]>
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 to point the hardware to the
6 * actual 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>
19 /*M:008 We hold reference to pages, which prevents them from being swapped.
20 * It'd be nice to have a callback in the "struct mm_struct" when Linux wants
21 * to swap out. If we had this, and a shrinker callback to trim PTE pages, we
22 * could probably consider launching Guests as non-root. :*/
27 * We use two-level page tables for the Guest. If you're not entirely
28 * comfortable with virtual addresses, physical addresses and page tables then
29 * I recommend you review lguest.c's "Page Table Handling" (with diagrams!).
31 * The Guest keeps page tables, but we maintain the actual ones here: these are
32 * called "shadow" page tables. Which is a very Guest-centric name: these are
33 * the real page tables the CPU uses, although we keep them up to date to
34 * reflect the Guest's. (See what I mean about weird naming? Since when do
35 * shadows reflect anything?)
37 * Anyway, this is the most complicated part of the Host code. There are seven
39 * (i) Setting up a page table entry for the Guest when it faults,
40 * (ii) Setting up the page table entry for the Guest stack,
41 * (iii) Setting up a page table entry when the Guest tells us it has changed,
42 * (iv) Switching page tables,
43 * (v) Flushing (thowing away) page tables,
44 * (vi) Mapping the Switcher when the Guest is about to run,
45 * (vii) Setting up the page tables initially.
49 /* 1024 entries in a page table page maps 1024 pages: 4MB. The Switcher is
50 * conveniently placed at the top 4MB, so it uses a separate, complete PTE
52 #define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1)
54 /* We actually need a separate PTE page for each CPU. Remember that after the
55 * Switcher code itself comes two pages for each CPU, and we don't want this
56 * CPU's guest to see the pages of any other CPU. */
57 static DEFINE_PER_CPU(pte_t
*, switcher_pte_pages
);
58 #define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu)
60 /*H:320 With our shadow and Guest types established, we need to deal with
61 * them: the page table code is curly enough to need helper functions to keep
64 * There are two functions which return pointers to the shadow (aka "real")
67 * spgd_addr() takes the virtual address and returns a pointer to the top-level
68 * page directory entry for that address. Since we keep track of several page
69 * tables, the "i" argument tells us which one we're interested in (it's
70 * usually the current one). */
71 static pgd_t
*spgd_addr(struct lguest
*lg
, u32 i
, unsigned long vaddr
)
73 unsigned int index
= pgd_index(vaddr
);
75 /* We kill any Guest trying to touch the Switcher addresses. */
76 if (index
>= SWITCHER_PGD_INDEX
) {
77 kill_guest(lg
, "attempt to access switcher pages");
80 /* Return a pointer index'th pgd entry for the i'th page table. */
81 return &lg
->pgdirs
[i
].pgdir
[index
];
84 /* This routine then takes the PGD entry given above, which contains the
85 * address of the PTE page. It then returns a pointer to the PTE entry for the
87 static pte_t
*spte_addr(struct lguest
*lg
, pgd_t spgd
, unsigned long vaddr
)
89 pte_t
*page
= __va(pgd_pfn(spgd
) << PAGE_SHIFT
);
90 /* You should never call this if the PGD entry wasn't valid */
91 BUG_ON(!(pgd_flags(spgd
) & _PAGE_PRESENT
));
92 return &page
[(vaddr
>> PAGE_SHIFT
) % PTRS_PER_PTE
];
95 /* These two functions just like the above two, except they access the Guest
96 * page tables. Hence they return a Guest address. */
97 static unsigned long gpgd_addr(struct lguest
*lg
, unsigned long vaddr
)
99 unsigned int index
= vaddr
>> (PGDIR_SHIFT
);
100 return lg
->pgdirs
[lg
->pgdidx
].gpgdir
+ index
* sizeof(pgd_t
);
103 static unsigned long gpte_addr(struct lguest
*lg
,
104 pgd_t gpgd
, unsigned long vaddr
)
106 unsigned long gpage
= pgd_pfn(gpgd
) << PAGE_SHIFT
;
107 BUG_ON(!(pgd_flags(gpgd
) & _PAGE_PRESENT
));
108 return gpage
+ ((vaddr
>>PAGE_SHIFT
) % PTRS_PER_PTE
) * sizeof(pte_t
);
111 /*H:350 This routine takes a page number given by the Guest and converts it to
112 * an actual, physical page number. It can fail for several reasons: the
113 * virtual address might not be mapped by the Launcher, the write flag is set
114 * and the page is read-only, or the write flag was set and the page was
115 * shared so had to be copied, but we ran out of memory.
