]>
git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - 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. :*/
8 /* Copyright (C) Rusty Russell IBM Corporation 2006.
9 * GPL v2 and any later version */
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>
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. :*/
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
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?)
39 * Anyway, this is the most complicated part of the Host code. There are seven
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.
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
54 #define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1)
56 /* For PAE we need the PMD index as well. We use the last 2MB, so we
57 * will need the last pmd entry of the last pmd page. */
59 #define SWITCHER_PMD_INDEX (PTRS_PER_PMD - 1)
60 #define RESERVE_MEM 2U
61 #define CHECK_GPGD_MASK _PAGE_PRESENT
63 #define RESERVE_MEM 4U
64 #define CHECK_GPGD_MASK _PAGE_TABLE
67 /* We actually need a separate PTE page for each CPU. Remember that after the
68 * Switcher code itself comes two pages for each CPU, and we don't want this
69 * CPU's guest to see the pages of any other CPU. */
70 static DEFINE_PER_CPU(pte_t
*, switcher_pte_pages
);
71 #define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu)
73 /*H:320 The page table code is curly enough to need helper functions to keep it
76 * There are two functions which return pointers to the shadow (aka "real")
79 * spgd_addr() takes the virtual address and returns a pointer to the top-level
80 * page directory entry (PGD) for that address. Since we keep track of several
81 * page tables, the "i" argument tells us which one we're interested in (it's
82 * usually the current one). */
83 static pgd_t
*spgd_addr(struct lg_cpu
*cpu
, u32 i
, unsigned long vaddr
)
85 unsigned int index
= pgd_index(vaddr
);
87 #ifndef CONFIG_X86_PAE
88 /* We kill any Guest trying to touch the Switcher addresses. */
89 if (index
>= SWITCHER_PGD_INDEX
) {
90 kill_guest(cpu
, "attempt to access switcher pages");
94 /* Return a pointer index'th pgd entry for the i'th page table. */
95 return &cpu
->lg
->pgdirs
[i
].pgdir
[index
];
99 /* This routine then takes the PGD entry given above, which contains the
100 * address of the PMD page. It then returns a pointer to the PMD entry for the
102 static pmd_t
*spmd_addr(struct lg_cpu
*cpu
, pgd_t spgd
, unsigned long vaddr
)
104 unsigned int index
= pmd_index(vaddr
);
107 /* We kill any Guest trying to touch the Switcher addresses. */
108 if (pgd_index(vaddr
) == SWITCHER_PGD_INDEX
&&
109 index
>= SWITCHER_PMD_INDEX
) {
110 kill_guest(cpu
, "attempt to access switcher pages");
114 /* You should never call this if the PGD entry wasn't valid */
115 BUG_ON(!(pgd_flags(spgd
) & _PAGE_PRESENT
));
116 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. */
125 static pte_t
*spte_addr(struct lg_cpu
*cpu
, pgd_t spgd
, unsigned long vaddr
)
127 #ifdef CONFIG_X86_PAE
128 pmd_t
*pmd
= spmd_addr(cpu
, spgd
, vaddr
);
129 pte_t
*page
= __va(pmd_pfn(*pmd
) << PAGE_SHIFT
);
131 /* You should never call this if the PMD entry wasn't valid */
132 BUG_ON(!(pmd_flags(*pmd
) & _PAGE_PRESENT
));
134 pte_t
*page
= __va(pgd_pfn(spgd
) << PAGE_SHIFT
);
135 /* You should never call this if the PGD entry wasn't valid */
136 BUG_ON(!(pgd_flags(spgd
) & _PAGE_PRESENT
));
139 return &page
[pte_index(vaddr
)];
142 /* These two functions just like the above two, except they access the Guest
143 * page tables. Hence they return a Guest address. */
144 static unsigned long gpgd_addr(struct lg_cpu
*cpu
, unsigned long vaddr
)
146 unsigned int index
= vaddr
>> (PGDIR_SHIFT
);
147 return cpu
->lg
->pgdirs
[cpu
->cpu_pgd
].gpgdir
+ index
* sizeof(pgd_t
);
150 #ifdef CONFIG_X86_PAE
151 static unsigned long gpmd_addr(pgd_t gpgd
, unsigned long vaddr
)
153 unsigned long gpage
= pgd_pfn(gpgd
) << PAGE_SHIFT
;
154 BUG_ON(!(pgd_flags(gpgd
) & _PAGE_PRESENT
));
155 return gpage
+ pmd_index(vaddr
) * sizeof(pmd_t
);
158 static unsigned long gpte_addr(struct lg_cpu
*cpu
,
159 pmd_t gpmd
, unsigned long vaddr
)
161 unsigned long gpage
= pmd_pfn(gpmd
) << PAGE_SHIFT
;
163 BUG_ON(!(pmd_flags(gpmd
) & _PAGE_PRESENT
));
164 return gpage
+ pte_index(vaddr
) * sizeof(pte_t
);
167 static unsigned long gpte_addr(struct lg_cpu
*cpu
,
168 pgd_t gpgd
, unsigned long vaddr
)
170 unsigned long gpage
= pgd_pfn(gpgd
) << PAGE_SHIFT
;
172 BUG_ON(!(pgd_flags(gpgd
) & _PAGE_PRESENT
));
173 return gpage
+ pte_index(vaddr
) * sizeof(pte_t
);
178 /*M:014 get_pfn is slow: we could probably try to grab batches of pages here as
179 * an optimization (ie. pre-faulting). :*/
181 /*H:350 This routine takes a page number given by the Guest and converts it to
182 * an actual, physical page number. It can fail for several reasons: the
183 * virtual address might not be mapped by the Launcher, the write flag is set
184 * and the page is read-only, or the write flag was set and the page was
185 * shared so had to be copied, but we ran out of memory.
