drivers/lguest/page_tables.c

   1 /*P:700 The pagetable code, on the other hand, still shows the scars of
   2  * previous encounters.  It's functional, and as neat as it can be in the
   3  * circumstances, but be wary, for these things are subtle and break easily.
   4  * The Guest provides a virtual to physical mapping, but we can neither trust
   5  * it nor use it: we verify and convert it here then point the CPU to the
   6  * converted Guest pages when running the Guest. :*/
   7
   8 /* Copyright (C) Rusty Russell IBM Corporation 2006.
   9  * GPL v2 and any later version */
  10 #include <linux/mm.h>
  11 #include <linux/types.h>
  12 #include <linux/spinlock.h>
  13 #include <linux/random.h>
  14 #include <linux/percpu.h>
  15 #include <asm/tlbflush.h>
  16 #include <asm/uaccess.h>
  17 #include "lg.h"
  18
  19 /*M:008 We hold reference to pages, which prevents them from being swapped.
  20  * It'd be nice to have a callback in the "struct mm_struct" when Linux wants
  21  * to swap out.  If we had this, and a shrinker callback to trim PTE pages, we
  22  * could probably consider launching Guests as non-root. :*/
  23
  24 /*H:300
  25  * The Page Table Code
  26  *
  27  * We use two-level page tables for the Guest.  If you're not entirely
  28  * comfortable with virtual addresses, physical addresses and page tables then
  29  * I recommend you review arch/x86/lguest/boot.c's "Page Table Handling" (with
  30  * diagrams!).
  31  *
  32  * The Guest keeps page tables, but we maintain the actual ones here: these are
  33  * called "shadow" page tables.  Which is a very Guest-centric name: these are
  34  * the real page tables the CPU uses, although we keep them up to date to
  35  * reflect the Guest's.  (See what I mean about weird naming?  Since when do
  36  * shadows reflect anything?)
  37  *
  38  * Anyway, this is the most complicated part of the Host code.  There are seven
  39  * parts to this:
  40  *  (i) Looking up a page table entry when the Guest faults,
  41  *  (ii) Making sure the Guest stack is mapped,
  42  *  (iii) Setting up a page table entry when the Guest tells us one has changed,
  43  *  (iv) Switching page tables,
  44  *  (v) Flushing (throwing away) page tables,
  45  *  (vi) Mapping the Switcher when the Guest is about to run,
  46  *  (vii) Setting up the page tables initially.
  47  :*/
  48
  49
  50 /* 1024 entries in a page table page maps 1024 pages: 4MB.  The Switcher is
  51  * conveniently placed at the top 4MB, so it uses a separate, complete PTE
  52  * page.  */
  53 #define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1)
  54
  55 /* We actually need a separate PTE page for each CPU.  Remember that after the
  56  * Switcher code itself comes two pages for each CPU, and we don't want this
  57  * CPU's guest to see the pages of any other CPU. */
  58 static DEFINE_PER_CPU(pte_t *, switcher_pte_pages);
  59 #define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu)
  60
  61 /*H:320 The page table code is curly enough to need helper functions to keep it
  62  * clear and clean.
  63  *
  64  * There are two functions which return pointers to the shadow (aka "real")
  65  * page tables.
