2 * The pagetable code, on the other hand, still shows the scars of
3 * previous encounters. It's functional, and as neat as it can be in the
4 * circumstances, but be wary, for these things are subtle and break easily.
5 * The Guest provides a virtual to physical mapping, but we can neither trust
6 * it nor use it: we verify and convert it here then point the CPU to the
7 * converted Guest pages when running the Guest.
10 /* Copyright (C) Rusty Russell IBM Corporation 2006.
11 * GPL v2 and any later version */
13 #include <linux/types.h>
14 #include <linux/spinlock.h>
15 #include <linux/random.h>
16 #include <linux/percpu.h>
17 #include <asm/tlbflush.h>
18 #include <asm/uaccess.h>
19 #include <asm/bootparam.h>
23 * We hold reference to pages, which prevents them from being swapped.
24 * It'd be nice to have a callback in the "struct mm_struct" when Linux wants
25 * to swap out. If we had this, and a shrinker callback to trim PTE pages, we
26 * could probably consider launching Guests as non-root.
32 * We use two-level page tables for the Guest, or three-level with PAE. If
33 * you're not entirely comfortable with virtual addresses, physical addresses
34 * and page tables then I recommend you review arch/x86/lguest/boot.c's "Page
35 * Table Handling" (with diagrams!).
37 * The Guest keeps page tables, but we maintain the actual ones here: these are
38 * called "shadow" page tables. Which is a very Guest-centric name: these are
39 * the real page tables the CPU uses, although we keep them up to date to
40 * reflect the Guest's. (See what I mean about weird naming? Since when do
41 * shadows reflect anything?)
43 * Anyway, this is the most complicated part of the Host code. There are seven
45 * (i) Looking up a page table entry when the Guest faults,
46 * (ii) Making sure the Guest stack is mapped,
47 * (iii) Setting up a page table entry when the Guest tells us one has changed,
48 * (iv) Switching page tables,
49 * (v) Flushing (throwing away) page tables,
50 * (vi) Mapping the Switcher when the Guest is about to run,
51 * (vii) Setting up the page tables initially.
55 * The Switcher uses the complete top PTE page. That's 1024 PTE entries (4MB)
56 * or 512 PTE entries with PAE (2MB).
58 #define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1)
61 * For PAE we need the PMD index as well. We use the last 2MB, so we
62 * will need the last pmd entry of the last pmd page.
65 #define SWITCHER_PMD_INDEX (PTRS_PER_PMD - 1)
66 #define RESERVE_MEM 2U
67 #define CHECK_GPGD_MASK _PAGE_PRESENT
69 #define RESERVE_MEM 4U
70 #define CHECK_GPGD_MASK _PAGE_TABLE
74 * We actually need a separate PTE page for each CPU. Remember that after the
75 * Switcher code itself comes two pages for each CPU, and we don't want this
76 * CPU's guest to see the pages of any other CPU.
78 static DEFINE_PER_CPU(pte_t
*, switcher_pte_pages
);
79 #define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu)
82 * The page table code is curly enough to need helper functions to keep it
83 * clear and clean. The kernel itself provides many of them; one advantage
84 * of insisting that the Guest and Host use the same CONFIG_PAE setting.
86 * There are two functions which return pointers to the shadow (aka "real")
89 * spgd_addr() takes the virtual address and returns a pointer to the top-level
90 * page directory entry (PGD) for that address. Since we keep track of several
91 * page tables, the "i" argument tells us which one we're interested in (it's
92 * usually the current one).
94 static pgd_t
*spgd_addr(struct lg_cpu
*cpu
, u32 i
, unsigned long vaddr
)
96 unsigned int index
= pgd_index(vaddr
);
98 #ifndef CONFIG_X86_PAE
99 /* We kill any Guest trying to touch the Switcher addresses. */
100 if (index
>= SWITCHER_PGD_INDEX
) {
101 kill_guest(cpu
, "attempt to access switcher pages");
105 /* Return a pointer index'th pgd entry for the i'th page table. */
106 return &cpu
->lg
->pgdirs
[i
].pgdir
[index
];
109 #ifdef CONFIG_X86_PAE
111 * This routine then takes the PGD entry given above, which contains the
112 * address of the PMD page. It then returns a pointer to the PMD entry for the
115 static pmd_t
*spmd_addr(struct lg_cpu
*cpu
, pgd_t spgd
, unsigned long vaddr
)
117 unsigned int index
= pmd_index(vaddr
);
120 /* We kill any Guest trying to touch the Switcher addresses. */
121 if (pgd_index(vaddr
) == SWITCHER_PGD_INDEX
&&
122 index
>= SWITCHER_PMD_INDEX
) {
123 kill_guest(cpu
, "attempt to access switcher pages");
127 /* You should never call this if the PGD entry wasn't valid */
128 BUG_ON(!(pgd_flags(spgd
) & _PAGE_PRESENT
));
129 page
= __va(pgd_pfn(spgd
) << PAGE_SHIFT
);
136 * This routine then takes the page directory entry returned above, which
137 * contains the address of the page table entry (PTE) page. It then returns a
138 * pointer to the PTE entry for the given address.
140 static pte_t
*spte_addr(struct lg_cpu
*cpu
, pgd_t spgd
, unsigned long vaddr
)
142 #ifdef CONFIG_X86_PAE
143 pmd_t
*pmd
= spmd_addr(cpu
, spgd
, vaddr
);
144 pte_t
*page
= __va(pmd_pfn(*pmd
) << PAGE_SHIFT
);
146 /* You should never call this if the PMD entry wasn't valid */
147 BUG_ON(!(pmd_flags(*pmd
) & _PAGE_PRESENT
));
149 pte_t
*page
= __va(pgd_pfn(spgd
) << PAGE_SHIFT
);
150 /* You should never call this if the PGD entry wasn't valid */
151 BUG_ON(!(pgd_flags(spgd
) & _PAGE_PRESENT
));
154 return &page
[pte_index(vaddr
)];
158 * These functions are just like the above two, except they access the Guest
159 * page tables. Hence they return a Guest address.
