git.oblomov.eu Git - linux-2.6/blob - 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
  29 #include <xen/interface/xen.h>
  30 #include <xen/interface/physdev.h>
  31 #include <xen/interface/vcpu.h>
  32 #include <xen/interface/sched.h>
  33 #include <xen/features.h>
  34 #include <xen/page.h>
  35
  36 #include <asm/paravirt.h>
  37 #include <asm/page.h>
  38 #include <asm/xen/hypercall.h>
  39 #include <asm/xen/hypervisor.h>
  40 #include <asm/fixmap.h>
  41 #include <asm/processor.h>
  42 #include <asm/setup.h>
  43 #include <asm/desc.h>
  44 #include <asm/pgtable.h>
  45 #include <asm/tlbflush.h>
  46 #include <asm/reboot.h>
  47
  48 #include "xen-ops.h"
  49 #include "mmu.h"
  50 #include "multicalls.h"
  51
  52 EXPORT_SYMBOL_GPL(hypercall_page);
  53
  54 DEFINE_PER_CPU(enum paravirt_lazy_mode, xen_lazy_mode);
  55
  56 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
  57 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
  58 DEFINE_PER_CPU(unsigned long, xen_cr3);
  59
  60 struct start_info *xen_start_info;
  61 EXPORT_SYMBOL_GPL(xen_start_info);
  62
  63 static /* __initdata */ struct shared_info dummy_shared_info;
  64
  65 /*
  66  * Point at some empty memory to start with. We map the real shared_info
  67  * page as soon as fixmap is up and running.
  68  */
  69 struct shared_info *HYPERVISOR_shared_info = (void *)&dummy_shared_info;
  70
  71 /*
  72  * Flag to determine whether vcpu info placement is available on all
  73  * VCPUs.  We assume it is to start with, and then set it to zero on
  74  * the first failure.  This is because it can succeed on some VCPUs
  75  * and not others, since it can involve hypervisor memory allocation,
  76  * or because the guest failed to guarantee all the appropriate
  77  * constraints on all VCPUs (ie buffer can't cross a page boundary).
  78  *
  79  * Note that any particular CPU may be using a placed vcpu structure,
  80  * but we can only optimise if the all are.
  81  *
  82  * 0: not available, 1: available
  83  */
  84 static int have_vcpu_info_placement = 1;
  85
  86 static void __init xen_vcpu_setup(int cpu)
  87 {
  88         struct vcpu_register_vcpu_info info;
  89         int err;
  90         struct vcpu_info *vcpup;
  91
  92         per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
  93
  94         if (!have_vcpu_info_placement)
  95                 return;         /* already tested, not available */
  96
  97         vcpup = &per_cpu(xen_vcpu_info, cpu);
  98
  99         info.mfn = virt_to_mfn(vcpup);
 100         info.offset = offset_in_page(vcpup);
 101
 102         printk(KERN_DEBUG "trying to map vcpu_info %d at %p, mfn %x, offset %d\n",
 103                cpu, vcpup, info.mfn, info.offset);
 104
 105         /* Check to see if the hypervisor will put the vcpu_info
 106            structure where we want it, which allows direct access via
 107            a percpu-variable. */
 108         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
 109
 110         if (err) {
 111                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
 112                 have_vcpu_info_placement = 0;
 113         } else {
 114                 /* This cpu is using the registered vcpu info, even if
 115                    later ones fail to. */
 116                 per_cpu(xen_vcpu, cpu) = vcpup;
 117
 118                 printk(KERN_DEBUG "cpu %d using vcpu_info at %p\n",
 119                        cpu, vcpup);
 120         }
 121 }
 122
 123 static void __init xen_banner(void)
 124 {
 125         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
 126                paravirt_ops.name);
 127         printk(KERN_INFO "Hypervisor signature: %s\n", xen_start_info->magic);
 128 }
 129
 130 static void xen_cpuid(unsigned int *eax, unsigned int *ebx,
 131                       unsigned int *ecx, unsigned int *edx)
 132 {
 133         unsigned maskedx = ~0;
 134
 135         /*
 136          * Mask out inconvenient features, to try and disable as many
 137          * unsupported kernel subsystems as possible.
