arch/i386/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(enum paravirt_lazy_mode, xen_lazy_mode);
  56
  57 DEFINE_PER_CPU(struct vcpu_info *, xen_vcpu);
  58 DEFINE_PER_CPU(struct vcpu_info, xen_vcpu_info);
  59 DEFINE_PER_CPU(unsigned long, xen_cr3);
  60
  61 struct start_info *xen_start_info;
  62 EXPORT_SYMBOL_GPL(xen_start_info);
  63
  64 static /* __initdata */ struct shared_info dummy_shared_info;
  65
  66 /*
  67  * Point at some empty memory to start with. We map the real shared_info
  68  * page as soon as fixmap is up and running.
  69  */
  70 struct shared_info *HYPERVISOR_shared_info = (void *)&dummy_shared_info;
  71
  72 /*
  73  * Flag to determine whether vcpu info placement is available on all
  74  * VCPUs.  We assume it is to start with, and then set it to zero on
  75  * the first failure.  This is because it can succeed on some VCPUs
  76  * and not others, since it can involve hypervisor memory allocation,
  77  * or because the guest failed to guarantee all the appropriate
  78  * constraints on all VCPUs (ie buffer can't cross a page boundary).
  79  *
  80  * Note that any particular CPU may be using a placed vcpu structure,
  81  * but we can only optimise if the all are.
  82  *
  83  * 0: not available, 1: available
  84  */
  85 static int have_vcpu_info_placement = 1;
  86
  87 static void __init xen_vcpu_setup(int cpu)
  88 {
  89         struct vcpu_register_vcpu_info info;
  90         int err;
  91         struct vcpu_info *vcpup;
  92
  93         per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
  94
  95         if (!have_vcpu_info_placement)
  96                 return;         /* already tested, not available */
  97
  98         vcpup = &per_cpu(xen_vcpu_info, cpu);
  99
 100         info.mfn = virt_to_mfn(vcpup);
 101         info.offset = offset_in_page(vcpup);
 102
 103         printk(KERN_DEBUG "trying to map vcpu_info %d at %p, mfn %x, offset %d\n",
 104                cpu, vcpup, info.mfn, info.offset);
 105
 106         /* Check to see if the hypervisor will put the vcpu_info
 107            structure where we want it, which allows direct access via
 108            a percpu-variable. */
 109         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
 110
 111         if (err) {
 112                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
 113                 have_vcpu_info_placement = 0;
 114         } else {
 115                 /* This cpu is using the registered vcpu info, even if
 116                    later ones fail to. */
 117                 per_cpu(xen_vcpu, cpu) = vcpup;
 118
 119                 printk(KERN_DEBUG "cpu %d using vcpu_info at %p\n",
 120                        cpu, vcpup);
 121         }
 122 }
 123
 124 static void __init xen_banner(void)
 125 {
 126         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
 127                paravirt_ops.name);
 128         printk(KERN_INFO "Hypervisor signature: %s\n", xen_start_info->magic);
 129 }
 130
 131 static void xen_cpuid(unsigned int *eax, unsigned int *ebx,
 132                       unsigned int *ecx, unsigned int *edx)
 133 {
 134         unsigned maskedx = ~0;
 135
 136         /*
 137          * Mask out inconvenient features, to try and disable as many
 138          * unsupported kernel subsystems as possible.
