linux/drivers/lguest/page_tables.c
<<
>>
Prefs
   1/*P:700
   2 * The pagetable code, on the other hand, still shows the scars of
   3 * previous encounters.  It's functional, and as neat as it can be in the
   4 * circumstances, but be wary, for these things are subtle and break easily.
   5 * The Guest provides a virtual to physical mapping, but we can neither trust
   6 * it nor use it: we verify and convert it here then point the CPU to the
   7 * converted Guest pages when running the Guest.
   8:*/
   9
  10/* Copyright (C) Rusty Russell IBM Corporation 2006.
  11 * GPL v2 and any later version */
  12#include <linux/mm.h>
  13#include <linux/gfp.h>
  14#include <linux/types.h>
  15#include <linux/spinlock.h>
  16#include <linux/random.h>
  17#include <linux/percpu.h>
  18#include <asm/tlbflush.h>
  19#include <asm/uaccess.h>
  20#include "lg.h"
  21
  22/*M:008
  23 * We hold reference to pages, which prevents them from being swapped.
  24 * It'd be nice to have a callback in the "struct mm_struct" when Linux wants
  25 * to swap out.  If we had this, and a shrinker callback to trim PTE pages, we
  26 * could probably consider launching Guests as non-root.
  27:*/
  28
  29/*H:300
  30 * The Page Table Code
  31 *
  32 * We use two-level page tables for the Guest, or three-level with PAE.  If
  33 * you're not entirely comfortable with virtual addresses, physical addresses
  34 * and page tables then I recommend you review arch/x86/lguest/boot.c's "Page
  35 * Table Handling" (with diagrams!).
  36 *
  37 * The Guest keeps page tables, but we maintain the actual ones here: these are
  38 * called "shadow" page tables.  Which is a very Guest-centric name: these are
  39 * the real page tables the CPU uses, although we keep them up to date to
  40 * reflect the Guest's.  (See what I mean about weird naming?  Since when do
  41 * shadows reflect anything?)
  42 *
  43 * Anyway, this is the most complicated part of the Host code.  There are seven
  44 * parts to this:
  45 *  (i) Looking up a page table entry when the Guest faults,
  46 *  (ii) Making sure the Guest stack is mapped,
  47 *  (iii) Setting up a page table entry when the Guest tells us one has changed,
  48 *  (iv) Switching page tables,
  49 *  (v) Flushing (throwing away) page tables,
  50 *  (vi) Mapping the Switcher when the Guest is about to run,
  51 *  (vii) Setting up the page tables initially.
  52:*/
  53
  54/*
  55 * The Switcher uses the complete top PTE page.  That's 1024 PTE entries (4MB)
  56 * or 512 PTE entries with PAE (2MB).
  57 */
  58#define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1)
  59
  60/*
  61 * For PAE we need the PMD index as well. We use the last 2MB, so we
  62 * will need the last pmd entry of the last pmd page.
  63 */
  64#ifdef CONFIG_X86_PAE
  65#define SWITCHER_PMD_INDEX      (PTRS_PER_PMD - 1)
  66#define RESERVE_MEM             2U
  67#define CHECK_GPGD_MASK         _PAGE_PRESENT
  68#else
  69#define RESERVE_MEM             4U
  70#define CHECK_GPGD_MASK         _PAGE_TABLE
  71#endif
  72
  73/*
  74 * We actually need a separate PTE page for each CPU.  Remember that after the
  75 * Switcher code itself comes two pages for each CPU, and we don't want this
  76 * CPU's guest to see the pages of any other CPU.
  77 */
  78static DEFINE_PER_CPU(pte_t *, switcher_pte_pages);
  79#define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu)
  80
  81/*H:320
  82 * The page table code is curly enough to need helper functions to keep it
  83 * clear and clean.  The kernel itself provides many of them; one advantage
  84 * of insisting that the Guest and Host use the same CONFIG_PAE setting.
  85 *
  86 * There are two functions which return pointers to the shadow (aka "real")
  87 * page tables.
  88 *
  89 * spgd_addr() takes the virtual address and returns a pointer to the top-level
  90 * page directory entry (PGD) for that address.  Since we keep track of several
  91 * page tables, the "i" argument tells us which one we're interested in (it's
  92 * usually the current one).
  93 */
  94static pgd_t *spgd_addr(struct lg_cpu *cpu, u32 i, unsigned long vaddr)
  95{
  96        unsigned int index = pgd_index(vaddr);
  97
  98#ifndef CONFIG_X86_PAE
  99        /* We kill any Guest trying to touch the Switcher addresses. */
 100        if (index >= SWITCHER_PGD_INDEX) {
 101                kill_guest(cpu, "attempt to access switcher pages");
 102                index = 0;
 103        }
 104#endif
 105        /* Return a pointer index'th pgd entry for the i'th page table. */
 106        return &cpu->lg->pgdirs[i].pgdir[index];
 107}
 108
 109#ifdef CONFIG_X86_PAE
 110/*
 111 * This routine then takes the PGD entry given above, which contains the
 112 * address of the PMD page.  It then returns a pointer to the PMD entry for the
 113 * given address.
 114 */
 115static pmd_t *spmd_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr)
 116{
 117        unsigned int index = pmd_index(vaddr);
 118        pmd_t *page;
 119
 120        /* We kill any Guest trying to touch the Switcher addresses. */
 121        if (pgd_index(vaddr) == SWITCHER_PGD_INDEX &&
 122                                        index >= SWITCHER_PMD_INDEX) {
 123                kill_guest(cpu, "attempt to access switcher pages");
 124                index = 0;
 125        }
 126
 127        /* You should never call this if the PGD entry wasn't valid */
 128        BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT));
 129        page = __va(pgd_pfn(spgd) << PAGE_SHIFT);
 130
 131        return &page[index];
 132}
 133#endif
 134
 135/*
 136 * This routine then takes the page directory entry returned above, which
 137 * contains the address of the page table entry (PTE) page.  It then returns a
 138 * pointer to the PTE entry for the given address.
