Commit | Line | Data |
---|---|---|
f938d2c8 RR |
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 | |
a6bd8e13 RR |
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. :*/ | |
f938d2c8 RR |
7 | |
8 | /* Copyright (C) Rusty Russell IBM Corporation 2006. | |
d7e28ffe RR |
9 | * GPL v2 and any later version */ |
10 | #include <linux/mm.h> | |
11 | #include <linux/types.h> | |
12 | #include <linux/spinlock.h> | |
13 | #include <linux/random.h> | |
14 | #include <linux/percpu.h> | |
15 | #include <asm/tlbflush.h> | |
47436aa4 | 16 | #include <asm/uaccess.h> |
58a24566 | 17 | #include <asm/bootparam.h> |
d7e28ffe RR |
18 | #include "lg.h" |
19 | ||
f56a384e RR |
20 | /*M:008 We hold reference to pages, which prevents them from being swapped. |
21 | * It'd be nice to have a callback in the "struct mm_struct" when Linux wants | |
22 | * to swap out. If we had this, and a shrinker callback to trim PTE pages, we | |
23 | * could probably consider launching Guests as non-root. :*/ | |
24 | ||
bff672e6 RR |
25 | /*H:300 |
26 | * The Page Table Code | |
27 | * | |
28 | * We use two-level page tables for the Guest. If you're not entirely | |
29 | * comfortable with virtual addresses, physical addresses and page tables then | |
e1e72965 RR |
30 | * I recommend you review arch/x86/lguest/boot.c's "Page Table Handling" (with |
31 | * diagrams!). | |
bff672e6 RR |
32 | * |
33 | * The Guest keeps page tables, but we maintain the actual ones here: these are | |
34 | * called "shadow" page tables. Which is a very Guest-centric name: these are | |
35 | * the real page tables the CPU uses, although we keep them up to date to | |
36 | * reflect the Guest's. (See what I mean about weird naming? Since when do | |
37 | * shadows reflect anything?) | |
38 | * | |
39 | * Anyway, this is the most complicated part of the Host code. There are seven | |
40 | * parts to this: | |
e1e72965 RR |
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, | |
bff672e6 | 44 | * (iv) Switching page tables, |
e1e72965 | 45 | * (v) Flushing (throwing away) page tables, |
bff672e6 RR |
46 | * (vi) Mapping the Switcher when the Guest is about to run, |
47 | * (vii) Setting up the page tables initially. | |
48 | :*/ | |
49 | ||
bff672e6 RR |
50 | |
51 | /* 1024 entries in a page table page maps 1024 pages: 4MB. The Switcher is | |
52 | * conveniently placed at the top 4MB, so it uses a separate, complete PTE | |
53 | * page. */ | |
df29f43e | 54 | #define SWITCHER_PGD_INDEX (PTRS_PER_PGD - 1) |
d7e28ffe | 55 | |
acdd0b62 MZ |
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. */ | |
58 | #ifdef CONFIG_X86_PAE | |
59 | #define SWITCHER_PMD_INDEX (PTRS_PER_PMD - 1) | |
60 | #define RESERVE_MEM 2U | |
61 | #define CHECK_GPGD_MASK _PAGE_PRESENT | |
62 | #else | |
63 | #define RESERVE_MEM 4U | |
64 | #define CHECK_GPGD_MASK _PAGE_TABLE | |
65 | #endif | |
66 | ||
bff672e6 RR |
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. */ | |
df29f43e | 70 | static DEFINE_PER_CPU(pte_t *, switcher_pte_pages); |
d7e28ffe RR |
71 | #define switcher_pte_page(cpu) per_cpu(switcher_pte_pages, cpu) |
72 | ||
e1e72965 RR |
73 | /*H:320 The page table code is curly enough to need helper functions to keep it |
74 | * clear and clean. | |
bff672e6 | 75 | * |
df29f43e | 76 | * There are two functions which return pointers to the shadow (aka "real") |
bff672e6 RR |
77 | * page tables. |
78 | * | |
79 | * spgd_addr() takes the virtual address and returns a pointer to the top-level | |
e1e72965 RR |
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 | |
bff672e6 | 82 | * usually the current one). */ |
382ac6b3 | 83 | static pgd_t *spgd_addr(struct lg_cpu *cpu, u32 i, unsigned long vaddr) |
d7e28ffe | 84 | { |
df29f43e | 85 | unsigned int index = pgd_index(vaddr); |
d7e28ffe | 86 | |
acdd0b62 | 87 | #ifndef CONFIG_X86_PAE |
bff672e6 | 88 | /* We kill any Guest trying to touch the Switcher addresses. */ |
d7e28ffe | 89 | if (index >= SWITCHER_PGD_INDEX) { |
382ac6b3 | 90 | kill_guest(cpu, "attempt to access switcher pages"); |
d7e28ffe RR |
91 | index = 0; |
92 | } | |
acdd0b62 | 93 | #endif |
bff672e6 | 94 | /* Return a pointer index'th pgd entry for the i'th page table. */ |
382ac6b3 | 95 | return &cpu->lg->pgdirs[i].pgdir[index]; |
d7e28ffe RR |
96 | } |
97 | ||
acdd0b62 MZ |
98 | #ifdef CONFIG_X86_PAE |
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 | |
101 | * given address. */ | |
102 | static pmd_t *spmd_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr) | |
103 | { | |
104 | unsigned int index = pmd_index(vaddr); | |
105 | pmd_t *page; | |
106 | ||
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"); | |
111 | index = 0; | |
112 | } | |
113 | ||
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); | |
117 | ||
118 | return &page[index]; | |
119 | } | |
120 | #endif | |
121 | ||
e1e72965 RR |
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. */ | |
acdd0b62 | 125 | static pte_t *spte_addr(struct lg_cpu *cpu, pgd_t spgd, unsigned long vaddr) |
d7e28ffe | 126 | { |
acdd0b62 MZ |
127 | #ifdef CONFIG_X86_PAE |
128 | pmd_t *pmd = spmd_addr(cpu, spgd, vaddr); | |
129 | pte_t *page = __va(pmd_pfn(*pmd) << PAGE_SHIFT); | |
130 | ||
131 | /* You should never call this if the PMD entry wasn't valid */ | |
132 | BUG_ON(!(pmd_flags(*pmd) & _PAGE_PRESENT)); | |
133 | #else | |
df29f43e | 134 | pte_t *page = __va(pgd_pfn(spgd) << PAGE_SHIFT); |
bff672e6 | 135 | /* You should never call this if the PGD entry wasn't valid */ |
df29f43e | 136 | BUG_ON(!(pgd_flags(spgd) & _PAGE_PRESENT)); |
acdd0b62 MZ |
137 | #endif |
138 | ||
90603d15 | 139 | return &page[pte_index(vaddr)]; |
d7e28ffe RR |
140 | } |
141 | ||
bff672e6 RR |
142 | /* These two functions just like the above two, except they access the Guest |
143 | * page tables. Hence they return a Guest address. */ | |
1713608f | 144 | static unsigned long gpgd_addr(struct lg_cpu *cpu, unsigned long vaddr) |
d7e28ffe | 145 | { |
df29f43e | 146 | unsigned int index = vaddr >> (PGDIR_SHIFT); |
1713608f | 147 | return cpu->lg->pgdirs[cpu->cpu_pgd].gpgdir + index * sizeof(pgd_t); |
d7e28ffe RR |
148 | } |
149 | ||
acdd0b62 MZ |
150 | #ifdef CONFIG_X86_PAE |
151 | static unsigned long gpmd_addr(pgd_t gpgd, unsigned long vaddr) | |
d7e28ffe | 152 | { |
df29f43e MZ |
