X86_CR4_PGE;
 684         cr4 &= ~X86_CR4_PSE;
 685
 686         native_write_cr4(cr4);
 687 }
 688
 689 static unsigned long xen_read_cr3(void)
 690 {
 691         return x86_read_percpu(xen_cr3);
 692 }
 693
 694 static void set_current_cr3(void *v)
 695 {
 696         x86_write_percpu(xen_current_cr3, (unsigned long)v);
 697 }
 698
 699 static void xen_write_cr3(unsigned long cr3)
 700 {
 701         struct mmuext_op *op;
 702         struct multicall_space mcs;
 703         unsigned long mfn = pfn_to_mfn(PFN_DOWN(cr3));
 704
 705         BUG_ON(preemptible());
 706
 707         mcs = xen_mc_entry(sizeof(*op));  /* disables interrupts */
 708
 709         /* Update while interrupts are disabled, so its atomic with
 710            respect to ipis */
 711         x86_write_percpu(xen_cr3, cr3);
 712
 713         op = mcs.args;
 714         op->cmd = MMUEXT_NEW_BASEPTR;
 715         op->arg1.mfn = mfn;
 716
 717         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 718
 719         /* Update xen_update_cr3 once the batch has actually
 720            been submitted. */
 721         xen_mc_callback(set_current_cr3, (void *)cr3);
 722
 723         xen_mc_issue(PARAVIRT_LAZY_CPU);  /* interrupts restored */
 724 }
 725
 726 /* Early in boot, while setting up the initial pagetable, assume
 727    everything is pinned. */
 728 static __init void xen_alloc_pte_init(struct mm_struct *mm, u32 pfn)
 729 {
 730 #ifdef CONFIG_FLATMEM
 731         BUG_ON(mem_map);        /* should only be used early */
 732 #endif
 733         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 734 }
 735
 736 /* Early release_pte assumes that all pts are pinned, since there's
 737    only init_mm and anything attached to that is pinned. */
 738 static void xen_release_pte_init(u32 pfn)
 739 {
 740         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 741 }
 742
 743 static void pin_pagetable_pfn(unsigned cmd, unsigned long pfn)
 744 {
 745         struct mmuext_op op;
 746         op.cmd = cmd;
 747         op.arg1.mfn = pfn_to_mfn(pfn);
 748         if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
 749                 BUG();
 750 }
 751
 752 /* This needs to make sure the new pte page is pinned iff its being
 753    attached to a pinned pagetable. */
 754 static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 755 {
 756         struct page *page = pfn_to_page(pfn);
 757
 758         if (PagePinned(virt_to_page(mm->pgd))) {
 759                 SetPagePinned(page);
 760
 761                 if (!PageHighMem(page)) {
 762                         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 763                         if (level == PT_PTE)
 764                                 pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
 765                 } else
 766                         /* make sure there are no stray mappings of
 767                            this page */
 768                         kmap_flush_unused();
 769         }
 770 }
 771
 772 static void xen_alloc_pte(struct mm_struct *mm, u32 pfn)
 773 {
 774         xen_alloc_ptpage(mm, pfn, PT_PTE);
 775 }
 776
 777 static void xen_alloc_pmd(struct mm_struct *mm, u32 pfn)
 778 {
 779         xen_alloc_ptpage(mm, pfn, PT_PMD);
 780 }
 781
 782 /* This should never happen until we're OK to use struct page */
 783 static void xen_release_ptpage(u32 pfn, unsigned level)
 784 {
 785         struct page *page = pfn_to_page(pfn);
 786
 787         if (PagePinned(page)) {
 788                 if (!PageHighMem(page)) {
 789                         if (level == PT_PTE)
 790                                 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
 791                         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 792                 }
