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/gfp.h>
14 #include <linux/types.h>
15 #include <linux/spinlock.h>
16 #include <linux/random.h>
17 #include <linux/percpu.h>
18 #include <asm/tlbflush.h>
19 #include <asm/uaccess.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 /* Return a pointer index'th pgd entry for the i'th page table. */
99 return &cpu
->lg
->pgdirs
[i
].pgdir
[index
];
102 #ifdef CONFIG_X86_PAE
104 * This routine then takes the PGD entry given above, which contains the
105 * address of the PMD page. It then returns a pointer to the PMD entry for the
108 static pmd_t
*spmd_addr(struct lg_cpu
*cpu
, pgd_t spgd
, unsigned long vaddr
)
110 unsigned int index
= pmd_index(vaddr
);
113 /* You should never call this if the PGD entry wasn't valid */
114 BUG_ON(!(pgd_flags(spgd
) & _PAGE_PRESENT
));
115 page
= __va(pgd_pfn(spgd
) << PAGE_SHIFT
);
122 * This routine then takes the page directory entry returned above, which
123 * contains the address of the page table entry (PTE) page. It then returns a
124 * pointer to the PTE entry for the given address.
126 static pte_t
*spte_addr(struct lg_cpu
*cpu
, pgd_t spgd
, unsigned long vaddr
)
128 #ifdef CONFIG_X86_PAE
129 pmd_t
*pmd
= spmd_addr(cpu
, spgd
, vaddr
);
130 pte_t
*page
= __va(pmd_pfn(*pmd
) << PAGE_SHIFT
);
132 /* You should never call this if the PMD entry wasn't valid */
133 BUG_ON(!(pmd_flags(*pmd
) & _PAGE_PRESENT
));
135 pte_t
*page
= __va(pgd_pfn(spgd
) << PAGE_SHIFT
);
136 /* You should never call this if the PGD entry wasn't valid */
137 BUG_ON(!(pgd_flags(spgd
) & _PAGE_PRESENT
));
140 return &page
[pte_index(vaddr
)];
144 * These functions are just like the above, except they access the Guest
145 * page tables. Hence they return a Guest address.
147 static unsigned long gpgd_addr(struct lg_cpu
*cpu
, unsigned long vaddr
)
149 unsigned int index
= vaddr
>> (PGDIR_SHIFT
);
150 return cpu
->lg
->pgdirs
[cpu
->cpu_pgd
].gpgdir
+ index
* sizeof(pgd_t
);
153 #ifdef CONFIG_X86_PAE
154 /* Follow the PGD to the PMD. */
155 static unsigned long gpmd_addr(pgd_t gpgd
, unsigned long vaddr
)
157 unsigned long gpage
= pgd_pfn(gpgd
) << PAGE_SHIFT
;
158 BUG_ON(!(pgd_flags(gpgd
) & _PAGE_PRESENT
));
159 return gpage
+ pmd_index(vaddr
) * sizeof(pmd_t
);
162 /* Follow the PMD to the PTE. */
163 static unsigned long gpte_addr(struct lg_cpu
*cpu
,
164 pmd_t gpmd
, unsigned long vaddr
)
166 unsigned long gpage
= pmd_pfn(gpmd
) << PAGE_SHIFT
;
168 BUG_ON(!(pmd_flags(gpmd
) & _PAGE_PRESENT
));
169 return gpage
+ pte_index(vaddr
) * sizeof(pte_t
);
172 /* Follow the PGD to the PTE (no mid-level for !PAE). */
173 static unsigned long gpte_addr(struct lg_cpu
*cpu
,
174 pgd_t gpgd
, unsigned long vaddr
)
176 unsigned long gpage
= pgd_pfn(gpgd
) << PAGE_SHIFT
;
178 BUG_ON(!(pgd_flags(gpgd
) & _PAGE_PRESENT
));
179 return gpage
+ pte_index(vaddr
) * sizeof(pte_t
);
185 * get_pfn is slow: we could probably try to grab batches of pages here as
186 * an optimization (ie. pre-faulting).
190 * This routine takes a page number given by the Guest and converts it to
191 * an actual, physical page number. It can fail for several reasons: the
192 * virtual address might not be mapped by the Launcher, the write flag is set
193 * and the page is read-only, or the write flag was set and the page was
194 * shared so had to be copied, but we ran out of memory.