187 * This holds a reference to the page, so release_pte() is careful to put that
189 static unsigned long get_pfn(unsigned long virtpfn
, int write
)
193 /* gup me one page at this address please! */
194 if (get_user_pages_fast(virtpfn
<< PAGE_SHIFT
, 1, write
, &page
) == 1)
195 return page_to_pfn(page
);
197 /* This value indicates failure. */
201 /*H:340 Converting a Guest page table entry to a shadow (ie. real) page table
202 * entry can be a little tricky. The flags are (almost) the same, but the
203 * Guest PTE contains a virtual page number: the CPU needs the real page
205 static pte_t
gpte_to_spte(struct lg_cpu
*cpu
, pte_t gpte
, int write
)
207 unsigned long pfn
, base
, flags
;
209 /* The Guest sets the global flag, because it thinks that it is using
210 * PGE. We only told it to use PGE so it would tell us whether it was
211 * flushing a kernel mapping or a userspace mapping. We don't actually
212 * use the global bit, so throw it away. */
213 flags
= (pte_flags(gpte
) & ~_PAGE_GLOBAL
);
215 /* The Guest's pages are offset inside the Launcher. */
216 base
= (unsigned long)cpu
->lg
->mem_base
/ PAGE_SIZE
;
218 /* We need a temporary "unsigned long" variable to hold the answer from
219 * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't
220 * fit in spte.pfn. get_pfn() finds the real physical number of the
221 * page, given the virtual number. */
222 pfn
= get_pfn(base
+ pte_pfn(gpte
), write
);
224 kill_guest(cpu
, "failed to get page %lu", pte_pfn(gpte
));
225 /* When we destroy the Guest, we'll go through the shadow page
226 * tables and release_pte() them. Make sure we don't think
227 * this one is valid! */
230 /* Now we assemble our shadow PTE from the page number and flags. */
231 return pfn_pte(pfn
, __pgprot(flags
));
234 /*H:460 And to complete the chain, release_pte() looks like this: */
235 static void release_pte(pte_t pte
)
237 /* Remember that get_user_pages_fast() took a reference to the page, in
238 * get_pfn()? We have to put it back now. */
239 if (pte_flags(pte
) & _PAGE_PRESENT
)
240 put_page(pte_page(pte
));
244 static void check_gpte(struct lg_cpu
*cpu
, pte_t gpte
)
246 if ((pte_flags(gpte
) & _PAGE_PSE
) ||
247 pte_pfn(gpte
) >= cpu
->lg
->pfn_limit
)
248 kill_guest(cpu
, "bad page table entry");
251 static void check_gpgd(struct lg_cpu
*cpu
, pgd_t gpgd
)
253 if ((pgd_flags(gpgd
) & ~CHECK_GPGD_MASK
) ||
254 (pgd_pfn(gpgd
) >= cpu
->lg
->pfn_limit
))
255 kill_guest(cpu
, "bad page directory entry");
258 #ifdef CONFIG_X86_PAE
259 static void check_gpmd(struct lg_cpu
*cpu
, pmd_t gpmd
)
261 if ((pmd_flags(gpmd
) & ~_PAGE_TABLE
) ||
262 (pmd_pfn(gpmd
) >= cpu
->lg
->pfn_limit
))
263 kill_guest(cpu
, "bad page middle directory entry");
268 * (i) Looking up a page table entry when the Guest faults.
270 * We saw this call in run_guest(): when we see a page fault in the Guest, we
271 * come here. That's because we only set up the shadow page tables lazily as
272 * they're needed, so we get page faults all the time and quietly fix them up
273 * and return to the Guest without it knowing.