 139          */
 140         if (*eax == 1)
 141                 maskedx = ~((1 << X86_FEATURE_APIC) |  /* disable APIC */
 142                             (1 << X86_FEATURE_ACPI) |  /* disable ACPI */
 143                             (1 << X86_FEATURE_ACC));   /* thermal monitoring */
 144
 145         asm(XEN_EMULATE_PREFIX "cpuid"
 146                 : "=a" (*eax),
 147                   "=b" (*ebx),
 148                   "=c" (*ecx),
 149                   "=d" (*edx)
 150                 : "0" (*eax), "2" (*ecx));
 151         *edx &= maskedx;
 152 }
 153
 154 static void xen_set_debugreg(int reg, unsigned long val)
 155 {
 156         HYPERVISOR_set_debugreg(reg, val);
 157 }
 158
 159 static unsigned long xen_get_debugreg(int reg)
 160 {
 161         return HYPERVISOR_get_debugreg(reg);
 162 }
 163
 164 static unsigned long xen_save_fl(void)
 165 {
 166         struct vcpu_info *vcpu;
 167         unsigned long flags;
 168
 169         vcpu = x86_read_percpu(xen_vcpu);
 170
 171         /* flag has opposite sense of mask */
 172         flags = !vcpu->evtchn_upcall_mask;
 173
 174         /* convert to IF type flag
 175            -0 -> 0x00000000
 176            -1 -> 0xffffffff
 177         */
 178         return (-flags) & X86_EFLAGS_IF;
 179 }
 180
 181 static void xen_restore_fl(unsigned long flags)
 182 {
 183         struct vcpu_info *vcpu;
 184
 185         /* convert from IF type flag */
 186         flags = !(flags & X86_EFLAGS_IF);
 187
 188         /* There's a one instruction preempt window here.  We need to
 189            make sure we're don't switch CPUs between getting the vcpu
 190            pointer and updating the mask. */
 191         preempt_disable();
 192         vcpu = x86_read_percpu(xen_vcpu);
 193         vcpu->evtchn_upcall_mask = flags;
 194         preempt_enable_no_resched();
 195
 196         /* Doesn't matter if we get preempted here, because any
 197            pending event will get dealt with anyway. */
 198
 199         if (flags == 0) {
 200                 preempt_check_resched();
 201                 barrier(); /* unmask then check (avoid races) */
 202                 if (unlikely(vcpu->evtchn_upcall_pending))
 203                         force_evtchn_callback();
 204         }
 205 }
 206
 207 static void xen_irq_disable(void)
 208 {
 209         /* There's a one instruction preempt window here.  We need to
 210            make sure we're don't switch CPUs between getting the vcpu
 211            pointer and updating the mask. */
 212         preempt_disable();
 213         x86_read_percpu(xen_vcpu)->evtchn_upcall_mask = 1;
 214         preempt_enable_no_resched();
 215 }
 216
 217 static void xen_irq_enable(void)
 218 {
 219         struct vcpu_info *vcpu;
 220
 221         /* There's a one instruction preempt window here.  We need to
 222            make sure we're don't switch CPUs between getting the vcpu
 223            pointer and updating the mask. */
 224         preempt_disable();
 225         vcpu = x86_read_percpu(xen_vcpu);
 226         vcpu->evtchn_upcall_mask = 0;
 227         preempt_enable_no_resched();
 228
 229         /* Doesn't matter if we get preempted here, because any
 230            pending event will get dealt with anyway. */
 231
 232         barrier(); /* unmask then check (avoid races) */
 233         if (unlikely(vcpu->evtchn_upcall_pending))
 234                 force_evtchn_callback();
 235 }
 236
 237 static void xen_safe_halt(void)
 238 {
 239         /* Blocking includes an implicit local_irq_enable(). */
 240         if (HYPERVISOR_sched_op(SCHEDOP_block, 0) != 0)
 241                 BUG();
 242 }
 243
 244 static void xen_halt(void)
 245 {
 246         if (irqs_disabled())
 247                 HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
 248         else
 249                 xen_safe_halt();
 250 }
 251
 252 static void xen_set_lazy_mode(enum paravirt_lazy_mode mode)
 253 {
 254         BUG_ON(preemptible());
 255
 256         switch (mode) {
 257         case PARAVIRT_LAZY_NONE:
 258                 BUG_ON(x86_read_percpu(xen_lazy_mode) == PARAVIRT_LAZY_NONE);
 259                 break;
 260
 261         case PARAVIRT_LAZY_MMU:
 262         case PARAVIRT_LAZY_CPU:
 263                 BUG_ON(x86_read_percpu(xen_lazy_mode) != PARAVIRT_LAZY_NONE);
 264                 break;
 265
 266         case PARAVIRT_LAZY_FLUSH:
 267                 /* flush if necessary, but don't change state */
 268                 if (x86_read_percpu(xen_lazy_mode) != PARAVIRT_LAZY_NONE)
 269                         xen_mc_flush();
 270                 return;
 271         }
 272
 273         xen_mc_flush();
 274         x86_write_percpu(xen_lazy_mode, mode);
 275 }
 276
 277 static unsigned long xen_store_tr(void)
 278 {
 279         return 0;
 280 }
 281
 282 static void xen_set_ldt(const void *addr, unsigned entries)
 283 {
 284         unsigned long linear_addr = (unsigned long)addr;
 285         struct mmuext_op *op;
 286         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 287
 288         op = mcs.args;
