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(struct vcpu_info *, xen_vcpu);
  55 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
  56 DEFINE_PER_CPU(unsigned long, xen_cr3);
  57
  58 struct start_info *xen_start_info;
  59 EXPORT_SYMBOL_GPL(xen_start_info);
  60
  61 static /* __initdata */ struct shared_info dummy_shared_info;
  62
  63 /*
  64  * Point at some empty memory to start with. We map the real shared_info
  65  * page as soon as fixmap is up and running.
  66  */
  67 struct shared_info *HYPERVISOR_shared_info = (void *)&dummy_shared_info;
  68
  69 /*
  70  * Flag to determine whether vcpu info placement is available on all
  71  * VCPUs.  We assume it is to start with, and then set it to zero on
  72  * the first failure.  This is because it can succeed on some VCPUs
  73  * and not others, since it can involve hypervisor memory allocation,
  74  * or because the guest failed to guarantee all the appropriate
  75  * constraints on all VCPUs (ie buffer can't cross a page boundary).
  76  *
  77  * Note that any particular CPU may be using a placed vcpu structure,
  78  * but we can only optimise if the all are.
  79  *
  80  * 0: not available, 1: available
  81  */
  82 static int have_vcpu_info_placement = 1;
  83
  84 static void __init xen_vcpu_setup(int cpu)
  85 {
  86         struct vcpu_register_vcpu_info info;
  87         int err;
  88         struct vcpu_info *vcpup;
  89
  90         per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
  91
  92         if (!have_vcpu_info_placement)
  93                 return;         /* already tested, not available */
  94
  95         vcpup = &per_cpu(xen_vcpu_info, cpu);
  96
  97         info.mfn = virt_to_mfn(vcpup);
  98         info.offset = offset_in_page(vcpup);
  99
 100         printk(KERN_DEBUG "trying to map vcpu_info %d at %p, mfn %x, offset %d\n",
 101                cpu, vcpup, info.mfn, info.offset);
 102
 103         /* Check to see if the hypervisor will put the vcpu_info
 104            structure where we want it, which allows direct access via
 105            a percpu-variable. */
 106         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
 107
 108         if (err) {
 109                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
 110                 have_vcpu_info_placement = 0;
 111         } else {
 112                 /* This cpu is using the registered vcpu info, even if
 113                    later ones fail to. */
 114                 per_cpu(xen_vcpu, cpu) = vcpup;
 115
 116                 printk(KERN_DEBUG "cpu %d using vcpu_info at %p\n",
 117                        cpu, vcpup);
 118         }
 119 }
 120
 121 static void __init xen_banner(void)
 122 {
 123         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
 124                pv_info.name);
 125         printk(KERN_INFO "Hypervisor signature: %s\n", xen_start_info->magic);
 126 }
 127
 128 static void xen_cpuid(unsigned int *eax, unsigned int *ebx,
 129                       unsigned int *ecx, unsigned int *edx)
 130 {
 131         unsigned maskedx = ~0;
 132
 133         /*
 134          * Mask out inconvenient features, to try and disable as many
 135          * unsupported kernel subsystems as possible.
 136          */
 137         if (*eax == 1)
 138                 maskedx = ~((1 << X86_FEATURE_APIC) |  /* disable APIC */
 139                             (1 << X86_FEATURE_ACPI) |  /* disable ACPI */
 140                             (1 << X86_FEATURE_ACC));   /* thermal monitoring */
 141
 142         asm(XEN_EMULATE_PREFIX "cpuid"
 143                 : "=a" (*eax),
 144                   "=b" (*ebx),
 145                   "=c" (*ecx),
 146                   "=d" (*edx)
 147                 : "0" (*eax), "2" (*ecx));
 148         *edx &= maskedx;
 149 }
 150
 151 static void xen_set_debugreg(int reg, unsigned long val)
 152 {
 153         HYPERVISOR_set_debugreg(reg, val);
 154 }
 155
 156 static unsigned long xen_get_debugreg(int reg)
 157 {
 158         return HYPERVISOR_get_debugreg(reg);
 159 }
 160
 161 static unsigned long xen_save_fl(void)
 162 {
 163         struct vcpu_info *vcpu;
 164         unsigned long flags;
 165
 166         vcpu = x86_read_percpu(xen_vcpu);
 167
 168         /* flag has opposite sense of mask */
 169         flags = !vcpu->evtchn_upcall_mask;
 170
 171         /* convert to IF type flag
 172            -0 -> 0x00000000
 173            -1 -> 0xffffffff
 174         */
 175         return (-flags) & X86_EFLAGS_IF;
 176 }
 177
 178 static void xen_restore_fl(unsigned long flags)
 179 {
 180         struct vcpu_info *vcpu;
 181
 182         /* convert from IF type flag */
 183         flags = !(flags & X86_EFLAGS_IF);
 184
 185         /* There's a one instruction preempt window here.  We need to
 186            make sure we're don't switch CPUs between getting the vcpu
 187            pointer and updating the mask. */
 188         preempt_disable();
 189         vcpu = x86_read_percpu(xen_vcpu);
 190         vcpu->evtchn_upcall_mask = flags;
 191         preempt_enable_no_resched();