117 * This holds a reference to the page, so release_pte() is careful to
119 static unsigned long get_pfn(unsigned long virtpfn
, int write
)
122 /* This value indicates failure. */
123 unsigned long ret
= -1UL;
125 /* get_user_pages() is a complex interface: it gets the "struct
126 * vm_area_struct" and "struct page" assocated with a range of pages.
127 * It also needs the task's mmap_sem held, and is not very quick.
128 * It returns the number of pages it got. */
129 down_read(¤t
->mm
->mmap_sem
);
130 if (get_user_pages(current
, current
->mm
, virtpfn
<< PAGE_SHIFT
,
131 1, write
, 1, &page
, NULL
) == 1)
132 ret
= page_to_pfn(page
);
133 up_read(¤t
->mm
->mmap_sem
);
137 /*H:340 Converting a Guest page table entry to a shadow (ie. real) page table
138 * entry can be a little tricky. The flags are (almost) the same, but the
139 * Guest PTE contains a virtual page number: the CPU needs the real page
141 static pte_t
gpte_to_spte(struct lguest
*lg
, pte_t gpte
, int write
)
143 unsigned long pfn
, base
, flags
;
145 /* The Guest sets the global flag, because it thinks that it is using
146 * PGE. We only told it to use PGE so it would tell us whether it was
147 * flushing a kernel mapping or a userspace mapping. We don't actually
148 * use the global bit, so throw it away. */
149 flags
= (pte_flags(gpte
) & ~_PAGE_GLOBAL
);
151 /* The Guest's pages are offset inside the Launcher. */
152 base
= (unsigned long)lg
->mem_base
/ PAGE_SIZE
;
154 /* We need a temporary "unsigned long" variable to hold the answer from
155 * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't
156 * fit in spte.pfn. get_pfn() finds the real physical number of the
157 * page, given the virtual number. */
158 pfn
= get_pfn(base
+ pte_pfn(gpte
), write
);
160 kill_guest(lg
, "failed to get page %lu", pte_pfn(gpte
));
161 /* When we destroy the Guest, we'll go through the shadow page
162 * tables and release_pte() them. Make sure we don't think
163 * this one is valid! */
166 /* Now we assemble our shadow PTE from the page number and flags. */
167 return pfn_pte(pfn
, __pgprot(flags
));
170 /*H:460 And to complete the chain, release_pte() looks like this: */
171 static void release_pte(pte_t pte
)
173 /* Remember that get_user_pages() took a reference to the page, in
174 * get_pfn()? We have to put it back now. */
175 if (pte_flags(pte
) & _PAGE_PRESENT
)
176 put_page(pfn_to_page(pte_pfn(pte
)));
180 static void check_gpte(struct lguest
*lg
, pte_t gpte
)
182 if ((pte_flags(gpte
) & (_PAGE_PWT
|_PAGE_PSE
))
183 || pte_pfn(gpte
) >= lg
->pfn_limit
)
184 kill_guest(lg
, "bad page table entry");
187 static void check_gpgd(struct lguest
*lg
, pgd_t gpgd
)
189 if ((pgd_flags(gpgd
) & ~_PAGE_TABLE
) || pgd_pfn(gpgd
) >= lg
->pfn_limit
)
190 kill_guest(lg
, "bad page directory entry");
194 * (i) Setting up a page table entry for the Guest when it faults
196 * We saw this call in run_guest(): when we see a page fault in the Guest, we
197 * come here. That's because we only set up the shadow page tables lazily as
198 * they're needed, so we get page faults all the time and quietly fix them up
199 * and return to the Guest without it knowing.
