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
  57 /*
  58  * Note about cr3 (pagetable base) values:
  59  *
  60  * xen_cr3 contains the current logical cr3 value; it contains the
  61  * last set cr3.  This may not be the current effective cr3, because
  62  * its update may be being lazily deferred.  However, a vcpu looking
  63  * at its own cr3 can use this value knowing that it everything will
  64  * be self-consistent.
  65  *
  66  * xen_current_cr3 contains the actual vcpu cr3; it is set once the
  67  * hypercall to set the vcpu cr3 is complete (so it may be a little
  68  * out of date, but it will never be set early).  If one vcpu is
  69  * looking at another vcpu's cr3 value, it should use this variable.
  70  */
  71 DEFINE_PER_CPU(unsigned long, xen_cr3);  /* cr3 stored as physaddr */
  72 DEFINE_PER_CPU(unsigned long, xen_current_cr3);  /* actual vcpu cr3 */
  73
  74 struct start_info *xen_start_info;
  75 EXPORT_SYMBOL_GPL(xen_start_info);
  76
  77 static /* __initdata */ struct shared_info dummy_shared_info;
  78
  79 /*
  80  * Point at some empty memory to start with. We map the real shared_info
  81  * page as soon as fixmap is up and running.
  82  */
  83 struct shared_info *HYPERVISOR_shared_info = (void *)&dummy_shared_info;
  84
  85 /*
  86  * Flag to determine whether vcpu info placement is available on all
  87  * VCPUs.  We assume it is to start with, and then set it to zero on
  88  * the first failure.  This is because it can succeed on some VCPUs
  89  * and not others, since it can involve hypervisor memory allocation,
  90  * or because the guest failed to guarantee all the appropriate
  91  * constraints on all VCPUs (ie buffer can't cross a page boundary).
  92  *
  93  * Note that any particular CPU may be using a placed vcpu structure,
  94  * but we can only optimise if the all are.
  95  *
  96  * 0: not available, 1: available
  97  */
  98 static int have_vcpu_info_placement = 0;
  99
 100 static void __init xen_vcpu_setup(int cpu)
 101 {
 102         struct vcpu_register_vcpu_info info;
 103         int err;
 104         struct vcpu_info *vcpup;
 105
 106         per_cpu(xen_vcpu, cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
 107
 108         if (!have_vcpu_info_placement)
 109                 return;         /* already tested, not available */
 110
 111         vcpup = &per_cpu(xen_vcpu_info, cpu);
 112
 113         info.mfn = virt_to_mfn(vcpup);
 114         info.offset = offset_in_page(vcpup);
 115
 116         printk(KERN_DEBUG "trying to map vcpu_info %d at %p, mfn %llx, offset %d\n",
 117                cpu, vcpup, info.mfn, info.offset);
 118
 119         /* Check to see if the hypervisor will put the vcpu_info
 120            structure where we want it, which allows direct access via
 121            a percpu-variable. */
 122         err = HYPERVISOR_vcpu_op(VCPUOP_register_vcpu_info, cpu, &info);
 123
 124         if (err) {
 125                 printk(KERN_DEBUG "register_vcpu_info failed: err=%d\n", err);
 126                 have_vcpu_info_placement = 0;
 127         } else {
 128                 /* This cpu is using the registered vcpu info, even if
 129                    later ones fail to. */
 130                 per_cpu(xen_vcpu, cpu) = vcpup;
 131
 132                 printk(KERN_DEBUG "cpu %d using vcpu_info at %p\n",
 133                        cpu, vcpup);
 134         }
 135 }
 136
 137 static void __init xen_banner(void)
 138 {
 139         printk(KERN_INFO "Booting paravirtualized kernel on %s\n",
 140                pv_info.name);
 141         printk(KERN_INFO "Hypervisor signature: %s\n", xen_start_info->magic);
 142 }
 143
 144 static void xen_cpuid(unsigned int *eax, unsigned int *ebx,
 145                       unsigned int *ecx, unsigned int *edx)
 146 {
 147         unsigned maskedx = ~0;
 148
 149         /*
 150          * Mask out inconvenient features, to try and disable as many
 151          * unsupported kernel subsystems as possible.