 192
 193         /* Doesn't matter if we get preempted here, because any
 194            pending event will get dealt with anyway. */
 195
 196         if (flags == 0) {
 197                 preempt_check_resched();
 198                 barrier(); /* unmask then check (avoid races) */
 199                 if (unlikely(vcpu->evtchn_upcall_pending))
 200                         force_evtchn_callback();
 201         }
 202 }
 203
 204 static void xen_irq_disable(void)
 205 {
 206         /* There's a one instruction preempt window here.  We need to
 207            make sure we're don't switch CPUs between getting the vcpu
 208            pointer and updating the mask. */
 209         preempt_disable();
 210         x86_read_percpu(xen_vcpu)->evtchn_upcall_mask = 1;
 211         preempt_enable_no_resched();
 212 }
 213
 214 static void xen_irq_enable(void)
 215 {
 216         struct vcpu_info *vcpu;
 217
 218         /* There's a one instruction preempt window here.  We need to
 219            make sure we're don't switch CPUs between getting the vcpu
 220            pointer and updating the mask. */
 221         preempt_disable();
 222         vcpu = x86_read_percpu(xen_vcpu);
 223         vcpu->evtchn_upcall_mask = 0;
 224         preempt_enable_no_resched();
 225
 226         /* Doesn't matter if we get preempted here, because any
 227            pending event will get dealt with anyway. */
 228
 229         barrier(); /* unmask then check (avoid races) */
 230         if (unlikely(vcpu->evtchn_upcall_pending))
 231                 force_evtchn_callback();
 232 }
 233
 234 static void xen_safe_halt(void)
 235 {
 236         /* Blocking includes an implicit local_irq_enable(). */
 237         if (HYPERVISOR_sched_op(SCHEDOP_block, 0) != 0)
 238                 BUG();
 239 }
 240
 241 static void xen_halt(void)
 242 {
 243         if (irqs_disabled())
 244                 HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
 245         else
 246                 xen_safe_halt();
 247 }
 248
 249 static void xen_leave_lazy(void)
 250 {
 251         paravirt_leave_lazy(paravirt_get_lazy_mode());
 252         xen_mc_flush();
 253 }
 254
 255 static unsigned long xen_store_tr(void)
 256 {
 257         return 0;
 258 }
 259
 260 static void xen_set_ldt(const void *addr, unsigned entries)
 261 {
 262         unsigned long linear_addr = (unsigned long)addr;
 263         struct mmuext_op *op;
 264         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 265
 266         op = mcs.args;
 267         op->cmd = MMUEXT_SET_LDT;
 268         if (linear_addr) {
 269                 /* ldt my be vmalloced, use arbitrary_virt_to_machine */
 270                 xmaddr_t maddr;
 271                 maddr = arbitrary_virt_to_machine((unsigned long)addr);
 272                 linear_addr = (unsigned long)maddr.maddr;
 273         }
 274         op->arg1.linear_addr = linear_addr;
 275         op->arg2.nr_ents = entries;
 276
 277         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 278
 279         xen_mc_issue(PARAVIRT_LAZY_CPU);
 280 }
 281
 282 static void xen_load_gdt(const struct Xgt_desc_struct *dtr)
 283 {
 284         unsigned long *frames;
 285         unsigned long va = dtr->address;
 286         unsigned int size = dtr->size + 1;
 287         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
 288         int f;
 289         struct multicall_space mcs;
 290
 291         /* A GDT can be up to 64k in size, which corresponds to 8192
 292            8-byte entries, or 16 4k pages.. */
 293
 294         BUG_ON(size > 65536);
 295         BUG_ON(va & ~PAGE_MASK);
 296
 297         mcs = xen_mc_entry(sizeof(*frames) * pages);
 298         frames = mcs.args;
 299
 300         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
 301                 frames[f] = virt_to_mfn(va);
 302                 make_lowmem_page_readonly((void *)va);
 303         }
 304
 305         MULTI_set_gdt(mcs.mc, frames, size / sizeof(struct desc_struct));
 306
 307         xen_mc_issue(PARAVIRT_LAZY_CPU);
 308 }
 309
 310 static void load_TLS_descriptor(struct thread_struct *t,
 311                                 unsigned int cpu, unsigned int i)
 312 {
 313         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
 314         xmaddr_t maddr = virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
 315         struct multicall_space mc = __xen_mc_entry(0);
 316
 317         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
 318 }
 319
 320 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
 321 {
 322         xen_mc_batch();
 323
 324         load_TLS_descriptor(t, cpu, 0);
 325         load_TLS_descriptor(t, cpu, 1);
 326         load_TLS_descriptor(t, cpu, 2);
 327
 328         xen_mc_issue(PARAVIRT_LAZY_CPU);
 329
 330         /*
 331          * XXX sleazy hack: If we're being called in a lazy-cpu zone,
 332          * it means we're in a context switch, and %gs has just been
 333          * saved.  This means we can zero it out to prevent faults on
 334          * exit from the hypervisor if the next process has no %gs.