153 | unsigned long gpage = pgd_pfn(gpgd) << PAGE_SHIFT; |
154 | BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT)); | |
acdd0b62 MZ |
155 | return gpage + pmd_index(vaddr) * sizeof(pmd_t); |
156 | } | |
157 | #endif | |
158 | ||
159 | static unsigned long gpte_addr(struct lg_cpu *cpu, | |
160 | pgd_t gpgd, unsigned long vaddr) | |
161 | { | |
162 | #ifdef CONFIG_X86_PAE | |
163 | pmd_t gpmd; | |
164 | #endif | |
165 | unsigned long gpage; | |
166 | ||
167 | BUG_ON(!(pgd_flags(gpgd) & _PAGE_PRESENT)); | |
168 | #ifdef CONFIG_X86_PAE | |
169 | gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t); | |
170 | gpage = pmd_pfn(gpmd) << PAGE_SHIFT; | |
171 | BUG_ON(!(pmd_flags(gpmd) & _PAGE_PRESENT)); | |
172 | #else | |
173 | gpage = pgd_pfn(gpgd) << PAGE_SHIFT; | |
174 | #endif | |
90603d15 | 175 | return gpage + pte_index(vaddr) * sizeof(pte_t); |
d7e28ffe | 176 | } |
a6bd8e13 RR |
177 | /*:*/ |
178 | ||
71a3f4ed RR |
179 | /*M:014 get_pfn is slow: we could probably try to grab batches of pages here as |
180 | * an optimization (ie. pre-faulting). :*/ | |
d7e28ffe | 181 | |
bff672e6 RR |
182 | /*H:350 This routine takes a page number given by the Guest and converts it to |
183 | * an actual, physical page number. It can fail for several reasons: the | |
184 | * virtual address might not be mapped by the Launcher, the write flag is set | |
185 | * and the page is read-only, or the write flag was set and the page was | |
186 | * shared so had to be copied, but we ran out of memory. | |
187 | * | |
a6bd8e13 RR |
188 | * This holds a reference to the page, so release_pte() is careful to put that |
189 | * back. */ | |
d7e28ffe RR |
190 | static unsigned long get_pfn(unsigned long virtpfn, int write) |
191 | { | |
192 | struct page *page; | |
71a3f4ed RR |
193 | |
194 | /* gup me one page at this address please! */ | |
195 | if (get_user_pages_fast(virtpfn << PAGE_SHIFT, 1, write, &page) == 1) | |
196 | return page_to_pfn(page); | |
197 | ||
bff672e6 | 198 | /* This value indicates failure. */ |
71a3f4ed | 199 | return -1UL; |
d7e28ffe RR |
200 | } |
201 | ||
bff672e6 RR |
202 | /*H:340 Converting a Guest page table entry to a shadow (ie. real) page table |
203 | * entry can be a little tricky. The flags are (almost) the same, but the | |
204 | * Guest PTE contains a virtual page number: the CPU needs the real page | |
205 | * number. */ | |
382ac6b3 | 206 | static pte_t gpte_to_spte(struct lg_cpu *cpu, pte_t gpte, int write) |
d7e28ffe | 207 | { |
df29f43e | 208 | unsigned long pfn, base, flags; |
d7e28ffe | 209 | |
bff672e6 RR |
210 | /* The Guest sets the global flag, because it thinks that it is using |
211 | * PGE. We only told it to use PGE so it would tell us whether it was | |
212 | * flushing a kernel mapping or a userspace mapping. We don't actually | |
213 | * use the global bit, so throw it away. */ | |
df29f43e | 214 | flags = (pte_flags(gpte) & ~_PAGE_GLOBAL); |
bff672e6 | 215 | |
3c6b5bfa | 216 | /* The Guest's pages are offset inside the Launcher. */ |
382ac6b3 | 217 | base = (unsigned long)cpu->lg->mem_base / PAGE_SIZE; |
3c6b5bfa | 218 | |
bff672e6 RR |
219 | /* We need a temporary "unsigned long" variable to hold the answer from |
220 | * get_pfn(), because it returns 0xFFFFFFFF on failure, which wouldn't | |
221 | * fit in spte.pfn. get_pfn() finds the real physical number of the | |
222 | * page, given the virtual number. */ | |
df29f43e | 223 | pfn = get_pfn(base + pte_pfn(gpte), write); |
d7e28ffe | 224 | if (pfn == -1UL) { |
382ac6b3 | 225 | kill_guest(cpu, "failed to get page %lu", pte_pfn(gpte)); |
bff672e6 RR |
226 | /* When we destroy the Guest, we'll go through the shadow page |
227 | * tables and release_pte() them. Make sure we don't think | |
228 | * this one is valid! */ | |
df29f43e | 229 | flags = 0; |
d7e28ffe | 230 | } |
df29f43e MZ |
231 | /* Now we assemble our shadow PTE from the page number and flags. */ |
232 | return pfn_pte(pfn, __pgprot(flags)); | |
d7e28ffe RR |
233 | } |
234 | ||
bff672e6 | 235 | /*H:460 And to complete the chain, release_pte() looks like this: */ |
df29f43e | 236 | static void release_pte(pte_t pte) |
d7e28ffe | 237 | { |
71a3f4ed | 238 | /* Remember that get_user_pages_fast() took a reference to the page, in |
bff672e6 | 239 | * get_pfn()? We have to put it back now. */ |
df29f43e | 240 | if (pte_flags(pte) & _PAGE_PRESENT) |
90603d15 | 241 | put_page(pte_page(pte)); |
d7e28ffe | 242 | } |
bff672e6 | 243 | /*:*/ |
d7e28ffe | 244 | |
382ac6b3 | 245 | static void check_gpte(struct lg_cpu *cpu, pte_t gpte) |
d7e28ffe | 246 | { |
31f4b46e AD |
247 | if ((pte_flags(gpte) & _PAGE_PSE) || |
248 | pte_pfn(gpte) >= cpu->lg->pfn_limit) | |
382ac6b3 | 249 | kill_guest(cpu, "bad page table entry"); |
d7e28ffe RR |
250 | } |
251 | ||
382ac6b3 | 252 | static void check_gpgd(struct lg_cpu *cpu, pgd_t gpgd) |
d7e28ffe | 253 | { |
acdd0b62 | 254 | if ((pgd_flags(gpgd) & ~CHECK_GPGD_MASK) || |
382ac6b3 GOC |
255 | (pgd_pfn(gpgd) >= cpu->lg->pfn_limit)) |
256 | kill_guest(cpu, "bad page directory entry"); | |
d7e28ffe RR |
257 | } |
258 | ||
acdd0b62 MZ |
259 | #ifdef CONFIG_X86_PAE |
260 | static void check_gpmd(struct lg_cpu *cpu, pmd_t gpmd) | |
261 | { | |
262 | if ((pmd_flags(gpmd) & ~_PAGE_TABLE) || | |
263 | (pmd_pfn(gpmd) >= cpu->lg->pfn_limit)) | |
264 | kill_guest(cpu, "bad page middle directory entry"); | |
265 | } | |
266 | #endif | |
267 | ||
bff672e6 | 268 | /*H:330 |
e1e72965 | 269 | * (i) Looking up a page table entry when the Guest faults. |
bff672e6 RR |
270 | * |
271 | * We saw this call in run_guest(): when we see a page fault in the Guest, we | |
272 | * come here. That's because we only set up the shadow page tables lazily as | |
273 | * they're needed, so we get page faults all the time and quietly fix them up | |
274 | * and return to the Guest without it knowing. | |
275 | * | |
276 | * If we fixed up the fault (ie. we mapped the address), this routine returns | |
e1e72965 | 277 | * true. Otherwise, it was a real fault and we need to tell the Guest. */ |
df1693ab | 278 | bool demand_page(struct lg_cpu *cpu, unsigned long vaddr, int errcode) |
d7e28ffe | 279 | { |
df29f43e MZ |
280 | pgd_t gpgd; |
281 | pgd_t *spgd; | |
d7e28ffe | 282 | unsigned long gpte_ptr; |
df29f43e MZ |
283 | pte_t gpte; |