 152          */
 153         if (*eax == 1)
 154                 maskedx = ~((1 << X86_FEATURE_APIC) |  /* disable APIC */
 155                             (1 << X86_FEATURE_ACPI) |  /* disable ACPI */
 156                             (1 << X86_FEATURE_ACC));   /* thermal monitoring */
 157
 158         asm(XEN_EMULATE_PREFIX "cpuid"
 159                 : "=a" (*eax),
 160                   "=b" (*ebx),
 161                   "=c" (*ecx),
 162                   "=d" (*edx)
 163                 : "0" (*eax), "2" (*ecx));
 164         *edx &= maskedx;
 165 }
 166
 167 static void xen_set_debugreg(int reg, unsigned long val)
 168 {
 169         HYPERVISOR_set_debugreg(reg, val);
 170 }
 171
 172 static unsigned long xen_get_debugreg(int reg)
 173 {
 174         return HYPERVISOR_get_debugreg(reg);
 175 }
 176
 177 static unsigned long xen_save_fl(void)
 178 {
 179         struct vcpu_info *vcpu;
 180         unsigned long flags;
 181
 182         vcpu = x86_read_percpu(xen_vcpu);
 183
 184         /* flag has opposite sense of mask */
 185         flags = !vcpu->evtchn_upcall_mask;
 186
 187         /* convert to IF type flag
 188            -0 -> 0x00000000
 189            -1 -> 0xffffffff
 190         */
 191         return (-flags) & X86_EFLAGS_IF;
 192 }
 193
 194 static void xen_restore_fl(unsigned long flags)
 195 {
 196         struct vcpu_info *vcpu;
 197
 198         /* convert from IF type flag */
 199         flags = !(flags & X86_EFLAGS_IF);
 200
 201         /* There's a one instruction preempt window here.  We need to
 202            make sure we're don't switch CPUs between getting the vcpu
 203            pointer and updating the mask. */
 204         preempt_disable();
 205         vcpu = x86_read_percpu(xen_vcpu);
 206         vcpu->evtchn_upcall_mask = flags;
 207         preempt_enable_no_resched();
 208
 209         /* Doesn't matter if we get preempted here, because any
 210            pending event will get dealt with anyway. */
 211
 212         if (flags == 0) {
 213                 preempt_check_resched();
 214                 barrier(); /* unmask then check (avoid races) */
 215                 if (unlikely(vcpu->evtchn_upcall_pending))
 216                         force_evtchn_callback();
 217         }
 218 }
 219
 220 static void xen_irq_disable(void)
 221 {
 222         /* There's a one instruction preempt window here.  We need to
 223            make sure we're don't switch CPUs between getting the vcpu
 224            pointer and updating the mask. */
 225         preempt_disable();
 226         x86_read_percpu(xen_vcpu)->evtchn_upcall_mask = 1;
 227         preempt_enable_no_resched();
 228 }
 229
 230 static void xen_irq_enable(void)
 231 {
 232         struct vcpu_info *vcpu;
 233
 234         /* There's a one instruction preempt window here.  We need to
 235            make sure we're don't switch CPUs between getting the vcpu
 236            pointer and updating the mask. */
 237         preempt_disable();
 238         vcpu = x86_read_percpu(xen_vcpu);
 239         vcpu->evtchn_upcall_mask = 0;
 240         preempt_enable_no_resched();
 241
 242         /* Doesn't matter if we get preempted here, because any
 243            pending event will get dealt with anyway. */
 244
 245         barrier(); /* unmask then check (avoid races) */
 246         if (unlikely(vcpu->evtchn_upcall_pending))
 247                 force_evtchn_callback();
 248 }
 249
 250 static void xen_safe_halt(void)
 251 {
 252         /* Blocking includes an implicit local_irq_enable(). */
 253         if (HYPERVISOR_sched_op(SCHEDOP_block, 0) != 0)
 254                 BUG();
 255 }
 256
 257 static void xen_halt(void)
 258 {
 259         if (irqs_disabled())
 260                 HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
 261         else
 262                 xen_safe_halt();
 263 }
 264
 265 static void xen_leave_lazy(void)
 266 {
 267         paravirt_leave_lazy(paravirt_get_lazy_mode());
 268         xen_mc_flush();
 269 }
 270
 271 static unsigned long xen_store_tr(void)
 272 {
 273         return 0;
 274 }
 275
 276 static void xen_set_ldt(const void *addr, unsigned entries)
 277 {
 278         unsigned long linear_addr = (unsigned long)addr;
 279         struct mmuext_op *op;
 280         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 281
 282         op = mcs.args;
 283         op->cmd = MMUEXT_SET_LDT;
 284         if (linear_addr) {
 285                 /* ldt my be vmalloced, use arbitrary_virt_to_machine */
 286                 xmaddr_t maddr;
 287                 maddr = arbitrary_virt_to_machine((unsigned long)addr);
 288                 linear_addr = (unsigned long)maddr.maddr;
 289         }
 290         op->arg1.linear_addr = linear_addr;
 291         op->arg2.nr_ents = entries;
 292
 293         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 294
 295         xen_mc_issue(PARAVIRT_LAZY_CPU);
 296 }
 297
 298 static void xen_load_gdt(const struct Xgt_desc_struct *dtr)
 299 {
 300         unsigned long *frames;
 301         unsigned long va = dtr->address;
 302         unsigned int size = dtr->size + 1;
 303         unsigned pages = (size + PAGE_SIZE - 1) / PAGE_SIZE;
 304         int f;
 305         struct multicall_space mcs;
 306
 307         /* A GDT can be up to 64k in size, which corresponds to 8192
 308            8-byte entries, or 16 4k pages.. */
 309
 310         BUG_ON(size > 65536);
 311         BUG_ON(va & ~PAGE_MASK);
 312
 313         mcs = xen_mc_entry(sizeof(*frames) * pages);
 314         frames = mcs.args;
 315
 316         for (f = 0; va < dtr->address + size; va += PAGE_SIZE, f++) {
 317                 frames[f] = virt_to_mfn(va);
 318                 make_lowmem_page_readonly((void *)va);
 319         }
 320
 321         MULTI_set_gdt(mcs.mc, frames, size / sizeof(struct desc_struct));
 322
 323         xen_mc_issue(PARAVIRT_LAZY_CPU);
 324 }
 325
 326 static void load_TLS_descriptor(struct thread_struct *t,
 327                                 unsigned int cpu, unsigned int i)
 328 {
 329         struct desc_struct *gdt = get_cpu_gdt_table(cpu);
 330         xmaddr_t maddr = virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
 331         struct multicall_space mc = __xen_mc_entry(0);
 332
 333         MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
 334 }
 335
 336 static void xen_load_tls(struct thread_struct *t, unsigned int cpu)
 337 {
 338         xen_mc_batch();
 339
 340         load_TLS_descriptor(t, cpu, 0);
 341         load_TLS_descriptor(t, cpu, 1);
 342         load_TLS_descriptor(t, cpu, 2);
 343
 344         xen_mc_issue(PARAVIRT_LAZY_CPU);
 345
 346         /*
 347          * XXX sleazy hack: If we're being called in a lazy-cpu zone,
 348          * it means we're in a context switch, and %gs has just been
 349          * saved.  This means we can zero it out to prevent faults on
 350          * exit from the hypervisor if the next process has no %gs.
