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 <asm/bootparam.h>
  18 #include "lg.h"
  19
  20 /*M:008 We hold reference to pages, which prevents them from being swapped.
  21  * It'd be nice to have a callback in the "struct mm_struct" when Linux wants
  22  * to swap out.  If we had this, and a shrinker callback to trim PTE pages, we
  23  * could probably consider launching Guests as non-root. :*/
  24
  25 /*H:300
  26  * The Page Table Code
  27  *
  28  * We use two-level page tables for the Guest.  If you're not entirely
  29  * comfortable with virtual addresses, physical addresses and page tables then
  30  * I recommend you review arch/x86/lguest/boot.c's "Page Table Handling" (with
  31  * diagrams!).
  32  *
  33  * The Guest keeps page tables, but we maintain the actual ones here: these are
  34  * called "shadow" page tables.  Which is a very Guest-centric name: these are
  35  * the real page tables the CPU uses, although we keep them up to date to
  36  * reflect the Guest's.  (See what I mean about weird naming?  Since when do
  37  * shadows reflect anything?)
  38  *
  39  * Anyway, this is the most complicated part of the Host code.  There are seven
  40  * parts to this:
  41  *  (i) Looking up a page table entry when the Guest faults,
  42  *  (ii) Making sure the Guest stack is mapped,
  43  *  (iii) Setting up a page table entry when the Guest tells us one has changed,
  44  *  (iv) Switching page tables,
  45  *  (v) Flushing (throwing away) page tables,
  46  *  (vi) Mapping the Switcher when the Guest is about to run,
  47  *  (vii) Setting up the page tables initially.
  48  :*/
  49
  50
  51 /* 1024 entries in a page table page maps 1024 pages: 4MB.  The Switcher is
  52  * conveniently placed at the top 4MB, so it uses a separate, complete PTE
  53  * page.  */
  54 #define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1)
  55
  56 /* We actually need a separate PTE page for each CPU.  Remember that after the
  57  * Switcher code itself comes two pages for each CPU, and we don't want this
  58  * CPU's guest to see the pages of any other CPU. */
  59 static DEFINE_PER_CPU(pte_t *, switcher_pte_pages);
  60 #define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu)
  61
  62 /*H:320 The page table code is curly enough to need helper functions to keep it
  63  * clear and clean.
  64  *
  65  * There are two functions which return pointers to the shadow (aka "real")
  66  * page tables.
  67  *
  68  * spgd_addr() takes the virtual address and returns a pointer to the top-level
  69  * page directory entry (PGD) for that address.  Since we keep track of several
  70  * page tables, the "i" argument tells us which one we're interested in (it's
  71  * usually the current one). */
  72 static pgd_t *spgd_addr(struct lg_cpu *cpu, u32 i, unsigned long vaddr)
  73 {
  74         unsigned int index = pgd_index(vaddr);
  75
  76         /* We kill any Guest trying to touch the Switcher addresses. */
  77         if (index >= SWITCHER_PGD_INDEX) {
  78                 kill_guest(cpu, "attempt to access switcher pages");
  79                 index = 0;
  80         }
  81         /* Return a pointer index'th pgd entry for the i'th page table. */
  82         return &cpu->lg->pgdirs[i].pgdir[index];
  83 }
  84
  85 /* This routine then takes the page directory entry returned above, which
  86  * contains the address of the page table entry (PTE) page.  It then returns a
  87  * pointer to the PTE entry for the given address. */
  88 static pte_t *spte_addr(pgd_t spgd, unsigned long vaddr)
  89 {
  90         pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT);
  91         /* You should never call this if the PGD entry wasn't valid */
  92         BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT));
  93         return &page[(vaddr >> PAGE_SHIFT) % PTRS_PER_PTE];
  94 }
  95
  96 /* These two functions just like the above two, except they access the Guest
  97  * page tables.  Hence they return a Guest address. */
