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/smp.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
  37 #include <asm/paravirt.h>
  38 #include <asm/page.h>
  39 #include <asm/xen/hypercall.h>
  40 #include <asm/xen/hypervisor.h>
  41 #include <asm/fixmap.h>
  42 #include <asm/processor.h>
  43 #include <asm/setup.h>
  44 #include <asm/desc.h>
  45 #include <asm/pgtable.h>
  46 #include <asm/tlbflush.h>
  47 #include <asm/reboot.h>
  48
  49 #include "xen-ops.h"
  50 #include "mmu.h"
  51 #include "multicalls.h"
  52
  53 EXPORT_SYMBOL_GPL(hypercall_page);
  54
  55 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
  56 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
  57 DEFINE_PER_CPU(unsigned long, xen_cr3);
  58
  59 struct start_info *xen_start_info;
  60 EXPORT_SYMBOL_GPL(xen_start_info);
  61
  62 static /* __initdata */ struct shared_info dummy_shared_info;
  63
  64 /*
  65  * Point at some empty memory to start with. We map the real shared_info
  66  * page as soon as fixmap is up and running.
  67  */
  68 struct shared_info *HYPERVISOR_shared_info = (void *)&dummy_shared_info;
  69
  70 /*
  71  * Flag to determine whether vcpu info placement is available on all
  72  * VCPUs.  We assume it is to start with, and then set it to zero on
  73  * the first failure.  This is because it can succeed on some VCPUs
  74  * and not others, since it can involve hypervisor memory allocation,
  75  * or because the guest failed to guarantee all the appropriate
  76  * constraints on all VCPUs (ie buffer can't cross a page boundary).
  77  *
  78  * Note that any particular CPU may be using a placed vcpu structure,
  79  * but we can only optimise if the all are.
  80  *
  81  * 0: not available, 1: available
  82  */
  83 static int have_vcpu_info_placement = 1;
  84
  85 static void __init xen_vcpu_setup(int cpu)
  86 {
  87         struct vcpu_register_vcpu_info info;
  88         int err;
  89         struct vcpu_info *vcpup;
  90
  91         per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
  92
  93         if (!have_vcpu_info_placement)
  94                 return;         /* already tested, not available */
  95
  96         vcpup = &per_cpu(xen_vcpu_info, cpu);
  97
  98         info.mfn = virt_to_mfn(vcpup);
  99         info.offset = offset_in_page(vcpup);
 100
 101         printk(KERN_DEBUG "trying to map vcpu_info %d at %p, mfn %x, offset %d\n",
 102                cpu, vcpup, info.mfn, info.offset);
 103
 104         /* Check to see if the hypervisor will put the vcpu_info
 105            structure where we want it, which allows direct access via
 106            a percpu-variable. */
 107         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
 108
 109         if (err) {
 110                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
 111                 have_vcpu_info_placement = 0;
 112         } else {
 113                 /* This cpu is using the registered vcpu info, even if
 114                    later ones fail to. */
 115                 per_cpu(xen_vcpu, cpu) = vcpup;
 116
 117                 printk(KERN_DEBUG "cpu %d using vcpu_info at %p\n",
 118                        cpu, vcpup);
 119         }
 120 }
 121
 122 static void __init xen_banner(void)
 123 {
 124         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
 125                pv_info.name);
 126         printk(KERN_INFO "Hypervisor signature: %s\n", xen_start_info->magic);
 127 }
 128
 129 static void xen_cpuid(unsigned int *eax, unsigned int *ebx,
 130                       unsigned int *ecx, unsigned int *edx)
 131 {
 132         unsigned maskedx = ~0;
 133
 134         /*
 135          * Mask out inconvenient features, to try and disable as many
 136          * unsupported kernel subsystems as possible.