 793                 ClearPagePinned(page);
 794         }
 795 }
 796
 797 static void xen_release_pte(u32 pfn)
 798 {
 799         xen_release_ptpage(pfn, PT_PTE);
 800 }
 801
 802 static void xen_release_pmd(u32 pfn)
 803 {
 804         xen_release_ptpage(pfn, PT_PMD);
 805 }
 806
 807 #if PAGETABLE_LEVELS == 4
 808 static void xen_alloc_pud(struct mm_struct *mm, u32 pfn)
 809 {
 810         xen_alloc_ptpage(mm, pfn, PT_PUD);
 811 }
 812
 813 static void xen_release_pud(u32 pfn)
 814 {
 815         xen_release_ptpage(pfn, PT_PUD);
 816 }
 817 #endif
 818
 819 #ifdef CONFIG_HIGHPTE
 820 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
 821 {
 822         pgprot_t prot = PAGE_KERNEL;
 823
 824         if (PagePinned(page))
 825                 prot = PAGE_KERNEL_RO;
 826
 827         if (0 && PageHighMem(page))
 828                 printk("mapping highpte %lx type %d prot %s\n",
 829                        page_to_pfn(page), type,
 830                        (unsigned long)pgprot_val(prot) & _PAGE_RW ? "WRITE" : "READ");
 831
 832         return kmap_atomic_prot(page, type, prot);
 833 }
 834 #endif
 835
 836 static __init pte_t mask_rw_pte(pte_t *ptep, pte_t pte)
 837 {
 838         /* If there's an existing pte, then don't allow _PAGE_RW to be set */
 839         if (pte_val_ma(*ptep) & _PAGE_PRESENT)
 840                 pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
 841                                pte_val_ma(pte));
 842
 843         return pte;
 844 }
 845
 846 /* Init-time set_pte while constructing initial pagetables, which
 847    doesn't allow RO pagetable pages to be remapped RW */
 848 static __init void xen_set_pte_init(pte_t *ptep, pte_t pte)
 849 {
 850         pte = mask_rw_pte(ptep, pte);
 851
 852         xen_set_pte(ptep, pte);
 853 }
 854
 855 static __init void xen_pagetable_setup_start(pgd_t *base)
 856 {
 857 }
 858
 859 void xen_setup_shared_info(void)
 860 {
 861         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 862                 set_fixmap(FIX_PARAVIRT_BOOTMAP,
 863                            xen_start_info->shared_info);
 864
 865                 HYPERVISOR_shared_info =
 866                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
 867         } else
 868                 HYPERVISOR_shared_info =
 869                         (struct shared_info *)__va(xen_start_info->shared_info);
 870
 871 #ifndef CONFIG_SMP
 872         /* In UP this is as good a place as any to set up shared info */
 873         xen_setup_vcpu_info_placement();
 874 #endif
 875
 876         xen_setup_mfn_list_list();
 877 }
 878
 879 static __init void xen_pagetable_setup_done(pgd_t *base)
 880 {
 881         xen_setup_shared_info();
 882 }
 883
 884 static __init void xen_post_allocator_init(void)
 885 {
 886         pv_mmu_ops.set_pte = xen_set_pte;
 887         pv_mmu_ops.set_pmd = xen_set_pmd;
 888         pv_mmu_ops.set_pud = xen_set_pud;
 889 #if PAGETABLE_LEVELS == 4
 890         pv_mmu_ops.set_pgd = xen_set_pgd;
 891 #endif
 892
 893         /* This will work as long as patching hasn't happened yet
 894            (which it hasn't) */
 895         pv_mmu_ops.alloc_pte = xen_alloc_pte;
 896         pv_mmu_ops.alloc_pmd = xen_alloc_pmd;
 897         pv_mmu_ops.release_pte = xen_release_pte;
 898         pv_mmu_ops.release_pmd = xen_release_pmd;
 899 #if PAGETABLE_LEVELS == 4
 900         pv_mmu_ops.alloc_pud = xen_alloc_pud;
 901         pv_mmu_ops.release_pud = xen_release_pud;
 902 #endif
 903
 904         xen_mark_init_mm_pinned();
 905 }
 906
 907 /* This is called once we have the cpu_possible_map */
 908 void xen_setup_vcpu_info_placement(void)
 909 {
 910         int cpu;
 911
 912         for_each_possible_cpu(cpu)
 913                 xen_vcpu_setup(cpu);