196 * This holds a reference to the page, so release_pte() is careful to put that
199 static unsigned long get_pfn(unsigned long virtpfn
, int write
)
203 /* gup me one page at this address please! */
204 if (get_user_pages_fast(virtpfn
<< PAGE_SHIFT
, 1, write
, &page
) == 1)
205 return page_to_pfn(page
);
207 /* This value indicates failure. */
212 * Converting a Guest page table entry to a shadow (ie. real) page table
213 * entry can be a little tricky. The flags are (almost) the same, but the
214 * Guest PTE contains a virtual page number: the CPU needs the real page
217 static pte_t
gpte_to_spte(struct lg_cpu
*cpu
, pte_t gpte
, int write
)
219 unsigned long pfn
, base
, flags
;
222 * The Guest sets the global flag, because it thinks that it is using
223 * PGE. We only told it to use PGE so it would tell us whether it was
224 * flushing a kernel mapping or a userspace mapping. We don't actually
225 * use the global bit, so throw it away.
227 flags
= (pte_flags(gpte
) & ~_PAGE_GLOBAL
);
229 /* The Guest's pages are offset inside the Launcher. */
230 base
= (unsigned long)cpu
->lg
->mem_base
/ PAGE_SIZE
;
233 * We need a temporary "unsigned long" variable to hold the answer from
234 * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't
235 * fit in spte.pfn. get_pfn() finds the real physical number of the
236 * page, given the virtual number.
238 pfn
= get_pfn(base
+ pte_pfn(gpte
), write
);
240 kill_guest(cpu
, "failed to get page %lu", pte_pfn(gpte
));
242 * When we destroy the Guest, we'll go through the shadow page
243 * tables and release_pte() them. Make sure we don't think
248 /* Now we assemble our shadow PTE from the page number and flags. */
249 return pfn_pte(pfn
, __pgprot(flags
));
252 /*H:460 And to complete the chain, release_pte() looks like this: */
253 static void release_pte(pte_t pte
)
256 * Remember that get_user_pages_fast() took a reference to the page, in
257 * get_pfn()? We have to put it back now.
259 if (pte_flags(pte
) & _PAGE_PRESENT
)
260 put_page(pte_page(pte
));
264 static void check_gpte(struct lg_cpu
*cpu
, pte_t gpte
)
266 if ((pte_flags(gpte
) & _PAGE_PSE
) ||
267 pte_pfn(gpte
) >= cpu
->lg
->pfn_limit
)
268 kill_guest(cpu
, "bad page table entry");
271 static void check_gpgd(struct lg_cpu
*cpu
, pgd_t gpgd
)
273 if ((pgd_flags(gpgd
) & ~CHECK_GPGD_MASK
) ||
274 (pgd_pfn(gpgd
) >= cpu
->lg
->pfn_limit
))
275 kill_guest(cpu
, "bad page directory entry");
278 #ifdef CONFIG_X86_PAE
279 static void check_gpmd(struct lg_cpu
*cpu
, pmd_t gpmd
)
281 if ((pmd_flags(gpmd
) & ~_PAGE_TABLE
) ||
282 (pmd_pfn(gpmd
) >= cpu
->lg
->pfn_limit
))
283 kill_guest(cpu
, "bad page middle directory entry");
288 * (i) Looking up a page table entry when the Guest faults.
290 * We saw this call in run_guest(): when we see a page fault in the Guest, we
291 * come here. That's because we only set up the shadow page tables lazily as
292 * they're needed, so we get page faults all the time and quietly fix them up
293 * and return to the Guest without it knowing.
295 * If we fixed up the fault (ie. we mapped the address), this routine returns
296 * true. Otherwise, it was a real fault and we need to tell the Guest.
298 bool demand_page(struct lg_cpu
*cpu
, unsigned long vaddr
, int errcode
)
302 unsigned long gpte_ptr
;
306 /* Mid level for PAE. */
307 #ifdef CONFIG_X86_PAE
312 /* We never demand page the Switcher, so trying is a mistake. */
313 if (vaddr
>= switcher_addr
)
316 /* First step: get the top-level Guest page table entry. */
317 if (unlikely(cpu
->linear_pages
)) {
318 /* Faking up a linear mapping. */
319 gpgd
= __pgd(CHECK_GPGD_MASK
);
321 gpgd
= lgread(cpu
, gpgd_addr(cpu
, vaddr
), pgd_t
);
322 /* Toplevel not present? We can't map it in. */
323 if (!(pgd_flags(gpgd
) & _PAGE_PRESENT
))
327 /* Now look at the matching shadow entry. */
328 spgd
= spgd_addr(cpu
, cpu
->cpu_pgd
, vaddr
);
329 if (!(pgd_flags(*spgd
) & _PAGE_PRESENT
)) {
330 /* No shadow entry: allocate a new shadow PTE page. */
331 unsigned long ptepage
= get_zeroed_page(GFP_KERNEL
);
333 * This is not really the Guest's fault, but killing it is
334 * simple for this corner case.
