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  71  * out of date, but it will never be set early).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 #ifdef CONFIG_X86_64
 389 static void xen_load_gs_index(unsigned int idx)
 390 {
 391         if (HYPERVISOR_set_segment_base(SEGBASE_GS_USER_SEL, idx))
 392                 BUG();
 393 }
 394 #endif
 395
 396 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
 397                                 const void *ptr)
 398 {
 399         unsigned long lp = (unsigned long)&dt[entrynum];
 400         xmaddr_t mach_lp = virt_to_machine(lp);
 401         u64 entry = *(u64 *)ptr;
 402
 403         preempt_disable();
 404
 405         xen_mc_flush();
 406         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
 407                 BUG();
 408
 409         preempt_enable();
 410 }
 411
 412 static int cvt_gate_to_trap(int vector, const gate_desc *val,
 413                             struct trap_info *info)
 414 {
 415         if (val->type != 0xf && val->type != 0xe)
 416                 return 0;
 417
 418         info->vector = vector;
 419         info->address = gate_offset(*val);
 420         info->cs = gate_segment(*val);
 421         info->flags = val->dpl;
 422         /* interrupt gates clear IF */
 423         if (val->type == 0xe)
 424                 info->flags |= 4;
 425
 426         return 1;
 427 }
 428
 429 /* Locations of each CPU's IDT */
 430 static DEFINE_PER_CPU(struct desc_ptr, idt_desc);
 431
 432 /* Set an IDT entry.  If the entry is part of the current IDT, then
 433    also update Xen. */
 434 static void xen_write_idt_entry(gate_desc *dt, int entrynum, const gate_desc *g)
 435 {
 436         unsigned long p = (unsigned long)&dt[entrynum];
 437         unsigned long start, end;
 438
 439         preempt_disable();
 440
 441         start = __get_cpu_var(idt_desc).address;
 442         end = start + __get_cpu_var(idt_desc).size + 1;
 443
 444         xen_mc_flush();
 445
 446         native_write_idt_entry(dt, entrynum, g);
 447
 448         if (p >= start && (p + 8) <= end) {
 449                 struct trap_info info[2];
 450
 451                 info[1].address = 0;
 452
 453                 if (cvt_gate_to_trap(entrynum, g, &info[0]))
 454                         if (HYPERVISOR_set_trap_table(info))
 455                                 BUG();
 456         }
 457
 458         preempt_enable();
 459 }
 460
 461 static void xen_convert_trap_info(const struct desc_ptr *desc,
 462                                   struct trap_info *traps)
 463 {
 464         unsigned in, out, count;
 465
 466         count = (desc->size+1) / sizeof(gate_desc);
 467         BUG_ON(count > 256);
 468
 469         for (in = out = 0; in < count; in++) {
 470                 gate_desc *entry = (gate_desc*)(desc->address) + in;
 471
 472                 if (cvt_gate_to_trap(in, entry, &traps[out]))
 473                         out++;
 474         }
 475         traps[out].address = 0;
 476 }
 477
 478 void xen_copy_trap_info(struct trap_info *traps)
 479 {
 480         const struct desc_ptr *desc = &__get_cpu_var(idt_desc);
 481
 482         xen_convert_trap_info(desc, traps);
 483 }
 484
 485 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
 486    hold a spinlock to protect the static traps[] array (static because
 487    it avoids allocation, and saves stack space). */
 488 static void xen_load_idt(const struct desc_ptr *desc)
 489 {
 490         static DEFINE_SPINLOCK(lock);
 491         static struct trap_info traps[257];
 492
 493         spin_lock(&lock);
 494
 495         __get_cpu_var(idt_desc) = *desc;
 496
 497         xen_convert_trap_info(desc, traps);
 498
 499         xen_mc_flush();
 500         if (HYPERVISOR_set_trap_table(traps))
 501                 BUG();
 502
 503         spin_unlock(&lock);
 504 }
 505
 506 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
 507    they're handled differently. */
 508 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
 509                                 const void *desc, int type)
 510 {
 511         preempt_disable();
 512
 513         switch (type) {
 514         case DESC_LDT:
 515         case DESC_TSS:
 516                 /* ignore */
 517                 break;
 518
 519         default: {
 520                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
 521
 522                 xen_mc_flush();
 523                 if (HYPERVISOR_update_descriptor(maddr.maddr, *(u64 *)desc))
 524                         BUG();
 525         }
 526
 527         }
 528
 529         preempt_enable();
 530 }
 531
 532 static void xen_load_sp0(struct tss_struct *tss,
 533                           struct thread_struct *thread)
 534 {
 535         struct multicall_space mcs = xen_mc_entry(0);
 536         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->sp0);
 537         xen_mc_issue(PARAVIRT_LAZY_CPU);
 538 }
 539
 540 static void xen_set_iopl_mask(unsigned mask)
 541 {
 542         struct physdev_set_iopl set_iopl;
 543
 544         /* Force the change at ring 0. */
 545         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
 546         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