 139 */
 140static pte_t *spte_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr)
 141{
 142#ifdef CONFIG_X86_PAE
 143        pmd_t *pmd = spmd_addr(cpu, spgd, vaddr);
 144        pte_t *page = __va(pmd_pfn(*pmd) << PAGE_SHIFT);
 145
 146        /* You should never call this if the PMD entry wasn't valid */
 147        BUG_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT));
 148#else
 149        pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT);
 150        /* You should never call this if the PGD entry wasn't valid */
 151        BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT));
 152#endif
 153
 154        return &page[pte_index(vaddr)];
 155}
 156
 157/*
 158 * These functions are just like the above, except they access the Guest
 159 * page tables.  Hence they return a Guest address.
 160 */
 161static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr)
 162{
 163        unsigned int index = vaddr >> (PGDIR_SHIFT);
 164        return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index * sizeof(pgd_t);
 165}
 166
 167#ifdef CONFIG_X86_PAE
 168/* Follow the PGD to the PMD. */
 169static unsigned long gpmd_addr(pgd_t gpgd, unsigned long vaddr)
 170{
 171        unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT;
 172        BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT));
 173        return gpage + pmd_index(vaddr) * sizeof(pmd_t);
 174}
 175
 176/* Follow the PMD to the PTE. */
 177static unsigned long gpte_addr(struct lg_cpu *cpu,
 178                               pmd_t gpmd, unsigned long vaddr)
 179{
 180        unsigned long gpage = pmd_pfn(gpmd) << PAGE_SHIFT;
 181
 182        BUG_ON(!(pmd_flags(gpmd) & _PAGE_PRESENT));
 183        return gpage + pte_index(vaddr) * sizeof(pte_t);
 184}
 185#else
 186/* Follow the PGD to the PTE (no mid-level for !PAE). */
 187static unsigned long gpte_addr(struct lg_cpu *cpu,
 188                                pgd_t gpgd, unsigned long vaddr)
 189{
 190        unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT;
 191
 192        BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT));
 193        return gpage + pte_index(vaddr) * sizeof(pte_t);
 194}
 195#endif
 196/*:*/
 197
 198/*M:007
 199 * get_pfn is slow: we could probably try to grab batches of pages here as
 200 * an optimization (ie. pre-faulting).
 201:*/
 202
 203/*H:350
 204 * This routine takes a page number given by the Guest and converts it to
 205 * an actual, physical page number.  It can fail for several reasons: the
 206 * virtual address might not be mapped by the Launcher, the write flag is set
 207 * and the page is read-only, or the write flag was set and the page was
 208 * shared so had to be copied, but we ran out of memory.
 209 *
 210 * This holds a reference to the page, so release_pte() is careful to put that
 211 * back.
 212 */
 213static unsigned long get_pfn(unsigned long virtpfn, int write)
 214{
 215        struct page *page;
 216
 217        /* gup me one page at this address please! */
 218        if (get_user_pages_fast(virtpfn << PAGE_SHIFT, 1, write, &page) == 1)
 219                return page_to_pfn(page);
 220
 221        /* This value indicates failure. */
 222        return -1UL;
 223}
 224
 225/*H:340
 226 * Converting a Guest page table entry to a shadow (ie. real) page table
 227 * entry can be a little tricky.  The flags are (almost) the same, but the
 228 * Guest PTE contains a virtual page number: the CPU needs the real page
 229 * number.
 230 */
 231static pte_t gpte_to_spte(struct lg_cpu *cpu, pte_t gpte, int write)
 232{
 233        unsigned long pfn, base, flags;
 234
 235        /*
 236         * The Guest sets the global flag, because it thinks that it is using
 237         * PGE.  We only told it to use PGE so it would tell us whether it was
 238         * flushing a kernel mapping or a userspace mapping.  We don't actually
 239         * use the global bit, so throw it away.
 240         */
 241        flags = (pte_flags(gpte) & ~_PAGE_GLOBAL);
 242
 243        /* The Guest's pages are offset inside the Launcher. */
 244        base = (unsigned long)cpu->lg->mem_base / PAGE_SIZE;
 245
 246        /*
 247         * We need a temporary "unsigned long" variable to hold the answer from
 248         * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't
 249         * fit in spte.pfn.  get_pfn() finds the real physical number of the
 250         * page, given the virtual number.
 251         */
 252        pfn = get_pfn(base + pte_pfn(gpte), write);
 253        if (pfn == -1UL) {
 254                kill_guest(cpu, "failed to get page %lu", pte_pfn(gpte));
 255                /*
 256                 * When we destroy the Guest, we'll go through the shadow page
 257                 * tables and release_pte() them.  Make sure we don't think
 258                 * this one is valid!
 259                 */
 260                flags = 0;
 261        }
 262        /* Now we assemble our shadow PTE from the page number and flags. */
 263        return pfn_pte(pfn, __pgprot(flags));
 264}
 265
 266/*H:460 And to complete the chain, release_pte() looks like this: */
 267static void release_pte(pte_t pte)
 268{
 269        /*
 270         * Remember that get_user_pages_fast() took a reference to the page, in
 271         * get_pfn()?  We have to put it back now.