337 kill_guest(cpu
, "out of memory allocating pte page");
340 /* We check that the Guest pgd is OK. */
341 check_gpgd(cpu
, gpgd
);
343 * And we copy the flags to the shadow PGD entry. The page
344 * number in the shadow PGD is the page we just allocated.
346 set_pgd(spgd
, __pgd(__pa(ptepage
) | pgd_flags(gpgd
)));
349 #ifdef CONFIG_X86_PAE
350 if (unlikely(cpu
->linear_pages
)) {
351 /* Faking up a linear mapping. */
352 gpmd
= __pmd(_PAGE_TABLE
);
354 gpmd
= lgread(cpu
, gpmd_addr(gpgd
, vaddr
), pmd_t
);
355 /* Middle level not present? We can't map it in. */
356 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 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 if (unlikely(cpu
->linear_pages
)) {
400 /* Linear? Make up a PTE which points to same page. */
401 gpte
= __pte((vaddr
& PAGE_MASK
) | _PAGE_RW
| _PAGE_PRESENT
);
403 /* Read the actual PTE value. */
404 gpte
= lgread(cpu
, gpte_ptr
, pte_t
);
407 /* If this page isn't in the Guest page tables, we can't page it in. */
408 if (!(pte_flags(gpte
) & _PAGE_PRESENT
))
412 * Check they're not trying to write to a page the Guest wants
413 * read-only (bit 2 of errcode == write).
415 if ((errcode
& 2) && !(pte_flags(gpte
) & _PAGE_RW
))
418 /* User access to a kernel-only page? (bit 3 == user access) */
419 if ((errcode
& 4) && !(pte_flags(gpte
) & _PAGE_USER
))
423 * Check that the Guest PTE flags are OK, and the page number is below
424 * the pfn_limit (ie. not mapping the Launcher binary).
426 check_gpte(cpu
, gpte
);
428 /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */
429 gpte
= pte_mkyoung(gpte
);
431 gpte
= pte_mkdirty(gpte
);
433 /* Get the pointer to the shadow PTE entry we're going to set. */
434 spte
= spte_addr(cpu
, *spgd
, vaddr
);
437 * If there was a valid shadow PTE entry here before, we release it.
438 * This can happen with a write to a previously read-only entry.
443 * If this is a write, we insist that the Guest page is writable (the
444 * final arg to gpte_to_spte()).
447 *spte
= gpte_to_spte(cpu
, gpte
, 1);
450 * If this is a read, don't set the "writable" bit in the page
451 * table entry, even if the Guest says it's writable. That way
452 * we will come back here when a write does actually occur, so
453 * we can update the Guest's _PAGE_DIRTY flag.
455 set_pte(spte
, gpte_to_spte(cpu
, pte_wrprotect(gpte
), 0));
458 * Finally, we write the Guest PTE entry back: we've set the
459 * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags.
461 if (likely(!cpu
->linear_pages
))
462 lgwrite(cpu
, gpte_ptr
, pte_t
, gpte
);
465 * The fault is fixed, the page table is populated, the mapping
466 * manipulated, the result returned and the code complete. A small
467 * delay and a trace of alliteration are the only indications the Guest
468 * has that a page fault occurred at all.
474 * (ii) Making sure the Guest stack is mapped.
476 * Remember that direct traps into the Guest need a mapped Guest kernel stack.
477 * pin_stack_pages() calls us here: we could simply call demand_page(), but as
478 * we've seen that logic is quite long, and usually the stack pages are already
479 * mapped, so it's overkill.
481 * This is a quick version which answers the question: is this virtual address
482 * mapped by the shadow page tables, and is it writable?