 289         op->cmd = MMUEXT_SET_LDT;
 290         if (linear_addr) {
 291                 /* ldt my be vmalloced, use arbitrary_virt_to_machine */
 292                 xmaddr_t maddr;
 293                 maddr = arbitrary_virt_to_machine((unsigned long)addr);
 294                 linear_addr = (unsigned long)maddr.maddr;
 295         }
 296         op->arg1.linear_addr = linear_addr;
 297         op->arg2.nr_ents = entries;
 298
 299         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 300
 301         xen_mc_issue(PARAVIRT_LAZY_CPU);
 302 }
 303
 304 static void xen_load_gdt(const struct Xgt_desc_struct *dtr)
 305 {
 306         unsigned long *frames;
 307         unsigned long va = dtr->address;
 308         unsigned int size = dtr->size + 1;
 309         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
 310         int f;
 311         struct multicall_space mcs;
 312
 313         /* A GDT can be up to 64k in size, which corresponds to 8192
 314            8-byte entries, or 16 4k pages.. */
 315
 316         BUG_ON(size > 65536);
 317         BUG_ON(va & ~PAGE_MASK);
 318
 319         mcs = xen_mc_entry(sizeof(*frames) * pages);
 320         frames = mcs.args;
 321
 322         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
 323                 frames[f] = virt_to_mfn(va);
 324                 make_lowmem_page_readonly((void *)va);
 325         }
 326
 327         MULTI_set_gdt(mcs.mc, frames, size / sizeof(struct desc_struct));
 328
 329         xen_mc_issue(PARAVIRT_LAZY_CPU);
 330 }
 331
 332 static void load_TLS_descriptor(struct thread_struct *t,
 333                                 unsigned int cpu, unsigned int i)
 334 {
 335         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
 336         xmaddr_t maddr = virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
 337         struct multicall_space mc = __xen_mc_entry(0);
 338
 339         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
 340 }
 341
 342 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
 343 {
 344         xen_mc_batch();
 345
 346         load_TLS_descriptor(t, cpu, 0);
 347         load_TLS_descriptor(t, cpu, 1);
 348         load_TLS_descriptor(t, cpu, 2);
 349
 350         xen_mc_issue(PARAVIRT_LAZY_CPU);
 351
 352         /*
 353          * XXX sleazy hack: If we're being called in a lazy-cpu zone,
 354          * it means we're in a context switch, and %gs has just been
 355          * saved.  This means we can zero it out to prevent faults on
 356          * exit from the hypervisor if the next process has no %gs.
 357          * Either way, it has been saved, and the new value will get
 358          * loaded properly.  This will go away as soon as Xen has been
 359          * modified to not save/restore %gs for normal hypercalls.
 360          */
 361         if (xen_get_lazy_mode() == PARAVIRT_LAZY_CPU)
 362                 loadsegment(gs, 0);
 363 }
 364
 365 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
 366                                 u32 low, u32 high)
 367 {
 368         unsigned long lp = (unsigned long)&dt[entrynum];
 369         xmaddr_t mach_lp = virt_to_machine(lp);
 370         u64 entry = (u64)high << 32 | low;
 371
 372         preempt_disable();
 373
 374         xen_mc_flush();
 375         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
 376                 BUG();
 377
 378         preempt_enable();
 379 }
 380
 381 static int cvt_gate_to_trap(int vector, u32 low, u32 high,
 382                             struct trap_info *info)
 383 {
 384         u8 type, dpl;
 385
 386         type = (high >> 8) & 0x1f;
 387         dpl = (high >> 13) & 3;
 388
 389         if (type != 0xf && type != 0xe)
 390                 return 0;
 391
 392         info->vector = vector;
 393         info->address = (high & 0xffff0000) | (low & 0x0000ffff);
 394         info->cs = low >> 16;
 395         info->flags = dpl;
 396         /* interrupt gates clear IF */
 397         if (type == 0xe)
 398                 info->flags |= 4;
 399
 400         return 1;
 401 }
 402
 403 /* Locations of each CPU's IDT */
 404 static DEFINE_PER_CPU(struct Xgt_desc_struct, idt_desc);
 405
 406 /* Set an IDT entry.  If the entry is part of the current IDT, then
 407    also update Xen. */
 408 static void xen_write_idt_entry(struct desc_struct *dt, int entrynum,
 409                                 u32 low, u32 high)
 410 {
 411         unsigned long p = (unsigned long)&dt[entrynum];
 412         unsigned long start, end;
 413
 414         preempt_disable();
 415
 416         start = __get_cpu_var(idt_desc).address;
 417         end = start + __get_cpu_var(idt_desc).size + 1;
 418
 419         xen_mc_flush();
 420
 421         write_dt_entry(dt, entrynum, low, high);
 422
 423         if (p >= start && (p + 8) <= end) {
 424                 struct trap_info info[2];
 425
 426                 info[1].address = 0;
 427
 428                 if (cvt_gate_to_trap(entrynum, low, high, &info[0]))
 429                         if (HYPERVISOR_set_trap_table(info))
 430                                 BUG();