 138          */
 139         if (*eax == 1)
 140                 maskedx = ~((1 << X86_FEATURE_APIC) |  /* disable APIC */
 141                             (1 << X86_FEATURE_ACPI) |  /* disable ACPI */
 142                             (1 << X86_FEATURE_ACC));   /* thermal monitoring */
 143
 144         asm(XEN_EMULATE_PREFIX "cpuid"
 145                 : "=a" (*eax),
 146                   "=b" (*ebx),
 147                   "=c" (*ecx),
 148                   "=d" (*edx)
 149                 : "0" (*eax), "2" (*ecx));
 150         *edx &= maskedx;
 151 }
 152
 153 static void xen_set_debugreg(int reg, unsigned long val)
 154 {
 155         HYPERVISOR_set_debugreg(reg, val);
 156 }
 157
 158 static unsigned long xen_get_debugreg(int reg)
 159 {
 160         return HYPERVISOR_get_debugreg(reg);
 161 }
 162
 163 static unsigned long xen_save_fl(void)
 164 {
 165         struct vcpu_info *vcpu;
 166         unsigned long flags;
 167
 168         vcpu = x86_read_percpu(xen_vcpu);
 169
 170         /* flag has opposite sense of mask */
 171         flags = !vcpu->evtchn_upcall_mask;
 172
 173         /* convert to IF type flag
 174            -0 -> 0x00000000
 175            -1 -> 0xffffffff
 176         */
 177         return (-flags) & X86_EFLAGS_IF;
 178 }
 179
 180 static void xen_restore_fl(unsigned long flags)
 181 {
 182         struct vcpu_info *vcpu;
 183
 184         /* convert from IF type flag */
 185         flags = !(flags & X86_EFLAGS_IF);
 186
 187         /* There's a one instruction preempt window here.  We need to
 188            make sure we're don't switch CPUs between getting the vcpu
 189            pointer and updating the mask. */
 190         preempt_disable();
 191         vcpu = x86_read_percpu(xen_vcpu);
 192         vcpu->evtchn_upcall_mask = flags;
 193         preempt_enable_no_resched();
 194
 195         /* Doesn't matter if we get preempted here, because any
 196            pending event will get dealt with anyway. */
 197
 198         if (flags == 0) {
 199                 preempt_check_resched();
 200                 barrier(); /* unmask then check (avoid races) */
 201                 if (unlikely(vcpu->evtchn_upcall_pending))
 202                         force_evtchn_callback();
 203         }
 204 }
 205
 206 static void xen_irq_disable(void)
 207 {
 208         /* There's a one instruction preempt window here.  We need to
 209            make sure we're don't switch CPUs between getting the vcpu
 210            pointer and updating the mask. */
 211         preempt_disable();
 212         x86_read_percpu(xen_vcpu)->evtchn_upcall_mask = 1;
 213         preempt_enable_no_resched();
 214 }
 215
 216 static void xen_irq_enable(void)
 217 {
 218         struct vcpu_info *vcpu;
 219
 220         /* There's a one instruction preempt window here.  We need to
 221            make sure we're don't switch CPUs between getting the vcpu
 222            pointer and updating the mask. */
 223         preempt_disable();
 224         vcpu = x86_read_percpu(xen_vcpu);
 225         vcpu->evtchn_upcall_mask = 0;
 226         preempt_enable_no_resched();
 227
 228         /* Doesn't matter if we get preempted here, because any
 229            pending event will get dealt with anyway. */
 230
 231         barrier(); /* unmask then check (avoid races) */
 232         if (unlikely(vcpu->evtchn_upcall_pending))
 233                 force_evtchn_callback();
 234 }
 235
 236 static void xen_safe_halt(void)
 237 {
 238         /* Blocking includes an implicit local_irq_enable(). */
 239         if (HYPERVISOR_sched_op(SCHEDOP_block, 0) != 0)
 240                 BUG();
 241 }
 242
 243 static void xen_halt(void)
 244 {
 245         if (irqs_disabled())
 246                 HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
 247         else
 248                 xen_safe_halt();
 249 }
 250
 251 static void xen_set_lazy_mode(enum paravirt_lazy_mode mode)
 252 {
 253         BUG_ON(preemptible());
 254
 255         switch (mode) {
 256         case PARAVIRT_LAZY_NONE:
 257                 BUG_ON(x86_read_percpu(xen_lazy_mode) == PARAVIRT_LAZY_NONE);
 258                 break;
 259
 260         case PARAVIRT_LAZY_MMU:
 261         case PARAVIRT_LAZY_CPU:
 262                 BUG_ON(x86_read_percpu(xen_lazy_mode) != PARAVIRT_LAZY_NONE);
 263                 break;
 264
 265         case PARAVIRT_LAZY_FLUSH:
 266                 /* flush if necessary, but don't change state */
 267                 if (x86_read_percpu(xen_lazy_mode) != PARAVIRT_LAZY_NONE)
 268                         xen_mc_flush();
 269                 return;
 270         }
 271
 272         xen_mc_flush();
 273         x86_write_percpu(xen_lazy_mode, mode);
 274 }
 275
 276 static unsigned long xen_store_tr(void)
 277 {
 278         return 0;
 279 }
 280
 281 static void xen_set_ldt(const void *addr, unsigned entries)
 282 {
 283         unsigned long linear_addr = (unsigned long)addr;
 284         struct mmuext_op *op;
 285         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 286
 287         op = mcs.args;
 288         op->cmd = MMUEXT_SET_LDT;
 289         if (linear_addr) {