 351          * Either way, it has been saved, and the new value will get
 352          * loaded properly.  This will go away as soon as Xen has been
 353          * modified to not save/restore %gs for normal hypercalls.
 354          */
 355         if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_CPU)
 356                 loadsegment(gs, 0);
 357 }
 358
 359 static void xen_write_ldt_entry(struct desc_struct *dt, int entrynum,
 360                                 u32 low, u32 high)
 361 {
 362         unsigned long lp = (unsigned long)&dt[entrynum];
 363         xmaddr_t mach_lp = virt_to_machine(lp);
 364         u64 entry = (u64)high << 32 | low;
 365
 366         preempt_disable();
 367
 368         xen_mc_flush();
 369         if (HYPERVISOR_update_descriptor(mach_lp.maddr, entry))
 370                 BUG();
 371
 372         preempt_enable();
 373 }
 374
 375 static int cvt_gate_to_trap(int vector, u32 low, u32 high,
 376                             struct trap_info *info)
 377 {
 378         u8 type, dpl;
 379
 380         type = (high >> 8) & 0x1f;
 381         dpl = (high >> 13) & 3;
 382
 383         if (type != 0xf && type != 0xe)
 384                 return 0;
 385
 386         info->vector = vector;
 387         info->address = (high & 0xffff0000) | (low & 0x0000ffff);
 388         info->cs = low >> 16;
 389         info->flags = dpl;
 390         /* interrupt gates clear IF */
 391         if (type == 0xe)
 392                 info->flags |= 4;
 393
 394         return 1;
 395 }
 396
 397 /* Locations of each CPU's IDT */
 398 static DEFINE_PER_CPU(struct Xgt_desc_struct, idt_desc);
 399
 400 /* Set an IDT entry.  If the entry is part of the current IDT, then
 401    also update Xen. */
 402 static void xen_write_idt_entry(struct desc_struct *dt, int entrynum,
 403                                 u32 low, u32 high)
 404 {
 405         unsigned long p = (unsigned long)&dt[entrynum];
 406         unsigned long start, end;
 407
 408         preempt_disable();
 409
 410         start = __get_cpu_var(idt_desc).address;
 411         end = start + __get_cpu_var(idt_desc).size + 1;
 412
 413         xen_mc_flush();
 414
 415         write_dt_entry(dt, entrynum, low, high);
 416
 417         if (p >= start && (p + 8) <= end) {
 418                 struct trap_info info[2];
 419
 420                 info[1].address = 0;
 421
 422                 if (cvt_gate_to_trap(entrynum, low, high, &info[0]))
 423                         if (HYPERVISOR_set_trap_table(info))
 424                                 BUG();
 425         }
 426
 427         preempt_enable();
 428 }
 429
 430 static void xen_convert_trap_info(const struct Xgt_desc_struct *desc,
 431                                   struct trap_info *traps)
 432 {
 433         unsigned in, out, count;
 434
 435         count = (desc->size+1) / 8;
 436         BUG_ON(count > 256);
 437
 438         for (in = out = 0; in < count; in++) {
 439                 const u32 *entry = (u32 *)(desc->address + in * 8);
 440
 441                 if (cvt_gate_to_trap(in, entry[0], entry[1], &traps[out]))
 442                         out++;
 443         }
 444         traps[out].address = 0;
 445 }
 446
 447 void xen_copy_trap_info(struct trap_info *traps)
 448 {
 449         const struct Xgt_desc_struct *desc = &__get_cpu_var(idt_desc);
 450
 451         xen_convert_trap_info(desc, traps);
 452 }
 453
 454 /* Load a new IDT into Xen.  In principle this can be per-CPU, so we
 455    hold a spinlock to protect the static traps[] array (static because
 456    it avoids allocation, and saves stack space). */
 457 static void xen_load_idt(const struct Xgt_desc_struct *desc)
 458 {
 459         static DEFINE_SPINLOCK(lock);
 460         static struct trap_info traps[257];
 461
 462         spin_lock(&lock);
 463
 464         __get_cpu_var(idt_desc) = *desc;
 465
 466         xen_convert_trap_info(desc, traps);
 467
 468         xen_mc_flush();
 469         if (HYPERVISOR_set_trap_table(traps))
 470                 BUG();
 471
 472         spin_unlock(&lock);
 473 }
 474
 475 /* Write a GDT descriptor entry.  Ignore LDT descriptors, since
 476    they're handled differently. */
 477 static void xen_write_gdt_entry(struct desc_struct *dt, int entry,
 478                                 u32 low, u32 high)
 479 {
 480         preempt_disable();
 481
 482         switch ((high >> 8) & 0xff) {
 483         case DESCTYPE_LDT:
 484         case DESCTYPE_TSS:
 485                 /* ignore */
 486                 break;
 487
 488         default: {
 489                 xmaddr_t maddr = virt_to_machine(&dt[entry]);
 490                 u64 desc = (u64)high << 32 | low;
 491
 492                 xen_mc_flush();
 493                 if (HYPERVISOR_update_descriptor(maddr.maddr, desc))
 494                         BUG();
 495         }
 496
 497         }
 498
 499         preempt_enable();
 500 }
 501
 502 static void xen_load_esp0(struct tss_struct *tss,