  66  *
  67  * spgd_addr() takes the virtual address and returns a pointer to the top-level
  68  * page directory entry (PGD) for that address.  Since we keep track of several
  69  * page tables, the "i" argument tells us which one we're interested in (it's
  70  * usually the current one). */
  71 static pgd_t *spgd_addr(struct lg_cpu *cpu, u32 i, unsigned long vaddr)
  72 {
  73         unsigned int index = pgd_index(vaddr);
  74
  75         /* We kill any Guest trying to touch the Switcher addresses. */
  76         if (index >= SWITCHER_PGD_INDEX) {
  77                 kill_guest(cpu, "attempt to access switcher pages");
  78                 index = 0;
  79         }
  80         /* Return a pointer index'th pgd entry for the i'th page table. */
  81         return &cpu->lg->pgdirs[i].pgdir[index];
  82 }
  83
  84 /* This routine then takes the page directory entry returned above, which
  85  * contains the address of the page table entry (PTE) page.  It then returns a
  86  * pointer to the PTE entry for the given address. */
  87 static pte_t *spte_addr(pgd_t spgd, unsigned long vaddr)
  88 {
  89         pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT);
  90         /* You should never call this if the PGD entry wasn't valid */
  91         BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT));
  92         return &page[(vaddr >> PAGE_SHIFT) % PTRS_PER_PTE];
  93 }
  94
  95 /* These two functions just like the above two, except they access the Guest
  96  * page tables.  Hence they return a Guest address. */
  97 static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr)
  98 {
  99         unsigned int index = vaddr >> (PGDIR_SHIFT);
 100         return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index * sizeof(pgd_t);
 101 }
 102
 103 static unsigned long gpte_addr(pgd_t gpgd, unsigned long vaddr)
 104 {
 105         unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT;
 106         BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT));
 107         return gpage + ((vaddr>>PAGE_SHIFT) % PTRS_PER_PTE) * sizeof(pte_t);
 108 }
 109 /*:*/
 110
 111 /*M:014 get_pfn is slow; it takes the mmap sem and calls get_user_pages.  We
 112  * could probably try to grab batches of pages here as an optimization
 113  * (ie. pre-faulting). :*/
 114
 115 /*H:350 This routine takes a page number given by the Guest and converts it to
 116  * an actual, physical page number.  It can fail for several reasons: the
 117  * virtual address might not be mapped by the Launcher, the write flag is set
 118  * and the page is read-only, or the write flag was set and the page was
 119  * shared so had to be copied, but we ran out of memory.
 120  *
 121  * This holds a reference to the page, so release_pte() is careful to put that
 122  * back. */
 123 static unsigned long get_pfn(unsigned long virtpfn, int write)
 124 {
 125         struct page *page;
 126         /* This value indicates failure. */
 127         unsigned long ret = -1UL;
 128
 129         /* get_user_pages() is a complex interface: it gets the "struct
 130          * vm_area_struct" and "struct page" assocated with a range of pages.
 131          * It also needs the task's mmap_sem held, and is not very quick.
 132          * It returns the number of pages it got. */
 133         down_read(&current->mm->mmap_sem);
 134         if (get_user_pages(current, current->mm, virtpfn << PAGE_SHIFT,
 135                            1, write, 1, &page, NULL) == 1)
 136                 ret = page_to_pfn(page);
 137         up_read(&current->mm->mmap_sem);
 138         return ret;
 139 }
 140
 141 /*H:340 Converting a Guest page table entry to a shadow (ie. real) page table
 142  * entry can be a little tricky.  The flags are (almost) the same, but the
 143  * Guest PTE contains a virtual page number: the CPU needs the real page
 144  * number. */
 145 static pte_t gpte_to_spte(struct lg_cpu *cpu, pte_t gpte, int write)
 146 {
 147         unsigned long pfn, base, flags;
 148
 149         /* The Guest sets the global flag, because it thinks that it is using
 150          * PGE.  We only told it to use PGE so it would tell us whether it was
 151          * flushing a kernel mapping or a userspace mapping.  We don't actually
 152          * use the global bit, so throw it away. */
 153         flags = (pte_flags(gpte) & ~_PAGE_GLOBAL);
 154
 155         /* The Guest's pages are offset inside the Launcher. */
 156         base = (unsigned long)cpu->lg->mem_base / PAGE_SIZE;
 157
 158         /* We need a temporary "unsigned long" variable to hold the answer from
 159          * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't
 160          * fit in spte.pfn.  get_pfn() finds the real physical number of the
 161          * page, given the virtual number. */
 162         pfn = get_pfn(base + pte_pfn(gpte), write);
 163         if (pfn == -1UL) {
 164                 kill_guest(cpu, "failed to get page %lu", pte_pfn(gpte));
 165                 /* When we destroy the Guest, we'll go through the shadow page
 166                  * tables and release_pte() them.  Make sure we don't think
 167                  * this one is valid! */
 168                 flags = 0;
 169         }
 170         /* Now we assemble our shadow PTE from the page number and flags. */
 171         return pfn_pte(pfn, __pgprot(flags));
 172 }
 173
 174 /*H:460 And to complete the chain, release_pte() looks like this: */
 175 static void release_pte(pte_t pte)
 176 {
 177         /* Remember that get_user_pages() took a reference to the page, in
 178          * get_pfn()?  We have to put it back now. */
 179         if (pte_flags(pte) & _PAGE_PRESENT)
 180                 put_page(pfn_to_page(pte_pfn(pte)));
 181 }
 182 /*:*/
 183
 184 static void check_gpte(struct lg_cpu *cpu, pte_t gpte)
 185 {
 186         if ((pte_flags(gpte) & _PAGE_PSE) ||
 187             pte_pfn(gpte) >= cpu->lg->pfn_limit)
 188                 kill_guest(cpu, "bad page table entry");
 189 }
 190
 191 static void check_gpgd(struct lg_cpu *cpu, pgd_t gpgd)
 192 {
 193         if ((pgd_flags(gpgd) & ~_PAGE_TABLE) ||
 194            (pgd_pfn(gpgd) >= cpu->lg->pfn_limit))
 195                 kill_guest(cpu, "bad page directory entry");
 196 }
 197
 198 /*H:330
 199  * (i) Looking up a page table entry when the Guest faults.