161 static unsigned long gpgd_addr(struct lg_cpu
*cpu
, unsigned long vaddr
)
163 unsigned int index
= vaddr
>> (PGDIR_SHIFT
);
164 return cpu
->lg
->pgdirs
[cpu
->cpu_pgd
].gpgdir
+ index
* sizeof(pgd_t
);
167 #ifdef CONFIG_X86_PAE
168 /* Follow the PGD to the PMD. */
169 static unsigned long gpmd_addr(pgd_t gpgd
, unsigned long vaddr
)
171 unsigned long gpage
= pgd_pfn(gpgd
) << PAGE_SHIFT
;
172 BUG_ON(!(pgd_flags(gpgd
) & _PAGE_PRESENT
));
173 return gpage
+ pmd_index(vaddr
) * sizeof(pmd_t
);
176 /* Follow the PMD to the PTE. */
177 static unsigned long gpte_addr(struct lg_cpu
*cpu
,
178 pmd_t gpmd
, unsigned long vaddr
)
180 unsigned long gpage
= pmd_pfn(gpmd
) << PAGE_SHIFT
;
182 BUG_ON(!(pmd_flags(gpmd
) & _PAGE_PRESENT
));
183 return gpage
+ pte_index(vaddr
) * sizeof(pte_t
);
186 /* Follow the PGD to the PTE (no mid-level for !PAE). */
187 static unsigned long gpte_addr(struct lg_cpu
*cpu
,
188 pgd_t gpgd
, unsigned long vaddr
)
190 unsigned long gpage
= pgd_pfn(gpgd
) << PAGE_SHIFT
;
192 BUG_ON(!(pgd_flags(gpgd
) & _PAGE_PRESENT
));
193 return gpage
+ pte_index(vaddr
) * sizeof(pte_t
);
199 * get_pfn is slow: we could probably try to grab batches of pages here as
200 * an optimization (ie. pre-faulting).
204 * This routine takes a page number given by the Guest and converts it to
205 * an actual, physical page number. It can fail for several reasons: the
206 * virtual address might not be mapped by the Launcher, the write flag is set
207 * and the page is read-only, or the write flag was set and the page was
208 * shared so had to be copied, but we ran out of memory.
210 * This holds a reference to the page, so release_pte() is careful to put that
213 static unsigned long get_pfn(unsigned long virtpfn
, int write
)
217 /* gup me one page at this address please! */
218 if (get_user_pages_fast(virtpfn
<< PAGE_SHIFT
, 1, write
, &page
) == 1)
219 return page_to_pfn(page
);
221 /* This value indicates failure. */
226 * Converting a Guest page table entry to a shadow (ie. real) page table
227 * entry can be a little tricky. The flags are (almost) the same, but the
228 * Guest PTE contains a virtual page number: the CPU needs the real page
231 static pte_t
gpte_to_spte(struct lg_cpu
*cpu
, pte_t gpte
, int write
)
233 unsigned long pfn
, base
, flags
;
236 * The Guest sets the global flag, because it thinks that it is using
237 * PGE. We only told it to use PGE so it would tell us whether it was
238 * flushing a kernel mapping or a userspace mapping. We don't actually
239 * use the global bit, so throw it away.
241 flags
= (pte_flags(gpte
) & ~_PAGE_GLOBAL
);
243 /* The Guest's pages are offset inside the Launcher. */
244 base
= (unsigned long)cpu
->lg
->mem_base
/ PAGE_SIZE
;
247 * We need a temporary "unsigned long" variable to hold the answer from
248 * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't
249 * fit in spte.pfn. get_pfn() finds the real physical number of the
250 * page, given the virtual number.
252 pfn
= get_pfn(base
+ pte_pfn(gpte
), write
);
254 kill_guest(cpu
, "failed to get page %lu", pte_pfn(gpte
));
256 * When we destroy the Guest, we'll go through the shadow page
257 * tables and release_pte() them. Make sure we don't think
262 /* Now we assemble our shadow PTE from the page number and flags. */
263 return pfn_pte(pfn
, __pgprot(flags
));
266 /*H:460 And to complete the chain, release_pte() looks like this: */
267 static void release_pte(pte_t pte
)
270 * Remember that get_user_pages_fast() took a reference to the page, in
271 * get_pfn()? We have to put it back now.
273 if (pte_flags(pte
) & _PAGE_PRESENT
)
274 put_page(pte_page(pte
));
278 static void check_gpte(struct lg_cpu
*cpu
, pte_t gpte
)
280 if ((pte_flags(gpte
) & _PAGE_PSE
) ||
281 pte_pfn(gpte
) >= cpu
->lg
->pfn_limit
)
282 kill_guest(cpu
, "bad page table entry");
285 static void check_gpgd(struct lg_cpu
*cpu
, pgd_t gpgd
)
287 if ((pgd_flags(gpgd
) & ~CHECK_GPGD_MASK
) ||
288 (pgd_pfn(gpgd
) >= cpu
->lg
->pfn_limit
))
289 kill_guest(cpu
, "bad page directory entry");
292 #ifdef CONFIG_X86_PAE
293 static void check_gpmd(struct lg_cpu
*cpu
, pmd_t gpmd
)
295 if ((pmd_flags(gpmd
) & ~_PAGE_TABLE
) ||
296 (pmd_pfn(gpmd
) >= cpu
->lg
->pfn_limit
))
297 kill_guest(cpu
, "bad page middle directory entry");
302 * (i) Looking up a page table entry when the Guest faults.