 290                 /* ldt my be vmalloced, use arbitrary_virt_to_machine */
 291                 xmaddr_t maddr;
 292                 maddr = arbitrary_virt_to_machine((unsigned long)addr);
 293                 linear_addr = (unsigned long)maddr.maddr;
 294         }
 295         op->arg1.linear_addr = linear_addr;
 296         op->arg2.nr_ents = entries;
 297
 298         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 299
 300         xen_mc_issue(PARAVIRT_LAZY_CPU);
 301 }
 302
 303 static void xen_load_gdt(const struct Xgt_desc_struct *dtr)
 304 {
 305         unsigned long *frames;
 306         unsigned long va = dtr->address;
 307         unsigned int size = dtr->size + 1;
 308         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
 309         int f;
 310         struct multicall_space mcs;
 311
 312         /* A GDT can be up to 64k in size, which corresponds to 8192
 313            8-byte entries, or 16 4k pages.. */
 314
 315         BUG_ON(size > 65536);
 316         BUG_ON(va & ~PAGE_MASK);
 317
 318         mcs = xen_mc_entry(sizeof(*frames) * pages);
 319         frames = mcs.args;
 320
 321         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
 322                 frames[f] = virt_to_mfn(va);
 323                 make_lowmem_page_readonly((void *)va);
 324         }
 325
 326         MULTI_set_gdt(mcs.mc, frames, size / sizeof(struct desc_struct));
 327
 328         xen_mc_issue(PARAVIRT_LAZY_CPU);
 329 }
 330
 331 static void load_TLS_descriptor(struct thread_struct *t,
 332                                 unsigned int cpu, unsigned int i)
 333 {
 334         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
 335         xmaddr_t maddr = virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
 336         struct multicall_space mc = __xen_mc_entry(0);
 337
 338         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
 339 }
 340
 341 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
 342 {
 343         xen_mc_batch();
 344
 345         load_TLS_descriptor(t, cpu, 0);
 346         load_TLS_descriptor(t, cpu, 1);
 347         load_TLS_descriptor(t, cpu, 2);
 348
 349         xen_mc_issue(PARAVIRT_LAZY_CPU);
 350
 351         /*
 352          * XXX sleazy hack: If we're being called in a lazy-cpu zone,
 353          * it means we're in a context switch, and %gs has just been
 354          * saved.  This means we can zero it out to prevent faults on
 355          * exit from the hypervisor if the next process has no %gs.
 356          * Either way, it has been saved, and the new value will get
 357          * loaded properly.  This will go away as soon as Xen has been
 358          * modified to not save/restore %gs for normal hypercalls.
 359          */
 360         if (xen_get_lazy_mode() == PARAVIRT_LAZY_CPU)
 361                 loadsegment(gs, 0);
 362 }
 363
 364 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
 365                                 u32 low, u32 high)
 366 {
 367         unsigned long lp = (unsigned long)&dt[entrynum];
 368         xmaddr_t mach_lp = virt_to_machine(lp);
 369         u64 entry = (u64)high << 32 | low;
 370
 371         preempt_disable();
 372
 373         xen_mc_flush();
 374         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
 375                 BUG();
 376
 377         preempt_enable();
 378 }
 379
 380 static int cvt_gate_to_trap(int vector, u32 low, u32 high,
 381                             struct trap_info *info)
 382 {
 383         u8 type, dpl;
 384
 385         type = (high >> 8) & 0x1f;
 386         dpl = (high >> 13) & 3;
 387
 388         if (type != 0xf && type != 0xe)
 389                 return 0;
 390
 391         info->vector = vector;
 392         info->address = (high & 0xffff0000) | (low & 0x0000ffff);
 393         info->cs = low >> 16;
 394         info->flags = dpl;
 395         /* interrupt gates clear IF */
 396         if (type == 0xe)
 397                 info->flags |= 4;
 398
 399         return 1;
 400 }
 401
 402 /* Locations of each CPU's IDT */
 403 static DEFINE_PER_CPU(struct Xgt_desc_struct, idt_desc);
 404
 405 /* Set an IDT entry.  If the entry is part of the current IDT, then
 406    also update Xen. */
 407 static void xen_write_idt_entry(struct desc_struct *dt, int entrynum,
 408                                 u32 low, u32 high)
 409 {
 410         unsigned long p = (unsigned long)&dt[entrynum];
 411         unsigned long start, end;
 412
 413         preempt_disable();
 414
 415         start = __get_cpu_var(idt_desc).address;
 416         end = start + __get_cpu_var(idt_desc).size + 1;
 417
 418         xen_mc_flush();
 419
 420         write_dt_entry(dt, entrynum, low, high);
 421
 422         if (p >= start && (p + 8) <= end) {
 423                 struct trap_info info[2];
 424
 425                 info[1].address = 0;
 426
 427                 if (cvt_gate_to_trap(entrynum, low, high, &info[0]))
 428                         if (HYPERVISOR_set_trap_table(info))
 429                                 BUG();
 430         }
 431
 432         preempt_enable();
 433 }
 434
 435 static void xen_convert_trap_info(const struct Xgt_desc_struct *desc,
 436                                   struct trap_info *traps)