 200  *
 201  * We saw this call in run_guest(): when we see a page fault in the Guest, we
 202  * come here.  That's because we only set up the shadow page tables lazily as
 203  * they're needed, so we get page faults all the time and quietly fix them up
 204  * and return to the Guest without it knowing.
 205  *
 206  * If we fixed up the fault (ie. we mapped the address), this routine returns
 207  * true.  Otherwise, it was a real fault and we need to tell the Guest. */
 208 int demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode)
 209 {
 210         pgd_t gpgd;
 211         pgd_t *spgd;
 212         unsigned long gpte_ptr;
 213         pte_t gpte;
 214         pte_t *spte;
 215
 216         /* First step: get the top-level Guest page table entry. */
 217         gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t);
 218         /* Toplevel not present?  We can't map it in. */
 219         if (!(pgd_flags(gpgd) & _PAGE_PRESENT))
 220                 return 0;
 221
 222         /* Now look at the matching shadow entry. */
 223         spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr);
 224         if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) {
 225                 /* No shadow entry: allocate a new shadow PTE page. */
 226                 unsigned long ptepage = get_zeroed_page(GFP_KERNEL);
 227                 /* This is not really the Guest's fault, but killing it is
 228                  * simple for this corner case. */
 229                 if (!ptepage) {
 230                         kill_guest(cpu, "out of memory allocating pte page");
 231                         return 0;
 232                 }
 233                 /* We check that the Guest pgd is OK. */
 234                 check_gpgd(cpu, gpgd);
 235                 /* And we copy the flags to the shadow PGD entry.  The page
 236                  * number in the shadow PGD is the page we just allocated. */
 237                 *spgd = __pgd(__pa(ptepage) | pgd_flags(gpgd));
 238         }
 239
 240         /* OK, now we look at the lower level in the Guest page table: keep its
 241          * address, because we might update it later. */
 242         gpte_ptr = gpte_addr(gpgd, vaddr);
 243         gpte = lgread(cpu, gpte_ptr, pte_t);
 244
 245         /* If this page isn't in the Guest page tables, we can't page it in. */
 246         if (!(pte_flags(gpte) & _PAGE_PRESENT))
 247                 return 0;
 248
 249         /* Check they're not trying to write to a page the Guest wants
 250          * read-only (bit 2 of errcode == write). */
 251         if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW))
 252                 return 0;
 253
 254         /* User access to a kernel-only page? (bit 3 == user access) */
 255         if ((errcode & 4) && !(pte_flags(gpte) & _PAGE_USER))
 256                 return 0;
 257
 258         /* Check that the Guest PTE flags are OK, and the page number is below
 259          * the pfn_limit (ie. not mapping the Launcher binary). */
 260         check_gpte(cpu, gpte);
 261
 262         /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */
 263         gpte = pte_mkyoung(gpte);
 264         if (errcode & 2)
 265                 gpte = pte_mkdirty(gpte);
 266
 267         /* Get the pointer to the shadow PTE entry we're going to set. */
 268         spte = spte_addr(*spgd, vaddr);
 269         /* If there was a valid shadow PTE entry here before, we release it.