304 * We saw this call in run_guest(): when we see a page fault in the Guest, we
305 * come here. That's because we only set up the shadow page tables lazily as
306 * they're needed, so we get page faults all the time and quietly fix them up
307 * and return to the Guest without it knowing.
309 * If we fixed up the fault (ie. we mapped the address), this routine returns
310 * true. Otherwise, it was a real fault and we need to tell the Guest.
312 bool demand_page(struct lg_cpu
*cpu
, unsigned long vaddr
, int errcode
)
316 unsigned long gpte_ptr
;
320 /* Mid level for PAE. */
321 #ifdef CONFIG_X86_PAE
326 /* First step: get the top-level Guest page table entry. */
327 gpgd
= lgread(cpu
, gpgd_addr(cpu
, vaddr
), pgd_t
);
328 /* Toplevel not present? We can't map it in. */
329 if (!(pgd_flags(gpgd
) & _PAGE_PRESENT
))
332 /* Now look at the matching shadow entry. */
333 spgd
= spgd_addr(cpu
, cpu
->cpu_pgd
, vaddr
);
334 if (!(pgd_flags(*spgd
) & _PAGE_PRESENT
)) {
335 /* No shadow entry: allocate a new shadow PTE page. */
336 unsigned long ptepage
= get_zeroed_page(GFP_KERNEL
);
338 * This is not really the Guest's fault, but killing it is
339 * simple for this corner case.
342 kill_guest(cpu
, "out of memory allocating pte page");
345 /* We check that the Guest pgd is OK. */
346 check_gpgd(cpu
, gpgd
);
348 * And we copy the flags to the shadow PGD entry. The page
349 * number in the shadow PGD is the page we just allocated.
351 set_pgd(spgd
, __pgd(__pa(ptepage
) | pgd_flags(gpgd
)));
354 #ifdef CONFIG_X86_PAE
355 gpmd
= lgread(cpu
, gpmd_addr(gpgd
, vaddr
), pmd_t
);
356 /* Middle level not present? We can't map it in. */
357 if (!(pmd_flags(gpmd
) & _PAGE_PRESENT
))
360 /* Now look at the matching shadow entry. */
361 spmd
= spmd_addr(cpu
, *spgd
, vaddr
);
363 if (!(pmd_flags(*spmd
) & _PAGE_PRESENT
)) {
364 /* No shadow entry: allocate a new shadow PTE page. */
365 unsigned long ptepage
= get_zeroed_page(GFP_KERNEL
);
368 * This is not really the Guest's fault, but killing it is
369 * simple for this corner case.
372 kill_guest(cpu
, "out of memory allocating pte page");
376 /* We check that the Guest pmd is OK. */
377 check_gpmd(cpu
, gpmd
);
380 * And we copy the flags to the shadow PMD entry. The page
381 * number in the shadow PMD is the page we just allocated.
383 native_set_pmd(spmd
, __pmd(__pa(ptepage
) | pmd_flags(gpmd
)));
387 * OK, now we look at the lower level in the Guest page table: keep its
388 * address, because we might update it later.
390 gpte_ptr
= gpte_addr(cpu
, gpmd
, vaddr
);
393 * OK, now we look at the lower level in the Guest page table: keep its
394 * address, because we might update it later.
396 gpte_ptr
= gpte_addr(cpu
, gpgd
, vaddr
);
399 /* Read the actual PTE value. */
400 gpte
= lgread(cpu
, gpte_ptr
, pte_t
);
402 /* If this page isn't in the Guest page tables, we can't page it in. */
403 if (!(pte_flags(gpte
) & _PAGE_PRESENT
))
407 * Check they're not trying to write to a page the Guest wants
408 * read-only (bit 2 of errcode == write).
410 if ((errcode
& 2) && !(pte_flags(gpte
) & _PAGE_RW
))
413 /* User access to a kernel-only page? (bit 3 == user access) */
414 if ((errcode
& 4) && !(pte_flags(gpte
) & _PAGE_USER
))
418 * Check that the Guest PTE flags are OK, and the page number is below
419 * the pfn_limit (ie. not mapping the Launcher binary).
421 check_gpte(cpu
, gpte
);
423 /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */
424 gpte
= pte_mkyoung(gpte
);
426 gpte
= pte_mkdirty(gpte
);
428 /* Get the pointer to the shadow PTE entry we're going to set. */
429 spte
= spte_addr(cpu
, *spgd
, vaddr
);
432 * If there was a valid shadow PTE entry here before, we release it.
433 * This can happen with a write to a previously read-only entry.
438 * If this is a write, we insist that the Guest page is writable (the
439 * final arg to gpte_to_spte()).
442 *spte
= gpte_to_spte(cpu
, gpte
, 1);
445 * If this is a read, don't set the "writable" bit in the page
446 * table entry, even if the Guest says it's writable. That way
447 * we will come back here when a write does actually occur, so
448 * we can update the Guest's _PAGE_DIRTY flag.
450 native_set_pte(spte
, gpte_to_spte(cpu
, pte_wrprotect(gpte
), 0));
453 * Finally, we write the Guest PTE entry back: we've set the
454 * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags.
456 lgwrite(cpu
, gpte_ptr
, pte_t
, gpte
);
459 * The fault is fixed, the page table is populated, the mapping
460 * manipulated, the result returned and the code complete. A small
461 * delay and a trace of alliteration are the only indications the Guest
462 * has that a page fault occurred at all.
468 * (ii) Making sure the Guest stack is mapped.
470 * Remember that direct traps into the Guest need a mapped Guest kernel stack.
471 * pin_stack_pages() calls us here: we could simply call demand_page(), but as
472 * we've seen that logic is quite long, and usually the stack pages are already
473 * mapped, so it's overkill.
475 * This is a quick version which answers the question: is this virtual address
476 * mapped by the shadow page tables, and is it writable?