 272         */
 273        if (pte_flags(pte) & _PAGE_PRESENT)
 274                put_page(pte_page(pte));
 275}
 276/*:*/
 277
 278static void check_gpte(struct lg_cpu *cpu, pte_t gpte)
 279{
 280        if ((pte_flags(gpte) & _PAGE_PSE) ||
 281            pte_pfn(gpte) >= cpu->lg->pfn_limit)
 282                kill_guest(cpu, "bad page table entry");
 283}
 284
 285static void check_gpgd(struct lg_cpu *cpu, pgd_t gpgd)
 286{
 287        if ((pgd_flags(gpgd) & ~CHECK_GPGD_MASK) ||
 288           (pgd_pfn(gpgd) >= cpu->lg->pfn_limit))
 289                kill_guest(cpu, "bad page directory entry");
 290}
 291
 292#ifdef CONFIG_X86_PAE
 293static void check_gpmd(struct lg_cpu *cpu, pmd_t gpmd)
 294{
 295        if ((pmd_flags(gpmd) & ~_PAGE_TABLE) ||
 296           (pmd_pfn(gpmd) >= cpu->lg->pfn_limit))
 297                kill_guest(cpu, "bad page middle directory entry");
 298}
 299#endif
 300
 301/*H:330
 302 * (i) Looking up a page table entry when the Guest faults.
 303 *
 304 * We saw this call in run_guest(): when we see a page fault in the Guest, we
 305 * come here.  That's because we only set up the shadow page tables lazily as
 306 * they're needed, so we get page faults all the time and quietly fix them up
 307 * and return to the Guest without it knowing.
 308 *
 309 * If we fixed up the fault (ie. we mapped the address), this routine returns
 310 * true.  Otherwise, it was a real fault and we need to tell the Guest.
 311 */
 312bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode)
 313{
 314        pgd_t gpgd;
 315        pgd_t *spgd;
 316        unsigned long gpte_ptr;
 317        pte_t gpte;
 318        pte_t *spte;
 319
 320        /* Mid level for PAE. */
 321#ifdef CONFIG_X86_PAE
 322        pmd_t *spmd;
 323        pmd_t gpmd;
 324#endif
 325
 326        /* First step: get the top-level Guest page table entry. */
 327        if (unlikely(cpu->linear_pages)) {
 328                /* Faking up a linear mapping. */
 329                gpgd = __pgd(CHECK_GPGD_MASK);
 330        } else {
 331                gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t);
 332                /* Toplevel not present?  We can't map it in. */
 333                if (!(pgd_flags(gpgd) & _PAGE_PRESENT))
 334                        return false;
 335        }
 336
 337        /* Now look at the matching shadow entry. */
 338        spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr);
 339        if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) {
 340                /* No shadow entry: allocate a new shadow PTE page. */
 341                unsigned long ptepage = get_zeroed_page(GFP_KERNEL);
 342                /*
 343                 * This is not really the Guest's fault, but killing it is
 344                 * simple for this corner case.
 345                 */
 346                if (!ptepage) {
 347                        kill_guest(cpu, "out of memory allocating pte page");
 348                        return false;
 349                }
 350                /* We check that the Guest pgd is OK. */
 351                check_gpgd(cpu, gpgd);
 352                /*
 353                 * And we copy the flags to the shadow PGD entry.  The page
 354                 * number in the shadow PGD is the page we just allocated.
 355                 */
 356                set_pgd(spgd, __pgd(__pa(ptepage) | pgd_flags(gpgd)));
 357        }
 358
 359#ifdef CONFIG_X86_PAE
 360        if (unlikely(cpu->linear_pages)) {
 361                /* Faking up a linear mapping. */
 362                gpmd = __pmd(_PAGE_TABLE);
 363        } else {
 364                gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t);
 365                /* Middle level not present?  We can't map it in. */
 366                if (!(pmd_flags(gpmd) & _PAGE_PRESENT))
 367                        return false;
 368        }
 369
 370        /* Now look at the matching shadow entry. */
 371        spmd = spmd_addr(cpu, *spgd, vaddr);
 372
 373        if (!(pmd_flags(*spmd) & _PAGE_PRESENT)) {
 374                /* No shadow entry: allocate a new shadow PTE page. */
 375                unsigned long ptepage = get_zeroed_page(GFP_KERNEL);
 376
 377                /*
 378                 * This is not really the Guest's fault, but killing it is
 379                 * simple for this corner case.
 380                 */
 381                if (!ptepage) {
 382                        kill_guest(cpu, "out of memory allocating pte page");
 383                        return false;
 384                }
 385
 386                /* We check that the Guest pmd is OK. */
 387                check_gpmd(cpu, gpmd);
 388
 389                /*
 390                 * And we copy the flags to the shadow PMD entry.  The page
 391                 * number in the shadow PMD is the page we just allocated.
 392                 */
 393                set_pmd(spmd, __pmd(__pa(ptepage) | pmd_flags(gpmd)));
 394        }
 395
 396        /*
 397         * OK, now we look at the lower level in the Guest page table: keep its
 398         * address, because we might update it later.
 399         */
 400        gpte_ptr = gpte_addr(cpu, gpmd, vaddr);
 401#else
 402        /*
 403         * OK, now we look at the lower level in the Guest page table: keep its
 404         * address, because we might update it later.
 405         */
 406        gpte_ptr = gpte_addr(cpu, gpgd, vaddr);
 407#endif
 408
 409        if (unlikely(cpu->linear_pages)) {
 410                /* Linear?  Make up a PTE which points to same page. */
 411                gpte = __pte((vaddr & PAGE_MASK) | _PAGE_RW | _PAGE_PRESENT);
 412        } else {
 413                /* Read the actual PTE value. */
 414                gpte = lgread(cpu, gpte_ptr, pte_t);
 415        }
 416
 417        /* If this page isn't in the Guest page tables, we can't page it in. */
 418        if (!(pte_flags(gpte) & _PAGE_PRESENT))
 419                return false;
 420
 421        /*
 422         * Check they're not trying to write to a page the Guest wants
 423         * read-only (bit 2 of errcode == write).
 424         */
 425        if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW))
 426                return false;
 427
 428        /* User access to a kernel-only page? (bit 3 == user access) */
 429        if ((errcode & 4) && !(pte_flags(gpte) & _PAGE_USER))
 430                return false;
 431
 432        /*
 433         * Check that the Guest PTE flags are OK, and the page number is below
 434         * the pfn_limit (ie. not mapping the Launcher binary).