484 static bool page_writable(struct lg_cpu
*cpu
, unsigned long vaddr
)
488 #ifdef CONFIG_X86_PAE
492 /* You can't put your stack in the Switcher! */
493 if (vaddr
>= switcher_addr
)
496 /* Look at the current top level entry: is it present? */
497 spgd
= spgd_addr(cpu
, cpu
->cpu_pgd
, vaddr
);
498 if (!(pgd_flags(*spgd
) & _PAGE_PRESENT
))
501 #ifdef CONFIG_X86_PAE
502 spmd
= spmd_addr(cpu
, *spgd
, vaddr
);
503 if (!(pmd_flags(*spmd
) & _PAGE_PRESENT
))
508 * Check the flags on the pte entry itself: it must be present and
511 flags
= pte_flags(*(spte_addr(cpu
, *spgd
, vaddr
)));
513 return (flags
& (_PAGE_PRESENT
|_PAGE_RW
)) == (_PAGE_PRESENT
|_PAGE_RW
);
517 * So, when pin_stack_pages() asks us to pin a page, we check if it's already
518 * in the page tables, and if not, we call demand_page() with error code 2
521 void pin_page(struct lg_cpu
*cpu
, unsigned long vaddr
)
523 if (!page_writable(cpu
, vaddr
) && !demand_page(cpu
, vaddr
, 2))
524 kill_guest(cpu
, "bad stack page %#lx", vaddr
);
528 #ifdef CONFIG_X86_PAE
529 static void release_pmd(pmd_t
*spmd
)
531 /* If the entry's not present, there's nothing to release. */
532 if (pmd_flags(*spmd
) & _PAGE_PRESENT
) {
534 pte_t
*ptepage
= __va(pmd_pfn(*spmd
) << PAGE_SHIFT
);
535 /* For each entry in the page, we might need to release it. */
536 for (i
= 0; i
< PTRS_PER_PTE
; i
++)
537 release_pte(ptepage
[i
]);
538 /* Now we can free the page of PTEs */
539 free_page((long)ptepage
);
540 /* And zero out the PMD entry so we never release it twice. */
541 set_pmd(spmd
, __pmd(0));
545 static void release_pgd(pgd_t
*spgd
)
547 /* If the entry's not present, there's nothing to release. */
548 if (pgd_flags(*spgd
) & _PAGE_PRESENT
) {
550 pmd_t
*pmdpage
= __va(pgd_pfn(*spgd
) << PAGE_SHIFT
);
552 for (i
= 0; i
< PTRS_PER_PMD
; i
++)
553 release_pmd(&pmdpage
[i
]);
555 /* Now we can free the page of PMDs */
556 free_page((long)pmdpage
);
557 /* And zero out the PGD entry so we never release it twice. */
558 set_pgd(spgd
, __pgd(0));
562 #else /* !CONFIG_X86_PAE */
564 * If we chase down the release_pgd() code, the non-PAE version looks like
565 * this. The PAE version is almost identical, but instead of calling
566 * release_pte it calls release_pmd(), which looks much like this.
568 static void release_pgd(pgd_t
*spgd
)
570 /* If the entry's not present, there's nothing to release. */
571 if (pgd_flags(*spgd
) & _PAGE_PRESENT
) {
574 * Converting the pfn to find the actual PTE page is easy: turn
575 * the page number into a physical address, then convert to a
576 * virtual address (easy for kernel pages like this one).
578 pte_t
*ptepage
= __va(pgd_pfn(*spgd
) << PAGE_SHIFT
);
579 /* For each entry in the page, we might need to release it. */
580 for (i
= 0; i
< PTRS_PER_PTE
; i
++)
581 release_pte(ptepage
[i
]);
582 /* Now we can free the page of PTEs */
583 free_page((long)ptepage
);
584 /* And zero out the PGD entry so we never release it twice. */
591 * We saw flush_user_mappings() twice: once from the flush_user_mappings()
592 * hypercall and once in new_pgdir() when we re-used a top-level pgdir page.
593 * It simply releases every PTE page from 0 up to the Guest's kernel address.
595 static void flush_user_mappings(struct lguest
*lg
, int idx
)
598 /* Release every pgd entry up to the kernel's address. */
599 for (i
= 0; i
< pgd_index(lg
->kernel_address
); i
++)
600 release_pgd(lg
->pgdirs
[idx
].pgdir
+ i
);
604 * (v) Flushing (throwing away) page tables,
606 * The Guest has a hypercall to throw away the page tables: it's used when a
607 * large number of mappings have been changed.