 270          * This can happen with a write to a previously read-only entry. */
 271         release_pte(*spte);
 272
 273         /* If this is a write, we insist that the Guest page is writable (the
 274          * final arg to gpte_to_spte()). */
 275         if (pte_dirty(gpte))
 276                 *spte = gpte_to_spte(cpu, gpte, 1);
 277         else
 278                 /* If this is a read, don't set the "writable" bit in the page
 279                  * table entry, even if the Guest says it's writable.  That way
 280                  * we will come back here when a write does actually occur, so
 281                  * we can update the Guest's _PAGE_DIRTY flag. */
 282                 *spte = gpte_to_spte(cpu, pte_wrprotect(gpte), 0);
 283
 284         /* Finally, we write the Guest PTE entry back: we've set the
 285          * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. */
 286         lgwrite(cpu, gpte_ptr, pte_t, gpte);
 287
 288         /* The fault is fixed, the page table is populated, the mapping
 289          * manipulated, the result returned and the code complete.  A small
 290          * delay and a trace of alliteration are the only indications the Guest
 291          * has that a page fault occurred at all. */
 292         return 1;
 293 }
 294
 295 /*H:360
 296  * (ii) Making sure the Guest stack is mapped.
 297  *
 298  * Remember that direct traps into the Guest need a mapped Guest kernel stack.
 299  * pin_stack_pages() calls us here: we could simply call demand_page(), but as
 300  * we've seen that logic is quite long, and usually the stack pages are already
 301  * mapped, so it's overkill.
 302  *
 303  * This is a quick version which answers the question: is this virtual address
 304  * mapped by the shadow page tables, and is it writable? */
 305 static int page_writable(struct lg_cpu *cpu, unsigned long vaddr)
 306 {
 307         pgd_t *spgd;
 308         unsigned long flags;
 309
 310         /* Look at the current top level entry: is it present? */
 311         spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr);
 312         if (!(pgd_flags(*spgd) & _PAGE_PRESENT))
 313                 return 0;
 314
 315         /* Check the flags on the pte entry itself: it must be present and
 316          * writable. */
 317         flags = pte_flags(*(spte_addr(*spgd, vaddr)));
 318
 319         return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW);
 320 }
 321
 322 /* So, when pin_stack_pages() asks us to pin a page, we check if it's already
 323  * in the page tables, and if not, we call demand_page() with error code 2
 324  * (meaning "write"). */
 325 void pin_page(struct lg_cpu *cpu, unsigned long vaddr)
 326 {
 327         if (!page_writable(cpu, vaddr) && !demand_page(cpu, vaddr, 2))
 328                 kill_guest(cpu, "bad stack page %#lx", vaddr);
 329 }
 330
 331 /*H:450 If we chase down the release_pgd() code, it looks like this: */
 332 static void release_pgd(struct lguest *lg, pgd_t *spgd)
 333 {
 334         /* If the entry's not present, there's nothing to release. */
 335         if (pgd_flags(*spgd) & _PAGE_PRESENT) {
 336                 unsigned int i;
 337                 /* Converting the pfn to find the actual PTE page is easy: turn
 338                  * the page number into a physical address, then convert to a
 339                  * virtual address (easy for kernel pages like this one). */
 340                 pte_t *ptepage = __va(pgd_pfn(*spgd) << PAGE_SHIFT);
 341                 /* For each entry in the page, we might need to release it. */
 342                 for (i = 0; i < PTRS_PER_PTE; i++)
 343                         release_pte(ptepage[i]);
 344                 /* Now we can free the page of PTEs */
 345                 free_page((long)ptepage);
 346                 /* And zero out the PGD entry so we never release it twice. */
 347                 *spgd = __pgd(0);
 348         }
 349 }
 350
 351 /*H:445 We saw flush_user_mappings() twice: once from the flush_user_mappings()
 352  * hypercall and once in new_pgdir() when we re-used a top-level pgdir page.