 435         */
 436        check_gpte(cpu, gpte);
 437
 438        /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */
 439        gpte = pte_mkyoung(gpte);
 440        if (errcode & 2)
 441                gpte = pte_mkdirty(gpte);
 442
 443        /* Get the pointer to the shadow PTE entry we're going to set. */
 444        spte = spte_addr(cpu, *spgd, vaddr);
 445
 446        /*
 447         * If there was a valid shadow PTE entry here before, we release it.
 448         * This can happen with a write to a previously read-only entry.
 449         */
 450        release_pte(*spte);
 451
 452        /*
 453         * If this is a write, we insist that the Guest page is writable (the
 454         * final arg to gpte_to_spte()).
 455         */
 456        if (pte_dirty(gpte))
 457                *spte = gpte_to_spte(cpu, gpte, 1);
 458        else
 459                /*
 460                 * If this is a read, don't set the "writable" bit in the page
 461                 * table entry, even if the Guest says it's writable.  That way
 462                 * we will come back here when a write does actually occur, so
 463                 * we can update the Guest's _PAGE_DIRTY flag.
 464                 */
 465                set_pte(spte, gpte_to_spte(cpu, pte_wrprotect(gpte), 0));
 466
 467        /*
 468         * Finally, we write the Guest PTE entry back: we've set the
 469         * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags.
 470         */
 471        if (likely(!cpu->linear_pages))
 472                lgwrite(cpu, gpte_ptr, pte_t, gpte);
 473
 474        /*
 475         * The fault is fixed, the page table is populated, the mapping
 476         * manipulated, the result returned and the code complete.  A small
 477         * delay and a trace of alliteration are the only indications the Guest
 478         * has that a page fault occurred at all.
 479         */
 480        return true;
 481}
 482
 483/*H:360
 484 * (ii) Making sure the Guest stack is mapped.
 485 *
 486 * Remember that direct traps into the Guest need a mapped Guest kernel stack.
 487 * pin_stack_pages() calls us here: we could simply call demand_page(), but as
 488 * we've seen that logic is quite long, and usually the stack pages are already
 489 * mapped, so it's overkill.
 490 *
 491 * This is a quick version which answers the question: is this virtual address
 492 * mapped by the shadow page tables, and is it writable?
 493 */
 494static bool page_writable(struct lg_cpu *cpu, unsigned long vaddr)
 495{
 496        pgd_t *spgd;
 497        unsigned long flags;
 498
 499#ifdef CONFIG_X86_PAE
 500        pmd_t *spmd;
 501#endif
 502        /* Look at the current top level entry: is it present? */
 503        spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr);
 504        if (!(pgd_flags(*spgd) & _PAGE_PRESENT))
 505                return false;
 506
 507#ifdef CONFIG_X86_PAE
 508        spmd = spmd_addr(cpu, *spgd, vaddr);
 509        if (!(pmd_flags(*spmd) & _PAGE_PRESENT))
 510                return false;
 511#endif
 512
 513        /*
 514         * Check the flags on the pte entry itself: it must be present and
 515         * writable.
 516         */
 517        flags = pte_flags(*(spte_addr(cpu, *spgd, vaddr)));
 518
 519        return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW);
 520}
 521
 522/*
 523 * So, when pin_stack_pages() asks us to pin a page, we check if it's already
 524 * in the page tables, and if not, we call demand_page() with error code 2
 525 * (meaning "write").
 526 */
 527void pin_page(struct lg_cpu *cpu, unsigned long vaddr)
 528{
 529        if (!page_writable(cpu, vaddr) && !demand_page(cpu, vaddr, 2))
 530                kill_guest(cpu, "bad stack page %#lx", vaddr);
 531}
 532/*:*/
 533
 534#ifdef CONFIG_X86_PAE
 535static void release_pmd(pmd_t *spmd)
 536{
 537        /* If the entry's not present, there's nothing to release. */
 538        if (pmd_flags(*spmd) & _PAGE_PRESENT) {
 539                unsigned int i;
 540                pte_t *ptepage = __va(pmd_pfn(*spmd) << PAGE_SHIFT);
 541                /* For each entry in the page, we might need to release it. */
 542                for (i = 0; i < PTRS_PER_PTE; i++)
 543                        release_pte(ptepage[i]);
 544                /* Now we can free the page of PTEs */
 545                free_page((long)ptepage);
 546                /* And zero out the PMD entry so we never release it twice. */
 547                set_pmd(spmd, __pmd(0));
 548        }
 549}
 550
 551static void release_pgd(pgd_t *spgd)
 552{
 553        /* If the entry's not present, there's nothing to release. */
 554        if (pgd_flags(*spgd) & _PAGE_PRESENT) {
 555                unsigned int i;
 556                pmd_t *pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT);
 557
 558                for (i = 0; i < PTRS_PER_PMD; i++)
 559                        release_pmd(&pmdpage[i]);
 560
 561                /* Now we can free the page of PMDs */
 562                free_page((long)pmdpage);
 563                /* And zero out the PGD entry so we never release it twice. */
 564                set_pgd(spgd, __pgd(0));
 565        }
 566}
 567
 568#else /* !CONFIG_X86_PAE */
 569/*H:450
 570 * If we chase down the release_pgd() code, the non-PAE version looks like
 571 * this.  The PAE version is almost identical, but instead of calling
 572 * release_pte it calls release_pmd(), which looks much like this.
 573 */
 574static void release_pgd(pgd_t *spgd)
 575{
 576        /* If the entry's not present, there's nothing to release. */
 577        if (pgd_flags(*spgd) & _PAGE_PRESENT) {
 578                unsigned int i;
 579                /*
 580                 * Converting the pfn to find the actual PTE page is easy: turn
 581                 * the page number into a physical address, then convert to a
 582                 * virtual address (easy for kernel pages like this one).