609 void guest_pagetable_flush_user(struct lg_cpu
*cpu
)
611 /* Drop the userspace part of the current page table. */
612 flush_user_mappings(cpu
->lg
, cpu
->cpu_pgd
);
616 /* We walk down the guest page tables to get a guest-physical address */
617 unsigned long guest_pa(struct lg_cpu
*cpu
, unsigned long vaddr
)
621 #ifdef CONFIG_X86_PAE
625 /* Still not set up? Just map 1:1. */
626 if (unlikely(cpu
->linear_pages
))
629 /* First step: get the top-level Guest page table entry. */
630 gpgd
= lgread(cpu
, gpgd_addr(cpu
, vaddr
), pgd_t
);
631 /* Toplevel not present? We can't map it in. */
632 if (!(pgd_flags(gpgd
) & _PAGE_PRESENT
)) {
633 kill_guest(cpu
, "Bad address %#lx", vaddr
);
637 #ifdef CONFIG_X86_PAE
638 gpmd
= lgread(cpu
, gpmd_addr(gpgd
, vaddr
), pmd_t
);
639 if (!(pmd_flags(gpmd
) & _PAGE_PRESENT
))
640 kill_guest(cpu
, "Bad address %#lx", vaddr
);
641 gpte
= lgread(cpu
, gpte_addr(cpu
, gpmd
, vaddr
), pte_t
);
643 gpte
= lgread(cpu
, gpte_addr(cpu
, gpgd
, vaddr
), pte_t
);
645 if (!(pte_flags(gpte
) & _PAGE_PRESENT
))
646 kill_guest(cpu
, "Bad address %#lx", vaddr
);
648 return pte_pfn(gpte
) * PAGE_SIZE
| (vaddr
& ~PAGE_MASK
);
652 * We keep several page tables. This is a simple routine to find the page
653 * table (if any) corresponding to this top-level address the Guest has given
656 static unsigned int find_pgdir(struct lguest
*lg
, unsigned long pgtable
)
659 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
660 if (lg
->pgdirs
[i
].pgdir
&& lg
->pgdirs
[i
].gpgdir
== pgtable
)
666 * And this is us, creating the new page directory. If we really do
667 * allocate a new one (and so the kernel parts are not there), we set
670 static unsigned int new_pgdir(struct lg_cpu
*cpu
,
671 unsigned long gpgdir
,
675 #ifdef CONFIG_X86_PAE
680 * We pick one entry at random to throw out. Choosing the Least
681 * Recently Used might be better, but this is easy.
683 next
= random32() % ARRAY_SIZE(cpu
->lg
->pgdirs
);
684 /* If it's never been allocated at all before, try now. */
685 if (!cpu
->lg
->pgdirs
[next
].pgdir
) {
686 cpu
->lg
->pgdirs
[next
].pgdir
=
687 (pgd_t
*)get_zeroed_page(GFP_KERNEL
);
688 /* If the allocation fails, just keep using the one we have */
689 if (!cpu
->lg
->pgdirs
[next
].pgdir
)
692 #ifdef CONFIG_X86_PAE
694 * In PAE mode, allocate a pmd page and populate the
697 pmd_table
= (pmd_t
*)get_zeroed_page(GFP_KERNEL
);
699 free_page((long)cpu
->lg
->pgdirs
[next
].pgdir
);
700 set_pgd(cpu
->lg
->pgdirs
[next
].pgdir
, __pgd(0));
703 set_pgd(cpu
->lg
->pgdirs
[next
].pgdir
+
705 __pgd(__pa(pmd_table
) | _PAGE_PRESENT
));
707 * This is a blank page, so there are no kernel
708 * mappings: caller must map the stack!
717 /* Record which Guest toplevel this shadows. */
718 cpu
->lg
->pgdirs
[next
].gpgdir
= gpgdir
;
719 /* Release all the non-kernel mappings. */
720 flush_user_mappings(cpu
->lg
, next
);
726 * Finally, a routine which throws away everything: all PGD entries in all
727 * the shadow page tables, including the Guest's kernel mappings. This is used
728 * when we destroy the Guest.
730 static void release_all_pagetables(struct lguest
*lg
)
734 /* Every shadow pagetable this Guest has */
735 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
736 if (lg
->pgdirs
[i
].pgdir
) {
737 #ifdef CONFIG_X86_PAE
742 /* Get the last pmd page. */
743 spgd
= lg
->pgdirs
[i
].pgdir
+ SWITCHER_PGD_INDEX
;
744 pmdpage
= __va(pgd_pfn(*spgd
) << PAGE_SHIFT
);
747 * And release the pmd entries of that pmd page,
748 * except for the switcher pmd.
750 for (k
= 0; k
< SWITCHER_PMD_INDEX
; k
++)
751 release_pmd(&pmdpage
[k
]);
753 /* Every PGD entry except the Switcher at the top */
754 for (j
= 0; j
< SWITCHER_PGD_INDEX
; j
++)
755 release_pgd(lg
->pgdirs
[i
].pgdir
+ j
);
760 * We also throw away everything when a Guest tells us it's changed a kernel
761 * mapping. Since kernel mappings are in every page table, it's easiest to
762 * throw them all away. This traps the Guest in amber for a while as
763 * everything faults back in, but it's rare.