 353  * It simply releases every PTE page from 0 up to the Guest's kernel address. */
 354 static void flush_user_mappings(struct lguest *lg, int idx)
 355 {
 356         unsigned int i;
 357         /* Release every pgd entry up to the kernel's address. */
 358         for (i = 0; i < pgd_index(lg->kernel_address); i++)
 359                 release_pgd(lg, lg->pgdirs[idx].pgdir + i);
 360 }
 361
 362 /*H:440 (v) Flushing (throwing away) page tables,
 363  *
 364  * The Guest has a hypercall to throw away the page tables: it's used when a
 365  * large number of mappings have been changed. */
 366 void guest_pagetable_flush_user(struct lg_cpu *cpu)
 367 {
 368         /* Drop the userspace part of the current page table. */
 369         flush_user_mappings(cpu->lg, cpu->cpu_pgd);
 370 }
 371 /*:*/
 372
 373 /* We walk down the guest page tables to get a guest-physical address */
 374 unsigned long guest_pa(struct lg_cpu *cpu, unsigned long vaddr)
 375 {
 376         pgd_t gpgd;
 377         pte_t gpte;
 378
 379         /* First step: get the top-level Guest page table entry. */
 380         gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t);
 381         /* Toplevel not present?  We can't map it in. */
 382         if (!(pgd_flags(gpgd) & _PAGE_PRESENT))
 383                 kill_guest(cpu, "Bad address %#lx", vaddr);
 384
 385         gpte = lgread(cpu, gpte_addr(gpgd, vaddr), pte_t);
 386         if (!(pte_flags(gpte) & _PAGE_PRESENT))
 387                 kill_guest(cpu, "Bad address %#lx", vaddr);
 388
 389         return pte_pfn(gpte) * PAGE_SIZE | (vaddr & ~PAGE_MASK);
 390 }
 391
 392 /* We keep several page tables.  This is a simple routine to find the page
 393  * table (if any) corresponding to this top-level address the Guest has given
 394  * us. */
 395 static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable)
 396 {
 397         unsigned int i;
 398         for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
 399                 if (lg->pgdirs[i].pgdir && lg->pgdirs[i].gpgdir == pgtable)
 400                         break;
 401         return i;
 402 }
 403
 404 /*H:435 And this is us, creating the new page directory.  If we really do
 405  * allocate a new one (and so the kernel parts are not there), we set
 406  * blank_pgdir. */
 407 static unsigned int new_pgdir(struct lg_cpu *cpu,
 408                               unsigned long gpgdir,
 409                               int *blank_pgdir)
 410 {
 411         unsigned int next;
 412
 413         /* We pick one entry at random to throw out.  Choosing the Least
 414          * Recently Used might be better, but this is easy. */
 415         next = random32() % ARRAY_SIZE(cpu->lg->pgdirs);
 416         /* If it's never been allocated at all before, try now. */
 417         if (!cpu->lg->pgdirs[next].pgdir) {
 418                 cpu->lg->pgdirs[next].pgdir =
 419                                         (pgd_t *)get_zeroed_page(GFP_KERNEL);
 420                 /* If the allocation fails, just keep using the one we have */
 421                 if (!cpu->lg->pgdirs[next].pgdir)
 422                         next = cpu->cpu_pgd;
 423                 else
 424                         /* This is a blank page, so there are no kernel
 425                          * mappings: caller must map the stack! */
 426                         *blank_pgdir = 1;
 427         }
 428         /* Record which Guest toplevel this shadows. */
 429         cpu->lg->pgdirs[next].gpgdir = gpgdir;
 430         /* Release all the non-kernel mappings. */
 431         flush_user_mappings(cpu->lg, next);
 432
 433         return next;
 434 }
 435
 436 /*H:430 (iv) Switching page tables
 437  *
 438  * Now we've seen all the page table setting and manipulation, let's see what
 439  * what happens when the Guest changes page tables (ie. changes the top-level
 440  * pgdir).  This occurs on almost every context switch. */
 441 void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable)
 442 {
 443         int newpgdir, repin = 0;
 444
 445         /* Look to see if we have this one already. */
 446         newpgdir = find_pgdir(cpu->lg, pgtable);
 447         /* If not, we allocate or mug an existing one: if it's a fresh one,
 448          * repin gets set to 1. */
 449         if (newpgdir == ARRAY_SIZE(cpu->lg->pgdirs))
 450                 newpgdir = new_pgdir(cpu, pgtable, &repin);
 451         /* Change the current pgd index to the new one. */
 452         cpu->cpu_pgd = newpgdir;
 453         /* If it was completely blank, we map in the Guest kernel stack */
 454         if (repin)
 455                 pin_stack_pages(cpu);
 456 }
 457
 458 /*H:470 Finally, a routine which throws away everything: all PGD entries in all
 459  * the shadow page tables, including the Guest's kernel mappings.  This is used
 460  * when we destroy the Guest. */
 461 static void release_all_pagetables(struct lguest *lg)
 462 {
 463         unsigned int i, j;
 464
 465         /* Every shadow pagetable this Guest has */
 466         for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
 467                 if (lg->pgdirs[i].pgdir)
 468                         /* Every PGD entry except the Switcher at the top */
 469                         for (j = 0; j < SWITCHER_PGD_INDEX; j++)
 470                                 release_pgd(lg, lg->pgdirs[i].pgdir + j);
 471 }
 472
 473 /* We also throw away everything when a Guest tells us it's changed a kernel
 474  * mapping.  Since kernel mappings are in every page table, it's easiest to
 475  * throw them all away.  This traps the Guest in amber for a while as
 476  * everything faults back in, but it's rare. */
 477 void guest_pagetable_clear_all(struct lg_cpu *cpu)
 478 {
 479         release_all_pagetables(cpu->lg);
 480         /* We need the Guest kernel stack mapped again. */
 481         pin_stack_pages(cpu);
 482 }
 483 /*:*/
 484 /*M:009 Since we throw away all mappings when a kernel mapping changes, our
 485  * performance sucks for guests using highmem.  In fact, a guest with
 486  * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is
 487  * usually slower than a Guest with less memory.