  98 static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr)
  99 {
 100         unsigned int index = vaddr >> (PGDIR_SHIFT);
 101         return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index * sizeof(pgd_t);
 102 }
 103
 104 static unsigned long gpte_addr(pgd_t gpgd, unsigned long vaddr)
 105 {
 106         unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT;
 107         BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT));
 108         return gpage + ((vaddr>>PAGE_SHIFT) % PTRS_PER_PTE) * sizeof(pte_t);
 109 }
 110 /*:*/
 111
 112 /*M:014 get_pfn is slow: we could probably try to grab batches of pages here as
 113  * an optimization (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
 127         /* gup me one page at this address please! */
 128         if (get_user_pages_fast(virtpfn << PAGE_SHIFT, 1, write, &page) == 1)
 129                 return page_to_pfn(page);
 130
 131         /* This value indicates failure. */
 132         return -1UL;
 133 }
 134
 135 /*H:340 Converting a Guest page table entry to a shadow (ie. real) page table
 136  * entry can be a little tricky.  The flags are (almost) the same, but the
 137  * Guest PTE contains a virtual page number: the CPU needs the real page
 138  * number. */
 139 static pte_t gpte_to_spte(struct lg_cpu *cpu, pte_t gpte, int write)
 140 {
 141         unsigned long pfn, base, flags;
 142
 143         /* The Guest sets the global flag, because it thinks that it is using
 144          * PGE.  We only told it to use PGE so it would tell us whether it was
 145          * flushing a kernel mapping or a userspace mapping.  We don't actually
 146          * use the global bit, so throw it away. */
 147         flags = (pte_flags(gpte) & ~_PAGE_GLOBAL);
 148
 149         /* The Guest's pages are offset inside the Launcher. */
 150         base = (unsigned long)cpu->lg->mem_base / PAGE_SIZE;
 151
 152         /* We need a temporary "unsigned long" variable to hold the answer from
 153          * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't
 154          * fit in spte.pfn.  get_pfn() finds the real physical number of the
 155          * page, given the virtual number. */
 156         pfn = get_pfn(base + pte_pfn(gpte), write);
 157         if (pfn == -1UL) {
 158                 kill_guest(cpu, "failed to get page %lu", pte_pfn(gpte));
 159                 /* When we destroy the Guest, we'll go through the shadow page
 160                  * tables and release_pte() them.  Make sure we don't think
 161                  * this one is valid! */
 162                 flags = 0;
 163         }
 164         /* Now we assemble our shadow PTE from the page number and flags. */
 165         return pfn_pte(pfn, __pgprot(flags));
 166 }
 167
 168 /*H:460 And to complete the chain, release_pte() looks like this: */
 169 static void release_pte(pte_t pte)
 170 {
 171         /* Remember that get_user_pages_fast() took a reference to the page, in
 172          * get_pfn()?  We have to put it back now. */
 173         if (pte_flags(pte) & _PAGE_PRESENT)
 174                 put_page(pfn_to_page(pte_pfn(pte)));
 175 }
 176 /*:*/
 177
 178 static void check_gpte(struct lg_cpu *cpu, pte_t gpte)
 179 {
 180         if ((pte_flags(gpte) & _PAGE_PSE) ||
 181             pte_pfn(gpte) >= cpu->lg->pfn_limit)
 182                 kill_guest(cpu, "bad page table entry");
 183 }
 184
 185 static void check_gpgd(struct lg_cpu *cpu, pgd_t gpgd)
 186 {
 187         if ((pgd_flags(gpgd) & ~_PAGE_TABLE) ||
 188            (pgd_pfn(gpgd) >= cpu->lg->pfn_limit))
 189                 kill_guest(cpu, "bad page directory entry");
 190 }
 191
 192 /*H:330
 193  * (i) Looking up a page table entry when the Guest faults.
 194  *
 195  * We saw this call in run_guest(): when we see a page fault in the Guest, we
 196  * come here.  That's because we only set up the shadow page tables lazily as
 197  * they're needed, so we get page faults all the time and quietly fix them up
 198  * and return to the Guest without it knowing.