 437 {
 438         unsigned in, out, count;
 439
 440         count = (desc->size+1) / 8;
 441         BUG_ON(count > 256);
 442
 443         for (in = out = 0; in < count; in++) {
 444                 const u32 *entry = (u32 *)(desc->address + in * 8);
 445
 446                 if (cvt_gate_to_trap(in, entry[0], entry[1], &traps[out]))
 447                         out++;
 448         }
 449         traps[out].address = 0;
 450 }
 451
 452 void xen_copy_trap_info(struct trap_info *traps)
 453 {
 454         const struct Xgt_desc_struct *desc = &__get_cpu_var(idt_desc);
 455
 456         xen_convert_trap_info(desc, traps);
 457 }
 458
 459 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
 460    hold a spinlock to protect the static traps[] array (static because
 461    it avoids allocation, and saves stack space). */
 462 static void xen_load_idt(const struct Xgt_desc_struct *desc)
 463 {
 464         static DEFINE_SPINLOCK(lock);
 465         static struct trap_info traps[257];
 466
 467         spin_lock(&lock);
 468
 469         __get_cpu_var(idt_desc) = *desc;
 470
 471         xen_convert_trap_info(desc, traps);
 472
 473         xen_mc_flush();
 474         if (HYPERVISOR_set_trap_table(traps))
 475                 BUG();
 476
 477         spin_unlock(&lock);
 478 }
 479
 480 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
 481    they're handled differently. */
 482 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
 483                                 u32 low, u32 high)
 484 {
 485         preempt_disable();
 486
 487         switch ((high >> 8) & 0xff) {
 488         case DESCTYPE_LDT:
 489         case DESCTYPE_TSS:
 490                 /* ignore */
 491                 break;
 492
 493         default: {
 494                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
 495                 u64 desc = (u64)high << 32 | low;
 496
 497                 xen_mc_flush();
 498                 if (HYPERVISOR_update_descriptor(maddr.maddr, desc))
 499                         BUG();
 500         }
 501
 502         }
 503
 504         preempt_enable();
 505 }
 506
 507 static void xen_load_esp0(struct tss_struct *tss,
 508                           struct thread_struct *thread)
 509 {
 510         struct multicall_space mcs = xen_mc_entry(0);
 511         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->esp0);
 512         xen_mc_issue(PARAVIRT_LAZY_CPU);
 513 }
 514
 515 static void xen_set_iopl_mask(unsigned mask)
 516 {
 517         struct physdev_set_iopl set_iopl;
 518
 519         /* Force the change at ring 0. */
 520         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
 521         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
 522 }
 523
 524 static void xen_io_delay(void)
 525 {
 526 }
 527
 528 #ifdef CONFIG_X86_LOCAL_APIC
 529 static unsigned long xen_apic_read(unsigned long reg)
 530 {
 531         return 0;
 532 }
 533
 534 static void xen_apic_write(unsigned long reg, unsigned long val)
 535 {
 536         /* Warn to see if there's any stray references */
 537         WARN_ON(1);
 538 }
 539 #endif
 540
 541 static void xen_flush_tlb(void)
 542 {
 543         struct mmuext_op *op;
 544         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 545
 546         op = mcs.args;
 547         op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
 548         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 549
 550         xen_mc_issue(PARAVIRT_LAZY_MMU);
 551 }
 552
 553 static void xen_flush_tlb_single(unsigned long addr)
 554 {
 555         struct mmuext_op *op;
 556         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 557
 558         op = mcs.args;
 559         op->cmd = MMUEXT_INVLPG_LOCAL;
 560         op->arg1.linear_addr = addr & PAGE_MASK;
 561         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 562
 563         xen_mc_issue(PARAVIRT_LAZY_MMU);
 564 }
 565
 566 static void xen_flush_tlb_others(const cpumask_t *cpus, struct mm_struct *mm,
 567                                  unsigned long va)
 568 {
 569         struct {
 570                 struct mmuext_op op;
 571                 cpumask_t mask;
 572         } *args;
 573         cpumask_t cpumask = *cpus;
 574         struct multicall_space mcs;
 575
 576         /*
 577          * A couple of (to be removed) sanity checks:
 578          *
 579          * - current CPU must not be in mask
 580          * - mask must exist :)
 581          */
 582         BUG_ON(cpus_empty(cpumask));
 583         BUG_ON(cpu_isset(smp_processor_id(), cpumask));
 584         BUG_ON(!mm);
 585
 586         /* If a CPU which we ran on has gone down, OK. */
 587         cpus_and(cpumask, cpumask, cpu_online_map);
 588         if (cpus_empty(cpumask))
 589                 return;
 590
 591         mcs = xen_mc_entry(sizeof(*args));
 592         args = mcs.args;
 593         args->mask = cpumask;
 594         args->op.arg2.vcpumask = &args->mask;
 595
 596         if (va == TLB_FLUSH_ALL) {
 597                 args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
 598         } else {
 599                 args->op.cmd = MMUEXT_INVLPG_MULTI;