 503                           struct thread_struct *thread)
 504 {
 505         struct multicall_space mcs = xen_mc_entry(0);
 506         MULTI_stack_switch(mcs.mc, __KERNEL_DS, thread->esp0);
 507         xen_mc_issue(PARAVIRT_LAZY_CPU);
 508 }
 509
 510 static void xen_set_iopl_mask(unsigned mask)
 511 {
 512         struct physdev_set_iopl set_iopl;
 513
 514         /* Force the change at ring 0. */
 515         set_iopl.iopl = (mask == 0) ? 1 : (mask >> 12) & 3;
 516         HYPERVISOR_physdev_op(PHYSDEVOP_set_iopl, &set_iopl);
 517 }
 518
 519 static void xen_io_delay(void)
 520 {
 521 }
 522
 523 #ifdef CONFIG_X86_LOCAL_APIC
 524 static unsigned long xen_apic_read(unsigned long reg)
 525 {
 526         return 0;
 527 }
 528
 529 static void xen_apic_write(unsigned long reg, unsigned long val)
 530 {
 531         /* Warn to see if there's any stray references */
 532         WARN_ON(1);
 533 }
 534 #endif
 535
 536 static void xen_flush_tlb(void)
 537 {
 538         struct mmuext_op *op;
 539         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 540
 541         op = mcs.args;
 542         op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
 543         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 544
 545         xen_mc_issue(PARAVIRT_LAZY_MMU);
 546 }
 547
 548 static void xen_flush_tlb_single(unsigned long addr)
 549 {
 550         struct mmuext_op *op;
 551         struct multicall_space mcs = xen_mc_entry(sizeof(*op));
 552
 553         op = mcs.args;
 554         op->cmd = MMUEXT_INVLPG_LOCAL;
 555         op->arg1.linear_addr = addr & PAGE_MASK;
 556         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 557
 558         xen_mc_issue(PARAVIRT_LAZY_MMU);
 559 }
 560
 561 static void xen_flush_tlb_others(const cpumask_t *cpus, struct mm_struct *mm,
 562                                  unsigned long va)
 563 {
 564         struct {
 565                 struct mmuext_op op;
 566                 cpumask_t mask;
 567         } *args;
 568         cpumask_t cpumask = *cpus;
 569         struct multicall_space mcs;
 570
 571         /*
 572          * A couple of (to be removed) sanity checks:
 573          *
 574          * - current CPU must not be in mask
 575          * - mask must exist :)
 576          */
 577         BUG_ON(cpus_empty(cpumask));
 578         BUG_ON(cpu_isset(smp_processor_id(), cpumask));
 579         BUG_ON(!mm);
 580
 581         /* If a CPU which we ran on has gone down, OK. */
 582         cpus_and(cpumask, cpumask, cpu_online_map);
 583         if (cpus_empty(cpumask))
 584                 return;
 585
 586         mcs = xen_mc_entry(sizeof(*args));
 587         args = mcs.args;
 588         args->mask = cpumask;
 589         args->op.arg2.vcpumask = &args->mask;
 590
 591         if (va == TLB_FLUSH_ALL) {
 592                 args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
 593         } else {
 594                 args->op.cmd = MMUEXT_INVLPG_MULTI;
 595                 args->op.arg1.linear_addr = va;
 596         }
 597
 598         MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
 599
 600         xen_mc_issue(PARAVIRT_LAZY_MMU);
 601 }
 602
 603 static void xen_write_cr2(unsigned long cr2)
 604 {
 605         x86_read_percpu(xen_vcpu)->arch.cr2 = cr2;
 606 }
 607
 608 static unsigned long xen_read_cr2(void)
 609 {
 610         return x86_read_percpu(xen_vcpu)->arch.cr2;
 611 }
 612
 613 static unsigned long xen_read_cr2_direct(void)
 614 {
 615         return x86_read_percpu(xen_vcpu_info.arch.cr2);
 616 }
 617
 618 static void xen_write_cr4(unsigned long cr4)
 619 {
 620         /* Just ignore cr4 changes; Xen doesn't allow us to do
 621            anything anyway. */
 622 }
 623
 624 static unsigned long xen_read_cr3(void)
 625 {
 626         return x86_read_percpu(xen_cr3);
 627 }
 628
 629 static void set_current_cr3(void *v)
 630 {
 631         x86_write_percpu(xen_current_cr3, (unsigned long)v);
 632 }
 633
 634 static void xen_write_cr3(unsigned long cr3)
 635 {
 636         struct mmuext_op *op;
 637         struct multicall_space mcs;
 638         unsigned long mfn = pfn_to_mfn(PFN_DOWN(cr3));
 639
 640         BUG_ON(preemptible());
 641
 642         mcs = xen_mc_entry(sizeof(*op));  /* disables interrupts */
 643
 644         /* Update while interrupts are disabled, so its atomic with
 645            respect to ipis */
 646         x86_write_percpu(xen_cr3, cr3);
 647
 648         op = mcs.args;
 649         op->cmd = MMUEXT_NEW_BASEPTR;
 650         op->arg1.mfn = mfn;
 651
 652         MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
 653
 654         /* Update xen_update_cr3 once the batch has actually
 655            been submitted. */
 656         xen_mc_callback(set_current_cr3, (void *)cr3);
 657
 658         xen_mc_issue(PARAVIRT_LAZY_CPU);  /* interrupts restored */
 659 }
 660
 661 /* Early in boot, while setting up the initial pagetable, assume
 662    everything is pinned. */