478 static bool page_writable(struct lg_cpu
*cpu
, unsigned long vaddr
)
483 #ifdef CONFIG_X86_PAE
486 /* Look at the current top level entry: is it present? */
487 spgd
= spgd_addr(cpu
, cpu
->cpu_pgd
, vaddr
);
488 if (!(pgd_flags(*spgd
) & _PAGE_PRESENT
))
491 #ifdef CONFIG_X86_PAE
492 spmd
= spmd_addr(cpu
, *spgd
, vaddr
);
493 if (!(pmd_flags(*spmd
) & _PAGE_PRESENT
))
498 * Check the flags on the pte entry itself: it must be present and
501 flags
= pte_flags(*(spte_addr(cpu
, *spgd
, vaddr
)));
503 return (flags
& (_PAGE_PRESENT
|_PAGE_RW
)) == (_PAGE_PRESENT
|_PAGE_RW
);
507 * So, when pin_stack_pages() asks us to pin a page, we check if it's already
508 * in the page tables, and if not, we call demand_page() with error code 2
511 void pin_page(struct lg_cpu
*cpu
, unsigned long vaddr
)
513 if (!page_writable(cpu
, vaddr
) && !demand_page(cpu
, vaddr
, 2))
514 kill_guest(cpu
, "bad stack page %#lx", vaddr
);
518 #ifdef CONFIG_X86_PAE
519 static void release_pmd(pmd_t
*spmd
)
521 /* If the entry's not present, there's nothing to release. */
522 if (pmd_flags(*spmd
) & _PAGE_PRESENT
) {
524 pte_t
*ptepage
= __va(pmd_pfn(*spmd
) << PAGE_SHIFT
);
525 /* For each entry in the page, we might need to release it. */
526 for (i
= 0; i
< PTRS_PER_PTE
; i
++)
527 release_pte(ptepage
[i
]);
528 /* Now we can free the page of PTEs */
529 free_page((long)ptepage
);
530 /* And zero out the PMD entry so we never release it twice. */
531 native_set_pmd(spmd
, __pmd(0));
535 static void release_pgd(pgd_t
*spgd
)
537 /* If the entry's not present, there's nothing to release. */
538 if (pgd_flags(*spgd
) & _PAGE_PRESENT
) {
540 pmd_t
*pmdpage
= __va(pgd_pfn(*spgd
) << PAGE_SHIFT
);
542 for (i
= 0; i
< PTRS_PER_PMD
; i
++)
543 release_pmd(&pmdpage
[i
]);
545 /* Now we can free the page of PMDs */
546 free_page((long)pmdpage
);
547 /* And zero out the PGD entry so we never release it twice. */
548 set_pgd(spgd
, __pgd(0));
552 #else /* !CONFIG_X86_PAE */
554 * If we chase down the release_pgd() code, the non-PAE version looks like
555 * this. The PAE version is almost identical, but instead of calling
556 * release_pte it calls release_pmd(), which looks much like this.
558 static void release_pgd(pgd_t
*spgd
)
560 /* If the entry's not present, there's nothing to release. */
561 if (pgd_flags(*spgd
) & _PAGE_PRESENT
) {
564 * Converting the pfn to find the actual PTE page is easy: turn
565 * the page number into a physical address, then convert to a
566 * virtual address (easy for kernel pages like this one).
568 pte_t
*ptepage
= __va(pgd_pfn(*spgd
) << PAGE_SHIFT
);
569 /* For each entry in the page, we might need to release it. */
570 for (i
= 0; i
< PTRS_PER_PTE
; i
++)
571 release_pte(ptepage
[i
]);
572 /* Now we can free the page of PTEs */
573 free_page((long)ptepage
);
574 /* And zero out the PGD entry so we never release it twice. */
581 * We saw flush_user_mappings() twice: once from the flush_user_mappings()
582 * hypercall and once in new_pgdir() when we re-used a top-level pgdir page.
583 * It simply releases every PTE page from 0 up to the Guest's kernel address.
585 static void flush_user_mappings(struct lguest
*lg
, int idx
)
588 /* Release every pgd entry up to the kernel's address. */
589 for (i
= 0; i
< pgd_index(lg
->kernel_address
); i
++)
590 release_pgd(lg
->pgdirs
[idx
].pgdir
+ i
);
594 * (v) Flushing (throwing away) page tables,
596 * The Guest has a hypercall to throw away the page tables: it's used when a
597 * large number of mappings have been changed.
599 void guest_pagetable_flush_user(struct lg_cpu
*cpu
)
601 /* Drop the userspace part of the current page table. */
602 flush_user_mappings(cpu
->lg
, cpu
->cpu_pgd
);
606 /* We walk down the guest page tables to get a guest-physical address */
607 unsigned long guest_pa(struct lg_cpu
*cpu
, unsigned long vaddr
)
611 #ifdef CONFIG_X86_PAE
614 /* First step: get the top-level Guest page table entry. */
615 gpgd
= lgread(cpu
, gpgd_addr(cpu
, vaddr
), pgd_t
);
616 /* Toplevel not present? We can't map it in. */
617 if (!(pgd_flags(gpgd
) & _PAGE_PRESENT
)) {
618 kill_guest(cpu
, "Bad address %#lx", vaddr
);
622 #ifdef CONFIG_X86_PAE
623 gpmd
= lgread(cpu
, gpmd_addr(gpgd
, vaddr
), pmd_t
);
624 if (!(pmd_flags(gpmd
) & _PAGE_PRESENT
))
625 kill_guest(cpu
, "Bad address %#lx", vaddr
);
626 gpte
= lgread(cpu
, gpte_addr(cpu
, gpmd
, vaddr
), pte_t
);
628 gpte
= lgread(cpu
, gpte_addr(cpu
, gpgd
, vaddr
), pte_t
);
630 if (!(pte_flags(gpte
) & _PAGE_PRESENT
))
631 kill_guest(cpu
, "Bad address %#lx", vaddr
);
633 return pte_pfn(gpte
) * PAGE_SIZE
| (vaddr
& ~PAGE_MASK
);
637 * We keep several page tables. This is a simple routine to find the page
638 * table (if any) corresponding to this top-level address the Guest has given
641 static unsigned int find_pgdir(struct lguest
*lg
, unsigned long pgtable
)
644 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
645 if (lg
->pgdirs
[i
].pgdir
&& lg
->pgdirs
[i
].gpgdir
== pgtable
)
651 * And this is us, creating the new page directory. If we really do
652 * allocate a new one (and so the kernel parts are not there), we set
655 static unsigned int new_pgdir(struct lg_cpu
*cpu
,
656 unsigned long gpgdir
,
660 #ifdef CONFIG_X86_PAE
665 * We pick one entry at random to throw out. Choosing the Least
666 * Recently Used might be better, but this is easy.