 431         }
 432
 433         preempt_enable();
 434 }
 435
 436 static void xen_convert_trap_info(const struct Xgt_desc_struct *desc,
 437                                   struct trap_info *traps)
 438 {
 439         unsigned in, out, count;
 440
 441         count = (desc->size+1) / 8;
 442         BUG_ON(count > 256);
 443
 444         for (in = out = 0; in < count; in++) {
 445                 const u32 *entry = (u32 *)(desc->address + in * 8);
 446
 447                 if (cvt_gate_to_trap(in, entry[0], entry[1], &traps[out]))
 448                         out++;
 449         }
 450         traps[out].address = 0;
 451 }
 452
 453 void xen_copy_trap_info(struct trap_info *traps)
 454 {
 455         const struct Xgt_desc_struct *desc = &__get_cpu_var(idt_desc);
 456
 457         xen_convert_trap_info(desc, traps);
 458 }
 459
 460 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
 461    hold a spinlock to protect the static traps[] array (static because
 462    it avoids allocation, and saves stack space). */
 463 static void xen_load_idt(const struct Xgt_desc_struct *desc)
 464 {
 465         static DEFINE_SPINLOCK(lock);
 466         static struct trap_info traps[257];
 467
 468         spin_lock(&lock);
 469
 470         __get_cpu_var(idt_desc) = *desc;
 471
 472         xen_convert_trap_info(desc, traps);
 473
 474         xen_mc_flush();
 475         if (HYPERVISOR_set_trap_table(traps))
 476                 BUG();
 477
 478         spin_unlock(&lock);
 479 }
 480
 481 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
 482    they're handled differently. */
 483 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
 484                                 u32 low, u32 high)
 485 {
 486         preempt_disable();
 487
 488         switch ((high >> 8) & 0xff) {
 489         case DESCTYPE_LDT:
 490         case DESCTYPE_TSS:
 491                 /* ignore */
 492                 break;
 493
 494         default: {
 495                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
 496                 u64 desc = (u64)high << 32 | low;
 497
 498                 xen_mc_flush();
 499                 if (HYPERVISOR_update_descriptor(maddr.maddr, desc))
 500                         BUG();
 501         }
 502
 503         }
 504
 505         preempt_enable();
 506 }
 507
 508 static void xen_load_esp0(struct tss_struct *tss,
 509                           struct thread_struct *thread)
 510 {
 511         struct multicall_space mcs = xen_mc_entry(0);
 512         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->esp0);
 513         xen_mc_issue(PARAVIRT_LAZY_CPU);
 514 }
 515
 516 static void xen_set_iopl_mask(unsigned mask)
 517 {
 518         struct physdev_set_iopl set_iopl;
 519
 520         /* Force the change at ring 0. */
 521         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
 522         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
 523 }
 524
 525 static void xen_io_delay(void)
 526 {
 527 }
 528
 529 #ifdef CONFIG_X86_LOCAL_APIC
 530 static unsigned long xen_apic_read(unsigned long reg)
 531 {
 532         return 0;
 533 }
 534
 535 static void xen_apic_write(unsigned long reg, unsigned long val)
 536 {
 537         /* Warn to see if there's any stray references */
 538         WARN_ON(1);
 539 }
 540 #endif
 541
 542 static void xen_flush_tlb(void)
 543 {
 544         struct mmuext_op *op;
 545         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 546
 547         op = mcs.args;
 548         op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
 549         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 550
 551         xen_mc_issue(PARAVIRT_LAZY_MMU);
 552 }
 553
 554 static void xen_flush_tlb_single(unsigned long addr)
 555 {
 556         struct mmuext_op *op;
 557         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 558
 559         op = mcs.args;
 560         op->cmd = MMUEXT_INVLPG_LOCAL;
 561         op->arg1.linear_addr = addr & PAGE_MASK;
 562         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 563
 564         xen_mc_issue(PARAVIRT_LAZY_MMU);
 565 }
 566
 567 static void xen_flush_tlb_others(const cpumask_t *cpus, struct mm_struct *mm,
 568                                  unsigned long va)
 569 {
 570         struct {
 571                 struct mmuext_op op;
 572                 cpumask_t mask;
 573         } *args;
 574         cpumask_t cpumask = *cpus;
 575         struct multicall_space mcs;
 576
 577         /*
 578          * A couple of (to be removed) sanity checks:
 579          *
 580          * - current CPU must not be in mask
 581          * - mask must exist :)
 582          */
 583         BUG_ON(cpus_empty(cpumask));
 584         BUG_ON(cpu_isset(smp_processor_id(), cpumask));
 585         BUG_ON(!mm);
 586
 587         /* If a CPU which we ran on has gone down, OK. */
 588         cpus_and(cpumask, cpumask, cpu_online_map);
 589         if (cpus_empty(cpumask))
 590                 return;
 591
 592         mcs = xen_mc_entry(sizeof(*args));
 593         args = mcs.args;
 594         args->mask = cpumask;