 199  *
 200  * If we fixed up the fault (ie. we mapped the address), this routine returns
 201  * true.  Otherwise, it was a real fault and we need to tell the Guest. */
 202 bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode)
 203 {
 204         pgd_t gpgd;
 205         pgd_t *spgd;
 206         unsigned long gpte_ptr;
 207         pte_t gpte;
 208         pte_t *spte;
 209
 210         /* First step: get the top-level Guest page table entry. */
 211         gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t);
 212         /* Toplevel not present?  We can't map it in. */
 213         if (!(pgd_flags(gpgd) & _PAGE_PRESENT))
 214                 return false;
 215
 216         /* Now look at the matching shadow entry. */
 217         spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr);
 218         if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) {
 219                 /* No shadow entry: allocate a new shadow PTE page. */
 220                 unsigned long ptepage = get_zeroed_page(GFP_KERNEL);
 221                 /* This is not really the Guest's fault, but killing it is
 222                  * simple for this corner case. */
 223                 if (!ptepage) {
 224                         kill_guest(cpu, "out of memory allocating pte page");
 225                         return false;
 226                 }
 227                 /* We check that the Guest pgd is OK. */
 228                 check_gpgd(cpu, gpgd);
 229                 /* And we copy the flags to the shadow PGD entry.  The page
 230                  * number in the shadow PGD is the page we just allocated. */
 231                 *spgd = __pgd(__pa(ptepage) | pgd_flags(gpgd));
 232         }
 233
 234         /* OK, now we look at the lower level in the Guest page table: keep its
 235          * address, because we might update it later. */
 236         gpte_ptr = gpte_addr(gpgd, vaddr);
 237         gpte = lgread(cpu, gpte_ptr, pte_t);
 238
 239         /* If this page isn't in the Guest page tables, we can't page it in. */
 240         if (!(pte_flags(gpte) & _PAGE_PRESENT))
 241                 return false;
 242
 243         /* Check they're not trying to write to a page the Guest wants
 244          * read-only (bit 2 of errcode == write). */
 245         if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW))
 246                 return false;
 247
 248         /* User access to a kernel-only page? (bit 3 == user access) */
 249         if ((errcode & 4) && !(pte_flags(gpte) & _PAGE_USER))
 250                 return false;
 251
 252         /* Check that the Guest PTE flags are OK, and the page number is below
 253          * the pfn_limit (ie. not mapping the Launcher binary). */
 254         check_gpte(cpu, gpte);
 255
 256         /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */
 257         gpte = pte_mkyoung(gpte);
 258         if (errcode & 2)
 259                 gpte = pte_mkdirty(gpte);
 260
 261         /* Get the pointer to the shadow PTE entry we're going to set. */
 262         spte = spte_addr(*spgd, vaddr);
 263         /* If there was a valid shadow PTE entry here before, we release it.
 264          * This can happen with a write to a previously read-only entry. */
 265         release_pte(*spte);
 266
 267         /* If this is a write, we insist that the Guest page is writable (the
 268          * final arg to gpte_to_spte()). */
 269         if (pte_dirty(gpte))
 270                 *spte = gpte_to_spte(cpu, gpte, 1);
 271         else
 272                 /* If this is a read, don't set the "writable" bit in the page
 273                  * table entry, even if the Guest says it's writable.  That way
 274                  * we will come back here when a write does actually occur, so
 275                  * we can update the Guest's _PAGE_DIRTY flag. */
 276                 *spte = gpte_to_spte(cpu, pte_wrprotect(gpte), 0);
 277
 278         /* Finally, we write the Guest PTE entry back: we've set the
 279          * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. */
 280         lgwrite(cpu, gpte_ptr, pte_t, gpte);
 281
 282         /* The fault is fixed, the page table is populated, the mapping
 283          * manipulated, the result returned and the code complete.  A small
 284          * delay and a trace of alliteration are the only indications the Guest
 285          * has that a page fault occurred at all. */
 286         return true;
 287 }
 288
 289 /*H:360
 290  * (ii) Making sure the Guest stack is mapped.
 291  *
 292  * Remember that direct traps into the Guest need a mapped Guest kernel stack.
 293  * pin_stack_pages() calls us here: we could simply call demand_page(), but as
 294  * we've seen that logic is quite long, and usually the stack pages are already
 295  * mapped, so it's overkill.
 296  *
 297  * This is a quick version which answers the question: is this virtual address
 298  * mapped by the shadow page tables, and is it writable? */
 299 static bool page_writable(struct lg_cpu *cpu, unsigned long vaddr)
 300 {
 301         pgd_t *spgd;
 302         unsigned long flags;
 303
 304         /* Look at the current top level entry: is it present? */
 305         spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr);
 306         if (!(pgd_flags(*spgd) & _PAGE_PRESENT))
 307                 return false;
 308
 309         /* Check the flags on the pte entry itself: it must be present and
 310          * writable. */
 311         flags = pte_flags(*(spte_addr(*spgd, vaddr)));
 312
 313         return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW);
 314 }
 315
 316 /* So, when pin_stack_pages() asks us to pin a page, we check if it's already
 317  * in the page tables, and if not, we call demand_page() with error code 2
 318  * (meaning "write"). */
 319 void pin_page(struct lg_cpu *cpu, unsigned long vaddr)
 320 {
 321         if (!page_writable(cpu, vaddr) && !demand_page(cpu, vaddr, 2))
 322                 kill_guest(cpu, "bad stack page %#lx", vaddr);
 323 }
 324
 325 /*H:450 If we chase down the release_pgd() code, it looks like this: */
 326 static void release_pgd(struct lguest *lg, pgd_t *spgd)
 327 {
 328         /* If the entry's not present, there's nothing to release. */
 329         if (pgd_flags(*spgd) & _PAGE_PRESENT) {
 330                 unsigned int i;
 331                 /* Converting the pfn to find the actual PTE page is easy: turn
 332                  * the page number into a physical address, then convert to a
 333                  * virtual address (easy for kernel pages like this one). */
 334                 pte_t *ptepage = __va(pgd_pfn(*spgd) << PAGE_SHIFT);
 335                 /* For each entry in the page, we might need to release it. */
 336                 for (i = 0; i < PTRS_PER_PTE; i++)
 337                         release_pte(ptepage[i]);
 338                 /* Now we can free the page of PTEs */
 339                 free_page((long)ptepage);
 340                 /* And zero out the PGD entry so we never release it twice. */
 341                 *spgd = __pgd(0);
 342         }
 343 }
 344
 345 /*H:445 We saw flush_user_mappings() twice: once from the flush_user_mappings()
 346  * hypercall and once in new_pgdir() when we re-used a top-level pgdir page.
