drivers/lguest/x86/core.c

   1 /*
   2  * Copyright (C) 2006, Rusty Russell <rusty@rustcorp.com.au> IBM Corporation.
   3  * Copyright (C) 2007, Jes Sorensen <jes@sgi.com> SGI.
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License as published by
   7  * the Free Software Foundation; either version 2 of the License, or
   8  * (at your option) any later version.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  13  * NON INFRINGEMENT.  See the GNU General Public License for more
  14  * details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20 /*P:450
  21  * This file contains the x86-specific lguest code.  It used to be all
  22  * mixed in with drivers/lguest/core.c but several foolhardy code slashers
  23  * wrestled most of the dependencies out to here in preparation for porting
  24  * lguest to other architectures (see what I mean by foolhardy?).
  25  *
  26  * This also contains a couple of non-obvious setup and teardown pieces which
  27  * were implemented after days of debugging pain.
  28 :*/
  29 #include <linux/kernel.h>
  30 #include <linux/start_kernel.h>
  31 #include <linux/string.h>
  32 #include <linux/console.h>
  33 #include <linux/screen_info.h>
  34 #include <linux/irq.h>
  35 #include <linux/interrupt.h>
  36 #include <linux/clocksource.h>
  37 #include <linux/clockchips.h>
  38 #include <linux/cpu.h>
  39 #include <linux/lguest.h>
  40 #include <linux/lguest_launcher.h>
  41 #include <asm/paravirt.h>
  42 #include <asm/param.h>
  43 #include <asm/page.h>
  44 #include <asm/pgtable.h>
  45 #include <asm/desc.h>
  46 #include <asm/setup.h>
  47 #include <asm/lguest.h>
  48 #include <linux/uaccess.h>
  49 #include <asm/fpu/internal.h>
  50 #include <asm/tlbflush.h>
  51 #include "../lg.h"
  52
  53 static int cpu_had_pge;
  54
  55 static struct {
  56         unsigned long offset;
  57         unsigned short segment;
  58 } lguest_entry;
  59
  60 /* Offset from where switcher.S was compiled to where we've copied it */
  61 static unsigned long switcher_offset(void)
  62 {
  63         return switcher_addr - (unsigned long)start_switcher_text;
  64 }
  65
  66 /* This cpu's struct lguest_pages (after the Switcher text page) */
  67 static struct lguest_pages *lguest_pages(unsigned int cpu)
  68 {
  69         return &(((struct lguest_pages *)(switcher_addr + PAGE_SIZE))[cpu]);
  70 }
  71
  72 static DEFINE_PER_CPU(struct lg_cpu *, lg_last_cpu);
  73
  74 /*S:010
  75  * We approach the Switcher.
  76  *
  77  * Remember that each CPU has two pages which are visible to the Guest when it
  78  * runs on that CPU.  This has to contain the state for that Guest: we copy the
  79  * state in just before we run the Guest.
  80  *
  81  * Each Guest has "changed" flags which indicate what has changed in the Guest
  82  * since it last ran.  We saw this set in interrupts_and_traps.c and
  83  * segments.c.
  84  */
  85 static void copy_in_guest_info(struct lg_cpu *cpu, struct lguest_pages *pages)
  86 {
  87         /*
  88          * Copying all this data can be quite expensive.  We usually run the
  89          * same Guest we ran last time (and that Guest hasn't run anywhere else
  90          * meanwhile).  If that's not the case, we pretend everything in the
  91          * Guest has changed.
  92          */
  93         if (__this_cpu_read(lg_last_cpu) != cpu || cpu->last_pages != pages) {
  94                 __this_cpu_write(lg_last_cpu, cpu);
  95                 cpu->last_pages = pages;
  96                 cpu->changed = CHANGED_ALL;
  97         }
  98
  99         /*
 100          * These copies are pretty cheap, so we do them unconditionally: */
 101         /* Save the current Host top-level page directory.
 102          */
 103         pages->state.host_cr3 = __pa(current->mm->pgd);
 104         /*
 105          * Set up the Guest's page tables to see this CPU's pages (and no
 106          * other CPU's pages).
 107          */
 108         map_switcher_in_guest(cpu, pages);
 109         /*
 110          * Set up the two "TSS" members which tell the CPU what stack to use
 111          * for traps which do directly into the Guest (ie. traps at privilege
 112          * level 1).