668 next
= random32() % ARRAY_SIZE(cpu
->lg
->pgdirs
);
669 /* If it's never been allocated at all before, try now. */
670 if (!cpu
->lg
->pgdirs
[next
].pgdir
) {
671 cpu
->lg
->pgdirs
[next
].pgdir
=
672 (pgd_t
*)get_zeroed_page(GFP_KERNEL
);
673 /* If the allocation fails, just keep using the one we have */
674 if (!cpu
->lg
->pgdirs
[next
].pgdir
)
677 #ifdef CONFIG_X86_PAE
679 * In PAE mode, allocate a pmd page and populate the
682 pmd_table
= (pmd_t
*)get_zeroed_page(GFP_KERNEL
);
684 free_page((long)cpu
->lg
->pgdirs
[next
].pgdir
);
685 set_pgd(cpu
->lg
->pgdirs
[next
].pgdir
, __pgd(0));
688 set_pgd(cpu
->lg
->pgdirs
[next
].pgdir
+
690 __pgd(__pa(pmd_table
) | _PAGE_PRESENT
));
692 * This is a blank page, so there are no kernel
693 * mappings: caller must map the stack!
702 /* Record which Guest toplevel this shadows. */
703 cpu
->lg
->pgdirs
[next
].gpgdir
= gpgdir
;
704 /* Release all the non-kernel mappings. */
705 flush_user_mappings(cpu
->lg
, next
);
711 * (iv) Switching page tables
713 * Now we've seen all the page table setting and manipulation, let's see
714 * what happens when the Guest changes page tables (ie. changes the top-level
715 * pgdir). This occurs on almost every context switch.
717 void guest_new_pagetable(struct lg_cpu
*cpu
, unsigned long pgtable
)
719 int newpgdir
, repin
= 0;
721 /* Look to see if we have this one already. */
722 newpgdir
= find_pgdir(cpu
->lg
, pgtable
);
724 * If not, we allocate or mug an existing one: if it's a fresh one,
725 * repin gets set to 1.
727 if (newpgdir
== ARRAY_SIZE(cpu
->lg
->pgdirs
))
728 newpgdir
= new_pgdir(cpu
, pgtable
, &repin
);
729 /* Change the current pgd index to the new one. */
730 cpu
->cpu_pgd
= newpgdir
;
731 /* If it was completely blank, we map in the Guest kernel stack */
733 pin_stack_pages(cpu
);
737 * Finally, a routine which throws away everything: all PGD entries in all
738 * the shadow page tables, including the Guest's kernel mappings. This is used
739 * when we destroy the Guest.
741 static void release_all_pagetables(struct lguest
*lg
)
745 /* Every shadow pagetable this Guest has */
746 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
747 if (lg
->pgdirs
[i
].pgdir
) {
748 #ifdef CONFIG_X86_PAE
753 /* Get the last pmd page. */
754 spgd
= lg
->pgdirs
[i
].pgdir
+ SWITCHER_PGD_INDEX
;
755 pmdpage
= __va(pgd_pfn(*spgd
) << PAGE_SHIFT
);
758 * And release the pmd entries of that pmd page,
759 * except for the switcher pmd.
761 for (k
= 0; k
< SWITCHER_PMD_INDEX
; k
++)
762 release_pmd(&pmdpage
[k
]);
764 /* Every PGD entry except the Switcher at the top */
765 for (j
= 0; j
< SWITCHER_PGD_INDEX
; j
++)
766 release_pgd(lg
->pgdirs
[i
].pgdir
+ j
);
771 * We also throw away everything when a Guest tells us it's changed a kernel
772 * mapping. Since kernel mappings are in every page table, it's easiest to
773 * throw them all away. This traps the Guest in amber for a while as
774 * everything faults back in, but it's rare.
776 void guest_pagetable_clear_all(struct lg_cpu
*cpu
)
778 release_all_pagetables(cpu
->lg
);
779 /* We need the Guest kernel stack mapped again. */
780 pin_stack_pages(cpu
);
785 * Since we throw away all mappings when a kernel mapping changes, our
786 * performance sucks for guests using highmem. In fact, a guest with
787 * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is
788 * usually slower than a Guest with less memory.
790 * This, of course, cannot be fixed. It would take some kind of... well, I
791 * don't know, but the term "puissant code-fu" comes to mind.
795 * This is the routine which actually sets the page table entry for then
796 * "idx"'th shadow page table.
798 * Normally, we can just throw out the old entry and replace it with 0: if they
799 * use it demand_page() will put the new entry in. We need to do this anyway:
800 * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page
801 * is read from, and _PAGE_DIRTY when it's written to.
803 * But Avi Kivity pointed out that most Operating Systems (Linux included) set
804 * these bits on PTEs immediately anyway. This is done to save the CPU from
805 * having to update them, but it helps us the same way: if they set
806 * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if
807 * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately.