 600                 args->op.arg1.linear_addr = va;
 601         }
 602
 603         MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
 604
 605         xen_mc_issue(PARAVIRT_LAZY_MMU);
 606 }
 607
 608 static void xen_write_cr2(unsigned long cr2)
 609 {
 610         x86_read_percpu(xen_vcpu)->arch.cr2 = cr2;
 611 }
 612
 613 static unsigned long xen_read_cr2(void)
 614 {
 615         return x86_read_percpu(xen_vcpu)->arch.cr2;
 616 }
 617
 618 static unsigned long xen_read_cr2_direct(void)
 619 {
 620         return x86_read_percpu(xen_vcpu_info.arch.cr2);
 621 }
 622
 623 static void xen_write_cr4(unsigned long cr4)
 624 {
 625         /* Just ignore cr4 changes; Xen doesn't allow us to do
 626            anything anyway. */
 627 }
 628
 629 static unsigned long xen_read_cr3(void)
 630 {
 631         return x86_read_percpu(xen_cr3);
 632 }
 633
 634 static void xen_write_cr3(unsigned long cr3)
 635 {
 636         BUG_ON(preemptible());
 637
 638         if (cr3 == x86_read_percpu(xen_cr3)) {
 639                 /* just a simple tlb flush */
 640                 xen_flush_tlb();
 641                 return;
 642         }
 643
 644         x86_write_percpu(xen_cr3, cr3);
 645
 646
 647         {
 648                 struct mmuext_op *op;
 649                 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 650                 unsigned long mfn = pfn_to_mfn(PFN_DOWN(cr3));
 651
 652                 op = mcs.args;
 653                 op->cmd = MMUEXT_NEW_BASEPTR;
 654                 op->arg1.mfn = mfn;
 655
 656                 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 657
 658                 xen_mc_issue(PARAVIRT_LAZY_CPU);
 659         }
 660 }
 661
 662 /* Early in boot, while setting up the initial pagetable, assume
 663    everything is pinned. */
 664 static __init void xen_alloc_pt_init(struct mm_struct *mm, u32 pfn)
 665 {
 666         BUG_ON(mem_map);        /* should only be used early */
 667         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 668 }
 669
 670 /* This needs to make sure the new pte page is pinned iff its being
 671    attached to a pinned pagetable. */
 672 static void xen_alloc_pt(struct mm_struct *mm, u32 pfn)
 673 {
 674         struct page *page = pfn_to_page(pfn);
 675
 676         if (PagePinned(virt_to_page(mm->pgd))) {
 677                 SetPagePinned(page);
 678
 679                 if (!PageHighMem(page))
 680                         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 681                 else
 682                         /* make sure there are no stray mappings of
 683                            this page */
 684                         kmap_flush_unused();
 685         }
 686 }
 687
 688 /* This should never happen until we're OK to use struct page */
 689 static void xen_release_pt(u32 pfn)
 690 {
 691         struct page *page = pfn_to_page(pfn);
 692
 693         if (PagePinned(page)) {
 694                 if (!PageHighMem(page))
 695                         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 696         }
 697 }
 698
 699 #ifdef CONFIG_HIGHPTE
 700 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
 701 {
 702         pgprot_t prot = PAGE_KERNEL;
 703
 704         if (PagePinned(page))
 705                 prot = PAGE_KERNEL_RO;
 706
 707         if (0 && PageHighMem(page))
 708                 printk("mapping highpte %lx type %d prot %s\n",
 709                        page_to_pfn(page), type,
 710                        (unsigned long)pgprot_val(prot) & _PAGE_RW ? "WRITE" : "READ");
 711
 712         return kmap_atomic_prot(page, type, prot);
 713 }
 714 #endif
 715
 716 static __init pte_t mask_rw_pte(pte_t *ptep, pte_t pte)
 717 {
 718         /* If there's an existing pte, then don't allow _PAGE_RW to be set */
 719         if (pte_val_ma(*ptep) & _PAGE_PRESENT)
 720                 pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
 721                                pte_val_ma(pte));
 722
 723         return pte;
 724 }
 725
 726 /* Init-time set_pte while constructing initial pagetables, which
 727    doesn't allow RO pagetable pages to be remapped RW */
 728 static __init void xen_set_pte_init(pte_t *ptep, pte_t pte)
 729 {
 730         pte = mask_rw_pte(ptep, pte);
 731
 732         xen_set_pte(ptep, pte);
 733 }
 734
 735 static __init void xen_pagetable_setup_start(pgd_t *base)
 736 {
 737         pgd_t *xen_pgd = (pgd_t *)xen_start_info->pt_base;
 738
 739         /* special set_pte for pagetable initialization */
 740         paravirt_ops.set_pte = xen_set_pte_init;
 741
 742         init_mm.pgd = base;
 743         /*
 744          * copy top-level of Xen-supplied pagetable into place.  For
 745          * !PAE we can use this as-is, but for PAE it is a stand-in
 746          * while we copy the pmd pages.
 747          */
 748         memcpy(base, xen_pgd, PTRS_PER_PGD * sizeof(pgd_t));
 749
 750         if (PTRS_PER_PMD > 1) {
 751                 int i;
 752                 /*
 753                  * For PAE, need to allocate new pmds, rather than
 754                  * share Xen's, since Xen doesn't like pmd's being
 755                  * shared between address spaces.