275 * If we fixed up the fault (ie. we mapped the address), this routine returns
276 * true. Otherwise, it was a real fault and we need to tell the Guest. */
277 bool demand_page(struct lg_cpu
*cpu
, unsigned long vaddr
, int errcode
)
281 unsigned long gpte_ptr
;
285 #ifdef CONFIG_X86_PAE
290 /* First step: get the top-level Guest page table entry. */
291 gpgd
= lgread(cpu
, gpgd_addr(cpu
, vaddr
), pgd_t
);
292 /* Toplevel not present? We can't map it in. */
293 if (!(pgd_flags(gpgd
) & _PAGE_PRESENT
))
296 /* Now look at the matching shadow entry. */
297 spgd
= spgd_addr(cpu
, cpu
->cpu_pgd
, vaddr
);
298 if (!(pgd_flags(*spgd
) & _PAGE_PRESENT
)) {
299 /* No shadow entry: allocate a new shadow PTE page. */
300 unsigned long ptepage
= get_zeroed_page(GFP_KERNEL
);
301 /* This is not really the Guest's fault, but killing it is
302 * simple for this corner case. */
304 kill_guest(cpu
, "out of memory allocating pte page");
307 /* We check that the Guest pgd is OK. */
308 check_gpgd(cpu
, gpgd
);
309 /* And we copy the flags to the shadow PGD entry. The page
310 * number in the shadow PGD is the page we just allocated. */
311 set_pgd(spgd
, __pgd(__pa(ptepage
) | pgd_flags(gpgd
)));
314 #ifdef CONFIG_X86_PAE
315 gpmd
= lgread(cpu
, gpmd_addr(gpgd
, vaddr
), pmd_t
);
316 /* middle level not present? We can't map it in. */
317 if (!(pmd_flags(gpmd
) & _PAGE_PRESENT
))
320 /* Now look at the matching shadow entry. */
321 spmd
= spmd_addr(cpu
, *spgd
, vaddr
);
323 if (!(pmd_flags(*spmd
) & _PAGE_PRESENT
)) {
324 /* No shadow entry: allocate a new shadow PTE page. */
325 unsigned long ptepage
= get_zeroed_page(GFP_KERNEL
);
327 /* This is not really the Guest's fault, but killing it is
328 * simple for this corner case. */
330 kill_guest(cpu
, "out of memory allocating pte page");
334 /* We check that the Guest pmd is OK. */
335 check_gpmd(cpu
, gpmd
);
337 /* And we copy the flags to the shadow PMD entry. The page
338 * number in the shadow PMD is the page we just allocated. */
339 native_set_pmd(spmd
, __pmd(__pa(ptepage
) | pmd_flags(gpmd
)));
342 /* OK, now we look at the lower level in the Guest page table: keep its
343 * address, because we might update it later. */
344 gpte_ptr
= gpte_addr(cpu
, gpmd
, vaddr
);
346 /* OK, now we look at the lower level in the Guest page table: keep its
347 * address, because we might update it later. */
348 gpte_ptr
= gpte_addr(cpu
, gpgd
, vaddr
);
350 gpte
= lgread(cpu
, gpte_ptr
, pte_t
);
352 /* If this page isn't in the Guest page tables, we can't page it in. */
353 if (!(pte_flags(gpte
) & _PAGE_PRESENT
))
356 /* Check they're not trying to write to a page the Guest wants
357 * read-only (bit 2 of errcode == write). */
358 if ((errcode
& 2) && !(pte_flags(gpte
) & _PAGE_RW
))
361 /* User access to a kernel-only page? (bit 3 == user access) */
362 if ((errcode
& 4) && !(pte_flags(gpte
) & _PAGE_USER
))
365 /* Check that the Guest PTE flags are OK, and the page number is below
366 * the pfn_limit (ie. not mapping the Launcher binary). */
367 check_gpte(cpu
, gpte
);
369 /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */
370 gpte
= pte_mkyoung(gpte
);
372 gpte
= pte_mkdirty(gpte
);
374 /* Get the pointer to the shadow PTE entry we're going to set. */
375 spte
= spte_addr(cpu
, *spgd
, vaddr
);
376 /* If there was a valid shadow PTE entry here before, we release it.
377 * This can happen with a write to a previously read-only entry. */
380 /* If this is a write, we insist that the Guest page is writable (the
381 * final arg to gpte_to_spte()). */
383 *spte
= gpte_to_spte(cpu
, gpte
, 1);
385 /* If this is a read, don't set the "writable" bit in the page
386 * table entry, even if the Guest says it's writable. That way
387 * we will come back here when a write does actually occur, so
388 * we can update the Guest's _PAGE_DIRTY flag. */
389 native_set_pte(spte
, gpte_to_spte(cpu
, pte_wrprotect(gpte
), 0));
391 /* Finally, we write the Guest PTE entry back: we've set the
392 * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. */
393 lgwrite(cpu
, gpte_ptr
, pte_t
, gpte
);
395 /* The fault is fixed, the page table is populated, the mapping
396 * manipulated, the result returned and the code complete. A small
397 * delay and a trace of alliteration are the only indications the Guest
398 * has that a page fault occurred at all. */
403 * (ii) Making sure the Guest stack is mapped.
405 * Remember that direct traps into the Guest need a mapped Guest kernel stack.
406 * pin_stack_pages() calls us here: we could simply call demand_page(), but as
407 * we've seen that logic is quite long, and usually the stack pages are already
408 * mapped, so it's overkill.