 663 static __init void xen_alloc_pt_init(struct mm_struct *mm, u32 pfn)
 664 {
 665         BUG_ON(mem_map);        /* should only be used early */
 666         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 667 }
 668
 669 static void pin_pagetable_pfn(unsigned level, unsigned long pfn)
 670 {
 671         struct mmuext_op op;
 672         op.cmd = level;
 673         op.arg1.mfn = pfn_to_mfn(pfn);
 674         if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
 675                 BUG();
 676 }
 677
 678 /* This needs to make sure the new pte page is pinned iff its being
 679    attached to a pinned pagetable. */
 680 static void xen_alloc_pt(struct mm_struct *mm, u32 pfn)
 681 {
 682         struct page *page = pfn_to_page(pfn);
 683
 684         if (PagePinned(virt_to_page(mm->pgd))) {
 685                 SetPagePinned(page);
 686
 687                 if (!PageHighMem(page)) {
 688                         make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
 689                         pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
 690                 } else
 691                         /* make sure there are no stray mappings of
 692                            this page */
 693                         kmap_flush_unused();
 694         }
 695 }
 696
 697 /* This should never happen until we're OK to use struct page */
 698 static void xen_release_pt(u32 pfn)
 699 {
 700         struct page *page = pfn_to_page(pfn);
 701
 702         if (PagePinned(page)) {
 703                 if (!PageHighMem(page)) {
 704                         pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
 705                         make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
 706                 }
 707         }
 708 }
 709
 710 #ifdef CONFIG_HIGHPTE
 711 static void *xen_kmap_atomic_pte(struct page *page, enum km_type type)
 712 {
 713         pgprot_t prot = PAGE_KERNEL;
 714
 715         if (PagePinned(page))
 716                 prot = PAGE_KERNEL_RO;
 717
 718         if (0 && PageHighMem(page))
 719                 printk("mapping highpte %lx type %d prot %s\n",
 720                        page_to_pfn(page), type,
 721                        (unsigned long)pgprot_val(prot) & _PAGE_RW ? "WRITE" : "READ");
 722
 723         return kmap_atomic_prot(page, type, prot);
 724 }
 725 #endif
 726
 727 static __init pte_t mask_rw_pte(pte_t *ptep, pte_t pte)
 728 {
 729         /* If there's an existing pte, then don't allow _PAGE_RW to be set */
 730         if (pte_val_ma(*ptep) & _PAGE_PRESENT)
 731                 pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
 732                                pte_val_ma(pte));
 733
 734         return pte;
 735 }
 736
 737 /* Init-time set_pte while constructing initial pagetables, which
 738    doesn't allow RO pagetable pages to be remapped RW */
 739 static __init void xen_set_pte_init(pte_t *ptep, pte_t pte)
 740 {
 741         pte = mask_rw_pte(ptep, pte);
 742
 743         xen_set_pte(ptep, pte);
 744 }
 745
 746 static __init void xen_pagetable_setup_start(pgd_t *base)
 747 {
 748         pgd_t *xen_pgd = (pgd_t *)xen_start_info->pt_base;
 749
 750         /* special set_pte for pagetable initialization */
 751         pv_mmu_ops.set_pte = xen_set_pte_init;
 752
 753         init_mm.pgd = base;
 754         /*
 755          * copy top-level of Xen-supplied pagetable into place.  For
 756          * !PAE we can use this as-is, but for PAE it is a stand-in
 757          * while we copy the pmd pages.
 758          */
 759         memcpy(base, xen_pgd, PTRS_PER_PGD * sizeof(pgd_t));
 760
 761         if (PTRS_PER_PMD > 1) {
 762                 int i;
 763                 /*
 764                  * For PAE, need to allocate new pmds, rather than
 765                  * share Xen's, since Xen doesn't like pmd's being
 766                  * shared between address spaces.
 767                  */
 768                 for (i = 0; i < PTRS_PER_PGD; i++) {
 769                         if (pgd_val_ma(xen_pgd[i]) & _PAGE_PRESENT) {
 770                                 pmd_t *pmd = (pmd_t *)alloc_bootmem_low_pages(PAGE_SIZE);
 771
 772                                 memcpy(pmd, (void *)pgd_page_vaddr(xen_pgd[i]),
 773                                        PAGE_SIZE);
 774
 775                                 make_lowmem_page_readonly(pmd);
 776
 777                                 set_pgd(&base[i], __pgd(1 + __pa(pmd)));
 778                         } else
 779                                 pgd_clear(&base[i]);
 780                 }
 781         }
 782
 783         /* make sure zero_page is mapped RO so we can use it in pagetables */
 784         make_lowmem_page_readonly(empty_zero_page);
 785         make_lowmem_page_readonly(base);
 786         /*
 787          * Switch to new pagetable.  This is done before
 788          * pagetable_init has done anything so that the new pages
 789          * added to the table can be prepared properly for Xen.