 547 }
 548
 549 static void xen_io_delay(void)
 550 {
 551 }
 552
 553 #ifdef CONFIG_X86_LOCAL_APIC
 554 static u32 xen_apic_read(unsigned long reg)
 555 {
 556         return 0;
 557 }
 558
 559 static void xen_apic_write(unsigned long reg, u32 val)
 560 {
 561         /* Warn to see if there's any stray references */
 562         WARN_ON(1);
 563 }
 564 #endif
 565
 566 static void xen_flush_tlb(void)
 567 {
 568         struct mmuext_op *op;
 569         struct multicall_space mcs;
 570
 571         preempt_disable();
 572
 573         mcs = xen_mc_entry(sizeof(*op));
 574
 575         op = mcs.args;
 576         op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
 577         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 578
 579         xen_mc_issue(PARAVIRT_LAZY_MMU);
 580
 581         preempt_enable();
 582 }
 583
 584 static void xen_flush_tlb_single(unsigned long addr)
 585 {
 586         struct mmuext_op *op;
 587         struct multicall_space mcs;
 588
 589         preempt_disable();
 590
 591         mcs = xen_mc_entry(sizeof(*op));
 592         op = mcs.args;
 593         op->cmd = MMUEXT_INVLPG_LOCAL;
 594         op->arg1.linear_addr = addr & PAGE_MASK;
 595         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 596
 597         xen_mc_issue(PARAVIRT_LAZY_MMU);
 598
 599         preempt_enable();
 600 }
 601
 602 static void xen_flush_tlb_others(const cpumask_t *cpus, struct mm_struct *mm,
 603                                  unsigned long va)
 604 {
 605         struct {
 606                 struct mmuext_op op;
 607                 cpumask_t mask;
 608         } *args;
 609         cpumask_t cpumask = *cpus;
 610         struct multicall_space mcs;
 611
 612         /*
 613          * A couple of (to be removed) sanity checks:
 614          *
 615          * - current CPU must not be in mask
 616          * - mask must exist :)
 617          */
 618         BUG_ON(cpus_empty(cpumask));
 619         BUG_ON(cpu_isset(smp_processor_id(), cpumask));
 620         BUG_ON(!mm);
 621
 622         /* If a CPU which we ran on has gone down, OK. */
 623         cpus_and(cpumask, cpumask, cpu_online_map);
 624         if (cpus_empty(cpumask))
 625                 return;
 626
 627         mcs = xen_mc_entry(sizeof(*args));
 628         args = mcs.args;
 629         args->mask = cpumask;
 630         args->op.arg2.vcpumask = &args->mask;
 631
 632         if (va == TLB_FLUSH_ALL) {
 633                 args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
 634         } else {
 635                 args->op.cmd = MMUEXT_INVLPG_MULTI;
 636                 args->op.arg1.linear_addr = va;
 637         }
 638
 639         MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
 640
 641         xen_mc_issue(PARAVIRT_LAZY_MMU);
 642 }
 643
 644 static void xen_clts(void)
 645 {
 646         struct multicall_space mcs;
 647
 648         mcs = xen_mc_entry(0);
 649
 650         MULTI_fpu_taskswitch(mcs.mc, 0);
 651
 652         xen_mc_issue(PARAVIRT_LAZY_CPU);
 653 }
 654
 655 static void xen_write_cr0(unsigned long cr0)
 656 {
 657         struct multicall_space mcs;
 658
 659         /* Only pay attention to cr0.TS; everything else is
 660            ignored. */
 661         mcs = xen_mc_entry(0);
 662
 663         MULTI_fpu_taskswitch(mcs.mc, (cr0 & X86_CR0_TS) != 0);
 664
 665         xen_mc_issue(PARAVIRT_LAZY_CPU);
 666 }
 667
 668 static void xen_write_cr2(unsigned long cr2)
 669 {
 670         x86_read_percpu(xen_vcpu)->arch.cr2 = cr2;
 671 }
 672
 673 static unsigned long xen_read_cr2(void)
 674 {
 675         return x86_read_percpu(xen_vcpu)->arch.cr2;
 676 }
 677
 678 static unsigned long xen_read_cr2_direct(void)
 679 {
 680         return x86_read_percpu(xen_vcpu_info.arch.cr2);
 681 }
 682
 683 static void xen_write_cr4(unsigned long cr4)
 684 {
 685         cr4 &= ~X86_CR4_PGE;
 686         cr4 &= ~X86_CR4_PSE;
 687
 688         native_write_cr4(cr4);
 689 }
 690
 691 static unsigned long xen_read_cr3(void)
 692 {
 693         return x86_read_percpu(xen_cr3);
 694 }
 695
 696 static void set_current_cr3(void *v)
 697 {
 698         x86_write_percpu(xen_current_cr3, (unsigned long)v);
 699 }
 700
 701 static void xen_write_cr3(unsigned long cr3)
 702 {
 703         struct mmuext_op *op;
 704         struct multicall_space mcs;
 705         unsigned long mfn = pfn_to_mfn(PFN_DOWN(cr3));
 706
 707         BUG_ON(preemptible());
 708
 709         mcs = xen_mc_entry(sizeof(*op));  /* disables interrupts */
 710
 711         /* Update while interrupts are disabled, so its atomic with
 712            respect to ipis */
 713         x86_write_percpu(xen_cr3, cr3);
 714
 715         op = mcs.args;
 716         op->cmd = MMUEXT_NEW_BASEPTR;
 717         op->arg1.mfn = mfn;
 718
 719         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 720
 721         /* Update xen_update_cr3 once the batch has actually
 722            been submitted. */
 723         xen_mc_callback(set_current_cr3, (void *)cr3);
 724
 725         xen_mc_issue(PARAVIRT_LAZY_CPU);  /* interrupts restored */
 726 }
 727
 728 /* Early in boot, while setting up the initial pagetable, assume