 335          * Either way, it has been saved, and the new value will get
 336          * loaded properly.  This will go away as soon as Xen has been
 337          * modified to not save/restore %gs for normal hypercalls.
 338          */
 339         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU)
 340                 loadsegment(gs, 0);
 341 }
 342
 343 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
 344                                 u32 low, u32 high)
 345 {
 346         unsigned long lp = (unsigned long)&dt[entrynum];
 347         xmaddr_t mach_lp = virt_to_machine(lp);
 348         u64 entry = (u64)high << 32 | low;
 349
 350         preempt_disable();
 351
 352         xen_mc_flush();
 353         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
 354                 BUG();
 355
 356         preempt_enable();
 357 }
 358
 359 static int cvt_gate_to_trap(int vector, u32 low, u32 high,
 360                             struct trap_info *info)
 361 {
 362         u8 type, dpl;
 363
 364         type = (high >> 8) & 0x1f;
 365         dpl = (high >> 13) & 3;
 366
 367         if (type != 0xf && type != 0xe)
 368                 return 0;
 369
 370         info->vector = vector;
 371         info->address = (high & 0xffff0000) | (low & 0x0000ffff);
 372         info->cs = low >> 16;
 373         info->flags = dpl;
 374         /* interrupt gates clear IF */
 375         if (type == 0xe)
 376                 info->flags |= 4;
 377
 378         return 1;
 379 }
 380
 381 /* Locations of each CPU's IDT */
 382 static DEFINE_PER_CPU(struct Xgt_desc_struct, idt_desc);
 383
 384 /* Set an IDT entry.  If the entry is part of the current IDT, then
 385    also update Xen. */
 386 static void xen_write_idt_entry(struct desc_struct *dt, int entrynum,
 387                                 u32 low, u32 high)
 388 {
 389         unsigned long p = (unsigned long)&dt[entrynum];
 390         unsigned long start, end;
 391
 392         preempt_disable();
 393
 394         start = __get_cpu_var(idt_desc).address;
 395         end = start + __get_cpu_var(idt_desc).size + 1;
 396
 397         xen_mc_flush();
 398
 399         write_dt_entry(dt, entrynum, low, high);
 400
 401         if (p >= start && (p + 8) <= end) {
 402                 struct trap_info info[2];
 403
 404                 info[1].address = 0;
 405
 406                 if (cvt_gate_to_trap(entrynum, low, high, &info[0]))
 407                         if (HYPERVISOR_set_trap_table(info))
 408                                 BUG();
 409         }
 410
 411         preempt_enable();
 412 }
 413
 414 static void xen_convert_trap_info(const struct Xgt_desc_struct *desc,
 415                                   struct trap_info *traps)
 416 {
 417         unsigned in, out, count;
 418
 419         count = (desc->size+1) / 8;
 420         BUG_ON(count > 256);
 421
 422         for (in = out = 0; in < count; in++) {
 423                 const u32 *entry = (u32 *)(desc->address + in * 8);
 424
 425                 if (cvt_gate_to_trap(in, entry[0], entry[1], &traps[out]))
 426                         out++;
 427         }
 428         traps[out].address = 0;
 429 }
 430
 431 void xen_copy_trap_info(struct trap_info *traps)
 432 {
 433         const struct Xgt_desc_struct *desc = &__get_cpu_var(idt_desc);
 434
 435         xen_convert_trap_info(desc, traps);
 436 }
 437
 438 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
 439    hold a spinlock to protect the static traps[] array (static because
 440    it avoids allocation, and saves stack space). */
 441 static void xen_load_idt(const struct Xgt_desc_struct *desc)
 442 {
 443         static DEFINE_SPINLOCK(lock);
 444         static struct trap_info traps[257];
 445
 446         spin_lock(&lock);
 447
 448         __get_cpu_var(idt_desc) = *desc;
 449
 450         xen_convert_trap_info(desc, traps);
 451
 452         xen_mc_flush();
 453         if (HYPERVISOR_set_trap_table(traps))
 454                 BUG();
 455
 456         spin_unlock(&lock);
 457 }
 458
 459 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
 460    they're handled differently. */
 461 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
 462                                 u32 low, u32 high)
 463 {
 464         preempt_disable();
 465
 466         switch ((high >> 8) & 0xff) {
 467         case DESCTYPE_LDT:
 468         case DESCTYPE_TSS:
 469                 /* ignore */
 470                 break;
 471
 472         default: {
 473                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
 474                 u64 desc = (u64)high << 32 | low;
 475
 476                 xen_mc_flush();
 477                 if (HYPERVISOR_update_descriptor(maddr.maddr, desc))
 478                         BUG();
 479         }
 480
 481         }
 482
 483         preempt_enable();
 484 }
 485
 486 static void xen_load_esp0(struct tss_struct *tss,