 756                  */
 757                 for (i = 0; i < PTRS_PER_PGD; i++) {
 758                         if (pgd_val_ma(xen_pgd[i]) & _PAGE_PRESENT) {
 759                                 pmd_t *pmd = (pmd_t *)alloc_bootmem_low_pages(PAGE_SIZE);
 760
 761                                 memcpy(pmd, (void *)pgd_page_vaddr(xen_pgd[i]),
 762                                        PAGE_SIZE);
 763
 764                                 make_lowmem_page_readonly(pmd);
 765
 766                                 set_pgd(&base[i], __pgd(1 + __pa(pmd)));
 767                         } else
 768                                 pgd_clear(&base[i]);
 769                 }
 770         }
 771
 772         /* make sure zero_page is mapped RO so we can use it in pagetables */
 773         make_lowmem_page_readonly(empty_zero_page);
 774         make_lowmem_page_readonly(base);
 775         /*
 776          * Switch to new pagetable.  This is done before
 777          * pagetable_init has done anything so that the new pages
 778          * added to the table can be prepared properly for Xen.
 779          */
 780         xen_write_cr3(__pa(base));
 781 }
 782
 783 static __init void xen_pagetable_setup_done(pgd_t *base)
 784 {
 785         /* This will work as long as patching hasn't happened yet
 786            (which it hasn't) */
 787         paravirt_ops.alloc_pt = xen_alloc_pt;
 788         paravirt_ops.set_pte = xen_set_pte;
 789
 790         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 791                 /*
 792                  * Create a mapping for the shared info page.
 793                  * Should be set_fixmap(), but shared_info is a machine
 794                  * address with no corresponding pseudo-phys address.
 795                  */
 796                 set_pte_mfn(fix_to_virt(FIX_PARAVIRT_BOOTMAP),
 797                             PFN_DOWN(xen_start_info->shared_info),
 798                             PAGE_KERNEL);
 799
 800                 HYPERVISOR_shared_info =
 801                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
 802
 803         } else
 804                 HYPERVISOR_shared_info =
 805                         (struct shared_info *)__va(xen_start_info->shared_info);
 806
 807         /* Actually pin the pagetable down, but we can't set PG_pinned
 808            yet because the page structures don't exist yet. */
 809         {
 810                 struct mmuext_op op;
 811 #ifdef CONFIG_X86_PAE
 812                 op.cmd = MMUEXT_PIN_L3_TABLE;
 813 #else
 814                 op.cmd = MMUEXT_PIN_L3_TABLE;
 815 #endif
 816                 op.arg1.mfn = pfn_to_mfn(PFN_DOWN(__pa(base)));
 817                 if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
 818                         BUG();
 819         }
 820 }
 821
 822 /* This is called once we have the cpu_possible_map */
 823 void __init xen_setup_vcpu_info_placement(void)
 824 {
 825         int cpu;
 826
 827         for_each_possible_cpu(cpu)
 828                 xen_vcpu_setup(cpu);
 829
 830         /* xen_vcpu_setup managed to place the vcpu_info within the
 831            percpu area for all cpus, so make use of it */
 832         if (have_vcpu_info_placement) {
 833                 printk(KERN_INFO "Xen: using vcpu_info placement\n");
 834
 835                 paravirt_ops.save_fl = xen_save_fl_direct;
 836                 paravirt_ops.restore_fl = xen_restore_fl_direct;
 837                 paravirt_ops.irq_disable = xen_irq_disable_direct;
 838                 paravirt_ops.irq_enable = xen_irq_enable_direct;
 839                 paravirt_ops.read_cr2 = xen_read_cr2_direct;
 840                 paravirt_ops.iret = xen_iret_direct;
 841         }
 842 }
 843
 844 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
 845                           unsigned long addr, unsigned len)
 846 {
 847         char *start, *end, *reloc;
 848         unsigned ret;
 849
 850         start = end = reloc = NULL;
 851
 852 #define SITE(x)                                                         \
 853         case PARAVIRT_PATCH(x):                                         \
 854         if (have_vcpu_info_placement) {                                 \
 855                 start = (char *)xen_##x##_direct;                       \
 856                 end = xen_##x##_direct_end;                             \
 857                 reloc = xen_##x##_direct_reloc;                         \
 858         }                                                               \
 859         goto patch_site
 860
 861         switch (type) {
 862                 SITE(irq_enable);
 863                 SITE(irq_disable);
 864                 SITE(save_fl);