 113          */
 114         pages->state.guest_tss.sp1 = cpu->esp1;
 115         pages->state.guest_tss.ss1 = cpu->ss1;
 116
 117         /* Copy direct-to-Guest trap entries. */
 118         if (cpu->changed & CHANGED_IDT)
 119                 copy_traps(cpu, pages->state.guest_idt, default_idt_entries);
 120
 121         /* Copy all GDT entries which the Guest can change. */
 122         if (cpu->changed & CHANGED_GDT)
 123                 copy_gdt(cpu, pages->state.guest_gdt);
 124         /* If only the TLS entries have changed, copy them. */
 125         else if (cpu->changed & CHANGED_GDT_TLS)
 126                 copy_gdt_tls(cpu, pages->state.guest_gdt);
 127
 128         /* Mark the Guest as unchanged for next time. */
 129         cpu->changed = 0;
 130 }
 131
 132 /* Finally: the code to actually call into the Switcher to run the Guest. */
 133 static void run_guest_once(struct lg_cpu *cpu, struct lguest_pages *pages)
 134 {
 135         /* This is a dummy value we need for GCC's sake. */
 136         unsigned int clobber;
 137
 138         /*
 139          * Copy the guest-specific information into this CPU's "struct
 140          * lguest_pages".
 141          */
 142         copy_in_guest_info(cpu, pages);
 143
 144         /*
 145          * Set the trap number to 256 (impossible value).  If we fault while
 146          * switching to the Guest (bad segment registers or bug), this will
 147          * cause us to abort the Guest.
 148          */
 149         cpu->regs->trapnum = 256;
 150
 151         /*
 152          * Now: we push the "eflags" register on the stack, then do an "lcall".
 153          * This is how we change from using the kernel code segment to using
 154          * the dedicated lguest code segment, as well as jumping into the
 155          * Switcher.
 156          *
 157          * The lcall also pushes the old code segment (KERNEL_CS) onto the
 158          * stack, then the address of this call.  This stack layout happens to
 159          * exactly match the stack layout created by an interrupt...
 160          */
 161         asm volatile("pushf; lcall *%4"
 162                      /*
 163                       * This is how we tell GCC that %eax ("a") and %ebx ("b")
 164                       * are changed by this routine.  The "=" means output.
 165                       */
 166                      : "=a"(clobber), "=b"(clobber)
 167                      /*
 168                       * %eax contains the pages pointer.  ("0" refers to the
 169                       * 0-th argument above, ie "a").  %ebx contains the
 170                       * physical address of the Guest's top-level page
 171                       * directory.
 172                       */
 173                      : "0"(pages),
 174                        "1"(__pa(cpu->lg->pgdirs[cpu->cpu_pgd].pgdir)),
 175                        "m"(lguest_entry)
 176                      /*
 177                       * We tell gcc that all these registers could change,
 178                       * which means we don't have to save and restore them in
 179                       * the Switcher.
 180                       */
 181                      : "memory", "%edx", "%ecx", "%edi", "%esi");
 182 }
 183 /*:*/
 184
 185 unsigned long *lguest_arch_regptr(struct lg_cpu *cpu, size_t reg_off, bool any)
 186 {
 187         switch (reg_off) {
 188         case offsetof(struct pt_regs, bx):
 189                 return &cpu->regs->ebx;
 190         case offsetof(struct pt_regs, cx):
 191                 return &cpu->regs->ecx;
 192         case offsetof(struct pt_regs, dx):
 193                 return &cpu->regs->edx;
 194         case offsetof(struct pt_regs, si):
 195                 return &cpu->regs->esi;
 196         case offsetof(struct pt_regs, di):
 197                 return &cpu->regs->edi;
 198         case offsetof(struct pt_regs, bp):
 199                 return &cpu->regs->ebp;
 200         case offsetof(struct pt_regs, ax):
 201                 return &cpu->regs->eax;
 202         case offsetof(struct pt_regs, ip):
 203                 return &cpu->regs->eip;
 204         case offsetof(struct pt_regs, sp):
 205                 return &cpu->regs->esp;
 206         }
 207
 208         /* Launcher can read these, but we don't allow any setting. */
 209         if (any) {
 210                 switch (reg_off) {
 211                 case offsetof(struct pt_regs, ds):
 212                         return &cpu->regs->ds;
 213                 case offsetof(struct pt_regs, es):
 214                         return &cpu->regs->es;
 215                 case offsetof(struct pt_regs, fs):
 216                         return &cpu->regs->fs;
 217                 case offsetof(struct pt_regs, gs):
 218                         return &cpu->regs->gs;
 219                 case offsetof(struct pt_regs, cs):
 220                         return &cpu->regs->cs;
 221                 case offsetof(struct pt_regs, flags):
 222                         return &cpu->regs->eflags;
 223                 case offsetof(struct pt_regs, ss):
 224                         return &cpu->regs->ss;
 225                 }
 226         }
 227
 228         return NULL;
 229 }
 230
 231 /*M:002
 232  * There are hooks in the scheduler which we can register to tell when we
 233  * get kicked off the CPU (preempt_notifier_register()).  This would allow us
 234  * to lazily disable SYSENTER which would regain some performance, and should
 235  * also simplify copy_in_guest_info().  Note that we'd still need to restore
 236  * things when we exit to Launcher userspace, but that's fairly easy.