410 * This is a quick version which answers the question: is this virtual address
411 * mapped by the shadow page tables, and is it writable? */
412 static bool page_writable(struct lg_cpu
*cpu
, unsigned long vaddr
)
417 #ifdef CONFIG_X86_PAE
420 /* Look at the current top level entry: is it present? */
421 spgd
= spgd_addr(cpu
, cpu
->cpu_pgd
, vaddr
);
422 if (!(pgd_flags(*spgd
) & _PAGE_PRESENT
))
425 #ifdef CONFIG_X86_PAE
426 spmd
= spmd_addr(cpu
, *spgd
, vaddr
);
427 if (!(pmd_flags(*spmd
) & _PAGE_PRESENT
))
431 /* Check the flags on the pte entry itself: it must be present and
433 flags
= pte_flags(*(spte_addr(cpu
, *spgd
, vaddr
)));
435 return (flags
& (_PAGE_PRESENT
|_PAGE_RW
)) == (_PAGE_PRESENT
|_PAGE_RW
);
438 /* So, when pin_stack_pages() asks us to pin a page, we check if it's already
439 * in the page tables, and if not, we call demand_page() with error code 2
440 * (meaning "write"). */
441 void pin_page(struct lg_cpu
*cpu
, unsigned long vaddr
)
443 if (!page_writable(cpu
, vaddr
) && !demand_page(cpu
, vaddr
, 2))
444 kill_guest(cpu
, "bad stack page %#lx", vaddr
);
447 #ifdef CONFIG_X86_PAE
448 static void release_pmd(pmd_t
*spmd
)
450 /* If the entry's not present, there's nothing to release. */
451 if (pmd_flags(*spmd
) & _PAGE_PRESENT
) {
453 pte_t
*ptepage
= __va(pmd_pfn(*spmd
) << PAGE_SHIFT
);
454 /* For each entry in the page, we might need to release it. */
455 for (i
= 0; i
< PTRS_PER_PTE
; i
++)
456 release_pte(ptepage
[i
]);
457 /* Now we can free the page of PTEs */
458 free_page((long)ptepage
);
459 /* And zero out the PMD entry so we never release it twice. */
460 native_set_pmd(spmd
, __pmd(0));
464 static void release_pgd(pgd_t
*spgd
)
466 /* If the entry's not present, there's nothing to release. */
467 if (pgd_flags(*spgd
) & _PAGE_PRESENT
) {
469 pmd_t
*pmdpage
= __va(pgd_pfn(*spgd
) << PAGE_SHIFT
);
471 for (i
= 0; i
< PTRS_PER_PMD
; i
++)
472 release_pmd(&pmdpage
[i
]);
474 /* Now we can free the page of PMDs */
475 free_page((long)pmdpage
);
476 /* And zero out the PGD entry so we never release it twice. */
477 set_pgd(spgd
, __pgd(0));
481 #else /* !CONFIG_X86_PAE */
482 /*H:450 If we chase down the release_pgd() code, it looks like this: */
483 static void release_pgd(pgd_t
*spgd
)
485 /* If the entry's not present, there's nothing to release. */
486 if (pgd_flags(*spgd
) & _PAGE_PRESENT
) {
488 /* Converting the pfn to find the actual PTE page is easy: turn
489 * the page number into a physical address, then convert to a
490 * virtual address (easy for kernel pages like this one). */
491 pte_t
*ptepage
= __va(pgd_pfn(*spgd
) << PAGE_SHIFT
);
492 /* For each entry in the page, we might need to release it. */
493 for (i
= 0; i
< PTRS_PER_PTE
; i
++)
494 release_pte(ptepage
[i
]);
495 /* Now we can free the page of PTEs */
496 free_page((long)ptepage
);
497 /* And zero out the PGD entry so we never release it twice. */
502 /*H:445 We saw flush_user_mappings() twice: once from the flush_user_mappings()
503 * hypercall and once in new_pgdir() when we re-used a top-level pgdir page.
504 * It simply releases every PTE page from 0 up to the Guest's kernel address. */
505 static void flush_user_mappings(struct lguest
*lg
, int idx
)
508 /* Release every pgd entry up to the kernel's address. */
509 for (i
= 0; i
< pgd_index(lg
->kernel_address
); i
++)
510 release_pgd(lg
->pgdirs
[idx
].pgdir
+ i
);
513 /*H:440 (v) Flushing (throwing away) page tables,
515 * The Guest has a hypercall to throw away the page tables: it's used when a
516 * large number of mappings have been changed. */
517 void guest_pagetable_flush_user(struct lg_cpu
*cpu
)
519 /* Drop the userspace part of the current page table. */
520 flush_user_mappings(cpu
->lg
, cpu
->cpu_pgd
);
524 /* We walk down the guest page tables to get a guest-physical address */
525 unsigned long guest_pa(struct lg_cpu
*cpu
, unsigned long vaddr
)
529 #ifdef CONFIG_X86_PAE
532 /* First step: get the top-level Guest page table entry. */
533 gpgd
= lgread(cpu
, gpgd_addr(cpu
, vaddr
), pgd_t
);
534 /* Toplevel not present? We can't map it in. */
535 if (!(pgd_flags(gpgd
) & _PAGE_PRESENT
)) {
536 kill_guest(cpu
, "Bad address %#lx", vaddr
);
540 #ifdef CONFIG_X86_PAE
541 gpmd
= lgread(cpu
, gpmd_addr(gpgd
, vaddr
), pmd_t
);
542 if (!(pmd_flags(gpmd
) & _PAGE_PRESENT
))
543 kill_guest(cpu
, "Bad address %#lx", vaddr
);
544 gpte
= lgread(cpu
, gpte_addr(cpu
, gpmd
, vaddr
), pte_t
);
546 gpte
= lgread(cpu
, gpte_addr(cpu
, gpgd
, vaddr
), pte_t
);
548 if (!(pte_flags(gpte
) & _PAGE_PRESENT
))
549 kill_guest(cpu
, "Bad address %#lx", vaddr
);
551 return pte_pfn(gpte
) * PAGE_SIZE
| (vaddr
& ~PAGE_MASK
);
554 /* We keep several page tables. This is a simple routine to find the page
555 * table (if any) corresponding to this top-level address the Guest has given
557 static unsigned int find_pgdir(struct lguest
*lg