201 * If we fixed up the fault (ie. we mapped the address), this routine returns
203 int demand_page(struct lguest
*lg
, unsigned long vaddr
, int errcode
)
207 unsigned long gpte_ptr
;
211 /* First step: get the top-level Guest page table entry. */
212 gpgd
= lgread(lg
, gpgd_addr(lg
, vaddr
), pgd_t
);
213 /* Toplevel not present? We can't map it in. */
214 if (!(pgd_flags(gpgd
) & _PAGE_PRESENT
))
217 /* Now look at the matching shadow entry. */
218 spgd
= spgd_addr(lg
, lg
->pgdidx
, vaddr
);
219 if (!(pgd_flags(*spgd
) & _PAGE_PRESENT
)) {
220 /* No shadow entry: allocate a new shadow PTE page. */
221 unsigned long ptepage
= get_zeroed_page(GFP_KERNEL
);
222 /* This is not really the Guest's fault, but killing it is
223 * simple for this corner case. */
225 kill_guest(lg
, "out of memory allocating pte page");
228 /* We check that the Guest pgd is OK. */
229 check_gpgd(lg
, gpgd
);
230 /* And we copy the flags to the shadow PGD entry. The page
231 * number in the shadow PGD is the page we just allocated. */
232 *spgd
= __pgd(__pa(ptepage
) | pgd_flags(gpgd
));
235 /* OK, now we look at the lower level in the Guest page table: keep its
236 * address, because we might update it later. */
237 gpte_ptr
= gpte_addr(lg
, gpgd
, vaddr
);
238 gpte
= lgread(lg
, gpte_ptr
, pte_t
);
240 /* If this page isn't in the Guest page tables, we can't page it in. */
241 if (!(pte_flags(gpte
) & _PAGE_PRESENT
))
244 /* Check they're not trying to write to a page the Guest wants
245 * read-only (bit 2 of errcode == write). */
246 if ((errcode
& 2) && !(pte_flags(gpte
) & _PAGE_RW
))
249 /* User access to a kernel page? (bit 3 == user access) */
250 if ((errcode
& 4) && !(pte_flags(gpte
) & _PAGE_USER
))
253 /* Check that the Guest PTE flags are OK, and the page number is below
254 * the pfn_limit (ie. not mapping the Launcher binary). */
255 check_gpte(lg
, gpte
);
256 /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */
257 gpte
= pte_mkyoung(gpte
);
260 gpte
= pte_mkdirty(gpte
);
262 /* Get the pointer to the shadow PTE entry we're going to set. */
263 spte
= spte_addr(lg
, *spgd
, vaddr
);
264 /* If there was a valid shadow PTE entry here before, we release it.
265 * This can happen with a write to a previously read-only entry. */
268 /* If this is a write, we insist that the Guest page is writable (the
269 * final arg to gpte_to_spte()). */
271 *spte
= gpte_to_spte(lg
, gpte
, 1);
273 /* If this is a read, don't set the "writable" bit in the page
274 * table entry, even if the Guest says it's writable. That way
275 * we come back here when a write does actually ocur, so we can
276 * update the Guest's _PAGE_DIRTY flag. */
277 *spte
= gpte_to_spte(lg
, pte_wrprotect(gpte
), 0);
279 /* Finally, we write the Guest PTE entry back: we've set the
280 * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. */
281 lgwrite(lg
, gpte_ptr
, pte_t
, gpte
);
283 /* We succeeded in mapping the page! */
287 /*H:360 (ii) Setting up the page table entry for the Guest stack.
289 * Remember pin_stack_pages() which makes sure the stack is mapped? It could
290 * simply call demand_page(), but as we've seen that logic is quite long, and
291 * usually the stack pages are already mapped anyway, so it's not required.