 865                 SITE(restore_fl);
 866 #undef SITE
 867
 868         patch_site:
 869                 if (start == NULL || (end-start) > len)
 870                         goto default_patch;
 871
 872                 ret = paravirt_patch_insns(insnbuf, len, start, end);
 873
 874                 /* Note: because reloc is assigned from something that
 875                    appears to be an array, gcc assumes it's non-null,
 876                    but doesn't know its relationship with start and
 877                    end. */
 878                 if (reloc > start && reloc < end) {
 879                         int reloc_off = reloc - start;
 880                         long *relocp = (long *)(insnbuf + reloc_off);
 881                         long delta = start - (char *)addr;
 882
 883                         *relocp += delta;
 884                 }
 885                 break;
 886
 887         default_patch:
 888         default:
 889                 ret = paravirt_patch_default(type, clobbers, insnbuf,
 890                                              addr, len);
 891                 break;
 892         }
 893
 894         return ret;
 895 }
 896
 897 static const struct paravirt_ops xen_paravirt_ops __initdata = {
 898         .paravirt_enabled = 1,
 899         .shared_kernel_pmd = 0,
 900
 901         .name = "Xen",
 902         .banner = xen_banner,
 903
 904         .patch = xen_patch,
 905
 906         .memory_setup = xen_memory_setup,
 907         .arch_setup = xen_arch_setup,
 908         .init_IRQ = xen_init_IRQ,
 909         .post_allocator_init = xen_mark_init_mm_pinned,
 910
 911         .time_init = xen_time_init,
 912         .set_wallclock = xen_set_wallclock,
 913         .get_wallclock = xen_get_wallclock,
 914         .get_cpu_khz = xen_cpu_khz,
 915         .sched_clock = xen_sched_clock,
 916
 917         .cpuid = xen_cpuid,
 918
 919         .set_debugreg = xen_set_debugreg,
 920         .get_debugreg = xen_get_debugreg,
 921
 922         .clts = native_clts,
 923
 924         .read_cr0 = native_read_cr0,
 925         .write_cr0 = native_write_cr0,
 926
 927         .read_cr2 = xen_read_cr2,
 928         .write_cr2 = xen_write_cr2,
 929
 930         .read_cr3 = xen_read_cr3,
 931         .write_cr3 = xen_write_cr3,
 932
 933         .read_cr4 = native_read_cr4,
 934         .read_cr4_safe = native_read_cr4_safe,
 935         .write_cr4 = xen_write_cr4,
 936
 937         .save_fl = xen_save_fl,
 938         .restore_fl = xen_restore_fl,
 939         .irq_disable = xen_irq_disable,
 940         .irq_enable = xen_irq_enable,
 941         .safe_halt = xen_safe_halt,
 942         .halt = xen_halt,
 943         .wbinvd = native_wbinvd,
 944
 945         .read_msr = native_read_msr_safe,
 946         .write_msr = native_write_msr_safe,
 947         .read_tsc = native_read_tsc,
 948         .read_pmc = native_read_pmc,
 949
 950         .iret = (void *)&hypercall_page[__HYPERVISOR_iret],
 951         .irq_enable_sysexit = NULL,  /* never called */
 952
 953         .load_tr_desc = paravirt_nop,
 954         .set_ldt = xen_set_ldt,
 955         .load_gdt = xen_load_gdt,
 956         .load_idt = xen_load_idt,
 957         .load_tls = xen_load_tls,
 958
 959         .store_gdt = native_store_gdt,
 960         .store_idt = native_store_idt,
 961         .store_tr = xen_store_tr,
 962
 963         .write_ldt_entry = xen_write_ldt_entry,
 964         .write_gdt_entry = xen_write_gdt_entry,
 965         .write_idt_entry = xen_write_idt_entry,
 966         .load_esp0 = xen_load_esp0,
 967
 968         .set_iopl_mask = xen_set_iopl_mask,
 969         .io_delay = xen_io_delay,
 970
 971 #ifdef CONFIG_X86_LOCAL_APIC
 972         .apic_write = xen_apic_write,
 973         .apic_write_atomic = xen_apic_write,
 974         .apic_read = xen_apic_read,
 975         .setup_boot_clock = paravirt_nop,
 976         .setup_secondary_clock = paravirt_nop,
 977         .startup_ipi_hook = paravirt_nop,
 978 #endif
 979
 980         .flush_tlb_user = xen_flush_tlb,
 981         .flush_tlb_kernel = xen_flush_tlb,
 982         .flush_tlb_single = xen_flush_tlb_single,
 983         .flush_tlb_others = xen_flush_tlb_others,
 984
 985         .pte_update = paravirt_nop,
 986         .pte_update_defer = paravirt_nop,
 987
 988         .pagetable_setup_start = xen_pagetable_setup_start,
 989         .pagetable_setup_done = xen_pagetable_setup_done,
 990
 991         .alloc_pt = xen_alloc_pt_init,
 992         .release_pt = xen_release_pt,
 993         .alloc_pd = paravirt_nop,
 994         .alloc_pd_clone = paravirt_nop,
 995         .release_pd = paravirt_nop,
 996
 997 #ifdef CONFIG_HIGHPTE
 998         .kmap_atomic_pte = xen_kmap_atomic_pte,
 999 #endif
1000