 137          */
 138         if (*eax == 1)
 139                 maskedx = ~((1 << X86_FEATURE_APIC) |  /* disable APIC */
 140                             (1 << X86_FEATURE_ACPI) |  /* disable ACPI */
 141                             (1 << X86_FEATURE_ACC));   /* thermal monitoring */
 142
 143         asm(XEN_EMULATE_PREFIX "cpuid"
 144                 : "=a" (*eax),
 145                   "=b" (*ebx),
 146                   "=c" (*ecx),
 147                   "=d" (*edx)
 148                 : "0" (*eax), "2" (*ecx));
 149         *edx &= maskedx;
 150 }
 151
 152 static void xen_set_debugreg(int reg, unsigned long val)
 153 {
 154         HYPERVISOR_set_debugreg(reg, val);
 155 }
 156
 157 static unsigned long xen_get_debugreg(int reg)
 158 {
 159         return HYPERVISOR_get_debugreg(reg);
 160 }
 161
 162 static unsigned long xen_save_fl(void)
 163 {
 164         struct vcpu_info *vcpu;
 165         unsigned long flags;
 166
 167         vcpu = x86_read_percpu(xen_vcpu);
 168
 169         /* flag has opposite sense of mask */
 170         flags = !vcpu->evtchn_upcall_mask;
 171
 172         /* convert to IF type flag
 173            -0 -> 0x00000000
 174            -1 -> 0xffffffff
 175         */
 176         return (-flags) & X86_EFLAGS_IF;
 177 }
 178
 179 static void xen_restore_fl(unsigned long flags)
 180 {
 181         struct vcpu_info *vcpu;
 182
 183         /* convert from IF type flag */
 184         flags = !(flags & X86_EFLAGS_IF);
 185
 186         /* There's a one instruction preempt window here.  We need to
 187            make sure we're don't switch CPUs between getting the vcpu
 188            pointer and updating the mask. */
 189         preempt_disable();
 190         vcpu = x86_read_percpu(xen_vcpu);
 191         vcpu->evtchn_upcall_mask = flags;
 192         preempt_enable_no_resched();
 193
 194         /* Doesn't matter if we get preempted here, because any
 195            pending event will get dealt with anyway. */
 196
 197         if (flags == 0) {
 198                 preempt_check_resched();
 199                 barrier(); /* unmask then check (avoid races) */
 200                 if (unlikely(vcpu->evtchn_upcall_pending))
 201                         force_evtchn_callback();
 202         }
 203 }
 204
 205 static void xen_irq_disable(void)
 206 {
 207         /* There's a one instruction preempt window here.  We need to
 208            make sure we're don't switch CPUs between getting the vcpu
 209            pointer and updating the mask. */
 210         preempt_disable();
 211         x86_read_percpu(xen_vcpu)->evtchn_upcall_mask = 1;
 212         preempt_enable_no_resched();
 213 }
 214
 215 static void xen_irq_enable(void)
 216 {
 217         struct vcpu_info *vcpu;
 218
 219         /* There's a one instruction preempt window here.  We need to
 220            make sure we're don't switch CPUs between getting the vcpu
 221            pointer and updating the mask. */
 222         preempt_disable();
 223         vcpu = x86_read_percpu(xen_vcpu);
 224         vcpu->evtchn_upcall_mask = 0;
 225         preempt_enable_no_resched();
 226
 227         /* Doesn't matter if we get preempted here, because any
 228            pending event will get dealt with anyway. */
 229
 230         barrier(); /* unmask then check (avoid races) */
 231         if (unlikely(vcpu->evtchn_upcall_pending))
 232                 force_evtchn_callback();
 233 }
 234
 235 static void xen_safe_halt(void)
 236 {
 237         /* Blocking includes an implicit local_irq_enable(). */
 238         if (HYPERVISOR_sched_op(SCHEDOP_block, 0) != 0)
 239                 BUG();
 240 }
 241
 242 static void xen_halt(void)
 243 {
 244         if (irqs_disabled())
 245                 HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
 246         else
 247                 xen_safe_halt();
 248 }
 249
 250 static void xen_leave_lazy(void)
 251 {
 252         paravirt_leave_lazy(paravirt_get_lazy_mode());
 253         xen_mc_flush();
 254 }
 255
 256 static unsigned long xen_store_tr(void)
 257 {
 258         return 0;
 259 }
 260
 261 static void xen_set_ldt(const void *addr, unsigned entries)
 262 {
 263         unsigned long linear_addr = (unsigned long)addr;
 264         struct mmuext_op *op;
 265         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 266
 267         op = mcs.args;
 268         op->cmd = MMUEXT_SET_LDT;
 269         if (linear_addr) {
 270                 /* ldt my be vmalloced, use arbitrary_virt_to_machine */
 271                 xmaddr_t maddr;
 272                 maddr = arbitrary_virt_to_machine((unsigned long)addr);
 273                 linear_addr = (unsigned long)maddr.maddr;
 274         }
 275         op->arg1.linear_addr = linear_addr;
 276         op->arg2.nr_ents = entries;
 277
 278         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 279
 280         xen_mc_issue(PARAVIRT_LAZY_CPU);
 281 }
 282
 283 static void xen_load_gdt(const struct Xgt_desc_struct *dtr)
 284 {
 285         unsigned long *frames;
 286         unsigned long va = dtr->address;
 287         unsigned int size = dtr->size + 1;
 288         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
 289         int f;
 290         struct multicall_space mcs;
 291
 292         /* A GDT can be up to 64k in size, which corresponds to 8192
 293            8-byte entries, or 16 4k pages.. */
 294
 295         BUG_ON(size > 65536);
 296         BUG_ON(va & ~PAGE_MASK);
 297
 298         mcs = xen_mc_entry(sizeof(*frames) * pages);
 299         frames = mcs.args;
 300
 301         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
 302                 frames[f] = virt_to_mfn(va);
 303                 make_lowmem_page_readonly((void *)va);
 304         }
 305
 306         MULTI_set_gdt(mcs.mc, frames, size / sizeof(struct desc_struct));
 307
 308         xen_mc_issue(PARAVIRT_LAZY_CPU);
 309 }
 310
 311 static void load_TLS_descriptor(struct thread_struct *t,
 312                                 unsigned int cpu, unsigned int i)
 313 {
 314         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
 315         xmaddr_t maddr = virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
 316         struct multicall_space mc = __xen_mc_entry(0);
 317
 318         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
 319 }
 320
 321 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
 322 {
 323         xen_mc_batch();
 324
 325         load_TLS_descriptor(t, cpu, 0);
 326         load_TLS_descriptor(t, cpu, 1);
 327         load_TLS_descriptor(t, cpu, 2);
 328
 329         xen_mc_issue(PARAVIRT_LAZY_CPU);
 330
 331         /*
 332          * XXX sleazy hack: If we're being called in a lazy-cpu zone,
 333          * it means we're in a context switch, and %gs has just been
 334          * saved.  This means we can zero it out to prevent faults on
 335          * exit from the hypervisor if the next process has no %gs.