293 * This is a quick version which answers the question: is this virtual address
294 * mapped by the shadow page tables, and is it writable? */
295 static int page_writable(struct lguest
*lg
, unsigned long vaddr
)
300 /* Look at the top level entry: is it present? */
301 spgd
= spgd_addr(lg
, lg
->pgdidx
, vaddr
);
302 if (!(pgd_flags(*spgd
) & _PAGE_PRESENT
))
305 /* Check the flags on the pte entry itself: it must be present and
307 flags
= pte_flags(*(spte_addr(lg
, *spgd
, vaddr
)));
309 return (flags
& (_PAGE_PRESENT
|_PAGE_RW
)) == (_PAGE_PRESENT
|_PAGE_RW
);
312 /* So, when pin_stack_pages() asks us to pin a page, we check if it's already
313 * in the page tables, and if not, we call demand_page() with error code 2
314 * (meaning "write"). */
315 void pin_page(struct lguest
*lg
, unsigned long vaddr
)
317 if (!page_writable(lg
, vaddr
) && !demand_page(lg
, vaddr
, 2))
318 kill_guest(lg
, "bad stack page %#lx", vaddr
);
321 /*H:450 If we chase down the release_pgd() code, it looks like this: */
322 static void release_pgd(struct lguest
*lg
, pgd_t
*spgd
)
324 /* If the entry's not present, there's nothing to release. */
325 if (pgd_flags(*spgd
) & _PAGE_PRESENT
) {
327 /* Converting the pfn to find the actual PTE page is easy: turn
328 * the page number into a physical address, then convert to a
329 * virtual address (easy for kernel pages like this one). */
330 pte_t
*ptepage
= __va(pgd_pfn(*spgd
) << PAGE_SHIFT
);
331 /* For each entry in the page, we might need to release it. */
332 for (i
= 0; i
< PTRS_PER_PTE
; i
++)
333 release_pte(ptepage
[i
]);
334 /* Now we can free the page of PTEs */
335 free_page((long)ptepage
);
336 /* And zero out the PGD entry we we never release it twice. */
341 /*H:440 (v) Flushing (thowing away) page tables,
343 * We saw flush_user_mappings() called when we re-used a top-level pgdir page.
344 * It simply releases every PTE page from 0 up to the kernel address. */
345 static void flush_user_mappings(struct lguest
*lg
, int idx
)
348 /* Release every pgd entry up to the kernel's address. */
349 for (i
= 0; i
< pgd_index(lg
->kernel_address
); i
++)
350 release_pgd(lg
, lg
->pgdirs
[idx
].pgdir
+ i
);
353 /* The Guest also has a hypercall to do this manually: it's used when a large
354 * number of mappings have been changed. */
355 void guest_pagetable_flush_user(struct lguest
*lg
)
357 /* Drop the userspace part of the current page table. */
358 flush_user_mappings(lg
, lg
->pgdidx
);
362 /* We walk down the guest page tables to get a guest-physical address */
363 unsigned long guest_pa(struct lguest
*lg
, unsigned long vaddr
)
368 /* First step: get the top-level Guest page table entry. */
369 gpgd
= lgread(lg
, gpgd_addr(lg
, vaddr
), pgd_t
);
370 /* Toplevel not present? We can't map it in. */
371 if (!(pgd_flags(gpgd
) & _PAGE_PRESENT
))
372 kill_guest(lg
, "Bad address %#lx", vaddr
);
374 gpte
= lgread(lg
, gpte_addr(lg
, gpgd
, vaddr
), pte_t
);
375 if (!(pte_flags(gpte
) & _PAGE_PRESENT
))
376 kill_guest(lg
, "Bad address %#lx", vaddr
);
378 return pte_pfn(gpte
) * PAGE_SIZE
| (vaddr
& ~PAGE_MASK
);
381 /* We keep several page tables. This is a simple routine to find the page
382 * table (if any) corresponding to this top-level address the Guest has given
384 static unsigned int find_pgdir(struct lguest
*lg
, unsigned long pgtable
)
387 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
388 if (lg
->pgdirs
[i
].gpgdir
== pgtable
)
393 /*H:435 And this is us, creating the new page directory. If we really do
394 * allocate a new one (and so the kernel parts are not there), we set
396 static unsigned int new_pgdir(struct lguest
*lg
,
397 unsigned long gpgdir
,
402 /* We pick one entry at random to throw out. Choosing the Least
403 * Recently Used might be better, but this is easy. */
404 next
= random32() % ARRAY_SIZE(lg