 790          */
 791         xen_write_cr3(__pa(base));
 792 }
 793
 794 static __init void xen_pagetable_setup_done(pgd_t *base)
 795 {
 796         /* This will work as long as patching hasn't happened yet
 797            (which it hasn't) */
 798         pv_mmu_ops.alloc_pt = xen_alloc_pt;
 799         pv_mmu_ops.set_pte = xen_set_pte;
 800
 801         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 802                 /*
 803                  * Create a mapping for the shared info page.
 804                  * Should be set_fixmap(), but shared_info is a machine
 805                  * address with no corresponding pseudo-phys address.
 806                  */
 807                 set_pte_mfn(fix_to_virt(FIX_PARAVIRT_BOOTMAP),
 808                             PFN_DOWN(xen_start_info->shared_info),
 809                             PAGE_KERNEL);
 810
 811                 HYPERVISOR_shared_info =
 812                         (struct shared_info *)fix_to_virt(FIX_PARAVIRT_BOOTMAP);
 813
 814         } else
 815                 HYPERVISOR_shared_info =
 816                         (struct shared_info *)__va(xen_start_info->shared_info);
 817
 818         /* Actually pin the pagetable down, but we can't set PG_pinned
 819            yet because the page structures don't exist yet. */
 820         {
 821                 unsigned level;
 822
 823 #ifdef CONFIG_X86_PAE
 824                 level = MMUEXT_PIN_L3_TABLE;
 825 #else
 826                 level = MMUEXT_PIN_L2_TABLE;
 827 #endif
 828
 829                 pin_pagetable_pfn(level, PFN_DOWN(__pa(base)));
 830         }
 831 }
 832
 833 /* This is called once we have the cpu_possible_map */
 834 void __init xen_setup_vcpu_info_placement(void)
 835 {
 836         int cpu;
 837
 838         for_each_possible_cpu(cpu)
 839                 xen_vcpu_setup(cpu);
 840
 841         /* xen_vcpu_setup managed to place the vcpu_info within the
 842            percpu area for all cpus, so make use of it */
 843         if (have_vcpu_info_placement) {
 844                 printk(KERN_INFO "Xen: using vcpu_info placement\n");
 845
 846                 pv_irq_ops.save_fl = xen_save_fl_direct;
 847                 pv_irq_ops.restore_fl = xen_restore_fl_direct;
 848                 pv_irq_ops.irq_disable = xen_irq_disable_direct;
 849                 pv_irq_ops.irq_enable = xen_irq_enable_direct;
 850                 pv_mmu_ops.read_cr2 = xen_read_cr2_direct;
 851                 pv_cpu_ops.iret = xen_iret_direct;
 852         }
 853 }
 854
 855 static unsigned xen_patch(u8 type, u16 clobbers, void *insnbuf,
 856                           unsigned long addr, unsigned len)
 857 {
 858         char *start, *end, *reloc;
 859         unsigned ret;
 860
 861         start = end = reloc = NULL;
 862
 863 #define SITE(op, x)                                                     \
 864         case PARAVIRT_PATCH(op.x):                                      \
 865         if (have_vcpu_info_placement) {                                 \
 866                 start = (char *)xen_##x##_direct;                       \
 867                 end = xen_##x##_direct_end;                             \
 868                 reloc = xen_##x##_direct_reloc;                         \
 869         }                                                               \
 870         goto patch_site
 871
 872         switch (type) {
 873                 SITE(pv_irq_ops, irq_enable);
 874                 SITE(pv_irq_ops, irq_disable);
 875                 SITE(pv_irq_ops, save_fl);
 876                 SITE(pv_irq_ops, restore_fl);
 877 #undef SITE
 878
 879         patch_site:
 880                 if (start == NULL || (end-start) > len)
 881                         goto default_patch;
 882
 883                 ret = paravirt_patch_insns(insnbuf, len, start, end);
 884
 885                 /* Note: because reloc is assigned from something that
 886                    appears to be an array, gcc assumes it's non-null,
 887                    but doesn't know its relationship with start and
 888                    end. */
 889                 if (reloc > start && reloc < end) {
 890                         int reloc_off = reloc - start;
 891                         long *relocp = (long *)(insnbuf + reloc_off);
 892                         long delta = start - (char *)addr;
 893
 894                         *relocp += delta;
 895                 }
 896                 break;
 897
 898         default_patch:
 899         default:
 900                 ret = paravirt_patch_default(type, clobbers, insnbuf,
 901                                              addr, len);
 902                 break;
 903         }
 904
 905         return ret;
 906 }
 907
 908 static const struct pv_info xen_info __initdata = {