, unsigned long pgtable
)
560 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
561 if (lg
->pgdirs
[i
].pgdir
&& lg
->pgdirs
[i
].gpgdir
== pgtable
)
566 /*H:435 And this is us, creating the new page directory. If we really do
567 * allocate a new one (and so the kernel parts are not there), we set
569 static unsigned int new_pgdir(struct lg_cpu
*cpu
,
570 unsigned long gpgdir
,
574 #ifdef CONFIG_X86_PAE
578 /* We pick one entry at random to throw out. Choosing the Least
579 * Recently Used might be better, but this is easy. */
580 next
= random32() % ARRAY_SIZE(cpu
->lg
->pgdirs
);
581 /* If it's never been allocated at all before, try now. */
582 if (!cpu
->lg
->pgdirs
[next
].pgdir
) {
583 cpu
->lg
->pgdirs
[next
].pgdir
=
584 (pgd_t
*)get_zeroed_page(GFP_KERNEL
);
585 /* If the allocation fails, just keep using the one we have */
586 if (!cpu
->lg
->pgdirs
[next
].pgdir
)
589 #ifdef CONFIG_X86_PAE
590 /* In PAE mode, allocate a pmd page and populate the
592 pmd_table
= (pmd_t
*)get_zeroed_page(GFP_KERNEL
);
594 free_page((long)cpu
->lg
->pgdirs
[next
].pgdir
);
595 set_pgd(cpu
->lg
->pgdirs
[next
].pgdir
, __pgd(0));
598 set_pgd(cpu
->lg
->pgdirs
[next
].pgdir
+
600 __pgd(__pa(pmd_table
) | _PAGE_PRESENT
));
601 /* This is a blank page, so there are no kernel
602 * mappings: caller must map the stack! */
610 /* Record which Guest toplevel this shadows. */
611 cpu
->lg
->pgdirs
[next
].gpgdir
= gpgdir
;
612 /* Release all the non-kernel mappings. */
613 flush_user_mappings(cpu
->lg
, next
);
618 /*H:430 (iv) Switching page tables
620 * Now we've seen all the page table setting and manipulation, let's see
621 * what happens when the Guest changes page tables (ie. changes the top-level
622 * pgdir). This occurs on almost every context switch. */
623 void guest_new_pagetable(struct lg_cpu
*cpu
, unsigned long pgtable
)
625 int newpgdir
, repin
= 0;
627 /* Look to see if we have this one already. */
628 newpgdir
= find_pgdir(cpu
->lg
, pgtable
);
629 /* If not, we allocate or mug an existing one: if it's a fresh one,
630 * repin gets set to 1. */
631 if (newpgdir
== ARRAY_SIZE(cpu
->lg
->pgdirs
))
632 newpgdir
= new_pgdir(cpu
, pgtable
, &repin
);
633 /* Change the current pgd index to the new one. */
634 cpu
->cpu_pgd
= newpgdir
;
635 /* If it was completely blank, we map in the Guest kernel stack */
637 pin_stack_pages(cpu
);
640 /*H:470 Finally, a routine which throws away everything: all PGD entries in all
641 * the shadow page tables, including the Guest's kernel mappings. This is used
642 * when we destroy the Guest. */
643 static void release_all_pagetables(struct lguest
*lg
)
647 /* Every shadow pagetable this Guest has */
648 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
649 if (lg
->pgdirs
[i
].pgdir
) {
650 #ifdef CONFIG_X86_PAE
655 /* Get the last pmd page. */
656 spgd
= lg
->pgdirs
[i
].pgdir
+ SWITCHER_PGD_INDEX
;
657 pmdpage
= __va(pgd_pfn(*spgd
) << PAGE_SHIFT
);
659 /* And release the pmd entries of that pmd page,
660 * except for the switcher pmd. */
661 for (k
= 0; k
< SWITCHER_PMD_INDEX
; k
++)
662 release_pmd(&pmdpage
[k
]);
664 /* Every PGD entry except the Switcher at the top */
665 for (j
= 0; j
< SWITCHER_PGD_INDEX
; j
++)
666 release_pgd(lg
->pgdirs
[i
].pgdir
+ j
);
670 /* We also throw away everything when a Guest tells us it's changed a kernel
671 * mapping. Since kernel mappings are in every page table, it's easiest to
672 * throw them all away. This traps the Guest in amber for a while as
673 * everything faults back in, but it's rare. */
674 void guest_pagetable_clear_all(struct lg_cpu
*cpu
)
676 release_all_pagetables(cpu
->lg
);
677 /* We need the Guest kernel stack mapped again. */
678 pin_stack_pages(cpu
);
681 /*M:009 Since we throw away all mappings when a kernel mapping changes, our
682 * performance sucks for guests using highmem. In fact, a guest with
683 * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is
684 * usually slower than a Guest with less memory.
686 * This, of course, cannot be fixed. It would take some kind of... well, I
687 * don't know, but the term "puissant code-fu" comes to mind. :*/
689 /*H:420 This is the routine which actually sets the page table entry for then
690 * "idx"'th shadow page table.
692 * Normally, we can just throw out the old entry and replace it with 0: if they
693 * use it demand_page() will put the new entry in. We need to do this anyway:
694 * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page
695 * is read from, and _PAGE_DIRTY when it's written to.
697 * But Avi Kivity pointed out that most Operating Systems (Linux included) set
698 * these bits on PTEs immediately anyway. This is done to save the CPU from
699 * having to update them, but it helps us the same way: if they set
700 * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if
701 * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately.