 347  * It simply releases every PTE page from 0 up to the Guest's kernel address. */
 348 static void flush_user_mappings(struct lguest *lg, int idx)
 349 {
 350         unsigned int i;
 351         /* Release every pgd entry up to the kernel's address. */
 352         for (i = 0; i < pgd_index(lg->kernel_address); i++)
 353                 release_pgd(lg, lg->pgdirs[idx].pgdir + i);
 354 }
 355
 356 /*H:440 (v) Flushing (throwing away) page tables,
 357  *
 358  * The Guest has a hypercall to throw away the page tables: it's used when a
 359  * large number of mappings have been changed. */
 360 void guest_pagetable_flush_user(struct lg_cpu *cpu)
 361 {
 362         /* Drop the userspace part of the current page table. */
 363         flush_user_mappings(cpu->lg, cpu->cpu_pgd);
 364 }
 365 /*:*/
 366
 367 /* We walk down the guest page tables to get a guest-physical address */
 368 unsigned long guest_pa(struct lg_cpu *cpu, unsigned long vaddr)
 369 {
 370         pgd_t gpgd;
 371         pte_t gpte;
 372
 373         /* First step: get the top-level Guest page table entry. */
 374         gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t);
 375         /* Toplevel not present?  We can't map it in. */
 376         if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) {
 377                 kill_guest(cpu, "Bad address %#lx", vaddr);
 378                 return -1UL;
 379         }
 380
 381         gpte = lgread(cpu, gpte_addr(gpgd, vaddr), pte_t);
 382         if (!(pte_flags(gpte) & _PAGE_PRESENT))
 383                 kill_guest(cpu, "Bad address %#lx", vaddr);
 384
 385         return pte_pfn(gpte) * PAGE_SIZE | (vaddr & ~PAGE_MASK);
 386 }
 387
 388 /* We keep several page tables.  This is a simple routine to find the page
 389  * table (if any) corresponding to this top-level address the Guest has given
 390  * us. */
 391 static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable)
 392 {
 393         unsigned int i;
 394         for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
 395                 if (lg->pgdirs[i].pgdir && lg->pgdirs[i].gpgdir == pgtable)
 396                         break;
 397         return i;
 398 }
 399
 400 /*H:435 And this is us, creating the new page directory.  If we really do
 401  * allocate a new one (and so the kernel parts are not there), we set
 402  * blank_pgdir. */
 403 static unsigned int new_pgdir(struct lg_cpu *cpu,
 404                               unsigned long gpgdir,
 405                               int *blank_pgdir)
 406 {
 407         unsigned int next;
 408
 409         /* We pick one entry at random to throw out.  Choosing the Least
 410          * Recently Used might be better, but this is easy. */
 411         next = random32() % ARRAY_SIZE(cpu->lg->pgdirs);
 412         /* If it's never been allocated at all before, try now. */
 413         if (!cpu->lg->pgdirs[next].pgdir) {
 414                 cpu->lg->pgdirs[next].pgdir =
 415                                         (pgd_t *)get_zeroed_page(GFP_KERNEL);
 416                 /* If the allocation fails, just keep using the one we have */
 417                 if (!cpu->lg->pgdirs[next].pgdir)
 418                         next = cpu->cpu_pgd;
 419                 else
 420                         /* This is a blank page, so there are no kernel
 421                          * mappings: caller must map the stack! */
 422                         *blank_pgdir = 1;
 423         }
 424         /* Record which Guest toplevel this shadows. */
 425         cpu->lg->pgdirs[next].gpgdir = gpgdir;
 426         /* Release all the non-kernel mappings. */
 427         flush_user_mappings(cpu->lg, next);
 428
 429         return next;
 430 }
 431
 432 /*H:430 (iv) Switching page tables
 433  *
 434  * Now we've seen all the page table setting and manipulation, let's see what
 435  * what happens when the Guest changes page tables (ie. changes the top-level
 436  * pgdir).  This occurs on almost every context switch. */
 437 void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable)
 438 {
 439         int newpgdir, repin = 0;
 440
 441         /* Look to see if we have this one already. */
 442         newpgdir = find_pgdir(cpu->lg, pgtable);
 443         /* If not, we allocate or mug an existing one: if it's a fresh one,