284 | pte_t *spte; | |
d7e28ffe | 285 | |
acdd0b62 MZ |
286 | #ifdef CONFIG_X86_PAE |
287 | pmd_t *spmd; | |
288 | pmd_t gpmd; | |
289 | #endif | |
290 | ||
bff672e6 | 291 | /* First step: get the top-level Guest page table entry. */ |
382ac6b3 | 292 | gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t); |
bff672e6 | 293 | /* Toplevel not present? We can't map it in. */ |
df29f43e | 294 | if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) |
df1693ab | 295 | return false; |
d7e28ffe | 296 | |
bff672e6 | 297 | /* Now look at the matching shadow entry. */ |
382ac6b3 | 298 | spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr); |
df29f43e | 299 | if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) { |
bff672e6 | 300 | /* No shadow entry: allocate a new shadow PTE page. */ |
d7e28ffe | 301 | unsigned long ptepage = get_zeroed_page(GFP_KERNEL); |
bff672e6 RR |
302 | /* This is not really the Guest's fault, but killing it is |
303 | * simple for this corner case. */ | |
d7e28ffe | 304 | if (!ptepage) { |
382ac6b3 | 305 | kill_guest(cpu, "out of memory allocating pte page"); |
df1693ab | 306 | return false; |
d7e28ffe | 307 | } |
bff672e6 | 308 | /* We check that the Guest pgd is OK. */ |
382ac6b3 | 309 | check_gpgd(cpu, gpgd); |
bff672e6 RR |
310 | /* And we copy the flags to the shadow PGD entry. The page |
311 | * number in the shadow PGD is the page we just allocated. */ | |
acdd0b62 | 312 | set_pgd(spgd, __pgd(__pa(ptepage) | pgd_flags(gpgd))); |
d7e28ffe RR |
313 | } |
314 | ||
acdd0b62 MZ |
315 | #ifdef CONFIG_X86_PAE |
316 | gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t); | |
317 | /* middle level not present? We can't map it in. */ | |
318 | if (!(pmd_flags(gpmd) & _PAGE_PRESENT)) | |
319 | return false; | |
320 | ||
321 | /* Now look at the matching shadow entry. */ | |
322 | spmd = spmd_addr(cpu, *spgd, vaddr); | |
323 | ||
324 | if (!(pmd_flags(*spmd) & _PAGE_PRESENT)) { | |
325 | /* No shadow entry: allocate a new shadow PTE page. */ | |
326 | unsigned long ptepage = get_zeroed_page(GFP_KERNEL); | |
327 | ||
328 | /* This is not really the Guest's fault, but killing it is | |
329 | * simple for this corner case. */ | |
330 | if (!ptepage) { | |
331 | kill_guest(cpu, "out of memory allocating pte page"); | |
332 | return false; | |
333 | } | |
334 | ||
335 | /* We check that the Guest pmd is OK. */ | |
336 | check_gpmd(cpu, gpmd); | |
337 | ||
338 | /* And we copy the flags to the shadow PMD entry. The page | |
339 | * number in the shadow PMD is the page we just allocated. */ | |
340 | native_set_pmd(spmd, __pmd(__pa(ptepage) | pmd_flags(gpmd))); | |
341 | } | |
342 | #endif | |
bff672e6 RR |
343 | /* OK, now we look at the lower level in the Guest page table: keep its |
344 | * address, because we might update it later. */ | |
acdd0b62 | 345 | gpte_ptr = gpte_addr(cpu, gpgd, vaddr); |
382ac6b3 | 346 | gpte = lgread(cpu, gpte_ptr, pte_t); |
d7e28ffe | 347 | |
bff672e6 | 348 | /* If this page isn't in the Guest page tables, we can't page it in. */ |
df29f43e | 349 | if (!(pte_flags(gpte) & _PAGE_PRESENT)) |
df1693ab | 350 | return false; |
d7e28ffe | 351 | |
bff672e6 RR |
352 | /* Check they're not trying to write to a page the Guest wants |
353 | * read-only (bit 2 of errcode == write). */ | |
df29f43e | 354 | if ((errcode & 2) && !(pte_flags(gpte) & _PAGE_RW)) |
df1693ab | 355 | return false; |
d7e28ffe | 356 | |
e1e72965 | 357 | /* User access to a kernel-only page? (bit 3 == user access) */ |
df29f43e | 358 | if ((errcode & 4) && !(pte_flags(gpte) & _PAGE_USER)) |
df1693ab | 359 | return false; |
d7e28ffe | 360 | |
bff672e6 RR |
361 | /* Check that the Guest PTE flags are OK, and the page number is below |
362 | * the pfn_limit (ie. not mapping the Launcher binary). */ | |
382ac6b3 | 363 | check_gpte(cpu, gpte); |
e1e72965 | 364 | |
bff672e6 | 365 | /* Add the _PAGE_ACCESSED and (for a write) _PAGE_DIRTY flag */ |
df29f43e | 366 | gpte = pte_mkyoung(gpte); |
d7e28ffe | 367 | if (errcode & 2) |
df29f43e | 368 | gpte = pte_mkdirty(gpte); |
d7e28ffe | 369 | |
bff672e6 | 370 | /* Get the pointer to the shadow PTE entry we're going to set. */ |
acdd0b62 | 371 | spte = spte_addr(cpu, *spgd, vaddr); |
bff672e6 RR |
372 | /* If there was a valid shadow PTE entry here before, we release it. |
373 | * This can happen with a write to a previously read-only entry. */ | |
d7e28ffe RR |
374 | release_pte(*spte); |
375 | ||
bff672e6 RR |
376 | /* If this is a write, we insist that the Guest page is writable (the |
377 | * final arg to gpte_to_spte()). */ | |
df29f43e | 378 | if (pte_dirty(gpte)) |
382ac6b3 | 379 | *spte = gpte_to_spte(cpu, gpte, 1); |
df29f43e | 380 | else |
bff672e6 RR |
381 | /* If this is a read, don't set the "writable" bit in the page |
382 | * table entry, even if the Guest says it's writable. That way | |
e1e72965 RR |
383 | * we will come back here when a write does actually occur, so |
384 | * we can update the Guest's _PAGE_DIRTY flag. */ | |
90603d15 | 385 | native_set_pte(spte, gpte_to_spte(cpu, pte_wrprotect(gpte), 0)); |
d7e28ffe | 386 | |
bff672e6 RR |
387 | /* Finally, we write the Guest PTE entry back: we've set the |
388 | * _PAGE_ACCESSED and maybe the _PAGE_DIRTY flags. */ | |
382ac6b3 | 389 | lgwrite(cpu, gpte_ptr, pte_t, gpte); |
bff672e6 | 390 | |
e1e72965 RR |
391 | /* The fault is fixed, the page table is populated, the mapping |
392 | * manipulated, the result returned and the code complete. A small | |
393 | * delay and a trace of alliteration are the only indications the Guest | |
394 | * has that a page fault occurred at all. */ | |
df1693ab | 395 | return true; |
d7e28ffe RR |
396 | } |
397 | ||
e1e72965 RR |
398 | /*H:360 |
399 | * (ii) Making sure the Guest stack is mapped. | |
bff672e6 | 400 | * |
e1e72965 RR |
401 | * Remember that direct traps into the Guest need a mapped Guest kernel stack. |
402 | * pin_stack_pages() calls us here: we could simply call demand_page(), but as | |
403 | * we've seen that logic is quite long, and usually the stack pages are already | |
404 | * mapped, so it's overkill. | |
bff672e6 RR |
405 | * |
406 | * This is a quick version which answers the question: is this virtual address | |
407 | * mapped by the shadow page tables, and is it writable? */ | |