 237  *
 238  * We could also try using these hooks for PGE, but that might be too expensive.
 239  *
 240  * The hooks were designed for KVM, but we can also put them to good use.
 241 :*/
 242
 243 /*H:040
 244  * This is the i386-specific code to setup and run the Guest.  Interrupts
 245  * are disabled: we own the CPU.
 246  */
 247 void lguest_arch_run_guest(struct lg_cpu *cpu)
 248 {
 249         /*
 250          * SYSENTER is an optimized way of doing system calls.  We can't allow
 251          * it because it always jumps to privilege level 0.  A normal Guest
 252          * won't try it because we don't advertise it in CPUID, but a malicious
 253          * Guest (or malicious Guest userspace program) could, so we tell the
 254          * CPU to disable it before running the Guest.
 255          */
 256         if (boot_cpu_has(X86_FEATURE_SEP))
 257                 wrmsr(MSR_IA32_SYSENTER_CS, 0, 0);
 258
 259         /*
 260          * Now we actually run the Guest.  It will return when something
 261          * interesting happens, and we can examine its registers to see what it
 262          * was doing.
 263          */
 264         run_guest_once(cpu, lguest_pages(raw_smp_processor_id()));
 265
 266         /*
 267          * Note that the "regs" structure contains two extra entries which are
 268          * not really registers: a trap number which says what interrupt or
 269          * trap made the switcher code come back, and an error code which some
 270          * traps set.
 271          */
 272
 273          /* Restore SYSENTER if it's supposed to be on. */
 274          if (boot_cpu_has(X86_FEATURE_SEP))
 275                 wrmsr(MSR_IA32_SYSENTER_CS, __KERNEL_CS, 0);
 276
 277         /*
 278          * If the Guest page faulted, then the cr2 register will tell us the
 279          * bad virtual address.  We have to grab this now, because once we
 280          * re-enable interrupts an interrupt could fault and thus overwrite
 281          * cr2, or we could even move off to a different CPU.
 282          */
 283         if (cpu->regs->trapnum == 14)
 284                 cpu->arch.last_pagefault = read_cr2();
 285         /*
 286          * Similarly, if we took a trap because the Guest used the FPU,
 287          * we have to restore the FPU it expects to see.
 288          * fpu__restore() may sleep and we may even move off to
 289          * a different CPU. So all the critical stuff should be done
 290          * before this.
 291          */
 292         else if (cpu->regs->trapnum == 7 && !fpregs_active())
 293                 fpu__restore(&current->thread.fpu);
 294 }
 295
 296 /*H:130
 297  * Now we've examined the hypercall code; our Guest can make requests.
 298  * Our Guest is usually so well behaved; it never tries to do things it isn't
 299  * allowed to, and uses hypercalls instead.  Unfortunately, Linux's paravirtual
 300  * infrastructure isn't quite complete, because it doesn't contain replacements
 301  * for the Intel I/O instructions.  As a result, the Guest sometimes fumbles
 302  * across one during the boot process as it probes for various things which are
 303  * usually attached to a PC.
 304  *
 305  * When the Guest uses one of these instructions, we get a trap (General
 306  * Protection Fault) and come here.  We queue this to be sent out to the
 307  * Launcher to handle.
 308  */
 309
 310 /*
 311  * The eip contains the *virtual* address of the Guest's instruction:
 312  * we copy the instruction here so the Launcher doesn't have to walk
 313  * the page tables to decode it.  We handle the case (eg. in a kernel
 314  * module) where the instruction is over two pages, and the pages are
 315  * virtually but not physically contiguous.