 444          * repin gets set to 1. */
 445         if (newpgdir == ARRAY_SIZE(cpu->lg->pgdirs))
 446                 newpgdir = new_pgdir(cpu, pgtable, &repin);
 447         /* Change the current pgd index to the new one. */
 448         cpu->cpu_pgd = newpgdir;
 449         /* If it was completely blank, we map in the Guest kernel stack */
 450         if (repin)
 451                 pin_stack_pages(cpu);
 452 }
 453
 454 /*H:470 Finally, a routine which throws away everything: all PGD entries in all
 455  * the shadow page tables, including the Guest's kernel mappings.  This is used
 456  * when we destroy the Guest. */
 457 static void release_all_pagetables(struct lguest *lg)
 458 {
 459         unsigned int i, j;
 460
 461         /* Every shadow pagetable this Guest has */
 462         for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
 463                 if (lg->pgdirs[i].pgdir)
 464                         /* Every PGD entry except the Switcher at the top */
 465                         for (j = 0; j < SWITCHER_PGD_INDEX; j++)
 466                                 release_pgd(lg, lg->pgdirs[i].pgdir + j);
 467 }
 468
 469 /* We also throw away everything when a Guest tells us it's changed a kernel
 470  * mapping.  Since kernel mappings are in every page table, it's easiest to
 471  * throw them all away.  This traps the Guest in amber for a while as
 472  * everything faults back in, but it's rare. */
 473 void guest_pagetable_clear_all(struct lg_cpu *cpu)
 474 {
 475         release_all_pagetables(cpu->lg);
 476         /* We need the Guest kernel stack mapped again. */
 477         pin_stack_pages(cpu);
 478 }
 479 /*:*/
 480 /*M:009 Since we throw away all mappings when a kernel mapping changes, our
 481  * performance sucks for guests using highmem.  In fact, a guest with
 482  * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is
 483  * usually slower than a Guest with less memory.
 484  *
 485  * This, of course, cannot be fixed.  It would take some kind of... well, I
 486  * don't know, but the term "puissant code-fu" comes to mind. :*/
 487
 488 /*H:420 This is the routine which actually sets the page table entry for then
 489  * "idx"'th shadow page table.
 490  *
 491  * Normally, we can just throw out the old entry and replace it with 0: if they
 492  * use it demand_page() will put the new entry in.  We need to do this anyway:
 493  * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page
 494  * is read from, and _PAGE_DIRTY when it's written to.
 495  *
 496  * But Avi Kivity pointed out that most Operating Systems (Linux included) set
 497  * these bits on PTEs immediately anyway.  This is done to save the CPU from
 498  * having to update them, but it helps us the same way: if they set
 499  * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if
 500  * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately.
 501  */
 502 static void do_set_pte(struct lg_cpu *cpu, int idx,
 503                        unsigned long vaddr, pte_t gpte)
 504 {
 505         /* Look up the matching shadow page directory entry. */
 506         pgd_t *spgd = spgd_addr(cpu, idx, vaddr);
 507
 508         /* If the top level isn't present, there's no entry to update. */
 509         if (pgd_flags(*spgd) & _PAGE_PRESENT) {
 510                 /* Otherwise, we start by releasing the existing entry. */
 511                 pte_t *spte = spte_addr(*spgd, vaddr);
 512                 release_pte(*spte);
 513
 514                 /* If they're setting this entry as dirty or accessed, we might
 515                  * as well put that entry they've given us in now.  This shaves
 516                  * 10% off a copy-on-write micro-benchmark. */
 517                 if (pte_flags(gpte) & (_PAGE_DIRTY | _PAGE_ACCESSED)) {
 518                         check_gpte(cpu, gpte);
 519                         *spte = gpte_to_spte(cpu, gpte,
 520                                              pte_flags(gpte) & _PAGE_DIRTY);
 521                 } else
 522                         /* Otherwise kill it and we can demand_page() it in
 523                          * later. */
 524                         *spte = __pte(0);
 525         }
 526 }
 527
 528 /*H:410 Updating a PTE entry is a little trickier.