 583                 */
 584                pte_t *ptepage = __va(pgd_pfn(*spgd) << PAGE_SHIFT);
 585                /* For each entry in the page, we might need to release it. */
 586                for (i = 0; i < PTRS_PER_PTE; i++)
 587                        release_pte(ptepage[i]);
 588                /* Now we can free the page of PTEs */
 589                free_page((long)ptepage);
 590                /* And zero out the PGD entry so we never release it twice. */
 591                *spgd = __pgd(0);
 592        }
 593}
 594#endif
 595
 596/*H:445
 597 * We saw flush_user_mappings() twice: once from the flush_user_mappings()
 598 * hypercall and once in new_pgdir() when we re-used a top-level pgdir page.
 599 * It simply releases every PTE page from 0 up to the Guest's kernel address.
 600 */
 601static void flush_user_mappings(struct lguest *lg, int idx)
 602{
 603        unsigned int i;
 604        /* Release every pgd entry up to the kernel's address. */
 605        for (i = 0; i < pgd_index(lg->kernel_address); i++)
 606                release_pgd(lg->pgdirs[idx].pgdir + i);
 607}
 608
 609/*H:440
 610 * (v) Flushing (throwing away) page tables,
 611 *
 612 * The Guest has a hypercall to throw away the page tables: it's used when a
 613 * large number of mappings have been changed.
 614 */
 615void guest_pagetable_flush_user(struct lg_cpu *cpu)
 616{
 617        /* Drop the userspace part of the current page table. */
 618        flush_user_mappings(cpu->lg, cpu->cpu_pgd);
 619}
 620/*:*/
 621
 622/* We walk down the guest page tables to get a guest-physical address */
 623unsigned long guest_pa(struct lg_cpu *cpu, unsigned long vaddr)
 624{
 625        pgd_t gpgd;
 626        pte_t gpte;
 627#ifdef CONFIG_X86_PAE
 628        pmd_t gpmd;
 629#endif
 630
 631        /* Still not set up?  Just map 1:1. */
 632        if (unlikely(cpu->linear_pages))
 633                return vaddr;
 634
 635        /* First step: get the top-level Guest page table entry. */
 636        gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t);
 637        /* Toplevel not present?  We can't map it in. */
 638        if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) {
 639                kill_guest(cpu, "Bad address %#lx", vaddr);
 640                return -1UL;
 641        }
 642
 643#ifdef CONFIG_X86_PAE
 644        gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t);
 645        if (!(pmd_flags(gpmd) & _PAGE_PRESENT))
 646                kill_guest(cpu, "Bad address %#lx", vaddr);
 647        gpte = lgread(cpu, gpte_addr(cpu, gpmd, vaddr), pte_t);
 648#else
 649        gpte = lgread(cpu, gpte_addr(cpu, gpgd, vaddr), pte_t);
 650#endif
 651        if (!(pte_flags(gpte) & _PAGE_PRESENT))
 652                kill_guest(cpu, "Bad address %#lx", vaddr);
 653
 654        return pte_pfn(gpte) * PAGE_SIZE | (vaddr & ~PAGE_MASK);
 655}
 656
 657/*
 658 * We keep several page tables.  This is a simple routine to find the page
 659 * table (if any) corresponding to this top-level address the Guest has given
 660 * us.
 661 */
 662static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable)
 663{
 664        unsigned int i;
 665        for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
 666                if (lg->pgdirs[i].pgdir && lg->pgdirs[i].gpgdir == pgtable)
 667                        break;
 668        return i;
 669}
 670
 671/*H:435
 672 * And this is us, creating the new page directory.  If we really do
 673 * allocate a new one (and so the kernel parts are not there), we set
 674 * blank_pgdir.
 675 */
 676static unsigned int new_pgdir(struct lg_cpu *cpu,
 677                              unsigned long gpgdir,
 678                              int *blank_pgdir)
 679{
 680        unsigned int next;
 681#ifdef CONFIG_X86_PAE
 682        pmd_t *pmd_table;
 683#endif
 684
 685        /*
 686         * We pick one entry at random to throw out.  Choosing the Least
 687         * Recently Used might be better, but this is easy.
 688         */
 689        next = random32() % ARRAY_SIZE(cpu->lg->pgdirs);
 690        /* If it's never been allocated at all before, try now. */
 691        if (!cpu->lg->pgdirs[next].pgdir) {
 692                cpu->lg->pgdirs[next].pgdir =
 693                                        (pgd_t *)get_zeroed_page(GFP_KERNEL);
 694                /* If the allocation fails, just keep using the one we have */
 695                if (!cpu->lg->pgdirs[next].pgdir)
 696                        next = cpu->cpu_pgd;
 697                else {
 698#ifdef CONFIG_X86_PAE
 699                        /*
 700                         * In PAE mode, allocate a pmd page and populate the
 701                         * last pgd entry.
 702                         */
 703                        pmd_table = (pmd_t *)get_zeroed_page(GFP_KERNEL);
 704                        if (!pmd_table) {
 705                                free_page((long)cpu->lg->pgdirs[next].pgdir);
 706                                set_pgd(cpu->lg->pgdirs[next].pgdir, __pgd(0));
 707                                next = cpu->cpu_pgd;
 708                        } else {
 709                                set_pgd(cpu->lg->pgdirs[next].pgdir +
 710                                        SWITCHER_PGD_INDEX,
 711                                        __pgd(__pa(pmd_table) | _PAGE_PRESENT));
 712                                /*
 713                                 * This is a blank page, so there are no kernel
 714                                 * mappings: caller must map the stack!