809 static void do_set_pte(struct lg_cpu
*cpu
, int idx
,
810 unsigned long vaddr
, pte_t gpte
)
812 /* Look up the matching shadow page directory entry. */
813 pgd_t
*spgd
= spgd_addr(cpu
, idx
, vaddr
);
814 #ifdef CONFIG_X86_PAE
818 /* If the top level isn't present, there's no entry to update. */
819 if (pgd_flags(*spgd
) & _PAGE_PRESENT
) {
820 #ifdef CONFIG_X86_PAE
821 spmd
= spmd_addr(cpu
, *spgd
, vaddr
);
822 if (pmd_flags(*spmd
) & _PAGE_PRESENT
) {
824 /* Otherwise, start by releasing the existing entry. */
825 pte_t
*spte
= spte_addr(cpu
, *spgd
, vaddr
);
829 * If they're setting this entry as dirty or accessed,
830 * we might as well put that entry they've given us in
831 * now. This shaves 10% off a copy-on-write
834 if (pte_flags(gpte
) & (_PAGE_DIRTY
| _PAGE_ACCESSED
)) {
835 check_gpte(cpu
, gpte
);
837 gpte_to_spte(cpu
, gpte
,
838 pte_flags(gpte
) & _PAGE_DIRTY
));
841 * Otherwise kill it and we can demand_page()
844 native_set_pte(spte
, __pte(0));
846 #ifdef CONFIG_X86_PAE
853 * Updating a PTE entry is a little trickier.
855 * We keep track of several different page tables (the Guest uses one for each
856 * process, so it makes sense to cache at least a few). Each of these have
857 * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for
858 * all processes. So when the page table above that address changes, we update
859 * all the page tables, not just the current one. This is rare.
861 * The benefit is that when we have to track a new page table, we can keep all
862 * the kernel mappings. This speeds up context switch immensely.
864 void guest_set_pte(struct lg_cpu
*cpu
,
865 unsigned long gpgdir
, unsigned long vaddr
, pte_t gpte
)
868 * Kernel mappings must be changed on all top levels. Slow, but doesn't
871 if (vaddr
>= cpu
->lg
->kernel_address
) {
873 for (i
= 0; i
< ARRAY_SIZE(cpu
->lg
->pgdirs
); i
++)
874 if (cpu
->lg
->pgdirs
[i
].pgdir
)
875 do_set_pte(cpu
, i
, vaddr
, gpte
);
877 /* Is this page table one we have a shadow for? */
878 int pgdir
= find_pgdir(cpu
->lg
, gpgdir
);
879 if (pgdir
!= ARRAY_SIZE(cpu
->lg
->pgdirs
))
880 /* If so, do the update. */
881 do_set_pte(cpu
, pgdir
, vaddr
, gpte
);
886 * (iii) Setting up a page table entry when the Guest tells us one has changed.
888 * Just like we did in interrupts_and_traps.c, it makes sense for us to deal
889 * with the other side of page tables while we're here: what happens when the
890 * Guest asks for a page table to be updated?
892 * We already saw that demand_page() will fill in the shadow page tables when
893 * needed, so we can simply remove shadow page table entries whenever the Guest
894 * tells us they've changed. When the Guest tries to use the new entry it will
895 * fault and demand_page() will fix it up.
897 * So with that in mind here's our code to to update a (top-level) PGD entry:
899 void guest_set_pgd(struct lguest
*lg
, unsigned long gpgdir
, u32 idx
)
903 if (idx
>= SWITCHER_PGD_INDEX
)
906 /* If they're talking about a page table we have a shadow for... */
907 pgdir
= find_pgdir(lg
, gpgdir
);
908 if (pgdir
< ARRAY_SIZE(lg
->pgdirs
))
909 /* ... throw it away. */
910 release_pgd(lg
->pgdirs
[pgdir
].pgdir
+ idx
);
913 #ifdef CONFIG_X86_PAE
914 /* For setting a mid-level, we just throw everything away. It's easy. */
915 void guest_set_pmd(struct lguest
*lg
, unsigned long pmdp
, u32 idx
)
917 guest_pagetable_clear_all(&lg
->cpus
[0]);
922 * To get through boot, we construct simple identity page mappings (which
923 * set virtual == physical) and linear mappings which will get the Guest far
924 * enough into the boot to create its own. The linear mapping means we
925 * simplify the Guest boot, but it makes assumptions about their PAGE_OFFSET,
928 * We lay them out of the way, just below the initrd (which is why we need to
929 * know its size here).
931 static unsigned long setup_pagetables(struct lguest
*lg
,
933 unsigned long initrd_size
)
936 pte_t __user
*linear
;
937 unsigned long mem_base
= (unsigned long)lg
->mem_base
;
938 unsigned int mapped_pages
, i
, linear_pages
;
939 #ifdef CONFIG_X86_PAE
945 unsigned int phys_linear
;
949 * We have mapped_pages frames to map, so we need linear_pages page
950 * tables to map them.
952 mapped_pages
= mem
/ PAGE_SIZE
;
953 linear_pages
= (mapped_pages
+ PTRS_PER_PTE
- 1) / PTRS_PER_PTE
;
955 /* We put the toplevel page directory page at the top of memory. */
956 pgdir
= (pgd_t
*)(mem
+ mem_base
- initrd_size
- PAGE_SIZE
);
958 /* Now we use the next linear_pages pages as pte pages */
959 linear
= (void *)pgdir
- linear_pages
* PAGE_SIZE
;
961 #ifdef CONFIG_X86_PAE
963 * And the single mid page goes below that. We only use one, but
964 * that's enough to map 1G, which definitely gets us through boot.