765 void guest_pagetable_clear_all(struct lg_cpu
*cpu
)
767 release_all_pagetables(cpu
->lg
);
768 /* We need the Guest kernel stack mapped again. */
769 pin_stack_pages(cpu
);
773 * (iv) Switching page tables
775 * Now we've seen all the page table setting and manipulation, let's see
776 * what happens when the Guest changes page tables (ie. changes the top-level
777 * pgdir). This occurs on almost every context switch.
779 void guest_new_pagetable(struct lg_cpu
*cpu
, unsigned long pgtable
)
781 int newpgdir
, repin
= 0;
784 * The very first time they call this, we're actually running without
785 * any page tables; we've been making it up. Throw them away now.
787 if (unlikely(cpu
->linear_pages
)) {
788 release_all_pagetables(cpu
->lg
);
789 cpu
->linear_pages
= false;
790 /* Force allocation of a new pgdir. */
791 newpgdir
= ARRAY_SIZE(cpu
->lg
->pgdirs
);
793 /* Look to see if we have this one already. */
794 newpgdir
= find_pgdir(cpu
->lg
, pgtable
);
798 * If not, we allocate or mug an existing one: if it's a fresh one,
799 * repin gets set to 1.
801 if (newpgdir
== ARRAY_SIZE(cpu
->lg
->pgdirs
))
802 newpgdir
= new_pgdir(cpu
, pgtable
, &repin
);
803 /* Change the current pgd index to the new one. */
804 cpu
->cpu_pgd
= newpgdir
;
805 /* If it was completely blank, we map in the Guest kernel stack */
807 pin_stack_pages(cpu
);
812 * Since we throw away all mappings when a kernel mapping changes, our
813 * performance sucks for guests using highmem. In fact, a guest with
814 * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is
815 * usually slower than a Guest with less memory.
817 * This, of course, cannot be fixed. It would take some kind of... well, I
818 * don't know, but the term "puissant code-fu" comes to mind.
822 * This is the routine which actually sets the page table entry for then
823 * "idx"'th shadow page table.
825 * Normally, we can just throw out the old entry and replace it with 0: if they
826 * use it demand_page() will put the new entry in. We need to do this anyway:
827 * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page
828 * is read from, and _PAGE_DIRTY when it's written to.
830 * But Avi Kivity pointed out that most Operating Systems (Linux included) set
831 * these bits on PTEs immediately anyway. This is done to save the CPU from
832 * having to update them, but it helps us the same way: if they set
833 * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if
834 * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately.
836 static void do_set_pte(struct lg_cpu
*cpu
, int idx
,
837 unsigned long vaddr
, pte_t gpte
)
839 /* Look up the matching shadow page directory entry. */
840 pgd_t
*spgd
= spgd_addr(cpu
, idx
, vaddr
);
841 #ifdef CONFIG_X86_PAE
845 /* If the top level isn't present, there's no entry to update. */
846 if (pgd_flags(*spgd
) & _PAGE_PRESENT
) {
847 #ifdef CONFIG_X86_PAE
848 spmd
= spmd_addr(cpu
, *spgd
, vaddr
);
849 if (pmd_flags(*spmd
) & _PAGE_PRESENT
) {
851 /* Otherwise, start by releasing the existing entry. */
852 pte_t
*spte
= spte_addr(cpu
, *spgd
, vaddr
);
856 * If they're setting this entry as dirty or accessed,
857 * we might as well put that entry they've given us in
858 * now. This shaves 10% off a copy-on-write
861 if (pte_flags(gpte
) & (_PAGE_DIRTY
| _PAGE_ACCESSED
)) {
862 check_gpte(cpu
, gpte
);
864 gpte_to_spte(cpu
, gpte
,
865 pte_flags(gpte
) & _PAGE_DIRTY
));
868 * Otherwise kill it and we can demand_page()
871 set_pte(spte
, __pte(0));
873 #ifdef CONFIG_X86_PAE
880 * Updating a PTE entry is a little trickier.
882 * We keep track of several different page tables (the Guest uses one for each
883 * process, so it makes sense to cache at least a few). Each of these have
884 * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for
885 * all processes. So when the page table above that address changes, we update
886 * all the page tables, not just the current one. This is rare.
888 * The benefit is that when we have to track a new page table, we can keep all
889 * the kernel mappings. This speeds up context switch immensely.