 595         args->op.arg2.vcpumask = &args->mask;
 596
 597         if (va == TLB_FLUSH_ALL) {
 598                 args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
 599         } else {
 600                 args->op.cmd = MMUEXT_INVLPG_MULTI;
 601                 args->op.arg1.linear_addr = va;
 602         }
 603
 604         MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
 605
 606         xen_mc_issue(PARAVIRT_LAZY_MMU);
 607 }
 608
 609 static void xen_write_cr2(unsigned long cr2)
 610 {
 611         x86_read_percpu(xen_vcpu)->arch.cr2 = cr2;
 612 }
 613
 614 static unsigned long xen_read_cr2(void)
 615 {
 616         return x86_read_percpu(xen_vcpu)->arch.cr2;
 617 }
 618
 619 static unsigned long xen_read_cr2_direct(void)
 620 {
 621         return x86_read_percpu(xen_vcpu_info.arch.cr2);
 622 }
 623
 624 static void xen_write_cr4(unsigned long cr4)
 625 {
 626         /* Just ignore cr4 changes; Xen doesn't allow us to do
 627            anything anyway. */
 628 }
 629
 630 static unsigned long xen_read_cr3(void)
 631 {
 632         return x86_read_percpu(xen_cr3);
 633 }
 634
 635 static void xen_write_cr3(unsigned long cr3)
 636 {
 637         BUG_ON(preemptible());
 638
 639         if (cr3 == x86_read_percpu(xen_cr3)) {
 640                 /* just a simple tlb flush */
 641                 xen_flush_tlb();
 642                 return;
 643         }
 644
 645         x86_write_percpu(xen_cr3, cr3);
 646
 647
 648         {
 649                 struct mmuext_op *op;
 650                 struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 651                 unsigned long mfn = pfn_to_mfn(PFN_DOWN(cr3));
 652
 653                 op = mcs.args;
 654                 op->cmd = MMUEXT_NEW_BASEPTR;
 655                 op->arg1.mfn = mfn;
 656
 657                 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 658
 659                 xen_mc_issue(PARAVIRT_LAZY_CPU);
 660         }
 661 }
 662
 663 /* Early in boot, while setting up the initial pagetable, assume
 664    everything is pinned. */
 665 static __init void xen_alloc_pt_init(struct mm_struct *mm, u32 pfn)
 666 {
 667         BUG_ON(mem_map);        /* should only be used early */
 668         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 669 }
 670
 671 /* This needs to make sure the new pte page is pinned iff its being
 672    attached to a pinned pagetable. */
 673 static void xen_alloc_pt(struct mm_struct *mm, u32 pfn)
 674 {
 675         struct page *page = pfn_to_page(pfn);
 676
 677         if (PagePinned(virt_to_page(mm->pgd))) {
 678                 SetPagePinned(page);
 679
 680                 if (!PageHighMem(page))
 681                         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 682                 else
 683                         /* make sure there are no stray mappings of
 684                            this page */
 685                         kmap_flush_unused();
 686         }
 687 }
 688
 689 /* This should never happen until we're OK to use struct page */
 690 static void xen_release_pt(u32 pfn)
 691 {
 692         struct page *page = pfn_to_page(pfn);
 693
 694         if (PagePinned(page)) {
 695                 if (!PageHighMem(page))
 696                         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 697         }
 698 }
 699
 700 #ifdef CONFIG_HIGHPTE
 701 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
 702 {
 703         pgprot_t prot = PAGE_KERNEL;
 704
 705         if (PagePinned(page))
 706                 prot = PAGE_KERNEL_RO;
 707
 708         if (0 && PageHighMem(page))
 709                 printk("mapping highpte %lx type %d prot %s\n",
 710                        page_to_pfn(page), type,
 711                        (unsigned long)pgprot_val(prot) & _PAGE_RW ? "WRITE" : "READ");
 712
 713         return kmap_atomic_prot(page, type, prot);
 714 }
 715 #endif
 716
 717 static __init pte_t mask_rw_pte(pte_t *ptep, pte_t pte)
 718 {
 719         /* If there's an existing pte, then don't allow _PAGE_RW to be set */
 720         if (pte_val_ma(*ptep) & _PAGE_PRESENT)
 721                 pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
 722                                pte_val_ma(pte));
 723
 724         return pte;
 725 }
 726
 727 /* Init-time set_pte while constructing initial pagetables, which
 728    doesn't allow RO pagetable pages to be remapped RW */
 729 static __init void xen_set_pte_init(pte_t *ptep, pte_t pte)
 730 {
 731         pte = mask_rw_pte(ptep, pte);
 732
 733         xen_set_pte(ptep, pte);
 734 }
 735
 736 static __init void xen_pagetable_setup_start(pgd_t *base)
 737 {
 738         pgd_t *xen_pgd = (pgd_t *)xen_start_info->pt_base;
 739
 740         /* special set_pte for pagetable initialization */
 741         paravirt_ops.set_pte = xen_set_pte_init;
 742
 743         init_mm.pgd = base;
 744         /*
 745          * copy top-level of Xen-supplied pagetable into place.  For
 746          * !PAE we can use this as-is, but for PAE it is a stand-in
 747          * while we copy the pmd pages.