df1693ab | 408 | static bool page_writable(struct lg_cpu *cpu, unsigned long vaddr) |
d7e28ffe | 409 | { |
df29f43e | 410 | pgd_t *spgd; |
d7e28ffe RR |
411 | unsigned long flags; |
412 | ||
acdd0b62 MZ |
413 | #ifdef CONFIG_X86_PAE |
414 | pmd_t *spmd; | |
415 | #endif | |
e1e72965 | 416 | /* Look at the current top level entry: is it present? */ |
382ac6b3 | 417 | spgd = spgd_addr(cpu, cpu->cpu_pgd, vaddr); |
df29f43e | 418 | if (!(pgd_flags(*spgd) & _PAGE_PRESENT)) |
df1693ab | 419 | return false; |
d7e28ffe | 420 | |
acdd0b62 MZ |
421 | #ifdef CONFIG_X86_PAE |
422 | spmd = spmd_addr(cpu, *spgd, vaddr); | |
423 | if (!(pmd_flags(*spmd) & _PAGE_PRESENT)) | |
424 | return false; | |
425 | #endif | |
426 | ||
bff672e6 RR |
427 | /* Check the flags on the pte entry itself: it must be present and |
428 | * writable. */ | |
acdd0b62 | 429 | flags = pte_flags(*(spte_addr(cpu, *spgd, vaddr))); |
df29f43e | 430 | |
d7e28ffe RR |
431 | return (flags & (_PAGE_PRESENT|_PAGE_RW)) == (_PAGE_PRESENT|_PAGE_RW); |
432 | } | |
433 | ||
bff672e6 RR |
434 | /* So, when pin_stack_pages() asks us to pin a page, we check if it's already |
435 | * in the page tables, and if not, we call demand_page() with error code 2 | |
436 | * (meaning "write"). */ | |
1713608f | 437 | void pin_page(struct lg_cpu *cpu, unsigned long vaddr) |
d7e28ffe | 438 | { |
1713608f | 439 | if (!page_writable(cpu, vaddr) && !demand_page(cpu, vaddr, 2)) |
382ac6b3 | 440 | kill_guest(cpu, "bad stack page %#lx", vaddr); |
d7e28ffe RR |
441 | } |
442 | ||
acdd0b62 MZ |
443 | #ifdef CONFIG_X86_PAE |
444 | static void release_pmd(pmd_t *spmd) | |
445 | { | |
446 | /* If the entry's not present, there's nothing to release. */ | |
447 | if (pmd_flags(*spmd) & _PAGE_PRESENT) { | |
448 | unsigned int i; | |
449 | pte_t *ptepage = __va(pmd_pfn(*spmd) << PAGE_SHIFT); | |
450 | /* For each entry in the page, we might need to release it. */ | |
451 | for (i = 0; i < PTRS_PER_PTE; i++) | |
452 | release_pte(ptepage[i]); | |
453 | /* Now we can free the page of PTEs */ | |
454 | free_page((long)ptepage); | |
455 | /* And zero out the PMD entry so we never release it twice. */ | |
456 | native_set_pmd(spmd, __pmd(0)); | |
457 | } | |
458 | } | |
459 | ||
460 | static void release_pgd(pgd_t *spgd) | |
461 | { | |
462 | /* If the entry's not present, there's nothing to release. */ | |
463 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { | |
464 | unsigned int i; | |
465 | pmd_t *pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); | |
466 | ||
467 | for (i = 0; i < PTRS_PER_PMD; i++) | |
468 | release_pmd(&pmdpage[i]); | |
469 | ||
470 | /* Now we can free the page of PMDs */ | |
471 | free_page((long)pmdpage); | |
472 | /* And zero out the PGD entry so we never release it twice. */ | |
473 | set_pgd(spgd, __pgd(0)); | |
474 | } | |
475 | } | |
476 | ||
477 | #else /* !CONFIG_X86_PAE */ | |
bff672e6 | 478 | /*H:450 If we chase down the release_pgd() code, it looks like this: */ |
90603d15 | 479 | static void release_pgd(pgd_t *spgd) |
d7e28ffe | 480 | { |
bff672e6 | 481 | /* If the entry's not present, there's nothing to release. */ |
df29f43e | 482 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { |
d7e28ffe | 483 | unsigned int i; |
bff672e6 RR |
484 | /* Converting the pfn to find the actual PTE page is easy: turn |
485 | * the page number into a physical address, then convert to a | |
486 | * virtual address (easy for kernel pages like this one). */ | |
df29f43e | 487 | pte_t *ptepage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); |
bff672e6 | 488 | /* For each entry in the page, we might need to release it. */ |
df29f43e | 489 | for (i = 0; i < PTRS_PER_PTE; i++) |
d7e28ffe | 490 | release_pte(ptepage[i]); |
bff672e6 | 491 | /* Now we can free the page of PTEs */ |
d7e28ffe | 492 | free_page((long)ptepage); |
e1e72965 | 493 | /* And zero out the PGD entry so we never release it twice. */ |
df29f43e | 494 | *spgd = __pgd(0); |
d7e28ffe RR |
495 | } |
496 | } | |
acdd0b62 | 497 | #endif |
e1e72965 RR |
498 | /*H:445 We saw flush_user_mappings() twice: once from the flush_user_mappings() |
499 | * hypercall and once in new_pgdir() when we re-used a top-level pgdir page. | |
500 | * It simply releases every PTE page from 0 up to the Guest's kernel address. */ | |
d7e28ffe RR |
501 | static void flush_user_mappings(struct lguest *lg, int idx) |
502 | { | |
503 | unsigned int i; | |
bff672e6 | 504 | /* Release every pgd entry up to the kernel's address. */ |
47436aa4 | 505 | for (i = 0; i < pgd_index(lg->kernel_address); i++) |
90603d15 | 506 | release_pgd(lg->pgdirs[idx].pgdir + i); |
d7e28ffe RR |
507 | } |
508 | ||
e1e72965 RR |
509 | /*H:440 (v) Flushing (throwing away) page tables, |
510 | * | |
511 | * The Guest has a hypercall to throw away the page tables: it's used when a | |
512 | * large number of mappings have been changed. */ | |
1713608f | 513 | void guest_pagetable_flush_user(struct lg_cpu *cpu) |
d7e28ffe | 514 | { |
bff672e6 | 515 | /* Drop the userspace part of the current page table. */ |
1713608f | 516 | flush_user_mappings(cpu->lg, cpu->cpu_pgd); |
d7e28ffe | 517 | } |
bff672e6 | 518 | /*:*/ |
d7e28ffe | 519 | |
47436aa4 | 520 | /* We walk down the guest page tables to get a guest-physical address */ |
1713608f | 521 | unsigned long guest_pa(struct lg_cpu *cpu, unsigned long vaddr) |
47436aa4 RR |
522 | { |
523 | pgd_t gpgd; | |
524 | pte_t gpte; | |
525 | ||
acdd0b62 MZ |
526 | #ifdef CONFIG_X86_PAE |
527 | pmd_t gpmd; | |
528 | #endif | |
47436aa4 | 529 | /* First step: get the top-level Guest page table entry. */ |
382ac6b3 | 530 | gpgd = lgread(cpu, gpgd_addr(cpu, vaddr), pgd_t); |
47436aa4 | 531 | /* Toplevel not present? We can't map it in. */ |
6afbdd05 | 532 | if (!(pgd_flags(gpgd) & _PAGE_PRESENT)) { |
382ac6b3 | 533 | kill_guest(cpu, "Bad address %#lx", vaddr); |
6afbdd05 RR |
534 | return -1UL; |
535 | } | |
47436aa4 | 536 | |
acdd0b62 MZ |
537 | gpte = lgread(cpu, gpte_addr(cpu, gpgd, vaddr), pte_t); |
538 | #ifdef CONFIG_X86_PAE | |
539 | gpmd = lgread(cpu, gpmd_addr(gpgd, vaddr), pmd_t); | |
540 | if (!(pmd_flags(gpmd) & _PAGE_PRESENT)) | |
541 | kill_guest(cpu, "Bad address %#lx", vaddr); | |
542 | #endif | |
543 | gpte = lgread(cpu, gpte_addr(cpu, gpgd, vaddr), pte_t); | |