 316  *
 317  * The longest possible x86 instruction is 15 bytes, but we don't handle
 318  * anything that strange.
 319  */
 320 static void copy_from_guest(struct lg_cpu *cpu,
 321                             void *dst, unsigned long vaddr, size_t len)
 322 {
 323         size_t to_page_end = PAGE_SIZE - (vaddr % PAGE_SIZE);
 324         unsigned long paddr;
 325
 326         BUG_ON(len > PAGE_SIZE);
 327
 328         /* If it goes over a page, copy in two parts. */
 329         if (len > to_page_end) {
 330                 /* But make sure the next page is mapped! */
 331                 if (__guest_pa(cpu, vaddr + to_page_end, &paddr))
 332                         copy_from_guest(cpu, dst + to_page_end,
 333                                         vaddr + to_page_end,
 334                                         len - to_page_end);
 335                 else
 336                         /* Otherwise fill with zeroes. */
 337                         memset(dst + to_page_end, 0, len - to_page_end);
 338                 len = to_page_end;
 339         }
 340
 341         /* This will kill the guest if it isn't mapped, but that
 342          * shouldn't happen. */
 343         __lgread(cpu, dst, guest_pa(cpu, vaddr), len);
 344 }
 345
 346
 347 static void setup_emulate_insn(struct lg_cpu *cpu)
 348 {
 349         cpu->pending.trap = 13;
 350         copy_from_guest(cpu, cpu->pending.insn, cpu->regs->eip,
 351                         sizeof(cpu->pending.insn));
 352 }
 353
 354 static void setup_iomem_insn(struct lg_cpu *cpu, unsigned long iomem_addr)
 355 {
 356         cpu->pending.trap = 14;
 357         cpu->pending.addr = iomem_addr;
 358         copy_from_guest(cpu, cpu->pending.insn, cpu->regs->eip,
 359                         sizeof(cpu->pending.insn));
 360 }
 361
 362 /*H:050 Once we've re-enabled interrupts, we look at why the Guest exited. */
 363 void lguest_arch_handle_trap(struct lg_cpu *cpu)
 364 {
 365         unsigned long iomem_addr;
 366
 367         switch (cpu->regs->trapnum) {
 368         case 13: /* We've intercepted a General Protection Fault. */
 369                 /* Hand to Launcher to emulate those pesky IN and OUT insns */
 370                 if (cpu->regs->errcode == 0) {
 371                         setup_emulate_insn(cpu);
 372                         return;
 373                 }
 374                 break;
 375         case 14: /* We've intercepted a Page Fault. */
 376                 /*
 377                  * The Guest accessed a virtual address that wasn't mapped.
 378                  * This happens a lot: we don't actually set up most of the page
 379                  * tables for the Guest at all when we start: as it runs it asks
 380                  * for more and more, and we set them up as required. In this
 381                  * case, we don't even tell the Guest that the fault happened.
 382                  *
 383                  * The errcode tells whether this was a read or a write, and
 384                  * whether kernel or userspace code.
 385                  */
 386                 if (demand_page(cpu, cpu->arch.last_pagefault,
 387                                 cpu->regs->errcode, &iomem_addr))
 388                         return;
 389
 390                 /* Was this an access to memory mapped IO? */
 391                 if (iomem_addr) {
 392                         /* Tell Launcher, let it handle it. */
 393                         setup_iomem_insn(cpu, iomem_addr);
 394                         return;
 395                 }
 396
 397                 /*
 398                  * OK, it's really not there (or not OK): the Guest needs to
 399                  * know.  We write out the cr2 value so it knows where the
 400                  * fault occurred.
 401                  *
 402                  * Note that if the Guest were really messed up, this could
 403                  * happen before it's done the LHCALL_LGUEST_INIT hypercall, so
 404                  * lg->lguest_data could be NULL
 405                  */
 406                 if (cpu->lg->lguest_data &&
 407                     put_user(cpu->arch.last_pagefault,
 408                              &cpu->lg->lguest_data->cr2))
 409                         kill_guest(cpu, "Writing cr2");
 410                 break;
 411         case 7: /* We've intercepted a Device Not Available fault. */
 412                 /* No special handling is needed here. */
 413                 break;
 414         case 32 ... 255:
 415                 /* This might be a syscall. */
 416                 if (could_be_syscall(cpu->regs->trapnum))
 417                         break;
 418
 419                 /*
 420                  * Other values mean a real interrupt occurred, in which case
 421                  * the Host handler has already been run. We just do a
 422                  * friendly check if another process should now be run, then
 423                  * return to run the Guest again.