 336          * Either way, it has been saved, and the new value will get
 337          * loaded properly.  This will go away as soon as Xen has been
 338          * modified to not save/restore %gs for normal hypercalls.
 339          */
 340         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU)
 341                 loadsegment(gs, 0);
 342 }
 343
 344 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
 345                                 u32 low, u32 high)
 346 {
 347         unsigned long lp = (unsigned long)&dt[entrynum];
 348         xmaddr_t mach_lp = virt_to_machine(lp);
 349         u64 entry = (u64)high << 32 | low;
 350
 351         preempt_disable();
 352
 353         xen_mc_flush();
 354         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
 355                 BUG();
 356
 357         preempt_enable();
 358 }
 359
 360 static int cvt_gate_to_trap(int vector, u32 low, u32 high,
 361                             struct trap_info *info)
 362 {
 363         u8 type, dpl;
 364
 365         type = (high >> 8) & 0x1f;
 366         dpl = (high >> 13) & 3;
 367
 368         if (type != 0xf && type != 0xe)
 369                 return 0;
 370
 371         info->vector = vector;
 372         info->address = (high & 0xffff0000) | (low & 0x0000ffff);
 373         info->cs = low >> 16;
 374         info->flags = dpl;
 375         /* interrupt gates clear IF */
 376         if (type == 0xe)
 377                 info->flags |= 4;
 378
 379         return 1;
 380 }
 381
 382 /* Locations of each CPU's IDT */
 383 static DEFINE_PER_CPU(struct Xgt_desc_struct, idt_desc);
 384
 385 /* Set an IDT entry.  If the entry is part of the current IDT, then
 386    also update Xen. */
 387 static void xen_write_idt_entry(struct desc_struct *dt, int entrynum,
 388                                 u32 low, u32 high)
 389 {
 390         unsigned long p = (unsigned long)&dt[entrynum];
 391         unsigned long start, end;
 392
 393         preempt_disable();
 394
 395         start = __get_cpu_var(idt_desc).address;
 396         end = start + __get_cpu_var(idt_desc).size + 1;
 397
 398         xen_mc_flush();
 399
 400         write_dt_entry(dt, entrynum, low, high);
 401
 402         if (p >= start && (p + 8) <= end) {
 403                 struct trap_info info[2];
 404
 405                 info[1].address = 0;
 406
 407                 if (cvt_gate_to_trap(entrynum, low, high, &info[0]))
 408                         if (HYPERVISOR_set_trap_table(info))
 409                                 BUG();
 410         }
 411
 412         preempt_enable();
 413 }
 414
 415 static void xen_convert_trap_info(const struct Xgt_desc_struct *desc,
 416                                   struct trap_info *traps)
 417 {
 418         unsigned in, out, count;
 419
 420         count = (desc->size+1) / 8;
 421         BUG_ON(count > 256);
 422
 423         for (in = out = 0; in < count; in++) {
 424                 const u32 *entry = (u32 *)(desc->address + in * 8);
 425
 426                 if (cvt_gate_to_trap(in, entry[0], entry[1], &traps[out]))
 427                         out++;
 428         }
 429         traps[out].address = 0;
 430 }
 431
 432 void xen_copy_trap_info(struct trap_info *traps)
 433 {
 434         const struct Xgt_desc_struct *desc = &__get_cpu_var(idt_desc);
 435
 436         xen_convert_trap_info(desc, traps);
 437 }
 438
 439 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
 440    hold a spinlock to protect the static traps[] array (static because
 441    it avoids allocation, and saves stack space). */
 442 static void xen_load_idt(const struct Xgt_desc_struct *desc)
 443 {
 444         static DEFINE_SPINLOCK(lock);
 445         static struct trap_info traps[257];
 446
 447         spin_lock(&lock);
 448
 449         __get_cpu_var(idt_desc) = *desc;
 450
 451         xen_convert_trap_info(desc, traps);
 452
 453         xen_mc_flush();
 454         if (HYPERVISOR_set_trap_table(traps))
 455                 BUG();
 456
 457         spin_unlock(&lock);
 458 }
 459
 460 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
 461    they're handled differently. */
 462 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
 463                                 u32 low, u32 high)
 464 {
 465         preempt_disable();
 466
 467         switch ((high >> 8) & 0xff) {
 468         case DESCTYPE_LDT:
 469         case DESCTYPE_TSS:
 470                 /* ignore */
 471                 break;
 472
 473         default: {
 474                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
 475                 u64 desc = (u64)high << 32 | low;
 476
 477                 xen_mc_flush();
 478                 if (HYPERVISOR_update_descriptor(maddr.maddr, desc))