47436aa4 | 544 | if (!(pte_flags(gpte) & _PAGE_PRESENT)) |
382ac6b3 | 545 | kill_guest(cpu, "Bad address %#lx", vaddr); |
47436aa4 RR |
546 | |
547 | return pte_pfn(gpte) * PAGE_SIZE | (vaddr & ~PAGE_MASK); | |
548 | } | |
549 | ||
bff672e6 RR |
550 | /* We keep several page tables. This is a simple routine to find the page |
551 | * table (if any) corresponding to this top-level address the Guest has given | |
552 | * us. */ | |
d7e28ffe RR |
553 | static unsigned int find_pgdir(struct lguest *lg, unsigned long pgtable) |
554 | { | |
555 | unsigned int i; | |
556 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) | |
4357bd94 | 557 | if (lg->pgdirs[i].pgdir && lg->pgdirs[i].gpgdir == pgtable) |
d7e28ffe RR |
558 | break; |
559 | return i; | |
560 | } | |
561 | ||
bff672e6 RR |
562 | /*H:435 And this is us, creating the new page directory. If we really do |
563 | * allocate a new one (and so the kernel parts are not there), we set | |
564 | * blank_pgdir. */ | |
1713608f | 565 | static unsigned int new_pgdir(struct lg_cpu *cpu, |
ee3db0f2 | 566 | unsigned long gpgdir, |
d7e28ffe RR |
567 | int *blank_pgdir) |
568 | { | |
569 | unsigned int next; | |
acdd0b62 MZ |
570 | #ifdef CONFIG_X86_PAE |
571 | pmd_t *pmd_table; | |
572 | #endif | |
d7e28ffe | 573 | |
bff672e6 RR |
574 | /* We pick one entry at random to throw out. Choosing the Least |
575 | * Recently Used might be better, but this is easy. */ | |
382ac6b3 | 576 | next = random32() % ARRAY_SIZE(cpu->lg->pgdirs); |
bff672e6 | 577 | /* If it's never been allocated at all before, try now. */ |
382ac6b3 GOC |
578 | if (!cpu->lg->pgdirs[next].pgdir) { |
579 | cpu->lg->pgdirs[next].pgdir = | |
580 | (pgd_t *)get_zeroed_page(GFP_KERNEL); | |
bff672e6 | 581 | /* If the allocation fails, just keep using the one we have */ |
382ac6b3 | 582 | if (!cpu->lg->pgdirs[next].pgdir) |
1713608f | 583 | next = cpu->cpu_pgd; |
acdd0b62 MZ |
584 | else { |
585 | #ifdef CONFIG_X86_PAE | |
586 | /* In PAE mode, allocate a pmd page and populate the | |
587 | * last pgd entry. */ | |
588 | pmd_table = (pmd_t *)get_zeroed_page(GFP_KERNEL); | |
589 | if (!pmd_table) { | |
590 | free_page((long)cpu->lg->pgdirs[next].pgdir); | |
591 | set_pgd(cpu->lg->pgdirs[next].pgdir, __pgd(0)); | |
592 | next = cpu->cpu_pgd; | |
593 | } else { | |
594 | set_pgd(cpu->lg->pgdirs[next].pgdir + | |
595 | SWITCHER_PGD_INDEX, | |
596 | __pgd(__pa(pmd_table) | _PAGE_PRESENT)); | |
597 | /* This is a blank page, so there are no kernel | |
598 | * mappings: caller must map the stack! */ | |
599 | *blank_pgdir = 1; | |
600 | } | |
601 | #else | |
d7e28ffe | 602 | *blank_pgdir = 1; |
acdd0b62 MZ |
603 | #endif |
604 | } | |
d7e28ffe | 605 | } |
bff672e6 | 606 | /* Record which Guest toplevel this shadows. */ |
382ac6b3 | 607 | cpu->lg->pgdirs[next].gpgdir = gpgdir; |
d7e28ffe | 608 | /* Release all the non-kernel mappings. */ |
382ac6b3 | 609 | flush_user_mappings(cpu->lg, next); |
d7e28ffe RR |
610 | |
611 | return next; | |
612 | } | |
613 | ||
bff672e6 RR |
614 | /*H:430 (iv) Switching page tables |
615 | * | |
90603d15 | 616 | * Now we've seen all the page table setting and manipulation, let's see |
e1e72965 RR |
617 | * what happens when the Guest changes page tables (ie. changes the top-level |
618 | * pgdir). This occurs on almost every context switch. */ | |
4665ac8e | 619 | void guest_new_pagetable(struct lg_cpu *cpu, unsigned long pgtable) |
d7e28ffe RR |
620 | { |
621 | int newpgdir, repin = 0; | |
622 | ||
bff672e6 | 623 | /* Look to see if we have this one already. */ |
382ac6b3 | 624 | newpgdir = find_pgdir(cpu->lg, pgtable); |
bff672e6 RR |
625 | /* If not, we allocate or mug an existing one: if it's a fresh one, |
626 | * repin gets set to 1. */ | |
382ac6b3 | 627 | if (newpgdir == ARRAY_SIZE(cpu->lg->pgdirs)) |
1713608f | 628 | newpgdir = new_pgdir(cpu, pgtable, &repin); |
bff672e6 | 629 | /* Change the current pgd index to the new one. */ |
1713608f | 630 | cpu->cpu_pgd = newpgdir; |
bff672e6 | 631 | /* If it was completely blank, we map in the Guest kernel stack */ |
d7e28ffe | 632 | if (repin) |
4665ac8e | 633 | pin_stack_pages(cpu); |
d7e28ffe RR |
634 | } |
635 | ||
bff672e6 | 636 | /*H:470 Finally, a routine which throws away everything: all PGD entries in all |
e1e72965 RR |
637 | * the shadow page tables, including the Guest's kernel mappings. This is used |
638 | * when we destroy the Guest. */ | |
d7e28ffe RR |
639 | static void release_all_pagetables(struct lguest *lg) |
640 | { | |
641 | unsigned int i, j; | |
642 | ||
bff672e6 | 643 | /* Every shadow pagetable this Guest has */ |
d7e28ffe | 644 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
acdd0b62 MZ |
645 | if (lg->pgdirs[i].pgdir) { |
646 | #ifdef CONFIG_X86_PAE | |
647 | pgd_t *spgd; | |
648 | pmd_t *pmdpage; | |
649 | unsigned int k; | |
650 | ||
651 | /* Get the last pmd page. */ | |
652 | spgd = lg->pgdirs[i].pgdir + SWITCHER_PGD_INDEX; | |
653 | pmdpage = __va(pgd_pfn(*spgd) << PAGE_SHIFT); | |
654 | ||
655 | /* And release the pmd entries of that pmd page, | |
656 | * except for the switcher pmd. */ | |
657 | for (k = 0; k < SWITCHER_PMD_INDEX; k++) | |
658 | release_pmd(&pmdpage[k]); | |
659 | #endif | |
bff672e6 | 660 | /* Every PGD entry except the Switcher at the top */ |
d7e28ffe | 661 | for (j = 0; j < SWITCHER_PGD_INDEX; j++) |
90603d15 | 662 | release_pgd(lg->pgdirs[i].pgdir + j); |
acdd0b62 | 663 | } |
d7e28ffe RR |
664 | } |
665 | ||
bff672e6 RR |
666 | /* We also throw away everything when a Guest tells us it's changed a kernel |
667 | * mapping. Since kernel mappings are in every page table, it's easiest to | |
e1e72965 RR |
668 | * throw them all away. This traps the Guest in amber for a while as |
669 | * everything faults back in, but it's rare. */ | |
4665ac8e | 670 | void guest_pagetable_clear_all(struct lg_cpu *cpu) |
d7e28ffe | 671 | { |
4665ac8e | 672 | release_all_pagetables(cpu->lg); |
bff672e6 | 673 | /* We need the Guest kernel stack mapped again. */ |
4665ac8e | 674 | pin_stack_pages(cpu); |
d7e28ffe | 675 | } |
e1e72965 RR |
676 | /*:*/ |
677 | /*M:009 Since we throw away all mappings when a kernel mapping changes, our | |
678 | * performance sucks for guests using highmem. In fact, a guest with | |
679 | * PAGE_OFFSET 0xc0000000 (the default) and more than about 700MB of RAM is | |