 424                  */
 425                 cond_resched();
 426                 return;
 427         case LGUEST_TRAP_ENTRY:
 428                 /*
 429                  * Our 'struct hcall_args' maps directly over our regs: we set
 430                  * up the pointer now to indicate a hypercall is pending.
 431                  */
 432                 cpu->hcall = (struct hcall_args *)cpu->regs;
 433                 return;
 434         }
 435
 436         /* We didn't handle the trap, so it needs to go to the Guest. */
 437         if (!deliver_trap(cpu, cpu->regs->trapnum))
 438                 /*
 439                  * If the Guest doesn't have a handler (either it hasn't
 440                  * registered any yet, or it's one of the faults we don't let
 441                  * it handle), it dies with this cryptic error message.
 442                  */
 443                 kill_guest(cpu, "unhandled trap %li at %#lx (%#lx)",
 444                            cpu->regs->trapnum, cpu->regs->eip,
 445                            cpu->regs->trapnum == 14 ? cpu->arch.last_pagefault
 446                            : cpu->regs->errcode);
 447 }
 448
 449 /*
 450  * Now we can look at each of the routines this calls, in increasing order of
 451  * complexity: do_hypercalls(), emulate_insn(), maybe_do_interrupt(),
 452  * deliver_trap() and demand_page().  After all those, we'll be ready to
 453  * examine the Switcher, and our philosophical understanding of the Host/Guest
 454  * duality will be complete.
 455 :*/
 456 static void adjust_pge(void *on)
 457 {
 458         if (on)
 459                 cr4_set_bits(X86_CR4_PGE);
 460         else
 461                 cr4_clear_bits(X86_CR4_PGE);
 462 }
 463
 464 /*H:020
 465  * Now the Switcher is mapped and every thing else is ready, we need to do
 466  * some more i386-specific initialization.
 467  */
 468 void __init lguest_arch_host_init(void)
 469 {
 470         int i;
 471
 472         /*
 473          * Most of the x86/switcher_32.S doesn't care that it's been moved; on
 474          * Intel, jumps are relative, and it doesn't access any references to
 475          * external code or data.
 476          *
 477          * The only exception is the interrupt handlers in switcher.S: their
 478          * addresses are placed in a table (default_idt_entries), so we need to
 479          * update the table with the new addresses.  switcher_offset() is a
 480          * convenience function which returns the distance between the
 481          * compiled-in switcher code and the high-mapped copy we just made.
 482          */
 483         for (i = 0; i < IDT_ENTRIES; i++)
 484                 default_idt_entries[i] += switcher_offset();
 485
 486         /*
 487          * Set up the Switcher's per-cpu areas.
 488          *
 489          * Each CPU gets two pages of its own within the high-mapped region
 490          * (aka. "struct lguest_pages").  Much of this can be initialized now,
 491          * but some depends on what Guest we are running (which is set up in
 492          * copy_in_guest_info()).
 493          */
 494         for_each_possible_cpu(i) {
 495                 /* lguest_pages() returns this CPU's two pages. */
 496                 struct lguest_pages *pages = lguest_pages(i);
 497                 /* This is a convenience pointer to make the code neater. */
 498                 struct lguest_ro_state *state = &pages->state;
 499
 500                 /*
 501                  * The Global Descriptor Table: the Host has a different one
 502                  * for each CPU.  We keep a descriptor for the GDT which says
 503                  * where it is and how big it is (the size is actually the last
 504                  * byte, not the size, hence the "-1").
 505                  */
 506                 state->host_gdt_desc.size = GDT_SIZE-1;
 507                 state->host_gdt_desc.address = (long)get_cpu_gdt_table(i);
 508
 509                 /*
 510                  * All CPUs on the Host use the same Interrupt Descriptor
 511                  * Table, so we just use store_idt(), which gets this CPU's IDT
 512                  * descriptor.
 513                  */
 514                 store_idt(&state->host_idt_desc);
 515
 516                 /*
 517                  * The descriptors for the Guest's GDT and IDT can be filled
 518                  * out now, too.  We copy the GDT & IDT into ->guest_gdt and
 519                  * ->guest_idt before actually running the Guest.