 487                           struct thread_struct *thread)
 488 {
 489         struct multicall_space mcs = xen_mc_entry(0);
 490         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->esp0);
 491         xen_mc_issue(PARAVIRT_LAZY_CPU);
 492 }
 493
 494 static void xen_set_iopl_mask(unsigned mask)
 495 {
 496         struct physdev_set_iopl set_iopl;
 497
 498         /* Force the change at ring 0. */
 499         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
 500         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
 501 }
 502
 503 static void xen_io_delay(void)
 504 {
 505 }
 506
 507 #ifdef CONFIG_X86_LOCAL_APIC
 508 static unsigned long xen_apic_read(unsigned long reg)
 509 {
 510         return 0;
 511 }
 512
 513 static void xen_apic_write(unsigned long reg, unsigned long val)
 514 {
 515         /* Warn to see if there's any stray references */
 516         WARN_ON(1);
 517 }
 518 #endif
 519
 520 static void xen_flush_tlb(void)
 521 {
 522         struct mmuext_op *op;
 523         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 524
 525         op = mcs.args;
 526         op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
 527         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 528
 529         xen_mc_issue(PARAVIRT_LAZY_MMU);
 530 }
 531
 532 static void xen_flush_tlb_single(unsigned long addr)
 533 {
 534         struct mmuext_op *op;
 535         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 536
 537         op = mcs.args;
 538         op->cmd = MMUEXT_INVLPG_LOCAL;
 539         op->arg1.linear_addr = addr & PAGE_MASK;
 540         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 541
 542         xen_mc_issue(PARAVIRT_LAZY_MMU);
 543 }
 544
 545 static void xen_flush_tlb_others(const cpumask_t *cpus, struct mm_struct *mm,
 546                                  unsigned long va)
 547 {
 548         struct {
 549                 struct mmuext_op op;
 550                 cpumask_t mask;
 551         } *args;
 552         cpumask_t cpumask = *cpus;
 553         struct multicall_space mcs;
 554
 555         /*
 556          * A couple of (to be removed) sanity checks:
 557          *
 558          * - current CPU must not be in mask
 559          * - mask must exist :)
 560          */
 561         BUG_ON(cpus_empty(cpumask));
 562         BUG_ON(cpu_isset(smp_processor_id(), cpumask));
 563         BUG_ON(!mm);
 564
 565         /* If a CPU which we ran on has gone down, OK. */
 566         cpus_and(cpumask, cpumask, cpu_online_map);
 567         if (cpus_empty(cpumask))
 568                 return;
 569
 570         mcs = xen_mc_entry(sizeof(*args));
 571         args = mcs.args;
 572         args->mask = cpumask;
 573         args->op.arg2.vcpumask = &args->mask;
 574
 575         if (va == TLB_FLUSH_ALL) {
 576                 args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
 577         } else {
 578                 args->op.cmd = MMUEXT_INVLPG_MULTI;
 579                 args->op.arg1.linear_addr = va;
 580         }
 581
 582         MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
 583
 584         xen_mc_issue(PARAVIRT_LAZY_MMU);
 585 }
 586
 587 static void xen_write_cr2(unsigned long cr2)
 588 {
 589         x86_read_percpu(xen_vcpu)->arch.cr2 = cr2;
 590 }
 591
 592 static unsigned long xen_read_cr2(void)
 593 {
 594         return x86_read_percpu(xen_vcpu)->arch.cr2;
 595 }
 596
 597 static unsigned long xen_read_cr2_direct(void)
 598 {
 599         return x86_read_percpu(xen_vcpu_info.arch.cr2);
 600 }
 601
 602 static void xen_write_cr4(unsigned long cr4)
 603 {
 604         /* Just ignore cr4 changes; Xen doesn't allow us to do
 605            anything anyway. */
 606 }
 607
 608 static unsigned long xen_read_cr3(void)
 609 {
 610         return x86_read_percpu(xen_cr3);
 611 }
 612
 613 static void xen_write_cr3(unsigned long cr3)
 614 {
 615         BUG_ON(preemptible());
 616
 617         if (cr3 == x86_read_percpu(xen_cr3)) {
 618                 /* just a simple tlb flush */
 619                 xen_flush_tlb();
 620                 return;
 621         }
 622
 623         x86_write_percpu(xen_cr3, cr3);
 624
 625
 626         {
 627                 struct mmuext_op *op;
 628                 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 629                 unsigned long mfn = pfn_to_mfn(PFN_DOWN(cr3));
 630
 631                 op = mcs.args;
 632                 op->cmd = MMUEXT_NEW_BASEPTR;
 633                 op->arg1.mfn = mfn;
 634
 635                 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 636
 637                 xen_mc_issue(PARAVIRT_LAZY_CPU);
 638         }
 639 }
 640
 641 /* Early in boot, while setting up the initial pagetable, assume
 642    everything is pinned. */
 643 static __init void xen_alloc_pt_init(struct mm_struct *mm, u32 pfn)
 644 {