 909         .paravirt_enabled = 1,
 910         .shared_kernel_pmd = 0,
 911
 912         .name = "Xen",
 913 };
 914
 915 static const struct pv_init_ops xen_init_ops __initdata = {
 916         .patch = xen_patch,
 917
 918         .banner = xen_banner,
 919         .memory_setup = xen_memory_setup,
 920         .arch_setup = xen_arch_setup,
 921         .post_allocator_init = xen_mark_init_mm_pinned,
 922 };
 923
 924 static const struct pv_time_ops xen_time_ops __initdata = {
 925         .time_init = xen_time_init,
 926
 927         .set_wallclock = xen_set_wallclock,
 928         .get_wallclock = xen_get_wallclock,
 929         .get_cpu_khz = xen_cpu_khz,
 930         .sched_clock = xen_sched_clock,
 931 };
 932
 933 static const struct pv_cpu_ops xen_cpu_ops __initdata = {
 934         .cpuid = xen_cpuid,
 935
 936         .set_debugreg = xen_set_debugreg,
 937         .get_debugreg = xen_get_debugreg,
 938
 939         .clts = native_clts,
 940
 941         .read_cr0 = native_read_cr0,
 942         .write_cr0 = native_write_cr0,
 943
 944         .read_cr4 = native_read_cr4,
 945         .read_cr4_safe = native_read_cr4_safe,
 946         .write_cr4 = xen_write_cr4,
 947
 948         .wbinvd = native_wbinvd,
 949
 950         .read_msr = native_read_msr_safe,
 951         .write_msr = native_write_msr_safe,
 952         .read_tsc = native_read_tsc,
 953         .read_pmc = native_read_pmc,
 954
 955         .iret = (void *)&hypercall_page[__HYPERVISOR_iret],
 956         .irq_enable_sysexit = NULL,  /* never called */
 957
 958         .load_tr_desc = paravirt_nop,
 959         .set_ldt = xen_set_ldt,
 960         .load_gdt = xen_load_gdt,
 961         .load_idt = xen_load_idt,
 962         .load_tls = xen_load_tls,
 963
 964         .store_gdt = native_store_gdt,
 965         .store_idt = native_store_idt,
 966         .store_tr = xen_store_tr,
 967
 968         .write_ldt_entry = xen_write_ldt_entry,
 969         .write_gdt_entry = xen_write_gdt_entry,
 970         .write_idt_entry = xen_write_idt_entry,
 971         .load_esp0 = xen_load_esp0,
 972
 973         .set_iopl_mask = xen_set_iopl_mask,
 974         .io_delay = xen_io_delay,
 975
 976         .lazy_mode = {
 977                 .enter = paravirt_enter_lazy_cpu,
 978                 .leave = xen_leave_lazy,
 979         },
 980 };
 981
 982 static const struct pv_irq_ops xen_irq_ops __initdata = {
 983         .init_IRQ = xen_init_IRQ,
 984         .save_fl = xen_save_fl,
 985         .restore_fl = xen_restore_fl,
 986         .irq_disable = xen_irq_disable,
 987         .irq_enable = xen_irq_enable,
 988         .safe_halt = xen_safe_halt,
 989         .halt = xen_halt,
 990 };
 991
 992 static const struct pv_apic_ops xen_apic_ops __initdata = {
 993 #ifdef CONFIG_X86_LOCAL_APIC
 994         .apic_write = xen_apic_write,
 995         .apic_write_atomic = xen_apic_write,
 996         .apic_read = xen_apic_read,
 997         .setup_boot_clock = paravirt_nop,
 998         .setup_secondary_clock = paravirt_nop,
 999         .startup_ipi_hook = paravirt_nop,
1000 #endif
1001 };
1002
1003 static const struct pv_mmu_ops xen_mmu_ops __initdata = {
1004         .pagetable_setup_start = xen_pagetable_setup_start,
1005         .pagetable_setup_done = xen_pagetable_setup_done,
1006
1007         .read_cr2 = xen_read_cr2,
1008         .write_cr2 = xen_write_cr2,
1009
1010         .read_cr3 = xen_read_cr3,
1011         .write_cr3 = xen_write_cr3,
1012
1013         .flush_tlb_user = xen_flush_tlb,
1014         .flush_tlb_kernel = xen_flush_tlb,
1015         .flush_tlb_single = xen_flush_tlb_single,
1016         .flush_tlb_others = xen_flush_tlb_others,
1017
1018         .pte_update = paravirt_nop,
1019         .pte_update_defer = paravirt_nop,
1020
1021         .alloc_pt = xen_alloc_pt_init,
1022         .release_pt = xen_release_pt,
1023         .alloc_pd = paravirt_nop,
1024         .alloc_pd_clone = paravirt_nop,
1025         .release_pd = paravirt_nop,
1026
1027 #ifdef CONFIG_HIGHPTE
1028         .kmap_atomic_pte = xen_kmap_atomic_pte,
1029 #endif
1030
1031         .set_pte = NULL,        /* see xen_pagetable_setup_* */
1032         .set_pte_at = xen_set_pte_at,
1033         .set_pmd = xen_set_pmd,
1034
1035         .pte_val = xen_pte_val,
1036         .pgd_val = xen_pgd_val,
1037
1038         .make_pte = xen_make_pte,
1039         .make_pgd = xen_make_pgd,
1040
1041 #ifdef CONFIG_X86_PAE
1042         .set_pte_atomic = xen_set_pte_atomic,
1043         .set_pte_present = xen_set_pte_at,
1044         .set_pud = xen_set_pud,
1045         .pte_clear = xen_pte_clear,
1046         .pmd_clear = xen_pmd_clear,
1047
1048         .make_pmd = xen_make_pmd,
1049         .pmd_val = xen_pmd_val,
1050 #endif  /* PAE */