 479                         BUG();
 480         }
 481
 482         }
 483
 484         preempt_enable();
 485 }
 486
 487 static void xen_load_esp0(struct tss_struct *tss,
 488                           struct thread_struct *thread)
 489 {
 490         struct multicall_space mcs = xen_mc_entry(0);
 491         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->esp0);
 492         xen_mc_issue(PARAVIRT_LAZY_CPU);
 493 }
 494
 495 static void xen_set_iopl_mask(unsigned mask)
 496 {
 497         struct physdev_set_iopl set_iopl;
 498
 499         /* Force the change at ring 0. */
 500         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
 501         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
 502 }
 503
 504 static void xen_io_delay(void)
 505 {
 506 }
 507
 508 #ifdef CONFIG_X86_LOCAL_APIC
 509 static unsigned long xen_apic_read(unsigned long reg)
 510 {
 511         return 0;
 512 }
 513
 514 static void xen_apic_write(unsigned long reg, unsigned long val)
 515 {
 516         /* Warn to see if there's any stray references */
 517         WARN_ON(1);
 518 }
 519 #endif
 520
 521 static void xen_flush_tlb(void)
 522 {
 523         struct mmuext_op *op;
 524         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 525
 526         op = mcs.args;
 527         op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
 528         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 529
 530         xen_mc_issue(PARAVIRT_LAZY_MMU);
 531 }
 532
 533 static void xen_flush_tlb_single(unsigned long addr)
 534 {
 535         struct mmuext_op *op;
 536         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 537
 538         op = mcs.args;
 539         op->cmd = MMUEXT_INVLPG_LOCAL;
 540         op->arg1.linear_addr = addr & PAGE_MASK;
 541         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 542
 543         xen_mc_issue(PARAVIRT_LAZY_MMU);
 544 }
 545
 546 static void xen_flush_tlb_others(const cpumask_t *cpus, struct mm_struct *mm,
 547                                  unsigned long va)
 548 {
 549         struct {
 550                 struct mmuext_op op;
 551                 cpumask_t mask;
 552         } *args;
 553         cpumask_t cpumask = *cpus;
 554         struct multicall_space mcs;
 555
 556         /*
 557          * A couple of (to be removed) sanity checks:
 558          *
 559          * - current CPU must not be in mask
 560          * - mask must exist :)
 561          */
 562         BUG_ON(cpus_empty(cpumask));
 563         BUG_ON(cpu_isset(smp_processor_id(), cpumask));
 564         BUG_ON(!mm);
 565
 566         /* If a CPU which we ran on has gone down, OK. */
 567         cpus_and(cpumask, cpumask, cpu_online_map);
 568         if (cpus_empty(cpumask))
 569                 return;
 570
 571         mcs = xen_mc_entry(sizeof(*args));
 572         args = mcs.args;
 573         args->mask = cpumask;
 574         args->op.arg2.vcpumask = &args->mask;
 575
 576         if (va == TLB_FLUSH_ALL) {
 577                 args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
 578         } else {
 579                 args->op.cmd = MMUEXT_INVLPG_MULTI;
 580                 args->op.arg1.linear_addr = va;
 581         }
 582
 583         MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
 584
 585         xen_mc_issue(PARAVIRT_LAZY_MMU);
 586 }
 587
 588 static void xen_write_cr2(unsigned long cr2)
 589 {
 590         x86_read_percpu(xen_vcpu)->arch.cr2 = cr2;
 591 }
 592
 593 static unsigned long xen_read_cr2(void)
 594 {
 595         return x86_read_percpu(xen_vcpu)->arch.cr2;
 596 }
 597
 598 static unsigned long xen_read_cr2_direct(void)
 599 {
 600         return x86_read_percpu(xen_vcpu_info.arch.cr2);
 601 }
 602
 603 static void xen_write_cr4(unsigned long cr4)
 604 {
 605         /* Just ignore cr4 changes; Xen doesn't allow us to do
 606            anything anyway. */
 607 }
 608
 609 static unsigned long xen_read_cr3(void)
 610 {
 611         return x86_read_percpu(xen_cr3);
 612 }
 613
 614 static void xen_write_cr3(unsigned long cr3)
 615 {
 616         BUG_ON(preemptible());
 617
 618         if (cr3 == x86_read_percpu(xen_cr3)) {
 619                 /* just a simple tlb flush */
 620                 xen_flush_tlb();
 621                 return;
 622         }
 623
 624         x86_write_percpu(xen_cr3, cr3);
 625
 626
 627         {
 628                 struct mmuext_op *op;
 629                 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 630                 unsigned long mfn = pfn_to_mfn(PFN_DOWN(cr3));
 631
 632                 op = mcs.args;
 633                 op->cmd = MMUEXT_NEW_BASEPTR;
 634                 op->arg1.mfn = mfn;
 635