703 static void do_set_pte(struct lg_cpu
*cpu
, int idx
,
704 unsigned long vaddr
, pte_t gpte
)
706 /* Look up the matching shadow page directory entry. */
707 pgd_t
*spgd
= spgd_addr(cpu
, idx
, vaddr
);
708 #ifdef CONFIG_X86_PAE
712 /* If the top level isn't present, there's no entry to update. */
713 if (pgd_flags(*spgd
) & _PAGE_PRESENT
) {
714 #ifdef CONFIG_X86_PAE
715 spmd
= spmd_addr(cpu
, *spgd
, vaddr
);
716 if (pmd_flags(*spmd
) & _PAGE_PRESENT
) {
718 /* Otherwise, we start by releasing
719 * the existing entry. */
720 pte_t
*spte
= spte_addr(cpu
, *spgd
, vaddr
);
723 /* If they're setting this entry as dirty or accessed,
724 * we might as well put that entry they've given us
725 * in now. This shaves 10% off a
726 * copy-on-write micro-benchmark. */
727 if (pte_flags(gpte
) & (_PAGE_DIRTY
| _PAGE_ACCESSED
)) {
728 check_gpte(cpu
, gpte
);
730 gpte_to_spte(cpu
, gpte
,
731 pte_flags(gpte
) & _PAGE_DIRTY
));
733 /* Otherwise kill it and we can demand_page()
735 native_set_pte(spte
, __pte(0));
736 #ifdef CONFIG_X86_PAE
742 /*H:410 Updating a PTE entry is a little trickier.
744 * We keep track of several different page tables (the Guest uses one for each
745 * process, so it makes sense to cache at least a few). Each of these have
746 * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for
747 * all processes. So when the page table above that address changes, we update
748 * all the page tables, not just the current one. This is rare.
750 * The benefit is that when we have to track a new page table, we can keep all
751 * the kernel mappings. This speeds up context switch immensely. */
752 void guest_set_pte(struct lg_cpu
*cpu
,
753 unsigned long gpgdir
, unsigned long vaddr
, pte_t gpte
)
755 /* Kernel mappings must be changed on all top levels. Slow, but doesn't
757 if (vaddr
>= cpu
->lg
->kernel_address
) {
759 for (i
= 0; i
< ARRAY_SIZE(cpu
->lg
->pgdirs
); i
++)
760 if (cpu
->lg
->pgdirs
[i
].pgdir
)
761 do_set_pte(cpu
, i
, vaddr
, gpte
);
763 /* Is this page table one we have a shadow for? */
764 int pgdir
= find_pgdir(cpu
->lg
, gpgdir
);
765 if (pgdir
!= ARRAY_SIZE(cpu
->lg
->pgdirs
))
766 /* If so, do the update. */
767 do_set_pte(cpu
, pgdir
, vaddr
, gpte
);
772 * (iii) Setting up a page table entry when the Guest tells us one has changed.
774 * Just like we did in interrupts_and_traps.c, it makes sense for us to deal
775 * with the other side of page tables while we're here: what happens when the
776 * Guest asks for a page table to be updated?
778 * We already saw that demand_page() will fill in the shadow page tables when
779 * needed, so we can simply remove shadow page table entries whenever the Guest
780 * tells us they've changed. When the Guest tries to use the new entry it will
781 * fault and demand_page() will fix it up.
783 * So with that in mind here's our code to to update a (top-level) PGD entry:
785 void guest_set_pgd(struct lguest
*lg
, unsigned long gpgdir
, u32 idx
)
789 if (idx
>= SWITCHER_PGD_INDEX
)
792 /* If they're talking about a page table we have a shadow for... */
793 pgdir
= find_pgdir(lg
, gpgdir
);
794 if (pgdir
< ARRAY_SIZE(lg
->pgdirs
))
795 /* ... throw it away. */
796 release_pgd(lg
->pgdirs
[pgdir
].pgdir
+ idx
);
798 #ifdef CONFIG_X86_PAE
799 void guest_set_pmd(struct lguest
*lg
, unsigned long pmdp
, u32 idx
)
801 guest_pagetable_clear_all(&lg
->cpus
[0]);
805 /* Once we know how much memory we have we can construct simple identity
806 * (which set virtual == physical) and linear mappings
807 * which will get the Guest far enough into the boot to create its own.
809 * We lay them out of the way, just below the initrd (which is why we need to
810 * know its size here). */
811 static unsigned long setup_pagetables(struct lguest
*lg
,
813 unsigned long initrd_size
)
816 pte_t __user
*linear
;
817 unsigned long mem_base
= (unsigned long)lg
->mem_base
;
818 unsigned int mapped_pages
, i
, linear_pages
;
819 #ifdef CONFIG_X86_PAE
825 unsigned int phys_linear
;
828 /* We have mapped_pages frames to map, so we need
829 * linear_pages page tables to map them. */
830 mapped_pages
= mem
/ PAGE_SIZE
;
831 linear_pages
= (mapped_pages
+ PTRS_PER_PTE
- 1) / PTRS_PER_PTE
;
833 /* We put the toplevel page directory page at the top of memory. */
834 pgdir
= (pgd_t
*)(mem
+ mem_base
- initrd_size
- PAGE_SIZE
);
836 /* Now we use the next linear_pages pages as pte pages */
837 linear
= (void *)pgdir
- linear_pages
* PAGE_SIZE
;
839 #ifdef CONFIG_X86_PAE
840 pmds
= (void *)linear
- PAGE_SIZE
;
842 /* Linear mapping is easy: put every page's address into the
843 * mapping in order. */
844 for (i
= 0; i
< mapped_pages
; i
++) {
846 pte
= pfn_pte(i
, __pgprot(_PAGE_PRESENT
|_PAGE_RW
|_PAGE_USER
));
847 if (copy_to_user(&linear
[i
], &pte
, sizeof(pte
)) != 0)
851 /* The top level points to the linear page table pages above.