->pgdirs
);
405 /* If it's never been allocated at all before, try now. */
406 if (!lg
->pgdirs
[next
].pgdir
) {
407 lg
->pgdirs
[next
].pgdir
= (pgd_t
*)get_zeroed_page(GFP_KERNEL
);
408 /* If the allocation fails, just keep using the one we have */
409 if (!lg
->pgdirs
[next
].pgdir
)
412 /* This is a blank page, so there are no kernel
413 * mappings: caller must map the stack! */
416 /* Record which Guest toplevel this shadows. */
417 lg
->pgdirs
[next
].gpgdir
= gpgdir
;
418 /* Release all the non-kernel mappings. */
419 flush_user_mappings(lg
, next
);
424 /*H:430 (iv) Switching page tables
426 * This is what happens when the Guest changes page tables (ie. changes the
427 * top-level pgdir). This happens on almost every context switch. */
428 void guest_new_pagetable(struct lguest
*lg
, unsigned long pgtable
)
430 int newpgdir
, repin
= 0;
432 /* Look to see if we have this one already. */
433 newpgdir
= find_pgdir(lg
, pgtable
);
434 /* If not, we allocate or mug an existing one: if it's a fresh one,
435 * repin gets set to 1. */
436 if (newpgdir
== ARRAY_SIZE(lg
->pgdirs
))
437 newpgdir
= new_pgdir(lg
, pgtable
, &repin
);
438 /* Change the current pgd index to the new one. */
439 lg
->pgdidx
= newpgdir
;
440 /* If it was completely blank, we map in the Guest kernel stack */
445 /*H:470 Finally, a routine which throws away everything: all PGD entries in all
446 * the shadow page tables. This is used when we destroy the Guest. */
447 static void release_all_pagetables(struct lguest
*lg
)
451 /* Every shadow pagetable this Guest has */
452 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
453 if (lg
->pgdirs
[i
].pgdir
)
454 /* Every PGD entry except the Switcher at the top */
455 for (j
= 0; j
< SWITCHER_PGD_INDEX
; j
++)
456 release_pgd(lg
, lg
->pgdirs
[i
].pgdir
+ j
);
459 /* We also throw away everything when a Guest tells us it's changed a kernel
460 * mapping. Since kernel mappings are in every page table, it's easiest to
461 * throw them all away. This is amazingly slow, but thankfully rare. */
462 void guest_pagetable_clear_all(struct lguest
*lg
)
464 release_all_pagetables(lg
);
465 /* We need the Guest kernel stack mapped again. */
469 /*H:420 This is the routine which actually sets the page table entry for then
470 * "idx"'th shadow page table.
472 * Normally, we can just throw out the old entry and replace it with 0: if they
473 * use it demand_page() will put the new entry in. We need to do this anyway:
474 * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page
475 * is read from, and _PAGE_DIRTY when it's written to.
477 * But Avi Kivity pointed out that most Operating Systems (Linux included) set
478 * these bits on PTEs immediately anyway. This is done to save the CPU from
479 * having to update them, but it helps us the same way: if they set
480 * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if
481 * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately.
483 static void do_set_pte(struct lguest
*lg
, int idx
,
484 unsigned long vaddr
, pte_t gpte
)
486 /* Look up the matching shadow page directot entry. */
487 pgd_t
*spgd
= spgd_addr(lg
, idx
, vaddr
);
489 /* If the top level isn't present, there's no entry to update. */
490 if (pgd_flags(*spgd
) & _PAGE_PRESENT
) {
491 /* Otherwise, we start by releasing the existing entry. */
492 pte_t
*spte
= spte_addr(lg
, *spgd
, vaddr
);
495 /* If they're setting this entry as dirty or accessed, we might
496 * as well put that entry they've given us in now. This shaves
497 * 10% off a copy-on-write micro-benchmark. */
498 if (pte_flags(gpte
) & (_PAGE_DIRTY
| _PAGE_ACCESSED
)) {
499 check_gpte(lg
, gpte
);
500 *spte
= gpte_to_spte(lg
, gpte
,
501 pte_flags(gpte
) & _PAGE_DIRTY
);
503 /* Otherwise we can demand_page() it in later. */
508 /*H:410 Updating a PTE entry is a little trickier.