 488  *
 489  * This, of course, cannot be fixed.  It would take some kind of... well, I
 490  * don't know, but the term "puissant code-fu" comes to mind. :*/
 491
 492 /*H:420 This is the routine which actually sets the page table entry for then
 493  * "idx"'th shadow page table.
 494  *
 495  * Normally, we can just throw out the old entry and replace it with 0: if they
 496  * use it demand_page() will put the new entry in.  We need to do this anyway:
 497  * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page
 498  * is read from, and _PAGE_DIRTY when it's written to.
 499  *
 500  * But Avi Kivity pointed out that most Operating Systems (Linux included) set
 501  * these bits on PTEs immediately anyway.  This is done to save the CPU from
 502  * having to update them, but it helps us the same way: if they set
 503  * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if
 504  * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately.
 505  */
 506 static void do_set_pte(struct lg_cpu *cpu, int idx,
 507                        unsigned long vaddr, pte_t gpte)
 508 {
 509         /* Look up the matching shadow page directory entry. */
 510         pgd_t *spgd = spgd_addr(cpu, idx, vaddr);
 511
 512         /* If the top level isn't present, there's no entry to update. */
 513         if (pgd_flags(*spgd) & _PAGE_PRESENT) {
 514                 /* Otherwise, we start by releasing the existing entry. */
 515                 pte_t *spte = spte_addr(*spgd, vaddr);
 516                 release_pte(*spte);
 517
 518                 /* If they're setting this entry as dirty or accessed, we might
 519                  * as well put that entry they've given us in now.  This shaves
 520                  * 10% off a copy-on-write micro-benchmark. */
 521                 if (pte_flags(gpte) & (_PAGE_DIRTY | _PAGE_ACCESSED)) {
 522                         check_gpte(cpu, gpte);
 523                         *spte = gpte_to_spte(cpu, gpte,
 524                                              pte_flags(gpte) & _PAGE_DIRTY);
 525                 } else
 526                         /* Otherwise kill it and we can demand_page() it in
 527                          * later. */
 528                         *spte = __pte(0);
 529         }
 530 }
 531
 532 /*H:410 Updating a PTE entry is a little trickier.
 533  *
 534  * We keep track of several different page tables (the Guest uses one for each
 535  * process, so it makes sense to cache at least a few).  Each of these have
 536  * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for
 537  * all processes.  So when the page table above that address changes, we update
 538  * all the page tables, not just the current one.  This is rare.
 539  *
 540  * The benefit is that when we have to track a new page table, we can keep all
 541  * the kernel mappings.  This speeds up context switch immensely. */
 542 void guest_set_pte(struct lg_cpu *cpu,
 543                    unsigned long gpgdir, unsigned long vaddr, pte_t gpte)
 544 {
 545         /* Kernel mappings must be changed on all top levels.  Slow, but doesn't
 546          * happen often. */
 547         if (vaddr >= cpu->lg->kernel_address) {
 548                 unsigned int i;
 549                 for (i = 0; i < ARRAY_SIZE(cpu->lg->pgdirs); i++)
 550                         if (cpu->lg->pgdirs[i].pgdir)
 551                                 do_set_pte(cpu, i, vaddr, gpte);
 552         } else {
 553                 /* Is this page table one we have a shadow for? */
 554                 int pgdir = find_pgdir(cpu->lg, gpgdir);
 555                 if (pgdir != ARRAY_SIZE(cpu->lg->pgdirs))
 556                         /* If so, do the update. */
 557                         do_set_pte(cpu, pgdir, vaddr, gpte);
 558         }
 559 }
 560
 561 /*H:400
 562  * (iii) Setting up a page table entry when the Guest tells us one has changed.