966 pmds
= (void *)linear
- PAGE_SIZE
;
969 * Linear mapping is easy: put every page's address into the
972 for (i
= 0; i
< mapped_pages
; i
++) {
974 pte
= pfn_pte(i
, __pgprot(_PAGE_PRESENT
|_PAGE_RW
|_PAGE_USER
));
975 if (copy_to_user(&linear
[i
], &pte
, sizeof(pte
)) != 0)
979 #ifdef CONFIG_X86_PAE
981 * Make the Guest PMD entries point to the corresponding place in the
982 * linear mapping (up to one page worth of PMD).
984 for (i
= j
= 0; i
< mapped_pages
&& j
< PTRS_PER_PMD
;
985 i
+= PTRS_PER_PTE
, j
++) {
986 /* FIXME: native_set_pmd is overkill here. */
987 native_set_pmd(&pmd
, __pmd(((unsigned long)(linear
+ i
)
988 - mem_base
) | _PAGE_PRESENT
| _PAGE_RW
| _PAGE_USER
));
990 if (copy_to_user(&pmds
[j
], &pmd
, sizeof(pmd
)) != 0)
994 /* One PGD entry, pointing to that PMD page. */
995 set_pgd(&pgd
, __pgd(((u32
)pmds
- mem_base
) | _PAGE_PRESENT
));
996 /* Copy it in as the first PGD entry (ie. addresses 0-1G). */
997 if (copy_to_user(&pgdir
[0], &pgd
, sizeof(pgd
)) != 0)
1000 * And the third PGD entry (ie. addresses 3G-4G).
1002 * FIXME: This assumes that PAGE_OFFSET for the Guest is 0xC0000000.
1004 if (copy_to_user(&pgdir
[3], &pgd
, sizeof(pgd
)) != 0)
1008 * The top level points to the linear page table pages above.
1009 * We setup the identity and linear mappings here.
1011 phys_linear
= (unsigned long)linear
- mem_base
;
1012 for (i
= 0; i
< mapped_pages
; i
+= PTRS_PER_PTE
) {
1015 * Create a PGD entry which points to the right part of the
1018 pgd
= __pgd((phys_linear
+ i
* sizeof(pte_t
)) |
1019 (_PAGE_PRESENT
| _PAGE_RW
| _PAGE_USER
));
1022 * Copy it into the PGD page at 0 and PAGE_OFFSET.
1024 if (copy_to_user(&pgdir
[i
/ PTRS_PER_PTE
], &pgd
, sizeof(pgd
))
1025 || copy_to_user(&pgdir
[pgd_index(PAGE_OFFSET
)
1026 + i
/ PTRS_PER_PTE
],
1033 * We return the top level (guest-physical) address: we remember where
1034 * this is to write it into lguest_data when the Guest initializes.
1036 return (unsigned long)pgdir
- mem_base
;
1040 * (vii) Setting up the page tables initially.
1042 * When a Guest is first created, the Launcher tells us where the toplevel of
1043 * its first page table is. We set some things up here:
1045 int init_guest_pagetable(struct lguest
*lg
)
1049 struct boot_params __user
*boot
= (struct boot_params
*)lg
->mem_base
;
1050 #ifdef CONFIG_X86_PAE
1055 * Get the Guest memory size and the ramdisk size from the boot header
1056 * located at lg->mem_base (Guest address 0).
1058 if (copy_from_user(&mem
, &boot
->e820_map
[0].size
, sizeof(mem
))
1059 || get_user(initrd_size
, &boot
->hdr
.ramdisk_size
))
1063 * We start on the first shadow page table, and give it a blank PGD
1066 lg
->pgdirs
[0].gpgdir
= setup_pagetables(lg
, mem
, initrd_size
);
1067 if (IS_ERR_VALUE(lg
->pgdirs
[0].gpgdir
))
1068 return lg
->pgdirs
[0].gpgdir
;
1069 lg
->pgdirs
[0].pgdir
= (pgd_t
*)get_zeroed_page(GFP_KERNEL
);
1070 if (!lg
->pgdirs
[0].pgdir
)
1073 #ifdef CONFIG_X86_PAE
1074 /* For PAE, we also create the initial mid-level. */
1075 pgd
= lg
->pgdirs
[0].pgdir
;
1076 pmd_table
= (pmd_t
*) get_zeroed_page(GFP_KERNEL
);
1080 set_pgd(pgd
+ SWITCHER_PGD_INDEX
,
1081 __pgd(__pa(pmd_table
) | _PAGE_PRESENT
));
1084 /* This is the current page table. */
1085 lg
->cpus
[0].cpu_pgd
= 0;
1089 /*H:508 When the Guest calls LHCALL_LGUEST_INIT we do more setup. */
1090 void page_table_guest_data_init(struct lg_cpu
*cpu
)
1092 /* We get the kernel address: above this is all kernel memory. */
1093 if (get_user(cpu
->lg
->kernel_address
,
1094 &cpu
->lg
->lguest_data
->kernel_address
)
1096 * We tell the Guest that it can't use the top 2 or 4 MB
1097 * of virtual addresses used by the Switcher.
1099 || put_user(RESERVE_MEM
* 1024 * 1024,
1100 &cpu
->lg
->lguest_data
->reserve_mem
)
1101 || put_user(cpu
->lg
->pgdirs
[0].gpgdir
,
1102 &cpu
->lg
->lguest_data
->pgdir
))
1103 kill_guest(cpu
, "bad guest page %p", cpu
->lg
->lguest_data
);
1106 * In flush_user_mappings() we loop from 0 to
1107 * "pgd_index(lg->kernel_address)". This assumes it won't hit the
1108 * Switcher mappings, so check that now.