 645         BUG_ON(mem_map);        /* should only be used early */
 646         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 647 }
 648
 649 /* This needs to make sure the new pte page is pinned iff its being
 650    attached to a pinned pagetable. */
 651 static void xen_alloc_pt(struct mm_struct *mm, u32 pfn)
 652 {
 653         struct page *page = pfn_to_page(pfn);
 654
 655         if (PagePinned(virt_to_page(mm->pgd))) {
 656                 SetPagePinned(page);
 657
 658                 if (!PageHighMem(page))
 659                         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 660                 else
 661                         /* make sure there are no stray mappings of
 662                            this page */
 663                         kmap_flush_unused();
 664         }
 665 }
 666
 667 /* This should never happen until we're OK to use struct page */
 668 static void xen_release_pt(u32 pfn)
 669 {
 670         struct page *page = pfn_to_page(pfn);
 671
 672         if (PagePinned(page)) {
 673                 if (!PageHighMem(page))
 674                         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 675         }
 676 }
 677
 678 #ifdef CONFIG_HIGHPTE
 679 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
 680 {
 681         pgprot_t prot = PAGE_KERNEL;
 682
 683         if (PagePinned(page))
 684                 prot = PAGE_KERNEL_RO;
 685
 686         if (0 && PageHighMem(page))
 687                 printk("mapping highpte %lx type %d prot %s\n",
 688                        page_to_pfn(page), type,
 689                        (unsigned long)pgprot_val(prot) & _PAGE_RW ? "WRITE" : "READ");
 690
 691         return kmap_atomic_prot(page, type, prot);
 692 }
 693 #endif
 694
 695 static __init pte_t mask_rw_pte(pte_t *ptep, pte_t pte)
 696 {
 697         /* If there's an existing pte, then don't allow _PAGE_RW to be set */
 698         if (pte_val_ma(*ptep) & _PAGE_PRESENT)
 699                 pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
 700                                pte_val_ma(pte));
 701
 702         return pte;
 703 }
 704
 705 /* Init-time set_pte while constructing initial pagetables, which
 706    doesn't allow RO pagetable pages to be remapped RW */
 707 static __init void xen_set_pte_init(pte_t *ptep, pte_t pte)
 708 {
 709         pte = mask_rw_pte(ptep, pte);
 710
 711         xen_set_pte(ptep, pte);
 712 }
 713
 714 static __init void xen_pagetable_setup_start(pgd_t *base)
 715 {
 716         pgd_t *xen_pgd = (pgd_t *)xen_start_info->pt_base;
 717
 718         /* special set_pte for pagetable initialization */
 719         pv_mmu_ops.set_pte = xen_set_pte_init;
 720
 721         init_mm.pgd = base;
 722         /*
 723          * copy top-level of Xen-supplied pagetable into place.  For
 724          * !PAE we can use this as-is, but for PAE it is a stand-in
 725          * while we copy the pmd pages.
 726          */
 727         memcpy(base, xen_pgd, PTRS_PER_PGD * sizeof(pgd_t));
 728
 729         if (PTRS_PER_PMD > 1) {
 730                 int i;
 731                 /*
 732                  * For PAE, need to allocate new pmds, rather than
 733                  * share Xen's, since Xen doesn't like pmd's being
 734                  * shared between address spaces.
 735                  */
 736                 for (i = 0; i < PTRS_PER_PGD; i++) {
 737                         if (pgd_val_ma(xen_pgd[i]) & _PAGE_PRESENT) {
 738                                 pmd_t *pmd = (pmd_t *)alloc_bootmem_low_pages(PAGE_SIZE);
 739
 740                                 memcpy(pmd, (void *)pgd_page_vaddr(xen_pgd[i]),
 741                                        PAGE_SIZE);
 742
 743                                 make_lowmem_page_readonly(pmd);
 744
 745                                 set_pgd(&base[i], __pgd(1 + __pa(pmd)));
 746                         } else
 747                                 pgd_clear(&base[i]);
 748                 }
 749         }
 750
 751         /* make sure zero_page is mapped RO so we can use it in pagetables */
 752         make_lowmem_page_readonly(empty_zero_page);
 753         make_lowmem_page_readonly(base);
 754         /*
 755          * Switch to new pagetable.  This is done before
 756          * pagetable_init has done anything so that the new pages
 757          * added to the table can be prepared properly for Xen.
 758          */
 759         xen_write_cr3(__pa(base));
 760 }
 761
 762 static __init void xen_pagetable_setup_done(pgd_t *base)
 763 {
 764         /* This will work as long as patching hasn't happened yet
 765            (which it hasn't) */
 766         pv_mmu_ops.alloc_pt = xen_alloc_pt;
 767         pv_mmu_ops.set_pte = xen_set_pte;
 768
 769         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 770                 /*
 771                  * Create a mapping for the shared info page.