 529  *
 530  * We keep track of several different page tables (the Guest uses one for each
 531  * process, so it makes sense to cache at least a few).  Each of these have
 532  * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for
 533  * all processes.  So when the page table above that address changes, we update
 534  * all the page tables, not just the current one.  This is rare.
 535  *
 536  * The benefit is that when we have to track a new page table, we can keep all
 537  * the kernel mappings.  This speeds up context switch immensely. */
 538 void guest_set_pte(struct lg_cpu *cpu,
 539                    unsigned long gpgdir, unsigned long vaddr, pte_t gpte)
 540 {
 541         /* Kernel mappings must be changed on all top levels.  Slow, but doesn't
 542          * happen often. */
 543         if (vaddr >= cpu->lg->kernel_address) {
 544                 unsigned int i;
 545                 for (i = 0; i < ARRAY_SIZE(cpu->lg->pgdirs); i++)
 546                         if (cpu->lg->pgdirs[i].pgdir)
 547                                 do_set_pte(cpu, i, vaddr, gpte);
 548         } else {
 549                 /* Is this page table one we have a shadow for? */
 550                 int pgdir = find_pgdir(cpu->lg, gpgdir);
 551                 if (pgdir != ARRAY_SIZE(cpu->lg->pgdirs))
 552                         /* If so, do the update. */
 553                         do_set_pte(cpu, pgdir, vaddr, gpte);
 554         }
 555 }
 556
 557 /*H:400
 558  * (iii) Setting up a page table entry when the Guest tells us one has changed.
 559  *
 560  * Just like we did in interrupts_and_traps.c, it makes sense for us to deal
 561  * with the other side of page tables while we're here: what happens when the
 562  * Guest asks for a page table to be updated?
 563  *
 564  * We already saw that demand_page() will fill in the shadow page tables when
 565  * needed, so we can simply remove shadow page table entries whenever the Guest
 566  * tells us they've changed.  When the Guest tries to use the new entry it will
 567  * fault and demand_page() will fix it up.
 568  *
 569  * So with that in mind here's our code to to update a (top-level) PGD entry:
 570  */
 571 void guest_set_pmd(struct lguest *lg, unsigned long gpgdir, u32 idx)
 572 {
 573         int pgdir;
 574
 575         /* The kernel seems to try to initialize this early on: we ignore its
 576          * attempts to map over the Switcher. */
 577         if (idx >= SWITCHER_PGD_INDEX)
 578                 return;
 579
 580         /* If they're talking about a page table we have a shadow for... */
 581         pgdir = find_pgdir(lg, gpgdir);
 582         if (pgdir < ARRAY_SIZE(lg->pgdirs))
 583                 /* ... throw it away. */
 584                 release_pgd(lg, lg->pgdirs[pgdir].pgdir + idx);
 585 }
 586
 587 /* Once we know how much memory we have we can construct simple identity
 588  * (which set virtual == physical) and linear mappings
 589  * which will get the Guest far enough into the boot to create its own.
 590  *
 591  * We lay them out of the way, just below the initrd (which is why we need to
 592  * know its size here). */
 593 static unsigned long setup_pagetables(struct lguest *lg,
 594                                       unsigned long mem,
 595                                       unsigned long initrd_size)
 596 {
 597         pgd_t __user *pgdir;
 598         pte_t __user *linear;
 599         unsigned int mapped_pages, i, linear_pages, phys_linear;
 600         unsigned long mem_base = (unsigned long)lg->mem_base;
 601
 602         /* We have mapped_pages frames to map, so we need
 603          * linear_pages page tables to map them. */
 604         mapped_pages = mem / PAGE_SIZE;
 605         linear_pages = (mapped_pages + PTRS_PER_PTE - 1) / PTRS_PER_PTE;
 606
 607         /* We put the toplevel page directory page at the top of memory. */
 608         pgdir = (pgd_t *)(mem + mem_base - initrd_size - PAGE_SIZE);
 609
 610         /* Now we use the next linear_pages pages as pte pages */
 611         linear = (void *)pgdir - linear_pages * PAGE_SIZE;
 612
 613         /* Linear mapping is easy: put every page's address into the
 614          * mapping in order. */
 615         for (i = 0; i < mapped_pages; i++) {
 616                 pte_t pte;
 617                 pte = pfn_pte(i, __pgprot(_PAGE_PRESENT|_PAGE_RW|_PAGE_USER));
 618                 if (copy_to_user(&linear[i], &pte, sizeof(pte)) != 0)
 619                         return -EFAULT;
 620         }
 621
 622         /* The top level points to the linear page table pages above.