 914
 915         /* xen_vcpu_setup managed to place the vcpu_info within the
 916            percpu area for all cpus, so make use of it */
 917 #ifdef CONFIG_X86_32
 918         if (have_vcpu_info_placement) {
 919                 printk(KERN_INFO "Xen: using vcpu_info placement\n");
 920
 921                 pv_irq_ops.save_fl = xen_save_fl_direct;
 922                 pv_irq_ops.restore_fl = xen_restore_fl_direct;
 923                 pv_irq_ops.irq_disable = xen_irq_disable_direct;
 924                 pv_irq_ops.irq_enable = xen_irq_enable_direct;
 925                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
 926         }
 927 #endif
 928 }
 929
 930 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
 931                           unsigned long addr, unsigned len)
 932 {
 933         char *start, *end, *reloc;
 934         unsigned ret;
 935
 936         start = end = reloc = NULL;
 937
 938 #define SITE(op, x)                                                     \
 939         case PARAVIRT_PATCH(op.x):                                      \
 940         if (have_vcpu_info_placement) {                                 \
 941                 start = (char *)xen_##x##_direct;                       \
 942                 end = xen_##x##_direct_end;                             \
 943                 reloc = xen_##x##_direct_reloc;                         \
 944         }                                                               \
 945         goto patch_site
 946
 947         switch (type) {
 948 #ifdef CONFIG_X86_32
 949                 SITE(pv_irq_ops, irq_enable);
 950                 SITE(pv_irq_ops, irq_disable);
 951                 SITE(pv_irq_ops, save_fl);
 952                 SITE(pv_irq_ops, restore_fl);
 953 #endif /* CONFIG_X86_32 */
 954 #undef SITE
 955
 956         patch_site:
 957                 if (start == NULL || (end-start) > len)
 958                         goto default_patch;
 959
 960                 ret = paravirt_patch_insns(insnbuf, len, start, end);
 961
 962                 /* Note: because reloc is assigned from something that
 963                    appears to be an array, gcc assumes it's non-null,
 964                    but doesn't know its relationship with start and
 965                    end. */
 966                 if (reloc > start && reloc < end) {
 967                         int reloc_off = reloc - start;
 968                         long *relocp = (long *)(insnbuf + reloc_off);
 969                         long delta = start - (char *)addr;
 970
 971                         *relocp += delta;
 972                 }
 973                 break;
 974
 975         default_patch:
 976         default:
 977                 ret = paravirt_patch_default(type, clobbers, insnbuf,
 978                                              addr, len);
 979                 break;
 980         }
 981
 982         return ret;
 983 }
 984
 985 static void xen_set_fixmap(unsigned idx, unsigned long phys, pgprot_t prot)
 986 {
 987         pte_t pte;
 988
 989         phys >>= PAGE_SHIFT;
 990
 991         switch (idx) {
 992         case FIX_BTMAP_END ... FIX_BTMAP_BEGIN:
 993 #ifdef CONFIG_X86_F00F_BUG
 994         case FIX_F00F_IDT:
 995 #endif
 996 #ifdef CONFIG_X86_32
 997         case FIX_WP_TEST:
 998         case FIX_VDSO:
 999         case FIX_KMAP_BEGIN ... FIX_KMAP_END:
1000 #else
1001         case VSYSCALL_LAST_PAGE ... VSYSCALL_FIRST_PAGE:
1002 #endif
1003 #ifdef CONFIG_X86_LOCAL_APIC
1004         case FIX_APIC_BASE:     /* maps dummy local APIC */
1005 #endif
1006                 pte = pfn_pte(phys, prot);
1007                 break;
1008
1009         default:
1010                 pte = mfn_pte(phys, prot);
1011                 break;
1012         }
1013