1051
1052         .activate_mm = xen_activate_mm,
1053         .dup_mmap = xen_dup_mmap,
1054         .exit_mmap = xen_exit_mmap,
1055
1056         .lazy_mode = {
1057                 .enter = paravirt_enter_lazy_mmu,
1058                 .leave = xen_leave_lazy,
1059         },
1060 };
1061
1062 #ifdef CONFIG_SMP
1063 static const struct smp_ops xen_smp_ops __initdata = {
1064         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
1065         .smp_prepare_cpus = xen_smp_prepare_cpus,
1066         .cpu_up = xen_cpu_up,
1067         .smp_cpus_done = xen_smp_cpus_done,
1068
1069         .smp_send_stop = xen_smp_send_stop,
1070         .smp_send_reschedule = xen_smp_send_reschedule,
1071         .smp_call_function_mask = xen_smp_call_function_mask,
1072 };
1073 #endif  /* CONFIG_SMP */
1074
1075 static void xen_reboot(int reason)
1076 {
1077 #ifdef CONFIG_SMP
1078         smp_send_stop();
1079 #endif
1080
1081         if (HYPERVISOR_sched_op(SCHEDOP_shutdown, reason))
1082                 BUG();
1083 }
1084
1085 static void xen_restart(char *msg)
1086 {
1087         xen_reboot(SHUTDOWN_reboot);
1088 }
1089
1090 static void xen_emergency_restart(void)
1091 {
1092         xen_reboot(SHUTDOWN_reboot);
1093 }
1094
1095 static void xen_machine_halt(void)
1096 {
1097         xen_reboot(SHUTDOWN_poweroff);
1098 }
1099
1100 static void xen_crash_shutdown(struct pt_regs *regs)
1101 {
1102         xen_reboot(SHUTDOWN_crash);
1103 }
1104
1105 static const struct machine_ops __initdata xen_machine_ops = {
1106         .restart = xen_restart,
1107         .halt = xen_machine_halt,
1108         .power_off = xen_machine_halt,
1109         .shutdown = xen_machine_halt,
1110         .crash_shutdown = xen_crash_shutdown,
1111         .emergency_restart = xen_emergency_restart,
1112 };
1113
1114
1115 static void __init xen_reserve_top(void)
1116 {
1117         unsigned long top = HYPERVISOR_VIRT_START;
1118         struct xen_platform_parameters pp;
1119
1120         if (HYPERVISOR_xen_version(XENVER_platform_parameters, &pp) == 0)
1121                 top = pp.virt_start;
1122
1123         reserve_top_address(-top + 2 * PAGE_SIZE);
1124 }
1125
1126 /* First C function to be called on Xen boot */
1127 asmlinkage void __init xen_start_kernel(void)
1128 {
1129         pgd_t *pgd;
1130
1131         if (!xen_start_info)
1132                 return;
1133
1134         BUG_ON(memcmp(xen_start_info->magic, "xen-3", 5) != 0);
1135
1136         /* Install Xen paravirt ops */
1137         pv_info = xen_info;
1138         pv_init_ops = xen_init_ops;
1139         pv_time_ops = xen_time_ops;
1140         pv_cpu_ops = xen_cpu_ops;
1141         pv_irq_ops = xen_irq_ops;
1142         pv_apic_ops = xen_apic_ops;
1143         pv_mmu_ops = xen_mmu_ops;
1144
1145         machine_ops = xen_machine_ops;
1146
1147 #ifdef CONFIG_SMP
1148         smp_ops = xen_smp_ops;
1149 #endif
1150
1151         xen_setup_features();
1152
1153         /* Get mfn list */
1154         if (!xen_feature(XENFEAT_auto_translated_physmap))
1155                 phys_to_machine_mapping = (unsigned long *)xen_start_info->mfn_list;
1156
1157         pgd = (pgd_t *)xen_start_info->pt_base;
1158
1159         init_pg_tables_end = __pa(pgd) + xen_start_info->nr_pt_frames*PAGE_SIZE;
1160
1161         init_mm.pgd = pgd; /* use the Xen pagetables to start */
1162
1163         /* keep using Xen gdt for now; no urgent need to change it */
1164
1165         x86_write_percpu(xen_cr3, __pa(pgd));
1166         x86_write_percpu(xen_current_cr3, __pa(pgd));
1167
1168 #ifdef CONFIG_SMP
1169         /* Don't do the full vcpu_info placement stuff until we have a
1170            possible map. */
1171         per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
1172 #else
1173         /* May as well do it now, since there's no good time to call
1174            it later on UP. */
1175         xen_setup_vcpu_info_placement();
1176 #endif
1177
1178         pv_info.kernel_rpl = 1;
1179         if (xen_feature(XENFEAT_supervisor_mode_kernel))
1180                 pv_info.kernel_rpl = 0;
1181
1182         /* set the limit of our address space */
1183         xen_reserve_top();
1184
1185         /* set up basic CPUID stuff */
1186         cpu_detect(&new_cpu_data);
1187         new_cpu_data.hard_math = 1;
1188         new_cpu_data.x86_capability[0] = cpuid_edx(1);
1189
1190         /* Poke various useful things into boot_params */
1191         boot_params.hdr.type_of_loader = (9 << 4) | 0;
1192         boot_params.hdr.ramdisk_image = xen_start_info->mod_start
1193                 ? __pa(xen_start_info->mod_start) : 0;
1194         boot_params.hdr.ramdisk_size = xen_start_info->mod_len;
1195
1196         /* Start the world */
1197         start_kernel();
1198 }