852 * We setup the identity and linear mappings here. */
853 #ifdef CONFIG_X86_PAE
854 for (i
= j
= 0; i
< mapped_pages
&& j
< PTRS_PER_PMD
;
855 i
+= PTRS_PER_PTE
, j
++) {
856 native_set_pmd(&pmd
, __pmd(((unsigned long)(linear
+ i
)
857 - mem_base
) | _PAGE_PRESENT
| _PAGE_RW
| _PAGE_USER
));
859 if (copy_to_user(&pmds
[j
], &pmd
, sizeof(pmd
)) != 0)
863 set_pgd(&pgd
, __pgd(((u32
)pmds
- mem_base
) | _PAGE_PRESENT
));
864 if (copy_to_user(&pgdir
[0], &pgd
, sizeof(pgd
)) != 0)
866 if (copy_to_user(&pgdir
[3], &pgd
, sizeof(pgd
)) != 0)
869 phys_linear
= (unsigned long)linear
- mem_base
;
870 for (i
= 0; i
< mapped_pages
; i
+= PTRS_PER_PTE
) {
872 pgd
= __pgd((phys_linear
+ i
* sizeof(pte_t
)) |
873 (_PAGE_PRESENT
| _PAGE_RW
| _PAGE_USER
));
875 if (copy_to_user(&pgdir
[i
/ PTRS_PER_PTE
], &pgd
, sizeof(pgd
))
876 || copy_to_user(&pgdir
[pgd_index(PAGE_OFFSET
)
883 /* We return the top level (guest-physical) address: remember where
885 return (unsigned long)pgdir
- mem_base
;
888 /*H:500 (vii) Setting up the page tables initially.
890 * When a Guest is first created, the Launcher tells us where the toplevel of
891 * its first page table is. We set some things up here: */
892 int init_guest_pagetable(struct lguest
*lg
)
896 struct boot_params __user
*boot
= (struct boot_params
*)lg
->mem_base
;
897 #ifdef CONFIG_X86_PAE
901 /* Get the Guest memory size and the ramdisk size from the boot header
902 * located at lg->mem_base (Guest address 0). */
903 if (copy_from_user(&mem
, &boot
->e820_map
[0].size
, sizeof(mem
))
904 || get_user(initrd_size
, &boot
->hdr
.ramdisk_size
))
907 /* We start on the first shadow page table, and give it a blank PGD
909 lg
->pgdirs
[0].gpgdir
= setup_pagetables(lg
, mem
, initrd_size
);
910 if (IS_ERR_VALUE(lg
->pgdirs
[0].gpgdir
))
911 return lg
->pgdirs
[0].gpgdir
;
912 lg
->pgdirs
[0].pgdir
= (pgd_t
*)get_zeroed_page(GFP_KERNEL
);
913 if (!lg
->pgdirs
[0].pgdir
)
915 #ifdef CONFIG_X86_PAE
916 pgd
= lg
->pgdirs
[0].pgdir
;
917 pmd_table
= (pmd_t
*) get_zeroed_page(GFP_KERNEL
);
921 set_pgd(pgd
+ SWITCHER_PGD_INDEX
,
922 __pgd(__pa(pmd_table
) | _PAGE_PRESENT
));
924 lg
->cpus
[0].cpu_pgd
= 0;
928 /* When the Guest calls LHCALL_LGUEST_INIT we do more setup. */
929 void page_table_guest_data_init(struct lg_cpu
*cpu
)
931 /* We get the kernel address: above this is all kernel memory. */
932 if (get_user(cpu
->lg
->kernel_address
,
933 &cpu
->lg
->lguest_data
->kernel_address
)
934 /* We tell the Guest that it can't use the top 2 or 4 MB
935 * of virtual addresses used by the Switcher. */
936 || put_user(RESERVE_MEM
* 1024 * 1024,
937 &cpu
->lg
->lguest_data
->reserve_mem
)
938 || put_user(cpu
->lg
->pgdirs
[0].gpgdir
,
939 &cpu
->lg
->lguest_data
->pgdir
))
940 kill_guest(cpu
, "bad guest page %p", cpu
->lg
->lguest_data
);
942 /* In flush_user_mappings() we loop from 0 to
943 * "pgd_index(lg->kernel_address)". This assumes it won't hit the
944 * Switcher mappings, so check that now. */
945 #ifdef CONFIG_X86_PAE
946 if (pgd_index(cpu
->lg
->kernel_address
) == SWITCHER_PGD_INDEX
&&
947 pmd_index(cpu
->lg
->kernel_address
) == SWITCHER_PMD_INDEX
)
949 if (pgd_index(cpu
->lg
->kernel_address
) >= SWITCHER_PGD_INDEX
)
951 kill_guest(cpu
, "bad kernel address %#lx",
952 cpu
->lg
->kernel_address
);
955 /* When a Guest dies, our cleanup is fairly simple. */
956 void free_guest_pagetable(struct lguest
*lg
)
960 /* Throw away all page table pages. */
961 release_all_pagetables(lg
);
962 /* Now free the top levels: free_page() can handle 0 just fine. */
963 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
964 free_page((long)lg
->pgdirs
[i
].pgdir
);
967 /*H:480 (vi) Mapping the Switcher when the Guest is about to run.