891 void guest_set_pte(struct lg_cpu
*cpu
,
892 unsigned long gpgdir
, unsigned long vaddr
, pte_t gpte
)
894 /* We don't let you remap the Switcher; we need it to get back! */
895 if (vaddr
>= switcher_addr
) {
896 kill_guest(cpu
, "attempt to set pte into Switcher pages");
901 * Kernel mappings must be changed on all top levels. Slow, but doesn't
904 if (vaddr
>= cpu
->lg
->kernel_address
) {
906 for (i
= 0; i
< ARRAY_SIZE(cpu
->lg
->pgdirs
); i
++)
907 if (cpu
->lg
->pgdirs
[i
].pgdir
)
908 do_set_pte(cpu
, i
, vaddr
, gpte
);
910 /* Is this page table one we have a shadow for? */
911 int pgdir
= find_pgdir(cpu
->lg
, gpgdir
);
912 if (pgdir
!= ARRAY_SIZE(cpu
->lg
->pgdirs
))
913 /* If so, do the update. */
914 do_set_pte(cpu
, pgdir
, vaddr
, gpte
);
919 * (iii) Setting up a page table entry when the Guest tells us one has changed.
921 * Just like we did in interrupts_and_traps.c, it makes sense for us to deal
922 * with the other side of page tables while we're here: what happens when the
923 * Guest asks for a page table to be updated?
925 * We already saw that demand_page() will fill in the shadow page tables when
926 * needed, so we can simply remove shadow page table entries whenever the Guest
927 * tells us they've changed. When the Guest tries to use the new entry it will
928 * fault and demand_page() will fix it up.
930 * So with that in mind here's our code to update a (top-level) PGD entry:
932 void guest_set_pgd(struct lguest
*lg
, unsigned long gpgdir
, u32 idx
)
936 if (idx
>= SWITCHER_PGD_INDEX
)
939 /* If they're talking about a page table we have a shadow for... */
940 pgdir
= find_pgdir(lg
, gpgdir
);
941 if (pgdir
< ARRAY_SIZE(lg
->pgdirs
))
942 /* ... throw it away. */
943 release_pgd(lg
->pgdirs
[pgdir
].pgdir
+ idx
);
946 #ifdef CONFIG_X86_PAE
947 /* For setting a mid-level, we just throw everything away. It's easy. */
948 void guest_set_pmd(struct lguest
*lg
, unsigned long pmdp
, u32 idx
)
950 guest_pagetable_clear_all(&lg
->cpus
[0]);
955 * (vii) Setting up the page tables initially.
957 * When a Guest is first created, set initialize a shadow page table which
958 * we will populate on future faults. The Guest doesn't have any actual
959 * pagetables yet, so we set linear_pages to tell demand_page() to fake it
962 int init_guest_pagetable(struct lguest
*lg
)
964 struct lg_cpu
*cpu
= &lg
->cpus
[0];
967 /* lg (and lg->cpus[]) starts zeroed: this allocates a new pgdir */
968 cpu
->cpu_pgd
= new_pgdir(cpu
, 0, &allocated
);
972 /* We start with a linear mapping until the initialize. */
973 cpu
->linear_pages
= true;
977 /*H:508 When the Guest calls LHCALL_LGUEST_INIT we do more setup. */
978 void page_table_guest_data_init(struct lg_cpu
*cpu
)
980 /* We get the kernel address: above this is all kernel memory. */
981 if (get_user(cpu
->lg
->kernel_address
,
982 &cpu
->lg
->lguest_data
->kernel_address
)
984 * We tell the Guest that it can't use the top 2 or 4 MB
985 * of virtual addresses used by the Switcher.
987 || put_user(RESERVE_MEM
* 1024 * 1024,
988 &cpu
->lg
->lguest_data
->reserve_mem
)) {
989 kill_guest(cpu
, "bad guest page %p", cpu
->lg
->lguest_data
);
994 * In flush_user_mappings() we loop from 0 to
995 * "pgd_index(lg->kernel_address)". This assumes it won't hit the
996 * Switcher mappings, so check that now.
998 if (cpu
->lg
->kernel_address
>= switcher_addr
)
999 kill_guest(cpu
, "bad kernel address %#lx",
1000 cpu
->lg
->kernel_address
);
1003 /* When a Guest dies, our cleanup is fairly simple. */
1004 void free_guest_pagetable(struct lguest
*lg
)
1008 /* Throw away all page table pages. */
1009 release_all_pagetables(lg
);
1010 /* Now free the top levels: free_page() can handle 0 just fine. */
1011 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
1012 free_page((long)lg
->pgdirs
[i
].pgdir
);
1016 * (vi) Mapping the Switcher when the Guest is about to run.
1018 * The Switcher and the two pages for this CPU need to be visible in the
1019 * Guest (and not the pages for other CPUs). We have the appropriate PTE pages
1020 * for each CPU already set up, we just need to hook them in now we know which
1021 * Guest is about to run on this CPU.