680 | * usually slower than a Guest with less memory. | |
681 | * | |
682 | * This, of course, cannot be fixed. It would take some kind of... well, I | |
683 | * don't know, but the term "puissant code-fu" comes to mind. :*/ | |
d7e28ffe | 684 | |
bff672e6 RR |
685 | /*H:420 This is the routine which actually sets the page table entry for then |
686 | * "idx"'th shadow page table. | |
687 | * | |
688 | * Normally, we can just throw out the old entry and replace it with 0: if they | |
689 | * use it demand_page() will put the new entry in. We need to do this anyway: | |
690 | * The Guest expects _PAGE_ACCESSED to be set on its PTE the first time a page | |
691 | * is read from, and _PAGE_DIRTY when it's written to. | |
692 | * | |
693 | * But Avi Kivity pointed out that most Operating Systems (Linux included) set | |
694 | * these bits on PTEs immediately anyway. This is done to save the CPU from | |
695 | * having to update them, but it helps us the same way: if they set | |
696 | * _PAGE_ACCESSED then we can put a read-only PTE entry in immediately, and if | |
697 | * they set _PAGE_DIRTY then we can put a writable PTE entry in immediately. | |
698 | */ | |
382ac6b3 | 699 | static void do_set_pte(struct lg_cpu *cpu, int idx, |
df29f43e | 700 | unsigned long vaddr, pte_t gpte) |
d7e28ffe | 701 | { |
e1e72965 | 702 | /* Look up the matching shadow page directory entry. */ |
382ac6b3 | 703 | pgd_t *spgd = spgd_addr(cpu, idx, vaddr); |
acdd0b62 MZ |
704 | #ifdef CONFIG_X86_PAE |
705 | pmd_t *spmd; | |
706 | #endif | |
bff672e6 RR |
707 | |
708 | /* If the top level isn't present, there's no entry to update. */ | |
df29f43e | 709 | if (pgd_flags(*spgd) & _PAGE_PRESENT) { |
acdd0b62 MZ |
710 | #ifdef CONFIG_X86_PAE |
711 | spmd = spmd_addr(cpu, *spgd, vaddr); | |
712 | if (pmd_flags(*spmd) & _PAGE_PRESENT) { | |
713 | #endif | |
714 | /* Otherwise, we start by releasing | |
715 | * the existing entry. */ | |
716 | pte_t *spte = spte_addr(cpu, *spgd, vaddr); | |
717 | release_pte(*spte); | |
718 | ||
719 | /* If they're setting this entry as dirty or accessed, | |
720 | * we might as well put that entry they've given us | |
721 | * in now. This shaves 10% off a | |
722 | * copy-on-write micro-benchmark. */ | |
723 | if (pte_flags(gpte) & (_PAGE_DIRTY | _PAGE_ACCESSED)) { | |
724 | check_gpte(cpu, gpte); | |
725 | native_set_pte(spte, | |
726 | gpte_to_spte(cpu, gpte, | |
727 | pte_flags(gpte) & _PAGE_DIRTY)); | |
728 | } else | |
729 | /* Otherwise kill it and we can demand_page() | |
730 | * it in later. */ | |
731 | native_set_pte(spte, __pte(0)); | |
732 | #ifdef CONFIG_X86_PAE | |
733 | } | |
734 | #endif | |
d7e28ffe RR |
735 | } |
736 | } | |
737 | ||
bff672e6 RR |
738 | /*H:410 Updating a PTE entry is a little trickier. |
739 | * | |
740 | * We keep track of several different page tables (the Guest uses one for each | |
741 | * process, so it makes sense to cache at least a few). Each of these have | |
742 | * identical kernel parts: ie. every mapping above PAGE_OFFSET is the same for | |
743 | * all processes. So when the page table above that address changes, we update | |
744 | * all the page tables, not just the current one. This is rare. | |
745 | * | |
a6bd8e13 RR |
746 | * The benefit is that when we have to track a new page table, we can keep all |
747 | * the kernel mappings. This speeds up context switch immensely. */ | |
382ac6b3 | 748 | void guest_set_pte(struct lg_cpu *cpu, |
ee3db0f2 | 749 | unsigned long gpgdir, unsigned long vaddr, pte_t gpte) |
d7e28ffe | 750 | { |
a6bd8e13 RR |
751 | /* Kernel mappings must be changed on all top levels. Slow, but doesn't |
752 | * happen often. */ | |
382ac6b3 | 753 | if (vaddr >= cpu->lg->kernel_address) { |
d7e28ffe | 754 | unsigned int i; |
382ac6b3 GOC |
755 | for (i = 0; i < ARRAY_SIZE(cpu->lg->pgdirs); i++) |
756 | if (cpu->lg->pgdirs[i].pgdir) | |
757 | do_set_pte(cpu, i, vaddr, gpte); | |
d7e28ffe | 758 | } else { |
bff672e6 | 759 | /* Is this page table one we have a shadow for? */ |
382ac6b3 GOC |
760 | int pgdir = find_pgdir(cpu->lg, gpgdir); |
761 | if (pgdir != ARRAY_SIZE(cpu->lg->pgdirs)) | |
bff672e6 | 762 | /* If so, do the update. */ |
382ac6b3 | 763 | do_set_pte(cpu, pgdir, vaddr, gpte); |
d7e28ffe RR |
764 | } |
765 | } | |
766 | ||
bff672e6 | 767 | /*H:400 |
e1e72965 | 768 | * (iii) Setting up a page table entry when the Guest tells us one has changed. |
bff672e6 RR |
769 | * |
770 | * Just like we did in interrupts_and_traps.c, it makes sense for us to deal | |
771 | * with the other side of page tables while we're here: what happens when the | |
772 | * Guest asks for a page table to be updated? | |
773 | * | |
774 | * We already saw that demand_page() will fill in the shadow page tables when | |
775 | * needed, so we can simply remove shadow page table entries whenever the Guest | |
776 | * tells us they've changed. When the Guest tries to use the new entry it will | |
777 | * fault and demand_page() will fix it up. | |
778 | * | |
779 | * So with that in mind here's our code to to update a (top-level) PGD entry: | |
780 | */ | |
ebe0ba84 | 781 | void guest_set_pgd(struct lguest *lg, unsigned long gpgdir, u32 idx) |
d7e28ffe RR |
782 | { |
783 | int pgdir; | |
784 | ||
785 | if (idx >= SWITCHER_PGD_INDEX) | |
786 | return; | |
787 | ||
bff672e6 | 788 | /* If they're talking about a page table we have a shadow for... */ |
ee3db0f2 | 789 | pgdir = find_pgdir(lg, gpgdir); |
d7e28ffe | 790 | if (pgdir < ARRAY_SIZE(lg->pgdirs)) |
bff672e6 | 791 | /* ... throw it away. */ |
90603d15 | 792 | release_pgd(lg->pgdirs[pgdir].pgdir + idx); |
d7e28ffe | 793 | } |
acdd0b62 MZ |
794 | #ifdef CONFIG_X86_PAE |
795 | void guest_set_pmd(struct lguest *lg, unsigned long pmdp, u32 idx) | |
796 | { | |
797 | guest_pagetable_clear_all(&lg->cpus[0]); | |
798 | } | |
799 | #endif | |
d7e28ffe | 800 | |
58a24566 MZ |
801 | /* Once we know how much memory we have we can construct simple identity |
802 | * (which set virtual == physical) and linear mappings | |
803 | * which will get the Guest far enough into the boot to create its own. | |
804 | * | |
805 | * We lay them out of the way, just below the initrd (which is why we need to | |
806 | * know its size here). */ | |
807 | static unsigned long setup_pagetables(struct lguest *lg, | |