 715                                 */
 716                                *blank_pgdir = 1;
 717                        }
 718#else
 719                        *blank_pgdir = 1;
 720#endif
 721                }
 722        }
 723        /* Record which Guest toplevel this shadows. */
 724        cpu->lg->pgdirs[next].gpgdir = gpgdir;
 725        /* Release all the non-kernel mappings. */
 726        flush_user_mappings(cpu->lg, next);
 727
 728        return next;
 729}
 730
 731/*H:470
 732 * Finally, a routine which throws away everything: all PGD entries in all
 733 * the shadow page tables, including the Guest's kernel mappings.  This is used
 734 * when we destroy the Guest.
 735 */
 736static void release_all_pagetables(struct lguest *lg)
 737{
 738        unsigned int i, j;
 739
 740        /* Every shadow pagetable this Guest has */
 741        for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
 742                if (lg->pgdirs[i].pgdir) {
 743#ifdef CONFIG_X86_PAE
 744                        pgd_t *spgd;
 745                        pmd_t *pmdpage;
 746                        unsigned int k;
 747
 748                        /* Get the last pmd page. */
 749                        spgd = lg->pgdirs[i].pgdir + SWITCHER_PGD_INDEX;
 750                        pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT);
 751
 752                        /*
 753                         * And release the pmd entries of that pmd page,
 754                         * except for the switcher pmd.
 755                         */
 756                        for (k = 0; k < SWITCHER_PMD_INDEX; k++)
 757                                release_pmd(&pmdpage[k]);
 758#endif
 759                        /* Every PGD entry except the Switcher at the top */
 760                        for (j = 0; j < SWITCHER_PGD_INDEX; j++)
 761                                release_pgd(lg->pgdirs[i].pgdir + j);
 762                }
 763}
 764
 765/*
 766 * We also throw away everything when a Guest tells us it's changed a kernel
 767 * mapping.  Since kernel mappings are in every page table, it's easiest to
 768 * throw them all away.  This traps the Guest in amber for a while as
 769 * everything faults back in, but it's rare.
 770 */
 771void guest_pagetable_clear_all(struct lg_cpu *cpu)
 772{
 773        release_all_pagetables(cpu->lg);
 774        /* We need the Guest kernel stack mapped again. */
 775        pin_stack_pages(cpu);
 776}
 777
 778/*H:430
 779 * (iv) Switching page tables
 780 *
 781 * Now we've seen all the page table setting and manipulation, let's see
 782 * what happens when the Guest changes page tables (ie. changes the top-level
 783 * pgdir).  This occurs on almost every context switch.
 784 */
 785void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable)
 786{
 787        int newpgdir, repin = 0;
 788
 789        /*
 790         * The very first time they call this, we're actually running without
 791         * any page tables; we've been making it up.  Throw them away now.
 792         */
 793        if (unlikely(cpu->linear_pages)) {
 794                release_all_pagetables(cpu->lg);
 795                cpu->linear_pages = false;
 796                /* Force allocation of a new pgdir. */
 797                newpgdir = ARRAY_SIZE(cpu->lg->pgdirs);
 798        } else {
 799                /* Look to see if we have this one already. */
 800                newpgdir = find_pgdir(cpu->lg, pgtable);
 801        }
 802
 803        /*
 804         * If not, we allocate or mug an existing one: if it's a fresh one,
 805         * repin gets set to 1.
 806         */
 807        if (newpgdir == ARRAY_SIZE(cpu->lg->pgdirs))
 808                newpgdir = new_pgdir(cpu, pgtable, &repin);
 809        /* Change the current pgd index to the new one. */
 810        cpu->cpu_pgd = newpgdir;
 811        /* If it was completely blank, we map in the Guest kernel stack */
 812        if (repin)
 813                pin_stack_pages(cpu);
 814}
 815/*:*/
 816
 817/*M:009
 818 * Since we throw away all mappings when a kernel mapping changes, our
 819 * performance sucks for guests using highmem.  In fact, a guest with
 820 * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is
 821 * usually slower than a Guest with less memory.
 822 *
 823 * This, of course, cannot be fixed.  It would take some kind of... well, I
 824 * don't know, but the term "puissant code-fu" comes to mind.
 825:*/
 826
 827/*H:420
 828 * This is the routine which actually sets the page table entry for then
 829 * "idx"'th shadow page table.
 830 *
 831 * Normally, we can just throw out the old entry and replace it with 0: if they
 832 * use it demand_page() will put the new entry in.  We need to do this anyway:
 833 * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page
 834 * is read from, and _PAGE_DIRTY when it's written to.
 835 *
 836 * But Avi Kivity pointed out that most Operating Systems (Linux included) set
 837 * these bits on PTEs immediately anyway.  This is done to save the CPU from
 838 * having to update them, but it helps us the same way: if they set
 839 * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if
 840 * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately.
 841 */
 842static void do_set_pte(struct lg_cpu *cpu, int idx,
 843                       unsigned long vaddr, pte_t gpte)
 844{
 845        /* Look up the matching shadow page directory entry. */
 846        pgd_t *spgd = spgd_addr(cpu, idx, vaddr);
 847#ifdef CONFIG_X86_PAE
 848        pmd_t *spmd;
 849#endif
 850
 851        /* If the top level isn't present, there's no entry to update. */
 852        if (pgd_flags(*spgd) & _PAGE_PRESENT) {
 853#ifdef CONFIG_X86_PAE
 854                spmd = spmd_addr(cpu, *spgd, vaddr);
 855                if (pmd_flags(*spmd) & _PAGE_PRESENT) {
 856#endif
 857                        /* Otherwise, start by releasing the existing entry. */
 858                        pte_t *spte = spte_addr(cpu, *spgd, vaddr);
 859                        release_pte(*spte);
 860
 861                        /*
 862                         * If they're setting this entry as dirty or accessed,
 863                         * we might as well put that entry they've given us in
 864                         * now.  This shaves 10% off a copy-on-write
 865                         * micro-benchmark.