 748          */
 749         memcpy(base, xen_pgd, PTRS_PER_PGD * sizeof(pgd_t));
 750
 751         if (PTRS_PER_PMD > 1) {
 752                 int i;
 753                 /*
 754                  * For PAE, need to allocate new pmds, rather than
 755                  * share Xen's, since Xen doesn't like pmd's being
 756                  * shared between address spaces.
 757                  */
 758                 for (i = 0; i < PTRS_PER_PGD; i++) {
 759                         if (pgd_val_ma(xen_pgd[i]) & _PAGE_PRESENT) {
 760                                 pmd_t *pmd = (pmd_t *)alloc_bootmem_low_pages(PAGE_SIZE);
 761
 762                                 memcpy(pmd, (void *)pgd_page_vaddr(xen_pgd[i]),
 763                                        PAGE_SIZE);
 764
 765                                 make_lowmem_page_readonly(pmd);
 766
 767                                 set_pgd(&base[i], __pgd(1 + __pa(pmd)));
 768                         } else
 769                                 pgd_clear(&base[i]);
 770                 }
 771         }
 772
 773         /* make sure zero_page is mapped RO so we can use it in pagetables */
 774         make_lowmem_page_readonly(empty_zero_page);
 775         make_lowmem_page_readonly(base);
 776         /*
 777          * Switch to new pagetable.  This is done before
 778          * pagetable_init has done anything so that the new pages
 779          * added to the table can be prepared properly for Xen.
 780          */
 781         xen_write_cr3(__pa(base));
 782 }
 783
 784 static __init void xen_pagetable_setup_done(pgd_t *base)
 785 {
 786         /* This will work as long as patching hasn't happened yet
 787            (which it hasn't) */
 788         paravirt_ops.alloc_pt = xen_alloc_pt;
 789         paravirt_ops.set_pte = xen_set_pte;
 790
 791         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 792                 /*
 793                  * Create a mapping for the shared info page.
 794                  * Should be set_fixmap(), but shared_info is a machine
 795                  * address with no corresponding pseudo-phys address.
 796                  */
 797                 set_pte_mfn(fix_to_virt(FIX_PARAVIRT_BOOTMAP),
 798                             PFN_DOWN(xen_start_info->shared_info),
 799                             PAGE_KERNEL);
 800
 801                 HYPERVISOR_shared_info =
 802                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
 803
 804         } else
 805                 HYPERVISOR_shared_info =
 806                         (struct shared_info *)__va(xen_start_info->shared_info);
 807
 808         /* Actually pin the pagetable down, but we can't set PG_pinned
 809            yet because the page structures don't exist yet. */
 810         {
 811                 struct mmuext_op op;
 812 #ifdef CONFIG_X86_PAE
 813                 op.cmd = MMUEXT_PIN_L3_TABLE;
 814 #else
 815                 op.cmd = MMUEXT_PIN_L3_TABLE;
 816 #endif
 817                 op.arg1.mfn = pfn_to_mfn(PFN_DOWN(__pa(base)));
 818                 if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
 819                         BUG();
 820         }
 821 }
 822
 823 /* This is called once we have the cpu_possible_map */
 824 void __init xen_setup_vcpu_info_placement(void)
 825 {
 826         int cpu;
 827
 828         for_each_possible_cpu(cpu)
 829                 xen_vcpu_setup(cpu);
 830
 831         /* xen_vcpu_setup managed to place the vcpu_info within the
 832            percpu area for all cpus, so make use of it */
 833         if (have_vcpu_info_placement) {
 834                 printk(KERN_INFO "Xen: using vcpu_info placement\n");
 835
 836                 paravirt_ops.save_fl = xen_save_fl_direct;
 837                 paravirt_ops.restore_fl = xen_restore_fl_direct;
 838                 paravirt_ops.irq_disable = xen_irq_disable_direct;
 839                 paravirt_ops.irq_enable = xen_irq_enable_direct;
 840                 paravirt_ops.read_cr2 = xen_read_cr2_direct;
 841                 paravirt_ops.iret = xen_iret_direct;
 842         }
 843 }
 844