1014         __native_set_fixmap(idx, pte);
1015 }
1016
1017 static const struct pv_info xen_info __initdata = {
1018         .paravirt_enabled = 1,
1019         .shared_kernel_pmd = 0,
1020
1021         .name = "Xen",
1022 };
1023
1024 static const struct pv_init_ops xen_init_ops __initdata = {
1025         .patch = xen_patch,
1026
1027         .banner = xen_banner,
1028         .memory_setup = xen_memory_setup,
1029         .arch_setup = xen_arch_setup,
1030         .post_allocator_init = xen_post_allocator_init,
1031 };
1032
1033 static const struct pv_time_ops xen_time_ops __initdata = {
1034         .time_init = xen_time_init,
1035
1036         .set_wallclock = xen_set_wallclock,
1037         .get_wallclock = xen_get_wallclock,
1038         .get_tsc_khz = xen_tsc_khz,
1039         .sched_clock = xen_sched_clock,
1040 };
1041
1042 static const struct pv_cpu_ops xen_cpu_ops __initdata = {
1043         .cpuid = xen_cpuid,
1044
1045         .set_debugreg = xen_set_debugreg,
1046         .get_debugreg = xen_get_debugreg,
1047
1048         .clts = xen_clts,
1049
1050         .read_cr0 = native_read_cr0,
1051         .write_cr0 = xen_write_cr0,
1052
1053         .read_cr4 = native_read_cr4,
1054         .read_cr4_safe = native_read_cr4_safe,
1055         .write_cr4 = xen_write_cr4,
1056
1057         .wbinvd = native_wbinvd,
1058
1059         .read_msr = native_read_msr_safe,
1060         .write_msr = native_write_msr_safe,
1061         .read_tsc = native_read_tsc,
1062         .read_pmc = native_read_pmc,
1063
1064         .iret = xen_iret,
1065         .irq_enable_sysexit = xen_sysexit,
1066
1067         .load_tr_desc = paravirt_nop,
1068         .set_ldt = xen_set_ldt,
1069         .load_gdt = xen_load_gdt,
1070         .load_idt = xen_load_idt,
1071         .load_tls = xen_load_tls,
1072
1073         .store_gdt = native_store_gdt,
1074         .store_idt = native_store_idt,
1075         .store_tr = xen_store_tr,
1076
1077         .write_ldt_entry = xen_write_ldt_entry,
1078         .write_gdt_entry = xen_write_gdt_entry,
1079         .write_idt_entry = xen_write_idt_entry,
1080         .load_sp0 = xen_load_sp0,
1081
1082         .set_iopl_mask = xen_set_iopl_mask,
1083         .io_delay = xen_io_delay,
1084
1085         .lazy_mode = {
1086                 .enter = paravirt_enter_lazy_cpu,
1087                 .leave = xen_leave_lazy,
1088         },
1089 };
1090
1091 static const struct pv_irq_ops xen_irq_ops __initdata = {
1092         .init_IRQ = xen_init_IRQ,
1093         .save_fl = xen_save_fl,
1094         .restore_fl = xen_restore_fl,
1095         .irq_disable = xen_irq_disable,
1096         .irq_enable = xen_irq_enable,
1097         .safe_halt = xen_safe_halt,
1098         .halt = xen_halt,
1099 #ifdef CONFIG_X86_64
1100         .adjust_exception_frame = paravirt_nop,
1101 #endif
1102 };
1103
1104 static const struct pv_apic_ops xen_apic_ops __initdata = {
1105 #ifdef CONFIG_X86_LOCAL_APIC
1106         .apic_write = xen_apic_write,
1107         .apic_write_atomic = xen_apic_write,
1108         .apic_read = xen_apic_read,
1109         .setup_boot_clock = paravirt_nop,
1110         .setup_secondary_clock = paravirt_nop,
1111         .startup_ipi_hook = paravirt_nop,
1112 #endif
1113 };
1114
1115 static const struct pv_mmu_ops xen_mmu_ops __initdata = {
1116         .pagetable_setup_start = xen_pagetable_setup_start,
1117         .pagetable_setup_done = xen_pagetable_setup_done,
1118
1119         .read_cr2 = xen_read_cr2,
1120         .write_cr2 = xen_write_cr2,
1121
1122         .read_cr3 = xen_read_cr3,
1123         .write_cr3 = xen_write_cr3,
1124
1125         .flush_tlb_user = xen_flush_tlb,
1126         .flush_tlb_kernel = xen_flush_tlb,
1127         .flush_tlb_single = xen_flush_tlb_single,