 772                  * Should be set_fixmap(), but shared_info is a machine
 773                  * address with no corresponding pseudo-phys address.
 774                  */
 775                 set_pte_mfn(fix_to_virt(FIX_PARAVIRT_BOOTMAP),
 776                             PFN_DOWN(xen_start_info->shared_info),
 777                             PAGE_KERNEL);
 778
 779                 HYPERVISOR_shared_info =
 780                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
 781
 782         } else
 783                 HYPERVISOR_shared_info =
 784                         (struct shared_info *)__va(xen_start_info->shared_info);
 785
 786         /* Actually pin the pagetable down, but we can't set PG_pinned
 787            yet because the page structures don't exist yet. */
 788         {
 789                 struct mmuext_op op;
 790 #ifdef CONFIG_X86_PAE
 791                 op.cmd = MMUEXT_PIN_L3_TABLE;
 792 #else
 793                 op.cmd = MMUEXT_PIN_L3_TABLE;
 794 #endif
 795                 op.arg1.mfn = pfn_to_mfn(PFN_DOWN(__pa(base)));
 796                 if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
 797                         BUG();
 798         }
 799 }
 800
 801 /* This is called once we have the cpu_possible_map */
 802 void __init xen_setup_vcpu_info_placement(void)
 803 {
 804         int cpu;
 805
 806         for_each_possible_cpu(cpu)
 807                 xen_vcpu_setup(cpu);
 808
 809         /* xen_vcpu_setup managed to place the vcpu_info within the
 810            percpu area for all cpus, so make use of it */
 811         if (have_vcpu_info_placement) {
 812                 printk(KERN_INFO "Xen: using vcpu_info placement\n");
 813
 814                 pv_irq_ops.save_fl = xen_save_fl_direct;
 815                 pv_irq_ops.restore_fl = xen_restore_fl_direct;
 816                 pv_irq_ops.irq_disable = xen_irq_disable_direct;
 817                 pv_irq_ops.irq_enable = xen_irq_enable_direct;
 818                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
 819                 pv_cpu_ops.iret = xen_iret_direct;
 820         }
 821 }
 822
 823 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
 824                           unsigned long addr, unsigned len)
 825 {
 826         char *start, *end, *reloc;
 827         unsigned ret;
 828
 829         start = end = reloc = NULL;
 830
 831 #define SITE(op, x)                                                     \
 832         case PARAVIRT_PATCH(op.x):                                      \
 833         if (have_vcpu_info_placement) {                                 \
 834                 start = (char *)xen_##x##_direct;                       \
 835                 end = xen_##x##_direct_end;                             \
 836                 reloc = xen_##x##_direct_reloc;                         \
 837         }                                                               \
 838         goto patch_site
 839
 840         switch (type) {
 841                 SITE(pv_irq_ops, irq_enable);
 842                 SITE(pv_irq_ops, irq_disable);
 843                 SITE(pv_irq_ops, save_fl);
 844                 SITE(pv_irq_ops, restore_fl);
 845 #undef SITE
 846
 847         patch_site:
 848                 if (start == NULL || (end-start) > len)
 849                         goto default_patch;
 850
 851                 ret = paravirt_patch_insns(insnbuf, len, start, end);
 852
 853                 /* Note: because reloc is assigned from something that
 854                    appears to be an array, gcc assumes it's non-null,
 855                    but doesn't know its relationship with start and
 856                    end. */
 857                 if (reloc > start && reloc < end) {
 858                         int reloc_off = reloc - start;
 859                         long *relocp = (long *)(insnbuf + reloc_off);
 860                         long delta = start - (char *)addr;
 861
 862                         *relocp += delta;
 863                 }
 864                 break;
 865
 866         default_patch:
 867         default:
 868                 ret = paravirt_patch_default(type, clobbers, insnbuf,
 869                                              addr, len);
 870                 break;
 871         }
 872
 873         return ret;
 874 }
 875
 876 static const struct pv_info xen_info __initdata = {
 877         .paravirt_enabled = 1,
 878         .shared_kernel_pmd = 0,
 879
 880         .name = "Xen",