1001         .set_pte = NULL,        /* see xen_pagetable_setup_* */
1002         .set_pte_at = xen_set_pte_at,
1003         .set_pmd = xen_set_pmd,
1004
1005         .pte_val = xen_pte_val,
1006         .pgd_val = xen_pgd_val,
1007
1008         .make_pte = xen_make_pte,
1009         .make_pgd = xen_make_pgd,
1010
1011 #ifdef CONFIG_X86_PAE
1012         .set_pte_atomic = xen_set_pte_atomic,
1013         .set_pte_present = xen_set_pte_at,
1014         .set_pud = xen_set_pud,
1015         .pte_clear = xen_pte_clear,
1016         .pmd_clear = xen_pmd_clear,
1017
1018         .make_pmd = xen_make_pmd,
1019         .pmd_val = xen_pmd_val,
1020 #endif  /* PAE */
1021
1022         .activate_mm = xen_activate_mm,
1023         .dup_mmap = xen_dup_mmap,
1024         .exit_mmap = xen_exit_mmap,
1025
1026         .set_lazy_mode = xen_set_lazy_mode,
1027 };
1028
1029 #ifdef CONFIG_SMP
1030 static const struct smp_ops xen_smp_ops __initdata = {
1031         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
1032         .smp_prepare_cpus = xen_smp_prepare_cpus,
1033         .cpu_up = xen_cpu_up,
1034         .smp_cpus_done = xen_smp_cpus_done,
1035
1036         .smp_send_stop = xen_smp_send_stop,
1037         .smp_send_reschedule = xen_smp_send_reschedule,
1038         .smp_call_function_mask = xen_smp_call_function_mask,
1039 };
1040 #endif  /* CONFIG_SMP */
1041
1042 static void xen_reboot(int reason)
1043 {
1044 #ifdef CONFIG_SMP
1045         smp_send_stop();
1046 #endif
1047
1048         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, reason))
1049                 BUG();
1050 }
1051
1052 static void xen_restart(char *msg)
1053 {
1054         xen_reboot(SHUTDOWN_reboot);
1055 }
1056
1057 static void xen_emergency_restart(void)
1058 {
1059         xen_reboot(SHUTDOWN_reboot);
1060 }
1061
1062 static void xen_machine_halt(void)
1063 {
1064         xen_reboot(SHUTDOWN_poweroff);
1065 }
1066
1067 static void xen_crash_shutdown(struct pt_regs *regs)
1068 {
1069         xen_reboot(SHUTDOWN_crash);
1070 }
1071
1072 static const struct machine_ops __initdata xen_machine_ops = {
1073         .restart = xen_restart,
1074         .halt = xen_machine_halt,
1075         .power_off = xen_machine_halt,
1076         .shutdown = xen_machine_halt,
1077         .crash_shutdown = xen_crash_shutdown,
1078         .emergency_restart = xen_emergency_restart,
1079 };
1080
1081
1082 /* First C function to be called on Xen boot */
1083 asmlinkage void __init xen_start_kernel(void)
1084 {
1085         pgd_t *pgd;
1086
1087         if (!xen_start_info)
1088                 return;
1089
1090         BUG_ON(memcmp(xen_start_info->magic, "xen-3.0", 7) != 0);
1091
1092         /* Install Xen paravirt ops */
1093         paravirt_ops = xen_paravirt_ops;
1094         machine_ops = xen_machine_ops;
1095
1096 #ifdef CONFIG_SMP
1097         smp_ops = xen_smp_ops;
1098 #endif
1099
1100         xen_setup_features();
1101
1102         /* Get mfn list */
1103         if (!xen_feature(XENFEAT_auto_translated_physmap))
1104                 phys_to_machine_mapping = (unsigned long *)xen_start_info->mfn_list;
1105
1106         pgd = (pgd_t *)xen_start_info->pt_base;
1107
1108         init_pg_tables_end = __pa(pgd) + xen_start_info->nr_pt_frames*PAGE_SIZE;
1109
1110         init_mm.pgd = pgd; /* use the Xen pagetables to start */
1111
1112         /* keep using Xen gdt for now; no urgent need to change it */
1113
1114         x86_write_percpu(xen_cr3, __pa(pgd));
1115
1116 #ifdef CONFIG_SMP
1117         /* Don't do the full vcpu_info placement stuff until we have a
1118            possible map. */
1119         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1120 #else
1121         /* May as well do it now, since there's no good time to call
1122            it later on UP. */
1123         xen_setup_vcpu_info_placement();
1124 #endif
1125
1126         paravirt_ops.kernel_rpl = 1;
1127         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1128                 paravirt_ops.kernel_rpl = 0;
1129
1130         /* set the limit of our address space */
1131         reserve_top_address(-HYPERVISOR_VIRT_START + 2 * PAGE_SIZE);
1132
1133         /* set up basic CPUID stuff */
1134         cpu_detect(&new_cpu_data);
1135         new_cpu_data.hard_math = 1;
1136         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1137
1138         /* Poke various useful things into boot_params */
1139         boot_params.hdr.type_of_loader = (9 << 4) | 0;
1140         boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1141                 ? __pa(xen_start_info->mod_start) : 0;
1142         boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1143
1144         /* Start the world */
1145         start_kernel();
1146 }