1128         .flush_tlb_others = xen_flush_tlb_others,
1129
1130         .pte_update = paravirt_nop,
1131         .pte_update_defer = paravirt_nop,
1132
1133         .pgd_alloc = __paravirt_pgd_alloc,
1134         .pgd_free = paravirt_nop,
1135
1136         .alloc_pte = xen_alloc_pte_init,
1137         .release_pte = xen_release_pte_init,
1138         .alloc_pmd = xen_alloc_pte_init,
1139         .alloc_pmd_clone = paravirt_nop,
1140         .release_pmd = xen_release_pte_init,
1141
1142 #ifdef CONFIG_HIGHPTE
1143         .kmap_atomic_pte = xen_kmap_atomic_pte,
1144 #endif
1145
1146 #ifdef CONFIG_X86_64
1147         .set_pte = xen_set_pte,
1148 #else
1149         .set_pte = xen_set_pte_init,
1150 #endif
1151         .set_pte_at = xen_set_pte_at,
1152         .set_pmd = xen_set_pmd_hyper,
1153
1154         .ptep_modify_prot_start = __ptep_modify_prot_start,
1155         .ptep_modify_prot_commit = __ptep_modify_prot_commit,
1156
1157         .pte_val = xen_pte_val,
1158         .pte_flags = native_pte_val,
1159         .pgd_val = xen_pgd_val,
1160
1161         .make_pte = xen_make_pte,
1162         .make_pgd = xen_make_pgd,
1163
1164 #ifdef CONFIG_X86_PAE
1165         .set_pte_atomic = xen_set_pte_atomic,
1166         .set_pte_present = xen_set_pte_at,
1167         .pte_clear = xen_pte_clear,
1168         .pmd_clear = xen_pmd_clear,
1169 #endif  /* CONFIG_X86_PAE */
1170         .set_pud = xen_set_pud_hyper,
1171
1172         .make_pmd = xen_make_pmd,
1173         .pmd_val = xen_pmd_val,
1174
1175 #if PAGETABLE_LEVELS == 4
1176         .pud_val = xen_pud_val,
1177         .make_pud = xen_make_pud,
1178         .set_pgd = xen_set_pgd_hyper,
1179
1180         .alloc_pud = xen_alloc_pte_init,
1181         .release_pud = xen_release_pte_init,
1182 #endif  /* PAGETABLE_LEVELS == 4 */
1183
1184         .activate_mm = xen_activate_mm,
1185         .dup_mmap = xen_dup_mmap,
1186         .exit_mmap = xen_exit_mmap,
1187
1188         .lazy_mode = {
1189                 .enter = paravirt_enter_lazy_mmu,
1190                 .leave = xen_leave_lazy,
1191         },
1192
1193         .set_fixmap = xen_set_fixmap,
1194 };
1195
1196 static void xen_reboot(int reason)
1197 {
1198         struct sched_shutdown r = { .reason = reason };
1199
1200 #ifdef CONFIG_SMP
1201         smp_send_stop();
1202 #endif
1203
1204         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1205                 BUG();
1206 }
1207
1208 static void xen_restart(char *msg)
1209 {
1210         xen_reboot(SHUTDOWN_reboot);
1211 }
1212
1213 static void xen_emergency_restart(void)
1214 {
1215         xen_reboot(SHUTDOWN_reboot);
1216 }
1217
1218 static void xen_machine_halt(void)
1219 {
1220         xen_reboot(SHUTDOWN_poweroff);
1221 }
1222
1223 static void xen_crash_shutdown(struct pt_regs *regs)
1224 {
1225         xen_reboot(SHUTDOWN_crash);
1226 }
1227
1228 static const struct machine_ops __initdata xen_machine_ops = {
1229         .restart = xen_restart,
1230         .halt = xen_machine_halt,
1231         .power_off = xen_machine_halt,
1232         .shutdown = xen_machine_halt,
1233         .crash_shutdown = xen_crash_shutdown,
1234         .emergency_restart = xen_emergency_restart,
1235 };
1236
1237
1238 static void __init xen_reserve_top(void)
1239 {
1240 #ifdef CONFIG_X86_32
1241         unsigned long top = HYPERVISOR_VIRT_START;
1242         struct xen_platform_parameters pp;
1243
1244         if (HYPERVISOR_xen_version(XENVER_platform_parameters, &pp) == 0)
1245                 top = pp.virt_start;
1246
1247         reserve_top_address(-top + 2 * PAGE_SIZE);
1248 #endif  /* CONFIG_X86_32 */
1249 }
1250
1251 /*
1252  * Like __va(), but returns address in the kernel mapping (which is
1253  * all we have until the physical memory mapping has been set up.