 729    everything is pinned. */
 730 static __init void xen_alloc_pte_init(struct mm_struct *mm, u32 pfn)
 731 {
 732 #ifdef CONFIG_FLATMEM
 733         BUG_ON(mem_map);        /* should only be used early */
 734 #endif
 735         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 736 }
 737
 738 /* Early release_pte assumes that all pts are pinned, since there's
 739    only init_mm and anything attached to that is pinned. */
 740 static void xen_release_pte_init(u32 pfn)
 741 {
 742         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 743 }
 744
 745 static void pin_pagetable_pfn(unsigned cmd, unsigned long pfn)
 746 {
 747         struct mmuext_op op;
 748         op.cmd = cmd;
 749         op.arg1.mfn = pfn_to_mfn(pfn);
 750         if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
 751                 BUG();
 752 }
 753
 754 /* This needs to make sure the new pte page is pinned iff its being
 755    attached to a pinned pagetable. */
 756 static void xen_alloc_ptpage(struct mm_struct *mm, u32 pfn, unsigned level)
 757 {
 758         struct page *page = pfn_to_page(pfn);
 759
 760         if (PagePinned(virt_to_page(mm->pgd))) {
 761                 SetPagePinned(page);
 762
 763                 if (!PageHighMem(page)) {
 764                         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 765                         if (level == PT_PTE)
 766                                 pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
 767                 } else
 768                         /* make sure there are no stray mappings of
 769                            this page */
 770                         kmap_flush_unused();
 771         }
 772 }
 773
 774 static void xen_alloc_pte(struct mm_struct *mm, u32 pfn)
 775 {
 776         xen_alloc_ptpage(mm, pfn, PT_PTE);
 777 }
 778
 779 static void xen_alloc_pmd(struct mm_struct *mm, u32 pfn)
 780 {
 781         xen_alloc_ptpage(mm, pfn, PT_PMD);
 782 }
 783
 784 /* This should never happen until we're OK to use struct page */
 785 static void xen_release_ptpage(u32 pfn, unsigned level)
 786 {
 787         struct page *page = pfn_to_page(pfn);
 788
 789         if (PagePinned(page)) {
 790                 if (!PageHighMem(page)) {
 791                         if (level == PT_PTE)
 792                                 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
 793                         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 794                 }
 795                 ClearPagePinned(page);
 796         }
 797 }
 798
 799 static void xen_release_pte(u32 pfn)
 800 {
 801         xen_release_ptpage(pfn, PT_PTE);
 802 }
 803
 804 static void xen_release_pmd(u32 pfn)
 805 {
 806         xen_release_ptpage(pfn, PT_PMD);
 807 }
 808
 809 #if PAGETABLE_LEVELS == 4
 810 static void xen_alloc_pud(struct mm_struct *mm, u32 pfn)
 811 {
 812         xen_alloc_ptpage(mm, pfn, PT_PUD);
 813 }
 814
 815 static void xen_release_pud(u32 pfn)
 816 {
 817         xen_release_ptpage(pfn, PT_PUD);
 818 }
 819 #endif
 820
 821 #ifdef CONFIG_HIGHPTE
 822 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
 823 {
 824         pgprot_t prot = PAGE_KERNEL;
 825
 826         if (PagePinned(page))
 827                 prot = PAGE_KERNEL_RO;
 828
 829         if (0 && PageHighMem(page))
 830                 printk("mapping highpte %lx type %d prot %s\n",
 831                        page_to_pfn(page), type,
 832                        (unsigned long)pgprot_val(prot) & _PAGE_RW ? "WRITE" : "READ");
 833
 834         return kmap_atomic_prot(page, type, prot);
 835 }
 836 #endif
 837
 838 static __init pte_t mask_rw_pte(pte_t *ptep, pte_t pte)
 839 {
 840         /* If there's an existing pte, then don't allow _PAGE_RW to be set */
 841         if (pte_val_ma(*ptep) & _PAGE_PRESENT)
 842                 pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
 843                                pte_val_ma(pte));
 844
 845         return pte;
 846 }
 847
 848 /* Init-time set_pte while constructing initial pagetables, which
 849    doesn't allow RO pagetable pages to be remapped RW */
 850 static __init void xen_set_pte_init(pte_t *ptep, pte_t pte)
 851 {
 852         pte = mask_rw_pte(ptep, pte);
 853
 854         xen_set_pte(ptep, pte);
 855 }
 856
 857 static __init void xen_pagetable_setup_start(pgd_t *base)
 858 {
 859 }
 860
 861 void xen_setup_shared_info(void)
 862 {
 863         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 864                 set_fixmap(FIX_PARAVIRT_BOOTMAP,
 865                            xen_start_info->shared_info);
 866
 867                 HYPERVISOR_shared_info =
 868                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
 869         } else
 870                 HYPERVISOR_shared_info =
 871                         (struct shared_info *)__va(xen_start_info->shared_info);
 872
 873 #ifndef CONFIG_SMP
 874         /* In UP this is as good a place as any to set up shared info */
 875         xen_setup_vcpu_info_placement();
 876 #endif
 877
 878         xen_setup_mfn_list_list();
 879 }
 880
 881 static __init void xen_pagetable_setup_done(pgd_t *base)
 882 {