 845 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
 846                           unsigned long addr, unsigned len)
 847 {
 848         char *start, *end, *reloc;
 849         unsigned ret;
 850
 851         start = end = reloc = NULL;
 852
 853 #define SITE(x)                                                         \
 854         case PARAVIRT_PATCH(x):                                         \
 855         if (have_vcpu_info_placement) {                                 \
 856                 start = (char *)xen_##x##_direct;                       \
 857                 end = xen_##x##_direct_end;                             \
 858                 reloc = xen_##x##_direct_reloc;                         \
 859         }                                                               \
 860         goto patch_site
 861
 862         switch (type) {
 863                 SITE(irq_enable);
 864                 SITE(irq_disable);
 865                 SITE(save_fl);
 866                 SITE(restore_fl);
 867 #undef SITE
 868
 869         patch_site:
 870                 if (start == NULL || (end-start) > len)
 871                         goto default_patch;
 872
 873                 ret = paravirt_patch_insns(insnbuf, len, start, end);
 874
 875                 /* Note: because reloc is assigned from something that
 876                    appears to be an array, gcc assumes it's non-null,
 877                    but doesn't know its relationship with start and
 878                    end. */
 879                 if (reloc > start && reloc < end) {
 880                         int reloc_off = reloc - start;
 881                         long *relocp = (long *)(insnbuf + reloc_off);
 882                         long delta = start - (char *)addr;
 883
 884                         *relocp += delta;
 885                 }
 886                 break;
 887
 888         default_patch:
 889         default:
 890                 ret = paravirt_patch_default(type, clobbers, insnbuf,
 891                                              addr, len);
 892                 break;
 893         }
 894
 895         return ret;
 896 }
 897
 898 static const struct paravirt_ops xen_paravirt_ops __initdata = {
 899         .paravirt_enabled = 1,
 900         .shared_kernel_pmd = 0,
 901
 902         .name = "Xen",
 903         .banner = xen_banner,
 904
 905         .patch = xen_patch,
 906
 907         .memory_setup = xen_memory_setup,
 908         .arch_setup = xen_arch_setup,
 909         .init_IRQ = xen_init_IRQ,
 910         .post_allocator_init = xen_mark_init_mm_pinned,
 911
 912         .time_init = xen_time_init,
 913         .set_wallclock = xen_set_wallclock,
 914         .get_wallclock = xen_get_wallclock,
 915         .get_cpu_khz = xen_cpu_khz,
 916         .sched_clock = xen_sched_clock,
 917
 918         .cpuid = xen_cpuid,
 919
 920         .set_debugreg = xen_set_debugreg,
 921         .get_debugreg = xen_get_debugreg,
 922
 923         .clts = native_clts,
 924
 925         .read_cr0 = native_read_cr0,
 926         .write_cr0 = native_write_cr0,
 927
 928         .read_cr2 = xen_read_cr2,
 929         .write_cr2 = xen_write_cr2,
 930
 931         .read_cr3 = xen_read_cr3,
 932         .write_cr3 = xen_write_cr3,
 933
 934         .read_cr4 = native_read_cr4,
 935         .read_cr4_safe = native_read_cr4_safe,
 936         .write_cr4 = xen_write_cr4,
 937
 938         .save_fl = xen_save_fl,
 939         .restore_fl = xen_restore_fl,
 940         .irq_disable = xen_irq_disable,
 941         .irq_enable = xen_irq_enable,
 942         .safe_halt = xen_safe_halt,
 943         .halt = xen_halt,
 944         .wbinvd = native_wbinvd,
 945
 946         .read_msr = native_read_msr_safe,
 947         .write_msr = native_write_msr_safe,
 948         .read_tsc = native_read_tsc,
 949         .read_pmc = native_read_pmc,
 950
 951         .iret = (void *)&hypercall_page[__HYPERVISOR_iret],
 952         .irq_enable_sysexit = NULL,  /* never called */
 953
 954         .load_tr_desc = paravirt_nop,
 955         .set_ldt = xen_set_ldt,
 956         .load_gdt = xen_load_gdt,
 957         .load_idt = xen_load_idt,