1254  */
1255 static void *__ka(phys_addr_t paddr)
1256 {
1257 #ifdef CONFIG_X86_64
1258         return (void *)(paddr + __START_KERNEL_map);
1259 #else
1260         return __va(paddr);
1261 #endif
1262 }
1263
1264 /* Convert a machine address to physical address */
1265 static unsigned long m2p(phys_addr_t maddr)
1266 {
1267         phys_addr_t paddr;
1268
1269         maddr &= PTE_MASK;
1270         paddr = mfn_to_pfn(maddr >> PAGE_SHIFT) << PAGE_SHIFT;
1271
1272         return paddr;
1273 }
1274
1275 /* Convert a machine address to kernel virtual */
1276 static void *m2v(phys_addr_t maddr)
1277 {
1278         return __ka(m2p(maddr));
1279 }
1280
1281 #ifdef CONFIG_X86_64
1282 static void walk(pgd_t *pgd, unsigned long addr)
1283 {
1284         unsigned l4idx = pgd_index(addr);
1285         unsigned l3idx = pud_index(addr);
1286         unsigned l2idx = pmd_index(addr);
1287         unsigned l1idx = pte_index(addr);
1288         pgd_t l4;
1289         pud_t l3;
1290         pmd_t l2;
1291         pte_t l1;
1292
1293         xen_raw_printk("walk %p, %lx -> %d %d %d %d\n",
1294                        pgd, addr, l4idx, l3idx, l2idx, l1idx);
1295
1296         l4 = pgd[l4idx];
1297         xen_raw_printk("  l4: %016lx\n", l4.pgd);
1298         xen_raw_printk("      %016lx\n", pgd_val(l4));
1299
1300         l3 = ((pud_t *)(m2v(l4.pgd)))[l3idx];
1301         xen_raw_printk("  l3: %016lx\n", l3.pud);
1302         xen_raw_printk("      %016lx\n", pud_val(l3));
1303
1304         l2 = ((pmd_t *)(m2v(l3.pud)))[l2idx];
1305         xen_raw_printk("  l2: %016lx\n", l2.pmd);
1306         xen_raw_printk("      %016lx\n", pmd_val(l2));
1307
1308         l1 = ((pte_t *)(m2v(l2.pmd)))[l1idx];
1309         xen_raw_printk("  l1: %016lx\n", l1.pte);
1310         xen_raw_printk("      %016lx\n", pte_val(l1));
1311 }
1312 #endif
1313
1314 static void set_page_prot(void *addr, pgprot_t prot)
1315 {
1316         unsigned long pfn = __pa(addr) >> PAGE_SHIFT;
1317         pte_t pte = pfn_pte(pfn, prot);
1318
1319         xen_raw_printk("addr=%p pfn=%lx mfn=%lx prot=%016llx pte=%016llx\n",
1320                        addr, pfn, get_phys_to_machine(pfn),
1321                        pgprot_val(prot), pte.pte);
1322
1323         if (HYPERVISOR_update_va_mapping((unsigned long)addr, pte, 0))
1324                 BUG();
1325 }
1326
1327 /*
1328  * Identity map, in addition to plain kernel map.  This needs to be
1329  * large enough to allocate page table pages to allocate the rest.
1330  * Each page can map 2MB.
1331  */
1332 static pte_t level1_ident_pgt[PTRS_PER_PTE * 4] __page_aligned_bss;
1333
1334 static __init void xen_map_identity_early(pmd_t *pmd, unsigned long max_pfn)
1335 {
1336         unsigned pmdidx, pteidx;
1337         unsigned ident_pte;
1338         unsigned long pfn;
1339
1340         ident_pte = 0;
1341         pfn = 0;
1342         for(pmdidx = 0; pmdidx < PTRS_PER_PMD && pfn < max_pfn; pmdidx++) {
1343                 pte_t *pte_page;
1344
1345                 /* Reuse or allocate a page of ptes */
1346                 if (pmd_present(pmd[pmdidx]))
1347                         pte_page = m2v(pmd[pmdidx].pmd);
1348                 else {
1349                         /* Check for free pte pages */
1350                         if (ident_pte == ARRAY_SIZE(level1_ident_pgt))
1351                                 break;
1352
1353                         pte_page = &level1_ident_pgt[ident_pte];
1354                         ident_pte += PTRS_PER_PTE;
1355
1356                         pmd[pmdidx] = __pmd(__pa(pte_page) | _PAGE_TABLE);
1357                 }
1358
1359                 /* Install mappings */
1360                 for(pteidx = 0; pteidx < PTRS_PER_PTE; pteidx++, pfn++) {
1361                         pte_t pte;
1362
1363                         if (pfn > max_pfn_mapped)
1364                                 max_pfn_mapped = pfn;
1365
1366                         if (!pte_none(pte_page[pteidx]))
1367                                 continue;
1368
1369                         pte = pfn_pte(pfn, PAGE_KERNEL_EXEC);
1370                         pte_page[pteidx] = pte;
1371                 }
1372         }
1373
1374         for(pteidx = 0; pteidx < ident_pte; pteidx += PTRS_PER_PTE)
1375                 set_page_prot(&level1_ident_pgt[pteidx], PAGE_KERNEL_RO);
1376
1377         set_page_prot(pmd, PAGE_KERNEL_RO);
1378 }
1379
1380 #ifdef CONFIG_X86_64
1381 static void convert_pfn_mfn(void *v)
1382 {
1383         pte_t *pte = v;
1384         int i;
1385
1386         /* All levels are converted the same way, so just treat them
1387            as ptes. */
1388         for(i = 0; i < PTRS_PER_PTE; i++)
1389                 pte[i] = xen_make_pte(pte[i].pte);
1390 }
1391
1392 /*
1393  * Set up the inital kernel pagetable.