 881 };
 882
 883 static const struct pv_init_ops xen_init_ops __initdata = {
 884         .patch = xen_patch,
 885
 886         .banner = xen_banner,
 887         .memory_setup = xen_memory_setup,
 888         .arch_setup = xen_arch_setup,
 889         .post_allocator_init = xen_mark_init_mm_pinned,
 890 };
 891
 892 static const struct pv_time_ops xen_time_ops __initdata = {
 893         .time_init = xen_time_init,
 894
 895         .set_wallclock = xen_set_wallclock,
 896         .get_wallclock = xen_get_wallclock,
 897         .get_cpu_khz = xen_cpu_khz,
 898         .sched_clock = xen_sched_clock,
 899 };
 900
 901 static const struct pv_cpu_ops xen_cpu_ops __initdata = {
 902         .cpuid = xen_cpuid,
 903
 904         .set_debugreg = xen_set_debugreg,
 905         .get_debugreg = xen_get_debugreg,
 906
 907         .clts = native_clts,
 908
 909         .read_cr0 = native_read_cr0,
 910         .write_cr0 = native_write_cr0,
 911
 912         .read_cr4 = native_read_cr4,
 913         .read_cr4_safe = native_read_cr4_safe,
 914         .write_cr4 = xen_write_cr4,
 915
 916         .wbinvd = native_wbinvd,
 917
 918         .read_msr = native_read_msr_safe,
 919         .write_msr = native_write_msr_safe,
 920         .read_tsc = native_read_tsc,
 921         .read_pmc = native_read_pmc,
 922
 923         .iret = (void *)&hypercall_page[__HYPERVISOR_iret],
 924         .irq_enable_sysexit = NULL,  /* never called */
 925
 926         .load_tr_desc = paravirt_nop,
 927         .set_ldt = xen_set_ldt,
 928         .load_gdt = xen_load_gdt,
 929         .load_idt = xen_load_idt,
 930         .load_tls = xen_load_tls,
 931
 932         .store_gdt = native_store_gdt,
 933         .store_idt = native_store_idt,
 934         .store_tr = xen_store_tr,
 935
 936         .write_ldt_entry = xen_write_ldt_entry,
 937         .write_gdt_entry = xen_write_gdt_entry,
 938         .write_idt_entry = xen_write_idt_entry,
 939         .load_esp0 = xen_load_esp0,
 940
 941         .set_iopl_mask = xen_set_iopl_mask,
 942         .io_delay = xen_io_delay,
 943
 944         .lazy_mode = {
 945                 .enter = paravirt_enter_lazy_cpu,
 946                 .leave = xen_leave_lazy,
 947         },
 948 };
 949
 950 static const struct pv_irq_ops xen_irq_ops __initdata = {
 951         .init_IRQ = xen_init_IRQ,
 952         .save_fl = xen_save_fl,
 953         .restore_fl = xen_restore_fl,
 954         .irq_disable = xen_irq_disable,
 955         .irq_enable = xen_irq_enable,
 956         .safe_halt = xen_safe_halt,
 957         .halt = xen_halt,
 958 };
 959
 960 static const struct pv_apic_ops xen_apic_ops __initdata = {
 961 #ifdef CONFIG_X86_LOCAL_APIC
 962         .apic_write = xen_apic_write,
 963         .apic_write_atomic = xen_apic_write,
 964         .apic_read = xen_apic_read,
 965         .setup_boot_clock = paravirt_nop,
 966         .setup_secondary_clock = paravirt_nop,
 967         .startup_ipi_hook = paravirt_nop,
 968 #endif
 969 };
 970
 971 static const struct pv_mmu_ops xen_mmu_ops __initdata = {
 972         .pagetable_setup_start = xen_pagetable_setup_start,
 973         .pagetable_setup_done = xen_pagetable_setup_done,
 974
 975         .read_cr2 = xen_read_cr2,
 976         .write_cr2 = xen_write_cr2,
 977
 978         .read_cr3 = xen_read_cr3,
 979         .write_cr3 = xen_write_cr3,
 980
 981         .flush_tlb_user = xen_flush_tlb,
 982         .flush_tlb_kernel = xen_flush_tlb,
 983         .flush_tlb_single = xen_flush_tlb_single,
 984         .flush_tlb_others = xen_flush_tlb_others,
 985
 986         .pte_update = paravirt_nop,
 987         .pte_update_defer = paravirt_nop,
 988
 989         .alloc_pt = xen_alloc_pt_init,
 990         .release_pt = xen_release_pt,
 991         .alloc_pd = paravirt_nop,
 992         .alloc_pd_clone = paravirt_nop,
 993         .release_pd = paravirt_nop,
 994
 995 #ifdef CONFIG_HIGHPTE
 996         .kmap_atomic_pte = xen_kmap_atomic_pte,
 997 #endif
 998
 999         .set_pte = NULL,        /* see xen_pagetable_setup_* */
1000         .set_pte_at = xen_set_pte_at,
1001         .set_pmd = xen_set_pmd,
1002
1003         .pte_val = xen_pte_val,
1004         .pgd_val = xen_pgd_val,
1005
1006         .make_pte = xen_make_pte,
1007         .make_pgd = xen_make_pgd,
1008
1009 #ifdef CONFIG_X86_PAE
1010         .set_pte_atomic = xen_set_pte_atomic,
1011         .set_pte_present = xen_set_pte_at,
1012         .set_pud = xen_set_pud,
1013         .pte_clear = xen_pte_clear,
