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