1394  *
1395  * We can construct this by grafting the Xen provided pagetable into
1396  * head_64.S's preconstructed pagetables.  We copy the Xen L2's into
1397  * level2_ident_pgt, level2_kernel_pgt and level2_fixmap_pgt.  This
1398  * means that only the kernel has a physical mapping to start with -
1399  * but that's enough to get __va working.  We need to fill in the rest
1400  * of the physical mapping once some sort of allocator has been set
1401  * up.
1402  */
1403 static __init pgd_t *xen_setup_kernel_pagetable(pgd_t *pgd, unsigned long max_pfn)
1404 {
1405         pud_t *l3;
1406         pmd_t *l2;
1407
1408         /* Zap identity mapping */
1409         init_level4_pgt[0] = __pgd(0);
1410
1411         /* Pre-constructed entries are in pfn, so convert to mfn */
1412         convert_pfn_mfn(init_level4_pgt);
1413         convert_pfn_mfn(level3_ident_pgt);
1414         convert_pfn_mfn(level3_kernel_pgt);
1415
1416         l3 = m2v(pgd[pgd_index(__START_KERNEL_map)].pgd);
1417         l2 = m2v(l3[pud_index(__START_KERNEL_map)].pud);
1418
1419         memcpy(level2_ident_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1420         memcpy(level2_kernel_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1421
1422         l3 = m2v(pgd[pgd_index(__START_KERNEL_map + PMD_SIZE)].pgd);
1423         l2 = m2v(l3[pud_index(__START_KERNEL_map + PMD_SIZE)].pud);
1424         memcpy(level2_fixmap_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1425
1426         /* Set up identity map */
1427         xen_map_identity_early(level2_ident_pgt, max_pfn);
1428
1429         /* Make pagetable pieces RO */
1430         set_page_prot(init_level4_pgt, PAGE_KERNEL_RO);
1431         set_page_prot(level3_ident_pgt, PAGE_KERNEL_RO);
1432         set_page_prot(level3_kernel_pgt, PAGE_KERNEL_RO);
1433         set_page_prot(level2_kernel_pgt, PAGE_KERNEL_RO);
1434         set_page_prot(level2_fixmap_pgt, PAGE_KERNEL_RO);
1435
1436         /* Pin down new L4 */
1437         pin_pagetable_pfn(MMUEXT_PIN_L4_TABLE,
1438                           PFN_DOWN(__pa_symbol(init_level4_pgt)));
1439
1440         /* Unpin Xen-provided one */
1441         pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1442
1443         /* Switch over */
1444         pgd = init_level4_pgt;
1445         xen_write_cr3(__pa(pgd));
1446
1447         reserve_early(__pa(xen_start_info->pt_base),
1448                       __pa(xen_start_info->pt_base +
1449                            xen_start_info->nr_pt_frames * PAGE_SIZE),
1450                       "XEN PAGETABLES");
1451
1452         return pgd;
1453 }
1454 #else   /* !CONFIG_X86_64 */
1455 static pmd_t level2_kernel_pgt[PTRS_PER_PMD] __page_aligned_bss;
1456
1457 static __init pgd_t *xen_setup_kernel_pagetable(pgd_t *pgd, unsigned long max_pfn)
1458 {
1459         pmd_t *kernel_pmd;
1460
1461         init_pg_tables_start = __pa(pgd);
1462         init_pg_tables_end = __pa(pgd) + xen_start_info->nr_pt_frames*PAGE_SIZE;
1463         max_pfn_mapped = PFN_DOWN(init_pg_tables_end + 512*1024);
1464
1465         kernel_pmd = m2v(pgd[KERNEL_PGD_BOUNDARY].pgd);
1466         memcpy(level2_kernel_pgt, kernel_pmd, sizeof(pmd_t) * PTRS_PER_PMD);
1467
1468         xen_map_identity_early(level2_kernel_pgt, max_pfn);
1469
1470         memcpy(swapper_pg_dir, pgd, sizeof(pgd_t) * PTRS_PER_PGD);
1471         set_pgd(&swapper_pg_dir[KERNEL_PGD_BOUNDARY],
1472                         __pgd(__pa(level2_kernel_pgt) | _PAGE_PRESENT));
1473
1474         set_page_prot(level2_kernel_pgt, PAGE_KERNEL_RO);
1475         set_page_prot(swapper_pg_dir, PAGE_KERNEL_RO);
1476         set_page_prot(empty_zero_page, PAGE_KERNEL_RO);
1477
1478         pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1479
1480         xen_write_cr3(__pa(swapper_pg_dir));
1481
1482         pin_pagetable_pfn(MMUEXT_PIN_L3_TABLE, PFN_DOWN(__pa(swapper_pg_dir)));