 636                 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 637
 638                 xen_mc_issue(PARAVIRT_LAZY_CPU);
 639         }
 640 }
 641
 642 /* Early in boot, while setting up the initial pagetable, assume
 643    everything is pinned. */
 644 static __init void xen_alloc_pt_init(struct mm_struct *mm, u32 pfn)
 645 {
 646         BUG_ON(mem_map);        /* should only be used early */
 647         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 648 }
 649
 650 /* This needs to make sure the new pte page is pinned iff its being
 651    attached to a pinned pagetable. */
 652 static void xen_alloc_pt(struct mm_struct *mm, u32 pfn)
 653 {
 654         struct page *page = pfn_to_page(pfn);
 655
 656         if (PagePinned(virt_to_page(mm->pgd))) {
 657                 SetPagePinned(page);
 658
 659                 if (!PageHighMem(page))
 660                         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 661                 else
 662                         /* make sure there are no stray mappings of
 663                            this page */
 664                         kmap_flush_unused();
 665         }
 666 }
 667
 668 /* This should never happen until we're OK to use struct page */
 669 static void xen_release_pt(u32 pfn)
 670 {
 671         struct page *page = pfn_to_page(pfn);
 672
 673         if (PagePinned(page)) {
 674                 if (!PageHighMem(page))
 675                         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 676         }
 677 }
 678
 679 #ifdef CONFIG_HIGHPTE
 680 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
 681 {
 682         pgprot_t prot = PAGE_KERNEL;
 683
 684         if (PagePinned(page))
 685                 prot = PAGE_KERNEL_RO;
 686
 687         if (0 && PageHighMem(page))
 688                 printk("mapping highpte %lx type %d prot %s\n",
 689                        page_to_pfn(page), type,
 690                        (unsigned long)pgprot_val(prot) & _PAGE_RW ? "WRITE" : "READ");
 691
 692         return kmap_atomic_prot(page, type, prot);
 693 }
 694 #endif
 695
 696 static __init pte_t mask_rw_pte(pte_t *ptep, pte_t pte)
 697 {
 698         /* If there's an existing pte, then don't allow _PAGE_RW to be set */
 699         if (pte_val_ma(*ptep) & _PAGE_PRESENT)
 700                 pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
 701                                pte_val_ma(pte));
 702
 703         return pte;
 704 }
 705
 706 /* Init-time set_pte while constructing initial pagetables, which
 707    doesn't allow RO pagetable pages to be remapped RW */
 708 static __init void xen_set_pte_init(pte_t *ptep, pte_t pte)
 709 {
 710         pte = mask_rw_pte(ptep, pte);
 711
 712         xen_set_pte(ptep, pte);
 713 }
 714
 715 static __init void xen_pagetable_setup_start(pgd_t *base)
 716 {
 717         pgd_t *xen_pgd = (pgd_t *)xen_start_info->pt_base;
 718
 719         /* special set_pte for pagetable initialization */
 720         pv_mmu_ops.set_pte = xen_set_pte_init;
 721
 722         init_mm.pgd = base;
 723         /*
 724          * copy top-level of Xen-supplied pagetable into place.  For
 725          * !PAE we can use this as-is, but for PAE it is a stand-in
 726          * while we copy the pmd pages.
 727          */
 728         memcpy(base, xen_pgd, PTRS_PER_PGD * sizeof(pgd_t));
 729
 730         if (PTRS_PER_PMD > 1) {
 731                 int i;
 732                 /*
 733                  * For PAE, need to allocate new pmds, rather than
 734                  * share Xen's, since Xen doesn't like pmd's being
 735                  * shared between address spaces.
 736                  */
 737                 for (i = 0; i < PTRS_PER_PGD; i++) {
 738                         if (pgd_val_ma(xen_pgd[i]) & _PAGE_PRESENT) {
 739                                 pmd_t *pmd = (pmd_t *)alloc_bootmem_low_pages(PAGE_SIZE);
 740
 741                                 memcpy(pmd, (void *)pgd_page_vaddr(xen_pgd[i]),
 742                                        PAGE_SIZE);
 743
 744                                 make_lowmem_page_readonly(pmd);
 745
 746                                 set_pgd(&base[i], __pgd(1 + __pa(pmd)));
 747                         } else
 748                                 pgd_clear(&base[i]);
 749                 }
 750         }
 751
 752         /* make sure zero_page is mapped RO so we can use it in pagetables */
 753         make_lowmem_page_readonly(empty_zero_page);
 754         make_lowmem_page_readonly(base);
 755         /*
 756          * Switch to new pagetable.  This is done before
 757          * pagetable_init has done anything so that the new pages
 758          * added to the table can be prepared properly for Xen.