 623          * We setup the identity and linear mappings here. */
 624         phys_linear = (unsigned long)linear - mem_base;
 625         for (i = 0; i < mapped_pages; i += PTRS_PER_PTE) {
 626                 pgd_t pgd;
 627                 pgd = __pgd((phys_linear + i * sizeof(pte_t)) |
 628                             (_PAGE_PRESENT | _PAGE_RW | _PAGE_USER));
 629
 630                 if (copy_to_user(&pgdir[i / PTRS_PER_PTE], &pgd, sizeof(pgd))
 631                     || copy_to_user(&pgdir[pgd_index(PAGE_OFFSET)
 632                                            + i / PTRS_PER_PTE],
 633                                     &pgd, sizeof(pgd)))
 634                         return -EFAULT;
 635         }
 636
 637         /* We return the top level (guest-physical) address: remember where
 638          * this is. */
 639         return (unsigned long)pgdir - mem_base;
 640 }
 641
 642 /*H:500 (vii) Setting up the page tables initially.
 643  *
 644  * When a Guest is first created, the Launcher tells us where the toplevel of
 645  * its first page table is.  We set some things up here: */
 646 int init_guest_pagetable(struct lguest *lg)
 647 {
 648         u64 mem;
 649         u32 initrd_size;
 650         struct boot_params __user *boot = (struct boot_params *)lg->mem_base;
 651
 652         /* Get the Guest memory size and the ramdisk size from the boot header
 653          * located at lg->mem_base (Guest address 0). */
 654         if (copy_from_user(&mem, &boot->e820_map[0].size, sizeof(mem))
 655             || get_user(initrd_size, &boot->hdr.ramdisk_size))
 656                 return -EFAULT;
 657
 658         /* We start on the first shadow page table, and give it a blank PGD
 659          * page. */
 660         lg->pgdirs[0].gpgdir = setup_pagetables(lg, mem, initrd_size);
 661         if (IS_ERR_VALUE(lg->pgdirs[0].gpgdir))
 662                 return lg->pgdirs[0].gpgdir;
 663         lg->pgdirs[0].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL);
 664         if (!lg->pgdirs[0].pgdir)
 665                 return -ENOMEM;
 666         lg->cpus[0].cpu_pgd = 0;
 667         return 0;
 668 }
 669
 670 /* When the Guest calls LHCALL_LGUEST_INIT we do more setup. */
 671 void page_table_guest_data_init(struct lg_cpu *cpu)
 672 {
 673         /* We get the kernel address: above this is all kernel memory. */
 674         if (get_user(cpu->lg->kernel_address,
 675                      &cpu->lg->lguest_data->kernel_address)
 676             /* We tell the Guest that it can't use the top 4MB of virtual
 677              * addresses used by the Switcher. */
 678             || put_user(4U*1024*1024, &cpu->lg->lguest_data->reserve_mem)
 679             || put_user(cpu->lg->pgdirs[0].gpgdir, &cpu->lg->lguest_data->pgdir))
 680                 kill_guest(cpu, "bad guest page %p", cpu->lg->lguest_data);
 681
 682         /* In flush_user_mappings() we loop from 0 to
 683          * "pgd_index(lg->kernel_address)".  This assumes it won't hit the
 684          * Switcher mappings, so check that now. */
 685         if (pgd_index(cpu->lg->kernel_address) >= SWITCHER_PGD_INDEX)
 686                 kill_guest(cpu, "bad kernel address %#lx",
 687                                  cpu->lg->kernel_address);
 688 }
 689
 690 /* When a Guest dies, our cleanup is fairly simple. */
 691 void free_guest_pagetable(struct lguest *lg)
 692 {
 693         unsigned int i;
 694
 695         /* Throw away all page table pages. */
 696         release_all_pagetables(lg);
 697         /* Now free the top levels: free_page() can handle 0 just fine. */
 698         for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++)
 699                 free_page((long)lg->pgdirs[i].pgdir);
 700 }
 701
 702 /*H:480 (vi) Mapping the Switcher when the Guest is about to run.