1483
1484         return swapper_pg_dir;
1485 }
1486 #endif  /* CONFIG_X86_64 */
1487
1488 /* First C function to be called on Xen boot */
1489 asmlinkage void __init xen_start_kernel(void)
1490 {
1491         pgd_t *pgd;
1492
1493         if (!xen_start_info)
1494                 return;
1495
1496         BUG_ON(memcmp(xen_start_info->magic, "xen-3", 5) != 0);
1497
1498         xen_setup_features();
1499
1500         /* Install Xen paravirt ops */
1501         pv_info = xen_info;
1502         pv_init_ops = xen_init_ops;
1503         pv_time_ops = xen_time_ops;
1504         pv_cpu_ops = xen_cpu_ops;
1505         pv_irq_ops = xen_irq_ops;
1506         pv_apic_ops = xen_apic_ops;
1507         pv_mmu_ops = xen_mmu_ops;
1508
1509         if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1510                 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1511                 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1512         }
1513
1514         machine_ops = xen_machine_ops;
1515
1516 #ifdef CONFIG_X86_64
1517         /* Disable until direct per-cpu data access. */
1518         have_vcpu_info_placement = 0;
1519         x86_64_init_pda();
1520 #endif
1521
1522         xen_smp_init();
1523
1524         /* Get mfn list */
1525         if (!xen_feature(XENFEAT_auto_translated_physmap))
1526                 xen_build_dynamic_phys_to_machine();
1527
1528         pgd = (pgd_t *)xen_start_info->pt_base;
1529
1530         /* Prevent unwanted bits from being set in PTEs. */
1531         __supported_pte_mask &= ~_PAGE_GLOBAL;
1532         if (!is_initial_xendomain())
1533                 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1534
1535         /* Don't do the full vcpu_info placement stuff until we have a
1536            possible map and a non-dummy shared_info. */
1537         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1538
1539         xen_raw_console_write("mapping kernel into physical memory\n");
1540         pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages);
1541
1542         init_mm.pgd = pgd;
1543
1544         /* keep using Xen gdt for now; no urgent need to change it */
1545
1546         pv_info.kernel_rpl = 1;
1547         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1548                 pv_info.kernel_rpl = 0;
1549
1550         /* set the limit of our address space */
1551         xen_reserve_top();
1552
1553 #ifdef CONFIG_X86_32
1554         /* set up basic CPUID stuff */
1555         cpu_detect(&new_cpu_data);
1556         new_cpu_data.hard_math = 1;
1557         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1558 #endif
1559
1560         /* Poke various useful things into boot_params */
1561         boot_params.hdr.type_of_loader = (9 << 4) | 0;
1562         boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1563                 ? __pa(xen_start_info->mod_start) : 0;
1564         boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1565
1566         if (!is_initial_xendomain()) {
1567                 add_preferred_console("xenboot", 0, NULL);
1568                 add_preferred_console("tty", 0, NULL);
1569                 add_preferred_console("hvc", 0, NULL);
1570         }
1571
1572         xen_raw_console_write("about to get started...\n");
1573
1574 #if 0
1575         xen_raw_printk("&boot_params=%p __pa(&boot_params)=%lx __va(__pa(&boot_params))=%lx\n",
1576                        &boot_params, __pa_symbol(&boot_params),
1577                        __va(__pa_symbol(&boot_params)));
1578
1579         walk(pgd, &boot_params);
1580         walk(pgd, __va(__pa(&boot_params)));
1581 #endif
1582
1583         /* Start the world */
1584 #ifdef CONFIG_X86_32
1585         i386_start_kernel();
1586 #else
1587         x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1588 #endif
1589 }