 759          */
 760         xen_write_cr3(__pa(base));
 761 }
 762
 763 static __init void xen_pagetable_setup_done(pgd_t *base)
 764 {
 765         /* This will work as long as patching hasn't happened yet
 766            (which it hasn't) */
 767         pv_mmu_ops.alloc_pt = xen_alloc_pt;
 768         pv_mmu_ops.set_pte = xen_set_pte;
 769
 770         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 771                 /*
 772                  * Create a mapping for the shared info page.
 773                  * Should be set_fixmap(), but shared_info is a machine
 774                  * address with no corresponding pseudo-phys address.
 775                  */
 776                 set_pte_mfn(fix_to_virt(FIX_PARAVIRT_BOOTMAP),
 777                             PFN_DOWN(xen_start_info->shared_info),
 778                             PAGE_KERNEL);
 779
 780                 HYPERVISOR_shared_info =
 781                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
 782
 783         } else
 784                 HYPERVISOR_shared_info =
 785                         (struct shared_info *)__va(xen_start_info->shared_info);
 786
 787         /* Actually pin the pagetable down, but we can't set PG_pinned
 788            yet because the page structures don't exist yet. */
 789         {
 790                 struct mmuext_op op;
 791 #ifdef CONFIG_X86_PAE
 792                 op.cmd = MMUEXT_PIN_L3_TABLE;
 793 #else
 794                 op.cmd = MMUEXT_PIN_L3_TABLE;
 795 #endif
 796                 op.arg1.mfn = pfn_to_mfn(PFN_DOWN(__pa(base)));
 797                 if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
 798                         BUG();
 799         }
 800 }
 801
 802 /* This is called once we have the cpu_possible_map */
 803 void __init xen_setup_vcpu_info_placement(void)
 804 {
 805         int cpu;
 806
 807         for_each_possible_cpu(cpu)
 808                 xen_vcpu_setup(cpu);
 809
 810         /* xen_vcpu_setup managed to place the vcpu_info within the
 811            percpu area for all cpus, so make use of it */
 812         if (have_vcpu_info_placement) {
 813                 printk(KERN_INFO "Xen: using vcpu_info placement\n");
 814
 815                 pv_irq_ops.save_fl = xen_save_fl_direct;
 816                 pv_irq_ops.restore_fl = xen_restore_fl_direct;
 817                 pv_irq_ops.irq_disable = xen_irq_disable_direct;
 818                 pv_irq_ops.irq_enable = xen_irq_enable_direct;
 819                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
 820                 pv_cpu_ops.iret = xen_iret_direct;
 821         }
 822 }
 823
 824 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
 825                           unsigned long addr, unsigned len)
 826 {
 827         char *start, *end, *reloc;
 828         unsigned ret;
 829
 830         start = end = reloc = NULL;
 831
 832 #define SITE(op, x)                                                     \
 833         case PARAVIRT_PATCH(op.x):                                      \
 834         if (have_vcpu_info_placement) {                                 \
 835                 start = (char *)xen_##x##_direct;                       \
 836                 end = xen_##x##_direct_end;                             \
 837                 reloc = xen_##x##_direct_reloc;                         \
 838         }                                                               \
 839         goto patch_site
 840
 841         switch (type) {
 842                 SITE(pv_irq_ops, irq_enable);
 843                 SITE(pv_irq_ops, irq_disable);
 844                 SITE(pv_irq_ops, save_fl);
 845                 SITE(pv_irq_ops, restore_fl);
 846 #undef SITE
 847
 848         patch_site:
 849                 if (start == NULL || (end-start) > len)
 850                         goto default_patch;
 851
 852                 ret = paravirt_patch_insns(insnbuf, len, start, end);
 853
 854                 /* Note: because reloc is assigned from something that
 855                    appears to be an array, gcc assumes it's non-null,
 856                    but doesn't know its relationship with start and
 857                    end. */
 858                 if (reloc > start && reloc < end) {
 859                         int reloc_off = reloc - start;
 860                         long *relocp = (long *)(insnbuf + reloc_off);
 861                         long delta = start - (char *)addr;
 862
 863                         *relocp += delta;
 864                 }
 865                 break;
 866
 867         default_patch:
 868         default:
 869                 ret = paravirt_patch_default(type, clobbers, insnbuf,