 563  *
 564  * Just like we did in interrupts_and_traps.c, it makes sense for us to deal
 565  * with the other side of page tables while we're here: what happens when the
 566  * Guest asks for a page table to be updated?
 567  *
 568  * We already saw that demand_page() will fill in the shadow page tables when
 569  * needed, so we can simply remove shadow page table entries whenever the Guest
 570  * tells us they've changed.  When the Guest tries to use the new entry it will
 571  * fault and demand_page() will fix it up.
 572  *
 573  * So with that in mind here's our code to to update a (top-level) PGD entry:
 574  */
 575 void guest_set_pmd(struct lguest *lg, unsigned long gpgdir, u32 idx)
 576 {
 577         int pgdir;
 578
 579         /* The kernel seems to try to initialize this early on: we ignore its
 580          * attempts to map over the Switcher. */
 581         if (idx >= SWITCHER_PGD_INDEX)
 582                 return;
 583
 584         /* If they're talking about a page table we have a shadow for... */
 585         pgdir = find_pgdir(lg, gpgdir);
 586         if (pgdir < ARRAY_SIZE(lg->pgdirs))
 587                 /* ... throw it away. */
 588                 release_pgd(lg, lg->pgdirs[pgdir].pgdir + idx);
 589 }
 590
 591 /*H:500 (vii) Setting up the page tables initially.
 592  *
 593  * When a Guest is first created, the Launcher tells us where the toplevel of
 594  * its first page table is.  We set some things up here: */
 595 int init_guest_pagetable(struct lguest *lg, unsigned long pgtable)
 596 {
 597         /* We start on the first shadow page table, and give it a blank PGD
 598          * page. */
 599         lg->pgdirs[0].gpgdir = pgtable;
 600         lg->pgdirs[0].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL);
 601         if (!lg->pgdirs[0].pgdir)
 602                 return -ENOMEM;
 603         lg->cpus[0].cpu_pgd = 0;
 604         return 0;
 605 }
 606
 607 /* When the Guest calls LHCALL_LGUEST_INIT we do more setup. */
 608 void page_table_guest_data_init(struct lg_cpu *cpu)
 609 {
 610         /* We get the kernel address: above this is all kernel memory. */
 611         if (get_user(cpu->lg->kernel_address,
 612                      &cpu->lg->lguest_data->kernel_address)
 613             /* We tell the Guest that it can't use the top 4MB of virtual
 614              * addresses used by the Switcher. */
 615             || put_user(4U*1024*1024, &cpu->lg->lguest_data->reserve_mem)
 616             || put_user(cpu->lg->pgdirs[0].gpgdir, &cpu->lg->lguest_data->pgdir))
 617                 kill_guest(cpu, "bad guest page %p", cpu->lg->lguest_data);
 618
 619         /* In flush_user_mappings() we loop from 0 to
 620          * "pgd_index(lg->kernel_address)".  This assumes it won't hit the
 621          * Switcher mappings, so check that now. */
 622         if (pgd_index(cpu->lg->kernel_address) >= SWITCHER_PGD_INDEX)
 623                 kill_guest(cpu, "bad kernel address %#lx",
 624                                  cpu->lg->kernel_address);
 625 }
 626
 627 /* When a Guest dies, our cleanup is fairly simple. */
 628 void free_guest_pagetable(struct lguest *lg)
 629 {
 630         unsigned int i;
 631
 632         /* Throw away all page table pages. */
 633         release_all_pagetables(lg);
 634         /* Now free the top levels: free_page() can handle 0 just fine. */
 635         for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
 636                 free_page((long)lg->pgdirs[i].pgdir);
 637 }
 638
 639 /*H:480 (vi) Mapping the Switcher when the Guest is about to run.