808 | unsigned long mem, | |
809 | unsigned long initrd_size) | |
810 | { | |
811 | pgd_t __user *pgdir; | |
812 | pte_t __user *linear; | |
58a24566 | 813 | unsigned long mem_base = (unsigned long)lg->mem_base; |
acdd0b62 MZ |
814 | unsigned int mapped_pages, i, linear_pages; |
815 | #ifdef CONFIG_X86_PAE | |
816 | pmd_t __user *pmds; | |
817 | unsigned int j; | |
818 | pgd_t pgd; | |
819 | pmd_t pmd; | |
820 | #else | |
821 | unsigned int phys_linear; | |
822 | #endif | |
58a24566 MZ |
823 | |
824 | /* We have mapped_pages frames to map, so we need | |
825 | * linear_pages page tables to map them. */ | |
826 | mapped_pages = mem / PAGE_SIZE; | |
827 | linear_pages = (mapped_pages + PTRS_PER_PTE - 1) / PTRS_PER_PTE; | |
828 | ||
829 | /* We put the toplevel page directory page at the top of memory. */ | |
830 | pgdir = (pgd_t *)(mem + mem_base - initrd_size - PAGE_SIZE); | |
831 | ||
832 | /* Now we use the next linear_pages pages as pte pages */ | |
833 | linear = (void *)pgdir - linear_pages * PAGE_SIZE; | |
834 | ||
acdd0b62 MZ |
835 | #ifdef CONFIG_X86_PAE |
836 | pmds = (void *)linear - PAGE_SIZE; | |
837 | #endif | |
58a24566 MZ |
838 | /* Linear mapping is easy: put every page's address into the |
839 | * mapping in order. */ | |
840 | for (i = 0; i < mapped_pages; i++) { | |
841 | pte_t pte; | |
842 | pte = pfn_pte(i, __pgprot(_PAGE_PRESENT|_PAGE_RW|_PAGE_USER)); | |
843 | if (copy_to_user(&linear[i], &pte, sizeof(pte)) != 0) | |
844 | return -EFAULT; | |
845 | } | |
846 | ||
847 | /* The top level points to the linear page table pages above. | |
848 | * We setup the identity and linear mappings here. */ | |
acdd0b62 MZ |
849 | #ifdef CONFIG_X86_PAE |
850 | for (i = 0, j; i < mapped_pages && j < PTRS_PER_PMD; | |
851 | i += PTRS_PER_PTE, j++) { | |
852 | native_set_pmd(&pmd, __pmd(((unsigned long)(linear + i) | |
853 | - mem_base) | _PAGE_PRESENT | _PAGE_RW | _PAGE_USER)); | |
854 | ||
855 | if (copy_to_user(&pmds[j], &pmd, sizeof(pmd)) != 0) | |
856 | return -EFAULT; | |
857 | } | |
858 | ||
859 | set_pgd(&pgd, __pgd(((u32)pmds - mem_base) | _PAGE_PRESENT)); | |
860 | if (copy_to_user(&pgdir[0], &pgd, sizeof(pgd)) != 0) | |
861 | return -EFAULT; | |
862 | if (copy_to_user(&pgdir[3], &pgd, sizeof(pgd)) != 0) | |
863 | return -EFAULT; | |
864 | #else | |
58a24566 MZ |
865 | phys_linear = (unsigned long)linear - mem_base; |
866 | for (i = 0; i < mapped_pages; i += PTRS_PER_PTE) { | |
867 | pgd_t pgd; | |
868 | pgd = __pgd((phys_linear + i * sizeof(pte_t)) | | |
869 | (_PAGE_PRESENT | _PAGE_RW | _PAGE_USER)); | |
870 | ||
871 | if (copy_to_user(&pgdir[i / PTRS_PER_PTE], &pgd, sizeof(pgd)) | |
872 | || copy_to_user(&pgdir[pgd_index(PAGE_OFFSET) | |
873 | + i / PTRS_PER_PTE], | |
874 | &pgd, sizeof(pgd))) | |
875 | return -EFAULT; | |
876 | } | |
acdd0b62 | 877 | #endif |
58a24566 MZ |
878 | |
879 | /* We return the top level (guest-physical) address: remember where | |
880 | * this is. */ | |
881 | return (unsigned long)pgdir - mem_base; | |
882 | } | |
883 | ||
bff672e6 RR |
884 | /*H:500 (vii) Setting up the page tables initially. |
885 | * | |
886 | * When a Guest is first created, the Launcher tells us where the toplevel of | |
887 | * its first page table is. We set some things up here: */ | |
58a24566 | 888 | int init_guest_pagetable(struct lguest *lg) |
d7e28ffe | 889 | { |
58a24566 MZ |
890 | u64 mem; |
891 | u32 initrd_size; | |
892 | struct boot_params __user *boot = (struct boot_params *)lg->mem_base; | |
acdd0b62 MZ |
893 | #ifdef CONFIG_X86_PAE |
894 | pgd_t *pgd; | |
895 | pmd_t *pmd_table; | |
896 | #endif | |
58a24566 MZ |
897 | /* Get the Guest memory size and the ramdisk size from the boot header |
898 | * located at lg->mem_base (Guest address 0). */ | |
899 | if (copy_from_user(&mem, &boot->e820_map[0].size, sizeof(mem)) | |
900 | || get_user(initrd_size, &boot->hdr.ramdisk_size)) | |
901 | return -EFAULT; | |
902 | ||
bff672e6 RR |
903 | /* We start on the first shadow page table, and give it a blank PGD |
904 | * page. */ | |
58a24566 MZ |
905 | lg->pgdirs[0].gpgdir = setup_pagetables(lg, mem, initrd_size); |
906 | if (IS_ERR_VALUE(lg->pgdirs[0].gpgdir)) | |
907 | return lg->pgdirs[0].gpgdir; | |
1713608f GOC |
908 | lg->pgdirs[0].pgdir = (pgd_t *)get_zeroed_page(GFP_KERNEL); |
909 | if (!lg->pgdirs[0].pgdir) | |
d7e28ffe | 910 | return -ENOMEM; |
acdd0b62 MZ |
911 | #ifdef CONFIG_X86_PAE |
912 | pgd = lg->pgdirs[0].pgdir; | |
913 | pmd_table = (pmd_t *) get_zeroed_page(GFP_KERNEL); | |
914 | if (!pmd_table) | |
915 | return -ENOMEM; | |
916 | ||
917 | set_pgd(pgd + SWITCHER_PGD_INDEX, | |
918 | __pgd(__pa(pmd_table) | _PAGE_PRESENT)); | |
919 | #endif | |
1713608f | 920 | lg->cpus[0].cpu_pgd = 0; |
d7e28ffe RR |
921 | return 0; |
922 | } | |
923 | ||
47436aa4 | 924 | /* When the Guest calls LHCALL_LGUEST_INIT we do more setup. */ |
382ac6b3 | 925 | void page_table_guest_data_init(struct lg_cpu *cpu) |
47436aa4 RR |
926 | { |
927 | /* We get the kernel address: above this is all kernel memory. */ | |
382ac6b3 | 928 | if (get_user(cpu->lg->kernel_address, |
acdd0b62 MZ |
929 | &cpu->lg->lguest_data->kernel_address) |
930 | /* We tell the Guest that it can't use the top 2 or 4 MB | |
931 | * of virtual addresses used by the Switcher. */ | |
932 | || put_user(RESERVE_MEM * 1024 * 1024, | |
933 | &cpu->lg->lguest_data->reserve_mem) | |
934 | || put_user(cpu->lg->pgdirs[0].gpgdir, | |
935 | &cpu->lg->lguest_data->pgdir)) | |
382ac6b3 | 936 | kill_guest(cpu, "bad guest page %p", cpu->lg->lguest_data); |
47436aa4 RR |
937 | |
938 | /* In flush_user_mappings() we loop from 0 to | |
939 | * "pgd_index(lg->kernel_address)". This assumes it won't hit the | |
940 | * Switcher mappings, so check that now. */ | |
acdd0b62 MZ |
941 | #ifdef CONFIG_X86_PAE |
942 | if (pgd_index(cpu->lg->kernel_address) == SWITCHER_PGD_INDEX && | |
943 | pmd_index(cpu->lg->kernel_address) == SWITCHER_PMD_INDEX) | |
944 | #else | |
382ac6b3 | 945 | if (pgd_index(cpu->lg->kernel_address) >= SWITCHER_PGD_INDEX) |
acdd0b62 | 946 | #endif |
382ac6b3 GOC |
947 | kill_guest(cpu, "bad kernel address %#lx", |
948 | cpu->lg->kernel_address); | |
47436aa4 RR |
949 | } |
950 | ||
bff672e6 | 951 | /* When a Guest dies, our cleanup is fairly simple. */ |
d7e28ffe RR |