1023 void map_switcher_in_guest(struct lg_cpu
*cpu
, struct lguest_pages
*pages
)
1025 pte_t
*switcher_pte_page
= __this_cpu_read(switcher_pte_pages
);
1028 #ifdef CONFIG_X86_PAE
1032 switcher_pmd
= pfn_pmd(__pa(switcher_pte_page
) >> PAGE_SHIFT
,
1035 /* Figure out where the pmd page is, by reading the PGD, and converting
1036 * it to a virtual address. */
1037 pmd_table
= __va(pgd_pfn(cpu
->lg
->
1038 pgdirs
[cpu
->cpu_pgd
].pgdir
[SWITCHER_PGD_INDEX
])
1040 /* Now write it into the shadow page table. */
1041 set_pmd(&pmd_table
[SWITCHER_PMD_INDEX
], switcher_pmd
);
1046 * Make the last PGD entry for this Guest point to the Switcher's PTE
1047 * page for this CPU (with appropriate flags).
1049 switcher_pgd
= __pgd(__pa(switcher_pte_page
) | __PAGE_KERNEL_EXEC
);
1051 cpu
->lg
->pgdirs
[cpu
->cpu_pgd
].pgdir
[SWITCHER_PGD_INDEX
] = switcher_pgd
;
1055 * We also change the Switcher PTE page. When we're running the Guest,
1056 * we want the Guest's "regs" page to appear where the first Switcher
1057 * page for this CPU is. This is an optimization: when the Switcher
1058 * saves the Guest registers, it saves them into the first page of this
1059 * CPU's "struct lguest_pages": if we make sure the Guest's register
1060 * page is already mapped there, we don't have to copy them out
1063 regs_pte
= pfn_pte(__pa(cpu
->regs_page
) >> PAGE_SHIFT
, PAGE_KERNEL
);
1064 set_pte(&switcher_pte_page
[pte_index((unsigned long)pages
)], regs_pte
);
1068 static void free_switcher_pte_pages(void)
1072 for_each_possible_cpu(i
)
1073 free_page((long)switcher_pte_page(i
));
1077 * Setting up the Switcher PTE page for given CPU is fairly easy, given
1078 * the CPU number and the "struct page"s for the Switcher code itself.
1080 * Currently the Switcher is less than a page long, so "pages" is always 1.
1082 static __init
void populate_switcher_pte_page(unsigned int cpu
,
1083 struct page
*switcher_page
[],
1087 pte_t
*pte
= switcher_pte_page(cpu
);
1089 /* The first entries are easy: they map the Switcher code. */
1090 for (i
= 0; i
< pages
; i
++) {
1091 set_pte(&pte
[i
], mk_pte(switcher_page
[i
],
1092 __pgprot(_PAGE_PRESENT
|_PAGE_ACCESSED
)));
1095 /* The only other thing we map is this CPU's pair of pages. */
1098 /* First page (Guest registers) is writable from the Guest */
1099 set_pte(&pte
[i
], pfn_pte(page_to_pfn(switcher_page
[i
]),
1100 __pgprot(_PAGE_PRESENT
|_PAGE_ACCESSED
|_PAGE_RW
)));
1103 * The second page contains the "struct lguest_ro_state", and is
1106 set_pte(&pte
[i
+1], pfn_pte(page_to_pfn(switcher_page
[i
+1]),
1107 __pgprot(_PAGE_PRESENT
|_PAGE_ACCESSED
)));
1111 * We've made it through the page table code. Perhaps our tired brains are
1112 * still processing the details, or perhaps we're simply glad it's over.
1114 * If nothing else, note that all this complexity in juggling shadow page tables
1115 * in sync with the Guest's page tables is for one reason: for most Guests this
1116 * page table dance determines how bad performance will be. This is why Xen
1117 * uses exotic direct Guest pagetable manipulation, and why both Intel and AMD
1118 * have implemented shadow page table support directly into hardware.
1120 * There is just one file remaining in the Host.
1124 * At boot or module load time, init_pagetables() allocates and populates
1125 * the Switcher PTE page for each CPU.
1127 __init
int init_pagetables(struct page
**switcher_page
, unsigned int pages
)
1131 for_each_possible_cpu(i
) {
1132 switcher_pte_page(i
) = (pte_t
*)get_zeroed_page(GFP_KERNEL
);
1133 if (!switcher_pte_page(i
)) {
1134 free_switcher_pte_pages();
1137 populate_switcher_pte_page(i
, switcher_page
, pages
);
1143 /* Cleaning up simply involves freeing the PTE page for each CPU. */
1144 void free_pagetables(void)
1146 free_switcher_pte_pages();
This page took 0.211695 seconds and 6 git commands to generate.