 870                                              addr, len);
 871                 break;
 872         }
 873
 874         return ret;
 875 }
 876
 877 static const struct pv_info xen_info __initdata = {
 878         .paravirt_enabled = 1,
 879         .shared_kernel_pmd = 0,
 880
 881         .name = "Xen",
 882 };
 883
 884 static const struct pv_init_ops xen_init_ops __initdata = {
 885         .patch = xen_patch,
 886
 887         .banner = xen_banner,
 888         .memory_setup = xen_memory_setup,
 889         .arch_setup = xen_arch_setup,
 890         .post_allocator_init = xen_mark_init_mm_pinned,
 891 };
 892
 893 static const struct pv_time_ops xen_time_ops __initdata = {
 894         .time_init = xen_time_init,
 895
 896         .set_wallclock = xen_set_wallclock,
 897         .get_wallclock = xen_get_wallclock,
 898         .get_cpu_khz = xen_cpu_khz,
 899         .sched_clock = xen_sched_clock,
 900 };
 901
 902 static const struct pv_cpu_ops xen_cpu_ops __initdata = {
 903         .cpuid = xen_cpuid,
 904
 905         .set_debugreg = xen_set_debugreg,
 906         .get_debugreg = xen_get_debugreg,
 907
 908         .clts = native_clts,
 909
 910         .read_cr0 = native_read_cr0,
 911         .write_cr0 = native_write_cr0,
 912
 913         .read_cr4 = native_read_cr4,
 914         .read_cr4_safe = native_read_cr4_safe,
 915         .write_cr4 = xen_write_cr4,
 916
 917         .wbinvd = native_wbinvd,
 918
 919         .read_msr = native_read_msr_safe,
 920         .write_msr = native_write_msr_safe,
 921         .read_tsc = native_read_tsc,
 922         .read_pmc = native_read_pmc,
 923
 924         .iret = (void *)&hypercall_page[__HYPERVISOR_iret],
 925         .irq_enable_sysexit = NULL,  /* never called */
 926
 927         .load_tr_desc = paravirt_nop,
 928         .set_ldt = xen_set_ldt,
 929         .load_gdt = xen_load_gdt,
 930         .load_idt = xen_load_idt,
 931         .load_tls = xen_load_tls,
 932
 933         .store_gdt = native_store_gdt,
 934         .store_idt = native_store_idt,
 935         .store_tr = xen_store_tr,
 936
 937         .write_ldt_entry = xen_write_ldt_entry,
 938         .write_gdt_entry = xen_write_gdt_entry,
 939         .write_idt_entry = xen_write_idt_entry,
 940         .load_esp0 = xen_load_esp0,
 941
 942         .set_iopl_mask = xen_set_iopl_mask,
 943         .io_delay = xen_io_delay,
 944
 945         .lazy_mode = {
 946                 .enter = paravirt_enter_lazy_cpu,
 947                 .leave = xen_leave_lazy,
 948         },
 949 };
 950
 951 static const struct pv_irq_ops xen_irq_ops __initdata = {
 952         .init_IRQ = xen_init_IRQ,
 953         .save_fl = xen_save_fl,
 954         .restore_fl = xen_restore_fl,
 955         .irq_disable = xen_irq_disable,
 956         .irq_enable = xen_irq_enable,
 957         .safe_halt = xen_safe_halt,
 958         .halt = xen_halt,
 959 };
 960
 961 static const struct pv_apic_ops xen_apic_ops __initdata = {
 962 #ifdef CONFIG_X86_LOCAL_APIC
 963         .apic_write = xen_apic_write,
 964         .apic_write_atomic = xen_apic_write,
 965         .apic_read = xen_apic_read,
 966         .setup_boot_clock = paravirt_nop,
 967         .setup_secondary_clock = paravirt_nop,
 968         .startup_ipi_hook = paravirt_nop,
 969 #endif
 970 };
 971
 972 static const struct pv_mmu_ops xen_mmu_ops __initdata = {
 973         .pagetable_setup_start = xen_pagetable_setup_start,
 974         .pagetable_setup_done = xen_pagetable_setup_done,
 975
 976         .read_cr2 = xen_read_cr2,
 977         .write_cr2 = xen_write_cr2,
 978
 979         .read_cr3 = xen_read_cr3,
 980         .write_cr3 = xen_write_cr3,
 981
 982         .flush_tlb_user = xen_flush_tlb,
 983         .flush_tlb_kernel = xen_flush_tlb,
 984         .flush_tlb_single = xen_flush_tlb_single,
 985         .flush_tlb_others = xen_flush_tlb_others,
 986
 987         .pte_update = paravirt_nop,
 988         .pte_update_defer = paravirt_nop,
 989
 990         .alloc_pt = xen_alloc_pt_init,
 991         .release_pt = xen_release_pt,
 992         .alloc_pd = paravirt_nop,
 993         .alloc_pd_clone = paravirt_nop,
 994         .release_pd = paravirt_nop,
 995
 996 #ifdef CONFIG_HIGHPTE
 997         .kmap_atomic_pte = xen_kmap_atomic_pte,
 998 #endif
 999
1000         .set_pte = NULL,        /* see xen_pagetable_setup_* */
1001         .set_pte_at = xen_set_pte_at,
1002         .set_pmd = xen_set_pmd,
1003
1004         .pte_val = xen_pte_val,
1005         .pgd_val = xen_pgd_val,
1006
1007         .make_pte = xen_make_pte,
1008         .make_pgd = xen_make_pgd,
1009
1010 #ifdef CONFIG_X86_PAE