510 * We keep track of several different page tables (the Guest uses one for each
511 * process, so it makes sense to cache at least a few). Each of these have
512 * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for
513 * all processes. So when the page table above that address changes, we update
514 * all the page tables, not just the current one. This is rare.
516 * The benefit is that when we have to track a new page table, we can copy keep
517 * all the kernel mappings. This speeds up context switch immensely. */
518 void guest_set_pte(struct lguest
*lg
,
519 unsigned long gpgdir
, unsigned long vaddr
, pte_t gpte
)
521 /* Kernel mappings must be changed on all top levels. Slow, but
522 * doesn't happen often. */
523 if (vaddr
>= lg
->kernel_address
) {
525 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
526 if (lg
->pgdirs
[i
].pgdir
)
527 do_set_pte(lg
, i
, vaddr
, gpte
);
529 /* Is this page table one we have a shadow for? */
530 int pgdir
= find_pgdir(lg
, gpgdir
);
531 if (pgdir
!= ARRAY_SIZE(lg
->pgdirs
))
532 /* If so, do the update. */
533 do_set_pte(lg
, pgdir
, vaddr
, gpte
);
538 * (iii) Setting up a page table entry when the Guest tells us it has changed.
540 * Just like we did in interrupts_and_traps.c, it makes sense for us to deal
541 * with the other side of page tables while we're here: what happens when the
542 * Guest asks for a page table to be updated?
544 * We already saw that demand_page() will fill in the shadow page tables when
545 * needed, so we can simply remove shadow page table entries whenever the Guest
546 * tells us they've changed. When the Guest tries to use the new entry it will
547 * fault and demand_page() will fix it up.
549 * So with that in mind here's our code to to update a (top-level) PGD entry:
551 void guest_set_pmd(struct lguest
*lg
, unsigned long gpgdir
, u32 idx
)
555 /* The kernel seems to try to initialize this early on: we ignore its
556 * attempts to map over the Switcher. */
557 if (idx
>= SWITCHER_PGD_INDEX
)
560 /* If they're talking about a page table we have a shadow for... */
561 pgdir
= find_pgdir(lg
, gpgdir
);
562 if (pgdir
< ARRAY_SIZE(lg
->pgdirs
))
563 /* ... throw it away. */
564 release_pgd(lg
, lg
->pgdirs
[pgdir
].pgdir
+ idx
);
567 /*H:500 (vii) Setting up the page tables initially.
569 * When a Guest is first created, the Launcher tells us where the toplevel of
570 * its first page table is. We set some things up here: */
571 int init_guest_pagetable(struct lguest
*lg
, unsigned long pgtable
)
573 /* We start on the first shadow page table, and give it a blank PGD
576 lg
->pgdirs
[lg
->pgdidx
].gpgdir
= pgtable
;
577 lg
->pgdirs
[lg
->pgdidx
].pgdir
= (pgd_t
*)get_zeroed_page(GFP_KERNEL
);
578 if (!lg
->pgdirs
[lg
->pgdidx
].pgdir
)
583 /* When the Guest calls LHCALL_LGUEST_INIT we do more setup. */
584 void page_table_guest_data_init(struct lguest
*lg
)
586 /* We get the kernel address: above this is all kernel memory. */
587 if (get_user(lg
->kernel_address
, &lg
->lguest_data
->kernel_address
)
588 /* We tell the Guest that it can't use the top 4MB of virtual
589 * addresses used by the Switcher. */
590 || put_user(4U*1024*1024, &lg
->lguest_data
->reserve_mem
)
591 || put_user(lg
->pgdirs
[lg
->pgdidx
].gpgdir
,&lg
->lguest_data
->pgdir
))
592 kill_guest(lg
, "bad guest page %p", lg
->lguest_data
);
594 /* In flush_user_mappings() we loop from 0 to
595 * "pgd_index(lg->kernel_address)". This assumes it won't hit the
596 * Switcher mappings, so check that now. */
597 if (pgd_index(lg
->kernel_address
) >= SWITCHER_PGD_INDEX
)
598 kill_guest(lg
, "bad kernel address %#lx", lg
->kernel_address
);
601 /* When a Guest dies, our cleanup is fairly simple. */
602 void free_guest_pagetable(struct lguest
*lg
)
606 /* Throw away all page table pages. */
607 release_all_pagetables(lg
);
608 /* Now free the top levels: free_page() can handle 0 just fine. */
609 for (i
= 0; i
< ARRAY_SIZE(lg
->pgdirs
); i
++)
610 free_page((long)lg
->pgdirs
[i
].pgdir
);
613 /*H:480 (vi) Mapping the Switcher when the Guest is about to run.