952 | void free_guest_pagetable(struct lguest *lg) |
953 | { | |
954 | unsigned int i; | |
955 | ||
bff672e6 | 956 | /* Throw away all page table pages. */ |
d7e28ffe | 957 | release_all_pagetables(lg); |
bff672e6 | 958 | /* Now free the top levels: free_page() can handle 0 just fine. */ |
d7e28ffe RR |
959 | for (i = 0; i < ARRAY_SIZE(lg->pgdirs); i++) |
960 | free_page((long)lg->pgdirs[i].pgdir); | |
961 | } | |
962 | ||
bff672e6 RR |
963 | /*H:480 (vi) Mapping the Switcher when the Guest is about to run. |
964 | * | |
e1e72965 | 965 | * The Switcher and the two pages for this CPU need to be visible in the |
bff672e6 | 966 | * Guest (and not the pages for other CPUs). We have the appropriate PTE pages |
e1e72965 RR |
967 | * for each CPU already set up, we just need to hook them in now we know which |
968 | * Guest is about to run on this CPU. */ | |
0c78441c | 969 | void map_switcher_in_guest(struct lg_cpu *cpu, struct lguest_pages *pages) |
d7e28ffe | 970 | { |
df29f43e | 971 | pte_t *switcher_pte_page = __get_cpu_var(switcher_pte_pages); |
df29f43e | 972 | pte_t regs_pte; |
a53a35a8 | 973 | unsigned long pfn; |
d7e28ffe | 974 | |
acdd0b62 MZ |
975 | #ifdef CONFIG_X86_PAE |
976 | pmd_t switcher_pmd; | |
977 | pmd_t *pmd_table; | |
978 | ||
979 | native_set_pmd(&switcher_pmd, pfn_pmd(__pa(switcher_pte_page) >> | |
980 | PAGE_SHIFT, PAGE_KERNEL_EXEC)); | |
981 | ||
982 | pmd_table = __va(pgd_pfn(cpu->lg-> | |
983 | pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX]) | |
984 | << PAGE_SHIFT); | |
985 | native_set_pmd(&pmd_table[SWITCHER_PMD_INDEX], switcher_pmd); | |
986 | #else | |
987 | pgd_t switcher_pgd; | |
988 | ||
bff672e6 RR |
989 | /* Make the last PGD entry for this Guest point to the Switcher's PTE |
990 | * page for this CPU (with appropriate flags). */ | |
ed1dc778 | 991 | switcher_pgd = __pgd(__pa(switcher_pte_page) | __PAGE_KERNEL_EXEC); |
df29f43e | 992 | |
1713608f | 993 | cpu->lg->pgdirs[cpu->cpu_pgd].pgdir[SWITCHER_PGD_INDEX] = switcher_pgd; |
d7e28ffe | 994 | |
acdd0b62 | 995 | #endif |
bff672e6 RR |
996 | /* We also change the Switcher PTE page. When we're running the Guest, |
997 | * we want the Guest's "regs" page to appear where the first Switcher | |
998 | * page for this CPU is. This is an optimization: when the Switcher | |
999 | * saves the Guest registers, it saves them into the first page of this | |
1000 | * CPU's "struct lguest_pages": if we make sure the Guest's register | |
1001 | * page is already mapped there, we don't have to copy them out | |
1002 | * again. */ | |
a53a35a8 | 1003 | pfn = __pa(cpu->regs_page) >> PAGE_SHIFT; |
90603d15 MZ |
1004 | native_set_pte(®s_pte, pfn_pte(pfn, PAGE_KERNEL)); |
1005 | native_set_pte(&switcher_pte_page[pte_index((unsigned long)pages)], | |
1006 | regs_pte); | |
d7e28ffe | 1007 | } |
bff672e6 | 1008 | /*:*/ |
d7e28ffe RR |
1009 | |
1010 | static void free_switcher_pte_pages(void) | |
1011 | { | |
1012 | unsigned int i; | |
1013 | ||
1014 | for_each_possible_cpu(i) | |
1015 | free_page((long)switcher_pte_page(i)); | |
1016 | } | |
1017 | ||
bff672e6 RR |
1018 | /*H:520 Setting up the Switcher PTE page for given CPU is fairly easy, given |
1019 | * the CPU number and the "struct page"s for the Switcher code itself. | |
1020 | * | |
1021 | * Currently the Switcher is less than a page long, so "pages" is always 1. */ | |
d7e28ffe RR |
1022 | static __init void populate_switcher_pte_page(unsigned int cpu, |
1023 | struct page *switcher_page[], | |
1024 | unsigned int pages) | |
1025 | { | |
1026 | unsigned int i; | |
df29f43e | 1027 | pte_t *pte = switcher_pte_page(cpu); |
d7e28ffe | 1028 | |
bff672e6 | 1029 | /* The first entries are easy: they map the Switcher code. */ |
d7e28ffe | 1030 | for (i = 0; i < pages; i++) { |
90603d15 MZ |
1031 | native_set_pte(&pte[i], mk_pte(switcher_page[i], |
1032 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED))); | |
d7e28ffe RR |
1033 | } |
1034 | ||
bff672e6 | 1035 | /* The only other thing we map is this CPU's pair of pages. */ |
d7e28ffe RR |
1036 | i = pages + cpu*2; |
1037 | ||
bff672e6 | 1038 | /* First page (Guest registers) is writable from the Guest */ |
90603d15 MZ |
1039 | native_set_pte(&pte[i], pfn_pte(page_to_pfn(switcher_page[i]), |
1040 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED|_PAGE_RW))); | |
df29f43e | 1041 | |
bff672e6 RR |
1042 | /* The second page contains the "struct lguest_ro_state", and is |
1043 | * read-only. */ | |
90603d15 MZ |
1044 | native_set_pte(&pte[i+1], pfn_pte(page_to_pfn(switcher_page[i+1]), |
1045 | __pgprot(_PAGE_PRESENT|_PAGE_ACCESSED))); | |
d7e28ffe RR |
1046 | } |
1047 | ||
e1e72965 RR |
1048 | /* We've made it through the page table code. Perhaps our tired brains are |
1049 | * still processing the details, or perhaps we're simply glad it's over. | |
1050 | * | |
a6bd8e13 RR |
1051 | * If nothing else, note that all this complexity in juggling shadow page tables |
1052 | * in sync with the Guest's page tables is for one reason: for most Guests this | |
1053 | * page table dance determines how bad performance will be. This is why Xen | |
1054 | * uses exotic direct Guest pagetable manipulation, and why both Intel and AMD | |
1055 | * have implemented shadow page table support directly into hardware. | |
e1e72965 RR |
1056 | * |
1057 | * There is just one file remaining in the Host. */ | |
1058 | ||
bff672e6 RR |
1059 | /*H:510 At boot or module load time, init_pagetables() allocates and populates |
1060 | * the Switcher PTE page for each CPU. */ | |
d7e28ffe RR |
1061 | __init int init_pagetables(struct page **switcher_page, unsigned int pages) |
1062 | { | |
1063 | unsigned int i; | |
1064 | ||
1065 | for_each_possible_cpu(i) { | |
df29f43e | 1066 | switcher_pte_page(i) = (pte_t *)get_zeroed_page(GFP_KERNEL); |
d7e28ffe RR |
1067 | if (!switcher_pte_page(i)) { |
1068 | free_switcher_pte_pages(); | |
1069 | return -ENOMEM; | |
1070 | } | |
1071 | populate_switcher_pte_page(i, switcher_page, pages); | |
1072 | } | |
1073 | return 0; | |
1074 | } | |
bff672e6 | 1075 | /*:*/ |
d7e28ffe | 1076 | |
bff672e6 | 1077 | /* Cleaning up simply involves freeing the PTE page for each CPU. */ |
d7e28ffe RR |
1078 | void free_pagetables(void) |
1079 | { | |
1080 | free_switcher_pte_pages(); | |
1081 | } |