 883         xen_setup_shared_info();
 884 }
 885
 886 static __init void xen_post_allocator_init(void)
 887 {
 888         pv_mmu_ops.set_pte = xen_set_pte;
 889         pv_mmu_ops.set_pmd = xen_set_pmd;
 890         pv_mmu_ops.set_pud = xen_set_pud;
 891 #if PAGETABLE_LEVELS == 4
 892         pv_mmu_ops.set_pgd = xen_set_pgd;
 893 #endif
 894
 895         /* This will work as long as patching hasn't happened yet
 896            (which it hasn't) */
 897         pv_mmu_ops.alloc_pte = xen_alloc_pte;
 898         pv_mmu_ops.alloc_pmd = xen_alloc_pmd;
 899         pv_mmu_ops.release_pte = xen_release_pte;
 900         pv_mmu_ops.release_pmd = xen_release_pmd;
 901 #if PAGETABLE_LEVELS == 4
 902         pv_mmu_ops.alloc_pud = xen_alloc_pud;
 903         pv_mmu_ops.release_pud = xen_release_pud;
 904 #endif
 905
 906         xen_mark_init_mm_pinned();
 907 }
 908
 909 /* This is called once we have the cpu_possible_map */
 910 void xen_setup_vcpu_info_placement(void)
 911 {
 912         int cpu;
 913
 914         for_each_possible_cpu(cpu)
 915                 xen_vcpu_setup(cpu);
 916
 917         /* xen_vcpu_setup managed to place the vcpu_info within the
 918            percpu area for all cpus, so make use of it */
 919 #ifdef CONFIG_X86_32
 920         if (have_vcpu_info_placement) {
 921                 printk(KERN_INFO "Xen: using vcpu_info placement\n");
 922
 923                 pv_irq_ops.save_fl = xen_save_fl_direct;
 924                 pv_irq_ops.restore_fl = xen_restore_fl_direct;
 925                 pv_irq_ops.irq_disable = xen_irq_disable_direct;
 926                 pv_irq_ops.irq_enable = xen_irq_enable_direct;
 927                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
 928         }
 929 #endif
 930 }
 931
 932 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
 933                           unsigned long addr, unsigned len)
 934 {
 935         char *start, *end, *reloc;
 936         unsigned ret;
 937
 938         start = end = reloc = NULL;
 939
 940 #define SITE(op, x)                                                     \
 941         case PARAVIRT_PATCH(op.x):                                      \
 942         if (have_vcpu_info_placement) {                                 \
 943                 start = (char *)xen_##x##_direct;                       \
 944                 end = xen_##x##_direct_end;                             \
 945                 reloc = xen_##x##_direct_reloc;                         \
 946         }                                                               \
 947         goto patch_site
 948
 949         switch (type) {
 950 #ifdef CONFIG_X86_32
 951                 SITE(pv_irq_ops, irq_enable);
 952                 SITE(pv_irq_ops, irq_disable);
 953                 SITE(pv_irq_ops, save_fl);
 954                 SITE(pv_irq_ops, restore_fl);
 955 #endif /* CONFIG_X86_32 */
 956 #undef SITE
 957
 958         patch_site:
 959                 if (start == NULL || (end-start) > len)
 960                         goto default_patch;
 961
 962                 ret = paravirt_patch_insns(insnbuf, len, start, end);
 963
 964                 /* Note: because reloc is assigned from something that
 965                    appears to be an array, gcc assumes it's non-null,
 966                    but doesn't know its relationship with start and
 967                    end. */
 968                 if (reloc > start && reloc < end) {
 969                         int reloc_off = reloc - start;
 970                         long *relocp = (long *)(insnbuf + reloc_off);
 971                         long delta = start - (char *)addr;
 972
 973                         *relocp += delta;
 974                 }
 975                 break;
 976
 977         default_patch:
 978         default:
 979                 ret = paravirt_patch_default(type, clobbers, insnbuf,
 980                                              addr, len);
 981                 break;
 982         }
 983
 984         return ret;
 985 }
 986
 987 static void xen_set_fixmap(unsigned idx, unsigned long phys, pgprot_t prot)
 988 {
 989         pte_t pte;
 990
 991         phys >>= PAGE_SHIFT;
 992
 993         switch (idx) {
 994         case FIX_BTMAP_END ... FIX_BTMAP_BEGIN:
 995 #ifdef CONFIG_X86_F00F_BUG
 996         case FIX_F00F_IDT:
 997 #endif
 998 #ifdef CONFIG_X86_32
 999         case FIX_WP_TEST:
1000         case FIX_VDSO:
1001         case FIX_KMAP_BEGIN ... FIX_KMAP_END:
1002 #else
1003         case VSYSCALL_LAST_PAGE ... VSYSCALL_FIRST_PAGE:
1004 #endif
1005 #ifdef CONFIG_X86_LOCAL_APIC
1006         case FIX_APIC_BASE:     /* maps dummy local APIC */
1007 #endif
1008                 pte = pfn_pte(phys, prot);
1009                 break;
1010
1011         default:
1012                 pte = mfn_pte(phys, prot);
1013                 break;
1014         }
1015
1016         __native_set_fixmap(idx, pte);
1017 }
1018
1019 static const struct pv_info xen_info __initdata = {
1020         .paravirt_enabled = 1,
1021         .shared_kernel_pmd = 0,
1022
1023         .name = "Xen",
1024 };
1025
1026 static const struct pv_init_ops xen_init_ops __initdata = {
1027         .patch = xen_patch,