969 * The Switcher and the two pages for this CPU need to be visible in the
970 * Guest (and not the pages for other CPUs). We have the appropriate PTE pages
971 * for each CPU already set up, we just need to hook them in now we know which
972 * Guest is about to run on this CPU. */
973 void map_switcher_in_guest(struct lg_cpu
*cpu
, struct lguest_pages
*pages
)
975 pte_t
*switcher_pte_page
= __get_cpu_var(switcher_pte_pages
);
979 #ifdef CONFIG_X86_PAE
983 native_set_pmd(&switcher_pmd
, pfn_pmd(__pa(switcher_pte_page
) >>
984 PAGE_SHIFT
, PAGE_KERNEL_EXEC
));
986 pmd_table
= __va(pgd_pfn(cpu
->lg
->
987 pgdirs
[cpu
->cpu_pgd
].pgdir
[SWITCHER_PGD_INDEX
])
989 native_set_pmd(&pmd_table
[SWITCHER_PMD_INDEX
], switcher_pmd
);
993 /* Make the last PGD entry for this Guest point to the Switcher's PTE
994 * page for this CPU (with appropriate flags). */
995 switcher_pgd
= __pgd(__pa(switcher_pte_page
) | __PAGE_KERNEL_EXEC
);
997 cpu
->lg
->pgdirs
[cpu
->cpu_pgd
].pgdir
[SWITCHER_PGD_INDEX
] = switcher_pgd
;
1000 /* We also change the Switcher PTE page. When we're running the Guest,
1001 * we want the Guest's "regs" page to appear where the first Switcher
1002 * page for this CPU is. This is an optimization: when the Switcher
1003 * saves the Guest registers, it saves them into the first page of this
1004 * CPU's "struct lguest_pages": if we make sure the Guest's register
1005 * page is already mapped there, we don't have to copy them out
1007 pfn
= __pa(cpu
->regs_page
) >> PAGE_SHIFT
;
1008 native_set_pte(®s_pte
, pfn_pte(pfn
, PAGE_KERNEL
));
1009 native_set_pte(&switcher_pte_page
[pte_index((unsigned long)pages
)],
1014 static void free_switcher_pte_pages(void)
1018 for_each_possible_cpu(i
)
1019 free_page((long)switcher_pte_page(i
));
1022 /*H:520 Setting up the Switcher PTE page for given CPU is fairly easy, given
1023 * the CPU number and the "struct page"s for the Switcher code itself.
1025 * Currently the Switcher is less than a page long, so "pages" is always 1. */
1026 static __init
void populate_switcher_pte_page(unsigned int cpu
,
1027 struct page
*switcher_page
[],
1031 pte_t
*pte
= switcher_pte_page(cpu
);
1033 /* The first entries are easy: they map the Switcher code. */
1034 for (i
= 0; i
< pages
; i
++) {
1035 native_set_pte(&pte
[i
], mk_pte(switcher_page
[i
],
1036 __pgprot(_PAGE_PRESENT
|_PAGE_ACCESSED
)));
1039 /* The only other thing we map is this CPU's pair of pages. */
1042 /* First page (Guest registers) is writable from the Guest */
1043 native_set_pte(&pte
[i
], pfn_pte(page_to_pfn(switcher_page
[i
]),
1044 __pgprot(_PAGE_PRESENT
|_PAGE_ACCESSED
|_PAGE_RW
)));
1046 /* The second page contains the "struct lguest_ro_state", and is
1048 native_set_pte(&pte
[i
+1], pfn_pte(page_to_pfn(switcher_page
[i
+1]),
1049 __pgprot(_PAGE_PRESENT
|_PAGE_ACCESSED
)));
1052 /* We've made it through the page table code. Perhaps our tired brains are
1053 * still processing the details, or perhaps we're simply glad it's over.
1055 * If nothing else, note that all this complexity in juggling shadow page tables
1056 * in sync with the Guest's page tables is for one reason: for most Guests this
1057 * page table dance determines how bad performance will be. This is why Xen
1058 * uses exotic direct Guest pagetable manipulation, and why both Intel and AMD
1059 * have implemented shadow page table support directly into hardware.
1061 * There is just one file remaining in the Host. */
1063 /*H:510 At boot or module load time, init_pagetables() allocates and populates
1064 * the Switcher PTE page for each CPU. */
1065 __init
int init_pagetables(struct page
**switcher_page
, unsigned int pages
)
1069 for_each_possible_cpu(i
) {
1070 switcher_pte_page(i
) = (pte_t
*)get_zeroed_page(GFP_KERNEL
);
1071 if (!switcher_pte_page(i
)) {
1072 free_switcher_pte_pages();
1075 populate_switcher_pte_page(i
, switcher_page
, pages
);
1081 /* Cleaning up simply involves freeing the PTE page for each CPU. */
1082 void free_pagetables(void)
1084 free_switcher_pte_pages();