1110 #ifdef CONFIG_X86_PAE
1111 if (pgd_index(cpu
->lg
->kernel_address
) == SWITCHER_PGD_INDEX
&&
1112 pmd_index(cpu
->lg
->kernel_address
) == SWITCHER_PMD_INDEX
)
1114 if (pgd_index(cpu
->lg
->kernel_address
) >= SWITCHER_PGD_INDEX
)
1116 kill_guest(cpu
, "bad kernel address %#lx",
1117 cpu
->lg
->kernel_address
);
1120 /* When a Guest dies, our cleanup is fairly simple. */
1121 void free_guest_pagetable(struct lguest
*lg
)
1125 /* Throw away all page table pages. */
1126 release_all_pagetables(lg
);
1127 /* Now free the top levels: free_page() can handle 0 just fine. */
1128 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
1129 free_page((long)lg
->pgdirs
[i
].pgdir
);
1133 * (vi) Mapping the Switcher when the Guest is about to run.
1135 * The Switcher and the two pages for this CPU need to be visible in the
1136 * Guest (and not the pages for other CPUs). We have the appropriate PTE pages
1137 * for each CPU already set up, we just need to hook them in now we know which
1138 * Guest is about to run on this CPU.
1140 void map_switcher_in_guest(struct lg_cpu
*cpu
, struct lguest_pages
*pages
)
1142 pte_t
*switcher_pte_page
= __get_cpu_var(switcher_pte_pages
);
1146 #ifdef CONFIG_X86_PAE
1150 /* FIXME: native_set_pmd is overkill here. */
1151 native_set_pmd(&switcher_pmd
, pfn_pmd(__pa(switcher_pte_page
) >>
1152 PAGE_SHIFT
, PAGE_KERNEL_EXEC
));
1154 /* Figure out where the pmd page is, by reading the PGD, and converting
1155 * it to a virtual address. */
1156 pmd_table
= __va(pgd_pfn(cpu
->lg
->
1157 pgdirs
[cpu
->cpu_pgd
].pgdir
[SWITCHER_PGD_INDEX
])
1159 /* Now write it into the shadow page table. */
1160 native_set_pmd(&pmd_table
[SWITCHER_PMD_INDEX
], switcher_pmd
);
1165 * Make the last PGD entry for this Guest point to the Switcher's PTE
1166 * page for this CPU (with appropriate flags).
1168 switcher_pgd
= __pgd(__pa(switcher_pte_page
) | __PAGE_KERNEL_EXEC
);
1170 cpu
->lg
->pgdirs
[cpu
->cpu_pgd
].pgdir
[SWITCHER_PGD_INDEX
] = switcher_pgd
;
1174 * We also change the Switcher PTE page. When we're running the Guest,
1175 * we want the Guest's "regs" page to appear where the first Switcher
1176 * page for this CPU is. This is an optimization: when the Switcher
1177 * saves the Guest registers, it saves them into the first page of this
1178 * CPU's "struct lguest_pages": if we make sure the Guest's register
1179 * page is already mapped there, we don't have to copy them out
1182 pfn
= __pa(cpu
->regs_page
) >> PAGE_SHIFT
;
1183 native_set_pte(®s_pte
, pfn_pte(pfn
, PAGE_KERNEL
));
1184 native_set_pte(&switcher_pte_page
[pte_index((unsigned long)pages
)],
1189 static void free_switcher_pte_pages(void)
1193 for_each_possible_cpu(i
)
1194 free_page((long)switcher_pte_page(i
));
1198 * Setting up the Switcher PTE page for given CPU is fairly easy, given
1199 * the CPU number and the "struct page"s for the Switcher code itself.
1201 * Currently the Switcher is less than a page long, so "pages" is always 1.
1203 static __init
void populate_switcher_pte_page(unsigned int cpu
,
1204 struct page
*switcher_page
[],
1208 pte_t
*pte
= switcher_pte_page(cpu
);
1210 /* The first entries are easy: they map the Switcher code. */
1211 for (i
= 0; i
< pages
; i
++) {
1212 native_set_pte(&pte
[i
], mk_pte(switcher_page
[i
],
1213 __pgprot(_PAGE_PRESENT
|_PAGE_ACCESSED
)));
1216 /* The only other thing we map is this CPU's pair of pages. */
1219 /* First page (Guest registers) is writable from the Guest */
1220 native_set_pte(&pte
[i
], pfn_pte(page_to_pfn(switcher_page
[i
]),
1221 __pgprot(_PAGE_PRESENT
|_PAGE_ACCESSED
|_PAGE_RW
)));
1224 * The second page contains the "struct lguest_ro_state", and is
1227 native_set_pte(&pte
[i
+1], pfn_pte(page_to_pfn(switcher_page
[i
+1]),
1228 __pgprot(_PAGE_PRESENT
|_PAGE_ACCESSED
)));
1232 * We've made it through the page table code. Perhaps our tired brains are
1233 * still processing the details, or perhaps we're simply glad it's over.
1235 * If nothing else, note that all this complexity in juggling shadow page tables
1236 * in sync with the Guest's page tables is for one reason: for most Guests this
1237 * page table dance determines how bad performance will be. This is why Xen
1238 * uses exotic direct Guest pagetable manipulation, and why both Intel and AMD
1239 * have implemented shadow page table support directly into hardware.
1241 * There is just one file remaining in the Host.
1245 * At boot or module load time, init_pagetables() allocates and populates
1246 * the Switcher PTE page for each CPU.
1248 __init
int init_pagetables(struct page
**switcher_page
, unsigned int pages
)
1252 for_each_possible_cpu(i
) {
1253 switcher_pte_page(i
) = (pte_t
*)get_zeroed_page(GFP_KERNEL
);
1254 if (!switcher_pte_page(i
)) {
1255 free_switcher_pte_pages();
1258 populate_switcher_pte_page(i
, switcher_page
, pages
);
1264 /* Cleaning up simply involves freeing the PTE page for each CPU. */
1265 void free_pagetables(void)
1267 free_switcher_pte_pages();
This page took 0.058041 seconds and 5 git commands to generate.