1028
1029         .banner = xen_banner,
1030         .memory_setup = xen_memory_setup,
1031         .arch_setup = xen_arch_setup,
1032         .post_allocator_init = xen_post_allocator_init,
1033 };
1034
1035 static const struct pv_time_ops xen_time_ops __initdata = {
1036         .time_init = xen_time_init,
1037
1038         .set_wallclock = xen_set_wallclock,
1039         .get_wallclock = xen_get_wallclock,
1040         .get_tsc_khz = xen_tsc_khz,
1041         .sched_clock = xen_sched_clock,
1042 };
1043
1044 static const struct pv_cpu_ops xen_cpu_ops __initdata = {
1045         .cpuid = xen_cpuid,
1046
1047         .set_debugreg = xen_set_debugreg,
1048         .get_debugreg = xen_get_debugreg,
1049
1050         .clts = xen_clts,
1051
1052         .read_cr0 = native_read_cr0,
1053         .write_cr0 = xen_write_cr0,
1054
1055         .read_cr4 = native_read_cr4,
1056         .read_cr4_safe = native_read_cr4_safe,
1057         .write_cr4 = xen_write_cr4,
1058
1059         .wbinvd = native_wbinvd,
1060
1061         .read_msr = native_read_msr_safe,
1062         .write_msr = native_write_msr_safe,
1063         .read_tsc = native_read_tsc,
1064         .read_pmc = native_read_pmc,
1065
1066         .iret = xen_iret,
1067         .irq_enable_sysexit = xen_sysexit,
1068
1069         .load_tr_desc = paravirt_nop,
1070         .set_ldt = xen_set_ldt,
1071         .load_gdt = xen_load_gdt,
1072         .load_idt = xen_load_idt,
1073         .load_tls = xen_load_tls,
1074 #ifdef CONFIG_X86_64
1075         .load_gs_index = xen_load_gs_index,
1076 #endif
1077
1078         .store_gdt = native_store_gdt,
1079         .store_idt = native_store_idt,
1080         .store_tr = xen_store_tr,
1081
1082         .write_ldt_entry = xen_write_ldt_entry,
1083         .write_gdt_entry = xen_write_gdt_entry,
1084         .write_idt_entry = xen_write_idt_entry,
1085         .load_sp0 = xen_load_sp0,
1086
1087         .set_iopl_mask = xen_set_iopl_mask,
1088         .io_delay = xen_io_delay,
1089
1090         /* Xen takes care of %gs when switching to usermode for us */
1091         .swapgs = paravirt_nop,
1092
1093         .lazy_mode = {
1094                 .enter = paravirt_enter_lazy_cpu,
1095                 .leave = xen_leave_lazy,
1096         },
1097 };
1098
1099 static void __init __xen_init_IRQ(void)
1100 {
1101 #ifdef CONFIG_X86_64
1102         int i;
1103
1104         /* Create identity vector->irq map */
1105         for(i = 0; i < NR_VECTORS; i++) {
1106                 int cpu;
1107
1108                 for_each_possible_cpu(cpu)
1109                         per_cpu(vector_irq, cpu)[i] = i;
1110         }
1111 #endif  /* CONFIG_X86_64 */
1112
1113         xen_init_IRQ();
1114 }
1115
1116 static const struct pv_irq_ops xen_irq_ops __initdata = {
1117         .init_IRQ = __xen_init_IRQ,
1118         .save_fl = xen_save_fl,
1119         .restore_fl = xen_restore_fl,
1120         .irq_disable = xen_irq_disable,
1121         .irq_enable = xen_irq_enable,
1122         .safe_halt = xen_safe_halt,
1123         .halt = xen_halt,
1124 #ifdef CONFIG_X86_64
1125         .adjust_exception_frame = xen_adjust_exception_frame,
1126 #endif
1127 };
1128
1129 static const struct pv_apic_ops xen_apic_ops __initdata = {
1130 #ifdef CONFIG_X86_LOCAL_APIC
1131         .apic_write = xen_apic_write,
1132         .apic_write_atomic = xen_apic_write,
1133         .apic_read = xen_apic_read,
1134         .setup_boot_clock = paravirt_nop,
1135         .setup_secondary_clock = paravirt_nop,
1136         .startup_ipi_hook = paravirt_nop,
1137 #endif
1138 };
1139
1140 static const struct pv_mmu_ops xen_mmu_ops __initdata = {
1141         .pagetable_setup_start = xen_pagetable_setup_start,
1142         .pagetable_setup_done = xen_pagetable_setup_done,
1143
1144         .read_cr2 = xen_read_cr2,
1145         .write_cr2 = xen_write_cr2,
1146
1147         .read_cr3 = xen_read_cr3,
1148         .write_cr3 = xen_write_cr3,
1149
1150         .flush_tlb_user = xen_flush_tlb,
1151         .flush_tlb_kernel = xen_flush_tlb,
1152         .flush_tlb_single = xen_flush_tlb_single,
1153         .flush_tlb_others = xen_flush_tlb_others,
1154
1155         .pte_update = paravirt_nop,
1156         .pte_update_defer = paravirt_nop,
1157
1158         .pgd_alloc = __paravirt_pgd_alloc,
1159         .pgd_free = paravirt_nop,
1160
1161         .alloc_pte = xen_alloc_pte_init,
1162         .release_pte = xen_release_pte_init,
1163         .alloc_pmd = xen_alloc_pte_init,
1164         .alloc_pmd_clone = paravirt_nop,
1165         .release_pmd = xen_release_pte_init,
1166
1167 #ifdef CONFIG_HIGHPTE
1168         .kmap_atomic_pte = xen_kmap_atomic_pte,
1169 #endif
1170
1171 #ifdef CONFIG_X86_64
1172         .set_pte = xen_set_pte,
1173 #else
1174         .set_pte = xen_set_pte_init,
1175 #endif
1176         .set_pte_at = xen_set_pte_at,
1177         .set_pmd = xen_set_pmd_hyper,
1178
1179         .ptep_modify_prot_start = __ptep_modify_prot_start,
1180         .ptep_modify_prot_commit = __ptep_modify_prot_commit,
1181
1182         .pte_val = xen_pte_val,
1183         .pte_flags = native_pte_val,
1184         .pgd_val = xen_pgd_val,
1185
1186         .make_pte = xen_make_pte,