 703  *
 704  * The Switcher and the two pages for this CPU need to be visible in the
 705  * Guest (and not the pages for other CPUs).  We have the appropriate PTE pages
 706  * for each CPU already set up, we just need to hook them in now we know which
 707  * Guest is about to run on this CPU. */
 708 void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages)
 709 {
 710         pte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages);
 711         pgd_t switcher_pgd;
 712         pte_t regs_pte;
 713         unsigned long pfn;
 714
 715         /* Make the last PGD entry for this Guest point to the Switcher's PTE
 716          * page for this CPU (with appropriate flags). */
 717         switcher_pgd = __pgd(__pa(switcher_pte_page) | __PAGE_KERNEL);
 718
 719         cpu->lg->pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd;
 720
 721         /* We also change the Switcher PTE page.  When we're running the Guest,
 722          * we want the Guest's "regs" page to appear where the first Switcher
 723          * page for this CPU is.  This is an optimization: when the Switcher
 724          * saves the Guest registers, it saves them into the first page of this
 725          * CPU's "struct lguest_pages": if we make sure the Guest's register
 726          * page is already mapped there, we don't have to copy them out
 727          * again. */
 728         pfn = __pa(cpu->regs_page) >> PAGE_SHIFT;
 729         regs_pte = pfn_pte(pfn, __pgprot(__PAGE_KERNEL));
 730         switcher_pte_page[(unsigned long)pages/PAGE_SIZE%PTRS_PER_PTE] = regs_pte;
 731 }
 732 /*:*/
 733
 734 static void free_switcher_pte_pages(void)
 735 {
 736         unsigned int i;
 737
 738         for_each_possible_cpu(i)
 739                 free_page((long)switcher_pte_page(i));
 740 }
 741
 742 /*H:520 Setting up the Switcher PTE page for given CPU is fairly easy, given
 743  * the CPU number and the "struct page"s for the Switcher code itself.
 744  *
 745  * Currently the Switcher is less than a page long, so "pages" is always 1. */
 746 static __init void populate_switcher_pte_page(unsigned int cpu,
 747                                               struct page *switcher_page[],
 748                                               unsigned int pages)
 749 {
 750         unsigned int i;
 751         pte_t *pte = switcher_pte_page(cpu);
 752
 753         /* The first entries are easy: they map the Switcher code. */
 754         for (i = 0; i < pages; i++) {
 755                 pte[i] = mk_pte(switcher_page[i],
 756                                 __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED));
 757         }
 758
 759         /* The only other thing we map is this CPU's pair of pages. */
 760         i = pages + cpu*2;
 761
 762         /* First page (Guest registers) is writable from the Guest */
 763         pte[i] = pfn_pte(page_to_pfn(switcher_page[i]),
 764                          __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW));
 765
 766         /* The second page contains the "struct lguest_ro_state", and is
 767          * read-only. */
 768         pte[i+1] = pfn_pte(page_to_pfn(switcher_page[i+1]),
 769                            __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED));
 770 }
 771
 772 /* We've made it through the page table code.  Perhaps our tired brains are
 773  * still processing the details, or perhaps we're simply glad it's over.
 774  *
 775  * If nothing else, note that all this complexity in juggling shadow page tables
 776  * in sync with the Guest's page tables is for one reason: for most Guests this
 777  * page table dance determines how bad performance will be.  This is why Xen
 778  * uses exotic direct Guest pagetable manipulation, and why both Intel and AMD
 779  * have implemented shadow page table support directly into hardware.
 780  *
 781  * There is just one file remaining in the Host. */
 782
 783 /*H:510 At boot or module load time, init_pagetables() allocates and populates
 784  * the Switcher PTE page for each CPU. */
 785 __init int init_pagetables(struct page **switcher_page, unsigned int pages)
 786 {
 787         unsigned int i;
 788
 789         for_each_possible_cpu(i) {
 790                 switcher_pte_page(i) = (pte_t *)get_zeroed_page(GFP_KERNEL);
 791                 if (!switcher_pte_page(i)) {
 792                         free_switcher_pte_pages();
 793                         return -ENOMEM;
 794                 }
 795                 populate_switcher_pte_page(i, switcher_page, pages);
 796         }
 797         return 0;
 798 }
 799 /*:*/
 800
 801 /* Cleaning up simply involves freeing the PTE page for each CPU. */
 802 void free_pagetables(void)
 803 {
 804         free_switcher_pte_pages();
 805 }