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