1187         .make_pgd = xen_make_pgd,
1188
1189 #ifdef CONFIG_X86_PAE
1190         .set_pte_atomic = xen_set_pte_atomic,
1191         .set_pte_present = xen_set_pte_at,
1192         .pte_clear = xen_pte_clear,
1193         .pmd_clear = xen_pmd_clear,
1194 #endif  /* CONFIG_X86_PAE */
1195         .set_pud = xen_set_pud_hyper,
1196
1197         .make_pmd = xen_make_pmd,
1198         .pmd_val = xen_pmd_val,
1199
1200 #if PAGETABLE_LEVELS == 4
1201         .pud_val = xen_pud_val,
1202         .make_pud = xen_make_pud,
1203         .set_pgd = xen_set_pgd_hyper,
1204
1205         .alloc_pud = xen_alloc_pte_init,
1206         .release_pud = xen_release_pte_init,
1207 #endif  /* PAGETABLE_LEVELS == 4 */
1208
1209         .activate_mm = xen_activate_mm,
1210         .dup_mmap = xen_dup_mmap,
1211         .exit_mmap = xen_exit_mmap,
1212
1213         .lazy_mode = {
1214                 .enter = paravirt_enter_lazy_mmu,
1215                 .leave = xen_leave_lazy,
1216         },
1217
1218         .set_fixmap = xen_set_fixmap,
1219 };
1220
1221 static void xen_reboot(int reason)
1222 {
1223         struct sched_shutdown r = { .reason = reason };
1224
1225 #ifdef CONFIG_SMP
1226         smp_send_stop();
1227 #endif
1228
1229         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, &r))
1230                 BUG();
1231 }
1232
1233 static void xen_restart(char *msg)
1234 {
1235         xen_reboot(SHUTDOWN_reboot);
1236 }
1237
1238 static void xen_emergency_restart(void)
1239 {
1240         xen_reboot(SHUTDOWN_reboot);
1241 }
1242
1243 static void xen_machine_halt(void)
1244 {
1245         xen_reboot(SHUTDOWN_poweroff);
1246 }
1247
1248 static void xen_crash_shutdown(struct pt_regs *regs)
1249 {
1250         xen_reboot(SHUTDOWN_crash);
1251 }
1252
1253 static const struct machine_ops __initdata xen_machine_ops = {
1254         .restart = xen_restart,
1255         .halt = xen_machine_halt,
1256         .power_off = xen_machine_halt,
1257         .shutdown = xen_machine_halt,
1258         .crash_shutdown = xen_crash_shutdown,
1259         .emergency_restart = xen_emergency_restart,
1260 };
1261
1262
1263 static void __init xen_reserve_top(void)
1264 {
1265 #ifdef CONFIG_X86_32
1266         unsigned long top = HYPERVISOR_VIRT_START;
1267         struct xen_platform_parameters pp;
1268
1269         if (HYPERVISOR_xen_version(XENVER_platform_parameters, &pp) == 0)
1270                 top = pp.virt_start;
1271
1272         reserve_top_address(-top + 2 * PAGE_SIZE);
1273 #endif  /* CONFIG_X86_32 */
1274 }
1275
1276 /*
1277  * Like __va(), but returns address in the kernel mapping (which is
1278  * all we have until the physical memory mapping has been set up.
1279  */
1280 static void *__ka(phys_addr_t paddr)
1281 {
1282 #ifdef CONFIG_X86_64
1283         return (void *)(paddr + __START_KERNEL_map);
1284 #else
1285         return __va(paddr);
1286 #endif
1287 }
1288
1289 /* Convert a machine address to physical address */
1290 static unsigned long m2p(phys_addr_t maddr)
1291 {
1292         phys_addr_t paddr;
1293
1294         maddr &= PTE_MASK;
1295         paddr = mfn_to_pfn(maddr >> PAGE_SHIFT) << PAGE_SHIFT;
1296
1297         return paddr;
1298 }
1299
1300 /* Convert a machine address to kernel virtual */
1301 static void *m2v(phys_addr_t maddr)
1302 {
1303         return __ka(m2p(maddr));
1304 }
1305
1306 #ifdef CONFIG_X86_64
1307 static void walk(pgd_t *pgd, unsigned long addr)
1308 {
1309         unsigned l4idx = pgd_index(addr);
1310         unsigned l3idx = pud_index(addr);
1311         unsigned l2idx = pmd_index(addr);
1312         unsigned l1idx = pte_index(addr);
1313         pgd_t l4;
1314         pud_t l3;
1315         pmd_t l2;
1316         pte_t l1;
1317
1318         xen_raw_printk("walk %p, %lx -> %d %d %d %d\n",
1319                        pgd, addr, l4idx, l3idx, l2idx, l1idx);
1320
1321         l4 = pgd[l4idx];
1322         xen_raw_printk("  l4: %016lx\n", l4.pgd);
1323         xen_raw_printk("      %016lx\n", pgd_val(l4));
1324
1325         l3 = ((pud_t *)(m2v(l4.pgd)))[l3idx];
1326         xen_raw_printk("  l3: %016lx\n", l3.pud);
1327         xen_raw_printk("      %016lx\n", pud_val(l3));
1328
1329         l2 = ((pmd_t *)(m2v(l3.pud)))[l2idx];
1330         xen_raw_printk("  l2: %016lx\n", l2.pmd);
1331         xen_raw_printk("      %016lx\n", pmd_val(l2));
1332
1333         l1 = ((pte_t *)(m2v(l2.pmd)))[l1idx];
1334         xen_raw_printk("  l1: %016lx\n", l1.pte);
1335         xen_raw_printk("      %016lx\n", pte_val(l1));
1336 }
1337 #endif
1338
1339 static void set_page_prot(void *addr, pgprot_t prot)
1340 {
1341         unsigned long pfn = __pa(addr) >> PAGE_SHIFT;
1342         pte_t pte = pfn_pte(pfn, prot);
1343
1344         xen_raw_printk("addr=%p pfn=%lx mfn=%lx prot=%016llx pte=%016llx\n",
1345                        addr, pfn, get_phys_to_machine(pfn),
1346                        pgprot_val(prot), pte.pte);
1347
1348         if (HYPERVISOR_update_va_mapping((unsigned long)addr, pte, 0))
1349                 BUG();
1350 }
1351
1352 /*
1353  * Identity map, in addition to plain kernel map.  This needs to be
1354  * large enough to allocate page table pages to allocate the rest.