 958         .load_tls = xen_load_tls,
 959
 960         .store_gdt = native_store_gdt,
 961         .store_idt = native_store_idt,
 962         .store_tr = xen_store_tr,
 963
 964         .write_ldt_entry = xen_write_ldt_entry,
 965         .write_gdt_entry = xen_write_gdt_entry,
 966         .write_idt_entry = xen_write_idt_entry,
 967         .load_esp0 = xen_load_esp0,
 968
 969         .set_iopl_mask = xen_set_iopl_mask,
 970         .io_delay = xen_io_delay,
 971
 972 #ifdef CONFIG_X86_LOCAL_APIC
 973         .apic_write = xen_apic_write,
 974         .apic_write_atomic = xen_apic_write,
 975         .apic_read = xen_apic_read,
 976         .setup_boot_clock = paravirt_nop,
 977         .setup_secondary_clock = paravirt_nop,
 978         .startup_ipi_hook = paravirt_nop,
 979 #endif
 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         .pagetable_setup_start = xen_pagetable_setup_start,
 990         .pagetable_setup_done = xen_pagetable_setup_done,
 991
 992         .alloc_pt = xen_alloc_pt_init,
 993         .release_pt = xen_release_pt,
 994         .alloc_pd = paravirt_nop,
 995         .alloc_pd_clone = paravirt_nop,
 996         .release_pd = paravirt_nop,
 997
 998 #ifdef CONFIG_HIGHPTE
 999         .kmap_atomic_pte = xen_kmap_atomic_pte,
1000 #endif
1001
1002         .set_pte = NULL,        /* see xen_pagetable_setup_* */
1003         .set_pte_at = xen_set_pte_at,
1004         .set_pmd = xen_set_pmd,
1005
1006         .pte_val = xen_pte_val,
1007         .pgd_val = xen_pgd_val,
1008
1009         .make_pte = xen_make_pte,
1010         .make_pgd = xen_make_pgd,
1011
1012 #ifdef CONFIG_X86_PAE
1013         .set_pte_atomic = xen_set_pte_atomic,
1014         .set_pte_present = xen_set_pte_at,
1015         .set_pud = xen_set_pud,
1016         .pte_clear = xen_pte_clear,
1017         .pmd_clear = xen_pmd_clear,
1018
1019         .make_pmd = xen_make_pmd,
1020         .pmd_val = xen_pmd_val,
1021 #endif  /* PAE */
1022
1023         .activate_mm = xen_activate_mm,
1024         .dup_mmap = xen_dup_mmap,
1025         .exit_mmap = xen_exit_mmap,
1026
1027         .set_lazy_mode = xen_set_lazy_mode,
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         paravirt_ops = xen_paravirt_ops;
1095         machine_ops = xen_machine_ops;
1096
1097 #ifdef CONFIG_SMP
1098         smp_ops = xen_smp_ops;
1099 #endif
1100
1101         xen_setup_features();
1102
1103         /* Get mfn list */
1104         if (!xen_feature(XENFEAT_auto_translated_physmap))
1105                 phys_to_machine_mapping = (unsigned long *)xen_start_info->mfn_list;
1106
1107         pgd = (pgd_t *)xen_start_info->pt_base;
1108
1109         init_pg_tables_end = __pa(pgd) + xen_start_info->nr_pt_frames*PAGE_SIZE;
1110
1111         init_mm.pgd = pgd; /* use the Xen pagetables to start */
1112
1113         /* keep using Xen gdt for now; no urgent need to change it */
1114
1115         x86_write_percpu(xen_cr3, __pa(pgd));
1116
1117 #ifdef CONFIG_SMP
1118         /* Don't do the full vcpu_info placement stuff until we have a
1119            possible map. */
1120         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1121 #else
1122         /* May as well do it now, since there's no good time to call
1123            it later on UP. */
1124         xen_setup_vcpu_info_placement();
1125 #endif
1126
1127         paravirt_ops.kernel_rpl = 1;
1128         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1129                 paravirt_ops.kernel_rpl = 0;
1130
1131         /* set the limit of our address space */
1132         reserve_top_address(-HYPERVISOR_VIRT_START + 2 * PAGE_SIZE);
1133
1134         /* set up basic CPUID stuff */
1135         cpu_detect(&new_cpu_data);
1136         new_cpu_data.hard_math = 1;
1137         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1138
1139         /* Poke various useful things into boot_params */
1140         LOADER_TYPE = (9 << 4) | 0;
1141         INITRD_START = xen_start_info->mod_start ? __pa(xen_start_info->mod_start) : 0;
1142         INITRD_SIZE = xen_start_info->mod_len;
1143
1144         /* Start the world */
1145         start_kernel();
1146 }