 520                  */
 521                 state->guest_idt_desc.size = sizeof(state->guest_idt)-1;
 522                 state->guest_idt_desc.address = (long)&state->guest_idt;
 523                 state->guest_gdt_desc.size = sizeof(state->guest_gdt)-1;
 524                 state->guest_gdt_desc.address = (long)&state->guest_gdt;
 525
 526                 /*
 527                  * We know where we want the stack to be when the Guest enters
 528                  * the Switcher: in pages->regs.  The stack grows upwards, so
 529                  * we start it at the end of that structure.
 530                  */
 531                 state->guest_tss.sp0 = (long)(&pages->regs + 1);
 532                 /*
 533                  * And this is the GDT entry to use for the stack: we keep a
 534                  * couple of special LGUEST entries.
 535                  */
 536                 state->guest_tss.ss0 = LGUEST_DS;
 537
 538                 /*
 539                  * x86 can have a finegrained bitmap which indicates what I/O
 540                  * ports the process can use.  We set it to the end of our
 541                  * structure, meaning "none".
 542                  */
 543                 state->guest_tss.io_bitmap_base = sizeof(state->guest_tss);
 544
 545                 /*
 546                  * Some GDT entries are the same across all Guests, so we can
 547                  * set them up now.
 548                  */
 549                 setup_default_gdt_entries(state);
 550                 /* Most IDT entries are the same for all Guests, too.*/
 551                 setup_default_idt_entries(state, default_idt_entries);
 552
 553                 /*
 554                  * The Host needs to be able to use the LGUEST segments on this
 555                  * CPU, too, so put them in the Host GDT.
 556                  */
 557                 get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_CS] = FULL_EXEC_SEGMENT;
 558                 get_cpu_gdt_table(i)[GDT_ENTRY_LGUEST_DS] = FULL_SEGMENT;
 559         }
 560
 561         /*
 562          * In the Switcher, we want the %cs segment register to use the
 563          * LGUEST_CS GDT entry: we've put that in the Host and Guest GDTs, so
 564          * it will be undisturbed when we switch.  To change %cs and jump we
 565          * need this structure to feed to Intel's "lcall" instruction.
 566          */
 567         lguest_entry.offset = (long)switch_to_guest + switcher_offset();
 568         lguest_entry.segment = LGUEST_CS;
 569
 570         /*
 571          * Finally, we need to turn off "Page Global Enable".  PGE is an
 572          * optimization where page table entries are specially marked to show
 573          * they never change.  The Host kernel marks all the kernel pages this
 574          * way because it's always present, even when userspace is running.
 575          *
 576          * Lguest breaks this: unbeknownst to the rest of the Host kernel, we
 577          * switch to the Guest kernel.  If you don't disable this on all CPUs,
 578          * you'll get really weird bugs that you'll chase for two days.
 579          *
 580          * I used to turn PGE off every time we switched to the Guest and back
 581          * on when we return, but that slowed the Switcher down noticibly.
 582          */
 583
 584         /*
 585          * We don't need the complexity of CPUs coming and going while we're
 586          * doing this.
 587          */
 588         get_online_cpus();
 589         if (boot_cpu_has(X86_FEATURE_PGE)) { /* We have a broader idea of "global". */
 590                 /* Remember that this was originally set (for cleanup). */
 591                 cpu_had_pge = 1;
 592                 /*
 593                  * adjust_pge is a helper function which sets or unsets the PGE
 594                  * bit on its CPU, depending on the argument (0 == unset).