1355  * Each page can map 2MB.
1356  */
1357 static pte_t level1_ident_pgt[PTRS_PER_PTE * 4] __page_aligned_bss;
1358
1359 static __init void xen_map_identity_early(pmd_t *pmd, unsigned long max_pfn)
1360 {
1361         unsigned pmdidx, pteidx;
1362         unsigned ident_pte;
1363         unsigned long pfn;
1364
1365         ident_pte = 0;
1366         pfn = 0;
1367         for(pmdidx = 0; pmdidx < PTRS_PER_PMD && pfn < max_pfn; pmdidx++) {
1368                 pte_t *pte_page;
1369
1370                 /* Reuse or allocate a page of ptes */
1371                 if (pmd_present(pmd[pmdidx]))
1372                         pte_page = m2v(pmd[pmdidx].pmd);
1373                 else {
1374                         /* Check for free pte pages */
1375                         if (ident_pte == ARRAY_SIZE(level1_ident_pgt))
1376                                 break;
1377
1378                         pte_page = &level1_ident_pgt[ident_pte];
1379                         ident_pte += PTRS_PER_PTE;
1380
1381                         pmd[pmdidx] = __pmd(__pa(pte_page) | _PAGE_TABLE);
1382                 }
1383
1384                 /* Install mappings */
1385                 for(pteidx = 0; pteidx < PTRS_PER_PTE; pteidx++, pfn++) {
1386                         pte_t pte;
1387
1388                         if (pfn > max_pfn_mapped)
1389                                 max_pfn_mapped = pfn;
1390
1391                         if (!pte_none(pte_page[pteidx]))
1392                                 continue;
1393
1394                         pte = pfn_pte(pfn, PAGE_KERNEL_EXEC);
1395                         pte_page[pteidx] = pte;
1396                 }
1397         }
1398
1399         for(pteidx = 0; pteidx < ident_pte; pteidx += PTRS_PER_PTE)
1400                 set_page_prot(&level1_ident_pgt[pteidx], PAGE_KERNEL_RO);
1401
1402         set_page_prot(pmd, PAGE_KERNEL_RO);
1403 }
1404
1405 #ifdef CONFIG_X86_64
1406 static void convert_pfn_mfn(void *v)
1407 {
1408         pte_t *pte = v;
1409         int i;
1410
1411         /* All levels are converted the same way, so just treat them
1412            as ptes. */
1413         for(i = 0; i < PTRS_PER_PTE; i++)
1414                 pte[i] = xen_make_pte(pte[i].pte);
1415 }
1416
1417 /*
1418  * Set up the inital kernel pagetable.
1419  *
1420  * We can construct this by grafting the Xen provided pagetable into
1421  * head_64.S's preconstructed pagetables.  We copy the Xen L2's into
1422  * level2_ident_pgt, level2_kernel_pgt and level2_fixmap_pgt.  This
1423  * means that only the kernel has a physical mapping to start with -
1424  * but that's enough to get __va working.  We need to fill in the rest
1425  * of the physical mapping once some sort of allocator has been set
1426  * up.
1427  */
1428 static __init pgd_t *xen_setup_kernel_pagetable(pgd_t *pgd, unsigned long max_pfn)
1429 {
1430         pud_t *l3;
1431         pmd_t *l2;
1432
1433         /* Zap identity mapping */
1434         init_level4_pgt[0] = __pgd(0);
1435
1436         /* Pre-constructed entries are in pfn, so convert to mfn */
1437         convert_pfn_mfn(init_level4_pgt);
1438         convert_pfn_mfn(level3_ident_pgt);
1439         convert_pfn_mfn(level3_kernel_pgt);
1440
1441         l3 = m2v(pgd[pgd_index(__START_KERNEL_map)].pgd);
1442         l2 = m2v(l3[pud_index(__START_KERNEL_map)].pud);
1443
1444         memcpy(level2_ident_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1445         memcpy(level2_kernel_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1446
1447         l3 = m2v(pgd[pgd_index(__START_KERNEL_map + PMD_SIZE)].pgd);
1448         l2 = m2v(l3[pud_index(__START_KERNEL_map + PMD_SIZE)].pud);
1449         memcpy(level2_fixmap_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1450
1451         /* Set up identity map */
1452         xen_map_identity_early(level2_ident_pgt, max_pfn);
1453
1454         /* Make pagetable pieces RO */
1455         set_page_prot(init_level4_pgt, PAGE_KERNEL_RO);
1456         set_page_prot(level3_ident_pgt, PAGE_KERNEL_RO);
1457         set_page_prot(level3_kernel_pgt, PAGE_KERNEL_RO);
1458         set_page_prot(level2_kernel_pgt, PAGE_KERNEL_RO);
1459         set_page_prot(level2_fixmap_pgt, PAGE_KERNEL_RO);
1460
1461         /* Pin down new L4 */
1462         pin_pagetable_pfn(MMUEXT_PIN_L4_TABLE,
1463                           PFN_DOWN(__pa_symbol(init_level4_pgt)));
1464
1465         /* Unpin Xen-provided one */
1466         pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1467
1468         /* Switch over */
1469         pgd = init_level4_pgt;
1470         xen_write_cr3(__pa(pgd));
1471
1472         reserve_early(__pa(xen_start_info->pt_base),
1473                       __pa(xen_start_info->pt_base +
1474                            xen_start_info->nr_pt_frames * PAGE_SIZE),
1475                       "XEN PAGETABLES");
1476
1477         return pgd;
1478 }
1479 #else   /* !CONFIG_X86_64 */
1480 static pmd_t level2_kernel_pgt[PTRS_PER_PMD] __page_aligned_bss;
1481
1482 static __init pgd_t *xen_setup_kernel_pagetable(pgd_t *pgd, unsigned long max_pfn)