 866                         */
 867                        if (pte_flags(gpte) & (_PAGE_DIRTY | _PAGE_ACCESSED)) {
 868                                check_gpte(cpu, gpte);
 869                                set_pte(spte,
 870                                        gpte_to_spte(cpu, gpte,
 871                                                pte_flags(gpte) & _PAGE_DIRTY));
 872                        } else {
 873                                /*
 874                                 * Otherwise kill it and we can demand_page()
 875                                 * it in later.
 876                                 */
 877                                set_pte(spte, __pte(0));
 878                        }
 879#ifdef CONFIG_X86_PAE
 880                }
 881#endif
 882        }
 883}
 884
 885/*H:410
 886 * Updating a PTE entry is a little trickier.
 887 *
 888 * We keep track of several different page tables (the Guest uses one for each
 889 * process, so it makes sense to cache at least a few).  Each of these have
 890 * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for
 891 * all processes.  So when the page table above that address changes, we update
 892 * all the page tables, not just the current one.  This is rare.
 893 *
 894 * The benefit is that when we have to track a new page table, we can keep all
 895 * the kernel mappings.  This speeds up context switch immensely.
 896 */
 897void guest_set_pte(struct lg_cpu *cpu,
 898                   unsigned long gpgdir, unsigned long vaddr, pte_t gpte)
 899{
 900        /*
 901         * Kernel mappings must be changed on all top levels.  Slow, but doesn't
 902         * happen often.
 903         */
 904        if (vaddr >= cpu->lg->kernel_address) {
 905                unsigned int i;
 906                for (i = 0; i < ARRAY_SIZE(cpu->lg->pgdirs); i++)
 907                        if (cpu->lg->pgdirs[i].pgdir)
 908                                do_set_pte(cpu, i, vaddr, gpte);
 909        } else {
 910                /* Is this page table one we have a shadow for? */
 911                int pgdir = find_pgdir(cpu->lg, gpgdir);
 912                if (pgdir != ARRAY_SIZE(cpu->lg->pgdirs))
 913                        /* If so, do the update. */
 914                        do_set_pte(cpu, pgdir, vaddr, gpte);
 915        }
 916}
 917
 918/*H:400
 919 * (iii) Setting up a page table entry when the Guest tells us one has changed.
 920 *
 921 * Just like we did in interrupts_and_traps.c, it makes sense for us to deal
 922 * with the other side of page tables while we're here: what happens when the
 923 * Guest asks for a page table to be updated?
 924 *
 925 * We already saw that demand_page() will fill in the shadow page tables when
 926 * needed, so we can simply remove shadow page table entries whenever the Guest
 927 * tells us they've changed.  When the Guest tries to use the new entry it will
 928 * fault and demand_page() will fix it up.
 929 *
 930 * So with that in mind here's our code to update a (top-level) PGD entry:
 931 */
 932void guest_set_pgd(struct lguest *lg, unsigned long gpgdir, u32 idx)
 933{
 934        int pgdir;
 935
 936        if (idx >= SWITCHER_PGD_INDEX)
 937                return;
 938
 939        /* If they're talking about a page table we have a shadow for... */
 940        pgdir = find_pgdir(lg, gpgdir);
 941        if (pgdir < ARRAY_SIZE(lg->pgdirs))
 942                /* ... throw it away. */
 943                release_pgd(lg->pgdirs[pgdir].pgdir + idx);
 944}
 945
 946#ifdef CONFIG_X86_PAE
 947/* For setting a mid-level, we just throw everything away.  It's easy. */
 948void guest_set_pmd(struct lguest *lg, unsigned long pmdp, u32 idx)
 949{
 950        guest_pagetable_clear_all(&lg->cpus[0]);
 951}
 952#endif
 953
 954/*H:500
 955 * (vii) Setting up the page tables initially.
 956 *
 957 * When a Guest is first created, set initialize a shadow page table which
 958 * we will populate on future faults.  The Guest doesn't have any actual
 959 * pagetables yet, so we set linear_pages to tell demand_page() to fake it
 960 * for the moment.
 961 */
 962int init_guest_pagetable(struct lguest *lg)
 963{
 964        struct lg_cpu *cpu = &lg->cpus[0];
 965        int allocated = 0;
 966
 967        /* lg (and lg->cpus[]) starts zeroed: this allocates a new pgdir */
 968        cpu->cpu_pgd = new_pgdir(cpu, 0, &allocated);
 969        if (!allocated)
 970                return -ENOMEM;
 971
 972        /* We start with a linear mapping until the initialize. */
 973        cpu->linear_pages = true;
 974        return 0;
 975}
 976
 977/*H:508 When the Guest calls LHCALL_LGUEST_INIT we do more setup. */
 978void page_table_guest_data_init(struct lg_cpu *cpu)
 979{
 980        /* We get the kernel address: above this is all kernel memory. */
 981        if (get_user(cpu->lg->kernel_address,
 982                &cpu->lg->lguest_data->kernel_address)
 983                /*
 984                 * We tell the Guest that it can't use the top 2 or 4 MB
 985                 * of virtual addresses used by the Switcher.
 986                 */
 987                || put_user(RESERVE_MEM * 1024 * 1024,
 988                            &cpu->lg->lguest_data->reserve_mem)) {
 989                kill_guest(cpu, "bad guest page %p", cpu->lg->lguest_data);
 990                return;
 991        }
 992
 993        /*
 994         * In flush_user_mappings() we loop from 0 to
 995         * "pgd_index(lg->kernel_address)".  This assumes it won't hit the
 996         * Switcher mappings, so check that now.