 595                  */
 596                 on_each_cpu(adjust_pge, (void *)0, 1);
 597                 /* Turn off the feature in the global feature set. */
 598                 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_PGE);
 599         }
 600         put_online_cpus();
 601 }
 602 /*:*/
 603
 604 void __exit lguest_arch_host_fini(void)
 605 {
 606         /* If we had PGE before we started, turn it back on now. */
 607         get_online_cpus();
 608         if (cpu_had_pge) {
 609                 set_cpu_cap(&boot_cpu_data, X86_FEATURE_PGE);
 610                 /* adjust_pge's argument "1" means set PGE. */
 611                 on_each_cpu(adjust_pge, (void *)1, 1);
 612         }
 613         put_online_cpus();
 614 }
 615
 616
 617 /*H:122 The i386-specific hypercalls simply farm out to the right functions. */
 618 int lguest_arch_do_hcall(struct lg_cpu *cpu, struct hcall_args *args)
 619 {
 620         switch (args->arg0) {
 621         case LHCALL_LOAD_GDT_ENTRY:
 622                 load_guest_gdt_entry(cpu, args->arg1, args->arg2, args->arg3);
 623                 break;
 624         case LHCALL_LOAD_IDT_ENTRY:
 625                 load_guest_idt_entry(cpu, args->arg1, args->arg2, args->arg3);
 626                 break;
 627         case LHCALL_LOAD_TLS:
 628                 guest_load_tls(cpu, args->arg1);
 629                 break;
 630         default:
 631                 /* Bad Guest.  Bad! */
 632                 return -EIO;
 633         }
 634         return 0;
 635 }
 636
 637 /*H:126 i386-specific hypercall initialization: */
 638 int lguest_arch_init_hypercalls(struct lg_cpu *cpu)
 639 {
 640         u32 tsc_speed;
 641
 642         /*
 643          * The pointer to the Guest's "struct lguest_data" is the only argument.
 644          * We check that address now.
 645          */
 646         if (!lguest_address_ok(cpu->lg, cpu->hcall->arg1,
 647                                sizeof(*cpu->lg->lguest_data)))
 648                 return -EFAULT;
 649
 650         /*
 651          * Having checked it, we simply set lg->lguest_data to point straight
 652          * into the Launcher's memory at the right place and then use
 653          * copy_to_user/from_user from now on, instead of lgread/write.  I put
 654          * this in to show that I'm not immune to writing stupid
 655          * optimizations.
 656          */
 657         cpu->lg->lguest_data = cpu->lg->mem_base + cpu->hcall->arg1;
 658
 659         /*
 660          * We insist that the Time Stamp Counter exist and doesn't change with
 661          * cpu frequency.  Some devious chip manufacturers decided that TSC
 662          * changes could be handled in software.  I decided that time going
 663          * backwards might be good for benchmarks, but it's bad for users.
 664          *
 665          * We also insist that the TSC be stable: the kernel detects unreliable
 666          * TSCs for its own purposes, and we use that here.
 667          */
 668         if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC) && !check_tsc_unstable())
 669                 tsc_speed = tsc_khz;
 670         else
 671                 tsc_speed = 0;
 672         if (put_user(tsc_speed, &cpu->lg->lguest_data->tsc_khz))
 673                 return -EFAULT;
 674
 675         /* The interrupt code might not like the system call vector. */
 676         if (!check_syscall_vector(cpu->lg))
 677                 kill_guest(cpu, "bad syscall vector");
 678
 679         return 0;
 680 }
 681 /*:*/
 682
 683 /*L:030
 684  * Most of the Guest's registers are left alone: we used get_zeroed_page() to
 685  * allocate the structure, so they will be 0.
 686  */
 687 void lguest_arch_setup_regs(struct lg_cpu *cpu, unsigned long start)
 688 {
 689         struct lguest_regs *regs = cpu->regs;
 690
 691         /*
 692          * There are four "segment" registers which the Guest needs to boot:
 693          * The "code segment" register (cs) refers to the kernel code segment
 694          * __KERNEL_CS, and the "data", "extra" and "stack" segment registers
 695          * refer to the kernel data segment __KERNEL_DS.
 696          *
 697          * The privilege level is packed into the lower bits.  The Guest runs
 698          * at privilege level 1 (GUEST_PL).
 699          */
 700         regs->ds = regs->es = regs->ss = __KERNEL_DS|GUEST_PL;
 701         regs->cs = __KERNEL_CS|GUEST_PL;
 702
 703         /*
 704          * The "eflags" register contains miscellaneous flags.  Bit 1 (0x002)
 705          * is supposed to always be "1".  Bit 9 (0x200) controls whether
 706          * interrupts are enabled.  We always leave interrupts enabled while
 707          * running the Guest.
 708          */
 709         regs->eflags = X86_EFLAGS_IF | X86_EFLAGS_FIXED;
 710
 711         /*
 712          * The "Extended Instruction Pointer" register says where the Guest is
 713          * running.
 714          */
 715         regs->eip = start;
 716
 717         /*
 718          * %esi points to our boot information, at physical address 0, so don't
 719          * touch it.
 720          */
 721
 722         /* There are a couple of GDT entries the Guest expects at boot. */
 723         setup_guest_gdt(cpu);
 724 }