1483 {
1484         pmd_t *kernel_pmd;
1485
1486         init_pg_tables_start = __pa(pgd);
1487         init_pg_tables_end = __pa(pgd) + xen_start_info->nr_pt_frames*PAGE_SIZE;
1488         max_pfn_mapped = PFN_DOWN(init_pg_tables_end + 512*1024);
1489
1490         kernel_pmd = m2v(pgd[KERNEL_PGD_BOUNDARY].pgd);
1491         memcpy(level2_kernel_pgt, kernel_pmd, sizeof(pmd_t) * PTRS_PER_PMD);
1492
1493         xen_map_identity_early(level2_kernel_pgt, max_pfn);
1494
1495         memcpy(swapper_pg_dir, pgd, sizeof(pgd_t) * PTRS_PER_PGD);
1496         set_pgd(&swapper_pg_dir[KERNEL_PGD_BOUNDARY],
1497                         __pgd(__pa(level2_kernel_pgt) | _PAGE_PRESENT));
1498
1499         set_page_prot(level2_kernel_pgt, PAGE_KERNEL_RO);
1500         set_page_prot(swapper_pg_dir, PAGE_KERNEL_RO);
1501         set_page_prot(empty_zero_page, PAGE_KERNEL_RO);
1502
1503         pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1504
1505         xen_write_cr3(__pa(swapper_pg_dir));
1506
1507         pin_pagetable_pfn(MMUEXT_PIN_L3_TABLE, PFN_DOWN(__pa(swapper_pg_dir)));
1508
1509         return swapper_pg_dir;
1510 }
1511 #endif  /* CONFIG_X86_64 */
1512
1513 /* First C function to be called on Xen boot */
1514 asmlinkage void __init xen_start_kernel(void)
1515 {
1516         pgd_t *pgd;
1517
1518         if (!xen_start_info)
1519                 return;
1520
1521         BUG_ON(memcmp(xen_start_info->magic, "xen-3", 5) != 0);
1522
1523         xen_setup_features();
1524
1525         /* Install Xen paravirt ops */
1526         pv_info = xen_info;
1527         pv_init_ops = xen_init_ops;
1528         pv_time_ops = xen_time_ops;
1529         pv_cpu_ops = xen_cpu_ops;
1530         pv_irq_ops = xen_irq_ops;
1531         pv_apic_ops = xen_apic_ops;
1532         pv_mmu_ops = xen_mmu_ops;
1533
1534         if (xen_feature(XENFEAT_mmu_pt_update_preserve_ad)) {
1535                 pv_mmu_ops.ptep_modify_prot_start = xen_ptep_modify_prot_start;
1536                 pv_mmu_ops.ptep_modify_prot_commit = xen_ptep_modify_prot_commit;
1537         }
1538
1539         machine_ops = xen_machine_ops;
1540
1541 #ifdef CONFIG_X86_64
1542         /* Disable until direct per-cpu data access. */
1543         have_vcpu_info_placement = 0;
1544         x86_64_init_pda();
1545 #endif
1546
1547         xen_smp_init();
1548
1549         /* Get mfn list */
1550         if (!xen_feature(XENFEAT_auto_translated_physmap))
1551                 xen_build_dynamic_phys_to_machine();
1552
1553         pgd = (pgd_t *)xen_start_info->pt_base;
1554
1555         /* Prevent unwanted bits from being set in PTEs. */
1556         __supported_pte_mask &= ~_PAGE_GLOBAL;
1557         if (!is_initial_xendomain())
1558                 __supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
1559
1560         /* Don't do the full vcpu_info placement stuff until we have a
1561            possible map and a non-dummy shared_info. */
1562         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1563
1564         xen_raw_console_write("mapping kernel into physical memory\n");
1565         pgd = xen_setup_kernel_pagetable(pgd, xen_start_info->nr_pages);
1566
1567         init_mm.pgd = pgd;
1568
1569         /* keep using Xen gdt for now; no urgent need to change it */
1570
1571         pv_info.kernel_rpl = 1;
1572         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1573                 pv_info.kernel_rpl = 0;
1574
1575         /* set the limit of our address space */
1576         xen_reserve_top();
1577
1578 #ifdef CONFIG_X86_32
1579         /* set up basic CPUID stuff */
1580         cpu_detect(&new_cpu_data);
1581         new_cpu_data.hard_math = 1;
1582         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1583 #endif
1584
1585         /* Poke various useful things into boot_params */
1586         boot_params.hdr.type_of_loader = (9 << 4) | 0;
1587         boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1588                 ? __pa(xen_start_info->mod_start) : 0;
1589         boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1590         boot_params.hdr.cmd_line_ptr = __pa(xen_start_info->cmd_line);
1591
1592         if (!is_initial_xendomain()) {
1593                 add_preferred_console("xenboot", 0, NULL);
1594                 add_preferred_console("tty", 0, NULL);
1595                 add_preferred_console("hvc", 0, NULL);
1596         }
1597
1598         xen_raw_console_write("about to get started...\n");
1599
1600 #if 0
1601         xen_raw_printk("&boot_params=%p __pa(&boot_params)=%lx __va(__pa(&boot_params))=%lx\n",
1602                        &boot_params, __pa_symbol(&boot_params),
1603                        __va(__pa_symbol(&boot_params)));
1604
1605         walk(pgd, &boot_params);
1606         walk(pgd, __va(__pa(&boot_params)));
1607 #endif
1608
1609         /* Start the world */
1610 #ifdef CONFIG_X86_32
1611         i386_start_kernel();
1612 #else
1613         x86_64_start_reservations((char *)__pa_symbol(&boot_params));
1614 #endif
1615 }