 640  *
 641  * The Switcher and the two pages for this CPU need to be visible in the
 642  * Guest (and not the pages for other CPUs).  We have the appropriate PTE pages
 643  * for each CPU already set up, we just need to hook them in now we know which
 644  * Guest is about to run on this CPU. */
 645 void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages)
 646 {
 647         pte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages);
 648         pgd_t switcher_pgd;
 649         pte_t regs_pte;
 650         unsigned long pfn;
 651
 652         /* Make the last PGD entry for this Guest point to the Switcher's PTE
 653          * page for this CPU (with appropriate flags). */
 654         switcher_pgd = __pgd(__pa(switcher_pte_page) | __PAGE_KERNEL);
 655
 656         cpu->lg->pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd;
 657
 658         /* We also change the Switcher PTE page.  When we're running the Guest,
 659          * we want the Guest's "regs" page to appear where the first Switcher
 660          * page for this CPU is.  This is an optimization: when the Switcher
 661          * saves the Guest registers, it saves them into the first page of this
 662          * CPU's "struct lguest_pages": if we make sure the Guest's register
 663          * page is already mapped there, we don't have to copy them out
 664          * again. */
 665         pfn = __pa(cpu->regs_page) >> PAGE_SHIFT;
 666         regs_pte = pfn_pte(pfn, __pgprot(__PAGE_KERNEL));
 667         switcher_pte_page[(unsigned long)pages/PAGE_SIZE%PTRS_PER_PTE] = regs_pte;
 668 }
 669 /*:*/
 670
 671 static void free_switcher_pte_pages(void)
 672 {
 673         unsigned int i;
 674
 675         for_each_possible_cpu(i)
 676                 free_page((long)switcher_pte_page(i));
 677 }
 678
 679 /*H:520 Setting up the Switcher PTE page for given CPU is fairly easy, given
 680  * the CPU number and the "struct page"s for the Switcher code itself.
 681  *
 682  * Currently the Switcher is less than a page long, so "pages" is always 1. */
 683 static __init void populate_switcher_pte_page(unsigned int cpu,
 684                                               struct page *switcher_page[],
 685                                               unsigned int pages)
 686 {
 687         unsigned int i;
 688         pte_t *pte = switcher_pte_page(cpu);
 689
 690         /* The first entries are easy: they map the Switcher code. */
 691         for (i = 0; i < pages; i++) {
 692                 pte[i] = mk_pte(switcher_page[i],
 693                                 __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED));
 694         }
 695
 696         /* The only other thing we map is this CPU's pair of pages. */
 697         i = pages + cpu*2;
 698
 699         /* First page (Guest registers) is writable from the Guest */
 700         pte[i] = pfn_pte(page_to_pfn(switcher_page[i]),
 701                          __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW));
 702
 703         /* The second page contains the "struct lguest_ro_state", and is
 704          * read-only. */
 705         pte[i+1] = pfn_pte(page_to_pfn(switcher_page[i+1]),
 706                            __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED));
 707 }
 708
 709 /* We've made it through the page table code.  Perhaps our tired brains are
 710  * still processing the details, or perhaps we're simply glad it's over.
 711  *
 712  * If nothing else, note that all this complexity in juggling shadow page tables
 713  * in sync with the Guest's page tables is for one reason: for most Guests this
 714  * page table dance determines how bad performance will be.  This is why Xen
 715  * uses exotic direct Guest pagetable manipulation, and why both Intel and AMD
 716  * have implemented shadow page table support directly into hardware.
 717  *
 718  * There is just one file remaining in the Host. */
 719
 720 /*H:510 At boot or module load time, init_pagetables() allocates and populates
 721  * the Switcher PTE page for each CPU. */
 722 __init int init_pagetables(struct page **switcher_page, unsigned int pages)
 723 {
 724         unsigned int i;
 725
 726         for_each_possible_cpu(i) {
 727                 switcher_pte_page(i) = (pte_t *)get_zeroed_page(GFP_KERNEL);
 728                 if (!switcher_pte_page(i)) {
 729                         free_switcher_pte_pages();
 730                         return -ENOMEM;
 731                 }
 732                 populate_switcher_pte_page(i, switcher_page, pages);
 733         }
 734         return 0;
 735 }
 736 /*:*/
 737
 738 /* Cleaning up simply involves freeing the PTE page for each CPU. */
 739 void free_pagetables(void)
 740 {
 741         free_switcher_pte_pages();
 742 }