1014         .pmd_clear = xen_pmd_clear,
1015
1016         .make_pmd = xen_make_pmd,
1017         .pmd_val = xen_pmd_val,
1018 #endif  /* PAE */
1019
1020         .activate_mm = xen_activate_mm,
1021         .dup_mmap = xen_dup_mmap,
1022         .exit_mmap = xen_exit_mmap,
1023
1024         .lazy_mode = {
1025                 .enter = paravirt_enter_lazy_mmu,
1026                 .leave = xen_leave_lazy,
1027         },
1028 };
1029
1030 #ifdef CONFIG_SMP
1031 static const struct smp_ops xen_smp_ops __initdata = {
1032         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
1033         .smp_prepare_cpus = xen_smp_prepare_cpus,
1034         .cpu_up = xen_cpu_up,
1035         .smp_cpus_done = xen_smp_cpus_done,
1036
1037         .smp_send_stop = xen_smp_send_stop,
1038         .smp_send_reschedule = xen_smp_send_reschedule,
1039         .smp_call_function_mask = xen_smp_call_function_mask,
1040 };
1041 #endif  /* CONFIG_SMP */
1042
1043 static void xen_reboot(int reason)
1044 {
1045 #ifdef CONFIG_SMP
1046         smp_send_stop();
1047 #endif
1048
1049         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, reason))
1050                 BUG();
1051 }
1052
1053 static void xen_restart(char *msg)
1054 {
1055         xen_reboot(SHUTDOWN_reboot);
1056 }
1057
1058 static void xen_emergency_restart(void)
1059 {
1060         xen_reboot(SHUTDOWN_reboot);
1061 }
1062
1063 static void xen_machine_halt(void)
1064 {
1065         xen_reboot(SHUTDOWN_poweroff);
1066 }
1067
1068 static void xen_crash_shutdown(struct pt_regs *regs)
1069 {
1070         xen_reboot(SHUTDOWN_crash);
1071 }
1072
1073 static const struct machine_ops __initdata xen_machine_ops = {
1074         .restart = xen_restart,
1075         .halt = xen_machine_halt,
1076         .power_off = xen_machine_halt,
1077         .shutdown = xen_machine_halt,
1078         .crash_shutdown = xen_crash_shutdown,
1079         .emergency_restart = xen_emergency_restart,
1080 };
1081
1082
1083 /* First C function to be called on Xen boot */
1084 asmlinkage void __init xen_start_kernel(void)
1085 {
1086         pgd_t *pgd;
1087
1088         if (!xen_start_info)
1089                 return;
1090
1091         BUG_ON(memcmp(xen_start_info->magic, "xen-3.0", 7) != 0);
1092
1093         /* Install Xen paravirt ops */
1094         pv_info = xen_info;
1095         pv_init_ops = xen_init_ops;
1096         pv_time_ops = xen_time_ops;
1097         pv_cpu_ops = xen_cpu_ops;
1098         pv_irq_ops = xen_irq_ops;
1099         pv_apic_ops = xen_apic_ops;
1100         pv_mmu_ops = xen_mmu_ops;
1101
1102         machine_ops = xen_machine_ops;
1103
1104 #ifdef CONFIG_SMP
1105         smp_ops = xen_smp_ops;
1106 #endif
1107
1108         xen_setup_features();
1109
1110         /* Get mfn list */
1111         if (!xen_feature(XENFEAT_auto_translated_physmap))
1112                 phys_to_machine_mapping = (unsigned long *)xen_start_info->mfn_list;
1113
1114         pgd = (pgd_t *)xen_start_info->pt_base;
1115
1116         init_pg_tables_end = __pa(pgd) + xen_start_info->nr_pt_frames*PAGE_SIZE;
1117
1118         init_mm.pgd = pgd; /* use the Xen pagetables to start */
1119
1120         /* keep using Xen gdt for now; no urgent need to change it */
1121
1122         x86_write_percpu(xen_cr3, __pa(pgd));
1123
1124 #ifdef CONFIG_SMP
1125         /* Don't do the full vcpu_info placement stuff until we have a
1126            possible map. */
1127         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1128 #else
1129         /* May as well do it now, since there's no good time to call
1130            it later on UP. */
1131         xen_setup_vcpu_info_placement();
1132 #endif
1133
1134         pv_info.kernel_rpl = 1;
1135         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1136                 pv_info.kernel_rpl = 0;
1137
1138         /* set the limit of our address space */
1139         reserve_top_address(-HYPERVISOR_VIRT_START + 2 * PAGE_SIZE);
1140
1141         /* set up basic CPUID stuff */
1142         cpu_detect(&new_cpu_data);
1143         new_cpu_data.hard_math = 1;
1144         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1145
1146         /* Poke various useful things into boot_params */
1147         LOADER_TYPE = (9 << 4) | 0;
1148         INITRD_START = xen_start_info->mod_start ? __pa(xen_start_info->mod_start) : 0;
1149         INITRD_SIZE = xen_start_info->mod_len;
1150
1151         /* Start the world */
1152         start_kernel();
1153 }