615 * The Switcher and the two pages for this CPU need to be available to the
616 * Guest (and not the pages for other CPUs). We have the appropriate PTE pages
617 * for each CPU already set up, we just need to hook them in. */
618 void map_switcher_in_guest(struct lguest
*lg
, struct lguest_pages
*pages
)
620 pte_t
*switcher_pte_page
= __get_cpu_var(switcher_pte_pages
);
624 /* Make the last PGD entry for this Guest point to the Switcher's PTE
625 * page for this CPU (with appropriate flags). */
626 switcher_pgd
= __pgd(__pa(switcher_pte_page
) | _PAGE_KERNEL
);
628 lg
->pgdirs
[lg
->pgdidx
].pgdir
[SWITCHER_PGD_INDEX
] = switcher_pgd
;
630 /* We also change the Switcher PTE page. When we're running the Guest,
631 * we want the Guest's "regs" page to appear where the first Switcher
632 * page for this CPU is. This is an optimization: when the Switcher
633 * saves the Guest registers, it saves them into the first page of this
634 * CPU's "struct lguest_pages": if we make sure the Guest's register
635 * page is already mapped there, we don't have to copy them out
637 regs_pte
= pfn_pte (__pa(lg
->regs_page
) >> PAGE_SHIFT
, __pgprot(_PAGE_KERNEL
));
638 switcher_pte_page
[(unsigned long)pages
/PAGE_SIZE
%PTRS_PER_PTE
] = regs_pte
;
642 static void free_switcher_pte_pages(void)
646 for_each_possible_cpu(i
)
647 free_page((long)switcher_pte_page(i
));
650 /*H:520 Setting up the Switcher PTE page for given CPU is fairly easy, given
651 * the CPU number and the "struct page"s for the Switcher code itself.
653 * Currently the Switcher is less than a page long, so "pages" is always 1. */
654 static __init
void populate_switcher_pte_page(unsigned int cpu
,
655 struct page
*switcher_page
[],
659 pte_t
*pte
= switcher_pte_page(cpu
);
661 /* The first entries are easy: they map the Switcher code. */
662 for (i
= 0; i
< pages
; i
++) {
663 pte
[i
] = mk_pte(switcher_page
[i
],
664 __pgprot(_PAGE_PRESENT
|_PAGE_ACCESSED
));
667 /* The only other thing we map is this CPU's pair of pages. */
670 /* First page (Guest registers) is writable from the Guest */
671 pte
[i
] = pfn_pte(page_to_pfn(switcher_page
[i
]),
672 __pgprot(_PAGE_PRESENT
|_PAGE_ACCESSED
|_PAGE_RW
));
674 /* The second page contains the "struct lguest_ro_state", and is
676 pte
[i
+1] = pfn_pte(page_to_pfn(switcher_page
[i
+1]),
677 __pgprot(_PAGE_PRESENT
|_PAGE_ACCESSED
));
680 /*H:510 At boot or module load time, init_pagetables() allocates and populates
681 * the Switcher PTE page for each CPU. */
682 __init
int init_pagetables(struct page
**switcher_page
, unsigned int pages
)
686 for_each_possible_cpu(i
) {
687 switcher_pte_page(i
) = (pte_t
*)get_zeroed_page(GFP_KERNEL
);
688 if (!switcher_pte_page(i
)) {
689 free_switcher_pte_pages();
692 populate_switcher_pte_page(i
, switcher_page
, pages
);
698 /* Cleaning up simply involves freeing the PTE page for each CPU. */
699 void free_pagetables(void)
701 free_switcher_pte_pages();