 997         */
 998#ifdef CONFIG_X86_PAE
 999        if (pgd_index(cpu->lg->kernel_address) == SWITCHER_PGD_INDEX &&
1000                pmd_index(cpu->lg->kernel_address) == SWITCHER_PMD_INDEX)
1001#else
1002        if (pgd_index(cpu->lg->kernel_address) >= SWITCHER_PGD_INDEX)
1003#endif
1004                kill_guest(cpu, "bad kernel address %#lx",
1005                                 cpu->lg->kernel_address);
1006}
1007
1008/* When a Guest dies, our cleanup is fairly simple. */
1009void free_guest_pagetable(struct lguest *lg)
1010{
1011        unsigned int i;
1012
1013        /* Throw away all page table pages. */
1014        release_all_pagetables(lg);
1015        /* Now free the top levels: free_page() can handle 0 just fine. */
1016        for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
1017                free_page((long)lg->pgdirs[i].pgdir);
1018}
1019
1020/*H:480
1021 * (vi) Mapping the Switcher when the Guest is about to run.
1022 *
1023 * The Switcher and the two pages for this CPU need to be visible in the
1024 * Guest (and not the pages for other CPUs).  We have the appropriate PTE pages
1025 * for each CPU already set up, we just need to hook them in now we know which
1026 * Guest is about to run on this CPU.
1027 */
1028void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages)
1029{
1030        pte_t *switcher_pte_page = __this_cpu_read(switcher_pte_pages);
1031        pte_t regs_pte;
1032
1033#ifdef CONFIG_X86_PAE
1034        pmd_t switcher_pmd;
1035        pmd_t *pmd_table;
1036
1037        switcher_pmd = pfn_pmd(__pa(switcher_pte_page) >> PAGE_SHIFT,
1038                               PAGE_KERNEL_EXEC);
1039
1040        /* Figure out where the pmd page is, by reading the PGD, and converting
1041         * it to a virtual address. */
1042        pmd_table = __va(pgd_pfn(cpu->lg->
1043                        pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX])
1044                                                                << PAGE_SHIFT);
1045        /* Now write it into the shadow page table. */
1046        set_pmd(&pmd_table[SWITCHER_PMD_INDEX], switcher_pmd);
1047#else
1048        pgd_t switcher_pgd;
1049
1050        /*
1051         * Make the last PGD entry for this Guest point to the Switcher's PTE
1052         * page for this CPU (with appropriate flags).
1053         */
1054        switcher_pgd = __pgd(__pa(switcher_pte_page) | __PAGE_KERNEL_EXEC);
1055
1056        cpu->lg->pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd;
1057
1058#endif
1059        /*
1060         * We also change the Switcher PTE page.  When we're running the Guest,
1061         * we want the Guest's "regs" page to appear where the first Switcher
1062         * page for this CPU is.  This is an optimization: when the Switcher
1063         * saves the Guest registers, it saves them into the first page of this
1064         * CPU's "struct lguest_pages": if we make sure the Guest's register
1065         * page is already mapped there, we don't have to copy them out
1066         * again.
1067         */
1068        regs_pte = pfn_pte(__pa(cpu->regs_page) >> PAGE_SHIFT, PAGE_KERNEL);
1069        set_pte(&switcher_pte_page[pte_index((unsigned long)pages)], regs_pte);
1070}
1071/*:*/
1072
1073static void free_switcher_pte_pages(void)
1074{
1075        unsigned int i;
1076
1077        for_each_possible_cpu(i)
1078                free_page((long)switcher_pte_page(i));
1079}
1080
1081/*H:520
1082 * Setting up the Switcher PTE page for given CPU is fairly easy, given
1083 * the CPU number and the "struct page"s for the Switcher code itself.
1084 *
1085 * Currently the Switcher is less than a page long, so "pages" is always 1.
1086 */
1087static __init void populate_switcher_pte_page(unsigned int cpu,
1088                                              struct page *switcher_page[],
1089                                              unsigned int pages)
1090{
1091        unsigned int i;
1092        pte_t *pte = switcher_pte_page(cpu);
1093
1094        /* The first entries are easy: they map the Switcher code. */
1095        for (i = 0; i < pages; i++) {
1096                set_pte(&pte[i], mk_pte(switcher_page[i],
1097                                __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED)));
1098        }
1099
1100        /* The only other thing we map is this CPU's pair of pages. */
1101        i = pages + cpu*2;
1102
1103        /* First page (Guest registers) is writable from the Guest */
1104        set_pte(&pte[i], pfn_pte(page_to_pfn(switcher_page[i]),
1105                         __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW)));
1106
1107        /*
1108         * The second page contains the "struct lguest_ro_state", and is
1109         * read-only.
1110         */
1111        set_pte(&pte[i+1], pfn_pte(page_to_pfn(switcher_page[i+1]),
1112                           __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED)));
1113}
1114
1115/*
1116 * We've made it through the page table code.  Perhaps our tired brains are
1117 * still processing the details, or perhaps we're simply glad it's over.
1118 *
1119 * If nothing else, note that all this complexity in juggling shadow page tables
1120 * in sync with the Guest's page tables is for one reason: for most Guests this
1121 * page table dance determines how bad performance will be.  This is why Xen
1122 * uses exotic direct Guest pagetable manipulation, and why both Intel and AMD
1123 * have implemented shadow page table support directly into hardware.
1124 *
1125 * There is just one file remaining in the Host.
1126 */
1127
1128/*H:510
1129 * At boot or module load time, init_pagetables() allocates and populates
1130 * the Switcher PTE page for each CPU.
1131 */
1132__init int init_pagetables(struct page **switcher_page, unsigned int pages)
1133{
1134        unsigned int i;
1135
1136        for_each_possible_cpu(i) {
1137                switcher_pte_page(i) = (pte_t *)get_zeroed_page(GFP_KERNEL);
1138                if (!switcher_pte_page(i)) {
1139                        free_switcher_pte_pages();
1140                        return -ENOMEM;
1141                }
1142                populate_switcher_pte_page(i, switcher_page, pages);
1143        }
1144        return 0;
1145}
1146/*:*/
1147
1148/* Cleaning up simply involves freeing the PTE page for each CPU. */
1149void free_pagetables(void)
1150{
1151        free_switcher_pte_pages();
1152}
1153
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.