1011         .set_pte_atomic = xen_set_pte_atomic,
1012         .set_pte_present = xen_set_pte_at,
1013         .set_pud = xen_set_pud,
1014         .pte_clear = xen_pte_clear,
1015         .pmd_clear = xen_pmd_clear,
1016
1017         .make_pmd = xen_make_pmd,
1018         .pmd_val = xen_pmd_val,
1019 #endif  /* PAE */
1020
1021         .activate_mm = xen_activate_mm,
1022         .dup_mmap = xen_dup_mmap,
1023         .exit_mmap = xen_exit_mmap,
1024
1025         .lazy_mode = {
1026                 .enter = paravirt_enter_lazy_mmu,
1027                 .leave = xen_leave_lazy,
1028         },
1029 };
1030
1031 #ifdef CONFIG_SMP
1032 static const struct smp_ops xen_smp_ops __initdata = {
1033         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
1034         .smp_prepare_cpus = xen_smp_prepare_cpus,
1035         .cpu_up = xen_cpu_up,
1036         .smp_cpus_done = xen_smp_cpus_done,
1037
1038         .smp_send_stop = xen_smp_send_stop,
1039         .smp_send_reschedule = xen_smp_send_reschedule,
1040         .smp_call_function_mask = xen_smp_call_function_mask,
1041 };
1042 #endif  /* CONFIG_SMP */
1043
1044 static void xen_reboot(int reason)
1045 {
1046 #ifdef CONFIG_SMP
1047         smp_send_stop();
1048 #endif
1049
1050         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, reason))
1051                 BUG();
1052 }
1053
1054 static void xen_restart(char *msg)
1055 {
1056         xen_reboot(SHUTDOWN_reboot);
1057 }
1058
1059 static void xen_emergency_restart(void)
1060 {
1061         xen_reboot(SHUTDOWN_reboot);
1062 }
1063
1064 static void xen_machine_halt(void)
1065 {
1066         xen_reboot(SHUTDOWN_poweroff);
1067 }
1068
1069 static void xen_crash_shutdown(struct pt_regs *regs)
1070 {
1071         xen_reboot(SHUTDOWN_crash);
1072 }
1073
1074 static const struct machine_ops __initdata xen_machine_ops = {
1075         .restart = xen_restart,
1076         .halt = xen_machine_halt,
1077         .power_off = xen_machine_halt,
1078         .shutdown = xen_machine_halt,
1079         .crash_shutdown = xen_crash_shutdown,
1080         .emergency_restart = xen_emergency_restart,
1081 };
1082
1083
1084 /* First C function to be called on Xen boot */
1085 asmlinkage void __init xen_start_kernel(void)
1086 {
1087         pgd_t *pgd;
1088
1089         if (!xen_start_info)
1090                 return;
1091
1092         BUG_ON(memcmp(xen_start_info->magic, "xen-3.0", 7) != 0);
1093
1094         /* Install Xen paravirt ops */
1095         pv_info = xen_info;
1096         pv_init_ops = xen_init_ops;
1097         pv_time_ops = xen_time_ops;
1098         pv_cpu_ops = xen_cpu_ops;
1099         pv_irq_ops = xen_irq_ops;
1100         pv_apic_ops = xen_apic_ops;
1101         pv_mmu_ops = xen_mmu_ops;
1102
1103         machine_ops = xen_machine_ops;
1104
1105 #ifdef CONFIG_SMP
1106         smp_ops = xen_smp_ops;
1107 #endif
1108
1109         xen_setup_features();
1110
1111         /* Get mfn list */
1112         if (!xen_feature(XENFEAT_auto_translated_physmap))
1113                 phys_to_machine_mapping = (unsigned long *)xen_start_info->mfn_list;
1114
1115         pgd = (pgd_t *)xen_start_info->pt_base;
1116
1117         init_pg_tables_end = __pa(pgd) + xen_start_info->nr_pt_frames*PAGE_SIZE;
1118
1119         init_mm.pgd = pgd; /* use the Xen pagetables to start */
1120
1121         /* keep using Xen gdt for now; no urgent need to change it */
1122
1123         x86_write_percpu(xen_cr3, __pa(pgd));
1124
1125 #ifdef CONFIG_SMP
1126         /* Don't do the full vcpu_info placement stuff until we have a
1127            possible map. */
1128         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1129 #else
1130         /* May as well do it now, since there's no good time to call
1131            it later on UP. */
1132         xen_setup_vcpu_info_placement();
1133 #endif
1134
1135         pv_info.kernel_rpl = 1;
1136         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1137                 pv_info.kernel_rpl = 0;
1138
1139         /* set the limit of our address space */
1140         reserve_top_address(-HYPERVISOR_VIRT_START + 2 * PAGE_SIZE);
1141
1142         /* set up basic CPUID stuff */
1143         cpu_detect(&new_cpu_data);
1144         new_cpu_data.hard_math = 1;
1145         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1146
1147         /* Poke various useful things into boot_params */
1148         LOADER_TYPE = (9 << 4) | 0;
1149         INITRD_START = xen_start_info->mod_start ? __pa(xen_start_info->mod_start) : 0;
1150         INITRD_SIZE = xen_start_info->mod_len;
1151
1152         /* Start the world */
1153         start_kernel();
1154 }